Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
Orthogonal Frequency
Division Multiplexing
Transmitters
BY
Gavin Hill
Thesis submitted in fulfilment of the
requirements for the degree of Doctor of
Philosophy
Victoria University of Technology
School of Communications & Informatics
March 2011
i
Abstract
Orthogonal Frequency Division Multiplexing (OFDM) is a digital transmission
method developed to meet the increasing demand for higher data rates in
communications which can be used in both wired and wireless environments. This
thesis describes the issue of the Peak to Average Power Ratio (PAPR) in OFDM
which is a major drawback, and presents new and variations to existing algorithms to
reduce it.
Initially the theoretical principles behind OFDM are discussed elaborating on the
advantages and disadvantages of OFDM. This is followed by analysis of the PAPR in
OFDM where it is shown through theoretical analysis and simulation that the
occurrence of large peaks in OFDM is actually quite rare. The effect on system
performance in terms of the Bit Error Rate (BER) and Power Spectral Density (PSD)
is simulated for an OFDM transceiver with a saturated High Power Amplifier. This is
followed by a study of published PAPR reduction methods
The first contribution is a low complexity variation of Partial Transmit Sequences
(PTS). In PTS several alternate transmit signals are seeded from the same source,
each alternate transmit signal has a reversible and different phase rotation performed
on the data. The transmit signal with the lowest PAPR is chosen for transmission. In
novel variations, called Cyclic Shifted Sequences (CSS) and Time Inversion (TI),
different shifts of the data are performed which avoid the need for complex
multiplications. In certain cases a whole IFFT operation can be removed with a
negligible effect on performance when CSS is combined with PTS. Furthermore it is
shown that the peak regrowth of TI and CSS after pulse shaping filtering is
considerably less than for PTS.
Next, new clipping techniques are presented which reduce substantially the
complexity of clipping algorithms by using novel methods to calculate the magnitude,
avoiding the use of multiplications. One method, called Sector clipping uses a rule
base to clip the signal, dividing the clipping region into a series of sectors. When the
rule base is expanded to include more sectors the performance is shown to approach
ii
more complex existing clipping methods. This algorithm is implemented in silicon in
a 3 metal layer 0.5µ process. Another clipping scheme called Vector Subtraction is a
variation of another low complexity magnitude estimate method which further
reduces complexity by alleviating the need for a scaling operation. The performance
of the new methods was ascertained through simulation of a whole OFDM transceiver
chain and shown to have relative BER’s. Finally Vector Subtraction was
implemented in a previously proposed clip and filter algorithm where its low latency
and accuracy proved it to be suitable for the algorithm.
iii
Acknowledgements
First of all I wish to thank Professor Mike Faulkner for his guidance, patience, advice,
and time thorough my many years of research. I owe the completion of this thesis to
his efforts.
I also wish to thank Dr Reza Berangi for his mathematical and simulation knowledge
and his willingness to share it. Special thanks goes to the post graduate studies
committee for the numerous second chances to complete my research and to Shirley
Herewyn for her omnipotent knowledge on dealing with university bureaucracy.
I would also like to acknowledge my research colleagues for the discussions on
everything from communications theory to ancient history. In particular I wish to
thank Abdi Waheed Mohammed, Tuan Nguyen, Andrew Mancuso, Trung Nguyen,
Olivia Hu (who never really saw the significance of being Dr ‘Who’, even after I
brought in pictures of Tom Baker and the Tardis) and Melvyn Pereira for their
company and generous natures.
Finally I wish to thank my parents and my brother for their support, both financially
and emotionally, and my best friends Dan, Spencer, and Mark for the good times and
social outlet.
Gavin Hill
7 February 2006
v
Contents
Abstract .......................................................................................................................... i
Acknowledgements ..................................................................................................... iii
Contents ........................................................................................................................ v
List of Tables .............................................................................................................. vii
List of Figures ............................................................................................................ viii
Acronyms and Symbols ............................................................................................ xvi
Chapter 1 Introduction ................................................................................................ 1
1.1 Structure of thesis .............................................................................................. 2
1.1.1 Contributions ................................................................................................. 4
Chapter 2 Orthogonal Frequency Division Multiplexing ........................................ 6
2.1 History of multicarrier networks .................................................................... 6
2.2 Multicarrier principle ........................................................................................ 8
2.3 OFDM implementation of multicarrier modulation ...................................... 8
2.3.1 Use of Fourier Transform for modulation and demodulation ..................... 11
2.3.2 Orthogonality in OFDM ............................................................................. 13
2.4 OFDM transmission over time varying channels ......................................... 14
2.4.1 Multipath propagation ................................................................................. 15
2.4.3 Frequency selective fading ......................................................................... 17
2.4.4 Equalization ................................................................................................ 18
2.5 Limitations in OFDM ...................................................................................... 18
2.5.1 Synchronization .......................................................................................... 19
2.5.1.1 Timing errors ................................................................................................................ 19
2.5.1.2 Carrier phase noise ....................................................................................................... 20
2.5.1.3 Frequency errors .......................................................................................................... 20
2.5.2 Non linearities ............................................................................................. 22
2.6 Applications of OFDM .................................................................................... 22
2.6.1 COFDM ...................................................................................................... 22
2.6.2 Digital Audio Broadcasting ........................................................................ 23
2.6.3 Digital Video Broadcasting ......................................................................... 24
2.6.4 HiperLan2/802.11a ..................................................................................... 24
2.6.5 ADSL .......................................................................................................... 25
2.6.6 MIMO OFDM ............................................................................................. 26
2.7 Conclusion ........................................................................................................ 26
Chapter 3 Peak to Average Power in OFDM .......................................................... 28
3.1 Peak to Average Power Ratio ......................................................................... 29
3.2 Statistical distribution of OFDM samples ..................................................... 31
3.3 Oversampling discrete OFDM symbols to find true (continuous) peaks.... 35
3.4 Effect of Non Linearity on OFDM ................................................................. 45
3.4.1 Description of memoryless Non Linearity .................................................. 45
3.4.2 Impact on Power Spectral Density .............................................................. 50
3.4.3 Impact on Bit Error Rate ............................................................................. 51
3.5 Conclusion ........................................................................................................ 53
Chapter 4 Peak to Average Power Solutions Distortionless Techniques ............ 55
4.1 Coding techniques ............................................................................................ 56
4.1.1 Block Codes ................................................................................................ 57
4.1.2 Bounds on PAPR ........................................................................................ 61
vi
4.1.3 Cyclic Codes ............................................................................................... 61
4.1.4 ShapiroRudin codes ................................................................................... 62
4.1.5 Golay complementary codes ....................................................................... 63
4.1.6 ReedMuller Codes ..................................................................................... 63
4.2 Multiple Signal Representation ...................................................................... 66
4.2.1 Partial Transmit Signals .............................................................................. 66
4.2.2 Oversampling PTS ...................................................................................... 72
4.2.3 Selective Mapping ...................................................................................... 77
4.3 Tone Reservation/Injection ............................................................................. 83
4.3.1 Tone Reservation ........................................................................................ 83
4.3.2 Tone Injection ............................................................................................. 87
4.4 Conclusion ........................................................................................................ 89
Chapter 5 Cyclic Shifted Sequences and Time Inversion of PTS .......................... 92
5.1 PTS subblock creation ..................................................................................... 92
5.2 New techniques for PTS subblock creation ................................................... 93
5.2.1 Cyclic Shifted Sequences ............................................................................ 93
5.2.2 PTS with CSS ............................................................................................. 96
5.2.3 Time Inversion ............................................................................................ 98
5.3 Filtering new techniques .................................................................................. 99
5.4 Oversampling new techniques ...................................................................... 104
5.5 Complexity evaluation ................................................................................... 108
5.6 Conclusion ...................................................................................................... 109
Chapter 6 Peak to Average Power Solutions Distorted Techniques .................. 111
6.1 Clipping in the Baseband .............................................................................. 112
6.1.1 Quantisation and Clipping ........................................................................ 121
6.2 Amplifier non linearities ................................................................................ 124
6.3 Windowing ...................................................................................................... 128
6.4 Conclusion ...................................................................................................... 130
Chapter 7 Reduced Complexity Clipping Algorithms ......................................... 132
7.1 Conventional clipping .................................................................................... 134
7.2 New Sector Clipping method ........................................................................ 143
7.2.1 Theoretical Analysis of Clipping Techniques ........................................... 143
7.2.1.1 SNR Analysis ............................................................................................................... 144
7.2.1.2 Conventional clipping ................................................................................................. 146
7.2.1.3 Sector Clipping ........................................................................................................... 149
7.2.1.4 Square Clipping .......................................................................................................... 158
7.2.1.5 Theoretical Results ..................................................................................................... 158
7.2.2 Extensions of Sector Clipping .................................................................. 160
7.2.3 Hardware Implementation ........................................................................ 162
7.3 New Vector Subtraction clipping method .................................................... 167
7.3.1 Lucent Algorithm ...................................................................................... 167
7.3.2 Vector Subtraction .................................................................................... 168
7.4 Comparison of new and existing clipping methods .................................... 173
7.4.1 BERF ......................................................................................................... 173
7.4.2 PSD results ................................................................................................ 184
7.5 New adaptive clipping method ...................................................................... 187
7.6 Conclusion ...................................................................................................... 196
Chapter 8 Conclusion .............................................................................................. 198
8.1 Future Work ................................................................................................... 200
Bibliography ............................................................................................................. 201
vii
List of Tables
2.1: DAB parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2: DVB system parameters for 2K mode . . . . . . . . . . . . . . . . . . . . . 24
2.3: HIPERLAN2 parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4: Data rates for HIPERLAN2 . . . . . . . . . . . . . . . . . . . . . . . . . . .25
4.1: Number of complex multiplications and magnitude operations required . . . 76
5.1: Complexity comparison of various techniques (adjacent partitioning) . . . . 109
7.1: Truth Table for 3 Sector clipping . . . . . . . . . . . . . . . . . . . . . . . .155
7.2: ‘Synopsys’ reports for Square and 3 sector clipping. . . . . . . . . . . . . . 164
7.3: ‘Cadence’ area utilization report on Square clipping . . . . . . . . . . . . . .165
7.4: ‘Cadence’ area utilization report on 3 Sector clipping . . . . . . . . . . . . 166
7.5: Baseband clip level required to maintain a BER=10
4
at varying IBO in
HPA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
viii
List of Figures
2.1: Block diagram of a basic multicarrier system . . . . . . . . . . . . . . . . . . . 9
2.2: Basic OFDM transmitter and receiver. Occupied frequency band shown in
between. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3: Basic OFDM transmitter and receiver pair utilizing Fourier transform. . . . . . 12
2.4: Frequency spectrum of 5 orthogonal subcarriers of an OFDM transmit signal . 14
2.5: OFDM symbol a) without cyclic prefix, and b) with cyclic prefix. . . . . . . . 16
2.6: Time and frequency properties of single carrier and OFDM techniques. . . . . 19
2.7: Effects of frequency offset F: reduction of signal amplitude (star), and ICI
(circle). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1: Simulated envelope for OFDM system (N=64) normalized by average power. 33
3.2: Simulated (solid line) and theoretical (3.13, dashed line) OFDM symbol CCDF
for N=32, 64, 128, and 256 subcarriers. QPSK, 30000 runs. . . . . . . . . . 33
3.3: Simulated OFDM sample CCDF for N=32, 64, 128, and 256 subcarriers.
QPSK, 30000 runs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4: Simulated OFDM CCDF for M=4, 16, and 64 constellation mapping. N=64,
30000 runs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5: Simulated OFDM symbol with no oversampling (dashed) with it’s oversampled
version (solid) overlaid on top. The solid circle represents the 6dB level with
respect to the average power. N=64, oversampling factors are 1 and 8. . . . 36
3.6: Simulated OFDM CCDF for oversampling rates of 1, 2, 4, and 8. N=64, QPSK,
15000 runs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.7: Theoretical OFDM CCDF from (3.15) and (3.21) for N=64, 512 with simulated
results: QPSK, oversampling factor rate of 16, 20000 runs, N=64, 512. . . . 40
3.8: Theoretical OFDM CCDF from (3.22) for N=64, 512 with simulated results:
QPSK, oversampling factor rate of 16, 20000 runs, N=64, 512. . . . . . . . . 41
3.9: Simulated OFDM CCDF QPSK, 64 point IFFT, 256 tap RRCF, α=0.15.
Clipped at 3dB after IFFT, then filtered (solid line). No clipping, then filtered
(dashed line). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.10: Zero padding of the IFFT, null carriers are set in the middle of the input. . . 44
3.11: Simulated OFDM CCDF after IFFT (os=1 and 2) and after filtering (os=1
and 2). QPSK, N=64, 256 tap RRCF, α=0.15, 15000 runs. . . . . . . . . . 44
ix
3.12: AM/AM properties of a Soft Limiter (SL). . . . . . . . . . . . . . . . . . . .47
3.13: AM/AM properties of a Traveling Wave Tube Amplifier (TWTA). . . . . . .48
3.14: AM/AM properties of a Solid State Amplifier (SSPA) for different values
of P, and ideal amplifier. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.15: PSD of 64 subcarrier OFDM signal with 64 point IFFT, RRCF with excess
bandwidth of 0.15. A SSPA with P=3 and various backoffs is used. . . . . 51
3.16: Signal constellation at the output of the HPA, P=3 after non linear
amplification a) 4 QAM, 6dB IBO; b) 4 QAM, 0dB IBO; c) 16 QAM 6dB
IBO; and d) 16 QAM 0dB IBO. . . . . . . . . . . . . . . . . . . . . . . . . 52
3.17: BER of 64 subcarrier OFDM signal with 64 point IFFT, RRCF with excess
bandwidth of 0.15 for 4, 16, and 64 QAM constellations. A SSPA with
P=3 and various backoffs is used. . . . . . . . . . . . . . . . . . . . . . . . 53
4.1: Block diagram of OFDM transmitter showing PAPR coding. . . . . . . . . . 56
4.2: Block diagram of PAPR reduction using the PTS approach. . . . . . . . . . . 68
4.3: An example of the 3 main PTS structures: (a) Interleaved (b) Adjacent (c)
Pseudorandom. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.4: Simulated CCDF for PTSOFDM with V=2 and varying W. N=64, adjacent
subblock partitioning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .71
4.5: Simulated CCDF for PTSOFDM with W=4 and varying V. N=64, adjacent
subblock partitioning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .71
4.6: Generation of subblocks for PTS using Concatenated Pseudo Random Sub
Block Partition Scheme (CPR SPS). . . . . . . . . . . . . . . . . . . . . . . 73
4.7: Block diagram of OLSPTS transmitter. . . . . . . . . . . . . . . . . . . . . 76
4.8: Block diagram of an SLM OFDM transmitter. . . . . . . . . . . . . . . . . . 78
4.9: Simulated CCDF for SLMOFDM for varying values of U. N=64, os=1. . . .79
4.10: Simulated CCDF for SLMOFDM. U=1, 3, and 8. Oversampling rates
are 1, 2, and 4, and 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.11: SLM transmitter block diagram employing technique to avoid explicit
transmission of side information, a) serial form, b) parallel form. . . . . . . 81
4.12: Scrambler polynomial for new SLM technique. . . . . . . . . . . . . . . . . 81
4.13: a) Receiver structure of the proposed SLM system, b) Descrambler
polynomial at the receiver. . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.14: Block diagram of a Tone Reservation (TR) OFDM transceiver. . . . . . . . .84
4.15: Block diagram of a Tone Injection (TI) OFDM transceiver. . . . . . . . . . 87
x
4.16: Example of possible expansions of constellation in TI for 16 QAM. . . . . .88
5.1: Block diagram showing the simulation model of Section 5.2. . . . . . . . . .93
5.2: Block diagram of PTS, CSS and TI transceiver. . . . . . . . . . . . . . . . .95
5.3: Simulated CCDF for PTS (V=2, W=8; V=4, W=4; V=4, W=8), CSS
(V=2, S=8; V=4, S=4; V=4, S=8) and uncoded OFDM. N=64, adjacent
partitioning, with no oversampling in IFFT, (os=1). . . . . . . . . . . . . . . 96
5.4: Simulated CCDF for PTS (V=2, W=8; V=3, W=4), PTS/CSS (V=2, W=4,
S=2; V=2, W=4, S=4), and uncoded OFDM. N=64, adjacent partitioning,
with no oversampling in IFFT, (os=1). . . . . . . . . . . . . . . . . . . . . .98
5.5: Simulated CCDF for PTS (V=2, W=8), PTS/TI (V=2, W=4, S=2) and
uncoded OFDM. N=64, adjacent partitioning, with no oversampling in
the IFFT, (os=1). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .99
5.6: Block diagram showing the simulation model of Section 5.2. . . . . . . . . 100
5.7: Simulated CCDF of discrete and filtered PTS (V=4, W=4), CSS (V=4,
S=4), and uncoded OFDM. N=64, adjacent partitioning, with no
oversampling in IFFT (os=1), interpolated by 8, filtered with RCF
(α=0.15). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.8: Simulated CCDF of discrete and filtered PTS (V=4, W=8), CSS (V=4,
S=8), and uncoded OFDM. N=64, adjacent partitioning, with no
oversampling in IFFT (os=1), interpolated by 8, filtered with RCF
(α=0.15). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.9: Simulated CCDF of discrete and filtered PTS (V=2, W=8), PTS/CSS
(V=2, W=4, S=2), and PTS/TI (V=2, W=4, S=2), and uncoded OFDM.
N=64, adjacent partitioning, with no oversampling in IFFT (os=1),
interpolated by 8, filtered with RCF (α=0.15). . . . . . . . . . . . . . . . . 103
5.10: Simulated CCDF of discrete and filtered PTS (V=3, W=4), PTS/CSS
(V=2, W=4, S=4), and uncoded OFDM. N=64, adjacent partitioning,
with no oversampling in the IFFT (os=1), interpolated by 8, filtered
with RCF (α=0.15). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.11: Simulated CCDF for PTS/CSS (V=2, W=4, S=4), and Uncoded OFDM.
Discrete oversampled curves (dashed) move from left to right and
oversampled (solid) filtered curves move from right to left. N=64,
adjacent partitioning, with oversampling rates in the IFFT of 1, 2, 4,
and 8, interpolated by 8, and filtered with RCF (α=0.15). . . . . . . . . . .105
xi
5.12: Simulated CCDF for PTS (V=3, W=4), and Uncoded OFDM. Discrete
oversampled (dashed) curves move from left to right and oversampled
(solid) filtered curves move from right to left. N=64, adjacent
partitioning, with oversampling rates in the IFFT of 1, 2, 4, and 8,
interpolated by 8, and filtered with RCF (α=0.15). . . . . . . . . . . . . . 106
5.13: Simulated CCDF for PTS (V=2, W=8), and Uncoded OFDM. Discrete
oversampled curves (dashed) move from left to right and filtered (solid)
oversampled curves move from right to left. N=64, adjacent partitioning,
with oversampling rates in the IFFT of 1, 2, and 4, interpolated by 8,
and filtered with RCF (α=0.15). . . . . . . . . . . . . . . . . . . . . . . . 107
5.14: Simulated CCDF for PTS/CSS (V=2, W=4, S=2), and Uncoded OFDM.
Discrete oversampled curves (dashed) move from left to right and
filtered (solid) oversampled curves move from right to left. N=64,
adjacent partitioning, with oversampling rates in the IFFT of 1, 2, and
4, interpolated by 8, and filtered with RCF (α=0.15). . . . . . . . . . . . . 107
5.15: Simulated CCDF for PTS/TI (V=2, W=4, S=2), and Uncoded OFDM.
Discrete oversampled curves (dashed) move from left to right and
filtered (solid) oversampled curves move from right to left. N=64,
adjacent partitioning, with oversampling rates in the IFFT of 1, 2, and
4, interpolated by 8, and filtered with RCF (α=0.15). . . . . . . . . . . . . 108
6.1: Analytical symbol error probability from (6.10) and (6.12) for
64 QAM and N=64. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.2: Probability of clipping DMT signal as a function of µ for p=1,2,3.
N=64 subcarriers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.1: Average noise in the channel vs. BER for 4, 16, and 64 QAM . . . . . . . 133
7.2: IQ diagram showing conventional clipping region. . . . . . . . . . . . . . 134
7.3: Block diagram of simulation model used for clipping models. . . . . . . . . 135
7.4: Baseband clip level vs the BERF for varying RRCF parameters. 64
QAM symbols, 64 point IFFT, LPA, no channel impairments. AWGN=0 . 136
7.5a: Demapped constellation, M=64, with no clipping or channel
impairments. 64 filter taps in RRCF, α=0.15. . . . . . . . . . . . . . . . . 137
7.5b: Demapped constellation, M=64, with no clipping or channel
impairments. 128 filter taps in RRCF, α=0.15. . . . . . . . . . . . . . . . .137
7.6: Baseband clip level vs the BERF with varying IBO in HPA. 64 QAM
xii
symbols, 64 point IFFT (os=1), RRCF with 128 taps and α=0.15. . . . . . . 139
7.7: Baseband clip level vs. the BERF with varying IBO in HPA. 64 QAM
symbols, 128 point IFFT (os=2), RRCF with 128 taps and α=0.15. . . . . . 139
7.8: CCDF clipped in baseband at 5dB, HPA backoff set to 8dB for 64 and
128 IFFT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.9: Baseband clip level vs the BERF with varying Mary constellations.
128 point IFFT (os=2), RRCF with 128 taps and α=0.15. AWGN=0. . . . . 141
7.10: Baseband clip level vs the BERF with varying p in the SSPA. 128 point
IFFT (os=2), RRCF with 128 taps and α=0.15. HPA backoff set equal
to baseband clipping level. . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.11: I Q diagram showing different sector clipping regions and the direction
of data reduction for 3 Sector Clipping. . . . . . . . . . . . . . . . . . . . 144
7.12: Input output relationship of clipping operation. . . . . . . . . . . . . . . . 144
7.13: I Q diagram of 1st quadrant of a 3 Sector clipping system showing
the vector of an unclipped and clipped sample. . . . . . . . . . . . . . . . 150
7.14: Block diagram of 3 Sector clipping. . . . . . . . . . . . . . . . . . . . . . 156
7.15: Flowchart for the LUT in Figure 7.14 (3 Sector clipping) . . . . . . . . . . 157
7.16: Theoretical Clip level vs. SNR for Conventional (Standard), 3 Sector,
and Square clipping. Theoretical (dashed), simulated (solid) . . . . . . . . 159
7.17: Clipping angle, θ, vs. the SNR for 3 Sector clipping, based on (7.41). . . . .160
7.18: I Q diagram showing the 1st quadrant Sector clipping regions of a)
4 Sector clipping and b) 5 Sector clipping. . . . . . . . . . . . . . . . . . . 161
7.19: Simulated clip level vs. SNR for Conventional, Square, 3, 4, and 5
Sector Clipping. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.20: Design flow for silicon implementation of 3 Sector clipping. . . . . . . . . 163
7.21: Block diagram of 3 Sector clipping implemented in VHDL. . . . . . . . . 163
7.22: Application Specific Integrated Circuit (ASIC) view of 3 Sector
clipping algorithm implemented in Silicon using ‘Cadence Silicon
Ensemble’. 0.5µ process, 3 metal layers. . . . . . . . . . . . . . . . . . . 166
7.23: Block diagram of new Vector Subtraction scaling operation. . . . . . . . .169
7.24: IQ plane for Vector Subtraction showing vector of sample being
clipped. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .170
7.25: Simulated CCDF for Lucent patent [101] for various iterations clipped
xiii
at 5dB showing the leakage of under clipped samples. . . . . . . . . . . . .171
7.26: Simulated CCDF for Vector Subtraction for various iterations clipped
at 5dB showing the leakage of under clipped samples. . . . . . . . . . . . 171
7.27: Simulated clip level vs. SNR for Lucent clipping technique with
varying iterations, as well as Conventional and Square clipping. . . . . . . 172
7.28: Simulated clip level vs. SNR for Vector Subtraction clipping with
varying iterations, as well as Conventional and Square clipping. . . . . . . 173
7.29: Simulated 3, 4, and 5 Sector, Conventional and Square clipping vs.
BERF with a LPA. 64 QAM symbols, 128 point IFFT (os=2), RRCF
with 128 taps, and α=0.15, AWGN=0 . . . . . . . . . . . . . . . . . . . . 174
7.30: Simulated Vector Subtraction (1, 2, 3, and 4 iterations), Conventional
and Square clipping vs. BERF with a LPA. 64 QAM symbols, 128
point IFFT (os=2), RRCF with 128 taps, and α=0.15, AWGN=0 . . . . . . 175
7.31: Simulated Lucent [101] clipping (1, 2, 3, and 4 iterations), Conventional
and Square clipping vs. BERF with a LPA. 64 QAM symbols, 128
point IFFT (os=2), RRCF with 128 taps, and α=0.15. AWGN=0 . . . . . . 176
7.32: Simulated 3 Sector clipping, vs. BERF with varying IBO in HPA,
64 QAM symbols, 128 point IFFT (os=2), RRCF with 128 taps and
α=0.15. AWGN=0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.33: Simulated 4 Sector clipping, vs. BERF with varying IBO in HPA
64 QAM symbols, 128 point IFFT (os=2), RRCF with 128 taps and
α=0.15. AWGN=0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.34: Simulated 5 Sector clipping, vs. BERF with varying IBO in HPA,
16 QAM symbols, 128 point IFFT (os=2), RRCF with 128 taps and
α=0.15. AWGN=0. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.35: Simulated Vector Subtraction (1 iteration), vs. BERF with varying
IBO in HPA. 64 QAM symbols, 128 point IFFT (os=2), RRCF
with 128 taps and α=0.15. AWGN=0. . . . . . . . . . . . . . . . . . . . .179
7.36: Simulated Vector Subtraction (4 iterations), vs. BERF with varying
IBO in HPA. 64 QAM symbols, 128 point IFFT (os=2), RRCF
with 128 taps and α=0.15. AWGN=0 . . . . . . . . . . . . . . . . . . . . 179
7.37: Simulated Lucent clipping (1 iteration), vs. BERF with varying IBO in
HPA. 64 QAM symbols, 128 point IFFT (os=2), RRCF with 128 taps,
and α=0.15. AWGN=0. . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
xiv
7.38: Simulated Lucent clipping (4 iterations), vs. BERF with varying IBO in
HPA. 64 QAM symbols, 128 point IFFT (os=2), RRCF
with 128 taps, and α=0.15. AWGN=0. . . . . . . . . . . . . . . . . . . . 181
7.39: Simulated Square clipping vs. BERF with varying IBO in HPA.
64 QAM symbols, 128 point IFFT (os=2), RRCF with 128 taps,
and α=0.15. AWGN=0. . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.40: PSD 0dB clipping in baseband with increasing amplifier backoffs
(HPA=CL, CL+1,CL+2,CL+3,CL+4, LPA). RRCF, alpha=0.15,
128 taps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.41: PSD 5dB clipping in baseband with increasing amplifier backoffs
(HPA=CL, CL+1,CL+2,CL+3,CL+4, LPA). RRCF, alpha=0.15,
128 taps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.42: PSD after 0dB clipping in baseband and receiver filtering, with
increasing amplifier backoffs (HPA=CL, CL+1,CL+2,CL+3,
CL+4, LPA). RRCF, alpha=0.15, 128 taps. . . . . . . . . . . . . . . . . . 186
7.43: PSD after 5dB clipping in baseband and receiver filtering, with
increasing amplifier backoffs (HPA=CL, CL+1,CL+2,CL+3,CL+4,
LPA). RRCF, alpha=0.15, 128 taps. . . . . . . . . . . . . . . . . . . . . .186
7.44: Block diagram of the Level Detection Algorithm (LDA). . . . . . . . . . .189
7.45: Detailed block diagram of the LDA. . . . . . . . . . . . . . . . . . . . . 189
7.46: a) Amplitude of zero padded input to filter b) amplitude of filtered
output c) LDA correction vectors. . . . . . . . . . . . . . . . . . . . . . . 190
7.47: Vector representation of filtering with 3 active taps and the required
correction vector to bring the output back to CL. . . . . . . . . . . . . . . 191
7.48: Block diagram of the simulation model used to evaluate LDA. . . . . . . .192
7.49: In band distortion for Conventional and Vector Subtraction (2 iter) with
64 and 128 taps in compensation filter. ACL=5dB. . . . . . . . . . . . . 193
7.50: In band distortion for Conventional and Vector Subtraction (2 iter) with
64 and 128 taps in compensation filter. ACL=8dB. . . . . . . . . . . . . 193
7.51: Clipping error vs. inband distortion. Performance curves for LDA
using Conventional clipping and Vector Subtraction with 2
iterations. Compensation filter has 64 and 128 taps. ACL=5dB. . . . 195
7.52: Clipping error vs. inband distortion. Performance curves for LDA
using Conventional clipping and Vector Subtraction with 2
xv
iterations. Compensation filter has 64 and 128 taps. ACL=8dB. . . . 195
xvi
Acronyms and Symbols
ADC Analog to Digital Converter
ACI Adjacent Channel Interference
ADSL Asymmetric Digital Subscriber Line
AGC Automatic Gain Control
AM/AM Amplitude Modulation to Amplitude Modulation
AM/PM Amplitude Modulation to Phase Modulation
AWGN Additive White Gaussian Noise
BER Bit Error Rate
BPSK Binary Phase Shift Keying
BSLM Blind Selected Mapping
BW Bandwidth
CCDF Complementary Cumulative Density Function
CCOFDM Combined Coded OFDM
CDF Cumulative Density Function
CDMA Code Division Multiple Access
CF Crest Factor
CL Clip Level
COFDM Coded Orthogonal Division Multiplexing
CP Cyclic Prefix
CPRSPS Concatenated Pseudo Random Subblock Partition Scheme
CSS Cyclic Shifted Sequences
DAB Digital Audio Broadcasting
DAC Digital to Analog Conversion
DAR Decision Aided Reconstruction
DFT Discrete Fourier Transform
DMT Discrete MultiTone
DQSK Differential Quadrature Phase Shift Keying
DRL Data Rate Loss
DSP Digital Signal Processing
DVB Digital Video Broadcasting
EDGE Enhanced Data rates for Global Evolution
xvii
FEC Forward Error Correction
FDM Frequency Division Multiplexing
FFT Fast Fourier Transform
FPGA Field Programmable Gate Array
GSM Global System Mobile
HDDIVINE High DefinitionDigital Video Narrowband Emission
HIPERLAN2 HiPERformance Local Area Network version 2
HF High Frequency
HPA High Power Amplifier
IBO Input BackOff
ICI Inter Channel Interference
IDFT Inverse Discrete Fourier Transform
IEEE Institute of Electrical and Electronic Engineering
IFFT Inverse Fast Fourier Transform
ISI Inter Symbol Interference
LAN Local Area Network
LDA Level Detection Algorithm
LP Linear program
LPA Linear Power Amplifier
LUT Look Up Table
MIMO Multiple Input Multiple Output
MLD Maximum Likelihood Detection
MMSE Minimum Mean Squared Error
MSE Mean Squared Error
MSR Multiple Signal Representation
MQAM Mary Quadrature Amplitude Modulation
OBO Output BackOff
OBR Out of Band Radiation
OLS Optimal Limited Search
PAP Peak to Average Power
PAPR Peak to Average Power Ratio
PDF Power Density Function
PEP Peak Envelope Power
PICR PeakIntercarrierto Carrier Interference
xviii
PMEPR Peak to Mean Envelope Power Ratio
PRT Peak Reduction Tones
PSD Power Spectral Density
PSK Phase Shift Keying
P/S Parallel to Serial
PTS Partial Transmit Sequences
QAM Quadrature Amplitude Modulation
QCQP Quadratically Constrained Quadratic Program
QPSK Quadrature Phase Shift Keying
RF Radio Frequency
RMS Root Mean Square
RRCF Root Raised Cosine Filter
RS ReedSolomon
SBC Sub Block Coding
SC Single Carrier
SER Symbol Error Rate
SES Suboptimal Exhaustive Search
SL Soft Limiter
SLM Selected Mapping
SNR Signal to Noise Ratio
S/P Serial to Parallel
SSPA Solid State Power Amplifier
STERNE System de Television En Radiodiffusion NumeriquE
SVD Singular Value Decomposition
TCM Trellis Coded Modulation
TI Time Inversion
TR Tone Reservation
TWTA Travelling Wave Tube Amplifier
VHDL Visual Hardware Design Language
VLSI Very Large Scale Integration
WCDMA Wide Band Code Division Multiplexing
WLAN Wireless Local Area Network
N Number of subcarriers
n nth subcarrier in OFDM symbol
xix
( )
m
x t m
th
continuous OFDM symbol
, m n
x n
th
discrete sample of m
th
OFDM symbol
W Total occupied frequency bandwidth of OFDM symbol
f ∆ Frequency separation between subcarriers
T Total OFDM symbol duration
T
s
Duration of one sample in OFDM symbol
, m k
X k
th
mapped transmit sample of m
th
OFDM symbol
, m k
Y k
th
mapped received sample of m
th
OFDM symbol
N
g
Length of Cyclic Prefix
T
E
Number of taps in equalization algorithm
P Mean envelope power
P
av
Average Power of an OFDM symbol
ζ Peak to Average Power Ratio of OFDM
CF
ζ Crest Factor of OFDM
mPB
x m
th
passband OFDM symbol
( )
n
P
ζ
ζ Probability density function of an OFDM symbol
[ ] F ρ Transfer properties of HPA
min
d Minimum code distance
d(C
i
, C
j
) Code distance
Rc Code rate
RM(2, m) 2
nd
order ReedMuller codes
, m k
p Phase rotation of PTS subblock
W Number of phase rotations in PTS
V Number of Subblocks
os Oversampling rate in IFFT
U Number of alternate SLM transmit symbols
v
∂ Cyclic shift in time domain for TI
R
xx
Autocorrelation of the input
R
xy
Cross correlation
R
yy
Autocorrelation of the output
α Roll off factor of pulse shaping filter
xx
R Clip level
SNR
conv
SNR of conventional clipping
SNR
3sec
SNR of 3 Sector clipping
d data stream
Chapter 1: Introduction
1
Chapter 1
Introduction
With the advance of communications technology comes the demand for higher data
rate services such as multimedia, voice, and data over both wired and wireless links.
New modulation schemes are required to transfer the large amounts of data which
existing 3
rd
generation schemes such as Global System Mobile (GSM), its enhanced
version Enhanced Data Rates for Global Evolution (EDGE), and Wideband Code
Division Multiple Access (WCDMA) cannot support. These new modulation
schemes must be able to act over point to point links and in broadcast mode, support
bidirectional communications, and be able to adapt to different requirements of
individual services in terms of their data rate, allowable Bit Error Rate (BER), and
maximum delay.
One new modulation scheme which has received significant attention over the last
few years is a form of multicarrier modulation called Orthogonal Frequency Division
Multiplexing (OFDM). OFDM has been used for Digital Audio Broadcasting (DAB)
and Digital Video Broadcasting (DVB) in Europe, and for Asymmetric Digital
Subscriber Line (ADSL) high data rate wired links. OFDM has also been
standardized as the physical layer for the wireless networking standard
‘HIPERLAN2’ in Europe and as the IEEE 802.11a, g standard in the US, promising
raw data rates of between 6 and 54Mbps.
OFDM has various properties that make it desirable over existing single carrier
systems, the main advantage is OFDM’s immunity to frequency selective fading.
"We are all interested in the future, for that is
where you and I are going to spend the rest of
our lives. And remember my friend, future
events such as these will affect you in the
future."
Criswell, Intro to "Plan 9 From Outer Space"
(1958)
Chapter 1: Introduction
2
Single carrier systems can increase their data rate by shortening the symbol time,
thereby increasing the occupied bandwidth. Wideband channels are sensitive to
frequency selective fading which require complex equalizers in the receiver to recover
the original signal. OFDM overcomes this problem by dividing the wideband channel
into a series of narrowband channels which each experience flat fading. Therefore
only 1 tap equalizers are required in the receiver, reducing complexity greatly.
Other factors such as advances in silicon and Digital Signal Processing (DSP) allow
the use of efficient Fourier transforms in the transmitter and receiver to perform the
modulation, demodulation respectively. Due to the orthogonality of the subcarriers
the transmission bandwidth is used efficiently as the subcarriers are allowed to
overlap each other and still be decoded at the receiver.
Despite the many advantages of OFDM it still suffers from some limitations such as
sensitivity to carrier frequency offset and a large Peak to Average Power Ratio
(PAPR). The large PAPR is due to the superposition of N independent equally spaced
subcarriers at the output of the Inverse Fast Fourier Transform (IFFT) in the
transmitter. A large PAPR is a problem as it requires increased complexity in the
wordlength at the output of the IFFT and the Digital to Analog Converter (DAC).
Perhaps the most serious problem is the reduced efficiency of the High Power
Amplifier (HPA) which must cater for these low probability large peaks.
If the high PAPR is allowed to saturate the HPA out of band radiation is produced
affecting adjacent channels and degrading the BER at the receiver. As portable
devices have a finite battery life it is important to find ways of reducing the PAPR
allowing for a smaller more efficient HPA, which in turn will mean a longer lasting
battery life.
1.1 Structure of thesis
This thesis analyses the principles of OFDM concentrating on the PAPR problem in
OFDM. The thesis is structured as follows:
Chapter 1: Introduction
3
• Chapter 2 provides an initial overview of OFDM starting with a brief history
of multicarrier networks and their evolution towards OFDM. The multicarrier
principle is explained mathematically encompassing the use of the Fourier
transform and the principle of orthogonality. OFDM in time varying channels
is discussed, its advantages in terms of multipath propagation, the use of a
cyclic prefix, frequency selective fading, and equalization. Problems with
OFDM are also discussed such as synchronization, which includes timing
errors, carrier phase noise, and frequency errors. The issue of the PAPR is
also briefly presented. The chapter concludes with a discussion on the
applications of OFDM in society.
• Chapter 3 explores the issue of the PAPR in more detail starting with a
mathematical definition of the PAPR. Theoretical Cumulative
Complementary Distribution Function (CCDF) results are compared to
simulated CCDF results identifying the processes which influence large peaks
such as the number of subcarriers and oversampling. Non linearities are
treated with a description of various models for the HPA, and finally the effect
of saturation of the HPA is analyzed in terms of the PSD and the BER.
• Chapter 4 begins the literature review for PAPR reduction techniques
reviewing distortionless techniques. Distortionless techniques do not corrupt
the data and encode it in such a way that it can be completely recovered at the
receiver. Initially the family of coding techniques such as block codes, cyclic
codes, ShapiroRudin Sequences, Golay complementary sequences, and Reed
Muller codes are presented. Multiple representation techniques such as
SeLective Mapping (SLM), Partial Transmit Sequences (PTS) are reviewed
with and without oversampling. Finally modified constellation techniques
Tone Reservation (TR) and Tone Insertion (TI) are examined.
• Chapter 5 introduces several new alterations to PTS called Cyclic Shifted
Sequences (CSS) and Time Inversion (TI). PTS produces alternative transmit
signals by dividing the bit source into a V subblocks which each have an
IFFT performed on them. Subblocks are then rotated by a set phase rotation
Chapter 1: Introduction
4
(which must be sent as side information to the receiver) and combined to
produce a possible transmit symbol, after a number of set phase rotations the
transmit symbol with the lowest PAPR is chosen for transmission. CSS and
TI reduce complexity and improve performance of PTS by using time shifts of
the data instead of phase rotations which can be combined with standard PTS
to reduce complexity and in some cases allow for the removal of a whole IFFT
operation without degrading performance. It is further shown that CSS and TI
perform better after oversampling and filtering than PTS. This work is
published in [1, 2].
• Chapter 6 continues the literature review for distorted PAPR reduction
techniques which do not attempt to create a transmit signal with a low crest
factor, instead they take the output of the IFFT and then limit the amplitude of
large samples which invariably causes distortion degrading the BER. Methods
reviewed are pulse shaping (or windowing), and clipping at every stage from
the output of the IFFT to limited backoffs in the amplifier. Results are
analyzed in terms of their BER and affect on the PSD.
• Chapter 7 introduces new low complexity clipping techniques starting with a
comprehensive analysis of an OFDM transceiver with clipping at various
points in the transmission chain and under other variable conditions such as
the amount of oversampling in the IFFT, pulse shaping taps and roll off, and
the HPA parameters. New low complexity clipping methods are introduced
which avoid complex hardware operations while maintaining similar
performance to conventional clipping. The new clipping algorithms called
Sector clipping and Vector Subtraction are then implemented in a new clip and
filter algorithm which is much less susceptible to peak regrowth after
baseband filtering. This work is published in [3].
1.1.1 Contributions
The contribution that this research work has made to the wireless communications
field is summarized as follows:
Chapter 1: Introduction
5
• A thorough analysis of the theory, principles, and techniques of OFDM based
wireless systems including a detailed analysis of PAPR reduction techniques
are presented (Chapters 2, 3, 4, and 6).
• A new method for producing Partial Transmit Sequences (PTS) signals and
their performance under oversampling conditions is proposed (Chapter 5).
• A detailed analysis of the effect of clipping on an OFDM transceiver under
various system conditions (Chapter 7).
• Several new low complexity clipping algorithms are proposed (Chapter 7).
• Implementation and analysis of a proposed clip and filter algorithm utilizing
one of the new low latency clipping algorithms (Chapter 7).
Chapter 2: Orthogonal Frequency Division Multiplexing
6
Chapter 2
Orthogonal Frequency Division
Multiplexing
This chapter provides an initial overview of Orthogonal Frequency Division
Multiplexing (OFDM). Section 2.1 provides a brief history of multicarrier networks
and their evolution towards OFDM. Section 2.2 explains the multicarrier principle
and Section 2.3 explains how it is applied to OFDM, detailing the use of the Fourier
transform, and the importance of orthogonality. Section 2.4 explores OFDM in time
varying channels describing its advantages in terms of multipath propagation, the use
of a cyclic prefix, frequency selective fading, and equalization. Limitations of
OFDM, such as synchronization, which includes timing errors, carrier phase noise,
and frequency errors are discussed in Section 2.5. Non linearities are also introduced
as a major hindrance to a practical OFDM system in this section. Section 2.6 looks at
applications of OFDM in society and discusses where this new communications
technology will be used. Finally Section 2.7 summarizes the chapter with a brief
recap of the chapter.
2.1 History of multicarrier networks
Multicarrier networks such as Frequency Division Multiplexing (FDM) have been
around since the late 1950’s [4], however due to their implementational complexity
and inefficient use of the frequency band they were restricted to military applications.
A multicarrier system is basically a number of information bearing carriers
Chapter 2: Orthogonal Frequency Division Multiplexing
7
transmitted in parallel. Multicarrier systems in wireless applications are less
susceptible to channel induced distortions than single carrier systems at corresponding
data rates.
Chang [5] and Saltzberg [6] further developed FDM in the mid 60’s by introducing
multiple carriers which overlap in the frequency domain without interfering with each
other, utilizing the frequency spectrum more efficiently, hence OFDM. However the
complexity issue still remained.
In the 1970’s Weinstein and Ebert [7] used an Inverse Discrete Fourier Transform
(IDFT) and Discrete Fourier Transform (DFT) to perform the modulation and
demodulation respectively, exploiting the sinusoidal nature of the Fourier Transform
and significantly reducing the complexity of an OFDM system.
In the last 10 years more advances in practical OFDM systems have been made,
particularly in Europe where various projects and prototypes were initiated such as
DIgital VIdeo Narrowband Emission (HDDIVINE), System de Television En
Radiodiffusion NumeriquE (STERNE), and digital Terrestrial Television broadcasting
(dTTb). This has led to the adoption of OFDM in many European standards.
OFDM has progressed to the point where it has now been used for various
communication applications such as Digital Audio Broadcasting (DAB) and Digital
Video Broadcasting (DVB) in Europe. It has also been adopted as the physical layer
modulation scheme for wireless networking standards such as Hiperlan2 in Europe
and the Institute of Electrical and Electronic Engineers (IEEE) 802.11a, g standards in
the United States.
However while OFDM successfully alleviates the problem of dispersive channels
there are still some problems which need to be addressed such as time and frequency
synchronization, frequency selective fading, and the Peak to Average Power Ratio
(PAPR).
Chapter 2: Orthogonal Frequency Division Multiplexing
8
2.2 Multicarrier principle
An early form of a multicarrier system is shown in Figure 2.1. The basic principle of
multicarrier modulation is to divide the data stream, d, into N parallel data streams
with a reduced data rate of d/N. Each low rate data steam is then modulated on a
separate narrow band subcarrier and summed together for transmission, thereby
providing the same data rate as an equivalent single carrier system. At the receiver a
set of filter banks separate the wideband signal into the original narrowband
subcarriers for demodulation. The advantage of this structure over single carrier
systems is that the extended symbol time (due to lower data rate) makes the signal less
susceptible to effects of the channel such as multipath propagation which introduces
Inter Symbol Interference (ISI). Each subchannel will therefore experience flat
fading reducing the equalization complexity in the receiver dramatically. This issue
will be explored in more depth in Section 2.4.
A disadvantage of the method shown in Figure 2.1 is the implementation complexity
due to the large number of filter banks required in the transmitter and receiver as well
as the inefficient use of the available frequency band [8]. The spectra of the different
carriers cannot overlap as this would introduce distortion degrading system
performance.
2.3 OFDM implementation of multicarrier modulation
A more spectrally efficient implementation of the aforementioned multicarrier system
is OFDM (Figure 2.2). In OFDM the transmit signals are constructed in such a way
that the frequency spectra of the individual subchannels are allowed to overlap
thereby utilising the frequency spectrum much more efficiently.
Mathematically the continuous time representation of the OFDM transmit signal
depicted in Figure 2.2 is
( ) ( )
2
2
,
1
2
1
. .
N
j k ft
m m k k
N
x t X e w t mT
N
π ∆
− +
= −
∑
0 t T > > (2.1)
Chapter 2: Orthogonal Frequency Division Multiplexing
9
Figure 2.1: Block diagram of a basic multicarrier system.
RF
Filter f
0
QAM
QAM
QAM
Filter f
1
Filter f
N1
f
0
f
1
f
N1
QAM Filter
RF QAM Filter
QAM Filter
f
0
f
1
f
N1
d/N b/s
d/N b/s
d/N b/s d
0
(t)
d
1
(t)
d
N1
(t)
f
0
f
1
f
2
Transmitter
Receiver
Chapter 2: Orthogonal Frequency Division Multiplexing
10
Figure 2.2: Basic OFDM transmitter and receiver. Occupied frequency band shown in between.
f
0
f
1
f
2
f
3
f
4
Transmitter
Parallel
to
serial
convertor
QAM
∫
∫
∫
RF
f
0
f
1
f
N1
0
ˆ
X
1
ˆ
X
1
ˆ
N
X
−
Receiver
QAM
Serial
to
parallel
converter
RF
0
f
1
f
1 N
f
−
.
.
.
.
.
.
( ) x t
1
0
( ) Re xp( )
N
n n
n
x t X e j t ω
−
=
= −
∑
Timelimited
signal
(block processing)
n
X
0
X
1
X
1 N
X
−
R b/s
Chapter 2: Orthogonal Frequency Division Multiplexing
11
where
, m k
X is the mapped (QAM, PSK, etc) data to be transmitted on the k
th
subcarrier of the m
th
transmitted symbol,
2 j k ft
e
π ∆
is the k
th
subcarrier, f ∆ is the
frequency spacing between subcarriers, and ( )
k
w t mT − is a rectangular window
applied to each subcarrier, N is the number of subcarriers, and T is the total time of
the transmit symbol. To ensure the orthogonal relationship between subcarriers f ∆ is
set as
1 W
N T
= (W is the total bandwidth of the signal).
In the receiver an integrate and dump operation is performed over time T to recover
the data.
2.3.1 Use of Fourier Transform for modulation and demodulation
In order to make multicarrier systems a more practical technology an IDFT and DFT
are used for the baseband modulation and demodulation respectively, as first
suggested in reference [7], where the sinusoidal nature of the Fourier transform basic
functions is exploited. Advances in silicon technology have made the production of
the DFT more cost efficient [911]. Figure 2.3 shows a block diagram of a basic
OFDM system in the baseband utilising the IDFT, DFT pair.
A discrete time representation of (2.1) can be obtained by sampling the continuous
signal. Under the condition that W N f = ∆ and
1
f
T
∆ = the signal can be determined
by its samples if sampled at
T
t
N
= . Under this condition (2.1) then becomes (2.2)
{ }
1
2
, , ,
0
1
.
N
j nk
N
m n m k m k
k
x X e IDFT X
N
π
−
=
= =
∑
0 1 n N ≤ ≤ − (2.2)
where ‘n’ are the discrete sampling points. This equation describes exactly the IDFT
operation. In hardware the more efficient form of the IDFT and DFT, the Inverse Fast
Fourier Transform (IFFT) and Fast Fourier Transform (FFT) is used for the
modulation, demodulation respectively, where N is set to be a power of 2.
Chapter 2: Orthogonal Frequency Division Multiplexing
12
Figure 2.3 shows a baseband transceiver structure for OFDM utilising the Fourier
transform for modulation and demodulation. Here the serial data stream is mapped to
complex data symbols (PSK, QAM, etc) with a symbol rate of
1
s
T
. The data is then
demultiplexed by a serial to parallel converter resulting in a block of N complex
symbols, X
0
to X
N1
. The parallel samples are then passed through an N point IFFT
(in this case no oversampling is assumed) with a rectangular window of length N.T
s
,
resulting in complex samples x
0
to x
N1.
Assuming the incoming complex data is
random it follows that the IFFT is a set of N independent random complex sinusoids
summed together. The samples,
0
x to
1 N
x
−
are then converted back into a serial data
stream producing a baseband OFDM transmit symbol of length T=N.T
s
.
Figure 2.3: Basic OFDM transmitter and receiver pair utilizing Fourier transform.
A Cyclic Prefix (CP), which is a copy of the last part of the samples is appended to
the front of the serial data stream before Radio Frequency (RF) up conversion and
Serial
to
Parallel
Parallel
to
Serial
IFFT
Channel
+ noise
Data
source
Data
sink
Parallel
to
Serial
Serial
to
Parallel
FFT
Constel
lation
mapping
Constel
lation
demappi
ng
CP
CP
Binary data
y
0
y
N1
y
1
X
0
X
N1
X
1
x
0
x
N1
x
1
Y
0
Y
N1
Y
1
Complex data
Chapter 2: Orthogonal Frequency Division Multiplexing
13
transmission. The CP combats the disrupting effects of the channel which introduce
Inter Symbol Interference (ISI) and is discussed in more detail in section 2.4.2.
In the receiver the whole process is reversed to recover the transmitted data, the CP is
removed prior to the FFT which reverses the effect of the IFFT. The complex
symbols at the output of the FFT, Y
0
.. Y
N1
are then decoded and the original bit
steam recovered.
Mathematically the demodulation process (assuming no CP and no channel
impairments) using the FFT is (2.3)
{ }
( )
( )
[ ]
, ,
1
2
,
0
1 1 2
,
0 0
1 1 2
,
0 0
1
,
0
,
1
1
1
1
m k m n
N
j nk
N
m n
n
N N j n d k
N
m d
n d
N N j n d k
N
m d
d n
N
m d
d
m k
Y FFT x
x e
N
X e
N
X e
N
X N d k
N
X
π
π
π
δ
−
−
=
− − −
= =
− − −
= =
−
=
=
=
=
=
= −
=
∑
∑∑
∑ ∑
∑
(2.3)
2.3.2 Orthogonality in OFDM
One of the key advantages of OFDM is its efficient use of the frequency band as the
subcarriers are allowed to overlap each other in the frequency domain. The N equally
spaced subcarriers will be orthogonal if the frequency separation between subcarriers
is
1 1
s
f
N T T
∆ = =
⋅
, where N.T
s
is symbol duration, and rectangular windowing of
the IFFT is performed. Under these conditions the subcarriers will have a sinc
waveform frequency response. Figure 2.4 shows the frequency response of a 5 carrier
system where it is seen that because of the orthogonal relationship the maximum of a
particular sample corresponds to a null in all other carriers, therefore eliminating the
effects of interference. Smoother window functions (eg. Raised Cosine Filter) reduce
Chapter 2: Orthogonal Frequency Division Multiplexing
14
the out of band emissions and Inter Carrier Interference (ICI) susceptibility to system
imperfections (e.g. frequency offset) but they increase the symbol period.
Mathematically, orthogonality of two signals, ( )
k
t ψ and ( )
l
t ψ over time period N.T
s
is described in reference [12] and expressed here as (2.4)
( ) ( )
0
0,
,
s
NT
k l
k l
t t dt
C k l
ψ ψ
∗
≠
=
=
∫
(2.4)
where C is a constant.
Figure 2.4: Frequency spectrum of 5 orthogonal subcarriers of an OFDM transmit signal
2.4 OFDM transmission over time varying channels
OFDM is being primarily deployed in the wireless environment. This section
describes properties of the wireless channel and describes the advantages and
disadvantages of OFDM in this environment.
6 4 2 0 2 4 6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Carrier Spectrum with no carrier offset
1/NT
s
Chapter 2: Orthogonal Frequency Division Multiplexing
15
2.4.1 Multipath propagation
The wireless channel is a harsh one, electromagnetic signals travelling through this
medium are fraught with disruptive and warping effects. The transmitted signal does
not only have a direct path to the receiver (in the case of line of sight). The signal is
reflected of buildings and mountains and other obstacles so that multiple delayed
copies of the same transmitted signal arrive at the receiver affecting other symbols.
This causes ISI which degrades the Bit Error Rate (BER). The longer the delay of the
paths the greater the ISI, a measure of the delay is given by the rootmeansquare
(rms) delay spread which is a measure of the delay experienced by a single pulse.
It is this effect which restricts single carrier systems from achieving high data rates.
The data rate in a single carrier system can be increased by shortening the symbol
time of the transmitted pulses, but they will be even more affected by the rms delay
spread and require more complex equalisation in the receiver. As the rms delay
spread is a result of the physical channel it cannot be changed and systems must be
designed to accommodate it. This phenomenon has prompted the use of multicarrier
techniques where the transmitted bandwidth is divided into many narrow band
channels which are then transmitted in parallel. Each subcarrier is modulated at a
sufficiently low data rate so that it is not affected by the delay spread.
2.4.2 Use of a Cyclic Prefix
In order to protect successive OFDM symbols from multipath a CP of length N
g
is
used which is a copy of the last part of the samples of a OFDM transmit block
appended to the front before transmission as depicted in Figure 2.5. The transmitted
signal is therefore N+N
g
samples. Provided that the length of the CP is chosen so that
it is longer than the longest expected delay path successive OFDM symbols will be
free of ISI [13].
At the receiver a window of N samples is chosen from the N+N
g
length block for
maximum power, the rest of the repeated samples are discarded. After cyclic shifting
to get the samples back into the original order a FFT is performed to demodulate the
data.
Chapter 2: Orthogonal Frequency Division Multiplexing
16
Obviously the use of a CP decreases the data rate by a factor of
g
N
N N +
(2.5)
as the repeated samples are discarded in the receiver so it is important to keep the
length of the CP as short as possible with respect to the rms delay spread. A loss in
the SNR of the received signal is also incurred due to the lost energy in the CP.
However there are techniques which use the CP for both frequency offset estimation
and symbol synchronization [14]. Also when filtering the signal there is a delay
before the filter is at full power, by using a CP the delay will occur in CP so that N of
the samples will be at full power. The CP (with repeated samples) retains the cyclic
nature of the symbol by creating a periodic received signal for processing, eliminating
ICI.
Figure 2.5: OFDM symbol a) without cyclic prefix, and b) with cyclic prefix.
0 N1
time
a) Original N point OFDM symbol
Original N samples
time
N
g
b) OFDM symbol with Cyclic Prefix (CP)
Chapter 2: Orthogonal Frequency Division Multiplexing
17
More recent work has shown that it is possible to use null values for the CP provided
additional processing is done at the receiver. The received signal in the null sample
positions is wrapped around and added to the samples at the start of the symbol to
restore orthogonality and eliminate ICI. This is one of the proposals for the new IEEE
802.15.3a ultra wideband standard.
Other forms of the CP have also been investigated, in particular reference [15]
examines the effect that using null values in the CP will have. It was concluded that
null samples have a detrimental effect through loss of orthogonality increasing ICI.
2.4.3 Frequency selective fading
Multipath propagation as discussed in the previous section can be combated
successfully through the use of a cyclic prefix. Frequency selective fading is the
reciprocal effect of multipath propagation in the time domain and can be defined thus.
If the channel has a constant magnitude and phase response over a bandwidth that is
smaller than the bandwidth of the transmitted signal the channel creates frequency
selective fading [16]. Under this condition the signal experiences multipath
introducing ISI, this effect shows itself in the frequency domain where certain
frequency components in the received spectrum have greater or less power than the
transmitted spectrum.
Narrow pulses in time (such as high data rate single carrier transmission) occupy a
wide frequency bandwidth, conversely pulses with a long duration (such as OFDM)
occupy a relatively narrow frequency band. Figure 2.6 compares a single and
multicarrier signal in the time and corresponding frequency domain with equivalent
data rates. Here we see that the many narrowband channels of the OFDM signal
experience fading, however each subchannel has a constant gain within its own
frequency band. Constant fading over the occupied bandwidth is known as frequency
flat fading and is a much easier effect to correct in the receiver than frequency
selective fading.
Chapter 2: Orthogonal Frequency Division Multiplexing
18
2.4.4 Equalization
In order for the receiver to correct the effect of fading equalization is performed in the
receiver which is the process of measuring the channel response and using this
information to correct the received signal. Several methods have been suggested, one
of the more popular methods utilize pilot tones [1719] which are certain (usually
evenly spaced) subcarriers with a known amplitude and phase at the receiver. By
measuring the difference between the received and the transmitted value a picture of
the channel can be extrapolated. Other methods use blind estimation techniques [20,
21] which do not require pilot tones. The equalization technique depends on the
modulation scheme and the channel properties. Equalization algorithms are usually
implemented as tapped delay lines. The relation between the number of taps required,
T
E
and the occupied bandwidth, W, of the signal is [22].
( )
2
E
T BW = (2.6)
Signals which experience frequency selective fading such as single carrier systems
require complex equalization (i.e. more taps) where the complexity is directly
proportional the bandwidth of the signal. Signals such as OFDM which experience
frequency flat fading only require a 1 tap equalizer. This reduction in equalization
complexity is a driving force for the use of OFDM.
2.5 Limitations in OFDM
Previous sections have detailed the advantages of OFDM, however the advantages are
offset by some problems that are unique to OFDM, namely time and frequency
synchronization problems and non linearities.
Chapter 2: Orthogonal Frequency Division Multiplexing
19
Figure 2.6: Time and frequency properties of single carrier and OFDM techniques.
2.5.1 Synchronization
Both time and frequency synchronization are a major drawback in OFDM, the
following sections detail the problem and provide a basic introduction into solutions
for these problems.
2.5.1.1 Timing errors
Timing synchronization is the process of finding the start of a symbol in the receiver.
If the timing mismatch is within the CP the demodulation produces a linear phase
rotation at the output of the FFT which can be corrected with a channel estimator.
( ) [ ]
( )
1
2
0
c
N
j k ft f k f t
b
k
x t a k e
π δ
−
∆ + + ∆
=
=
∑
ɶ (2.7)
freq
freq
time
time
time
time
OFDM time and frequency properties
Single carrier time and frequency properties
Frequency
fading
envelope
Frequency
fading
envelope
NT
s
T
s
N
1
2
1
s
NT
1
s
T
Chapter 2: Orthogonal Frequency Division Multiplexing
20
where
[ ] [ ]
( ) 2
ˆ
c
j f n f t
t
a k a k e
π δ
δ
+ ∆
= (2.8)
is the phase shift.
If the timing mismatch is not corrected additional interference (ISI) is generated.
Therefore a sufficient length of the CP needs to be chosen. An alternative approach is
to use pilot based methods [13] which uses certain carriers with a known amplitude
and phase at the receiver. By analyzing the phase rotation and amplitude change, an
estimate of the channel can be made.
2.5.1.2 Carrier phase noise
Carrier phase noise is caused by a mismatch in the RF oscillators in the transmitter
and receiver and manifests itself in the baseband as additional phase rotation and
amplitude attenuation [13]. No distinction can be made between phase rotations
introduced by timing errors and carrier phase offset [14]. The effect of phase noise is
more pronounced in differential detection schemes than coherent detection schemes
[13]. Several references [23, 24] have analyzed the effect of carrier phase noise on
the performance of OFDM schemes.
2.5.1.3 Frequency errors
Frequency offset errors are caused by mismatch between the RF oscillators, Doppler
shifts, and phase noise introduced by non linear channels [14]. Frequency offset
causes the received signal to not be sampled at the peak, this means that the sample
under consideration is not at maximum power. Power from adjacent subcarriers is
also sampled as well. A simplistic representation of this effect in the frequency
domain is visualized in Figure 2.7.
Chapter 2: Orthogonal Frequency Division Multiplexing
21
−4 −3 −2 −1 0 1 2 3 4
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
frequency (normalised)
A
m
p
l
i
t
u
d
e
Figure 2.7: Effects of frequency offset ∆F: reduction of signal amplitude (star), and ICI (circle).
OFDM is more sensitive to frequency offset than single carrier systems due to the
tight orthogonal packing of the subcarriers. Reference [14] concludes that to avoid
severe degradation the frequency accuracy should be better than 2%.
Suggested solutions to frequency synchronization (like symbol synchronization) are
based on pilot symbols and the cyclic prefix are used in reference [14] where it is
noted that time and frequency synchronization are closely related. Frequency
sensitivity can be made more robust by reducing the number of subcarriers within a
set bandwidth thereby increasing the frequency distance between subcarriers.
However this shortens the symbol time which increases the demands on timing
synchronization, therefore a trade off must be made.
F ∆
Chapter 2: Orthogonal Frequency Division Multiplexing
22
2.5.2 Non linearities
Another limiting aspect of multicarrier and OFDM modulation is the high
instantaneous signal peak with respect to the signals average power. Large peaks are
due to the superposition of N random phase sine waves in the IFFT. Hardware
components such as the DigitaltoAnalog Converter (DAC), IFFT/FFT with limited
word length and most importantly the High Power Amplifier (HPA) will be driven
into saturation unless they are designed to operate over large dynamic ranges. If the
signal is allowed to go into saturation both in band noise which degrades the BER and
out of band radiation introducing ICI will result.
Therefore many papers (refer to Chapters 4, 5, 6, and 7) have been published on ways
to overcome the PAPR and can be divided into two methods: distortionless techniques
which attempt to create a transmit signal with a low PAPR without affecting BER of
the data, and distorted techniques which deliberately reduce peaks but increase
distortion and therefore the BER. This problem is the area of research of this thesis
and will be treated with more detail in the following chapters.
2.6 Applications of OFDM
The previous section detailed some of the problems with OFDM, it should be noted
that depending on the application and medium different design issues take
precedence. This section identifies some of the current and future applications of
OFDM. OFDM takes its place in the next generation of communication systems
because of its high data rates and low complexity.
2.6.1 COFDM
Coded OFDM (COFDM) is a practical form of OFDM where redundant bits are
inserted into the bit stream at the transmitter. These specially chosen bits allow
powerful error correction codes in the receiver to reduce the BER. The more bits used
for error correction the better the error correction properties, however the useful data
rate is decreased.
Chapter 2: Orthogonal Frequency Division Multiplexing
23
Types of error correction codes used for example DABOFDM are Trellis Coded
Modulation (TCM) combined with frequency and time interleaving. In practice all
the following technologies use some form of COFDM.
2.6.2 Digital Audio Broadcasting
Digital Audio Broadcasting (DAB) using OFDM has been standardized in Europe
[25] and is the next step in evolution beyond FM radio broadcasting providing
interference free transmssion. The standard for DAB is known as Eureka147 [26]
and is a multiservice digital broadcasting method transmitting at around 1.5Mbps in
the 1.536MHz band. In DAB between 192 and 1536 carriers are used with
Differential Quadrature Phase Shift Keying (DQPSK), which allows the system to
avoid channel estimation techniques. The very long symbol time means that large
echo’s can be tolerated and that the redundancy due to the CP is not that great. Large
echoes are expected as the broadcasting is over large distances so that long delay
paths will be present. The PAPR is a problem but as DAB only uses DQPSK
modulation it is more impervious to noise generated through saturation of the
amplifier. The DAB data payload contains audio, data associated with audio, and
other optional data services. Table 2.1 displays system parameters for DAB.
Table 2.1: DAB parameters.
Parameters Mode
I II III
Application SFN Terrestial Satellite
Modulation DQPSK DQPSK DQPSK
Total number of subcarriers 1536 384 192
OFDM symbol duration 1246µS 312µS 156µS
Guard interval 246µS 62µS 31µS
Frequency range ≤375MHz ≤1.5GHz ≤3GHz
Chapter 2: Orthogonal Frequency Division Multiplexing
24
2.6.3 Digital Video Broadcasting
Digital Video Broadcasting (DVB) [27, 28] is also using OFDM as the carrier
modulation scheme. DVB promises to deliver full multimedia in digital form in a
broadcast format. DVB adapts the baseband TV signal from the output of the MPEG
2 [29] transport multiplexer to the terrestrial channel characteristics. Maximum
spectral efficiency within the VHF and UHF bands is achieved by utilizing Single
Frequency Network (SFN) operation. There are two modes defined in DVB: 1/ 2K
mode, and 2/ 8K mode. The 2K mode is used for single transmitters and small SFN’s
where the distance for transmission is limited. The 8K mode encompasses the 2K
mode as well as larger SFN’s. One of the many advantages of OFDM is that different
mapping types can be used on different subcarriers, this aspect is taken advantage of
in DVB so that the data rate on a channel mirrors its quality. Table 2.2 shows the
system parameters for DVB in 2K mode [28].
Table 2.2: DVB system parameters for 2K mode.
Parameters Value
Information data rate 530 Mbps
Modulation QPSK, 16 QAM, 64 QAM
FEC code Reed Solomon outer code
Convolutional inner code
Code Rates ½, 2/3, ¾
Total number of subcarriers 1705 (2K mode)
OFDM symbol duration 303µS
Guard interval 75.9µS
Signal bandwidth 5.62MHz
2.6.4 HiperLan2/802.11a
Wireless networking standards such as HIPERLAN2 and 802.11a use OFDM as the
physical layer modulation scheme and operate in the unlicensed 5GHz frequency
band. Hiperlan2 promises to deliver raw data rates of up to 56Mbps which puts them
in the ballpark of wired LANs which have data rates of up to 100Mbps. Wireless
Chapter 2: Orthogonal Frequency Division Multiplexing
25
LANs applications are for home and office networking over short distances (<50
metres) as well as community spaces such as Starbucks which operates the 802.11b
wireless standard free of charge for customers and provides data rates up to 11Mbps
[52]. Tables 2.3 and 2.4 list Hiperlan2/802.11a specifications and data rates
respectively.
Table 2.3: HIPERLAN2 parameters
Parameter Value
Sampling Rate f
s
=1/T 20MHz
Useful symbol part duration T
U
64 × T
3.2µS
Cyclic Prefix duration T
CP
16 × T
0.8µS (mandatory)
8 × T
0.4µS
Symbol Interval T
S
80 × T
4.0µS (T
U
+T
CP
)
72 × T
3.6µS (T
U
+T
CP
)
Number of data subcarriers N
SD
48
Number of pilot subcarriers N
SP
4
Total number of subcarriers N
ST
52 (N
SD
+N
SP
)
Subcarrier spacing ∆f 0.3125MHz (1/T
U
)
Spacing between two outmost subcarriers N
ST
16.25MHz (N
ST
× ∆f)
Table 2.4: Data rates for HIPERLAN2
Modulation Coding rate R Nominal bit rate (Mbps)
BPSK 1/2 6
BPSK 3/4 9
QPSK 1/2 12
QPSK 3/4 18
16QAM 9/16 27
16QAM 3/4 36
64QAM 3/4 54 (optional)
2.6.5 ADSL
Asynchronous Digital Subscriber Lines (ADSL) utilize OFDM over wired links [30].
Data rates for ADSL standard [14] are 1.54Mbps to 6.1Mbps in the downlink and 9.6
to 192Kbps in the uplink over several kilometers of ordinary twisted pair telephone
line, while still supporting the standard telephone. The unbalanced data rates make
Chapter 2: Orthogonal Frequency Division Multiplexing
26
ADSL particularly applicable to internet type applications where the downlink rate is
typically much larger than the uplink rate.
Stationary channels like wireless links do not change over time, therefore a technique
called bit loading is used. Bit loading assigns a mapping type to subcarriers
depending on its quality, using the available bandwidth efficiently. Bit loading used
in conjunction with OFDM over wired links is usually called Discrete MultiTone
(DMT).
2.6.6 MIMO OFDM
Multiple In Multiple Out (MIMO) [31] OFDM combines OFDM with multiple
antennas at the transmitter and receiver. This structure allows greater diversity when
techniques such as Singular Value Decomposition (SVD) are used. This process,
called spatial multiplexing, proportionally boosts the datatransmission speed by a
factor equal to the number of transmitting antennas. In addition, since all data is
transmitted both in the same frequency band and with separate spatial signatures, this
technique utilizes spectrum very efficiently.
2.7 Conclusion
This chapter introduced fundamental properties of OFDM, identifying its advantages
and discussing its limitations. Specifically the history of muticarrier networks was
discussed and its evolution towards OFDM. The introduction of the Fourier
transform, advances in silicon technology, and the efficient use of the frequency
spectrum with orthogonally spaced subcarriers were shown to make OFDM a
practical technology for the next generation of digital communications.
OFDM transmission over wireless channels was discussed focusing on multipath
propagation and the advantages of OFDM in this medium. The use of the cyclic
prefix in OFDM was shown to reduce the equalizer complexity dramatically down to
one tap per subcarrier. This is one of the great advantages of OFDM over single
Chapter 2: Orthogonal Frequency Division Multiplexing
27
carrier networks which require prohibitively high complexity equalizer structures as
the carrier frequency is increased to cater for higher data rates. Frequency selective
fading was also introduced as a major advantage of OFDM where due to the long
effective symbol time OFDM subcarriers experience flat fading.
Limitations of OFDM were analyzed next, with the two main disadvantages;
synchronization errors and non linearities treated. Synchronization errors were
shown to include timing errors, carrier phase noise, and frequency errors. Non
linearities due to the Rayleigh distributed samples at the output of the IFFT were also
briefly introduced and shown to have a degrading affect on the quality of OFDM
systems.
Finally applications of OFDM were presented detailing where this new technology
has manifested itself in society. Areas of application in the wireless field were shown
to be COFDM, DAB, DVB where OFDM is used in a broadcast mode,
Hiperlan2/802.11a for wireless networking potentially taking the place of large wired
networks. OFDM in DMT form is being used in wired networks for ADSL.
Extensions of OFDM such as MIMO OFDM were also shown to expand the reach of
OFDM systems.
Chapter 3: Peak to Average Power in OFDM
28
Chapter 3
Peak to Average Power in OFDM
Chapter 2 discussed fundamental principles of OFDM and showed how it is a
practical technology for the next generation of high data rate communication systems.
However several design issues need to be addressed, one of the most important being
the Peak to Average Power Ratio (PAPR) of the highly fluctuating transmit signal
envelope.
Due to the nature of the IFFT which, as described in Section 2.3, sums N sinusoids
through superposition, some combinations of the sinusoids create large peaks. The
drawback of a large dynamic range is that it places pressure on the design of
components such as the word length of the IFFT/FFT pair, DAC and ADC, mixer
stages, and most importantly the HPA which must be designed to handle irregularly
occurring large peaks. Failure to design components with a sufficiently large linear
range results in saturation of the HPA. Saturation creates both in band distortion,
increasing the BER and out of band distortion, or spectral splatter, which causes ACI.
One obvious solution is to design the components to operate within large linear
regions, however this is impractical as the components will be operating inefficiently
and the cost becomes prohibitively high. This is especially apparent in the HPA
where much of the cost and ~50% of the size of a transmitter lies.
This chapter provides a mathematical definition of the PAPR and identifies the
processes which contribute to large peaks. Specifically section 3.1 gives a
Chapter 3: Peak to Average Power in OFDM
29
mathematical definition of the PAPR, section 3.2 provides a statistical analysis of
PAPR and identifies contributing factors to large peaks. Section 3.3 introduces
oversampling of OFDM signals while section 3.4 investigates the effect of non
linearities on OFDM, finally section 3.5 summarizes the chapter reiterating the main
points of this section. Note that the terms subcarrier, tones, and N will be used
interchangeably to signify the number of subcarriers in an OFDM symbol.
3.1 Peak to Average Power Ratio
The PAPR is the relation between the maximum power of a sample in a given OFDM
transmit symbol divided by the average power of that OFDM symbol. The mean
envelope power of the baseband expression (assuming same constellation on each
subcarrier) is defined as (3.1)
( )
1
2 2
,
0
0
1 1
N
T
m m k
t
k
P x t dt X
T N
−
=
=
= =
∑
∫
(3.1)
where ( )
m
x t is defined in (2.1), X
m,k
are assumed to be complex Quadature
Amplitude Modulated (QAM) data which are statistically independent, identically
distributed (i.i.d) random variables with 0 mean and variance
2
2
, m k
E X σ
≜ . The
average power is defined as (3.2)
[ ] ( )
2
av m
P E P E x t
= =
(3.2)
The PAPR can then be defined
( )
2
0
max
t T m
av
x t
P
ζ
≤ ≤
= (3.3)
where ( )
2
0
max
t T m
x t
≤ ≤
is the maximum instantaneous power within the period
0 t T ≤ ≤ .
Chapter 3: Peak to Average Power in OFDM
30
Another definition for the PAPR is crest factor, defined in reference [32] as (3.4)
( )
2
0
max
t T m
CF
av
x t
P
ξ
≤ ≤
= (3.4)
and results in a 3dB shift in results compared to (3.3). Throughout this thesis the
definition of the PAPR given in (3.3) will used unless specified otherwise.
For passband transmission the OFDM symbol is modulated onto a carrier frequency,
f
c
,
( ) { }
2
c
j f t
mPB m
x x t e
π
= ℜ
( ) { } ( ) ( ) { } ( ) cos 2 sin 2
m c m c
x t f t j x t f t π π = ℜ − ℑ
( ) ( ) ( ) ( ) cos 2 sin 2
mI c mQ c
x t j f t jx t j f t π π = − (3.5)
The carrier frequency is usually much higher than the signal bandwidth, i.e. f
c
>>∆f,
therefore the maximum of the passband signal is approximately equal to the baseband
expression, i.e.
( ) ( ) max max
mPB m
x t x t ≈ (3.6)
Most OFDM schemes usually employ QAM mapping for the modulation where
( ) ( )
2 2
mI mQ
x t x t = therefore
( )
{ }
( )
{ }
( )
{ }
2 2 2
2 2
m mI mQ
E x t E x t E x t = = (3.7)
The average RF power of the passband signal can be derived as
( )
{ }
( ) ( ) ( ) ( )
{ }
2 2
cos 2 sin 2
mPB mI c mQ c
E x t E x t f t jx t f t π π = −
( ) ( )
{ }
( ) ( )
{ }
2 2
cos 2 sin 2
mI c mQ c
E x t f t E x t f t π π = +
Chapter 3: Peak to Average Power in OFDM
31
( )
{ }
( )
{ }
2 2 1 1
2 2
mI mQ
E x t E x t = +
( )
{ }
2 1
2
m
E x t =
2
av
P
= (3.8)
Substituting (3.8) back into (3.3) gives the PAPR in the passband.
( )
( )
{ }
2
0
2
max
t T mPB
PB
mPB
x t
E x t
ζ
≤ ≤
=
( )
2
0
max
2
t T mPB
av
x t
P
≤ ≤
= (3.9)
The problem with OFDM is that theoretically the PAPR can be up to log
2
(N), which
is huge. But as will be shown in the next section the general distribution of samples is
much lower.
3.2 Statistical distribution of OFDM samples
Section 3.1 provided a worst case scenario or upper bound of the PAPR, but only a
few combinations of input data sequences produce large peaks, therefore it is more
pertinent to define the statistical distribution of the PAPR in OFDM. (2.2) describes a
Nyquist sampled baseband OFDM symbol with N subcarriers, from the central limit
theorem [33] the sum of these elements are zero mean complex random near Gaussian
(provided N>64) distributed variables with variance, σ
2
, of ½. It then follows that the
amplitude, a
n
, of the OFDM symbol has a Rayleigh distribution [13] with a
Probability Density Function (PDF) of
( )
2
2
2
2 n
p e
ζ
σ
ζ
ζ
ζ
σ
−
= (3.10)
Chapter 3: Peak to Average Power in OFDM
32
Substituting σ
2
= ½
( )
2
2
n
p e
ζ
ζ
ζ ζ
−
= (3.11)
Figure 3.1 shows the simulated envelope (N=64) of a baseband OFDM system which
as expected follows a Rayleigh distribution. It is seen here that the probability of any
given sample having a magnitude above 3dB decays rapidly. The probability that the
magnitude of a sample is below a certain threshold,
0
ζ , is given by the Cumulative
Distribution Function (CDF)
{ } ( )
0
Pr
n
p
ζ
ζ ζ ζ ζ
∞
−∞
≤ = ∂
∫
2
0
0
2 e
ζ
ζ
ζ ζ
−
= ∂
∫
2
0
1 e
ζ −
= − (3.12)
Under the assumption of statistically independent samples the Complementary
Cumulative Distribution Function (CCDF) can be found for the case where at least
one sample in an OFDM symbol exceeds the magnitude,
0
ζ .
0 0
0 0
Pr 1 Pr
max max
n N n N
ζ ζ ζ ζ
≤ < ≤ <
> = − ≤
{ } ( )
0
1 Pr
N
ζ ζ = − ≤
( )
2
0
1 1
N
e
ζ −
= − − (3.13)
Figure 3.2 displays simulated and theoretical results of (3.13) with varying N. The
simulation model passes N QPSK symbols through a N point IFFT, the maximum
sample of each OFDM symbol is stored and plotted. As in Figure 3.1 the probability
that large peaks occur is very irregular, doubling the number of subcarriers results in a
modest increase in the PAPR leading to the assumption that using a large number of
subcarriers makes sense as this will allow for greater data throughput (provided the
CP length is constant). However a larger number of subcarriers leads to increased
sensitivity to carrier and sampling frequency offsets as described in section 2.5.1.2
and 2.5.1.3. Therefore, balance must be met between these design constraints.
Comparing the theoretical and simulated results we see that the results only converge
Chapter 3: Peak to Average Power in OFDM
33
at N=128, the theoretical results for N<128 are slightly more pessimistic at higher
PAPR levels.
Figure 3.1: Simulated envelope for OFDM system (N=64) normalized by average power.
.
2 3 4 5 6 7 8 9 10 11 12
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
N=32
N=64
N=128
N=256
Figure 3.2: Simulated (solid line) and theoretical (3.13, dashed line) OFDM symbol CCDF for N=32,
64, 128, and 256 subcarriers. QPSK, 30000 runs
0 0.5 1 1.5 2 2.5 3 3.5 4
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Amplitude, volts
P
r
o
b
a
b
i
l
i
t
y
Chapter 3: Peak to Average Power in OFDM
34
To provide further insight into the total distribution of the PAPR Figure 3.3 plots all
simulated OFDM samples for the same case as Figure 3.2. Here it is seen that the
sample distribution is lower than the symbol distribution as is predicted by (3.13).
4 2 0 2 4 6 8 10 12
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
N=256
N=128
N=64
N=32
Figure 3.3: Simulated OFDM sample CCDF for N=32, 64, 128, and 256 subcarriers. QPSK, 30000
runs.
The effect of various QAM mapping constellations is simulated in Figure 3.4 where it
is seen that changing the constellation has a minimum affect on the PAPR which is to
be expected considering the Mary constellations are normalized to have the same
average power. This same principle applies to non active subcarriers which also do
not influence the PAPR as the average power decreases in line with a reduction in the
number of active subcarriers. However null subcarriers have the advantageous side
effect of increasing the resolution of the OFDM symbol due to the oversampling
effect.
Chapter 3: Peak to Average Power in OFDM
35
2 3 4 5 6 7 8 9 10 11 12
10
4
10
3
10
2
10
1
10
0
P
r
(
ζ
>
ζ
0
)
ζ
0
(dB)
M=4
M=16
M=64
Figure 3.4: Simulated OFDM CCDF for M=4, 16, and 64 constellation mapping. N=64, 30000 runs.
This section has shown that the PAPR per symbol is only a function of the length, N,
of the IFFT. The constellation type and number of active subcarriers have a
negligible affect on the PAPR after modulation with the IFFT.
3.3 Oversampling discrete OFDM symbols to find true (continuous) peaks
Section 3.2 provided an analysis of the PAPR for critically sampled baseband OFDM
symbols. However this analysis does not reveal the peak of the band limited OFDM
signal. Oversampling the data in the IFFT increases the resolution of the OFDM
symbol giving a closer approximation to the band limited signal after filtering. This is
best explained in Figure 3.5 where the complex components of one OFDM symbol
with no oversampling is overlaid with the same symbol oversampled at the IFFT by a
factor of 8 (to approximate the continuous filtered signal). Note the parabolic
trajectory from one discrete sample to another, the growth in new peaks occurs in
between the discretely sampled peaks.
Chapter 3: Peak to Average Power in OFDM
36
0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
I channel
Q
c
h
a
n
n
e
l
P
3
P
1
P
2
Figure 3.5: Simulated OFDM symbol with no oversampling (dashed) with it’s oversampled version
(solid) overlaid on top. The solid circle represents the 6dB level with respect to the average power.
N=64, oversampling factors are 1 and 8.
Some interesting observations that can be made viewing Figure 3.5 are at P
1
where
the only sample of the critically sampled OFDM symbol is above 6dB, when
oversampled the true peak grows slightly larger still. At P
2
it is seen that the two
critical samples that make up its end points are well under 6dB, but after
oversampling a peak is produced in between the 2 critical samples. At P
3
we see that
the second largest peak in the oversampled case occurs at a position where no peak
existed in the critically sampled symbol. These results show how the critically
sampled OFDM symbol and its oversampled version can diverge greatly in PAPR.
The CCDF for various oversampling rates at the IFFT is shown in Figure 3.6, here it
is seen that an oversampling factor of 8 is sufficient to represent the continuous signal
and results in around 0.5dB increase in the PAPR.
Chapter 3: Peak to Average Power in OFDM
37
2 3 4 5 6 7 8 9 10 11
10
3
10
2
10
1
10
0
os=1
os=2
os=4
os=8
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
Figure 3.6: Simulated OFDM CCDF for oversampling rates of 1, 2, 4, and 8. N=64, QPSK, 15000
runs.
Equation (3.13) assumes that samples are mutually independent and uncorrelated,
however Parsevals theorem states that
1
2
0
N
m
m
x N
−
=
=
∑
(3.14)
Therefore the independent assumption of (3.13) is not true, especially in the
oversampling case where adjacent samples are highly correlated to each other.
Various papers [3436] have been published which address the issue of oversampling.
Reference [37] suggests that adding a number of extra independent samples to (3.13)
will give a closer approximation to the oversampled signal, (3.15)
( )
2
0
0
0
Pr 1 1
max
N
n N
e
α
ζ
ζ ζ
−
≤ <
> = − −
(3.15)
Chapter 3: Peak to Average Power in OFDM
38
where α=2.8 gives a good approximation to oversampled signals. Reference [35]
takes exception to this non theoretical approximation of the over sampled signal and
states that the bound is not close to the theoretical bound for large numbers of N.
Reference [35] developed a method for finding the exact peak distribution of band
limited Rayleigh processes giving an expression for the CCDF of the PAPR as (3.16)
( )
( )
( ) ( )
( ) 0 0
0 0
Pr 1 Pr
p p
N N
ζ ζ ζ ζ < = − >
( )
( )
( ) 0
0
1
0
p
N
p
p
N
N
ζ
= −
(3.16)
where
( )
( )
( )
( ) ( )
2
2 2
2 2
0
5
1
1
2 2 2
4
0
0
4 5 5
1 1
2 2 15
uerfc
u
u
p
a
N
N u e e u d du
φ
φ π
ζ φ φ φ
π
−
−
∞ ∞
− +
= − − −
¦  ¦
´ ` 
¹ \ ¹)
∫ ∫
which is the mean number of peaks above the level
0
ζ in one OFDM symbol, and
( ) 0 0.64
p
N N ≈ (3.17)
which is the mean number of total peaks.
As this method is numerically cumbersome to solve due to the double integration in
(3.16) a simpler approximation of the distribution is developed which is as accurate as
the previous method for a large number of N.
The simpler method derives the peak distribution of the band limited Rayleigh process
based on the level crossing rate approximation and then applies the result to the
derivation of the distribution of the CF in OFDM signals. This is made under the
assumption that 1) the complex components of the signal, x(t) are ideally band limited
Gaussian processes (i.e. N>64) and 2) the peaks are statistically uncorrelated. Also a
suitiably high level for a
0
must be chosen well above 0 to make the assumption valid,
Chapter 3: Peak to Average Power in OFDM
39
i.e. each positive crossing of the level
0
ζ has a single positive peak that is above the
level
0
ζ . The CDF of the CF is given as (3.18)
( ) ( )
( ) 0
0 0 0
Pr
p
N
C
F C
ζ
ζ ζ ζ ζ ζ ζ > = < > (3.18)
where
( )
0 p
N ζ is the mean number of peaks above
0
ζ , and can be approximated for
high
0
ζ by (3.19)
( )
2
0
0 0
3
p
N N e
ζ
π
ζ ζ
−
= (3.19)
An expression for the CDF, ( )
0 C
F ζ is then obtained:
( )
( )
0 0 0 C C
F F C ζ ζ ζ ≈ >
2
0
0 2
0
2
0
3
0
0
1
0
N e
e
e
ζ
π
ζ
ζ
ζ
ζ
ζ
−
−
−
¦
 
¹
¹
− 
=
 ´
\ ¹
¹
¹
¹
0 0
0 0
ζ ζ
ζ ζ
>
≤
(3.20)
The CCDF can then be expressed as
( )
0
1
C
F ζ − (3.21)
[35] suggests
0
ζ π = for QPSK modulation and a marginally lower value for 16
QAM. Figure 3.7 plots simulated CCDF of the PAPR for QPSK with N=64 and 512
with an oversampling factor of 16, and 20000 OFDM symbols against (3.15) and
(3.21). Here we see that the bound from (3.15) is closer to the simulated results for
N=64 as a small number of subcarriers does not have a completely Gaussian
distribution. On the other hand (3.21) is an excellent bound for N=512 where the
Gaussian assumption is true.
for
for
Chapter 3: Peak to Average Power in OFDM
40
4 5 6 7 8 9 10 11 12 13
10
4
10
3
10
2
10
1
10
0
P
r
(
ζ
>
ζ
0
)
Eq (3.15)
Eq (3.21)
Simulated
N=64
N=512
ζ
0
(dB)
Figure 3.7: Theoretical OFDM CCDF from (3.15) and (3.21) for N=64, 512 with simulated results:
QPSK, oversampling factor rate of 16, 20000 runs, N=64, 512.
Reference [36] finds bounds for the peak of the continuous envelope based on the
maximum of the oversampled sequence. This bound is used to derive a closed form
expression for the upper bound of the CCDF in an uncoded OFDM system with large
N. Unlike [35] where the PMEPR is derived under the assumption that OFDM
signals behave as band limited Gaussian processes, only the Gaussian assumption for
each sample is used, there is no assumption on the joint distribution of the samples.
The CCDF is given as
{ }
2
2
1
2
0
Pr
k
PMEPR kNe
π
λ
ζ
− −
> < (3.22)
where k is the oversampling factor, N is the number of subcarriers, and
0
ζ is the clip
value. Note that k must be > 2 π . Figure 3.8 plots the simulated results (QPSK,
os=16) against the theoretical results of (3.22). This method predicts a much higher
distribution of the peaks than is seen in simulation and is therefore very pessimistic.
Chapter 3: Peak to Average Power in OFDM
41
4 6 8 10 12 14 16 18 20
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
(3.22)
Simulated
N=64
N=512
Figure 3.8: Theoretical OFDM CCDF from (3.22) for N=64, 512 with simulated results: QPSK,
oversampling factor rate of 16, 20000 runs, N=64, 512.
Reference [38] extends the theory of [37], [35], and [36] to find new upper bounds for
different constellation types such as QAM and PSK, rather than just QPSK. The
theoretical results are useful at low probability regions where simulations are time
consuming. The upper bound on the CDF for QAM constellations is given as
( )
( )
( )
2 2
0
2 2 2
4 1
3 4
0
, ; 1, 2
min
k L
M a
M C C
L K L K
F KLNe
γ
ζ
− −
+
∈ > >
≤
ℕ
(3.23)
0
0 ζ ≥
where
1
cos
2
L
C
L
π
=
1 L > (3.24)
Chapter 3: Peak to Average Power in OFDM
42
( )
( )
( )
1
cos
3 cos
1 cos
k
K
C
K
K
π
π
π
=
−
+
3
3
K
K
>
≥
Keven
Kodd
(3.25)
where
0
ζ is the clip level, L is the oversampling factor, N is the number of
subcarriers, M is the constellation type: M=2 (4 QAM), 4 (16 QAM), or 8 (64 QAM),
γ =0 (4QAM), 1 (16 QAM), or 5 (64 QAM).
For BPSK constellation the CDF is given by
( ) ( )
2
2 2 1 2
1
1 2
1
1 2 2 , 2 , 2
0 0
, ; 1, 2
1 0
4 min
K L
N
LN
l l
K
C N C C
LN K
L K L K
l l
F B e
ζ
π π
ζ ζ
−
−
−
∈ > >
= =
≤ +
∑ ∑
ℕ
0
0 ζ ≥ (3.26)
where
( )
( )
( )
( ) ( )
1
sin
1
, , cos 1 2
2 2 sin
N
N
C N N
θ
θ α θ α
θ
= + − + (3.27)
and
( )
( )
0
0
/
1
/ 2 2
L
N
N
k N C keven
N
B
N k
ζ
ζ
=
=
−
∑
(3.28)
Results comparing the new CDF calculation against the Gaussian model of (3.13)
show that for BPSK the new bound is tighter below 10
4
probability for large N.
Interestingly the Gaussian approximation provides better results than the new bound
for all N. For 16 QAM and 64 QAM the Gaussian bound again looks better.
The CCDF results from Figure 3.6 imply that using an oversampling factor of 1
predicts the PAPR within 0.5dB, however this is not true when PAPR reduction
Chapter 3: Peak to Average Power in OFDM
43
techniques are applied to the OFDM symbol [30],[39]. Figure 3.9 shows the CCDF
of a 64 point IFFT OFDM modulator clipped at 3dB, interpolated by 8 and then
filtered with a 256 tap RRCF, as well as the same OFDM data set left unclipped at the
IFFT and then interpolated and filtered with the RRCF.
2 3 4 5 6 7 8 9 10 11 12
10
3
10
2
10
1
10
0
ζ
0
P
r
(
ζ
>
ζ
0
)
Clip at 3dB
and filter
No clipping and filter
Figure 3.9: Simulated OFDM CCDF QPSK, 64 point IFFT, 256 tap RRCF, α=0.15. Clipped at 3dB
after IFFT, then filtered (solid line). No clipping, then filtered (dashed line).
Peak regrowth after clipping and filtering is dramatic, much worse than without
clipping and filtering. The degree of peak regrowth is determined by the sequence of
the filtered data, the value of the excess bandwidth, α, of the RRCF (smaller α greater
peak regrowth), the length of zero padding in the IFFT, and the degree of clipping
(harder the clipping the greater the regrowth). Oversampling means the introduction
of null samples at the input of the IFFT as shown in Figure 3.10. The null samples
are introduced in the centre of the IFFT input to ensure that they occupy the outer
samples in the frequency spectrum.
Chapter 3: Peak to Average Power in OFDM
44
Figure 3.10: Zero padding of the IFFT, null carriers are set in the middle of the input
Figure 3.11 shows the simulated CCDF for a non oversampled IFFT compared to a 2
times oversampled IFFT, both of which are interpolated by a factor of 8 and filtered
with a RRCF.
4 5 6 7 8 9 10 11 12
10
3
10
2
10
1
10
0
P
r
(
ζ
>
ζ
0
)
after IFFT 128 pts
after filtering 128 pts
after filtering 64 pts
after IFFT 64 pts
ζ
0
(dB)
Figure 3.11: Simulated OFDM CCDF after IFFT (os=1 and 2) and after filtering (os=1 and 2). QPSK,
N=64, 256 tap RRCF, α=0.15, 15000 runs.
IFFT
X
0
X
1
X
(N/2)1
X
(N/2)
0
0
X
(N/2)+1
X
(N/2)2
X
N1
X
N
x
0
x
1
x
N×os
X
(N×os)1
Chapter 3: Peak to Average Power in OFDM
45
This time we notice that the oversampled IFFT has a slightly higher PAPR
distribution, but what is of most importance is that the PAPR distribution after
filtering has less peak regrowth and is below the critically sampled IFFT. We can
conclude that oversampling of the data at the IFFT is an effective way to counter the
peak regrowth after filtering. Oversampling beyond a factor of 2 provides a law of
diminishing returns in terms of peak regrowth (see Figure 4.10: U=1).
3.4 Effect of Non Linearity on OFDM
The previous sections described the causes of large PAPR and the problems of non
linearity, this section describes common models for HPA’s which are used in wireless
communications and the effect they have on the OFDM signal in terms of the Power
Spectral Density (PSD) and the increase in the BER. Section 3.4.1 describes
commonly used HPA models, section 3.4.2 details the corruption of the frequency
spectrum due to changing backoffs in the HPA, and finally section 3.4.3 describes the
degradation in BER due to changing backoffs in the HPA.
3.4.1 Description of memoryless Non Linearity
Non linearities provide the greatest obstacle to OFDM as a practical system due to
their distorting effect on the quality of the system. Here we concentrate on the most
common form of non linearity, distortion in the RF amplifier due to a limited linear
range in the amplifier. Papers [4044] tend to focus on the distortion due to the RF
amplifier stage as this is the most expensive component in a transmitter and takes up
to 50% of the cost and space in a unit. The RF amplifier must be driven as close as
possible to the maximum signal in the linear region to make it efficient, however
when operating near the saturation point it exhibits non linear behavior distorting the
transmitted signal. This distortion causes spectral regrowth in the transmitter which
can adversely affect adjacent frequency bands, and an increased BER at the receiver.
A balance must be met between allowable distortion and the linear region of an
amplifier. Therefore it is pertinent to evaluate the performance of OFDM signals
through different non linear devices.
Chapter 3: Peak to Average Power in OFDM
46
A convenient form of expression for non linear devices is in polar coordinates
reference [45]. The input can be expressed as (3.29)
( ) arg x j
x x e e
φ
ρ = = (3.29)
Therefore the complex envelope of the output signal can be expressed as (3.30)
( ) [ ]
[ ] ( ) j
g x F e
φ ψ ρ
ρ
+
= (3.30)
where [ ] F ρ and [ ] ψ ρ represent the AM/AM and AM/PM conversion
characteristics of the memoryless non linear amplifier respectively. Some commonly
used models follow.
Soft Limiter (SL)
The Amplitude Modulation to Amplitude Modulation (AM/AM) and Amplitude
Modulation to Phase Modulation (AM/PM) characteristics of a SL can be expressed
as [43]
[ ]
A
F
A
ρ ρ
−
=
if A
if A A
if A
ρ
ρ
ρ
< −
− ≤ ≤
>
(3.31)
and [ ] 0 ψ ρ = .
There is no phase distortion in this model, only amplitude distortion, A is the clipping
level of the amplifier. The AM/AM characteristics of a SL are plotted in Figure 3.12.
Chapter 3: Peak to Average Power in OFDM
47
0
0
Input voltage
O
u
t
p
u
t
v
o
l
t
a
g
e
A
A
A
A
Figure 3.12: AM/AM properties of a Soft Limiter (SL)
Travelling Wave Tube Amplifier (TWTA)
According to reference [40] the AM/AM and AM/PM functions are
[ ]
( )
2
1
a
F
ρ
ρ
β ρ
=
+
(3.32)
and
[ ]
( )
2
2
1
ϕ
ϕ
α ρ
ψ ρ
β ρ
=
+
(3.33)
A common choice for the above parameters is 0.25
a
β = ,
12
ϕ
π
α = , and 0.25
ϕ
β = .
The AM/AM characteristics of a TWTA are plotted in Figure 3.13.
Chapter 3: Peak to Average Power in OFDM
48
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Input voltage
O
u
t
p
u
t
v
o
l
t
a
g
e
Ideal
TWT
Figure 3.13: AM/AM properties of a Traveling Wave Tube Amplifier (TWTA)
Solid State Power Amplifier (SSPA)
Probably the most common and practical model for amplifiers is the SSPA [40]. The
AM/AM and AM/PM transfer characteristics can be modeled as
[ ]
1
2
2
1
P
P
F
A
ρ
ρ
ρ
=
1
 
+
1

\ ¹
1
¸ ]
(3.34)
and
[ ] 0 ψ ρ = (3.35)
The parameter P controls the smoothness of the transition from the linear region into
the saturation region. When P →∞ the SSPA acts as a SL, P=3 is a good
Chapter 3: Peak to Average Power in OFDM
49
approximation of a practical amplifier. The AM/AM characteristics of a SSPA and an
ideal amplifier are plotted in Figure 3.14.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Input voltage
O
u
t
p
u
t
v
o
l
t
a
g
e
Ideal
P=12
P=3
P=1
Figure 3.14: AM/AM properties of a Solid State Amplifier (SSPA) for different values of P, and ideal
amplifier.
In all amplifiers discussed in this section ‘A’ represents the saturating amplitude of
the amplifier. The non linear distortion depends on the backoff of the amplifier and
can be calculated as either the Input BackOff (IBO) or the Output BackOff (OBO),
and is defined as
2
10
10log
s
IN
A
IBO
P
≜ (3.36)
2
10
10log
OUT
A
OBO
P
≜ (3.37)
Chapter 3: Peak to Average Power in OFDM
50
where A
s
is the amplifier input saturation voltage, P
IN
is the average power at the
input, A is the saturating amplitude at the output, and P
OUT
is the average power at the
output.
3.4.2 Impact on Power Spectral Density
The simulated spectrum for N=64 subcarriers is shown in Figure 3.15. The PSD is
measured for each OFDM block then averaged over 2000 blocks to eliminate the
effect of
the rectangular time window [46]. Blocks of 4 QAM data were mapped to a 64 point
IFFT for modulation, the data is then interpolated by a factor of 8 and filtered. A
RRCF was used for pulse shaping with 128 taps and an excess bandwidth of 0.15.
The low rolloff factor of the filter results in a fast drop off of the spectral splatter
outside the normalised FFT bandwidth. A SSPA with P=3 is used as the amplifier
model. The frequency axis is normalized.
Changing the IBO of the SSPA results in inband distortion and spectral regrowth, or
splatter outside the normalized frequency bandwidth of the filter. An amplifier
backoff of 0dB results in noise power that is only 16dB lower than signal power.
With a backoff of 3dB the noise power is 19.5dB below the signal power. A 6dB
amplifier backoff results in a slight amount of spectral splatter 23dB below the signal
power and an almost indistinguishable amount of in band distortion. With an infinite
amplifier backoff the PSD matches exactly the PSD after transmit filtering.
OFDM standards such as Hiperlan2 and 802.11a require adjacent OFDM carriers to
be closely packed in the frequency domain with overlapping spectrums. This is
achieved by
Chapter 3: Peak to Average Power in OFDM
51
4 3 2 1 0 1 2 3 4
50
45
40
35
30
25
20
15
10
5
0
Hz
A
b
s
o
l
u
t
e
p
o
w
e
r
(
d
B
)
HPA=0dB
HPA=3dB
HPA=6dB
16dB 19.5dB 23dB
LPA
Figure 3.15: PSD of 64 subcarrier OFDM signal with 64 point IFFT, RRCF with excess bandwidth of
0.15. A SSPA (HPA in figure) with P=3 and various backoffs is used.
setting null subcarriers at the edge of the spectrum, where the rolloff of the filters
occurs in the unused subcarriers, easing the design constraints on the filters.
Filtering after the SSPA to reduce spectral splattering is complex, therefore reference
[46] proposed clipping in the baseband after the IFFT followed by filtering. This
successfully reduces the out of band distortion but the BER due to the in band
distortion remains. Also peak regrowth after filtering remains a problem. These
issues are explored in more depth in chapters 5, 7.
3.4.3 Impact on Bit Error Rate
The inband distortion due to non linear amplification at the transmitter results in ISI
when filtered with the matched filter at the receiver, resulting in an increase in the
BER when the data is decoded. Two observed effects of clipping (whether due to the
amplifier non lineraity or baseband clipping) are a Gaussian like spreading of the
Chapter 3: Peak to Average Power in OFDM
52
decoded constellation points and a compaction of the whole received constellation,
shown in Figure 3.16 where the two constellation types (4 and 16 QAM) are
normalized to have the same average power.
Figure 3.16: Signal constellation at the output of the SSPA, P=3 after non linear amplification a) 4
QAM, 6dB IBO; b) 4 QAM, 0dB IBO; c) 16 QAM 6dB IBO; and d) 16 QAM 0dB IBO.
The noise due to clipping is evenly spread across all subcarriers [47], this is because
each sample that is clipped is a conglomeration of all the input samples to the IFFT.
Therefore the BER is nearly equal on all subcarriers.
Figure 3.17 shows the BER due to a non linearity in the HPA for 3 Mary
constellation types. 4 QAM is very impervious to clipping noise as there is a greater
Euclidean distance between constellation points. Indeed 4 QAM has a very
acceptable BER performance without any coding or PAPR correction. Mapping types
16 and 64 QAM are much more susceptible to clipping in the amplifier. The BER of
Figure 3.17 could be improved with Automatic Gain Control (AGC) in the receiver
which would expand the constellation to the correct size before decoding [48].
Chapter 3: Peak to Average Power in OFDM
53
4 2 0 2 4 6 8
10
6
10
5
10
4
10
3
10
2
10
1
10
0
HPA backoff (dB)
B
E
R
4 QAM
16 QAM
64 QAM
Figure 3.17: BER of 64 subcarrier OFDM signal with 64 point IFFT, RRCF with excess bandwidth of
0.15 for 4, 16, and 64 QAM constellations. A SSPA (HPA on xaxis) with P=3 and various backoffs is
used.
3.5 Conclusion
This chapter defined the problem of PAPR in OFDM beginning with a mathematical
analysis of the PAPR in both the baseband and passband. Next the stochastic
distribution of OFDM samples was shown through analytical and simulated means to
only be a function of the number of subcarriers. It was also shown that the general
distribution of samples is quite low and that the transmit envelope follows a Rayleigh
distribution.
Oversampling of the IFFT was also introduced as an important issue, it was shown
that oversampling the IFFT increased the CCDF by ~0.5dB. These results were
supported analytically by various papers which analysed the issue of oversampling.
Chapter 3: Peak to Average Power in OFDM
54
Furthermore, it was shown through simulation that filtering has a dramatic affect on
the CCDF when combined with PAPR reduction techniques such as clipping. The
peak regrowth is much worse than when no PAPR reduction techniques are used.
Oversampling at the IFFT was shown to both increase the discrete CCDF and reduce
the continuous, filtered CCDF.
Non linearities were also examined with models of the most common form of non
linearity, the SSPA, described mathematically. The impact of non linearity on the
PSD and BER was simulated. It was shown that reducing the IBO of the SSPA
causes both spectral spreading, affecting the adjacent channels, and in band distortion
corrupting the BER at the receiver. Higher order constellations were shown to be
more susceptible to lower backoffs in the SSPA.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
55
Chapter 4
Peak to Average Power Solutions 
Distortionless Techniques
The previous chapter outlined the disruptive effects of an uncontrolled OFDM signal
envelope on system performance. This chapter and chapter 6 provide an analysis of
solutions to the PAPR problem, each solution has advantages and disadvantages in
terms of PAPR reduction, distortion of data, and complexity. This chapter reviews
distortionless PAPR reduction techniques. Distortionless techniques do not corrupt
the data and encode it in such a way that it can be completely recovered at the
receiver, however they are usually more complex. Specifically Section 4.1 explains
the family of coding techniques for PAPR reduction, Section 4.2 elaborates on
Multiple Signal Representation (MSR) and phase rotating techniques, Section 4.3
details modified constellation techniques, and finally Section 4.4 summarizes the
chapter and provides a comparison in terms of complexity and performance of the
detailed techniques.
The other side of PAPR reduction are distortion introducing techniques, these
techniques deliberately attenuate the envelope of the signal corrupting the BER as
described in Section 3.4, however things can be done to lessen the effect of the
introduced distortion. These techniques will be explored in Chapter 6.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
56
4.1 Coding techniques
Many early papers considered how standard coding techniques could be applied to
OFDM. The basic premise of coding is to insert redundant bits into the data stream
which can be used for error correction at the receiver. Their application to PAPR
reduction is in creating sequences of bits which will exhibit low PAPR after the IFFT.
There are 2 types of error detection and correction codes, block codes and
convolutional codes. Most papers relate to the block coding family for PAPR
reduction. During the encoding process k information bits are encoded into n code d
bits, therefore (nk) redundant non information bits are added to the k information bits
[16]. The block code is referred to as an (n,k) code, and the rate of the code as
R
c
=k/n. Figure 4.1 is a block diagram showing where coding for PAPR reduction is
located in an OFDM transmitter.
Figure 4.1: Block diagram of OFDM transmitter showing PAPR coding
The ability of a code to correct errors is a function of the code distance, (4.1)
( ) ( )
, ,
1
, mod
N
i j i l j l
l
d C C C C q
=
= ⊕ −
∑
(4.1)
where d is the distance of the codeword and q is the number of possible values of C
j
and C
l
. The smallest distance d
min
is the minimum distance for a given set, (4.2)
( ) { }
min
,
i j
d Min d C C = (4.2)
Bit
source
PAPR
coding
IFFT
Serial
To
parallel
n
k n+k
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
57
Different codes exhibit different degrees of error correction ability. Another
important property of codes is the weight of the code, which is the number of non zero
elements in the codeword. Types of block codes are Hamming, Golay, and Reed
Solomon, some of which are used for PAPR reduction.
4.1.1 Block Codes
The first paper to apply coding techniques for PAPR reduction in OFDM was
reference [49]. The basic premise of this paper was to determine which combinations
of data at the IFFT input produced large peaks at the output and to avoid transmitting
these sequences by adding redundant bits to the input. Initially a simple (impractical)
OFDM system with 4 subcarriers and BPSK modulation is considered. A 3 bit data
word is mapped on to a 4 bit code word, ie (3,4) block code, so that the set of
allowable code words does not create excessive envelope spikes. They identify the
code as an odd parity code and state that the PEP is reduced from 6.02dB to 2.48dB, a
reduction of 3.54dB.
An 8 subcarrier system with BPSK constellation is also evaluated to illustrate how
block coding can be traded of against PAPR reduction. The permissible number of
codewords CW
perm
is traded off against the total possible number of codewords,
CW
poss
. With no block coding the PAPR is 9.03dB. If half the code words are
allowed, i.e. a (7,8) rate code, the resultant PAPR is 4.45dB. If a quarter of the code
words are allowed, a (3,4) rate code the PAPR is 3.01dB.
The PAPR reduction comes at the cost of an increase of bandwidth for the same data
rate and a reduction of the energy per transmitted bit for the same transmit power.
However the increase in bandwidth is small and is offset by the high spectral
efficiency of OFDM, as is the reduction in energy per transmitted bit which is offset
by the possible error detection/correction potential of block coding. No results on the
minimum distance of the code are given but the authors indicate that a large number
of the codes found are Golay complementary sequences. These codes are explored in
greater depth in later papers. Practical OFDM systems employ at least 64 subcarriers
and higher order mapping types which make the complexity of a block coding scheme
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
58
grow exponentially as the coding must be done in combinational logic or a Look Up
Table (LUT).
Aware of the limitations of the previous paper Jones and Wilkinson extended the
block coding principle in reference [50]. 4 and 8 subcarrier systems are evaluated
using QPSK mapping. They prove that the code rate is unaffected by changing the
number of carriers and mapping types. 3 areas are identified to make coding
techniques more practical
• Selection of suitable code words for any number of carriers and any Mary
mapping type.
• Selection of code words that allow efficient implementation of the
coding/decoding.
• Selection of code words that also offer error detection/correction properties.
The authors review a set of papers published in the 80’s which addressed a similar
problem for multitone test signals with low PAPR. Although certain parameters
differ much of this work can be applied to the OFDM case.
The reverse of a code word will result in the same PAPR, i.e. for the 4 subcarrier
BPSK case ‘1000’ will give the same PAPR as ‘0001’. Codes such as these are
derived from ShapiroRudin sequences which are a subset of Golay complementary
sequences. A definition of Golay complementary sequences is given in the paper as
“a pair of equally long, finite sequences whose aperiodic autocorrelations sum to zero
for all nonzero displacements”. Analysis of the code words that exhibited low
PAPR’s for 4 and 8 subcarriers revealed them to be Golay complementary sequences,
however they did not give the optimum minimum solution, but they are amenable to
mathematical encoding and decoding and the authors note that they may have error
correction properties.
For an 8 subcarrier BPSK OFDM signal based on Golay complementary sequences a
(5,8) rate code can give a PAPR of 3.01dB. Larger code sets can be found by
rewriting the definition of an OFDM signal (4.3)
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
59
( ) ( )
( ) 2
1
n n
n N
j f t
n
n
s t d t e
π φ
=
+
=
=
∑
(4.3)
as (4.4)
( ) ( )
( )
( )
( )
/ 2
2 1 2 1
1
s s
n N
j n f t j n f t
n n
n
x t d t e d t e
π π
=
+ − − −
−
=
= +
∑
(4.4)
This form gives a term in the summation as 2 carriers that are equidistant from and on
either side of the centre frequency of the complex envelope representation. If the d
n
data words are chosen so that the resultant vector lie on one of two orthogonal lines,
then the PAPR is less than or equal to N/2. The principle of these sequences can be
extended to give polyphase sequences suitable for multilevel phase modulation. The
search for better sets of code words with error detection/correction properties is
addressed in later papers.
The authors of [49, 50] presented another paper, [51], in which they describe
Combined Coded OFDM (CCOFDM). CCOFDM attempts to exploit the error
detection/correction properties presented in reference [50] while still maintaining the
PAPR suppression. Polyphase weighting codes are applied to the encoded data
(chosen from a low PAPR set) which are known at the receiver. Thus they can be
compensated for without affecting the distance properties of the code. The BER in a
non linear channel for CCOFDM is compared to the COFDM channel and significant
improvement in the BER is seen in a highly non linear channel. Still, the work is
limited to 8 subcarriers.
Reference [52] expanded on the work of reference [51] by developing an algorithm to
compute the phases that minimize the PAPR for larger sets of data of practical
interest. The authors found sets of phase values (calculated offline) which are known
at both the transmitter and receiver that reduce the PAPR without affecting the error
correction properties. These phase shifts that calculated for various coding rates in
the Hiperlan2 [53] standard where there are 48 information bearing subcarriers. For
½ rate BPSK with 90° phase shifts a reduction of 4.09dB is reported. The reduction
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
60
can be improved slightly by using smaller phase shifts such as 8PSK and 16PSK. The
authors elaborated on the results given in [52] in [54] where the computation of the
PAPR is given in a rigorous mathematical proof.
Going back in time another early paper [55] extended the work in reference [49] by
identifying Quadature Phase Shift Keying (QPSK) message structures which produce
high PAPR. They are grouped into equivalence classes where messages which have
the same PAPR are in the same equivalence class. Up to N=5 subcarriers are
analyzed with different coding rates. They show that that a small amount of
redundancy can significantly reduce the PAPR.
The authors of reference [55] developed their results in references [55, 56] to provide
bounds for the PEP of OFDM using basic coding techniques as described in reference
[49] for up to 16 subcarriers. They prove through analytic and simulated results that
only 4 bits of redundancy are required to reduce the PEP to within 10% of its
optimum value as the number of subcarriers is increased. They show that further
redundancy provides little benefit in terms of PAPR reduction.
Reference [57] presented an idea again based on reference [49] where vectors or
messages of data which exhibit a high PAPR are attenuated. The amplitude of
subcarriers which are above a given threshold are uniformly reduced to achieve
equality between the maximum of the envelope power and the threshold level. Also
where the envelope power is below the threshold these subcarriers are increased to
obtain equality. This allows the PAPR to be reduced without affecting the net bit rate.
Results are given for a BPSK and QPSK with less than 20 subcarriers. The PAPR is
reduced significantly however this comes at the cost of a reduction in the SNR which
becomes more pronounced as the number of subcarriers is increased.
Reference [58] is another early paper which draws a link between the number of
subcarriers and the mapping type used (4 Phase Shift Keying (PSK), 8 PSK, etc). It is
stated that if sub carriers in the same group (where a group is a set of subcarrier
frequencies sharing some relation) are phase shifted by the same amount, then the
envelope remains unchanged. This concept is used to create code sets with low PAPR
properties. ReedSolomon codes are examined with parameters N=16 and M=4.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
61
They also note that the constraints on the code can be loosened up if at least some
variation in the envelope is allowed.
4.1.2 Bounds on PAPR
Reference [52] provided an efficient computational method for finding offsets which
can be used with coding in OFDM, which could allow large reductions in the PAPR,
but no concrete level on the PAPR has been proven. Reference [59] addressed this
issue by providing bounds for the PAPR for different error correction coding schemes
coined trace codes. Duals of primitive BCH codes are identified as good error
correction codes.
Reference [60] presented an interesting paper which clarified the relation between the
PAPR and the out of phase aperiodic autocorrelation of the message or data sequence.
The PAPR as defined in (3.3) can be bounded as (4.5)
( )
1
1
2
1
N
k
k
N
ζ ρ
−
=
≤ +
∑
(4.5)
where
( )
*
1
N k
n k n
n
k a a ρ
−
+
=
=
∑
for k=0, …N1 (4.6)
and a
n
is the input data sequence to the IFFT. (4.6) shows that binary or polyphase
sequences with low out of phase aperiodic correlation values can be used to construct
low PAPR signals. The problem remains to find sequences that reduce the PAPR.
(4.6) is useful as it is a much less complex process to calculate than (3.3).
4.1.3 Cyclic Codes
Reference [61] developed a simple method based on ¾ rate code rate cyclic coding
which can reduce the PAPR by 3dB when the number of subcarriers is a multiple of 4.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
62
The phase of every 4
th
subcarrier is calculated so as to minimize the amplitude giving
a ¾ rate code. The method is simple to implement.
Reference [62] achieves the same results as [61] but with reduced complexity. The
authors also introduced Sub Block Coding (SBC) where systems with a large number
of subcarriers are divided into subblocks with the last bit of each subblock altered
according to the method described in [49]. By dividing the OFDM frame into sub
blocks larger number of subcarriers can be used while still maintaining a reduced
PAPR, which is a problem with other earlier coding schemes. The idea is developed
to optimize the positions of the odd parity checking bits for further PAPR reduction at
the cost of the introduction of side information to inform the receiver of the positions
of odd parity checking bits.
4.1.4 ShapiroRudin codes
A very early paper [63] applied ShapiroRudin and Newmen phases to multitone
frequency response testing. Although the application was not for multicarrier
applications the theory still holds with multicarrier signals. Reference [63] stated that
a multitone signal can achieve a CF under 6dB for ShapiroRudin phases and around
4.6dB using Newman phases for an arbitrarily large number of subcarriers. Many
definitions which were to become convention in later papers as far as the statement of
the problem of large PAPR, as well as identifying some very pertinent parameters in
the makeup of multitone signals. It was also noted that the set of tones needs to be a
power of 2 (which is the case in all OFDM standards) in order for the codes to be
optimum. Reference [63] closes the paper with an open question as to whether the CF
can be reduced lower than 3dB. The authors also state that Newman phases which
vary quadratically exhibit a lower CF than ShapiroRudin sequences which vary
linearly for all N.
Eleven years later reference [64] revisited the application of ShapiroRudin sequences
to reduce the PAPR in OFDM. In [64] a ½ rate code is employed for QPSK signals
with 16 information carriers giving 32 subcarriers in total. 4 bit input messages are
concatenated to 4 bit codewords which are determined according to the message data.
A simple digital circuit using eight two input XOR gates can be used for the encoder.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
63
Results taken from a limited subset of possible messages suggest that the coded signal
will reduce the PAPR to ¼ of the original uncoded message.
4.1.5 Golay complementary codes
Golay codes are linear binary block codes which are the only non trivial example of a
perfect code. As every codeword lies within distance 3 of any codeword they can be
used in conjunction with maximum likelihood detection [16] for decoding which is
not overly complex to implement.
In reference [65] the use of Golay complementary codes is examined. Golay
sequences were first recognized to have good PAPR properties for application in
OFDM in reference [50]. An algorithm was developed where certain subsets of codes
up to length 16 have a minimum distance of half the code length and have a PAPR of
3dB. Existing Forward Error Correction (FEC) codes are incorporated into PAPR
techniques with a new decoding algorithm developed which utilizes the efficient
inverse WalshHadamard transform. However the new decoding algorithm has 3dB
worse performance than the optimum maximum likelihood detection.
The author notes that the scheme may be unfeasible for a large number of subcarriers
as the length of the codes is the same as the number of subcarriers. However
reference [65] proposed ways to nullify this effect at the cost of the PAPR reduction
made and error correction capability by breaking the total number of subchannels into
smaller groups and applying a complementary code to each group.
4.1.6 ReedMuller Codes
Golay complementary sequences were further developed for PAPR reduction in
reference [66] where again the PMEPR is found to be at most 2 (3dB) when the data
sequence is constrained to be a member of a Golay complementary pair. However
Golay pairs have a high overhead in terms of redundancy and may not be a practical
coding solution for OFDM. Therefore the authors developed a more suitable coding
method by recognizing the relation between Golay complementary sequences and
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
64
second order ReedMuller codes (RM(2,m)), i.e. in the binary case Golay sequences
occur as cosets of the first order ReedMuller code within the second order Reed
Muller code. Standard decoders can be used in the receiver for the RM(2,m).
RM(2,m) allow the code rate to be improved by allowing a slight increase in the
PMEPR. RM(2,m) further developed the theory of reference [65] by allowing a trade
off to be made between the Hamming distance, PMEPR reduction made, code rate,
and the number of phases allowed in the PSK mapping type.
The authors of reference [66] advanced their work further in reference [67] providing
many mathematical proofs for Golay sequences, ReedMuller codes, decoding using
fast Hadamard transforms. Basically this paper formalized the results given in
reference [66] and extended the results for larger sets of variables such as an increase
in the PSK mapping type. However a limitation which the authors note is that the
codes are limited to 32 subcarriers where the resulting code rate will be high.
Reference [68] examined the ReedMuller coding scheme presented in reference [67]
through simulation of an end to end system with various non linearity’s and 16
subcarriers. It was shown that at 40dB ACI the new coding scheme had around a
12dB gain in IBO over an uncoded system in experimental results, however this gain
drops to 4dB for the simulated results.
Reference [69] also uses ReedMuller codes in a simulation environment to test their
performance in the presence of AWGN to determine the BER with QPSK, and 8 PSK
mapping types, and different code lengths, m. The number of subcarriers is chosen
such that N=2
m
, the length of the code. Results indicated that increasing the code
length improves error correction capabilities for high SNR values, but at low values of
SNR the uncoded system has better performance. As the code rate decreases
performance is improved. QPSK and 8PSK are compared in terms of their BER, and
it was shown that to maintain the same BER the code length for 8PSK must be
increased.
Reference [70] also enlarged on the work of [67] by developing new decoding
algorithms with generalized fast Hadamard transforms. The complexity and
performance of their decoding algorithm is compared to the standard maximum
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
65
likelihood method. Suboptimal algorithms are also presented with reduced
complexity which were shown to have minimal degradation over optimal methods.
Another paper to address decoding issues is [71] which used soft decision decoding
methods for block codes. The performance is evaluated in an AWGN channel with
PSK modulation and 8 and 16 subcarriers. The authors point out the inherent problem
with coding schemes i.e. the decrease in coding rate with the increase in the number
of subcarriers but state as others have that the PAPR reduction gained can be traded
off for reduction in complexity and code length. The new block coding method
coined Ordered Statistic Codewords (OSD). BERs for (8,4) and (16,5) block codes
utilizing 8 and 16 PSK show that with partial order decoding almost optimum results
are achieved compared to the union bound. Complexity compared to MLD methods
is greatly reduced at the expense of a minor error increase.
A limitation of references [67, 68] is that PSK mapping types are used. Most
practical OFDM standards use QAM mapping [53]. Reference [72] addresses this
limitation by combining QPSK constellations to form any Mary QAM constellation.
This is achieved by using shift and rotation operations as defined in (4.7)
( ) ( )
1 1
1
2
0
2
2 exp
2 4
i
i
n
x
i
M
i
j
QAM j
π
−
=
=
∑
(4.7)
where the QPSK constellations can be expressed as
i
x
j , where { }
4
0,1, 2, 3
i
x Z ∈ = .
Any QPSK sequence ( )
0 1 1
...
N
a a a
−
= a can be associated with another sequence
( )
0 1 1
...
N
i i i i
x x x
−
= x where the elements of x
i
are in Z
4.
Golay and ReedMuller
sequences can now be created using the theory developed in reference [67].
Reference [73] extended on the coding theory of references [49, 50] by finding codes
that simultaneously reduce both the PAPR and the ICI which is introduced by
frequency offset between the transmitter and receiver. They define a new
measurement term, PeakIntercarriertoCarrier Interference (PICR) to quantify their
results. As minimum PICR and PAPR do not occur in the same code word, a balance
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
66
is found where both are at an acceptable level. Simulations performed on a
rudimentary system with 16 subcarriers and BPSK modulation in AWGN show that
the PAPR is reduced by 3dB with a decrease in the PICR. The authors note that
finding codes for higher mapping types and more subcarriers is difficult.
Coding techniques while popular in early OFDM papers have since fallen out of
favour. Their performance would be largely negated in a practical communications
system due to the interleaving stage which follows the coding. It is also worth noting
that none of the papers referenced consider the effect of over sampling.
4.2 Multiple Signal Representation
Multiple Signal Representation (MSR) techniques are another distortionless method
for PAPR reduction. The basic premise of MSR is to produce a set of alternative
transmit signals seeded from the same data source. Various techniques are used to
encode the alternative sets of transmit signals, which are encoded in such a way so
that they will have different PAPR properties. The transmit signal with the lowest
PAPR is chosen for transmission.
4.2.1 Partial Transmit Signals
Reference [74] first proposed Partial Transmit Sequences (PTS). PTS generates a
signal with a low PAPR through the addition of appropriately phase rotated signal
parts. The original signal is given by (2.2) and is reproduced in (4.8)
1
2
0
1
.
N
j nk
N
m m
k
x X e
N
π
−
=
=
∑
where 0 1 n N ≤ ≤ − (4.8)
The signal to be transmitted is broken up into several sub blocks,
, m k
X , of length N/V
(where N is the number of sub carriers and V is the number of sub blocks). All
subcarrier positions which are occupied in another block are set to 0, i.e.
,
1
V
m m k
k
X X
=
=
∑
Next a constant phase rotation,
,
,
m k
j
m k
p e
φ +
= , [ )
,
0, 2
m k
φ π ∈ 1<k<V is
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
67
performed on each subblock except for the first one which is kept constant, giving
(4.9)
, ,
1
.
V
m m k m k
k
X p X
=
=
∑
ɶ
(4.9)
The information in
m
X
ɶ
is the same as
m
X but with an added phase rotation, which
must be known at the receiver. An IFFT is performed on each subblock which are
then all summed together to create a possible transmit symbol, (4.10)
{ }
, , , ,
1 1
. .
V V
m m k m k m k m k
k k
x p IFFT X p x
= =
= =
∑ ∑
ɶ (4.10)
The process is repeated again with a different phase rotation,
, m k
p , to produce another
alternative transmit signal. The optimum parameters for the transmit symbol are
(4.11)
{ }
{ } ,1 ,
,1 , , ,
0 1
1
max .
argmin
m m V
V
m m V m k m k
n N
k
p p
p p p x
≤ < −
=
=
∑
ɶ ɶ …
ɶ ɶ … (4.11)
The optimum phase angle,
, m k
pɶ , is obviously the one where the PAPR is minimized.
Therefore the actual transmit signal is given as (4.12)
, ,
1
.
V
m m k m k
k
x p x
=
=
∑
ɶ ɶ (4.12)
Figure 4.2 shows a block diagram of a PTS transmitter.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
68
Figure 4.2: Block diagram of PAPR reduction using the PTS approach.
PTS requires side information to be sent to the receiver to inform it of the phase
rotation used so the data can be decoded. Reference [74] noted that the number of
angles should be kept low to keep the side information to a minimum. If each phase
rotation is chosen from a set of W admissible angles then the required number of bits
for side information is, R
ap
=(V1)log
2
W bits per OFDM symbol. In order to reduce
complexity the phase angles should be restricted to { } 1, j ± ± , i.e. W=4, this allows
multiplications to be performed with sign changes. Simulations shown in Chapter 5
reveal that increasing the number of allowed phase angles has a minimum impact of
PAPR reduction. Reference [75] noted that explicit side information can be avoided
if differential encoding is used for the modulation across the subcarriers within each
subblock. In this case only the block partitioning need to be known at the receiver
and one subcarrier in each subblock must be left unmodulated as a reference carrier.
Side Information
(if needed)
m
xɶ
Bit source
Coding
Mapping
Sub block
partitioning
IFFT
+
IFFT
IFFT
Peak value optimization
,1 m
X
,2 m
X
, mV
X
,1 m
x
,2 m
x
, mV
x
,2 m
pɶ
, mV
pɶ
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
69
PTS is flexible as the number of blocks and phase rotations can be increased
providing more alternative transmit signals to choose from. The disadvantage of this
scheme is the complexity, especially with an increase in V and W. Also, a large
amount of memory is required to store the alternative transmit signals (if check
performed in parallel) in order to compare them to find the one with the lowest peak
value. Alternatively the optimisation can be performed in an iterative fashion where
the current best transit signal is stored until a better one is found, at the cost of
increased latency.
It should also be noted that the data can be divided into sub blocks in different ways
as noted in reference [76] and shown in Figure 4.3. Different PTS sub block
structures have varying performance with pseudo random having the best and
interleaving having the worst. Of course there is a trade off with complexity,
interleaved sub blocks are the least complex PTS structure to implement (The size of
the IFFT’s can be halved by interleaving the input data to the IFFT and performing
the last stage of the IFFT operation at the IFFT output [77].) and pseudo random PTS
is the most complex (random sequences require more hardware complexity to
implement in this case). Also computationally efficient IFFT’s can be used to exploit
the number of zero’s in the PTS subblocks. Results from reference [74] indicate that
pseudorandom partitioning is 0.5 to 0.9dB better than adjacent partitioning.
It should be noted that no guaranteed level of PAPR reduction can be provided with
MSR techniques, all they can do is reduce the probability of large peaks. Figures 4.4
and 4.5 show CCDF results with varying factors of V and W, N is set at 64. The
simulation model is described in Figure 5.1.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
70
X
(0 )
n
X
(1)
n
X
(2)
N1 n
(a)
X
( 0 )
n
X
( 1 )
n
X
( 2 )
N  1 n
( b )
X
(0 )
n
X
(1)
n
X
(2)
N1 n
( c)
Figure 4.3: An example of the 3 main PTS structures: (a) Interleaved (b) Adjacent (c) Pseudorandom.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
71
2 3 4 5 6 7 8 9 10 11 12
10
4
10
3
10
2
10
1
10
0
ζ
o
(dB)
P
r
(
ζ
>
ζ
0
)
Uncoded
V=2, W=4
V=2, W=8
Figure 4.4: Simulated CCDF for PTSOFDM with V=2 and varying W. N=64, adjacent subblock
partitioning.
2 3 4 5 6 7 8 9 10 11 12
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
Uncoded
V=2, W=4
V=3, W=4
V=4, W=4
Figure 4.5: Simulated CCDF for PTSOFDM with W=4 and varying V. N=64, adjacent subblock
partitioning.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
72
4.2.2 Oversampling PTS
Early on in the development of PTS a problem was identified which skewed the true
gains of the proposed technique. As already mentioned in section 3.3 oversampling at
the IFFT is important to get a true reflection of PAPR reduction. Reference [34] first
identified this issue and showed that for a PTS system with V=4, W=4 the PAPR
reduction gained using the True Peak Factor (TPF) i.e. oversampled signal is only
1dB better than the uncoded case. In reference [34] it is shown that the true peak will
move away from the discrete sampled points. There is a 3dB difference in the
reduction between TPF and Lower Peak Factor (LPF), i.e. the discrete PAPR.
Reference [34] goes on to suggest a new optimisation technique based on the
aperiodic autocorrelation of
, m k
x ,
k
λ .
{ }
{ } ,1 ,
1
,1 ,
1
argmin
m m V
N
m m V k
k
p p
p p λ
−
=
=
∑
ɶ ɶ …
ɶ ɶ … (4.10)
A reduction of 2.5dB can be achieved when using the optimisation criteria of (4.10) at
Pr(ζ>ζ
0
)=10
5
. However this process is complex, for QPSK the calculation of
k
λ can
be achieved with less multiplications and replaced with integer additions giving a total
complexity of O(4
V1
N
2
).
Reference [39] acknowledges the limitations of discrete sampled PTS and notes that
the mismatch between discrete and continuous CCDF’s occurs due to DSP filtering
after the IFFT, peak regrowth becomes more pronounced with a sharper rolloff, α, of
the filter. Reference [39] suggests that oversampling the IFFT by a factor of 2 (zero
padding) will reduce the peak regrowth affects after filtering. Oversampling of PTS is
explored in more depth in Chapter 5.
Another paper, [78], to explore the PTS approach looked at alternative ways to create
subblocks. The new method coined Concatenated Pseudo Random Subblock
Partition Scheme (CPR SPS) divides the OFDM symbol into multiple disjoint sub
blocks and assigns signals randomly in each subblock. The partial sub band is then
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
73
duplicated and assigned to the rest of the subbands repetitively to form a complete sub
block. This procedure is depicted in Figure 4.6 for 3 subcarriers.
partial subband duplication
X
(0 )
n
X
(1)
n
X
(2)
N’1 N1 n
Figure 4.6: Generation of subblocks for PTS using Concatenated Pseudo Random Sub Block Partition
Scheme (CPR SPS).
C is the concatenation factor, i.e. the number of subblocks. C=1 is equivalent to
pseudo random PTS. Results indicate that CPR SPS can achieve similar results to the
optimum pseudo random block allocation method but with a decrease in complexity,
especially when efficient FFT structures such as the CooleyTukey algorithm are
used. Best results are obtained with low values of C (2 or 4) which results in a
marginal decrease in complexity compared to pseudo random PTS. For a significant
reduction in complexity C=16 has similar performance to the adjacent subblock
allocation method.
References [79, 80] also examined the PTS complexity reduction. The basic principle
of these papers is to produce a suboptimal iterative algorithm which uses a similar
structure to PTS but with reduced complexity. This method is coined Suboptimal
Exhaustive Search (SES) algorithm and exhibits good performance with a minimal
number of trials. One of the problems of PTS is that in order to work out which
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
74
transmit signal has the largest peak, every sample in each possible signal has to be
checked to work out its peak value. The process of finding large peaks contributes
significantly to the complexity of PTS. If the time domain signal is given as (4.9)
then the phase optimization values are restricted to [ ]
,
1
m k
p = ± . The SES algorithm
works as follows:
1/ Assume p
m,k
=1 for all k and compute the PAPR of the combined signal (ie. no
optimisation).
2/ Next, invert the first phase factor (p
m,k
=1) and recompute the resulting PAP for the
first value in each sub block, V. If the new PAP is lower than the previous value use
b
1
as part of the final phase sequence, otherwise use the original value for p
m,1
(ie. 1).
3/ The algorithm continues until all V points have been given this treatment.
Simulations for V=16 sub blocks, and N=256 subcarriers exhibited around a 1dB
degradation compared to standard PTS with adjacent subblock partitioning at
Pr(ζ>ζ
0
)=10
3
. However the optimisation process has been reduced to N
2
additions, a
considerable saving in terms of hardware operations. Note that in the simulations an
oversampling factor of 4 is used, which is sufficient to catch all peaks.
Various parameters were varied in the simulation such as a) the number of sub blocks,
V. b) number of allowed phases, p
m,k
for optimization and c) data constellation size.
It was concluded that the performance improves with an increase in V. However as V
is increased the performance improvement becomes less pronounced. Increasing the
number of phases to four [±1, ±j] provides a minor improvement for higher values of
V and up to a 1dB improvement at low values of V. The data constellation size has a
minimal effect on performance.
The SES algorithm has similar performance to traditional PTS with optimum phase
selection (within 1dB of optimum solution at Pr(ζ>ζ
0
)=10
4
) but with reduced
complexity. However complexity increases substantially with an increase in V.
Reference [81] again looked at developing optimum phase factors for PTS with
reduced complexity. The new method called Optimal Limited Search (OLS) also
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
75
addresses the oversampling issue by including an oversampling factor, L. The output
of the PTS algorithm can be written as (4.11)
1
2
11 21 1
12 22 2
1 2
...
...
... ... ... ... ...
...
M
j
M
j
M
j
LN LN MLN
A A A e
A A A e
S
A A A e
φ
φ
φ
=
(4.11)
where
( ) ( )
1
,.....,
T
LN
S S S φ φ =
(4.12)
contains the optimized signal samples. Figure 4.7 shows a block diagram of an OLS
PTS transmitter. OLS sorts same subcarrier positions over V IFFTs in order of
magnitude as depicted in (4.13)
1 2
.....
r i r i rMi
A A A > > > (4.13)
then chooses
rli
rl
rli
A
A
φ
π
−∠
=
−∠
1, 3,...
2, 4,...
l
l
=
=
(4.14)
The minimum amplitude sum for one IFFT point is then given by (4.15)
( )
1 2 3
...
i r i r i r i
S A A A φ = − + − (4.15)
The phase selection (4.14) nearly always yields the maximum amount of amplitude
cancellation for the i
th
signal sample. Next all LN solutions are calculated and the one
that minimises the maximum signal samples is chosen.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
76
Figure 4.7: Block diagram of OLSPTS transmitter.
The effect of quantization and the PSD is also examined for OLS. Quantization is
shown to have a negligible effect on performance and the out of band radiation is also
shown to be better than the OBPS case.
OLS out performs the methods of references [79, 80] for larger values of V (>8), also
the number of iterations is independent of V. Simulations (QPSK N=256) show OLS
to be slightly better than the OBPS of reference [74] for V>4 and much better for
V<4. The complexity of the 3 preceding methods is summarized in Table 4.1.
Reference [82] developed another scheme to find optimal phases based on orthogonal
vectors called Orthogonal Projectionbased PTS (OPPTS). The authors state that
OPPTS can reduce the PAPR within 0.40.8dB (using adjacent partitioning) of OBPS
using pseudo random partitioning and is suitable for a large number of subcarriers.
IFFT
IFFT
A
11
A
1LN
A
13
A
12
S/P
S/P
X
1
X
V
Phase
optimisation
A
v1
A
vLN
A
v3
A
v2
X
Data source
S
1
(Φ)
S
LN
(Φ)
S
3
(Φ)
S
2
(Φ)
Table 4.1: Number of complex multiplications and magnitude operations required.
L=4; N=256; W=2
V=2 V=16
OBPS (Huber et al) 33,554,432 5.52*10
70
SES (Cimini, et al) 2048 33,554,432
OLS (Tellambura et al) 1,048,576 1,048,576
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
77
The complexity of OPPTS is better than OBPS when a large number of subcarriers
and subblocks are used (N=128, W=4, V<=5), but improvement using OPPTS
reduces with an increase in V. The complexity of OPPTS is also slightly lower than
SES for high N and V. Note that no oversampling is used in OPPTS. A suboptimum
solution for OPPTS approach to find near optimum phases with negligible affect on
performance is also presented. The optimum phases are chosen from a set of 32
selected phases.
4.2.3 Selective Mapping
In reference [83] another multiple signal representation method is presented called
SeLected Mapping (SLM). The basic idea of SLM is to produce U alternative
transmit sequences seeded from the same data source and then to select the transmit
signal exhibiting the lowest PAPR. The idea stems from the fact that as the PAPR is
determined by the sequence of the transmit data vectors, X
m
, multiplying the data
vectors by some random phase will change the PAPR properties after the IFFT.
Mathematically, a set of U markedly different, pseudo random fixed vectors are
generated,
( ) ( )
0 1
, ,
u u u
D
P P P
−
=
… ,
( )
( )
( )
[ ) , 0, 2
u
k
u u
k k
P e
ϕ
ϕ π
+
= ∈ , 0 ,1 k N u U ≤ < ≤ < .
The data, X
m
, is multiplied subcarrier wise with each one of the U vectors,
( ) u
P ,
resulting in a set of U different possible transmit symbols,
( ) u
m
X as depicted in (4.16)
( ) ( )
, ,
.
u u
m k m k k
X X P = , 0 ,1 k N u U ≤ < ≤ < (4.16)
Next, all U possible transmit vectors are transferred to the to the time domain via the
IFFT,
( ) ( )
{ }
u u
m m
x IFFT X = , and the transmit symbol with the lowest PAPR,
m
xɶ , is
chosen for transmission. A SLM transmitter block structure is depicted in Figure 4.8.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
78
Figure 4.8: Block diagram of an SLM OFDM transmitter.
Simulations performed with N=128 and 4PSK mapping indicate that at Pr(ζ>ζ
0
)=10

4
, with U=4 a PAPR reduction of ~3dB is gained compared to the uncoded case.
Complexity reduction can also be achieved by restricting the random generated data
to
( )
{ } 1,
u
m
P j ∈ ± ± avoiding complex multiplications.
Figure 4.9 shows the CCDF of SLM with N=64, os=1, and U=1 to 32. At
Pr(ζ>ζ
0
)=10
4
the PAPR is reduced from 2 to 4.5dB for U=1 to U=32 respectively.
Figure 4.10 shows the CCDF of SLM with N=64 and U=1 (uncoded) 3, and 8. The
oversampling rate is set at 1, 2, and 4, and 8. Here it is seen that oversampling has a
minimal affect on the PAPR, only increasing it by ~0.5dB. An oversampling rate of 4
is sufficient. This is due to all the alternative transmit signals being uncorrelated as
shown in Figure 3.6. This is one of the main advantages of SLM, i.e. that
oversampling and filtering does not increase the PAPR dramatically as it does in PTS
(shown in Chapter 5). Again oversampling is seen to have a negligible affect on the
PAPR, the purpose of this plot is to show that oversampling still has a minimal affect
at higher values of U.
Side information
Bit source
Coding
Mapping
Serial to
parrallel
IFFT
Select
transmit
symbol
with
lowest
PAR
IFFT
IFFT
( ) U
p
( ) 2
p
m
X
( ) 1
m
X
( ) 2
m
X
( ) U
m
X
( ) 1
m
x
( ) 2
m
x
( ) U
m
x
m
xɶ
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
79
Side information is also an important issue in SLM as the receiver needs to be
informed which vector,
( ) u
P , was used. Log
2
(U) bits are required to send this
information increasing redundancy. As loss of this information (in a fading
channel) means the complete loss of the transmit symbol channel coding is required
to ensure correct recovery of the data at the receiver, increasing redundancy further.
In reference [74] SLM is compared to PTS for various combinations of V, W (PTS)
and U (SLM). In terms of redundancy (side information) SLM with U=1..5
outperforms PTS using pseudorandom subblock partitioning (W=4, V=1..5) at the
cost of greater system complexity. In terms of performance PTS has better PAPR
reduction if the number of IFFT’s is fixed as PTS can vary W with no additional
IFFT’s. 4 alternative signals produced with PTS require 3 IFFT’s while the same
performance with SLM can only be achieved with 4 IFFT’s.
2 3 4 5 6 7 8 9 10 11 12
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
U=1
U=2
U=3
U=4
U=8
U=16
U=32
Figure 4.9: Simulated CCDF for SLMOFDM for varying values of U. N=64, os=1.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
80
3 4 5 6 7 8 9 10 11 12
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
os=1
os=2
os=4
os=8
U=1
U=3
U=8
Figure 4.10: Simulated CCDF for SLMOFDM. U=1, 3, and 8. Oversampling rates are 1, 2, and 4,
and 8.
The issue of side information for SLM is further explored in [84] where the need for
explicitly sent side information is avoided. The new technique employs a
scrambling sequence and inserts U different ‘labels’,
( ) u
b , of length log
2
U as a prefix
to the data sequence as shown in Figure 4.11a. The data with prefix is then fed into a
scrambler polynomial as shown in Figure 4.12. The labels drive the scrambler into
one of U different states before scrambling the data itself. The scrambled data is then
processed as usual. The process is repeated with the other different U labels to
produce U alternative transmit sequences as in standard SLM.
Figure 4.11b is a variation where the linearity of the scrambler is exploited, here only
a single codeword
( ) 0
q
⌣
needs to be generated. U different subcarrier vectors can then
be generated by applying U different vector mappings to
( ) 0
q
⌣
. The advantage of this
structure is that u
th
calculated vector mapping can be calculated once from the label
( ) u
b and can be stored in memory. The receiver structure complexity is hardly
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
81
increased at all, only the label needs to be removed after descrambling. The
complexity of the new scheme is increased by around U compared to standard SLM.
(a)
(b)
Figure 4.11: SLM transmitter block diagram employing technique to avoid explicit transmission of side
information, a) serial form, b) parallel form.
Figure 4.12: Scrambler polynomial for new SLM technique.
Simulations were performed with N=256 carriers, 219 of which are active (to
decrease the design constraints of the transmit filter), 16 QAM mapping is used. The
IFFT was oversampled by a factor of 2 yielding more accurate results. The data was
then interpolated by a factor of 8 and a RRCF with a rolloff of 0.12 was used, results
were shown after the transmit filter. At Pr(ζ>ζ
0
)=10
5
the new SLM algorithm can
Label
inserter
Scrambler
Mapper
IFFT
Select
RF
up
conv.
S/P
P/S
( ) 0
b
( ) 1 U
b
−
q
( ) u
q
⌣ ( ) u
A
( ) u
a
( ) 0
a
( ) 1 U
a
−
aɶ
( ) u
b
Label
inserter
Scrambler
Mapper
(0)
IFFT
Select
RF
up
conv.
Mapper
(U
IFFT
S/P
S/P
P/S
P/S
0
( ) 0
q
⌣
( ) 0
A
( ) 1 U
A
−
( ) 1 U
a
−
( ) 0
a
aɶ
Delay Delay Delay Delay
( ) u
q
⌣
( ) u
q
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
82
reduce the PAPR by 1.8 and 2.5dB for U=4 and 8 respectively. Figure 4.13 shows the
receiver structure of the new SLM technique.
Figure 4.13: a) Receiver structure of the proposed SLM system, b) Descrambler polynomial at the
receiver.
The impact on the BER is also analyzed where perfect knowledge of the channel and
noise is assumed. The descrambler can multiply errors in receiver, but the effect is
negligible as the BER is degraded by 0.2dB. PSD plots showed that between 1 and
2dB can be saved in power backoff of the LPA with 4 bits redundancy per OFDM
symbol.
Reference [85] again looks at the issue of side information in SLM and proposes a
variation of reference [84] called Blind SLM (BSLM) which does not require the
labels to be sent with the data for descrambling in the receiver. The U sets pseudo
scrambling noise vectors are restricted to a known set at the receiver and all U sets are
sufficiently different. In order for the method to work
u
n
j
n
c e Q
φ
∈ , where Q is the
constellation mapping type.
At the receiver the decision metric is (4.17)
RF
down
conversi
on
FFT
Demapper
Descramble
r
Label
dumper
P/S
S/P
Y
y
ˆ
q
⌣
ˆ
q
ˆ q
a)
Delay Delay Delay Delay
ˆ
q
⌣
ˆ
q
b)
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
83
{ }
ˆ
0, 1, 1,
ˆ
1
2
ˆ ˆ ˆ ...
0
ˆ , 0,1,... 1
ˆ min
u
n
N
u
N
j
n n n
c c c C
n
P u U
D r e H c
φ
−
−
−
∈
=
∈ −
= −
∑
(4.17)
is performed for P
1
, P
2
,….P
U1
, the global minimumdistance solution provides the
best estimate for the transmitted data. A suboptimal metric is also presented to reduce
complexity in the receiver.
Simulations of the BER revealed that the new technique has almost the same
performance as SLM with perfect side information available with a infinite backoff in
the LPA. When a SL is employed with varying backoffs, BSLM slightly outperforms
the standard SLM approach. However there is a slight degradation when the new
algorithm is used in a fading channel.
4.3 Tone Reservation/Injection
Tone Reservation (TR) and Tone Injection (TI) were first introduced in reference [86]
and further detailed in reference [30]. These methods use an iterative algorithm to
provide increasingly better CCDF results.
4.3.1 Tone Reservation
In TR subcarriers, called Peak Reduction Tones (PRT’s), are set aside for PAPR
reduction as shown in the transceiver block diagram in Figure 4.14.
The signal plus Peak Reduction Tones (PRTs) are represented in (4.18)
( ) C X Q c x + = + (4.18)
where Q is the IFFT matrix, X is the transmit data before the IFFT, and C are the
PRT’s.
ℜ ∉
ℜ ∈
= +
C m
k
m
k m
k
m
k
k X
k C
C X
,
,
(4.19)
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
84
Figure 4.14: Block diagram of a Tone Reservation (TR) OFDM transceiver.
where
0 =
k
X , { }
L
i i k ,....,
1
∈ and
0 =
k
C , { }
L
i i k ,....,
1
∉
Coefficients for C
k
can be found with a reduced complexity iterative algorithm which
achieves a solution close to the optimum in a few steps to reduce the PAPR. This
method is distortionless as the data lies in disjoint frequency bins, i.e. 0 =
m
k
m
k
C X ,
which introduces redundancy. The new transmit signal has PAPR defined as (4.21)
{ }
[ ] N x
c x
c x PAR
/
2
2
ε
+
= + (4.21)
Vector c is computed so that the maximum peak value is reduced as (4.22)
∞
∞
+ = +
^ ^
^
min min C Q x c x
C
c
(4.22)
Solving
2
c x + is a convex problem, where the optimum solution lies at the bottom
of the parabola. When the terms are expanded out it can be solved as a Quadratically
Constrained Quadratic Program (QCQP), which can also be solved as a Linear
Program (LP). The complexity of the LP is O(NlogN).
FFT
X0
X2
C1
CN1
S/P
N
IFFT
X0
0
X2
0
P/S
N
IFFT
0
CN1
0
C1
P/S
N
x(t)
c(t)
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
85
L/N is the ratio of redundancy where L is the number of PRT’s and N is the total
number of data tones or subcarriers. To reduce the bit redundancy, PRTs can be
assigned to subcarriers which have a small number of bits assigned to them (in the
case of ADSL), i.e. QPSK as opposed to 16 QAM subchannels. The Data Rate Loss
(DRL) can be given as (4.23)
∑
∑
−
=
=
=
1
0
1
N
k
k
L
k
i
b
b
DRL
k
(4.23)
Simulation results indicated that for a L/N ratio of 5%, the clip probability is reduced
from 15dB to 9dB at Pr(ζ>ζ
0
)=10
5
. For a L/N ratio of 20% the clip probability is
reduced from 15dB to 5dB at Pr(ζ>ζ
0
)=10
5
. Basically an increase in the number of
PRT’s improves the PAPR reduction made. All simulations are based on N=512
subcarriers.
The performance of TR is influenced by the position of the peak reduction tones. As
with PTS it was found that random positions of PRC gave the best peak reduction
results of 6.2dB at Pr(ζ>ζ
0
)=10
5
probability. Adjacent PRTs on the other hand only
had a peak reduction of 3.4dB.
The number of iterations required to achieve various peak reduction values for
random PRT also has a pronounced affect on the performance of TR. For example
with L/N= 5%, 1 iteration gives a 2dB reduction at Pr(ζ>ζ
0
)=10
5
, 5 iterations
provides a 3.8dB reduction at Pr(ζ>ζ
0
)=10
5
, which is 2.1dB away from optimum
solution. Note that law of diminishing returns applies in that the more iterations
performed the less the PAPR reduction achieved, i.e. after 40 iterations the new peak
value is still 0.5dB away from optimum solution.
These results are for the optimum solution where the complexity is quite high. The
complexity can be reduced using general purpose iterative algorithms which find sub
optimal solutions such as the gradient algorithm.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
86
Gradient algorithm for computing c
Reference [86] uses gradient algorithms with reduced complexity, to reduce the
hardware requirements while still providing a good approximation to the optimum
solution.
By taking the gradient of the clipping noise Mean Square Error (MSE), simple
iterative algorithms were produced which achieved a solution close to the optimum in
a few steps. The gradient algorithm searches for the largest terms in x+c and subtracts
scaled replicas of them to cancel large peaks. The complexity of this algorithm is
O(N) per iteration compared to O(NlogN) in the optimum case.
The clipping noise of the gradient algorithm is defined as (4.24)
( ) ( ) ( )
∑
−
=
− = −
1
0
2 2
N
n
n A n A
x clip x x clip x (4.24)
Including the PRT’s in the transmitted time signal x+c, the Signal to Clipping noise
Ratio (SCR) is given by (4.25)
( )
2
2
c x clip c x
x
SCR
A
+ − +
= (4.25)
In order to maximize the SCR the denominator is minimized. Instead of solving for c
(c=QC) [30] solves for the gradient of c giving (4.26)
∑
> +
+
− =
A c x
n
row
k
n
k k
k
i i
q Q c c
) (
^
^
) ( ) ( ) 1 (
α µ (4.26)
where ( )( ) A c x c x sign
k
n n
k
n n
k
n
− + + =
) ( ) ( ) (
α . As is seen in (4.26), each iteration
requires an IFFT to be performed. Basically the algorithm searches for the largest
terms in
) (k
c x + and subtracts scaled, circularly shifted replicas of the vector
0
p to
cancel large peaks (
0 0
ˆ
ˆ
row
q Q p = ).
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
87
4.3.2 Tone Injection
Another method developed in reference [86] called TI maps data that cause large
peaks to new positions which will not produce peaks when the IFFT is performed.
The data can then be easily decoded correctly with a modulo D operation in the
receiver as shown in the block diagram in Figure 4.15.
Figure 4.15: Block diagram of a Tone Injection (TI) OFDM transceiver.
If
2
3
2
k k
k
d j d
A X + = = (4.27)
and A is changed so that
jqD pD A A + + =
ˆ
(4.28)
where p and q are any integer values and D is a positive real number known at the
receiver. This process effectively increases the size of the constellation, as well as the
average power of the transmit symbol. Reference [30] indicated that careful selection
ˆ
A can reduce the PAPR by over 5dB with only a 2% increases in the average power.
S/P
N
X0
IFFT
XN1
X2
P/S
N
IFFT
C0
CN1
C2
C1
P/S
N
x(t)
c(t)
Mod
D
X0
X2
X1
XN1
FFT
X0+C0
X1+C1
X2+C2
XN1+CN1
X1
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
88
The possible expansions of the selected tone for a 16QAM constellation are shown in
Figure 4.16. The new transmitted signal can be given as (4.29)
Figure 4.16: Example of possible expansions of constellation in TI for 16 QAM.
( )
∑
−
=
+ + =
1
0
/ 2
1
ˆ
N
k
N kn j
k k k k k n
e D jq D p X
N
x
π
(4.29)
The maximum peak reduction per tone shift is (4.30)
N M
M
k
k ρ
δ
2
1
6
2
2
0
0
−
= (4.30)
×
® ®
¸
×
Z
⊗
©
¶ ¶
×
® ®
¸
×
Z
⊗
©
¶ ¶
×
® ®
¸
×
Z
⊗
©
¶ ¶
×
® ®
¸
×
Z
⊗
©
¶ ¶
×
® ®
¸
×
Z
⊗
©
¶ ¶
×
® ®
¸
×
Z
⊗
©
¶ ¶
×
® ®
¸
×
Z
⊗
©
¶ ¶
×
® ®
¸
×
Z
⊗
©
¶ ¶
×
® ®
¸
×
Z
⊗
©
¶ ¶
I
Q
I
I
Q Q
D
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
89
where
k
M is the number of levels per dimension. The maximum peak reduction
factor (assuming normalized data) is ∂ K , where K is the number of real/imaginary
dimensions. The peak reduction factor δ decreases as N increases, so to keep the peak
reduction factor constant K must be increased.
There is no redundancy in TI as the receiver uses a modulo D operation to decode
information, thereby separating the original constellation point from the expanded
one. The complexity in TI comes from finding values of p
k
and q
k
which produce low
PAPR which in turn require (as in TR) the solving of an integer programming
problem, which has exponential complexity.
Assuming L duplicate points per constellation, if K dimensions are to be modified, the
algorithm must search over ( )
K
NL combinations for vectors p
k
and q
k
, where L is the
oversampling factor. Low complexity iterative algorithms are used to find
approximations close to the optimum solution.
The algorithm starts with the original multicarrier symbol (pk=0, qk=0). After the
maximum sample, n0, is located, the tone, k0, is found (4.29) is updated. If more
than 1 value of ˆ
n
x is large, a new value for k0 that reduces as many peaks of possible
is found. This procedure can be repeated several times until the desired PAPR is met
or the maximum of iterations is reached. Each iteration decreases the PAPR by
around 1 dB up to a max of 6dB of reduction.
4.4 Conclusion
This chapter introduced distortionless techniques for PAPR reduction in OFDM.
Distortionless techniques have the advantage of not corrupting the data thereby
maintaining a low BER but come at the cost of increased complexity and bandwidth.
Coding techniques for PAPR reduction were first discussed in Section 4.1 where
redundant bits are added to the bit stream before the IFFT. Properly chosen, these
codewords ensure that the PAPR after the IFFT is kept low. These codes can be
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
90
combined with existing COFDM to reduce the redundancy and complexity inherent in
coding. A disadvantage of coding is that the complexity becomes prohibitively high
with an increase in the number of subcarriers (>32). Various codewords were
presented such as cyclic codes, ShapiroRudin Sequences, Golay Complementary
codes, and ReedMuller codes. Golay codes and their subset, second order Reed
Muller codes were found to have excellent PAPR properties restricting the PAPR to
3dB. This reduction could be traded off with reductions in complexity and the code
length. Still complexity remains a restrictive issue in coding.
Section 5.2 introduced MSR methods, where a set of alternative transmit signals are
seeded from the same data source. Various techniques are used to encode the
alternative sets of transmit signals, which are constructed in such a way so that they
will have different PAPR properties. The transmit signal with the lowest PAPR is
chosen for transmission.
The effect of MSR OFDM is to shift the CCDF curve from the right to left, i.e. reduce
the amplitude peaks in a uniform manner to lower PAPR’s. Both PTS and SLM
require side information to be sent with the chosen transmit signal. This information
must be protected as loss of the side information means loss of the whole transmit
symbol. The redundancy required for PTS is (V1)log
2
W while SLM requires log
2
U
bits per OFDM symbol. Variations of SLM and PTS have also been presented which
alleviate the need for explicit side information. Increasing the number of alternative
transmit symbols reduces the PAPR but also increases the complexity, reductions in
the PAPR also become less pronounced with larger numbers of alternative transmit
symbols. The number of alternative transmit symbols that can be produced for 1
OFDM symbol is W
V1
for PTS and U for SLM.
An advantage of MSR over coding techniques is that the reductions are independent
of the constellation type, and only marginally affected by the number of subcarriers.
However unlike coding no specific level of PAPR can be guaranteed with MSR
methods, only a shift to lower PAPR as shown in the CCDF curves.
The effect of oversampling PTS is examined in the next chapter where filtering is
shown to degrade the gains made with PTS by ‘regrowing’ peaks. Oversampling at
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques
91
the IFFT is required to maintain the reductions gained. SLM is much more robust to
filtering due to the greater independence of the alternative transmit symbols.
Generally SLM outperforms PTS in terms of reduction versus redundancy, but PTS is
much better in terms of PAPR reduction versus additional system complexity.
Section 4.3 described TR and TI for PAPR reduction, both methods prevent distortion
by reducing the PAPR before the HPA. In TR the additional complexity is only in the
transmitter. Suboptimal solutions to the PAPR minimization problem were presented
which successfully reduced the complexity without sacrificing the amount of PAPR
reduction to a great degree.
In TI a similar method to TR is presented where most of the complexity is in the
transmitter with some additional complexity in the receiver which is a simple modulo
operation of the demodulated complex vectors. Again suboptimal solutions were
proposed with similar performance to the optimal case.
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
92
Chapter 5
Cyclic Shifted Sequences and Time
Inversion of PTS
This chapter introduces new techniques to produce alternative transmit signals in PTS.
Section 5.1 reviews standard PTS subblock creation methods. Two new techniques
for subblock creation in PTS are described in section 5.2. Section 5.2.1 describes
Cyclic Shifted Sequences (CSS) where shifts in the data are used in place of phase
rotations. Section 5.2.2 looks at a combination of traditional PTS using phase
rotations and CSS to produce alternative transmit signals. Section 5.2.3 describes
Time Inversion (TI) where the data sequence is reversed to provide an alternative
transmit signal. Section 5.3 analyses the effect of filtering on traditional phase
rotated PTS as well as CSS and TI. Section 5.4 provides results for the
aforementioned techniques when oversampling of the IFFT is done before filtering.
Section 5.5 compares the complexity of standard PTS and the new techniques and
Section 5.6 closes the chapter with a conclusion.
5.1 PTS subblock creation
As was shown in section 4.2.1 PTS can provide promising reductions in the PAPR.
However issues such as complexity and filtering after PTS limit the techniques
usefulness. In traditional PTS alternative subblocks are created by increasing V, the
number of IFFT’s, and W, the number of allowable phase rotations. Increasing the
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
93
number of phase rotations provides increasingly less PAPR reduction for a set V. In
addition the hardware operations are non trivial for W>4. This limitation in PTS is
the impetus for finding alternatives to phase rotations in order to create unique
transmit sequences. Note that all simulations use adjacent block partitioning.
5.2 New techniques for PTS subblock creation
Of all the MSR techniques PTS suffers the most after filtering [39] as the alternative
sequences are not necessarily independent of each other as in the case of SLM. The
following sections propose new techniques to create alternative sequences for PTS
and shows through simulation that the peak regrowth of the proposed techniques after
filtering is not as severe as in traditional PTS. This advantage is combined with a
reduction in complexity of the new algorithms.
Figure 5.1 shows a block diagram of the simulation model, note that the CCDF is
measured after the nyquist sampled IFFT.
Figure 5.1: Block diagram showing the simulation model of Section 5.2.
5.2.1 Cyclic Shifted Sequences
As detailed in Section 4.2.1 PTS produces alternative signals by breaking up the
transmit bit stream and phase rotating whole parts before performing the IFFT. These
phase rotations can be kept trivial if W is restricted to 4 or less. In other words for
phase rotations of 90, 180, and 270 degrees only the sign of the I or Q value needs to
be changed, making them trivial hardware operations. If the number of phase
Data
source
Mary
Mapping
PTS
/CSS
/TI
Measure
CCDF
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
94
rotations, W, is greater than four, then the rotation becomes non trivial as the I and Q
values have to be manipulated requiring a more complex circuit. CSS does not suffer
from this limitation, the number of positions that can be shifted to provide alternative
transmit signals can be up to N/2 (before identical signals are produced) where N is
the number of subcarriers.
Figure 5.2 shows a block diagram of a PTS system depicting PTS, CSS, and TI which
is introduced in the next section. The system represents a PTS/CSS/TI system with
V=2 subblocks and W/S=4 rotations/shifts.
Figure 5.3 compares the CCDF of CSS to PTS where the phase shifts are replaced
with cyclic shifts, they are equivalent in all other aspects. Unless otherwise stated all
simulations use N=64 subcarriers with QPSK mapping. No oversampling is
performed. Here we see that CSS outperforms PTS in all cases, also the complexity
of CSS when W>4 is reduced. CSS: V=2, S=8 performs 0.4dB better at Pr(ζ>ζ
0
)=10
4
than PTS: V=2, W=8. CSS: V=4, S=4 has a minor improvement of 0.2dB over PTS:
V=4, W=4 at Pr(ζ>ζ
0
)=10
4
. CSS: V=4, S=8 has 0.6dB better performance than PTS:
V=4, S=8 at Pr(ζ>ζ
0
)=10
4
. CSS has a PAPR reduction of 3dB, 4.5dB, and 5.5dB
compared to the uncoded conventional case for CSS: V=2, S=8; V=4, S=4; and V=4,
S=8 respectively at Pr(ζ>ζ
0
)=10
4
.
Mathematically CSS can be expressed as (5.1):
( )
( )
~ ~
1
v
v
V
n n
v
a a δ +
=
=
∑
(5.1)
Where δ
(v)
is a cyclic shift in the time domain.
It seems that with CSS not only is complexity reduced but the independence of the
generated alternative signals is greater thereby giving better PAPR reduction. The
performance of CSS at lower probability regions should be increasingly better than
PTS due to the steeper slope of the curve, especially for V=4, S=8.
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
95
Figure 5.2: Block diagram of PTS, CSS and TI transceiver.
For a CSS structured OFDM transmitter with V=2 and 64 subcarriers, 32 alternative
transmit signals can be created. In a PTS transmitter, with V=3 and W=4 only 16
alternative transmit signals with trivial phase rotations can be produced. Increasing
the number of subcarriers in CSS allows for more alternative transmit signals to be
produced with trivial operations. The effect of double shifts has also been analysed
FFT
A’,A’,A’,A’
B, B, B, B
Correction
A, A, A, A, B, B, B, B
a) Transmitter
b) Receiver
'
n
n
a a
−
=
0,0,0,0,B,B,B,B
A,A,A,A,0,0,0,0
A,A,A,A,B,B,B,B
Store
Store
Modify
Select
IFFT
IFFT
Correction
Store
v=0
v=1
v=2
v=3
'
v
j
a ae
θ
=
'
n
n m
a a
+
=
PTS
CSS
TI
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
96
where it was found that large shifts produce better results. This comes at the cost of
reducing the possible number of transmit signals that can be constructed.
3 4 5 6 7 8 9 10 11 12
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
Uncoded
CSS
PTS
V=4
S/W=8
V=4
S/W=4
V=2
S/W=8
Figure 5.3: Simulated CCDF for PTS (V=2, W=8; V=4, W=4; V=4, W=8), CSS (V=2, S=8; V=4, S=4;
V=4, S=8) and uncoded OFDM. N=64, adjacent partitioning, with no oversampling in IFFT, (os=1).
5.2.2 PTS with CSS
Another variation of CSS combining the trivial phase rotations (W=4) in standard
PTS with CSS shifts to create more alternative transmit sequences was first presented
in [1].
The PTS signal can be expressed as (5.2)
∑
=
=
V
1 v
j
) v (
n
~
n
~
) v (
e . a a
φ
(5.2)
Adding cyclic shifts to PTS gives (5.3)
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
97
) v (
) v (
j
V
1 v
) v (
n
~
n
~
e . a a
φ
δ ∑
=
+
= (5.3)
Where δ
(v)
is a cyclic shift in the time domain. The number of alternative transmit
signals is now WS, where S is the number of cyclic shifting options. Cyclic shifting
generates a linear phase shift at the output of the FFT in the receiver. This linear
phase component must be removed prior to demodulation and requires one complex
multiplication for each data symbol. This process can be made trivial if the linear
phase shift is constrained to
2
gπ
radians per frequency bin (g=0,1,2,or 3). This
corresponds to a signal shift of
4
g . N
) v (
= δ samples (g=0,1,2, or 3). Therefore for
each of the W different phase rotations, S=4 additional signals can be produced by
cyclic shifting with trivial operations in the transmitter and receiver.
Figure 5.4 shows CCDF plots comparing standard PTS to PTS/CSS. There is a
negligible performance difference in the two methods. The advantage here is that for
8 alternative transmit sequences (PTS: V=2, W=8; PTS/CSS: V=2, W=4, S=2) 4
complex multiplications per sample are avoided in the transmitter. For 16 alternative
transmit sequences (PTS: V=3, W=4; PTS/CSS: V=2, W=4, S=4) a whole IFFT can
be avoided, reducing complexity significantly. The reduction in PAPR from the
uncoded case is ~2.5dB and ~3.5dB for (PTS: V=2, W=8; PTS/CSS: V=2, W=4,
S=2) and (PTS: V=3, W=4; PTS/CSS: V=2, W=4, S=4) respectively at Pr(ζ>ζ
0
)=10
4
.
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
98
3 4 5 6 7 8 9 10 11 12
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
Uncoded
PTS/CSS
PTS
V=3 W=4
V=2 W=4, S=4
V=2 W=8
V=2 W=4, S=2
Figure 5.4: Simulated CCDF for PTS (V=2, W=8; V=3, W=4), PTS/CSS (V=2, W=4, S=2; V=2, W=4,
S=4), and uncoded OFDM. N=64, adjacent partitioning, with no oversampling in IFFT, (os=1).
5.2.3 Time Inversion
Another new proposed PAPR technique based on a variation of PTS first presented in
[2] is Time Inversion (TI), a block diagram description of which is shown in Figure
5.2. TI involves reversing the output sequence of the IFFT sub block before addition
and transmission of the sub blocks. Mathematically TI can be expressed as (5.4)
) . (
) (
) (
) (
) (
1
) (
) ( ) 1 (
~
~
v
v
v
i
v
j
V
v
v
n
i
n e a conj a
φ
δ ∑
=
+ −
= (5.4)
where i takes the value 0 or 1 for a time inversion of the V
th
sub block in the time
domain. TI of a subblock must be accompanied by conjugation of its elements to
stop the information from jumping into image subchannels after the demodulation
process. The effect that this has on the demodulated subchannels can be reversed by
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
99
taking the conjugate of the output after the FFT process in the receiver circuit. No
extra multiply operations are required in the transmitter or receiver.
Simulation results for PTS with TI are shown in Figure 5.5 where it is seen that TI has
slightly worse performance than PTS with an equivalent number of alternative
transmit sequences. TI has ~0.5dB degradation compared to PTS at Pr(ζ>ζ
0
)=10
4
. TI
reduces the PAPR by ~2.2dB over the uncoded case while PTS has a reduction of
~2.7dB at Pr(ζ>ζ
0
)=10
4
.
3 4 5 6 7 8 9 10 11 12
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
Uncoded
PTS/TI: V=2, W=4, S=2
PTS: V=2, W=8
Figure 5.5: Simulated CCDF for PTS (V=2, W=8), PTS/TI (V=2, W=4, S=2) and uncoded OFDM.
N=64, adjacent partitioning, with no oversampling in the IFFT, (os=1).
5.3 Filtering new techniques
As noted in [34, 39] the reductions made with PTS are drastically reduced when
passed through a pulse shaping filter due to peak regrowth. Furthermore as shown in
Figure 3.10 peak regrowth after some form of PAPR reduction is more severe than
when no PAPR reduction is performed. Certain PAPR reduction schemes such as
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
100
PTS and clipping display greater peak regrowth after interpolation and filtering than
coding and SLM (Figure 4.11). This section analyses the effect of interpolation and
pulse shaping filtering on CSS and TI, comparing them to standard PTS through
simulation.
Simulations performed in this section use N=64 subcarriers with no oversampling in
the IFFT (i.e. os=1) and adjacent partitioning. As shown in reference [39] the
partitioning type has a minimal effect on the peak regrowth after filtering. A RCF is
used with a rolloff factor of 0.15 and 128 filter taps with a normalized sampling rate.
The data is interpolated by a factor of 8 before filtering. The simulation model is
represented in Figure 5.6.
Figure 5.6: Block diagram showing the simulation model of Section 5.2.
In Figure 5.7 the discrete CCDF after the IFFT is plotted with the CCDF after filtering
for PTS and CSS. Both PTS and CSS have 16 alternative transmit symbols. CSS has
a peak regrowth of ~3.5dB from the discrete level and PTS has a peak regrowth of
~4dB at Pr(ζ>ζ
0
)=10
4
. This is contrasted with only 1dB of peak regrowth in the
uncoded case at Pr(ζ>ζ
0
)=10
4
. So the PAPR reduction after filtering is only ~1.5dB
and ~2dB for PTS and CSS respectively at Pr(ζ>ζ
0
)=10
4
. It is also worth noting that
CSS has a steeper rolloff than PTS leading to the increasingly better performance than
PTS at low probability levels.
RCF
Measure
CCDF
Measure
CCDF
Data
source
Mary
mapping
PTS
/CSS
/TI
(Figure 5.2)
8
IFFT size = N×os
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
101
3 4 5 6 7 8 9 10 11 12 13
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
discrete
filtered
CSS
PTS
Uncoded
Figure 5.7: Simulated CCDF of discrete and filtered PTS (V=4, W=4), CSS (V=4, S=4), and uncoded
OFDM. N=64, adjacent partitioning, with no oversampling in IFFT (os=1), interpolated by 8, filtered
with RCF (α=0.15).
Figure 5.8 maintains the same number of IFFT’s (V=4) as Figure 5.7 while increasing
the number of phase rotations/shifts from 4 to 8. The peak regrowth of CSS above the
discrete level after filtering is 4.3dB while PTS peak regrowth is 4.1dB at
Pr(ζ>ζ
0
)=10
4
. The peak power reduction after filtering is only ~1.7dB and ~2.2dB
for PTS and CSS. Comparing these results with Figure 5.7 it is seen that increasing
the number of phase rotations/shifts to 8 provides only a further 0.2dB PAPR
reduction for both PTS and CSS after filtering.
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
102
3 4 5 6 7 8 9 10 11 12 13
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
discrete
filtered
Uncoded
PTS
CSS
Figure 5.8: Simulated CCDF of discrete and filtered PTS (V=4, W=8), CSS (V=4, S=8), and uncoded
OFDM. N=64, adjacent partitioning, with no oversampling in IFFT (os=1), interpolated by 8, filtered
with RCF (α=0.15).
Figure 5.9 compares combined PTS/CSS and PTS/TI with standard PTS, with 8
alternative transmit signals. Both PTS/CSS and PTS/TI have very similar
performance after filtering with a peak regrowth of ~2.5dB and 2.1dB respectively
from the discrete level, while standard PTS is ~0.4dB worse with a peak regrowth of
~3dB at Pr(ζ>ζ
0
)=10
4
. Note that again the slope at Pr(ζ>ζ
0
)=10
4
is slightly steeper
for the new techniques after filtering implying further PAPR improvement over PTS
at lower probability levels. The PAPR reduction after filtering is very poor in all
cases, only ~1dB for the new techniques and 0.6dB for standard PTS at Pr(ζ>ζ
0
)=10
4
.
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
103
3 4 5 6 7 8 9 10 11 12 13
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
discrete
filtered
PTS/TI
PTS/CSS
PTS
Uncoded
Figure 5.9: Simulated CCDF of discrete and filtered PTS (V=2, W=8), PTS/CSS (V=2, W=4, S=2),
and PTS/TI (V=2, W=4, S=2), and uncoded OFDM. N=64, adjacent partitioning, with no
oversampling in IFFT (os=1), interpolated by 8, filtered with RCF (α=0.15).
In Figure 5.10 the CCDF of PTS and PTS/CSS is compared, where 16 alternative
transmit sequences produced. The peak regrowth is ~2.8dB and ~3.1dB above the
discrete level for PTS and PTS/CSS respectively at Pr(ζ>ζ
0
)=10
4
. Again the PAPR
reduction after filtering is heavily penalized with only ~1.8dB and ~1.5dB gained in
PAPR reduction at Pr(ζ>ζ
0
)=10
4
for PTS and PTS/CSS respectively. As stated earlier
the advantage of PTS/CSS in this case is the removal of 1 IFFT operation per transmit
symbol.
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
104
3 4 5 6 7 8 9 10 11 12 13
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
discrete
filtered
PTS/CSS
Uncoded
PTS
Figure 5.10: Simulated CCDF of discrete and filtered PTS (V=3, W=4), PTS/CSS (V=2, W=4, S=4),
and uncoded OFDM. N=64, adjacent partitioning, with no oversampling in the IFFT (os=1),
interpolated by 8, filtered with RCF (α=0.15).
5.4 Oversampling new techniques
Oversampling [39] is required to improve the performance of PTS techniques.
Oversampling the IFFT increases the convergence between the discrete PAPR and the
filtered PAPR. Oversampling increases the PAPR in the discrete domain, while
reducing the PAPR after filtering. The simulation model here is the same as Figure
5.6 except that oversampling is performed at the IFFT before filtering, i.e. os=1, 2, 4,
or 8. Again N=64 subcarriers, and adjacent partitioning is used. The data is
interpolated by 8 before filtering with a RCF with 128 taps and a rolloff factor of
0.15.
Figure 5.11 and 5.12 shows PTS/CSS (V=2, W=4, S=4) and PTS (V=3, W=4) under
oversampled and filtered conditions respectively, with 16 alternative transmit signals.
It is seen that oversampling by 2 brings the discrete and filtered CCDF curves to
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
105
within 1dB of each other at Pr(ζ>ζ
0
)=10
4
. Furthermore oversampling by a factor of 8
brings them almost to convergence with PTS/CSS slightly closer (~0.1dB) than PTS
(~0.3dB) at Pr(ζ>ζ
0
)=10
4
. The PAPR reduction for PTS/CSS and PTS after filtering
is 3.3dB and 3dB respectively at Pr(ζ>ζ
0
)=10
4
. As oversampling the IFFT by a factor
of 8 provides minimal improvement over 4, all further simulations will use
oversampling rates of os=1, 2, and 4.
3 4 5 6 7 8 9 10 11 12 13
10
4
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
discrete
filtered
Uncoded
os=1
os=2
os=4
os=8
os=1
os=2
os=4
os=8
Figure 5.11: Simulated CCDF for PTS/CSS (V=2, W=4, S=4), and Uncoded OFDM. Discrete
oversampled curves (dashed) move from left to right and oversampled (solid) filtered curves move
from right to left. N=64, adjacent partitioning, with oversampling rates in the IFFT of 1, 2, 4, and 8,
interpolated by 8, and filtered with RCF (α=0.15).
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
106
3 4 5 6 7 8 9 10 11 12 13
10
4
10
3
10
2
10
1
10
0
discrete
filtered
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
Uncoded
os=1
os=2
os=4
os=8 os=8
os=4
os=2
os=1
Figure 5.12: Simulated CCDF for PTS (V=3, W=4), and Uncoded OFDM. Discrete oversampled
(dashed) curves move from left to right and oversampled (solid) filtered curves move from right to left.
N=64, adjacent partitioning, with oversampling rates in the IFFT of 1, 2, 4, and 8, interpolated by 8,
and filtered with RCF (α=0.15).
Figures 5.13, 5.14, and 5.15 compare PTS (V=2, W=8), PTS/CSS (V=2, W=4, S=2),
and PTS/TI (V=2, W=4, S=2) under oversampled and filtered conditions respectively,
with 8 alternative transmit signals. Oversampling by a factor of 2 brings the discrete
and filtered CCDF curves to within 0.8dB of each other at Pr(ζ>ζ
0
)=10
4
.
Oversampling by a factor of 4 brings them almost to convergence (0.2dB). PTS/CSS
and PTS/TI have slightly better performance at any given oversampling rate. For
example, with an oversampling factor of 2 PTS has a PAPR of 10.2dB at Pr(ζ>ζ
0
)=10

4
, while PTS/CSS and PTS/TI has a PAPR of 9.9dB and 10dB respectively at
Pr(ζ>ζ
0
)=10
4
. The PAPR reduction for PTS, PTS/CSS, and PTS/TI after filtering
(os=4) is 2.2dB, 2.5dB, and 2.3dB respectively at Pr(ζ>ζ
0
)=10
4
.
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
107
3 4 5 6 7 8 9 10 11 12 13
10
4
10
3
10
2
10
1
10
0
discrete
filtered
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
Uncoded
os=1
os=2
os=4 os=4
os=2
os=1
3 4 5 6 7 8 9 10 11 12 13
10
4
10
3
10
2
10
1
10
0
discrete
filtered
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
Uncoded
os=1
os=2
os=4
os=1
os=2
os=4
Figure 5.13: Simulated CCDF for PTS (V=2, W=8), and Uncoded OFDM. Discrete oversampled curves
(dashed) move from left to right and filtered (solid) oversampled curves move from right to left. N=64,
adjacent partitioning, with oversampling rates in the IFFT of 1, 2, and 4, interpolated by 8, and filtered
with RCF (α=0.15).
Figure 5.14: Simulated CCDF for PTS/CSS (V=2, W=4, S=2), and Uncoded OFDM. Discrete
oversampled curves (dashed) move from left to right and filtered (solid) oversampled curves move from
right to left. N=64, adjacent partitioning, with oversampling rates in the IFFT of 1, 2, and 4, interpolated
by 8, and filtered with RCF (α=0.15).
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
108
3 4 5 6 7 8 9 10 11 12 13
10
4
10
3
10
2
10
1
10
0
discrete
filtered
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
Uncoded
os=1
os=2
os=4
os=4
os=2
os=1
Figure 5.15: Simulated CCDF for PTS/TI (V=2, W=4, S=2), and Uncoded OFDM. Discrete
oversampled curves (dashed) move from left to right and filtered (solid) oversampled curves move
from right to left. N=64, adjacent partitioning, with oversampling rates in the IFFT of 1, 2, and 4,
interpolated by 8, and filtered with RCF (α=0.15).
5.5 Complexity evaluation
The performance in terms of the CCDF and PSD of CSS, PTS/CSS, and PTS/TI has
been established. Table 5.1 compares them in terms of hardware operations where it
is seen that multiplications are avoided when appropriate values of V, W, and S are
chosen.
For example with a PTS system (N=64, V=3, W=4) 85 non complex multiplications,
170 squaring operations, and 178 compares are required. The equivalent PTS/CSS
system (N=64, V=2, W=4, S=2) requires no multiplications in exchange for an
increase in 512 squaring operations, and 528 comparisons as well as the removal of a
whole IFFT operation. A PTS system (N=64, V=2, W=8) requires 256
multiplications, half of which
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
109
Table 5.1: Complexity comparison of various techniques (adjacent partitioning)
Type Comments Operations
PTS
 V blocks
 W rotations
 (N/V)W multiplications (trivial if W<4)
 (N/V)W(V1) squaring
 (N/V+1)W(V1) comparisons
 V*(N/V) IFFT’S
PTS/CSS
 V blocks
 W rotations
 S shifts
 (N/V)WS(V1) squaring
 (N/V+1)WS(V1) comparisons
 V*(N/V) IFFT’S
PTS/TI
 V blocks
 W rotations
 T(=2)
inversion
 (N/V)TW(V1) squaring
 (N/V+1)TW(V1) comparisons
 V*(N/V) IFFT’S
are complex, 256 squaring operations, and 264 comparisons. The equivalent
PTS/CSS system (V=2, W=4, S=2) requires no multiplications, and has the same
number of squaring and comparison operations as the PTS system.
5.6 Conclusion
This chapter introduced new techniques for the creation of PTS OFDM signals. In
section 5.2 it was shown through simulation of the CCDF that CSS has better
performance (up to 1dB) than PTS for the same number of alternative signals over
various combinations of V and W. Combining PTS with CSS and TI was shown to
provide similar performance to standard PTS but with the advantage of a reduction in
complexity. PTS (V=3, W=4) was shown to have the same performance as PTS/CSS
(V=2, W=4, S=4) but with the added advantage of the removal of one whole IFFT
operation.
The performance after filtering was examined in Section 5.3 where it was shown that
the gains made with PTS techniques are dramatically affected when passed through a
pulse shaping filter. In the simulations a RCF with 128 taps and a roll off factor of
0.15 was used. The resultant reduction in PAPR after filtering was between 1 and
2dB for various combinations of V, W, and S with the new techniques having up to
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
110
0.7dB better PAPR for all combinations. The only exception is PTS/CSS (V=2, W=4,
S=4) where the performance was slightly worse than PTS (V=3, W=4) for the same
number of alternative signals. This small degradation is a trade off for removing 1 of
the IFFT’s.
The effect of oversampling on PTS and the new techniques was also examined, where
it was seen that an oversampling factor of 2 was sufficient to bring the discrete PAPR
to within 1dB of the filtered PAPR. The effect of oversampling was to increase the
PAPR of the discrete CCDF while reducing the CCDF of the filtered CCDF.
Oversampling by a factor of 8 (implying IFFT sizes of N×8)was necessary to bring
both curves almost to convergence. However in a practical system an oversampling
factor of 2 was shown to be sufficient bringing the discrete and filtered CCDF curves
to within 1dB of each other.
The avoidance of multiplications and in some cases IFFT operations, as well as a
modest performance gain combine to make the new techniques viable alternatives to
standard PTS OFDM. Two publications resulted from this chapter.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
111
Chapter 6
Peak to Average Power Solutions 
Distorted Techniques
This chapter details another approach to PAPR reduction, distortion introducing
techniques. These methods do not attempt to create a transmit signal with a low crest
factor, instead they take the output of the IFFT and then limit the amplitude of large
samples which invariably causes distortion. These methods include approaches such
as pulse shaping (or windowing), and clipping at every stage from the output of the
IFFT to limited backoffs in the amplifier.
The advantages of clipping are a reduction in complexity and ease of implementation,
the disadvantages are the inband distortion (increasing BER) and spectral splatter,
affecting adjacent channels by increasing out of band distortion.
Section 6.1 describes the most common form of clipping in the baseband and reviews
papers which seek to quantify the effect of clipping on the BER and PSD under
different OFDM system conditions. Section 6.1.1 looks at the relation between
clipping and quantization. Section 6.2 looks at amplifier limiting where the
unconstrained OFDM signal is allowed to saturate the amplifier introducing spectral
regrowth. Section 6.3 analyses windowing and pulse shaping techniques.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
112
6.1 Clipping in the Baseband
Reference [46] is an early paper looking at the effect of clipping and filtering on
OFDM. The authors note that clipping causes both inband distortion, deteriorating
the BER and out of band noise which reduces the spectral efficiency. Filtering after
clipping was noted to reduce the spectral splatter but at the cost of peak regrowth.
The simulation model used N=128 subcarriers with a guard interval, T
g
, of 32
samples. QPSK modulation was applied to the data samples before modulation with
an IFFT. As direct clipping of the samples will cause all the noise to fall in band the
OFDM symbol is oversampled by a factor of 8 before clipping. The complex
baseband samples were modulated up to a carrier frequency 1/4 of the sampling
frequency in order to reduce the complexity of the simulation model. The real valued
bandpass samples, x, are then clipped at amplitude, A, according to (6.1)
,
,
A
y x
A
−
=
if
if
if
x A
A x A
x A
< −
− ≤ ≤
>
(6.1)
where the clipping level A is determined as
A
CL
σ
= , where σ is the rms level of
the OFDM signal. Filtering is performed with an equiripple bandpass Finite Impulse
Response (FIR) filter with 103 taps, a stopband of 40dB, and 1dB ripple in the
passband. PSD results for clipping (CL=0.8 to 1.6) with no filtering displayed both in
band distortion and spectral splatter. For a CL=1.4 the out of band noise was only
16dB lower than the signal power. This demonstrated the need for filtering to
suppress the sidelobes. When the filter was applied the sidelobes were suppressed to
~50dB below the signal power for the same CL.
Results also indicated that the peak regrowth after filtering was significant. For
example, for a CL=1.4 (3dB) the clipped and filtered signal had a CF of ~9dB at the
99.999 percentile, i.e. 6dB of peak regrowth. The unclipped signal at the same
percentile had a CF of 13dB. Finally the BER was examined after clipping and
filtering in an AWGN channel. For reasonable clipping levels (CL>1.4) it was shown
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
113
that less than 1dB of degradation at the 10
2
BER level is encountered. However it
should be noted as shown in Figure 3.17 that QPSK, unlike higher modulation types is
inherently impervious to clipping.
Reference [87] presented a theoretical analysis of both cartesian and envelope
clipping for various oversampling rates. Cartesian clipping is where the I and Q
values of the complex sample are clipped independently. As the magnitude is not
required this method is much less complex to implement than envelope clipping,
however it introduces more distortion than envelope clipping at a set clipping level.
Considering z as the I or Q component of the complex OFDM signal the output of the
Cartesian clipper is expressed as (6.2)
( ) ( )
0
0 0
0
y
y g z y z z
y
−
= =
+
0
0 0
0
,
,
,
z z
z z z
z z
<
− < < +
>
(6.2)
Under the assumption of a Gaussian like nature of the samples (i.e. high number of
subcarriers) the impact of non linear distortion was analytically derived , and included
the effects of AWGN as well as clipping noise.
The analytical results were compared to a simulated system with N=128 subcarriers
and QPSK modulation. Square, or Cartesian clipping was shown to have worse
performance by around 1dB than envelope clipping (as expected). An oversampling
factor of 2 was shown to be enough to have the same performance as an infinitely
oversampled signal. A moderate deterioration was seen when the oversampling factor
was set at 1 for both square and envelope clipping. Optimum clipping levels were
shown to be 2.5
clip
A
σ
= for square clippers and 2.0
clip
A
σ
= for magnitude clippers.
Another reference [88] to analyze the performance of Cartesian clipping developed
analytical expressions for the PSD and SER for any Mary constellation with various
clipping backoffs. Clipping is performed according to the rule set in (6.2)
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
114
Clipping is defined as
0
10
20log
x
ibo
σ
=
(6.3)
The SER was found to be
( )
2
1 1
s e
SER p p = = − − (6.4)
where
1 3
2 Pr
1
e
M
p r
M M
ασ
−
= >
−
(6.5)
and
0 0
0
1 2
y x
Q
x
α
σ
= −
, Q is the Gaussian error function, σ is the variance of the
input signal, and M is the Mary constellation mapping type.
A simple simulation model was created where an IFFT was used to modulate the
complex samples, which were then clipped using the Cartesian method. A FFT was
used to demodulate the data.
Many interesting aspects of clipping were revealed analytically and supported through
simulation, namely
• The BER is not uniform across all the subcarriers, some subcarriers have
slightly worse performance.
• Increasing the number of subcarriers (which increases the CF) distributes the
clipping noise over more subcarriers improving the BER, especially in higher
order Mary constellation (M>64).
• The authors note the limitation of a Gaussian assumption of noise as
identified in reference [89]. The Gaussian assumption only holds for hard
clipping, at higher IBO backoffs, the noise tends to have an impulsive
distribution.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
115
Reference [89] also analytically examined clipping in the baseband focusing on
magnitude clipping and derived an expression for the SER and BER versus the clip
level, as well as performance in both AWGN and Rayleigh fading channels. Bounds
on the probability of error due to clipping are derived for both the transmitter and the
receiver. The analysis treats clipping noise differently to the standard AWGN
assumption which is sufficient if the clipping level is set low enough so that there are
a number of clips per symbol. In practice, the clipping level is set higher so that a low
probability of error is maintained making clipping a rare event. Under this condition
the clipping noise is of an impulsive nature as identified in reference [88] leading to a
different type of error mechanism.
Clipping is performed as (6.6)
( )
l
y h x x
l
−
= =
x l
x l
x l
≤
<
≥
(6.6)
where x(t) is a continuous time baseband multicarrier signal, y(t) is the clipped output,
and l is the clipping level.
Reference [89] identified from the distortion spectrum analysis that the probability of
error varies across the subcarriers with the lower subcarriers dominating the errors,
(assuming a constant constellation size on all subcarriers). The probability of symbol
error for a discretely sampled signal is given as (6.7)
( )
( )
1
3
2
2
8 1 3
Pr( ) ( ).
8 1
N L l
error Q l Q
L
L
π
 
1
 −
1
=

1
− 
 1
¸ ]
\ ¹
(6.7)
where N is the number of subcarriers, L is the constellation type (eg: L=2 is 4 QAM;
L=8 is 64 QAM), and Q(.) is the Gaussian error function. An expression for the BER
is given as (6.8)
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
116
( )
( )
1
3
2
2
2
4 1 3
Pr ( ) ( ).
log
8 1
b
N L l
error Q l Q
L L
L
π
 
1
 −
1
=

1
− 
 1
¸ ]
\ ¹
(6.8)
The AWGN approach is also presented for comparison and is given as (6.9)
2
1 3
Pr( ) 4
1
c
L
error Q
L
L σ
−
=
−
(6.9)
where
( ) ( )
2
2 2
2
exp 2 1
2
c
l
l l Q l σ
π
= − − + + (6.10)
Reference [89] noted that the Gaussian model is pessimistic as it does not take into
account that some of the noise power will fall out of band, and will therefore not
contribute to the in band distortion.
Comparing the two analytical methods in Figure 6.1 it is seen that even with the
pessimistic assumption of (6.9) the error probability is much lower than (6.7). This is
because instead of being spread uniformly over time, as is assumed in (6.9), clipping
noise is actually concentrated in time as impulses, which leads to a greater error
probability. The AWGN model is appropriate for hard clipping levels but it
underestimates the error probability by several magnitudes at higher clipping levels.
It should be noted that the yaxes in Figure 6.1 exaggerates the difference between the
2 equations. A probability of error of 10
6
is probably sufficient.
The advantage of (6.7) is that the error probabilities for high clipping levels can be
calculated analytically avoiding laborious simulations times which would be required
to get accurate results at these levels.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
117
2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4
10
16
10
14
10
12
10
10
10
8
10
6
10
4
10
2
10
0
Clip level
P
r
o
b
a
b
i
l
i
t
y
o
f
e
r
r
o
r
Pe vs clip level using AWGN approachMetsgadh
AWGN model
Impulsive noise model
Figure 6.1: Analytical symbol error probability from (6.7) and (6.9) for 64 QAM and N=64.
The affect of clipping in the presence of channel impairments is also analyzed both
through analytical methods and simulation to observe the effect at the receiver. The
simulation model uses the Hiperlan2 [52] specification with a guard interval length of
16 samples, perfect synchronization and a 1 tap equalizer. Analytical analysis
revealed that the error probability is further degraded in the presence of channel fades,
together with clipping at the receiver. The BER performance was degraded by 1 to
1.5dB on the lower subcarriers. The analytical model is within an order of magnitude
of the simulated results.
Reference [90] presented another analytical derivation of the SER resulting from
clipping in the baseband extending on the impulsive nature of noise at higher clip
levels first presented in reference [89]. Reference [90] claims that the SER curves of
reference [89] are too pessimistic and that their claim that approximations become
tight in the higher OBO region is unsubstantiated.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
118
Reference [90] proposed a different technique to limit the upper bound by using the
Chernoff bounding technique. The expression for the average SER due to clipping is
given as (6.11)
( ) ( ) ( ) ( )
( ) 0
0 0
4min
N
d N
SER e b J r p r dr
ε
ε
λ λ ε λ
∞
−
>
≤ + +
∫
(6.11)
where ( )
2
cos
0
0
1
:
2
x
J x e d
π
ϕ
ϕ
π
=
∫
is the modified Bessel function of the first kind.
Unfortunately no closed form solution is provided, however the computational effort
is much less compared to long simulation times required for smooth curves below 10
6
probability. Due to the Gaussian assumption of the non linearity the results are only
within an order of magnitude of simulated results provided N>256 and the clip level,
λ , is greater than 7dB.
Further expressions for the SER in AWGN and Rayleigh fading channels using the
previously described Chernoff method are provided. Oversampling at both the
transmitter and receiver is also treated, resulting in the realization that the side lobes
generated by the transmitting non linearity are suppressed at the receiver by the filter.
Furthermore, reference [90] claimed that the average SER is the same for the nyquist
and oversampling case, only the out of band radiation is reduced by oversampling.
An interesting claim which is not supported by simulations of the BER in the next
chapter. The asymptotic behavior is treated and it was found that for N →∞,
N
λ →∞ the Gaussian model matches the Chernoff bound of this paper in that the
0 SER → .
Reference [91] looked at the out of band radiation produced by clipping and presented
correcting functions which limit the signal while avoiding out of band radiation, and
keeping the in band interference to a minimum. This is relevant when oversampling
is performed at the IFFT as out of band radiation is created by the clipping process.
Filtering after clipping reduces the out of band radiation but regrows previously
clipped peaks.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
119
Two correction functions were suggested, the first is a Gaussian correcting function
k(t) which is an additive correction of the OFDM signal. If the signal exceeds the
amplitude threshold A
0
at times t
n
, then the corrected signal is (6.12)
( ) ( ) ( ) c t s t k t = + (6.12)
where ( ) ( )
n n
n
k t A g t t = −
∑
, ( )
2 2
2 t
g t e
σ −
= , and ( ) ( )
( )
( )
0
n
n n
n
s t
A s t A
s t
= − − . The
correcting function must be normalized so that g(0)=1, which limits the signal s(t) to
A
0
at the positions t
n
. However, the correction function may cause peaks in other
positions, but this consequence is shown to have a minor effect. Other functions for
g(t) are developed which cause no out of band interference and keep the in band
interference to a minimum. A Gaussian function is defined as (6.13)
( )
( )
1
2
0
1
0
0 1
N
j k ft
k
k
N
k
k
g t G e
g G
π
−
∆
=
−
=
=
= =
∑
∑
(6.13)
And a sinc correcting function is defined as (6.14)
( )
1
2
0
1
N
j k ft
k
g t e
N
π
−
∆
=
=
∑
( ) sin
j Bt
c Bt e
π
π ≈ (6.14)
The correcting function (6.14) can correct an amplitude peak in an OFDM signal with
minimal in band distortion and no out of band radiation. Note that if the signal is not
oversampled then the correction scheme is the same as normal clipping.
Simulations with the correcting functions were performed with N=128 subcarriers and
an oversampling rate of 4. The signal is corrected with k(t) and any peak regrowth
after the correction is clipped at A
0.
An IBO level of 4dB was used and the algorithm
was tested in both an AWGN and fading environment. It was shown that the
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
120
correction functions introduce more inband noise than standard clipping, however the
slight increase in inband noise is offset by the alleviation of out of band interference.
For example, at an IBO of 4dB standard clipping has a Signal to Interference Ratio
(SIR) of 21dB, Gaussian 13dB, and Sinc 16dB. The BER vs IBO in an AWGN
channel with SNR of 18dB shows a magnitude greater degradation for the sinc
correction function at 4dB IBO, but in the fading channel the BER degradation is
negligible for the sinc case and only marginally worse for the Gaussian case.
Clipping and filtering issues were addressed in reference [92] where an oversampled
(LN where L is the oversampling factor) IFFT zero padded in the middle is used to
modulate the OFDM symbol. The resultant samples are then clipped by a SL in the
normal way, as this results in out of band radiation as described section 3.4.2 the data
is filtered by a FFT/IFFT pair of size LN. The filter passes the wanted in band
samples while nulling out the out of band components. The advantage of this
technique is twofold, firstly out of band radiation is greatly attenuated and secondly,
by oversampling peak regrowth after filtering is greatly reduced (refer to Figure 3.11).
Simulations were performed with 4QAM data and N=64 subcarriers. CCDF results
for a CR of 6dB compared the new algorithm (L=2, N=64) with standard non
oversampled clipping (L=1, N=64) where it is seen that the L=2 case has 1dB less
peak regrowth than the L=1 case at 10
5
probability region. It was noted that
increasing the oversampling rate (L>2) at the IFFT provided minimal further
improvement. Out of band radiation is also analyzed in the form of the PSD where a
perfectly linear amplifier with a CR 1dB higher than the baseband clipping level is
used after modulation with a carrier frequency. Out of band radiation was reduced
down to 65dB using the new technique (L=2), compared to 55dB for clipping before
interpolation (L=1) and 45dB with no clipping before amplification. In band distortion
is also an affect of clipping, reference [92] stated that clipping adds a noise like
component and a reduction in the constellation size (refer to Figure 3.16) which can
be corrected at the receiver with AGC. Also as the noise from clipping is created at
the transmitter it will lessen its effect in a fading channel. These properties will
improve the BER. This technique can be implemented in existing OFDM systems
and requires no redesign at the receiver only replacing the IFFT at the transmitter.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
121
Another technique to lessen the effects of clipping [93] called Decision Aided
Reconstruction (DAR) reduces clipping noise with an algorithm in the baseband at the
receiver. Like the method of [92] it uses a FFT/IFFT pair in an iterative algorithm at
the receiver to try to estimate which samples were clipped at the transmitter. When
the clipping noise is large compared to the AWGN in the channel, performance is
limited by the clipping noise. Using the FFT/IFFT pair to make decisions in the
frequency domain regrows samples that were clipped at the transmitter, and although
they are still distorted decisions made on the new symbols are much less affected by
clipping noise. However much of the gain from DAR may be achieved by simply
correcting the constellation shrinkage which clipping causes (refer to Figure 3.16).
The algorithm can predict false clipped peaks when the clipping level is set too low
(<2dB) worsening performance more than standard clipping. Through simulation
reference [92] found that DAR worked best when a small number of samples were
clipped, and that it worked better with higher order constellations with a clip level
>4dB. For example in a 64QAM OFDM system with a CL of 5dB in an AWGN
channel the improvement was quite dramatic being only ~1dB lower than the
theoretical lower bound. It was also noted that a slight further improvement is seen as
N is increased. The number of iterations required for good performance was shown to
be around 3. The problem with the methods of references [92, 93] is the latency and
complexity of performing extra FFT/IFFT operations.
6.1.1 Quantisation and Clipping
An early paper to investigate the relation between clipping and quantization in DMT
transceivers is reference [94] where an analytical expression was developed to find
the minimum number of bits required in the A/D, D/A converters when the signal is
clipped to a predefined level. Due to the Rayleigh nature of the envelope of the DMT
envelope 2 to 3 bits can be saved in the A/D, D/A operation without changing the
SNR. An expression to calculate the number of bits that can be saved is (6.15)
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
122
( )
2
1
2 2 3
2
2 2
3 8 .2 . .
1
.log
2
R
e
µ
ν π µ
µ
−
−
−
∆ =
(6.15)
where
1 2
R R ∆ = − is the number of bits saved, R
1
is the number of bits required for
the A/DD/A when no clipping is performed, R
2
is the number of bits required for the
A/DD/A with clipping to keep the same SNR as when no clipping is
performed,. ( )
1 2 1
. 3 .
2 1
L
N
L
ν
+ −
=
+
is a parameter set by the number of
subcarriers, N, the QAM constellation size L
2
, (L=4 equates to 16 QAM), and
max
A
µ
σ
= is the clip level. A
max
is the clip level in volts and σ is the rms voltage of
the DMT transmit symbol. It was assumed in the paper that all subcarriers have the
same constellation type, however the authors claim that different mapping types on
subchannels would have a minimum effect.
Clipping and quantization are further explored in reference [95] for DMT based
transceivers where a improved clipping technique allows for up to an 8dB
improvement in the SNR over standard clipping. The new method analyzed the
samples after the IFFT, if a sample is above the clipping threshold, A
clip
, then the
phase of each QAM modulated carrier is changed by means of a fixed phasor rotation,
and a new DMT symbol is generated by the IFFT. By careful selection of the phasor
rotation, the probability of the new symbol requiring clipping at A
clip
will be reduced.
Otherwise the symbol is sent on to further processing unmolested.
The overall probability of clipping for the 2 pass method described above is (6.16)
2
/ /1 / 2
.
Clip Total clip clip clip
P P P P = = (6.16)
which is determined to be
2
2
/
1
2
N
Clip Total
P erf
µ 1  
= −
1 
\ ¹ ¸ ]
(6.17)
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
123
where /
Clip
A µ σ = .
Recalculating the IFFT can create a bottleneck slowing down system performance,
but does not necessarily require a factor 2 increase of the IFFT as not every symbol
will require clipping, especially at higher levels of A
clip
. Side information is also
required to inform the receiver of the number of passes, p, used (if any) of log
2
p bits
per transmitted symbol.
As seen in Figure 6.2, the clipping probability drops with an increase in the number of
passes. At µ =4 the 2 pass and 3 pass methods reduce the probability of clipping
down to ~10
4
and ~10
6
respectively.
2.5 3 3.5 4 4.5 5
10
14
10
12
10
10
10
8
10
6
10
4
10
2
10
0
A
cl i p
/sigma
P
c
l
i
p
/
p
p=1
p=2
p=3
Figure 6.2: Probability of clipping DMT signal as a function of µ for p=1,2,3. N=64 subcarriers.
Quantization effects are also examined in terms of noise from the DAC and ADC in
the transmitter and receiver respectively for various wordlengths, b. For 16 QAM,
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
124
N=256, b=12, 3.7 µ ≥ with p=1 and 3.4 µ ≥ with p=2 an improvement of 3 and 8dB
can be achieved.
Another reference [96] analyzed quantization effects on OFDM used a simulation
model with Hiperlan2 [53] specifications. Clipping is performed on the I and Q
outputs of the IFFT. As the wordlength at the IFFT output is decreased, the power
consumption and complexity of the DAC/ADC decreases at the expense of
quantization noise, which increases the BER. However as noted in reference [94] the
wordlength can be reduced with minimal affect on the BER. Also clipping at the
IFFT output increases the resolution giving a better average signal/quantization noise
power ratio, of course this is at the expense of clipping noise.
In a nutshell, lowering the clipping level increases the clipping noise while at the
same time reducing the quantization noise. Results from reference [96] indicated that
the optimum clipping level for wordlengths between 6 and 9 bits occurs at around 4σ
(slightly lower for smaller wordlengths). An 8 bit wordlength is also recommended
with an extra 2 bits for the receiver ADC to compensate for peak regrowth affects
after transmit filtering (upsampled by 4 before filtering) and imperfect AGC in the
receiver.
Reference [97] examined the effect of rounding and saturation in fixedpoint DSP
implementation of the IFFT and FFT where optimum tradeoffs are found between
saturation and rounding. Results from simulation revealed that performance for fixed
point FFT’s is improved when overflow is allowed to occur with low probability. The
distribution of error was shown to depend on the ratio of the maximum quantization
level to the RMS power of the random variable. For 16bit arithmetic the headroom
was shown to be around 15dB. Doubling the size of the FFT resulted in an
improvement in the new scaling method of 3dB.
6.2 Amplifier non linearities
As shown in Section 3.4 if no attempt is made to control the peak excursions of an
OFDM symbol the HPA will saturate causing spectral regrowth and an increase in the
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
125
BER at the receiver. This section reviews papers which analyze the effect of an
uncontrolled OFDM symbol on the HPA.
An early reference [42] to look at the effect of a non linearity on QPSK OFDM
compared analytically derived and simulated results for the BER versus SNR.
Specifically they assumed that the intermodulation products were generated by non
linearities in the receiver IF module. These results are stated to be applicable to the
transmitter amplifier, and that only 3
rd
order distortion will affect the system.
The BER is calculated under the assumption that the intermodulation products cause
an additive Gaussian interference and that the BER is approximately equal on all
subcarriers. The non linearity is set at the 1dB compression point of the in band
output and is related to the output signal, Vo, by
1
0.27
dB o
V V = . The simulated results
for the amplifier which is linear up to 3dB, with 1dB backoff show that even with a
1dB backoff the BER is within 1 order of magnitude of the linear amplifier at
SNR=10dB. The simulated results were shown to be in good agreement with
theoretical results.
Reference [98] focused on simulation comparing Single Carrier (SC) and OFDM
systems with clipping in the baseband and RF amplifier non linearities. The
simulation model used N=2K and 8K (DVB, DVA) subcarriers with QPSK and 16
QAM mapping to produce SER versus clipping and AWGN plots in a Monte Carlo
simulation. It was shown that the SER versus SNR is almost the same for both
systems with OFDM having slightly better performance due to the frequency guard
interval used in OFDM which reduces the equivalent noise bandwidth. The number
of subcarriers has a negligible affect on OFDM due to the Gaussian distribution of
samples.
The baseband clipping effect was simulated with the SNR set at 16dB for 16 QAM
and the backoff was defined as
2
2
2
A
BO
σ
= . Again the number of subcarriers was
shown to have a negligible effect on the BER. The performance of the SC and
OFDM system converges at a baseband clipping of BO=6dB for OFDM when 16
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
126
QAM mapping was used, a backoff of 3dB was only required for QPSK due to the
larger Euclidean distance of QPSK constellation points.
For the RF amplifier the low pass equivalent was used, where the signal is of the form
( )
( ) j t
A t e
θ
, the HPA output can then be expressed as (6.18)
( ) ( )
( ) ( ) ( ) j t g A t
y t f A t e
θ +
=
(6.18)
Where [ ] f A is the AM/AM characteristic of the amplifier and [ ] g A is the AM/PM
characteristic of the HPA. The phase distortion was considered to be linear.
Performance for 16 QAM OFDM under different amplifier backoffs is almost
identical to the baseband clipping effects with a 6dB BO required to bring
performance in line with its equivalent SC counterpart. It was also shown that the
degradation increases as the third order interception point approaches the 1dB
compression point. QPSK was again shown to be more robust to amplifier non
linearity.
Reference [40] both analytically derived and simulated an OFDM system in a AWGN
channel with and without equalization (1 tap) to produce BER versus SNR plots.
Amplifier non linearity (using TWT as described in Section 3.4.1) is stated to cause
two effects on the detected samples:
• constellation warping (amplitude and phase distortion)
• non linear distortion which generates a Gaussian spread like cluster of
received values around each constellation point (refer to Figure 3.16)
For 16 QAM OFDM system it was shown that at high IBO (25dB) equalization had
no effect on the BER. As the IBO is reduced (14dB) the equalized BER degraded
only a little, however the non equalized OFDM model lost a further 7dB of SNR. At
12dB IBO the non equalized system lost so much SNR that its error floor became 10

3
. For all cases a good agreement is shown between analytical and simulated results.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
127
Reference [47] provided an analytical analysis of smooth non linear distortion on the
SER as a function of third order distortion (the most dominant distortion for smooth
non linearities) in a memoryless non linear power amplifier. The analysis was
performed on a matched filter pair. Initially the spectrum of the distorted OFDM
signal is examined, which is then used to determine the detection error for the
matched filter detection. It was found through analytical means that the variance is
approximately equal on all subcarriers with the middle subcarriers experiencing the
most noise, this means that the SER is similar on all subcarriers, a result supported by
reference [88].
Simulated results are compared to analytical results for QPSK and 16 QAM OFDM.
The simulation model used an equivalent low pass representation to avoid RF up and
down conversion, N=1024 subcarriers were processed by an IFFT at which point the
non linearity was performed. The data was demodulated with an FFT and the
transmit symbols were compared to the received samples. The difference was
squared and stored and then averaged to find the noise variance. The simulation
results for 16 QAM were found to better fit the analytical results than QPSK, the tails
in both cases fall of quicker in the simulated case because they are not exactly
Gaussian. However a good agreement is seen between the analytical and simulated
curves.
Reference [41] extended on earlier work presented in reference [40] to theoretically
analyze the effect of non linear amplifiers in conjunction with phase noise on M
QAM OFDM. Phase noise is caused by the oscillators in the RF stage and becomes a
more dominant source of noise at higher carrier frequencies (up to 40GHz). The
theoretical expressions were supported by simulations using SSPA and TWTA with
different values of phase noise. Under the assumption of modeling the phase and
amplifier distortions as additive Gaussian noise, the computed variances are used to
get an estimate of the BER in an AWGN channel.
Results indicated that while QPSK is rather impervious to both amplifier non linearity
and phase noise, the performance of 16 and 64 QAM is greatly diminished in terms of
the BER even with a large OBO in the amplifier. A SSPA with p=2 introduced 4
th
order distortions which further diminished the performance of 16 and 64 QAM. In
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
128
general it was shown that the joint effects of amplifier non linearity and phase noise
have two major effects. Amplifier non linearity generates a uniform amplitude
attenuation and phase rotation which can be corrected at the receiver by AGC. Phase
noise introduced a constant phase rotation within 1 OFDM symbol which can be
estimated and corrected using the pilot tones. The second effect is constellation
clustering due to the interference produced by the HPA to the ICI caused by the phase
noise.
Finally it was shown that the phase noise impairment was dependant on the
relationship between the phase noise rate and the OFDM symbol period. It was
shown that the phase noise can become a limiting factor if a large number of
subcarriers and a high frequency carrier are used.
6.3 Windowing
Windowing or pulse shaping are similar to clipping in that they attenuate large peaks.
However in windowing a corrective window is multiplied with the data so that not
only the peak cancelled but surrounding samples are also affected. The advantage of
this process is to keep the OBR lower than in standard clipping. Windows should be
as narrowband as possible in the frequency spectrum domain, so as to have good OBR
properties. However narrowband windows have the reciprocal affect of being long in
the time domain which implies many signal samples being affected, which increases
the BER.
Reference [99] uses window types such as such as cosine, Kaiser, and Hamming,
comparing their use to standard clipping in terms of the frequency spectrum and BER.
The simulation model uses Hiperlan2 [53] specifications with a ½ rate convolutional
code and 16 QAM where it is shown that clipping the signal at 5dB has a minor
affect on the BER with a 0.2dB loss in SNR. Windowing is shown to have almost
identical affect on the BER above 5dB clipping but interestingly has worse
performance at harder clipping levels.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
129
To simulate the affect of the required backoff in the HPA a SSPA with p=3 as
described in Section 3.4.1 is used. In order to keep the OBR to below 30dB for 64
subcarriers a backoff of 6.3dB was required, which could be reduced by 1 dB to
5.3dB when peak windowing is used. When 256 subcarriers are used the backoff of
6.3dB can be reduced by 0.8dB to 5.5dB with peak windowing, showing that
windowing is independent of N.
A later reference [100] uses broadband pulse shaping on individual subcarriers as a
way to reduce the PAPR. By making the cross correlation between samples in the
same block close an OFDM signal with a low PAPR can be created. The new OFDM
signal is given as
( ) ( ) ( )
2 j m T
n m
x t X m p t e
π
=
∑
( ) 1 nT t n T ≤ ≤ + (6.19)
Where ( )
n
X m is the modulated data symbol of subcarrier m, T is the duration of the
OFDM block, and the waveform ( )
m
p t is a pulse shape of duration T, on subcarrier
m which has a bandwidth less than or equal to the bandwidth of the OFDM signal
x(t).
Unique RRC waveforms are multiplied with each sample, which are cyclic shifts of
each other within the same time interval 0<=t<T. As each RRC pulse is seeded from
the same source they are easy to create, the RRC pulse for each subcarrier is defined
as
( ) ( )
1
2 2
mk k m
N L
j j t
N N
m
k L
p t C k e e
π π
−
+ −
− −
=−
=
∑
0 t T ≤ < (6.20)
where
( ) ( )
2
0
1 1
k
j t T
T
m
C k p t e dt P
T T T
π −
= =
∫
(6.21)
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
130
is the Fourier series of p(t) and
( )
( ) 2 ,
0,
rc
p t T
p t
−
=
0 t T
elsewhere
≤ <
(6.22)
where ( ) / 2
rc
p t T − are the samples of the time domain RRC pulse.
When the rolloff factor is increased to 0.5 the reduction in the CCDF at Pr(ζ>ζ
0
)=10
3
is around 6.5dB. However the effect of oversampling and the frequency spectrum is
not treated. I anticipate that the spectrum is at least 50% larger than normal OFDM
because the spectrum of the broadband pulse would have to be convolved with the
basic OFDM linear spectrum.
Note that reference [100] is actually mathematically equivalent to clipping and
filtering with a linear time invariant filter.
6.4 Conclusion
This chapter presented distorted techniques for the reduction of PAPR in OFDM.
Clipping in the baseband was first introduced as this is the simplest and most widely
examined area in distorted PAPR reduction techniques. Clipping was shown both
analytically and through simulated means to have a minor effect on the BER when
QPSK modulation was used due to the large Euclidean distance between constellation
points. Higher order mapping types such as 16 and 64 QAM were much more
susceptible to clipping. The BER on individual subcarriers was also treated where it
was shown that the probability of error was almost equal on all subcarriers.
Increasing the number of subcarriers was shown to have a beneficial affect as the
noise introduced by clipping would be spread over more subcarriers.
The Gaussian like assumption of the noise which is assumed in most analysis of
clipping noise was also shown to be unsubstantiated resulting in optimistic error
probabilities. Clipping noise was shown to have an impulsive nature at higher clip
levels resulting in a much smoother decay in error probability as the clip level was
increased.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
131
Clipping at the nyquist rate (no oversampling) does not cause out of band radiation as
all the noise was shown to fall in band. Oversampling before clipping was shown
[46] to produce less in band noise but to increase the out of band noise, requiring
filtering.
Quantization in hardware and clipping was also treated where it was shown that some
bits could be saved after clipping, due to the Rayleigh distribution of samples without
degrading system performance. This also improves the resolution of the clipped
samples.
Furthermore it was shown that several steps could be taken to mitigate the errors
caused by clipping. As clipping noise caused shrinking in the constellation size as
well as a Gaussian like spreading the AGC in the receiver could be used to correct this
depending on the clip level. Also, the noise due to clipping will also experience
fading along with the signal further lessening its effect.
Early windowing and pulse shaping techniques displayed little improvement and
sometimes a further degradation in the BER while attempting to reduce the OBR.
Later work in this area produced a markedly greater improvement in the PAPR by the
selection of appropriate pulse shapes which were applied to individual subcarriers
rather than the whole transmit signal.
Chapter 7: Reduced Complexity Clipping Algorithms
132
Chapter 7
Reduced Complexity Clipping
Algorithms
The previous chapter covered distorted techniques for PAPR reduction. As described
clipping suffers after filtering as clipped peaks can regrow resulting in saturation of
the HPA. Also, although clipping is less complex in terms of hardware operations
than distortionless techniques, estimates of the magnitude still need to be made in
order to decide whether a sample needs to be clipped or not and multiplications have
to be made to correct the signal. This chapter presents new low complexity clipping
techniques which avoid complex hardware operations while maintaining similar
performance to conventional clipping. The new clipping algorithms are then
implemented in a new clip and filter algorithm which is much less susceptible to peak
regrowth after baseband filtering.
Section 7.1 describes conventional clipping and a simulation model is developed to
quantify the effects that various transceiver components have on the BER
1
. This
model is then used to test the new clipping algorithms. Section 7.2 details a new
technique coined Sector Clipping and provides theoretical and simulated analysis of
the new method. Section 7.3 presents another new technique which is similar to the
CORDIC algorithm but with reduced complexity called Vector Subtraction. Section
1
In this work the BER will be plotted against clipping level with the noise set to zero. All the errors
are therefore caused by clipping noise. The BER’s in these plots therefore represent the ‘error floor’ of
the more commonly used BER vs. SNR plots.
Chapter 7: Reduced Complexity Clipping Algorithms
133
7.4 compares the new and existing clipping methods in terms of their baseband
Clipping Level (CL) vs. the BER. Section 7.5 implements the new algorithms in a
new clip and filter algorithm which is less susceptible to peak regrowth. Finally
section 7.6 concludes the chapter with a review of the advantages of the proposed
clipping techniques.
Note that the results in Sections 7.1 and 7.4.1 show the Bit Error Rate Floor (BERF),
i.e. the BER due to clipping as that is the focus of this chapter. For comparison
Figure 7.1 shows the BER with Additive White Gaussian Noise (AWGN) in the
channel and the CL set at 3, 5dB and no clipping. 4, 16, and 64 Mary QAM symbols
are modulated with a 64 point IFFT and then pulse shape filtered with a Root Raised
Cosine Filter (RRCF) with α=0.35, and 128 taps. It can be observed that 16 and 64
QAM mapping is much more susceptible to clipping noise than 4 QAM. The error
floors for 64 QAM can be cross referenced with figure 7.4. The OFDM transceiver
system is shown in Figure 7.3.
0 5 10 15 20 25 30
10
4
10
3
10
2
10
1
10
0
AWGN (dB)
B
E
R
No clipping
CL=5dB
CL=3dB
QPSK
16 QAM
64 QAM
Figure 7.1: Average noise in the channel vs. BER for 4, 16, and 64 QAM.
Chapter 7: Reduced Complexity Clipping Algorithms
134
7.1 Conventional clipping
Conventional clipping is defined here as any hardware method which reduces the
amplitude of the signal to a predefined level in line with the origin as shown in Figure
7.2.
Figure 7.2: IQ diagram showing conventional clipping region. The vector r is reduced to r
clip
.
Clipping in this way only introduces amplitude distortion, the phase is unaffected.
This method is the most hardware intensive of all methods described in the following
sections. Some of the operations require iterative techniques when implemented in
fixed point processes (e.g. Division and square roots) and therefore take a number of
clock cycles. Other methods require vast LUT’s which consume chip area or memory
space. Mathematically conventional clipping can be described as
( )
''
r
clip
r
x
r e
=
∡
clip
clip
r r
r r
≤
>
(7.1)
In order to analyze the performance of conventional clipping it is useful to quantify
the effects of the different system components involved, i.e. the mapping type, IFFT
size, number of filter taps, filter rolloff factor, and IBO of the HPA. A block diagram
of the simulation model used is shown in Figure 7.3.
r
clip
r
i
q
x’’
Chapter 7: Reduced Complexity Clipping Algorithms
135
Figure 7.3 Block diagram of simulation model used for clipping models.
The baseband Monte Carlo simulation model of Figure 7.3 randomly generates Mary
mapped data and then modulates the signal with the IFFT. Note the data is buffered
at the input to the IFFT so that N samples are fed into the IFFT, after modulation the
data is converted back into serial form and sent to the baseband clipping block. After
clipping to a predefined level relative to the mean power of the transmit symbol a
cyclic prefix can be added to the data. The data is then interpolated by a factor of 8
before being filtered by a matched Root Raised Cosine Filter (RRCF). A SSPA (as
described in Section 3.4.1) models the HPA. AWGN and multipath delayed versions
of the signal can then be added in the channel.
At the receiver side the process is reversed, a matched RRCF filters the received
samples which are then decimated by 8 to retrieve the transmitted samples. The
cyclic prefix (if used) is removed prior to demodulation with the FFT. Again note
that the data is buffered at the input to the FFT until all N samples are ready. Finally,
a simple Least Square (LS) algorithm is used to make decisions on the decoded data.
A comparison between the gray encoded transmitted bits and the received bits is made
to determine the BER.
In the results that follow the conventional baseband clipping algorithm is used as
described in Figure 7.2. Note that 64 QAM mapping is used unless otherwise stated
Data
gen.
Mary
mapping
IFFT Clip
block
CP
Interp
by 8.
RRC
Filter
AWGN
Matched
RRC
Filter
Deci.
by 8
FFT
Remove
CP
Demap Data
decision
BER
calc
SSPA
Chapter 7: Reduced Complexity Clipping Algorithms
136
since the more traditional QPSK has such a high tolerance to clipping excessive
simulation times are required to get BER plots. In the following simulations 10,000
OFDM symbols are transmitted, each symbol has 52 information bearing subcarriers.
For 64 QAM this means around 3 million bits are transmitted.
Effect of filter on the BER
Figure 7.4 shows the baseband clip level vs. the BER for a RRCF with different
numbers of coefficient taps and roll off factors (named alpha in figure 7.4). There are
no other sources of distortion or noise. The number of taps is set at 64, 128 and 256,
and the roll off factor or excess bandwidth is set at 0.15 and 0.35.
0 1 2 3 4 5 6 7 8
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
alpha=0.35
alpha=0.15
64 filter taps
128 filter taps
256 filter taps
Figure 7.4: Baseband clip level vs. the BERF for varying RRCF parameters. 64 QAM symbols, 64
point IFFT, LPA, no channel impairments. AWGN=0.
Two points can be made reviewing Figure 7.4, the first is that when the easier roll off
or excess bandwidth of α=0.35 is used the number of taps has little effect on the BER.
The second point is that when a tighter roll off factor is used (α=0.15) the BER is
substantially affected by the number of taps. 64 filter taps has a BER almost 2
magnitudes worse than 128 taps at CL=7dB. This is due to the inband amplitude
Chapter 7: Reduced Complexity Clipping Algorithms
137
distortion (ripple) created when insufficient taps are used with a small excess
bandwidth (refer to Figures 7.5a and b). This creates a linear spreading of the
demodulated samples away from the origin. In all cases α=0.35 has worse
performance than α=0.15 for the same number of filter taps.
The 0.35 roll off factor curves will have a wide transition bandwidth in the frequency
domain. The extreme subcarriers will then be effected by additional attenuation
which will reduce the noise margin in the receiver decision (slicer). The 0.15 roll off
curves have a steeper transition band and avoid the problem of the extreme
subcarriers. However passband ripple will be introduced when the number of taps is
low (taps=64).
The Hiperlan2/802.11a physical layer specification requires a tight roll off, (α≈0.15)
so that the filter skirts occur in the null subcarriers. In all further simulations 128 filter
taps with a roll off of α=0.15 will be used unless stated otherwise.
Effect of HPA amplifier backoff and IFFT on the BERF
Figure 7.6 and Figure 7.7 again shows the clip level vs. BERF with 64 QAM
mapping. This time the IFFT size is set to 64 and 128 while maintaining the same
Figure 7.5b: Demapped constellation, M=64, with
no clipping or channel impairments. 128 filter taps
in RRCF, α=0.15.
Figure 7.5a: Demapped constellation, M=64, with
no clipping or channel impairments. 64 filter taps
in RRCF, α=0.15.
Chapter 7: Reduced Complexity Clipping Algorithms
138
number of information bearing subcarriers, i.e. the data is therefore critically sampled
(almost) and oversampled by a factor of 2 respectively. The other variation is the
inclusion of a HPA with IBO. The HPA follows the Rapp model of (3.34) with p=3.
This allows us to see the effect of peak regrowth on the BERF with and without
oversampling.
In Figure 7.6 (os=1) even with an IBO of 4dB above the baseband clipping level
(IBO=CL+4dB) peak regrowth still causes a small amount of saturation in the HPA,
degrading the BERF. With no extra IBO in the HPA (IBO=CL) the BERF is between
10
2
and 10
3
at CL=6dB while with a LPA the BERF is below 10
4
at 6dB baseband
clipping, a 1.5 order of magnitude improvement.
In Figure 7.7 (os=2) the performance is better for all IBO across the board. Some of
the clipping noise falls into the null bins and is subsequently filtered away while os=1
systems would cause this noise to fold back into the inband subcarriers. For no extra
IBO in the HPA the 64 point IFFT has a BERF between 10
2
and 10
3
at 6dB baseband
clipping (Figure 7.6) while the 128 point IFFT has a BERF just above 10
3
at the same
clip level, an improvement of half a magnitude. Note in Figure 7.7 that for the curve
IBO=CL+4 to the LPA curve (IBO=CL+∞) there are no errors occurring in over 3
million transmitted bits at 7dB clipping level.
In order to further highlight the effect of oversampling, the CCDF in Figure 7.8 is also
shown for the case described in Figures 7.6 and 7.7, showing the peak regrowth under
critically and oversampled conditions. In this case the baseband clip level is set at
5dB, peak regrowth for the critically sampled case is extreme with almost 5dB peak
regrowth at Pr(ζ>ζ
0
)=10
4
. Peak regrowth is 3dB (2dB less) for the oversampled case
(os=2) at Pr(ζ>ζ
0
)=10
4
. Results not shown here indicate that the harder the clip level
the more extreme the peak regrowth after filtering.
Chapter 7: Reduced Complexity Clipping Algorithms
139
0 1 2 3 4 5 6 7 8
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
IBO=CL
IBO=CL+2
IBO=CL+4
LPA
Figure 7.6: Baseband clip level vs. the BERF with varying IBO in HPA. 64 QAM symbols, 64 point
IFFT (os=1), RRCF with 128 taps and α=0.15. p=3 in HPA. AWGN=0.
0 1 2 3 4 5 6 7 8
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
IBO=CL
IBO=CL+2
IBO=CL+4
LPA
Figure 7.7: Baseband clip level vs. the BERF with varying IBO in HPA. 64 QAM symbols, 128 point
IFFT (os=2), RRCF with 128 taps and α=0.15. p=3 in HPA. AWGN=0.
Chapter 7: Reduced Complexity Clipping Algorithms
140
3 4 5 6 7 8 9 10
10
4
10
3
10
2
10
1
10
0
P
r
(
ζ
>
ζ
0
)
64 point IFFT
128 point IFFT
After Filtering
After HPA
After Baseband
Clipping
ζ
0
(dB)
Figure 7.8: Simulated CCDF clipped in baseband at 5dB, IBO in HPA set to 8dB for 64 and 128 IFFT.
Effect of constellation size on the BERF.
Figure 7.9 shows the effect of changing the mapping constellation M with an
oversampling factor of 2 in conjunction with a LPA. Although 4 QAM was
simulated, it is extremely impervious to clipping with no detected errors at 0dB and
above. 16 QAM also has rather robust performance in the presence of clipping with
no errors being detected above 4dB clipping at 10
5
probability. 64 QAM clipping has
performance around 2 magnitudes worse at equivalent clipping levels to 16 QAM.
Chapter 7: Reduced Complexity Clipping Algorithms
141
0 1 2 3 4 5 6 7
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
16 QAM
64 QAM
Figure 7.9: Baseband clip level vs. the BERF with varying Mary constellations. 128 point IFFT
(os=2), RRCF with 128 taps and α=0.15. AWGN=0.
Effect of changing P in HPA
The model used to simulate the HPA is the SSPA described in Section 3.4.1. Varying
the value of p in the SSPA controls the input to output curve of the amplifier as shown
in Figure 3.14. Figure 7.10 demonstrates the effect of changing p on the BERF, the
saturation level of the SSPA is set equal to the baseband clipping level. Here it is
seen that p=1 has an extreme effect on the BERF. P=3 causes a magnitude of
degradation over the absolutely linear region of p=1000. As p=3 is a practical value
used in many designs [40] it will be used in all further simulations.
Chapter 7: Reduced Complexity Clipping Algorithms
142
0 1 2 3 4 5 6 7 8
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
p=1
p=3
p=5
p=1000
Figure 7.10: Baseband clip level vs the BERF with varying p in the SSPA. 128 point IFFT (os=2),
RRCF with 128 taps and α=0.15. HPA backoff set equal to baseband clipping level. AWGN=0.
This section provided a description and simulations of an OFDM system with
baseband clipping. The effects of oversampling the IFFT, filter parameters, HPA
parameters, and mapping type were simulated to see their effect on the baseband CL
vs. the BERF. It can be concluded that a good set of parameters for further analysis
of an OFDM system are:
• Oversampling factor of 2 in the IFFT.
• RRCF with 128 filter taps and a roll off factor of 0.15.
• A SSPA with p=3.
Chapter 7: Reduced Complexity Clipping Algorithms
143
7.2 New Sector Clipping method
A new technique for clipping developed by the author which reduces hardware
complexity is Sector clipping [3] Sector clipping avoids magnitude estimates which
require hardware multiplications to perform corrective scaling. The decision to clip is
based on the I and Q values in conjunction with comparisons between them, and so
divides the clipping region into different ‘sectors’. Figure 7.11 shows the I Q plane of
Sector clipping with the clipping regions clearly identified. The new method requires
only comparators and can be implemented in hardware as either an iterative (to reduce
complexity) or parallel structure (to increase speed). The number of sectors can vary
from 2 (square clipping) to 5 or more, although no discernable improvement in
performance is seen above this number. Note that increasing the number of decisions
beyond 5 Sector will clipping increase the complexity as much as multiplications.
As seen in Figure 7.11, Sector clipping not only introduces extra amplitude distortion
over conventional clipping, but also phase distortion as data outside the clipping
regions is not reduced in line with the origin. The symmetry of Sector clipping can be
exploited to further reduce hardware complexity, the data can be ‘folded’ into the first
octant by removing the sign bits (making it positive) and making the largest value of
the complex signal the real component, the clipping operation is then performed on
the new value. Sign bits and the relative size of the real and imaginary components
can then be used to extrapolate the original position of the clipped sample. As there
are 3 unique sectors in Figure 7.11 this structure is known as 3 Sector clipping.
7.2.1 Theoretical Analysis of Clipping Techniques
In order to compare the performance of the various clipping techniques the Clip Level
vs. SNR is mathematically derived for 3 cases: Conventional clipping, Sector clipping
with 3 sectors, and Square clipping. Square clipping can be construed as a DAC with
limited word length. Note that for the theoretical analysis, no oversampling, filtering,
or amplifier is assumed.
Chapter 7: Reduced Complexity Clipping Algorithms
144
Figure 7.11: I Q diagram showing different sector clipping regions and the direction of data reduction
for 3 Sector Clipping.
7.2.1.1 SNR Analysis
The relation between the input and output of the clipping operation is expressed
pictorially in Figure 7.11 and can be used to find an expression for the SNR.
Figure 7.12: Input output relationship of clipping operation.
Mathematically, this relation between the input, x, and the output, y can be expressed
as Bussgang’s theorem
Noise
Amp
x y
n
r
clip
θ
i
q
Chapter 7: Reduced Complexity Clipping Algorithms
145
n x y + =α (7.2)
where x in the input signal, y is the output signal and α is chosen so that the input
signal and the noise, n, is uncorrelated. Finding the 2
nd
moment of y (7.2) gives
[ ] ( )( ) [ ]
nn xx yy
R xn E R R
n x n x E yy E
+ + =
+ + =
] [ 2
2
α α
α α
(7.3)
where R
xx
is the autocorrelation of the input signal giving the input power, R
yy
is the
autocorrelation of the output signal giving the output power. As x is assumed to be
uncorrelated to the noise the term 2 [ ] E xn α can be removed and (7.3) can be
rearranged and solved for the noise power, R
nn
.
xx yy nn
R R R
2
α − = (7.4)
Rearranging (7.2)
x y n α − = (7.5)
Taking the expected moments of (7.5) and then the correlation functions yields
[ ] ( )( ) [ ]
xx xy yy nn
R R R R
x y x y E nn E
2
2 α α
α α
+ − =
− − =
(7.6)
Equating (7.4) and (7.6) and solving for α:
xx
xy
xy xx
xx xy yy xx yy
R
R
R R
R R R R R
=
− =
+ − = −
α
α α
α α α
2 2 0
2
2
2 2
(7.7)
Chapter 7: Reduced Complexity Clipping Algorithms
146
The signal, S is equal to:
xx
R S
2
α = (7.8)
and the noise is given in (7.4), the SNR is then (7.9).
xx yy
xx
R R
R
N
S
2
2
α
α
−
= (7.9)
To reduce mathematical complexity substitutions are made to factor out R
xx
. The new
SNR is given by (7.10).
2
2
α
α
−
=
xx
xy
R
R
N
S
(7.10)
The expressions for the correlation functions need to be defined for each of the
clipping techniques, which are solved to find a closed form solution.
7.2.1.2 Conventional clipping
The definitions for the autocorrelation of the input, autocorrelation of the output and
cross correlation for conventional clipping are respectively given below.
( )
∫
∞
∂ =
0
2
) 0 (
. r r f r R
xx
(7.11)
( ) ( )
∫ ∫
∞
∂ + ∂ =
clip
clip
r
clip
r
yy
r r f r r r f r R . .
2
0
2
) 0 (
(7.12)
( ) ( )
∫ ∫
∞
∂ + ∂ =
clip
clip
r
clip
r
xy
r r f r r r r f r R . . .
0
2
) 0 (
(7.13)
Chapter 7: Reduced Complexity Clipping Algorithms
147
where r is the input data (i.e. the magnitude of the data at the output of the IFFT), r
clip
is the clipping level, and f(r) is the probability distribution of the data, r, which is
assumed to be Rayleigh distributed.
( )
2
2
2
2
σ
σ
r
e
r
r f
−
=
(7.14)
Note that for the Rayleigh assumption of the distribution of ‘r’ to hold the number of
subcarriers is assumed to be greater or equal to 64 [39]. In order to reduce the
mathematical complexity of the correlation functions they are normalized by the
average input voltage, 2 σ .
2 σ
r
R = (7.15a)
2 σ
clip
clip
r
R = (7.15b)
2 σ
r
R
∂
= ∂ (7.15c)
Substituting (7.15a, b, c) into (7.11), (7.12), and (7.13) yields the normalized
expressions for R
xx
, R
xy
, and R
yy
given in (7.16), (7.17), and (7.18).
( )
∫
∞
∂ =
0
3 2
) 0 (
. 4 R R f R R
xx
σ (7.16)
( ) ( )
∫ ∫
∞
∂ + ∂ =
clip
clip
r
clip
r
yy
R R f R R R R f R R . . 4 . 4
2 2
0
3 2
) 0 (
σ σ (7.17)
( ) ( )
∫ ∫
∞
∂ + ∂ =
clip
clip
R
clip
R
xy
R R f R R R R f R R . . 4 . 4
2 2
0
3 2
) 0 (
σ σ (7.18)
Chapter 7: Reduced Complexity Clipping Algorithms
148
where f(R) is the normalized Rayleigh probability distribution of the data, R.
( )
2
R
e R f
−
= (7.19)
The normalised clip level is also required (7.20)
( ) R dB
r
dB
clip
clip
10
10
log 20
2
log 20
=
=
σ
20
10
clip
dB
R =
(7.20)
The evaluated correlation expressions of (7.16), (7.17), and (7.18) are shown in
(7.21), (7.22), and (7.23). These are substituted into (7.10) to calculate the SNR at
various clipping levels.
( )
2
0
2
xx
R σ = (7.21)
( )
( )
2
0
0
1
clip
yy R
xx
R
e
R
−
= − + (7.22)
( )
( )
( )
2
0
0
1 .
2 2
clip
xy R
clip clip clip
xx
R
e R R erf R
R
π π
α
−
= = − + + − (7.23)
The SNR vs. clipping level for conventional clipping reduces to (7.24)
( )
( )
( )
( )
( )
2
2 2
2
2
0
0
1 .
2 2
1 1 .
2 2
clip
clip clip
R
clip clip clip
conv
xy R R
clip clip clip
xx
e R R erf R
SNR
R
e e R R erf R
R
π π
π π
α
−
− −
− + + −
=
− + − = = − + + −
(7.24)
A plot of the theoretical SNR vs. conventional clip level is shown in Figure 7.16.
Chapter 7: Reduced Complexity Clipping Algorithms
149
7.2.1.3 Sector Clipping
In sector clipping the reduction in hardware complexity comes at a cost of introducing
extra amplitude distortion and phase distortion as the components of the complex
signal are not attenuated by the same scaling factor. This leads to more complex
equations for R
xx
, R
xy
, and R
yy
as the data must be represented in terms of its
Cartesian coordinates.
A way to realize the transition from polar to Cartesian is to recognize that the joint
probability of two independent, zero mean, quadrature shifted, Gaussiandistributed
variables, x and y, with the same σ create a Rayleigh distribution (7.25).
2 2
2 2
2 2
1 1
( , )
2 2
x y
f x y e e
σ σ
σ π σ π
− −
= × (7.25)
2
2 2
2
2
2
1
) , (
σ
πσ
y x
e y x f
− −
=
3 sector clipping levels are defined relative to r
clip
(7.26a, b, c) of the conventional
clipping method. The clipping regions for 3 sector clipping are shown in Figure 7.13
showing the vectors of an unclipped signal and the resultant clipped sample.
0
sin
clip
l r θ = (7.26a)
1
2
clip
r
l = (7.26b)
2
cos
clip
l r θ = (7.26c)
Chapter 7: Reduced Complexity Clipping Algorithms
150
Figure 7.13: I Q diagram of 1
st
quadrant of a 3 Sector clipping system showing the vector of an
unclipped and clipped sample.
The new sector clipping correlation equations for R
xx
, R
yy
, and R
xy
are listed (7.27,
7.28, 7.29) respectively below. Due to symmetry, only the first quadrant is used
( ) ( )
2 2
(0)
0 0
. ,
xx
R x y f x y x y
∞ ∞
= + ∂ ∂
∫ ∫
(7.27)
The autocorrelation of the output equations are set out in (7.28) where the symmetry
of the clipping regions is exploited. The limits of the integrals of each part (7.28)
dictate what will happen at the output. Hence the first 2 parts of (7.28) do nothing to
the data as they are inside the clipping region. The last 3 parts of (7.28) (1, 2, and 3 in
Figure 7.13)) perform the attenuation as evidenced by the limits.
I
Q
l
0
l
1
l
2
l
0
l
1
l
2
1 2
3
1
2
Clipped
vector
Original
vector
Chapter 7: Reduced Complexity Clipping Algorithms
151
( ) ( )
( ) ( )
( )
( )
( )
( )
( )
( )
1 1
0 2
1
0
2
1
1 0
1 1
2 2
(0)
0 0
2 2
0
2 2
2
0
2 2
1
2 2
1 1
4 . ,
8 . ,
8 . ,
8 . ,
4 . ,
l l
yy
l l
l
l
l
l
l l
l l
R x y f x y x y
x y f x y x y
l y f x y x y
l y f x y x y
l l f x y x y
∞
∞
∞ ∞
= + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
∫ ∫
∫ ∫
∫ ∫
∫ ∫
∫ ∫
(7.28)
The cross correlation (7.29) follows the same form as (7.28) where the first 2 parts are
inside the clipping regions. The last 3 parts of (7.29) show the translation of the input
sample to the output. Again the symmetry of Figure 7.13 is exploited.
( ) ( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
1 1
0 2
1
0
2
1
1 0
1 1
2 2
(0)
0 0
2 2
0
2
2
0
2
1
1 1
4 . ,
8 . ,
8 . ,
8 . ,
4 . ,
l l
xy
l l
l
l
l
l
l l
l l
R x y f x y x y
x y f x y x y
xl y f x y x y
xl y f x y x y
xl yl f x y x y
∞
∞
∞ ∞
= + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
∫ ∫
∫ ∫
∫ ∫
∫ ∫
∫ ∫
(7.29)
As in the normalized conventional clipping method, substitutions are made to remove
σ and reduce the complexity of the correlation functions.
2 σ
x
X = (7.30a)
2 σ
y
Y = (7.30b)
2 σ
x
X
∂
= ∂ (7.30c)
2 σ
y
Y
∂
= ∂ (7.30d)
Chapter 7: Reduced Complexity Clipping Algorithms
152
2
0
0
σ
l
L = (7.30e)
2
1
1
σ
l
L = (7.30f)
2
2
2
σ
l
L = (7.30g)
Substituting (7.30) into (7.27), (7.28), and (7.29) yields the simplified expressions for
the 3 sector clipping correlation functions. The normalized equation for R
xx(0)
is
( )
( ) ( )
2 2 2
0
2 . ,
xx
R X Y f X Y X Y σ
∞ ∞
−∞ −∞
= + ∂ ∂
∫ ∫
(7.31)
The normalized equation for α is
( )
( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
1 1
2 0
1
0
2
1
1 0
1 1
0
2 2
0 0
0
2 2
0
2
2
0
2
1
1 1
4 . ,
8 . ,
8 . ,
8 . ,
4 . ,
L L
xy
xx
L L
L
L
L
L
L L
L L
R
X Y f X Y X Y
R
X Y f X Y X Y
XL Y f X Y X Y
XL Y f X Y X Y
XL YL f X Y X Y
α
∞
∞
∞ ∞
= = + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
∫ ∫
∫ ∫
∫ ∫
∫ ∫
∫ ∫
(7.32)
The normalized equation for R
yy(0)
is
( )
( )
( ) ( )
( ) ( )
( ) ( )
( ) ( )
( )
1 1
2 0
1
0
2
1
1 0
1 1
0
2 2
0 0
0
2 2
0
2 2
2
0
2 2
1
2
1
4 . ,
8 . ,
8 . ,
8 . ,
8 . ,
L L
yy
xx
L L
L
L
L
L
L L
L L
R
X Y f X Y X Y
R
X Y f X Y X Y
L Y f X Y X Y
L Y f X Y X Y
L f X Y X Y
∞
∞
∞ ∞
= + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
+ + ∂ ∂
+ ∂ ∂
∫ ∫
∫ ∫
∫ ∫
∫ ∫
∫ ∫
(7.33)
Where f(X,Y) is given by (7.34)
( )
2 2
,
X Y
e
f X Y
π
− −
= (7.34)
Chapter 7: Reduced Complexity Clipping Algorithms
153
The normalized clipping levels are
( ) R dB
r
dB
clip
clip
10
10
log 20
2
log 20
=

¹

\

=
σ
20
10
clip
dB
R = (7.35)
The evaluated correlation expressions of (7.31), (7.32), and (7.33) are shown in
(7.36), (7.38), and (7.40). These are substituted into (7.10) to calculate the SNR at
various clipping levels.
( )
2
0
2
xx
R σ = (7.36)
α was calculated to be (7.37)
( )
( )
( )
( )
( ) ( ) ( )
( )
( ) ( )
( )
( ) ( ) ( ) ( )
2
1
2
0
2 2
0 2
2 2
0 1
0
1
1
1
0
0 0
2 1
0
0 2
0 2
1 0 1 1 0
2
2
2 .
2
2
2 1
2
2
1 .
2 2
xy
L
xx
L
L L
L L
R
erf L
L
erf L e
R
erf L L
erf L erf L e
erf L
L L
e erf L erf L e
erf L L e L e erf L erf L
α
π
π
π π
π π
π
−
−
− −
− −
1  
−
= = +
1 
1
\ ¹ ¸ ]
1   −
+ − +
1 
1
\ ¹ ¸ ]
1 −
+ + − + 1
1
¸ ]
¸ ]
1  
+ − − + −
1 

1
\ ¹ ¸ ]
+ ( ) ( ) ( ) ( ) ( )
2 2
1 1 1
1 1 0 1
2 2
1
L L
L
L e erf L erf L e erf L
π π
− −
1
− + −
¸ ]
(7.37)
Which after expansion and reduction becomes
( )
( )
( ) ( ) ( ) ( ) ( )
0
2 0 1 0 1
0
xy
xx
R
erf L erf L erf L erf L erf L
R
α = = − + (7.38)
Chapter 7: Reduced Complexity Clipping Algorithms
154
The normalized autocorrelation of the output,
(0)
(0)
yy
xx
R
R
was derived to be (7.39)
( )
( )
( )
( )
( ) ( ) ( )
( )
( )
( ) ( )
( ) ( ) ( ) ( ) ( )
( )
( ) ( )
( )
2
1
2
0
2 2
2 1
2
0
2
1
0
1 1
1
0
0
0
2 1
2 1
2 1
0
2 0 0
2 0 2 2
1 1
1
2
2
2
2
2
2 2
2 1 2 1
2
2 1
2
yy
L
xx
L
L L
L
L
R
erf L L
erf L e
R
erf L
L
erf L erf L e
erf L erf L
L L
erf L e e
erf L L
L erf L erf L erf L e
erf L L L
erf L e
π
π
π π
π
π
−
−
− −
−
−
  −
= +

\ ¹
1  
−
+ − +
1 
1
\ ¹ ¸ ]
1
−
+ + + −
1
¸ ]
  −
+ − + − +

\ ¹
−
+ − + +
( )
( ) ( ) ( ) ( ) ( ) ( ) ( )
2
0
0 0
2
2 2
1 1 1 0 1 1
2
2 1 2 1
L
erf L
e
L erf L erf L erf L L erf L
π
−
1  
−
1 
1
\ ¹ ¸ ]
1
1 + − − + −
¸ ]
¸ ]
(7.39)
Which after expansion and reduction becomes (7.40)
( )
( )
( )
( ) ( ) ( )
( ) ( ) ( )
( ) ( ) ( )
2 2
2 1
2
1
0
2 1
0
0
0 2 1
2
1
2 0 2 1
2 2
1 1 0 1 0
2
2
2 1
2 1 2 1
yy
L L
xx
L
R
L L
erf L e e
R
erf L erf L erf L
L
L erf L erf L e erf L
L erf L erf L L erf L
π π
π
− −
−
−
= +
+ −
+ − − +
+ − + −
(7.40)
The closed form solution is shown in (7.41).
Chapter 7: Reduced Complexity Clipping Algorithms
155
( ) ( ) ( ) ( ) ( ) ( )
( )
( ) ( ) ( )
( ) ( )
( )
( ) ( )
( )
( )
2 2
2 1
2
1
2
2 0 1 0 1
3
2 1
0
0 2 1
2 2
2 0 2
1
1
2
1 1 0
2
1 0
2
2 1
2
2 1
2 1
Sec
L L
L
erf L erf L erf L erf L erf L
SNR
L L
erf L e e
erf L erf L erf L
erf L erf
L erf L erf L
L
e erf L
L erf L erf L
L erf L
π π
π
− −
−
− +
=
  − 1
+

1
¸ ]


+ − 1
¸ ]


+ − 1
¸ ]

−

− +



+ − 1
¸ ]



+ − 1
¸ ] \ ¹
( )
( ) ( )
( )
2
0
1 0
1
L
erf L erf L
erf L
 

−


+
\ ¹
(7.41)
A block diagram showing 3 Sector clipping implemented with a LUT is shown in
Figure 7.14. Note that the sign bits are removed in I
in
and Q
in
and attached back on at
I
out
and Q
out
. A flowchart detailing the decision matrix in the algorithm is shown in
Figure 7.15. From this the LUT in Figure 7.14 is derived and shown in Table 7.1,
where it is seen that the LUT requires a 8 bit input and a (N×2)+2 bit output. A plot
of the theoretical SNR vs. 3 Sector clipping is shown in Figure 7.16. The latency of
the structure in Figure 7.14 is low as the level comparisons are made in parallel and
fed into the LUT. Further more when reviewing the truth table in Table 7.1 the inputs
for ‘0’ and L2 can replaced with logic 0 and 1 respectively reducing the number of
inputs to the LUT to 4 and by implication the size of the LUT.
Table 7.1: Truth Table for 3 Sector clipping
Input Output
I
in
Q
in
I
clip
Q
clip
Ctrl
I
Ctrl
Q
0 L
0
L
1
L
2
0 L
0
L
1
L
2
X X 1 X X X 1 X L
1
L
1
1 1
X X X 1 1 0 X X L
2
Q
in
1 0
X X 1 X 1 1 0 X L
1
Q
in
1 0
1 0 X X X X X 1 I
in
L
2
0 1
1 1 0 X X X 1 X I
in
L
1
0 1
For all other combinations data is passed through I
in
Q
in
0 0
Chapter 7: Reduced Complexity Clipping Algorithms
156
Figure 7.14: Block diagram of 3 Sector clipping.
LUT
L
0
L
1
L
2
0
L
2
L
0
L
1
0
N
N
I
clip
Control I
Control Q
N
N
I
in
Q
in
I
out
Q
out
Q
clip
N
N
Chapter 7: Reduced Complexity Clipping Algorithms
157
Figure 7.15: Flowchart for the LUT in Figure 7.14 (3 Sector clipping).
Abs(Q)≥0 Abs(Q)≤L
0
Abs(I)≥L
2
L
2
+j×abs(Q)
Abs(Q)>L
0
Abs(Q) ≤L
1
Abs(I)≥L
1
L
1
+j×abs(Q)
Abs(I)≥L
1
Abs(Q)≥L
1
Abs(I)≥0 Abs(I)≤L
0
Abs(Q)≥L
2
Abs(I)+j×L
2
Store
sign bits
I+j×Q
Abs(I)>L
0
Abs(I)≤L
1
Abs(Q)≥L
1
Abs(I)+j×L
1
Abs(I)+j×abs(Q)
Reattach
sign bits
L
1
+j×L
1
yes
yes
yes
yes
yes
yes
yes yes
yes
yes
yes
yes
yes
yes
no
no
no
no
no
no
no
no
no
no
no
no
no
no
Chapter 7: Reduced Complexity Clipping Algorithms
158
7.2.1.4 Square Clipping
Square clipping is the least complex form of clipping and is shown in the following
theoretical results as a lower bound on the presented clipping techniques. The
theoretical SNR for Square clipping can easily be found by using the same method as
for 3 Sector clipping and changing the limits on (7.32) and (7.33) where appropriate.
That is by setting both I
0
and I
2
equal to I
1
in Figure 7.13. The equations are simpler
to derive and can be construed as 2 sector clipping which is in fact Cartesian clipping.
7.2.1.5 Theoretical Results
The theoretical results for Conventional, 3 Sector, and Square clipping SNR vs. clip
level are shown in Figure 7.16 together with the equivalent simulated results. Note
that for the simulated results no filtering, oversampling, or amplification is performed,
the noise is measured after clipping. As expected Conventional clipping has the best
performance and Square clipping the worst. 3 Sector clipping has performance in the
middle, but with greatly reduced complexity requiring only a few comparators and a
simple LUT, making it only marginally more complex than square clipping.
Generally 3 Sector clipping suffers a 1dB penalty in SNR compared to conventional
clipping levels above 0dB. In other words in order for 3 sector clipping to achieve the
same SNR as Conventional clipping the clip level needs to be set 1dB higher than the
Conventional case. Also of note it is seen that sector clipping performance
approaches the same SNR as Square clipping at clip levels below 0dB. This makes
sense as harder clip levels reduce the size of regions 1 and 2 in Figure 7.13, making
the corner sector the dominant clip region, just as in Square clipping. Therefore 3
Sector and Square clipping share the same lower SNR bound of 2.439dB.
Conventional clipping has a lower SNR bound of 5.634dB.
Chapter 7: Reduced Complexity Clipping Algorithms
159
10 8 6 4 2 0 2 4 6 8 10
0
10
20
30
40
50
60
70
Clip Level (dB)
S
N
R
(
d
B
)
Theoretical
Simulated
Conventional
3 Sector
Square
SNR
l i mi t
=5.64dB
SNR
l i mi t
=2.44dB
Figure 7.16: Theoretical Clip Level vs. SNR for Conventional (Standard), 3 Sector, and Square
Clipping. Theoretical (dashed), simulated (solid).
Comparing the simulated curves with the theoretical ones it is seen that they are well
matched at clip levels below 6dB. At higher levels of clipping the simulated results
have
slightly better SNR, this is due to the theoretical Rayleigh distribution assumption of
the signal not holding at higher amplitude levels [89]. Simulated OFDM symbols
have lower PDF values in the tail of the distribution.
Figure 7.17 plots the theoretical 3 sector clipping angle, θ (shown in Figure 7.11) vs.
SNR for clip levels ranging from 10dB (at the bottom of Figure 7.16) to 10dB (at the
top of Figure 7.16). The angle θ is varied from 20° to 40°. The choice of θ has no
impact on harder clipping levels, but it does affect SNR performance at weaker levels
(CL>5dB). It is seen that the optimum angle for θ is 27.5°, i.e. ( )
0
2
tan 27.5
L
L
=
,
for all practical clipping levels as shown in Figure 7.17.
Chapter 7: Reduced Complexity Clipping Algorithms
160
20 22 24 26 28 30 32 34 36 38 40
0
5
10
15
20
25
30
35
40
45
50
θ (°)
S
N
R
(
d
B
)
CL=10dB
CL=10dB
Figure 7.17: Clipping angle, θ (of Figure 7.11) vs. the SNR for 3 sector clipping, based on (7.41).
7.2.2 Extensions of Sector Clipping
3 Sector clipping can be extended to incorporate more sectors thereby improving
performance. Figure 7.18 shows the I Q plot of the first quadrant of 4 and 5 Sector
clipping. Figure 7.19 depicts simulated results for Sector Clipping with 3, 4, and 5
sectors as well as Conventional and Square clipping for comparison. Increasing the
numbers of sectors from 3 to 4 improves the SNR by ~2dB at a clipping level of 6dB.
5 sectors provides a further 0.5dB gain in SNR at the same clipping level. Generally,
the SNR difference between the schemes increases with clipping level. The
complexity increase for 4 and 5 sectors is minimal over 3 sector clipping requiring
only a few extra comparators and a doubling of the LUT size.
Chapter 7: Reduced Complexity Clipping Algorithms
161
(a) (b)
Figure 7.18: I Q diagram showing the 1
st
quadrant Sector clipping regions of a) 4 Sector clipping and
b) 5 sector clipping.
The degradation in SNR for Sector clipping from optimal Conventional clipping is
~3dB and the improvement in SNR over square clipping is ~4dB at a clip level of
6dB. Results not shown here indicate that increasing the number of sectors above 5
shows no discernable improvement in the SNR. The choice of angles for 4 and 5
sector clipping has been found to be optimum when the angles which determine the
sectors are equally spaced, i.e. for 4 sectors: θ=28.13° and β=39.75°; and 5 sectors:
θ=11.25°, β=22.5°, and α=33.75°.
This section detailed a new method for clipping OFDM symbols called Sector
Clipping. The new method is very simple to implement requiring only a few extra
comparators compared to square clipping, however its performance is closer to
conventional clipping where traditionally more complex circuits are required.
Chapter 7: Reduced Complexity Clipping Algorithms
162
10 8 6 4 2 0 2 4 6 8 10
0
10
20
30
40
50
60
70
Clip Level (dB)
S
N
R
(
d
B
)
Conventional
5 Sector
4 Sector
3 Sector
Square
Figure 7.19: Simulated clip level vs. SNR for Conventional, Square, 3, 4, and 5 Sector Clipping.
7.2.3 Hardware Implementation
3 Sector and Square clipping was implemented in digital form via Visual Hardware
Design Language (VHDL). The two algorithms were then simulated using
‘Synopsys’ to ensure proper operation. Next the algorithms were compiled into
‘Verilog’ (VHDL code expressed in terms of gates, flip flops, etc.) code and
simulated again in ‘Synopsys’. The ‘Verilog’ code was then exported into the
‘Cadence’ silicon design package, ‘Silicon Ensemble’, where routing and cell
placement was performed with a 0.5µ standard cell and port library. After routing a
‘gds2’ file was produced which could then be sent to the foundry for production.
However, this was not done, therefore the algorithm was only proved through
hardware simulation. The design flow is summarized in Figure 7.20.
Figure 7.21 depicts the layout of the 3 sector clipping algorithm with N=8 bit inputs.
A gain table was included to allow 16 different Clipping levels (CL) set from 3dB to
12dB.
Chapter 7: Reduced Complexity Clipping Algorithms
163
Figure 7.20: Design flow for silicon implementation of 3 sector clipping.
Figure 7.21: Block diagram of 3 Sector clipping implemented in VHDL.
VHDL
Simulate
(Synopsys)
Verilog
Silicon
Ensemble
(Cadence)
Gds2 file
Foundry
Cmos 0.5µ
standard cell
library
testing
Standard port
library
8
8
7
7
7
8
8
4
I shift reg.
Q shift reg.
Input data
I
serial in
erial
Q
serial in
erial
I
in
eri
Q
in
eri
Q
in
eri
Q
out
erial
I
out
eri
CL
eri
Sector
Clipper
(Fig 7.14)
Gain table
Chapter 7: Reduced Complexity Clipping Algorithms
164
Due to the clipping operation the Most Significant Bit (MSB) on each of the I and Q
channels can be removed at the output. In order to save the number of input/output
pins required, the I and Q input data is fed serially (I
in serial
and Q
in serial
) into shift
registers which then feed 8 bit words (I
in
and Q
in
) into the 3 Sector clipping algorithm.
The input data word is stored and output in parallel fashion for comparison to the
clipped sample.
Table 7.2 shows a summary of various ‘Synopsys’ reports for the power consumption,
delay path, and cell area of Square and 3 Sector clipping. ‘System with buffer’ refers
to the whole system shown in Figure 7.21 with the input and output buffers included.
The buffers are necessary when the system is implemented as a stand alone device in
silicon. In practice however the buffers are not necessary as only the clipping
algorithm is implemented as a block in the whole OFDM transmitter.
Table 7.2: ‘Synopsys’ reports for Square and 3 sector clipping. (Refer to Figure 7.21)
Hierarchy level Report Square clipping 3 sector clipping
System with buffer
Total cell area 13593.52 dbu 13827.55 dbu
Total dynamic power 23.268mW 35.3874mW
System
Total cell area 357.16 dbu 591.1800 dbu
Total dynamic power 11.5896mW 17.5455mW
Sector Clipper
Total cell area 110.37 dbu 385.38 dbu
Total dynamic power 6.8317mW 8.7933mW
Gain table
Total cell area 62.70 dbu 131.08 dbu
Total dynamic power 4.4682mW 8.3737mW
I and Q shift
registers
Total cell area 37.36 dbu 37.36 dbu
Total dynamic power 0.1893mW 0.1893mW
‘System’ again refers to Figure 7.21 without the input and output buffers. The cell
area is greatly reduced in this case as the buffers take up a lot of space. The algorithm
only takes up 4.27% of the total cell area for 3 Sector clipping and only 2.62% for
Chapter 7: Reduced Complexity Clipping Algorithms
165
Square clipping. The clipping algorithms themselves (‘Sector Clipper’ in Table 7.2)
have an area of 385.38 dbu and 110.37 dbu for 3 Sector and Square clipping
respectively, the power consumption is 8.7933mW and 6.8317mW for 3 Sector and
Square clipping respectively. The 3 Sector Gain Table requires 50% more cells than
Square clipping and the power consumption is doubled due to the extra clip levels
required for sector clipping. The shift registers are the same in both systems.
Table 7.3 and 7.4 show the ‘Cadence’ reports for Square and 3 Sector clipping
respectively. A 0.5µ process with 3 metal layers was used. As is seen in the
Application Specific Integrated Circuit (ASIC) view of 3 Sector clipping Figure 7.22
the size of the chip is determined by the input/output pads which are abutted to make
the area as small as possible. This made the wire routing process easy as there was a
lot of room to work with. The area of utilization (occupied chip area) is 74.78% in
Square clipping and 75.62% for 3 Sector clipping, a small difference. Comparing the
number of ‘CORE Rows’ and ‘CORE Cells’ for Square and 3 Sector clipping
algorithms it is seen that Square clipping requires 22 rows and 148 cells while 3
Sector clipping requires 26 rows and 319 cells.
Table 7.3: ‘Cadence’ area utilization report on Square clipping
Type Number Length Area
%_Row_Space
CORE Rows 22 1205160 3012900000
CORE Cells 148 237380 593450000
19.70
CORNERSITE_495 Rows 4 198000 9801000000
CORNERSITE_495 Cells 4 198000 9801000000
100.00
IOPADSITE_495 Rows 4 420000 20790000000
IOPADSITE_495 Cells 28 420000 20790000000
100.00
Area of chip: 41699680000 (square DBU)
Area required for all cells: 31184450000 (square DBU)
Area utilization of all cells: 74.78%
Chapter 7: Reduced Complexity Clipping Algorithms
166
Table 7.4: ‘Cadence’ area utilization report on 3 Sector clipping
Type Number Length Area
%_Row_Space
CORE Rows 26 1693120 4232800000
CORE Cells 319 390060 975150000
23.04
CORNERSITE_495 Rows 4 198000 9801000000
CORNERSITE_495 Cells 4 198000 9801000000
100.00
IOPADSITE_495 Rows 4 420000 20790000000
IOPADSITE_495 Cells 28 420000 20790000000
100.00
Area of chip: 41744615000 (square DBU)
Area required for all cells: 31566150000 (square DBU)
Area utilization of all cells: 75.62%
Figure 7.22: Application Specific Integrated Circuit (ASIC) view of 3sector clipping algorithm
implemented in Silicon using ‘Cadence Silicon Ensemble’. 0.5µ process, 3 metal layers.
Chapter 7: Reduced Complexity Clipping Algorithms
167
7.3 New Vector Subtraction clipping method
This section describes another new method called Vector Subtraction which is an
enhancement of an existing algorithm [101] called the Lucent algorithm here, to clip
samples. Vector Subtraction reduces complexity of the Lucent algorithm by removing
the need for divisions, which are complex operations in hardware and add
significantly to the complexity of the original algorithm.
7.3.1 Lucent Algorithm
The Lucent algorithm produces good estimates of the magnitude in K iterations. The
Lucent algorithm works as follows, first the complex sample,
i q
x x jx = + is folded
into the first octant to give
( ) ( )
'
max , . min ,
i q i q
x x x j x x = + (7.42)
x’ is then rotated by a number, K, of fixed phase angles, θ
k
, towards the real axis. The
phase angle which returns the largest real part gives the closest approximation to the
direction (phase) of the vector x’, and the magnitude of the real part, ˆ x , is the closest
approximation to the actual magnitude. The K phase values are spaced in the octant
as
16
k
k
l
K
π
θ = (7.43)
where { } 1, 3, 5,..., 2 1
k
l K = − . For all K values of θ
k
the magnitude estimate is given
by
( ) ( )
'
ˆ max Re .
k
k
j
x x e
θ
θ
−
= (7.44)
Chapter 7: Reduced Complexity Clipping Algorithms
168
After the magnitude has been estimated, any samples exceeding the clipping
threshold,
clip
x , are multiplied by a scaling factor (which reintroduces complexity
that was mitigated by the iterative magnitude estimator), the clipped output signal is
given by
( )
ˆ
clip
out
x x x
x
x
=
ˆ
ˆ
clip
clip
x x
x x
>
≤
(7.45)
The only error in this technique is a slight under estimate of ˆ x which reduces as K
increases. The scheme has many similarities to the CORDIC [102] method, but gives
better estimates of the magnitude at low values of K. The scaling operation requires a
division which is a complex hardware operation and can be avoided by using the new
Vector Subtraction technique described below.
7.3.2 Vector Subtraction
The process of finding the magnitude estimate, ˆ x , is the same as [101] described in
(7.44), but the scaling operation is replaced by a subtraction. First, the overshoot, o
s
,
is calculated
ˆ
s clip
o x x = − (7.46)
and then subtracted from the signal, x’. However the phase of the main signal is not
known therefore the best estimate of the phase is used, this is θ
k,max.
The overshoot, o
s,
is rotated by
,max k
j
e
θ
giving the correction vectors y
i
and y
q
.
, k m
j
i q s
y y jy o e
θ
= + = (7.46a)
The correction vectors are octant adjusted prior to subtraction from the original signal,
x.
Chapter 7: Reduced Complexity Clipping Algorithms
169
( )
q i i i
y y x x ,
' '
± =
(7.47)
( )
i q q q
y y x x ,
' '
± =
(7.48)
The addition/subtraction of y
i
and y
q
can be extrapolated from the maximum and
minimum values of x
i
, x
q
and the sign bits of the original data. A block diagram of
Vector Subtraction is shown in Figure 7.23.
Figure 7.23: Block diagram of new Vector Subtraction scaling operation.
Figure 7.24 shows the I Q diagram for Vector Subtraction with K=2 iterations
corresponding to phases of 16 π and 3 16 π . The vector x’ is closest to the
16 π vector. Note that the clipped value, x’’, has an additional phase error compared
to the Lucent method. It has both amplitude and phase error compared to the ideal
clipped value.
Figure 7.25 and 7.26 show the CCDF of the Lucent algorithm [101] and the new
Vector Subtraction variation respectively where 1000 OFDM symbols (N=64) are
clipped at 5dB. Here it is seen that a slight underestimate in the magnitude means that
samples are not always clipped back to the desired level. This is a function of the
number of iterations in the algorithm with 1 iteration underestimating the magnitude
by 0.7dB and 1.2dB to pass through in the Lucent algorithm and Vector Subtraction
respectively. This problem is exacerbated in Vector Subtraction where the new
scaling operation magnifies the error in the magnitude estimate.
y
i
y
q
i
x
q
x
xˆ ˆ
clip
x x >
ˆ
s clip
o x x = −
( ) 1..k
j
e
θ
(,+)
' '
i
x
(,+)
' '
q
x
( )
,max
Re
j
k
e
θ
( )
,max
Im
j
k
e
θ
Chapter 7: Reduced Complexity Clipping Algorithms
170
Figure 7.24: IQ plane for the new Vector Subtraction method showing vector, x’ being clipped to x”.
The amount of error in the magnitude estimate increases with the decrease in clipping
level for Vector Subtraction with clipping under 3dB requiring a prohibitive amount
of backoff. This is not an issue with the Lucent patent where the magnitude estimate
error is constant irrespective of clip level. The upshot of this is that the clip level will
need to be backed off to avoid saturation of the amplifier. Alternatively more than 1
iteration can be used making the error in the magnitude estimate small. This small
error can normally be neglected as filtering will cause substantial peak regrowth
anyway.
I
Q
y
i
y
q
θ
k,max
x’
x’’
Chapter 7: Reduced Complexity Clipping Algorithms
171
4 5 6 7 8 9 10 11
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
Unclipped
1 iteration
2 iterations
4 iterations
3 iterations
Figure 7.25: Simulated CCDF for Lucent patent [101] for various iterations clipped at 5dB showing the
leakage of under clipped samples.
4 5 6 7 8 9 10 11
10
3
10
2
10
1
10
0
ζ
0
(dB)
P
r
(
ζ
>
ζ
0
)
1 iteration
4 iterations
3 iterations
2 iterations
Unclipped
Figure 7.26: Simulated CCDF for Vector Subtraction for various iterations clipped at 5dB showing the leakage
of under clipped samples.
Chapter 7: Reduced Complexity Clipping Algorithms
172
The simulated clip level vs. SNR results are shown in Figure 7.27 and 7.28 for the
Lucent algorithm and Vector Subtraction method respectively. Note that the clip level
is adjusted depending on the iteration and clipping method according to the CCDF
results of Figures 7.25 and 7.26, for example the Lucent patent with 1 iteration
requires the clip level to be set to CL0.7. This ensures that the samples will not
saturate the HPA if it were present.
When 2, 3, or 4 iterations are used, both the Lucent patent and Vector subtraction
have very similar performance to the conventional clipping method. 1 iteration
results in a 5dB degradation from the optimum clipping method for the Lucent
algorithm while 1 iteration in Vector Subtraction results in a more serious degradation
across the board, making it impractical; In fact its performance goes below that of
Square clipping for clip levels below 4dB. The reason for this is the increasing
amount of back off required at harder clip levels to avoid saturation of the amplifier.
0 1 2 3 4 5 6 7 8 9 10
0
10
20
30
40
50
60
70
Clip Level (dB)
S
N
R
(
d
B
)
Conventional clipping
Lucent  4 iterations
Lucent  3 iterations
Lucent  2 iterations
Lucent  1 iteration
Square clipping
Figure 7.27: Simulated clip level vs. SNR for Lucent clipping technique with varying iterations, as
well as Conventional and Square clipping.
Chapter 7: Reduced Complexity Clipping Algorithms
173
0 1 2 3 4 5 6 7 8 9 10
0
10
20
30
40
50
60
70
Clip Level (dB)
S
N
R
(
d
B
)
Conventional Clipping
Vector Subtraction  4 iterations
Vector Subtraction  3 iterations
Vector Subtraction  2 iterations
Vector Subtraction  1 iteration
Square Clipping
Figure 7.28: Simulated clip level vs. SNR for Vector Subtraction clipping with varying iterations, as
well as Conventional and Square clipping.
7.4 Comparison of new and existing clipping methods
The previous sections described new low complexity clipping algorithms and
evaluated them both theoretically and through simulation. Section 7.4.1 analyses their
performance in the OFDM system described in Section 7.1. Section 7.4.2 compares
them in terms of their complexity by comparing hardware operations. The mapping
type used in the following simulations is 64 QAM, the number of taps in the RRCF is
128 with a roll off factor of 0.15, the oversampling rate in the IFFT is set to 2. A LPA
is used initially and later simulations use a HPA (P=3) with increasing backoffs.
Conventional and Square clipping are also shown as a reference.
7.4.1 BERF
Figure 7.29 shows the baseband clipping vs. BERF for 3, 4, and 5 Sector clipping
where it is seen that increasing the number of sectors to 4 provides an improvement of
1 and a half magnitudes over Square clipping at 4dB clipping. Increasing the number
of sectors to 5 provides a more modest decrease in the BERF beyond 4 Sector
Chapter 7: Reduced Complexity Clipping Algorithms
174
clipping at 4 dB clipping. 5 sector clipping is 1 magnitude worse than conventional
clipping at 4 dB clipping. At 5 dB clipping Conventional, 4, and 5 Sector clipping
have a BERF below 10
6
. Note that Sector clipping over clips the data therefore
performance after a HPA with a limited backoff may be better due to the extra
regrowth allowed. This will be explored further later in this section.
0 1 2 3 4 5 6 7 8
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
Conventional
5 sectors
4 sectors
3 sectors
Square
Figure 7.29: Simulated 3, 4, and 5 Sector, Conventional and Square clipping vs. BERF with a LPA. 64
QAM symbols, 128 point IFFT (os=2), RRCF with 128 taps, and α=0.15. AWGN=0.
Figure 7.30 shows baseband clipping vs. BERF for Vector Subtraction with 1, 2, 3,
and 4 iterations. Performance is nearly the same for all methods, however curiously 1
iteration has the best performance while conventional clipping has the worst. This
can be explained by reviewing Figure 7.25 where it seen that Vector Subtraction
under clips some samples which would lead to saturation of the amplifier had it not
been a LPA.
Chapter 7: Reduced Complexity Clipping Algorithms
175
0 1 2 3 4 5 6 7 8
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
Conventional
4 iterations
3 iterations
2 iterations
1 iteration
Square
Figure 7.30: Simulated Vector Subtraction (1, 2, 3, and 4 iterations), Conventional and Square clipping
vs. BERF with a LPA. 64 QAM symbols, 128 point IFFT (os=2), RRCF with 128 taps, and α=0.15.
AWGN=0.
Figure 7.31 shows baseband clipping vs. BERF for the Lucent algorithm [101] where
it seen that the performance of all iterations is almost the same as Conventional
clipping. Again 1 iteration has slightly better performance than the other methods due
to the under clipping of some samples as seen in Figure 7.25. Vector Subtraction has
an almost identical performance to the Lucent method except for the trivial case of 1
iteration.
Chapter 7: Reduced Complexity Clipping Algorithms
176
0 1 2 3 4 5 6 7 8
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
Conventional
4 iterations
3 iterations
2 iterations
1 iteration
Square
Figure 7.31: Simulated Lucent [101] clipping (1, 2, 3, and 4 iterations), Conventional and Square
clipping vs. BERF with a LPA. 64 QAM symbols, 128 point IFFT (os=2), RRCF with 128 taps, and
α=0.15. AWGN=0.
This section simulated the performance of the new and existing clipping algorithms in
an OFDM environment with a LPA. The rest of this section uses a HPA with
different IBO’s relative to the baseband clipping level. Note that Sector Clipping
overclips the signal in some instances while Vector Subtraction and the Lucent
algorithm under clip the signal which explains the higher BERF’s of the Sector
clipping when the HPA is taken into account.
Changing HPA backoff
Figures 7.32, 7.33, and 7.34 show baseband clipping vs. BERF for 3, 4, and 5 Sector
clipping, respectively, with increasing backoff in the HPA. Increasing the IBO in the
HPA improves the BERF in all cases.
Chapter 7: Reduced Complexity Clipping Algorithms
177
0 1 2 3 4 5 6 7 8
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
LPA
IBO=CL+2
IBO=CL+1
IBO=CL
0.8dB
Figure 7.32: Simulated 3 Sector clipping, vs. BERF with varying IBO in HPA. 64 QAM symbols, 128
point IFFT (os=2), RRCF with 128 taps and α=0.15. AWGN=0.
0 1 2 3 4 5 6 7 8
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
LPA
IBO=CL+2
IBO=CL+1
IBO=CL
0.8dB
Figure 7.33: Simulated 4 Sector clipping, vs. BERF with varying IBO in HPA. 64 QAM symbols, 128
point IFFT (os=2), RRCF with 128 taps and α=0.15. AWGN=0.
Chapter 7: Reduced Complexity Clipping Algorithms
178
0 1 2 3 4 5 6 7 8
10
6
10
5
10
4
10
3
10
2
10
1
10
0
LPA
IBO=CL+2
IBO=CL+1
IBO=CL
0.8dB
Clip level (dB)
B
E
R
Figure 7.34: Simulated 5 Sector clipping, vs. BERF with varying IBO in HPA. 64 QAM symbols, 128
point IFFT (os=2), RRCF with 128 taps and α=0.15. AWGN=0.
Comparing Figures 7.32 and 7.34, 5 Sector clipping is just under 1 magnitude better
in terms of the BERF than 3 Sector clipping at 5dB baseband clipping with an IBO of
2dB above the baseband clip level (i.e. IBO=7dB). At an IBO=CL+2 the BERF is
within 1 magnitude of the LPA at 5dB clipping for all methods and within half a
magnitude at 4dB clipping. 1 dB of additional baseband backoff is required for 3 and
4 Sector and slightly more for 5 Sector clipping to maintain the same BERF as an
LPA at a BERF=10
4
with no additional IBO in the HPA.
The nonlinear characteristics of a HPA adds another source of distortion to the
transmitted signal. Figures 7.32 to 7.34 show that this additional distortion is
equivalent to a 0.8dB change in the CL (at BERF=10
4
) when the HPA saturation
level is equal to the
Chapter 7: Reduced Complexity Clipping Algorithms
179
0 1 2 3 4 5 6 7 8
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
LPA
IBO=CL+2
IBO=CL+1
IBO=CL
1.5dB
Figure 7.35: Simulated Vector Subtraction (1 iteration), vs. BERF with varying IBO in HPA. 64 QAM
symbols, 128 point IFFT (os=2), RRCF with 128 taps and α=0.15. AWGN=0.
0 1 2 3 4 5 6 7 8
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
LPA
IBO=CL+2
IBO=CL+1
IBO=CL
1.2dB
Figure 7.36: Simulated Vector Subtraction (4 iterations), vs. BERF with varying IBO in HPA. 64
QAM symbols, 128 point IFFT (os=2), RRCF with 128 taps and α=0.15. AWGN=0.
clipping level (IBO=CL). Obviously, as the amplifier is backed off, IBO=CL+2, then
the additional distortion is reduced.
Chapter 7: Reduced Complexity Clipping Algorithms
180
Figures 7.35 and 7.36 show baseband clipping vs. BERF for Vector Subtraction with
1 and 4 iterations, with increasing backoff in the HPA. Increasing the IBO improves
the BER in all cases. There is little difference between 1 and 4 iterations when
HPA=CL but this increases with the larger IBO in the HPA. However 1 iteration out
performs 4 iterations for all HPA backoffs. This is due to the underestimation of the
magnitude reducing the clipping distortion. At an amplifier backoff of HPA=CL+2
the BERF is within 1 magnitude of LPA performance at 5dB clipping for all methods
and within half a magnitude at 4dB clip level. When no extra backoff is allowed in
the HPA above the baseband clip level the performance of 1 and 4 iterations is almost
the same, with an additional backoff of 2dB in the HPA (HPA=CL+2) 1 iteration
outperforms 4 iterations by 0.2 dB at a BERF=10
4
.
Figures 7.37 and 7.38 show baseband clipping vs. BERF for the Lucent patent with 1
and 4 iterations respectively, and increasing backoff in the HPA. As with the 2
previous methods increasing the IBO improves the BERF. In Figure 7.37 (1 iteration)
at a baseband clip level of 6dB there is a difference of 1.5 magnitudes between
HPA=CL and a LPA. Under the same conditions the difference is under 2
magnitudes for Figure 7.38 (4 iterations). When no extra backoff is allowed in the
HPA both 1 and 4 iterations have approximately the same performance. The
performance of 4 iterations is slightly worse for IBO=CL+1 and IBO=CL+2.
For comparison Square clipping with different backoffs in the HPA is plotted in
Figure 7.39. Here it is seen that like Sector clipping Square clipping is more robust
against a reduction in the amplifier backoff with only half a magnitude in difference
between a LPA and a HPA with no additional IBO (IBO=CL).
Comparing the 3 methods it is interesting to note that Sector clipping has the least
increase in BERF from IBO=CL to LPA with a difference of 1.2, 1.5, and 1.5
magnitudes
Chapter 7: Reduced Complexity Clipping Algorithms
181
0 1 2 3 4 5 6 7 8
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
LPA
IBO=CL+2
IBO=CL+1
IBO=CL
1.4dB
Figure 7.37: Simulated Lucent clipping (1 iteration), vs. BERF with varying IBO in HPA. 64 QAM
symbols, 128 point IFFT (os=2), RRCF with 128 taps, and α=0.15. AWGN=0.
0 1 2 3 4 5 6 7 8
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
LPA
IBO=CL+2
IBO=CL+1
IBO=CL
1.2dB
Figure 7.38: Simulated Lucent clipping (4 iterations), vs. BERF with varying IBO in HPA. 64 QAM
symbols, 128 point IFFT (os=2), RRCF with 128 taps, and α=0.15. AWGN=0.
Chapter 7: Reduced Complexity Clipping Algorithms
182
for 3, 4, and 5 Sector clipping respectively. Vector Subtraction has a difference of 2.5
and 2 magnitudes for 1 and 4 iterations respectively, while the Lucent algorithm has a
difference of 2.2 and 2 magnitudes.
0 1 2 3 4 5 6 7 8
10
4
10
3
10
2
10
1
10
0
Clip Level (dB)
B
E
R
LPA
IBO=CL+2
IBO=CL+1
IBO=CL
0.4dB
Figure 7.39: Simulated Square clipping vs. BERF with varying IBO in HPA. 64 QAM symbols, 128
point IFFT (os=2), RRCF with 128 taps, and α=0.15. AWGN=0.
Table 7.5 compares the baseband clip level for the aforementioned clipping
techniques required to maintain a BER of 10
4
for different backoffs in the HPA. For
3 and 4 Sector clipping decreasing the difference from a LPA to no additional backoff
in the HPA (IBO=CL) comes at a cost of around an additional 1dB increase in the
baseband clip level to maintain the same BERF=10
4
, and a 1.1dB increase for 5
Sector clipping. For Vector Subtraction and the Lucent clipping method the increase
required is around 1.5dB for 1 iteration and 1.3 dB for 4 iterations. Conventional
(Figure 7.6) clipping requires a 1.2dB increase in clipping level and Square clipping
only requires an extra 0.4dB to maintain the same BERF.
Reviewing the results in terms of performance at a set HPA backoff it is seen that for
3 Sector clipping there is an additional 0.4dB of extra baseband clipping backoff
Chapter 7: Reduced Complexity Clipping Algorithms
183
required than for 5 Sector clipping with no additional backoff in the HPA, and an
additional 0.6dB required under LPA conditions. Note that for 3 and 4 Sector
clipping an extra backoff of 2dB in the HPA is sufficient to provide near optimal
results (i.e. under LPA conditions), however for 5 Sector clipping an extra backoff of
3dB is required to approach the LPA results. Performance of Vector Subtraction and
the Lucent method for both 1 and 4 iterations have similar performance at a set
amplifier backoff, interestingly 1 iteration slightly outperforms 4 iterations.
Underestimation of the amplitude must therefore dominate y
je
phase error in the two
schemes.
Table 7.5: Baseband clip level required to maintain a BER=10
4
at varying IBO in HPA.
Baseband clip level required for BER=10
4
at relevant HPA
backoff
HPA backoff CL CL+1 CL+2 LPA
Sector Clipping
3 sectors (Fig 7.32)
7.5 dB 7 dB 6.9 dB 6.6 dB
4 sectors (Fig 7.33)
7.3 dB 7 dB 6.7 dB 6.3 dB
5 sectors (Fig 7.34)
7.2 dB 6.9 dB 6.6 dB 6.1 dB
Vector subtraction
1 iteration (Fig 7.35)
7 dB 6.4 dB 6 dB 5.5 dB
4 iterations (Fig 7.36)
7 dB 6.6 dB 6.2 dB 5.8 dB
Lucent
1 iteration (Fig 7.37)
7 dB 6.4 dB 6.1 dB 5.6 dB
4 iterations (Fig 7.38)
7 dB 6.6 dB 6.2 dB 5.8 dB
Conventional
Conventional (Fig
7.7)
7 dB  6.2 dB 5.8 dB
Square
Square (Fig 7.39)
8.4 dB 8.2 dB 8.1 dB 8 dB
Chapter 7: Reduced Complexity Clipping Algorithms
184
7.4.2 PSD results
The PSD after the HPA and receiver filtering is shown in the following figures. The
filter used is the same as the previous simulations, i.e. Matched RRCF, α=0.15, 128
filter taps and the HPA is a SSPA with p=3. As in Section 3.4.2 the PSD is measured
for each OFDM block and then averaged over many blocks.
In Figures 7.40 and 7.41 the baseband clip level is set at 0dB and 5dB respectively
and the amplifier backoff is set at increasing levels above this clip level (IBO=CL,
CL+1, CL+2, CL+3, CL+4, and a LPA). Baseband clipping at 0dB results in a large
amount of both in band and out of band distortion (ACI). With no additional backoff
in the HPA (IBO=CL) the ACI is only 18dB below the signal power and the inband
distortion is 3dB below the LPA case. Increasing the IBO by 1dB (IBO=CL+1)
results in spectral splatter 20dB below the signal power and an inband distortion of
2.5dB. With IBO=CL+2 the ACI is 22dB below the signal power and the inband
distortion is 1.5dB. An IBO level set at CL+3 results in 23dB ACI and 1.3dB inband
distortion. Even with a backoff of 4dB in the amplifier (IBO=CL+4) 24dB in ACI is
present and the inband distortion is 1dB below the LPA case.
In Figure 7.41 the baseband clip level is set at 5dB and the amplifier backoff is set at
increasing levels above the clip level as in Figure 7.40. Here it is seen that there is
much less distortion, both in band and out of band. There is only 0.5dB difference in
inband distortion between no additional IBO in the HPA and a LPA, and the ACI is
24dB for the IBO=CL case, and for higher backoffs is almost non existent.
These PSD results of 7.40 and 7.41 show that harder clipping in the baseband results
in greater peak regrowth which leads to heavier saturation of the amplifier, hence the
greater ACI and in band distortion. The inband distortion leads to a worse BERF as
the difference between signal power and noise is reduced, i.e. SNR is reduced. The
out of band distortion will lead to interference with adjacent channels.
Chapter 7: Reduced Complexity Clipping Algorithms
185
4 3 2 1 0 1 2 3 4
50
45
40
35
30
25
20
15
10
5
0
Normalized Frequency (Hz)
A
b
s
o
l
u
t
e
P
o
w
e
r
(
d
B
)
LPA
HPA=CL
Figure 7.40: PSD; 0dB clipping in baseband with increasing amplifier backoffs (HPA=CL,
CL+1,CL+2,CL+3,CL+4, LPA). RRCF, alpha=0.15, 128 taps.
4 3 2 1 0 1 2 3 4
45
40
35
30
25
20
15
10
5
0
Normalized Frequency (Hz)
A
b
s
o
l
u
t
e
P
o
w
e
r
(
d
B
)
LPA
HPA=CL
Figure 7.41: PSD; 5dB clipping in baseband with increasing amplifier backoffs (HPA=CL,
CL+1,CL+2,CL+3,CL+4, LPA). RRCF, alpha=0.15, 128 taps.
Chapter 7: Reduced Complexity Clipping Algorithms
186
4 3 2 1 0 1 2 3 4
50
45
40
35
30
25
20
15
10
5
Normalized Frequency (Hz)
A
b
s
o
l
u
t
e
P
o
w
e
r
(
d
B
)
HPA=CL
LPA
Figure 7.42: PSD; after 0dB clipping in baseband and receiver filtering, with increasing amplifier
backoffs (HPA=CL, CL+1,CL+2,CL+3,CL+4, LPA). RRCF, alpha=0.15, 128 taps.
4 3 2 1 0 1 2 3 4
50
45
40
35
30
25
20
15
10
5
0
Normalized Frequency (Hz)
A
b
s
o
l
u
t
e
P
o
w
e
r
(
d
B
)
HPA=CL
LPA
Figure 7.43 : PSD; after 5dB clipping in baseband and receiver filtering, with increasing amplifier
backoffs (HPA=CL, CL+1,CL+2,CL+3,CL+4, LPA). RRCF, alpha=0.15, 128 taps.
Chapter 7: Reduced Complexity Clipping Algorithms
187
Figures 7.42 and 7.43 show the PSD after receiver filtering for 0dB and 5dB baseband
clipping respectively. Here as expected the out of band distortion is mitigated but the
in band distortion remains.
7.5 New adaptive clipping method
Clipping after interpolation and filtering will ensure that the amplifier does not
saturate eliminating AM to PM distortions in the amplifier, but results in ACI
affecting adjacent channels. This effect can be compensated by putting the clipper
before the filter as done earlier in this chapter. However, as shown clipping and
filtering can regrow peaks causing saturation of the amplifier resulting in an increase
in the BERF.
This section introduces a new adaptive clip and filter algorithm, Level Detection
Algorithm (LDA) first presented in [103] which overclips the signal at certain times
avoiding the peak regrowth issue and sparing the amplifier from saturation. However
LDA as presented in [103] requires the use of a conventional clipper which adds
significantly to the overall complexity and latency of the algorithm. Latency is an
important issue in LDA, therefore Vector Subtraction is very useful in this algorithm.
In this section Vector Subtraction with 2 iterations is used in place of the conventional
clipper in LDA and their performance is compared through simulation. Vector
Subtraction has the advantage of providing good estimates of the error magnitude
with low latency, and low complexity, both of which are important in LDA.
LDA uses an extra matched filter before the standard pulse shaping filter to predict the
response of the signal from which the amount of compensation required can be
calculated. A block diagram showing the LDA algorithm is shown in Figure 7.44
with a more detailed view of the filtering operation shown in Figure 7.45. The input
data is modeled as a complex Gaussian process which describes accurately either an
OFDM (N>64) or CDMA distribution of data where the algorithm could be
implemented. The input data is interpolated to form the signal, x(n), (Figure 7.46a)
and then fed to 2 identical filters. The first filter is used for peak detection and for
generation of the correction vector, v(n). The suitably delayed correction vector is
subtracted from the signal, x, before being passed to the second filter, and on to the
Chapter 7: Reduced Complexity Clipping Algorithms
188
rest of the transmission chain. The filter used is a linear phase RRCF, in order to
compensate for the group delay of the filter the subtraction point is set at the centre of
the first filters delay line. This reduces the potential for additional peaks by including
the latter half of the smeared v(n) waveform in the output estimate y(n).
The clipping processor compares the amplitude of the first filter output with the
Clipping Level (CL) threshold to detect peaks. When the filter output magnitude,
y(n), exceeds the CL as shown in (Fig 7.46b) the correction vector, v(n) is
subtracted from the second filter input. The correction vectors (Figure 7.46c) show
the positions where the correction takes place at n1, n, and n+1.
Figure 7.47 shows the I Q diagram for LDA showing how the vector, v(n), is
calculated based on the filter input samples. The correction required at the filter
output to stop saturation at the amplifier is α(n) which is in phase with y(n) and has
amplitude
( ) ( ) n y n CL α = − (7.49)
When the correction vector v(n) is added to the centre of the filter delay line, α(n) will
be
( ) ( )
0
n h v n α = (7.50)
where h
0
is the central filter tap value. Combining (7.49) and(7.50) gives v(n) as
( )
( )
( )
0
1
y n
CL
v n
h y n
= −
(7.51)
Chapter 7: Reduced Complexity Clipping Algorithms
189
Figure 7.44: Block diagram of system with Level Detection Algorithm (LDA).
Figure 7.45: Detailed block diagram of the LDA.
A limitation with LDA is that the correction vector smears into other parts of the
signal, which can introduce new peaks where none existed before. This problem is
especially apparent when a number of peaks appear successively. To combat this
problem the baseband clip level CL is set lower than the Amplifier Clip Level (ACL),
the saturation level of the amplifier. The lower CL reduces the chance of regrown
peaks saturating the amplifier.
Clipping processor
z
1
RRCF
z
1
+
CL
h
0
h
1
v(n) y(n)
x(n1) x(n)
delay
1
st
filter
Clipping
processor
8
2
nd
filter
I/Q
I/P
y
x
v(n)
CL
I/Q
O/P
2
nd
filter
1
st
filter
Chapter 7: Reduced Complexity Clipping Algorithms
190
Figure 7.46: a) Amplitude of zero padded input to filter b) amplitude of filtered output c) LDA
correction vectors.
In order to compare the performance of the LDA 2 parameters are used: the Clipping
Error (CE) produced by the amplifier and the Mean Square Error (MSE) introduced
by the compensation vector. The CE is the noise power of the clipped part in the
amplifier, and is wideband and spread over many channels causing ACI. The MSE
distortion introduced by the compensation filters is filtered by the second filter so no
extra ACI is generated. The MSE introduces in band distortion that interferes with
the desired signal.
n+1∆
n+1+∆ n+1 n
x(n)
time
CL
CL+α(n)
y(n)
time
n
time
n
v(n)
(a)
(b)
(c)
n1 n+1
Chapter 7: Reduced Complexity Clipping Algorithms
191
Figure 7.47: Vector representation of filtering with 3 active taps and the required correction vector to
bring the output back to CL.
A block diagram of the simulation model used to evaluate LDA is shown in Figure
7.48. The interpolation factor is 8 and the RRCF has a roll off factor of 0.2. The
HPA amplifier used in this case is just a linear limiter which saturates at ACL while
the baseband clipper saturates at CL. A matched RRCF simulates the receiver filter
and a decimator follows to sample the received data symbols. The measured clipped
power is the average power of the difference between the amplifier input and output
signals. Similarly, the MSE is the average power of the difference between the
original data samples and the received ones.
x(n+1)h
0
x(n+1∆)h
(1∆)
x(n+1+∆)h
(1+∆)
y(n)
α(n)=v(n)h
0
Correction vector
CL
I
Q
Chapter 7: Reduced Complexity Clipping Algorithms
192
Figure 7.48: Block diagram of the simulation model used to evaluate LDA.
Figures 7.49 and 7.50 plot the simulated CL versus the inband distortion at the
receiver. The ACL is set at 5dB and CL is varied from 2 to 9dB in Figure 7.49, and in
Figure 7.50 the ACL is set at 8dB and the CL is varied from 7 to 12dB. Curves are
plotted for Vector Subtraction with 2 iterations and Conventional clipping, both of
which are implemented in a 64 tap and 128 tap filter.
Random
data
source
8
LDA
(Figure7.45)
Amp
Measure clip
noise
RRCF
8
Measure
noise
RRCF RRCF
Delay
Clipping Error (CE)
Transmit signal power
In band distortion (MSE)
Receive signal power
CL ACL
Chapter 7: Reduced Complexity Clipping Algorithms
193
2 3 4 5 6 7 8 9
26
24
22
20
18
16
14
12
10
SCL to input power
I
n
b
a
n
d
d
i
s
t
o
r
t
i
o
n
t
o
R
x
s
i
g
n
a
l
Conventional  64 taps
Vector Sub  2 iter  64 taps
Conventional  128 taps
Vector Sub  2 iter  128 taps
Figure 7.49: In band distortion for Conventional and Vector Subtraction (2 iter) with 64 and 128 taps in
compensation filter. ACL=5dB.
7 7.5 8 8.5 9 9.5 10 10.5 11 11.5 12
44
42
40
38
36
34
32
30
28
SCL to input power
I
n
b
a
n
d
d
i
s
t
o
r
t
i
o
n
t
o
R
x
s
i
g
n
a
l
Conventional  64 taps
Vector Sub  2 iter  64 taps
Conventional  128 taps
Vector Sub  2 iter  128 taps
Figure 7.50: In band distortion for Conventional and Vector Subtraction (2 iter) with 64 and 128 taps in
compensation filter. ACL=8dB.
Chapter 7: Reduced Complexity Clipping Algorithms
194
In all cases when the CL is set below the ACL more inband noise is produced as the
noise is dominated by the baseband clipping (CL). When the CL is set above the
ACL the clipping noise is dominated by the amplifier and levels off at CL=9dB for
ACL=5dB and at CL=10dB ACL=8dB. When the ACL=CL=5dB the inband
distortion is 20dB below the received signal and when ACL=CL=8dB the inband
distortion is around 35dB below the receive signal. Little variation is seen in
performance between Conventional and Vector Subtraction with the 64 tap filter
having slightly worse performance in the conventional case.
Figures 7.51 and 7.52 show the simulated results for the Clipping Error (CE) versus
the inband distortion, or Mean Squared Error (MSE) for a constant ACL. The results
are obtained by varying the CL and recording the two error sources (CE and MSE).
When the CL is set larger than the ACL, all the clipping is performed by the amplifier
and the most noise is produced as is evidenced by the top of the curves in Figures 7.51
and 7.52.
The bottom part of the curves in Figures 7.51 and 7.52 show noise produced when the
CL is set smaller than the ACL, as a result clipping is mostly performed by the
baseband clipper rather than the amplifier. The noise produced at the bottom of the
curves is due to peak regrowth after baseband clipping saturating the amplifier. In the
ACL=5dB case it is seen that at the top of the curves there is no discernable difference
in performance between Conventional and Vector Subtraction, the number of taps
also has a negligible effect on performance. At the bottom of the curves where the
noise is due to peak regrowth there is a divergence in performance between
Conventional and Vector Subtraction with Conventional clipping out performing
Vector Subtraction by around 6dB from 55dB to 61dB in terms of the clipping
noise. When the ACL is increased to 8dB the performance improves across the board
as expected. Again Conventional clipping outperforms Vector Subtraction by at least
1dB, with Vector Subtraction displaying 64dB of noise and Conventional clipping
displaying no noise below 65dB. The number of taps in the filter has a negligible
effect. The amplifier CE is low enough for LDA with Vector Subtraction to meet the
ACI specifications of most standards.
Chapter 7: Reduced Complexity Clipping Algorithms
195
Figure 7.51: Clipping error vs. inband distortion. Performance curves for LDA using Conventional
clipping and Vector Subtraction with 2 iterations. Compensation filter has 64 and 128 taps.
ACL=5dB.
Figure 7.52: Clipping error vs. inband distortion. Performance curves for LDA using Conventional
clipping and Vector Subtraction with 2 iterations. Compensation filter has 64 and 128 taps.
ACL=8dB.
26 24 22 20 18 16 14 12 10
70
65
60
55
50
45
40
35
30
25
20
CL > ACL
Conventional  64
Vector Sub  2 iter  64
Conventional  128
Vector Sub  2 iter  128 taps
CL = ACL = 5dB
CL < ACL
44 42 40 38 36 34 32 30 28
65
60
55
50
45
40
35
Inband distortion to Rx signal
CL > ACL
Conventional  64
Vector Sub  2 iter  64
Conventional  128
Vector Sub  2 iter 128 taps
CL = ACL = 8dB
CL < ACL
Chapter 7: Reduced Complexity Clipping Algorithms
196
7.6 Conclusion
This chapter presented new low complexity clipping methods for OFDM. Initially a
Monte Carlo OFDM transceiver simulation model was used to establish the basic
parameters on which the new clipping algorithms were tested. Various parameters
such as the mapping type, the amount of oversampling in the IFFT, the baseband clip
level, the number of taps in the pulse shaping filter, and the HPA backoff were varied
to quantify their affect on the BERF. In this chapter the BERF was plotted against the
clipping level with the noise set to zero, this meant that all the errors produced were
due to clipping noise alone. Good parameters for further simulation were found to be
oversampling the IFFT by a factor 2, a RRCF with 128 filter taps for a rolloff factor
of 0.15, and a SSPA with p=3. The effect of filtering after clipping was also studied,
it was shown through simulation how peak regrowth was more severe the harder the
signal is clipped in the baseband. The peak regrowth could be countered to a certain
degree by oversampling the IFFT by a factor of 2, oversampling beyond this point
provided little discernable improvement. QPSK mapping was shown to be extremely
robust to clipping, therefore making hard clipping a viable alternative at lower data
rates.
In Section 7.2 Sector clipping was introduced as a low complexity alternative to
conventional clipping. Theoretical analysis of Conventional, Sector, and Square
clipping SNR was shown to be in good agreement with the simulated results. At
higher clip levels the theoretical and simulated results diverged as the theoretical SNR
used the assumption that the clipping noise was Gaussian in nature. This was shown
to be untrue at higher clip levels where the clipping noise is more impulsive in nature,
hence the divergence in results. Both 3 Sector and square clipping were implemented
in silicon using a 0.5µ, 3 metal layer process. Reports generated by Synopsys and
cadence proved that the complexity increase of the Sector clipping algorithm was
small compared to Square clipping with the advantage of having much better SNR
properties. Sector clipping was expanded to include more sectors thereby improving
performance. 4 and 5 Sector clipping provided a marginal improvement over 3 sector
in terms of the BERF with a minor increase in complexity.
Chapter 7: Reduced Complexity Clipping Algorithms
197
Section 7.3 introduced another low complexity enhancement to an existing Lucent
algorithm. An iterative method similar to the CORDIC algorithm to calculate the
magnitude of a complex sample. The variation called Vector Subtraction further
reduced complexity by the removal of a complex multiplication for the price of 2
extra additions and a comparison operation. Its performance was found to be as good
as conventional clipping with only 2 iterations.
Section 7.4 compared simulated BERF results of the Sector, Square, Lucent, Vector
Subtraction, and Conventional clipping under various backoffs in the HPA. It was
shown that while the Vector Subtraction and Lucent algorithms have better
performance across the board, Sector clipping was more tolerant to harder backoff
levels in the HPA. The PSD results for conventional clipping at 0dB and 5dB at
various HPA backoffs was shown. While the inband distortion remains in both
clipping modes the ACI present at 0dB is up to 17dB below the signal power while
ACI is almost non existent at 5dB baseband clipping. As most WCDMA and OFDM
standards require an ACI of 20dB below the signal power it was shown that even with
hard clipping the ACI specifications are met.
For a baseband clipping system with 16 QAM mapping, an IFFT with an
oversampling factor of 2, a pulse shaping RRCF with 128 taps and α=0.15, to
maintain a BER of 10
4
clipping at 3.8 to 4dB above the average signal power with a
HPA with backoff 2dB above the baseband clip level is recommended. In this
situation the HPA must have a peak power some 6dB above the average signal power.
Finally in Section 7.5 Vector Subtraction with 2 iterations was implemented into an
existing adaptive clipping system which required the clipping operation to take place
within the filter. A low latency magnitude estimate was therefore required.
Simulated results showed that Vector Subtraction had similar performance to
conventional clipping in this algorithm but with less latency and complexity.
Chapter 8: Conclusion
198
Chapter 8
Conclusion
This thesis analysed and proposed new methods to deal with the PAPR in OFDM.
Initially Chapter 2 introduced the theory and principles behind OFDM and detailed
scenarios where it is used. Chapter 3 identified contributing factors to the PAPR,
these were that peaks are a function of the IFFT operation, where in phase waveforms
add to create a large peak. The general distribution of OFDM samples was shown to
have a Rayleigh distribution. The number of subcarriers, and to a lesser degree the
mapping type were also shown to contribute to the PAPR. Simulations of an OFDM
system revealed how uncontrolled large peaks will saturate the HPA creating ACI and
an increased BER at the receiver.
Chapter 4 began the literature review with an analysis of non distorted PAPR
reduction techniques which included coding techniques, PTS, SLM, and Tone
Reservation /Insertion. Coding introduced the most redundancy and became
extremely complex at a higher number of subcarriers (N>64), however later papers
identified promising code sets such as second order ReedMuller codes. PTS and
SLM reduced the PAPR by producing a series of alternative transmit signals seeded
from the same data source which are altered before the IFFT process so that they will
have different PAPR properties. The waveform with the lowest PAR is chosen for
transmission. These methods are complex and the amount of PAPR is not guaranteed.
Tone Insertion/variation used peak reduction carriers which introduced redundancy
and required some additional processing at the receiver.
Chapter 5 proposed two new low complexity variations to PTS, called Cyclic Shifted
Sequences (CSS) and Time Inversion (TI). CSS and TI were shown to be less
complex than PTS when the number of phase rotations was greater than 4.
Chapter 8: Conclusion
199
Combining shifts of CSS and TI with non complex phase rotations of standard PTS
allowed a whole IFFT operation to be removed at the expense of some extra non
complex operations. CSS and TI where shown to display less peak regrowth after
pulse shaping filtering. Oversampling at the IFFT by a factor of 2 was shown to
improve the performance of PTS, CSS, and TI bringing the discrete and filtered
CCDF to within 1dB of each other. These methods provided reduction in the CCDF
of between 2 and 3dB at Pr(ζ>ζ
0
)=10
4
.
Chapter 6 picked up the literature review again from Chapter 4 for distorted PAPR
solutions. Distorted PAPR methods were defined here as methods which intentionally
limit the excess peaks at the transmitter, usually in the baseband so that the HPA
would not saturate. These papers revealed that a backoff of around 6dB is sufficient
to maintain a respectable error rate for 4 and 16 QAM mapping. Other papers
analysed the affect on the amplifier of saturation and the resultant in band and out of
band distortion. Windowing was also examined which is the process of using a pulse
shape to clip the signal and surrounding samples to give better spectrum properties.
The gains in reduction in ACI were shown to be minimal when an acceptable amount
of backoff in the HPA was used.
Chapter 7 presented the next set of new PAPR solutions where a series of low
complexity and low latency clipping algorithms were proposed. This chapter started
with a detailed Monte Carlo simulation of an OFDM transceiver were different stages
of the transmission chain were modified to ascertain their effect on the BERF. The
motivation behind clipping as a solution to the PAPR was that with an acceptable
HPA backoff of around 6dB, clipping is a very rare occurrence affecting the BER
negligibly. The first method, Sector clipping uses a rule base to perform the clipping
operation. It was implemented in silicon and shown to have negligible extra
complexity to square clipping but with much better performance. The rule base was
extended up to 5 sectors where performance was close to conventional clipping. The
second method used a variation of a CORDIC like magnitude estimator and was
called Vector Subtraction. The complexity was further reduced by the removal of a
multiplication at the expense of 2 additions and comparisons. Vector Subtraction was
shown to have identical performance to conventional clipping with only 2 iterations.
Finally Vector Subtraction with 2 iterations was implemented in a previously
Chapter 8: Conclusion
200
proposed clip and filter algorithm where its low latency was important as the
magnitude had to be found quickly as it was operating in a feedback loop on filter
samples.
8.1 Future Work
The PAPR problem in OFDM is still an ongoing issue, especially for portable devices
where the need to minimise the power amplifier linear range is paramount. The
PTS/CSS/TI methods developed in this thesis to reduce the PAPR can be combined
with other PTS techniques such as adaptive PTS and variations of the blind SLM
techniques to further reduce complexity and the peak power. A hybrid system
utilizing clipping techniques could also be added as a last stage so as to have an upper
bound for the PTS signal.
Further enhancements of Sector clipping can be made where the error introduced by
the phase distortion could be minimised so that only the amplitude distortion remains.
All of the clipping techniques detailed in this thesis could be combined with coding
techniques to further improve the BER performance. Code sets for larger numbers of
subcarriers are an open ended problem and there is much ongoing research in this
area.
Analysis of the proposed algorithms in a MIMO OFDM system is an area that has
gained a lot of focus recently and analysis of the performance of the proposed
techniques in such an environment would be valid.
The LDA algorithm with Vector Subtraction could be implemented in a complete
OFDM transceiver and then implemented in hardware via a FPGA or silicon to
ascertain whether the latency requirements can be met with the wider bandwidth
systems now being proposed (40 MHz for 802.11n, and up to 100 MHz for 4G LTE).
Bibliography
201
Bibliography
[1] G. Hill, M. Faulkner, and J. Singh, "Reducing the peaktoaverage power ratio
in OFDM by cyclically shifting partial transmit sequences," Electronics
Letters, vol. 36, pp. 560561, Mar 16 2000.
[2] G. Hill, M. Faulkner, and J. Singh, "Cyclic Shifting and Time Inversion of
Partial Transmit Sequences to Reduce the PeaktoAverage Power Ratio in
OFDM," in PIMRC, London, 2000.
[3] G. Hill, "Comparison of Low complexity clipping algorithms for OFDM," in
PIMRC, Portugal, 2002.
[4] M. L. Doelz, E. T. Heald, and D. L. Martin, "Binary Data Transmission
Techniques for Linear Systems," Proc IRE, vol. 45, pp. 656661, May 1957.
[5] R. W. Chang, "Synthesis of Band Limited Orthogonal Signals for
Multichannel Data Transmission," Bell Syst. Tech. Jo., vol. 45, pp. 1,775
1,796, Dec 1966.
[6] B. R. Saltzberg, "Performance of an efficient parallel data transmission
system," IEEE Transactions on Communication Technologies, vol. COM15,
Dec 1967 1967.
[7] S. Weinstein and P. Ebert, "Data transmission by frequency division
multiplexing using the discrete fourier transform," IEEE Transactions on
Communication Techniques, vol. COM19, pp. 628634, October 1971.
[8] L. J. Cimini Jr, Ye, L., "Orthogonal Frequency Division Multiplexing for
Wireless Channels," in IEEE Global Telecommunications Conference
GLOBECOM, Sydney, Australia, 1998, p. 82.
[9] C. H. Chang, C. L. Wang, and Y. T. Chang, "Efficient VLSI architectures for
fast computation of the discrete fourier transform and its inverse," IEEE
Transactions on Signal Processing, vol. 48, pp. 32063216, Nov 2000.
[10] Y. Jung, Y. Tak, J. Kim, J. Park, D. Kim, and H. Park, "Efficient FFT
algorithm for OFDM modulation," in Electrical and Electronic Technology,
2001. TENCON. Proceedings of IEEE Region 10 International Conference on,
2001, pp. 676678 vol.2.
[11] A. Sadat and W. B. Mikhael, "Fast Fourier Transform for high speed OFDM
wireless multimedia system," in Circuits and Systems, 2001. MWSCAS 2001.
Proceedings of the 44th IEEE 2001 Midwest Symposium on, 2001, pp. 938
942 vol.2.
[12] A. D. S. Jayalath, "OFDM for Wireless Broadband Communications (Peak
Power Reduction, Spectrum and Coding)," in School of Computer Science and
Software Engineering Melbourne: Monash University, 2002, p. 257.
[13] L. Hanzo, Webb, W., Keller, T., Single and Multicarrier Quadrature
Amplitude Modulation. Chichester: Wiley, 2000.
[14] O. Edfors, Sandell, M., van de Beek, J.J., Landström, D., Sjöberg, "An
Introduction to Orthogonal Frequency Division Multiplexing," Division of
Bibliography
202
Signal Processing, Luleå University of Technology, Research Report TULEA
1996:16 1996.
[15] R. Morrison, L. J. Cimini, Jr., and S. K. Wilson, "On the use of a cyclic
extension in OFDM," in Vehicular Technology Conference, 2001. VTC 2001
Fall. IEEE VTS 54th, 2001, pp. 664668 vol.2.
[16] T. S. Rappaport, Wireless Communications, Principles and Practice. New
Jersey: Prentice Hall, 1996.
[17] S. Coleri, M. Ergen, A. Puri, and A. Bahai, "Channel estimation techniques
based on pilot arrangement in OFDM systems," Broadcasting, IEEE
Transactions on, vol. 48, pp. 223229, 2002.
[18] M. Morelli and U. Mengali, "A comparison of pilotaided channel estimation
methods for OFDM systems," IEEE Transactions on Signal Processing, vol.
49, pp. 30653073, Dec 2001.
[19] W. G. Jeon, K. H. Paik, and Y. S. Cho, "Twodimensional pilotsymbolaided
channel estimation for OFDM systems with transmitter diversity," IEICE
Transactions on Communications, vol. E85B, pp. 840844, Apr 2002.
[20] J. Shentu and J. Armstrong, "Blind frequency offset estimation for PCC
OFDM with symbols overlapped in the time domain," in Circuits and Systems,
2001. ISCAS 2001. The 2001 IEEE International Symposium on, 2001, pp.
570573 vol. 4.
[21] F. Tufvesson, "Design of Wireless Communication Systems  Issues on
Synchronization, Channel Estimation and MultiCarrier Systems," in
Department of Applied Electronics Lund: Lund University, 2000.
[22] M. Faulkner, "OFDM: Overview," Melbourne, 2002.
[23] A. G. Armada, "Understanding the effects of phase noise in orthogonal
frequency division multiplexing (OFDM)," IEEE Transactions on
Broadcasting, vol. 47, pp. 153159, Jun 2001.
[24] T. Pollet, M. Vanbladel, and M. Moeneclaey, "Ber sensitivity of ofdm systems
to carrier frequency offset and wiener phase noise," IEEE Transactions on
Communications, vol. 43, pp. 191193, 1995.
[25] ETSI, "Radio Broadcast Systems: Digital Audio Broadcasting (DAB) to
mobile, portable and fixed receivers," ETSI final draft ETS 300 401, Nov
1994.
[26] P. Shelswell, "The COFDM modulation system: The heart of digital audio
broadcasting," Electronic and Communication Engineering Journal, vol. 7,
pp. 127135, June 1995.
[27] J. Stott, "The DVB terrestial (DVBT) specification and its implementation in
a practical modem," International Broadcasting Convention, pp. 255260,
1996.
[28] ETSI, "Digital Video Broadcasting (DVB); framing structure, channel coding,
and modulation for digital terrestial television (DVBT)," European
Telecommunications Standard Institute ETSI EN 300 744, Nov 1999.
[29] ISO/IEC, "Information technology  generic coding of moving pictures and
associated audio information: Part 2 video," ISO/IEC 13818, 2000.
[30] J. TelladoMourelo, "Peak to Average Power Reduction for Multicarrier
Modulation," in Department of Electrical Engineering: Stanford University,
1999.
[31] H. Bolckei, D. Gesbert, and A. J. Paulraj, "On the capacity of OFDMbased
spatial multiplexing systems," Communications, IEEE Transactions on, vol.
50, pp. 225234, 2002.
Bibliography
203
[32] S. H. Müller, R. W. Bauml, R. F. H. Fischer, and J. B. Huber, "Ofdm with
reduced peaktoaverage power ratio by multiple signal representation,"
Annales des Telecommunications Annals of Telecommunications, vol. 52, pp.
5867, 1997.
[33] C. Annis, "Central Limit Theorem,
http://www.statisticalengineering.com/central_limit_theorem.htm," 2004.
[34] C. Tellambura, "Phase optimisation criterion for reducing peaktoaverage
power ratio in ofdm," Electronics Letters, vol. 34, pp. 169170, 1998.
[35] H. Ochiai and H. Imai, "On the distribution of the peaktoaverage power ratio
in OFDM signals," IEEE Transactions on Communications, vol. 49, pp. 282
289, Feb 2001.
[36] M. Sharif, M. GharaviAlkhansari, and B. H. Khalaj, "On the peaktoaverage
power of ofdm signals based on oversampling," Communications, IEEE
Transactions on, vol. 51, pp. 7278, 2003.
[37] R. v. Nee, OFDM for wireless multimedia communications. Boston ; London:
Artech House, 2000.
[38] G. Wunder and H. Boche, "Upper bounds on the statistical distribution of the
crestfactor in ofdm transmission," Information Theory, IEEE Transactions
on, vol. 49, pp. 488494, 2003.
[39] S. MüllerWeinfurtner, "OFDM for Wireless Communications: Nyquist
Windowing, PeakPower Reduction, and Synchronization," in Reihe
Kommunikations und Informationstechnik Nürnberg: Universität Erlangen,
2000.
[40] E. Costa, M. Midrio, and S. Pupolin, "Impact of amplifier nonlinearities on
OFDM transmission system performance," Communications Letters, vol. 3,
pp. 3739, 1999.
[41] E. Costa and S. Pupolin, "MQAMOFDM system performance in the
presence of a nonlinear amplifier and phase noise," IEEE Transactions on
Communications, vol. 50, pp. 462472, Mar 2002.
[42] C. L. Liu, "The effect of nonlinearity on a qpskofdmqam signal," IEEE
Transactions on Consumer Electronics, vol. 43, pp. 443447, 1997.
[43] S. Merchan, A. G. Armada, and J. L. Garcia, "Ofdm performance in amplifier
nonlinearity," IEEE Transactions on Broadcasting, vol. 44, pp. 106114,
1998.
[44] G. T. Zhou and J. S. Kenney, "Predicting spectral regrowth of nonlinear power
amplifiers," Communications, IEEE Transactions on, vol. 50, pp. 718722,
2002.
[45] G. Santella and F. Mazzenga, "A hybrid analyticalsimulation procedure for
performance evaluation in mqamofdm schemes in presence of nonlinear
distortions," IEEE Transactions on Vehicular Technology, vol. 47, pp. 142
151, 1998.
[46] X. Li and L. J. Cimini, "Effects of clipping and filtering in the performance of
OFDM," IEEE Communications Letters, vol. 2, pp. 131133, May 1998.
[47] C. van den Bos, M. H. L. Kouwenhoven, and W. A. Serdijn, "Effect of smooth
nonlinear distortion on OFDM symbol error rate," IEEE Transactions on
Communications, vol. 49, pp. 15101514, Sep 2001.
[48] M. Friese, "On the degradation of OFDM signals due to peak clipping in
optimally predistorted power amplifiers," in Globecom, Sydney, Australia,
1998.
Bibliography
204
[49] A. E. Jones, T. A. Wilkinson, and S. K. Barton, "Block coding scheme for
reduction of peak to mean envelope power ratio of multicarrier transmission
schemes," Electronics Letters, vol. 30, pp. 20982099, 1994.
[50] T. A. Wilkinson and A. E. Jones, "Minimisation of the Peak to Mean
Envelope Power Ratio of Multicarrier Transmission Schemes by Block
Coding," in VTC, Chicago, USA, 1995.
[51] A. E. Jones, "Combined Coding for Error Control and Increased Robusteness
to System Nonlinearities in OFDM," in IEEE 46th VTC, Atlanta, USA, 1996,
pp. pp 904908.
[52] V. Tarokh and H. Jafarkhani, "An Algorithm for Reducing the Peak to
Average Power Ratio in a Multicarrier Communication System," in IEEE
VTC, Piscataway, NJ, USA, 1999, pp. pp 680684.
[53] ETSI, "Broadband Radio Access Networks (BRAN); HIPERLAN Type 2;
Physical (PHY) layer," European Telecommunications Standards Institute
ETSI TS 101 475 v1.1.1, 2000.
[54] V. Tarokh and H. Jafarkhani, "On the computation and reduction of the peak
toaverage power ratio in multicarrier communications," Communications,
IEEE Transactions on, vol. 48, pp. 3744, 2000.
[55] P. W. J. VanEetvelt, S. J. Shepherd, and S. K. Barton, "The Distribution of
Peak Factor in QPSK MultiCarrier Modulation," in Wireless Communications
Netherlands: Kluwer Academic Publishers, 1995, pp. 8795.
[56] S. Shepherd, J. Orriss, and S. Barton, "Asymptotic limits in peak envelope
power reduction by redundant coding in orthogonal frequencydivision
multiplex modulation," IEEE Transactions on Communications, vol. 46, pp. 5
10, 1998.
[57] D. Wulich, "Peak Factor in Orthogonal Multicarrier Modulation with Variable
Levels," Electronic Letters, vol. 32, pp. 18591861, September 1996.
[58] J. E. M. Nilsson, "Coding to control the signal waveform in Mary PSK
multicarrier communications," in Radio Vetenskap och Kommunikation '96
(RVK), Luleå, Sweden, 1996, pp. pp 658662.
[59] K. G. Paterson, "Bounds on Envelope Power of Trace Codes for OFDM,"
Hewlett Packard Laboratories, Bristol HPL98176, October 1998.
[60] C. Tellambura, "Upper bound on peak factor on nmultiple carriers,"
Electronics Letters, vol. 33, pp. 16081609, 1997.
[61] C. Wulich, "Reduction of Peak to Mean Ratio of Multicarrier Modulation
using Cyclic Coding," Electronic Letters, vol. 32, pp. 432433, February 1996.
[62] Y. Y. Zhang, A.; Chouinard, J.Y.; Zhang, L., "OFDM peak power reduction
by subblockcoding and its extended versions," in Vehicular Technology
Conference, 1999 IEEE 49th, 1999, pp. 695699 vol.1.
[63] S. Boyd, "Multitone signals with low crest factor," Circuits and Systems, IEEE
Transactions on, vol. 33, pp. 10181022, 1986.
[64] S. J. V. E. Shepherd, P.W.J.; WyattMillington, C.W.; Barton, S.K., "Simple
coding scheme to reduce peak factor in QPSK multicarrier modulation," in
Electronics Letters. vol. 31, 1995, pp. 11311132.
[65] R. D. J. van Nee, "OFDM codes for peaktoaverage power reduction and
error correction," in Global Telecommunications Conference, 1996.
GLOBECOM '96. 'Communications: The Key to Global Prosperity, 1996, pp.
740744 vol.1.
Bibliography
205
[66] J. A. Davis and J. Jedwab, "Peaktomean power control and error correction
for ofdm transmission using golay sequences and reedmuller codes,"
Electronics Letters, vol. 33, pp. 267268, 1997.
[67] J. A. Davis and J. Jedwab, "Peaktomean power control in OFDM, Golay
complementary sequences, and ReedMuller codes," IEEE Transactions on
Information Theory, vol. 45, pp. 23972417, 1999.
[68] A. E. Jones and T. A. Wilkinson, "Performance of ReedMuller codes and a
maximumlikelihood decoding algorithm for OFDM," IEEE Transactions on
Communications, vol. 47, pp. 949952, 1999.
[69] A. G. Armada, V. Gil, and J. L. Garcia, "Improving Peak to Average power
Ratio (PAR) and Probability of Error in OFDMbased WLAN," in IST
Summit, United Kingdom, 2001.
[70] A. J. Grant and R. van Nee, "Efficient maximum likelihood decoding of peak
power limiting codes for OFDM," in Vehicular Technology Conference, 1998.
VTC 98. 48th IEEE, 1998, pp. 20812084 vol.3.
[71] I. Ochiai, M. P. C. Fossorier, and H. Imai, "On decoding of block codes with
peakpower reduction in OFDM systems," IEEE Communications Letters, vol.
4, pp. 226228, 2000.
[72] B. S. Tarokh, H.R., "Construction of ofdm m qam sequences with low peak
toaverage power ratio," Communications, IEEE Transactions on, vol. 51, pp.
2528, 2003.
[73] K. Sathananthan and C. Tellambura, "Coding to reduce both par and picr of an
ofdm signal," IEEE Communications Letters, vol. 6, pp. 316318, 2002.
[74] S. H. Müller and J. B. Huber, "A comparison of Peak Power Reduction
Schemes for OFDM," in Globecom '97, Phoenix, USA, 1997.
[75] S. H. Müller and J. B. Huber, "Ofdm with reduced peaktoaverage power
ratio by optimum combination of partial transmit sequences," Electronics
Letters, vol. 33, pp. 368369, 1997.
[76] S. H. Müller and J. B. Huber, "A novel peak power reduction scheme for
OFDM," in IEEE Personal Indoor Mobile Radio Communications, Helsinki,
Finland, 1997, pp. 10901094.
[77] S. H. Müller and J. B. Huber, "OFDM with Reduced Crest Factor," ETSI
BRAN  European Telecommunications Standards Institute Broadband Radio
Access Network, Ulm, Germany WG3 temporary document 25, 16 September
1997.
[78] S. G. Kang, J. G. Kim, and E. K. Joo, "A novel subblock partition scheme for
partial transmit sequence OFDM," IEEE Transactions on Broadcasting, vol.
45, pp. 333338, 1999.
[79] L. J. Cimini and N. R. Sollenberger, "Peaktoaverage power ratio reduction of
an OFDM signal using partial transmit sequences," in ICC'99, 1999.
[80] L. J. Cimini and N. R. Sollenberger, "Peaktoaverage power ratio reduction of
an OFDM signal using partial transmit sequences," Communications Letters,
vol. 4, pp. 8688, Mar 2000.
[81] C. Tellambura and A. D. S. Jayalath, "PAR reduction of an OFDM signal
using partial transmit sequences," in Vehicular Technology Conference, 2001.
VTC 2001 Fall. IEEE VTS 54th, 2001, pp. 465469 vol.1.
[82] H. Chen and G. J. Pottie, "An orthogonal projectionbased approach for PAR
reduction in OFDM," Communications Letters, vol. 6, pp. 169171, May 2002.
Bibliography
206
[83] R. W. Bauml, R. F. H. Fischer, and J. B. Huber, "Reducing the peakto
average power ratio of multicarrier modulation by selected mapping,"
Electronic Letters, vol. 32, pp. 20562057, 1996.
[84] H. Breiling, S. H. MullerWeinfurtner, and J. B. Huber, "SLM peakpower
reduction without explicit side information," IEEE Communications Letters,
vol. 5, pp. 239241, 2001.
[85] A. D. S. T. Jayalath, C., "A blind slm receiver for parreduced ofdm," in
Vehicular Technology Conference, 2002. Proceedings. VTC 2002Fall. 2002
IEEE 56th, 2002, pp. 219222.
[86] J. Tellado and J. Cioffi, "Peak Power Reduction of Multicarrier
Transmission," in Globecom 1998, Sydney, Australia, 1998.
[87] R. Dinis and A. Gusmao, "On the performance evaluation of OFDM
transmission using clipping techniques," in VTC'99, 1999.
[88] H. Nikopour and S. H. Jamali, "Effects of Cartesian Clipping Noise on the
Performance of Orthogonal Frequency Division Multiplexing System: A
Theoretical Approach," in IEEE GLOBECOM 2002, 2002, pp. 12541258.
[89] A. R. S. Bahai, M. Singh, A. J. Goldsmith, and B. R. Saltzberg, "A new
approach for evaluating clipping distortion in multicarrier systems," IEEE
Journal on Selected Areas in Communications, vol. 20, pp. 10371046, Jun
2002.
[90] G. Wunder and H. Boche, "The Impact of Nonlinear Devices on the Symbol
Error Rate in Broadband OFDM Transmission,"?, vol.?, 2002.
[91] T. May and H. Rohling, "Reducing the Peak to Average Power Ratio in
OFDM Radio Transmission Systems," in VTC '98, ?, 1998, pp. 24742478.
[92] J. Armstrong, "Peaktoaverage power reduction for OFDM by repeated
clipping and frequency domain filtering," Electronics Letters, vol. 38, pp. 246
247, Feb 28 2002.
[93] D. Kim and G. L. Stuber, "Clipping noise mitigation for OFDM by decision
aided reconstruction," Communications Letters, vol. 3, pp. 46, 1999.
[94] D. J. G. Mestdagh, P. M. P. Spruyt, and B. Biran, "Effect of Amplitude
Clipping in DMTADSL Transceivers," Electronic Letters, vol. 29, pp. 1354
1355, 22nd July 1993.
[95] D. J. G. Mestdagh and P. M. P. Spruyt, "A Method to Reduce the Probability
of Clipping in DMTBased Transceivers," IEEE Transactions on
Communications, vol. 44, pp. 12341238, October 1996.
[96] B. Come, R. Ness, S. Donnay, L. V. d. Perre, W. Eberle, P. Wambacq, M.
Engels, and I. Bolsens, "Impact of FrontEnd NonIdealities on Bit Error Rate
of WLANOFDM Transceivers," in Microwave Journal, 2001, pp. 126140.
[97] J. Armstrong, H. A. Suraweera, S. Brewer, and R. Slaviero, "Effect of
Rouunding and Saturation in FizedPoint DSP Implementation of IFFT and
FFT for OFDM Applications," in The Embedded Signal Processing
Conference (GSPx 2004), Santa Clara, CA, USA, 2004.
[98] S. Merchan, A. G. Armada, and J. L. Garcia, "OFDM performance in
amplifier nonlinearity," Broadcasting, IEEE Transactions on, vol. 44, pp. 106
114, 1998.
[99] R. v. Nee and A. d. Wild, "Reducing the PeaktoAverage Power Ratio of
OFDM," in Vehicular Technology Conference, Ottawa, Canada, 1998, pp.
20722076.
[100] S. B. Slimane, "Peaktoaverage power ratio reduction of OFDM signals using
broadband pulse shaping," in IEEE VTC '02, Canada, 2002, pp. 889893.
Bibliography
207
[101] R. S. Creighton, "System and method to reduce the peaktoaverage power
ratio in DSCDMA transmitter," in European Patent Application Great
Britain: Lucent Technologies (UK), 1999.
[102] J. E. Volder, "The CORDIC Trigonmetric Computing Technique," IRE Trans.
Electronic Computing, vol. 8, pp. 330334, Sept 1959.
[103] R. Berangi and M. Faulkner, "Peak Power Reduction for MultiCode CDMA
and Critically Sampled Complex Gaussian Signals," in IST 2001, Iran, 2001.
Abstract
Orthogonal Frequency Division Multiplexing (OFDM) is a digital transmission method developed to meet the increasing demand for higher data rates in communications which can be used in both wired and wireless environments. This thesis describes the issue of the Peak to Average Power Ratio (PAPR) in OFDM which is a major drawback, and presents new and variations to existing algorithms to reduce it.
Initially the theoretical principles behind OFDM are discussed elaborating on the advantages and disadvantages of OFDM. This is followed by analysis of the PAPR in OFDM where it is shown through theoretical analysis and simulation that the occurrence of large peaks in OFDM is actually quite rare. The effect on system performance in terms of the Bit Error Rate (BER) and Power Spectral Density (PSD) is simulated for an OFDM transceiver with a saturated High Power Amplifier. This is followed by a study of published PAPR reduction methods
The first contribution is a low complexity variation of Partial Transmit Sequences (PTS). In PTS several alternate transmit signals are seeded from the same source, each alternate transmit signal has a reversible and different phase rotation performed on the data. The transmit signal with the lowest PAPR is chosen for transmission. In novel variations, called Cyclic Shifted Sequences (CSS) and Time Inversion (TI), different shifts of the data are performed which avoid the need for complex multiplications. In certain cases a whole IFFT operation can be removed with a negligible effect on performance when CSS is combined with PTS. Furthermore it is shown that the peak regrowth of TI and CSS after pulse shaping filtering is considerably less than for PTS.
Next, new clipping techniques are presented which reduce substantially the complexity of clipping algorithms by using novel methods to calculate the magnitude, avoiding the use of multiplications. One method, called Sector clipping uses a rule base to clip the signal, dividing the clipping region into a series of sectors. When the rule base is expanded to include more sectors the performance is shown to approach i
more complex existing clipping methods. This algorithm is implemented in silicon in a 3 metal layer 0.5µ process. Another clipping scheme called Vector Subtraction is a variation of another low complexity magnitude estimate method which further reduces complexity by alleviating the need for a scaling operation. The performance of the new methods was ascertained through simulation of a whole OFDM transceiver chain and shown to have relative BER’s. Finally Vector Subtraction was
implemented in a previously proposed clip and filter algorithm where its low latency and accuracy proved it to be suitable for the algorithm.
ii
Acknowledgements
First of all I wish to thank Professor Mike Faulkner for his guidance, patience, advice, and time thorough my many years of research. I owe the completion of this thesis to his efforts.
I also wish to thank Dr Reza Berangi for his mathematical and simulation knowledge and his willingness to share it. Special thanks goes to the post graduate studies committee for the numerous second chances to complete my research and to Shirley Herewyn for her omnipotent knowledge on dealing with university bureaucracy.
I would also like to acknowledge my research colleagues for the discussions on everything from communications theory to ancient history. In particular I wish to thank Abdi Waheed Mohammed, Tuan Nguyen, Andrew Mancuso, Trung Nguyen, Olivia Hu (who never really saw the significance of being Dr ‘Who’, even after I brought in pictures of Tom Baker and the Tardis) and Melvyn Pereira for their company and generous natures.
Finally I wish to thank my parents and my brother for their support, both financially and emotionally, and my best friends Dan, Spencer, and Mark for the good times and social outlet.
Gavin Hill 7 February 2006
iii
.
..................................4..........4.....3 Impact on Bit Error Rate ...............2 Non linearities .............................................................................................6.......................................2 Impact on Power Spectral Density ......5 Limitations in OFDM ..............................................3......... i Acknowledgements .............. 13 2... 20 2.......................................................................................................7 Conclusion ....................... vii List of Figures ........................................................................................1 Synchronization .................... 61 v ..................................... 26 2...............1..................3..........................................11a .5................................................................................................. 6 2.......................................4..1 Block Codes .............5.................. 31 3...............................................................................................................................2 Statistical distribution of OFDM samples ...................................................................................................................... 29 3.......................................................... 20 2.................................................................................... 8 2.......... 18 2...............4............................................................... 28 3...... 57 4...............................................6................4 Effect of Non Linearity on OFDM ................1 Structure of thesis .............................1............................................................................1..................... 26 Chapter 3 Peak to Average Power in OFDM ......... 15 2.................................................................1 Description of memoryless Non Linearity .....5..............3 Oversampling discrete OFDM symbols to find true (continuous) peaks..... 22 2.......... 51 3..................................3 Digital Video Broadcasting................................. 24 2..........................................................................................................................................................................4 HiperLan2/802............................ 45 3.........1 Contributions................................ 6 2...............................................................................................................4.....4 OFDM transmission over time varying channels ........................................................................................................................................ 25 2..................................................................... 1 1.............................iii Contents ....................................6 Applications of OFDM .............. 23 2.....3 Frequency selective fading .....................5.................................................................................... 2 1....... 17 2.........1 History of multicarrier networks . xvi Chapter 1 Introduction.......................3 OFDM implementation of multicarrier modulation ............6................................................................... 35 3..................................................................6................................2 Carrier phase noise ................1 COFDM .......................1......1 Use of Fourier Transform for modulation and demodulation .................................................2 Orthogonality in OFDM ...........................................................viii Acronyms and Symbols .......... v List of Tables ..........................................3 Frequency errors ............. 22 2..........2 Digital Audio Broadcasting .....................................................................................................................4..............4 Equalization ..1 Multipath propagation .............................................................5........ 50 3...................................................................................6 MIMO OFDM ........ 18 2........ 22 2...................................2 Bounds on PAPR ........ 55 4.. 24 2............................1 Timing errors ............................................................................................2 Multicarrier principle ............1................................ 19 2................ 53 Chapter 4 Peak to Average Power Solutions Distortionless Techniques ..........................................Contents Abstract ...............................................................................6.........................1 Peak to Average Power Ratio ....................... 4 Chapter 2 Orthogonal Frequency Division Multiplexing ............................... 8 2...........5 Conclusion .1........ 45 3..........1 Coding techniques ..................... 19 2......................... 56 4.........5 ADSL ........................................................ 11 2................. 14 2.................................6.......
...............................................................................3.....6 Conclusion ........... 184 7.................... 112 6................................................................................. 66 4..........4 Conclusion ........................................... 87 4.....................................................2 New Sector Clipping method ............................. 77 4.....1 Future Work .................2.................................................2 Tone Injection ...................................................................................................................... 93 5....... 62 4........3 Filtering new techniques..... 200 Bibliography ..............2...........................4...........................................................4 Comparison of new and existing clipping methods ......................................... 168 7........................................................................................2 PSD results ....2.2......................... 196 Chapter 8 Conclusion ..............................3 Hardware Implementation ..........................................4.............................................................................. 158 7....... 121 6.................................................................1............................................................................. 83 4.........2............................................................... 132 7....................1...................................................................................2....1 Clipping in the Baseband . 93 5..........................1....................................................................................................... 158 7..1 Conventional clipping ......................4................................................................................1 Cyclic Shifted Sequences ......................................2 New techniques for PTS subblock creation ...........................................................1....................................................2.................................................1 Partial Transmit Signals ........................ 144 7.......................2...................................................................................1 PTS subblock creation ......... 99 5........ 146 7................ 134 7...................3 Selective Mapping .........4 Oversampling new techniques ............................................................................ 83 4.............................................................................................1..................3 Cyclic Codes ... 173 7...................................... 201 vi ..5 Theoretical Results ...........................1.....2 Multiple Signal Representation ...........................................1..............2.........4 Conclusion ......... 167 7.. 124 6.......... 198 8......6 ReedMuller Codes .......2......2 Extensions of Sector Clipping ............................ 63 4............2 Conventional clipping ..................................................1 Theoretical Analysis of Clipping Techniques ..................................................................2...................................................................................................5 Complexity evaluation ............ 108 5............................... 61 4. 109 Chapter 6 Peak to Average Power Solutions Distorted Techniques ......................5 New adaptive clipping method .......................................................... 96 5.......................1.................................................................................................................... 111 6............................................................................................................................ 92 5.1 Quantisation and Clipping .......................................................................................3..........................2........................................................................................................................................... 149 7..................3............................2.............................................................................................................................................................................4 ShapiroRudin codes .. 72 4.......1 BERF....................3 Tone Reservation/Injection ........... 187 7.....................3 Windowing ...............................................1......................................................... 104 5................1 Lucent Algorithm ....................3 Sector Clipping ................................... 167 7....... 162 7............2 Oversampling PTS ...... 173 7.................. 98 5..............2....................................... 160 7..............................................................3 New Vector Subtraction clipping method.......................................... 143 7........................1 SNR Analysis ................ 89 Chapter 5 Cyclic Shifted Sequences and Time Inversion of PTS .......... 63 4........1....... 130 Chapter 7 Reduced Complexity Clipping Algorithms ..3 Time Inversion .........3.............................................1 Tone Reservation ............................................ 66 4............2 Vector Subtraction .............2 PTS with CSS .........4 Square Clipping .6 Conclusion ........... 143 7......................2 Amplifier non linearities...............................5 Golay complementary codes ................. 128 6.. 92 5....................................
. . . 109 7. . . . 24 2. 183 vii .3: HIPERLAN2 parameters . . . . . .4: ‘Cadence’ area utilization report on 3 Sector clipping . . . . . .155 7. . . . . . . . 166 7. . . . . . . . . .25 4. . . . . . . . . . . . . . . . . . .1: Complexity comparison of various techniques (adjacent partitioning) . . . . . .1: Truth Table for 3 Sector clipping . . . . . . . . . . . . . . . . . . . 164 7. . . . . . . . . . . .1: DAB parameters . . .2: DVB system parameters for 2K mode .2: ‘Synopsys’ reports for Square and 3 sector clipping. . . . . . . .5: Baseband clip level required to maintain a BER=10 at varying IBO in HPA.4: Data rates for HIPERLAN2 . . . . . . . . . . . . . . . . . . . . . . . .165 7. . . 23 2. . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2. . . . .List of Tables 2. . . . . . . . . . . . . . . . . 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 5. . . . . . . . . . .3: ‘Cadence’ area utilization report on Square clipping . . . . .1: Number of complex multiplications and magnitude operations required . . . .
36 3. . . . 19 2. . N=64. . 4.List of Figures 2. . . . . . .2: Simulated (solid line) and theoretical (3. . 40 3. . .5: OFDM symbol a) without cyclic prefix. . . . . . . 2. . 16 2. . N=64. Occupied frequency band shown in between. . and 256 subcarriers. . . . . . . . . 128.6: Time and frequency properties of single carrier and OFDM techniques. . . . . .10: Zero padding of the IFFT. 33 3. . . . . 12 2. and b) with cyclic prefix. . . . . . .15. . . . . . . . . . . . . . . . . .2: Basic OFDM transmitter and receiver. . . . . . . . . . 16. α=0. 20000 runs. . .6: Simulated OFDM CCDF for oversampling rates of 1. .15) and (3. . . . . . . 30000 runs. . . . . 44 viii . . and 256 subcarriers. . . . . . . . 20000 runs. and 8. . . . 256 tap RRCF. . . . .9: Simulated OFDM CCDF QPSK. . . . 512 with simulated results: QPSK. . . . . . . . . 30000 runs.7: Effects of frequency offset F: reduction of signal amplitude (star). . . . . . 33 3. . and ICI (circle). . . . . . . The solid circle represents the 6dB level with respect to the average power. . . QPSK. . . .1: Block diagram of a basic multicarrier system . . . N=64. . 512. .7: Theoretical OFDM CCDF from (3. . . oversampling factor rate of 16. . . . .3: Simulated OFDM sample CCDF for N=32. . . . 256 tap RRCF. . . . . .22) for N=64. . . .13. . . . . . . . α=0. . . 15000 runs. . . . . N=64. QPSK. . . . . . and 64 constellation mapping. . . . . . . 14 2. .1: Simulated envelope for OFDM system (N=64) normalized by average power. . 9 2. . . . . . . . Clipped at 3dB after IFFT. . . . . .5: Simulated OFDM symbol with no oversampling (dashed) with it’s oversampled version (solid) overlaid on top. 43 3. 44 3. . . . . . . No clipping. . . . . . . . oversampling factor rate of 16. . . . . . . . . 37 3. . . . .15. . oversampling factors are 1 and 8. . 128. . . . . 34 3. . . . QPSK. 64. N=64.11: Simulated OFDM CCDF after IFFT (os=1 and 2) and after filtering (os=1 and 2). . N=64. . . . .8: Theoretical OFDM CCDF from (3. . . . . . . . . QPSK. . . 41 3.3: Basic OFDM transmitter and receiver pair utilizing Fourier transform. . . then filtered (solid line). . .21) for N=64. then filtered (dashed line). . dashed line) OFDM symbol CCDF for N=32. . .4: Frequency spectrum of 5 orthogonal subcarriers of an OFDM transmit signal . . 64. . . . . 30000 runs. . . 21 3. null carriers are set in the middle of the input. . . . . 36 3. . 10 2. . . . . 512. . . .4: Simulated OFDM CCDF for M=4. . . 64 point IFFT. . . . . . 15000 runs. . . . . . 512 with simulated results: QPSK. .
6dB IBO. . . . . . . . . . . . . . . . . . . . . . . . .15: Block diagram of a Tone Injection (TI) OFDM transceiver. . . . . .71 4. . . and 64 QAM constellations. . .79 4. . . . . 53 4. RRCF with excess bandwidth of 0. . . . . b) 4 QAM. . . . . . . . . . . . .47 3. . . . RRCF with excess bandwidth of 0. . . . . P=3 after non linear amplification a) 4 QAM. . . . . . . . . . 82 4. . .14: AM/AM properties of a Solid State Amplifier (SSPA) for different values of P. . c) 16 QAM 6dB IBO.15. . 81 4. . .13: a) Receiver structure of the proposed SLM system. a) serial form. A SSPA with P=3 and various backoffs is used. . . 78 4. . . .12: AM/AM properties of a Soft Limiter (SL). 56 4. . . . . .7: Block diagram of OLSPTS transmitter. . and 8. . . . . . . 68 4. . . . . . 51 3. . . . . . . . Oversampling rates are 1. 16. 81 4. . . . .71 4. . . . and d) 16 QAM 0dB IBO. . . . .16: Signal constellation at the output of the HPA.13: AM/AM properties of a Traveling Wave Tube Amplifier (TWTA). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .48 3. .17: BER of 64 subcarrier OFDM signal with 64 point IFFT. . . . . . . . .8: Block diagram of an SLM OFDM transmitter. . . . . . . . . . . . . . . . . . . . os=1. .2: Block diagram of PAPR reduction using the PTS approach. . . . . . . . . . . . . . . . . . .12: Scrambler polynomial for new SLM technique. . 76 4. .5: Simulated CCDF for PTSOFDM with W=4 and varying V. . . . . . . . . . . . . . and 8.3. . . . . . . . . . . . .4: Simulated CCDF for PTSOFDM with V=2 and varying W. . . . . . . . . . . . 80 4. . . . . . .9: Simulated CCDF for SLMOFDM for varying values of U. . . . . . . . . . . .1: Block diagram of OFDM transmitter showing PAPR coding. U=1. . . . . . 73 4. . . . 3. . . . . b) parallel form. . .14: Block diagram of a Tone Reservation (TR) OFDM transceiver. adjacent subblock partitioning. . . b) Descrambler polynomial at the receiver. . . . . 2. . 52 3. . .84 4. . . . . . . . .15 for 4. . . . . . N=64. N=64. . . . . N=64.15: PSD of 64 subcarrier OFDM signal with 64 point IFFT. . . . . . . . . 49 3. . . . . 70 4. .10: Simulated CCDF for SLMOFDM. . . . . . and ideal amplifier. . . . . . . . . . . . . . . . . . . . . . . . . . 87 ix . . . . 0dB IBO. . .3: An example of the 3 main PTS structures: (a) Interleaved (b) Adjacent (c) Pseudorandom. . . . . . . .11: SLM transmitter block diagram employing technique to avoid explicit transmission of side information. . . . . . . A SSPA with P=3 and various backoffs is used. . . . and 4. . . . . . adjacent subblock partitioning. . . .6: Generation of subblocks for PTS using Concatenated Pseudo Random Sub Block Partition Scheme (CPR SPS). . .
. Discrete oversampled curves (dashed) move from left to right and oversampled (solid) filtered curves move from right to left. . . adjacent partitioning. . . . . adjacent partitioning. PTS/TI (V=2. . . . . . . . .8: Simulated CCDF of discrete and filtered PTS (V=4. with no oversampling in IFFT (os=1). 102 5.3: Simulated CCDF for PTS (V=2. . . . CSS (V=2.9: Simulated CCDF of discrete and filtered PTS (V=2. W=8). interpolated by 8. 100 5. adjacent partitioning. interpolated by 8. . . . . . . V=4. with no oversampling in the IFFT. . . . . . . .2: Block diagram of PTS. . . adjacent partitioning. . . . . W=8). . . . . S=8). . .7: Simulated CCDF of discrete and filtered PTS (V=4. . S=2. . W=4. N=64. . with no oversampling in IFFT. . and uncoded OFDM. . . . . . . . . . . . adjacent partitioning. . (os=1). . . W=4. . . with no oversampling in the IFFT (os=1). . . . W=8. with no oversampling in IFFT (os=1). W=4. . . and 8. . S=4). . . filtered with RCF (α=0. . S=4).11: Simulated CCDF for PTS/CSS (V=2. . . . . . CSS (V=4. N=64. . N=64. V=4. .15). .15). . . . . . . . W=4). . and uncoded OFDM. . . . . 2. . . . . . . . . . N=64. filtered with RCF (α=0. . . . . . . PTS/CSS (V=2. . interpolated by 8. . (os=1). . . . . W=8.2. S=4.16: Example of possible expansions of constellation in TI for 16 QAM. W=4. W=4. . . W=4). . and filtered with RCF (α=0. . 103 5. .105 x . . and PTS/TI (V=2. S=2) and uncoded OFDM. . . . .93 5. . .98 5. . with no oversampling in IFFT. . . W=8). . . . S=4). . . . adjacent partitioning. V=4.6: Block diagram showing the simulation model of Section 5. . S=2). N=64. . . . . CSS and TI transceiver. . W=4. PTS/CSS (V=2. . . . . interpolated by 8. . V=2. . . .2. . . . . PTS/CSS (V=2. . CSS (V=4. adjacent partitioning. . and Uncoded OFDM.99 5. . with oversampling rates in the IFFT of 1. . . S=8. and uncoded OFDM. . . . (os=1). . . interpolated by 8. . . .1: Block diagram showing the simulation model of Section 5. . . . N=64. and uncoded OFDM. . . . S=2). . . .4: Simulated CCDF for PTS (V=2. . . .15). . . . W=4. . filtered with RCF (α=0. . . . . W=4.10: Simulated CCDF of discrete and filtered PTS (V=3. . . . adjacent partitioning. filtered with RCF (α=0. N=64. . . . S=4). . . . .15). W=4). . . . S=8) and uncoded OFDM. . . . with no oversampling in IFFT (os=1). .95 5. . . . . . . 101 5. 96 5. . . . . . 4.4.5: Simulated CCDF for PTS (V=2. . 104 5.88 5. . W=8). . . V=4. . . . . . . . . . and uncoded OFDM. . . N=64. V=3.15). . . .
. . α=0. . . . .15).5. with oversampling rates in the IFFT of 1. . . .5b: Demapped constellation. N=64. 107 5. . 2. . 64 QAM symbols. . . . . . 134 7. . . . .15). . W=4. and Uncoded OFDM. . . . and Uncoded OFDM. . S=2). . . . . . .4: Baseband clip level vs the BERF for varying RRCF parameters. . . M=64.15: Simulated CCDF for PTS/TI (V=2. . . . . . . . and 4.2: Probability of clipping DMT signal as a function of µ for p=1. . and 64 QAM . . . . . . 4. . . and 4. BER for 4. and filtered with RCF (α=0. .15).5a: Demapped constellation. . . . . . 2. . . with oversampling rates in the IFFT of 1. 137 7. . . . . . 2. 64 point IFFT. . and filtered with RCF (α=0. 135 7. . .3. . Discrete oversampled (dashed) curves move from left to right and oversampled (solid) filtered curves move from right to left. . . . . . .15. . . . . AWGN=0 . 107 5. . . . . . . Discrete oversampled curves (dashed) move from left to right and filtered (solid) oversampled curves move from right to left. . W=4). . . . . . .10) and (6. . with no clipping or channel impairments. . .15. . . . N=64. . N=64. adjacent partitioning. .14: Simulated CCDF for PTS/CSS (V=2. . and filtered with RCF (α=0. . . . . and filtered with RCF (α=0. 108 6. W=4. . . . and Uncoded OFDM. Discrete oversampled curves (dashed) move from left to right and filtered (solid) oversampled curves move from right to left. . . . . . 117 6. 16. with oversampling rates in the IFFT of 1. 123 7.15). . . adjacent partitioning. . 128 filter taps in RRCF. . . N=64 subcarriers.12: Simulated CCDF for PTS (V=3. 64 QAM xi .137 7.6: Baseband clip level vs the BERF with varying IBO in HPA. . . . LPA. . . . . . . . . adjacent partitioning. . interpolated by 8. . no channel impairments.1: Average noise in the channel vs. . .13: Simulated CCDF for PTS (V=2. . . . . . and Uncoded OFDM. 133 7. . N=64. . α=0. . and 4. . . . . .1: Analytical symbol error probability from (6. . interpolated by 8.2: IQ diagram showing conventional clipping region. . . . . . . 106 5. . . . . . . .3: Block diagram of simulation model used for clipping models.12) for 64 QAM and N=64. . M=64. . W=8). . . S=2). . . interpolated by 8. . . with no clipping or channel impairments. . . . . . . adjacent partitioning.2. Discrete oversampled curves (dashed) move from left to right and filtered (solid) oversampled curves move from right to left. 64 filter taps in RRCF. . . . . . . . and 8. 2. . . . 136 7. interpolated by 8. with oversampling rates in the IFFT of 1.
. . . 163 7. . . .16: Theoretical Clip level vs. . 144 7. . . . . . . . . . . . . .24: IQ plane for Vector Subtraction showing vector of sample being clipped. . . . AWGN=0. Theoretical (dashed). . . 0. . . . . .11: I Q diagram showing different sector clipping regions and the direction of data reduction for 3 Sector Clipping. . . . . . . . . . . . . . . . . . .19: Simulated clip level vs. . .10: Baseband clip level vs the BERF with varying p in the SSPA. . . 139 7. . 139 7. . . . 162 7. . . . . . 141 7. . . 150 7. . 128 point IFFT (os=2). . . . . . . the SNR for 3 Sector clipping. . . . . . . . . . . . 128 point IFFT (os=2). . 4. . . . .21: Block diagram of 3 Sector clipping implemented in VHDL. . 3. 7. 161 7. . . . . . . . . . . . θ. . . . . . . . . . . . . . . . 64 QAM symbols. . . . .18: I Q diagram showing the 1st quadrant Sector clipping regions of a) 4 Sector clipping and b) 5 Sector clipping. . . . . . . . .7: Baseband clip level vs. . . . . HPA backoff set to 8dB for 64 and 128 IFFT. . SNR for Conventional. 128 point IFFT (os=2). HPA backoff set equal to baseband clipping level. . . . . .23: Block diagram of new Vector Subtraction scaling operation.20: Design flow for silicon implementation of 3 Sector clipping. . . .169 7. .160 7. . . . . . . . . . . . . . .170 7. . . . .13: I Q diagram of 1st quadrant of a 3 Sector clipping system showing the vector of an unclipped and clipped sample.15. 156 7. . . . . . . . the BERF with varying IBO in HPA. . . . . . . .15.17: Clipping angle.15. . . SNR for Conventional (Standard). . . and 5 Sector Clipping. simulated (solid) . RRCF with 128 taps and α=0. . . . . . . 64 point IFFT (os=1). . . 159 7. . RRCF with 128 taps and α=0. . . . . . . . . 3 Sector. . . . . . . . . . . RRCF with 128 taps and α=0.22: Application Specific Integrated Circuit (ASIC) view of 3 Sector clipping algorithm implemented in Silicon using ‘Cadence Silicon Ensemble’. . . . . . . . . . . 144 7. . . . . .14: Block diagram of 3 Sector clipping. . . . . 142 7. . . .41). . . . . . . . . . . RRCF with 128 taps and α=0. . . . . . . . . .symbols. . 3 metal layers. . . .9: Baseband clip level vs the BERF with varying Mary constellations. 166 7. . based on (7. . .15: Flowchart for the LUT in Figure 7. . . . . .25: Simulated CCDF for Lucent patent [101] for various iterations clipped 140 xii . 157 7. . . . . . . . . . . . . . . . . .5µ process. . . . . Square. . . .8: CCDF clipped in baseband at 5dB. . and Square clipping.15. . . . vs. . . .12: Input output relationship of clipping operation. . . . 163 7. . . . .14 (3 Sector clipping) .
vs. . . 128 point IFFT (os=2).15. . 175 7. . . 64 QAM symbols. . . . . . . . AWGN=0. . 177 7. . vs. . SNR for Lucent clipping technique with varying iterations. AWGN=0 . . . . . . . . 171 7. . RRCF with 128 taps. Conventional and Square clipping vs. . . . . .at 5dB showing the leakage of under clipped samples.34: Simulated 5 Sector clipping. . .32: Simulated 3 Sector clipping. . . RRCF with 128 taps. . . . . . . . . . . . RRCF with 128 taps and α=0. . .179 7. . . . . . 64 QAM symbols.15. 128 point IFFT (os=2). . . . . . . RRCF with 128 taps and α=0. . . . AWGN=0 . . . . SNR for Vector Subtraction clipping with varying iterations. 64 QAM symbols. . 64 QAM symbols. 64 QAM symbols. .30: Simulated Vector Subtraction (1. and 5 Sector. 174 7. BERF with a LPA. . 177 7. . . 128 point IFFT (os=2). . . BERF with a LPA. . . . . . . 4. . . . 3. . 128 point IFFT (os=2). . . . . 128 point IFFT (os=2). .15. 16 QAM symbols. . 172 7.171 7. . 128 point IFFT (os=2). .26: Simulated CCDF for Vector Subtraction for various iterations clipped at 5dB showing the leakage of under clipped samples. 128 point IFFT (os=2). . . . 128 point IFFT (os=2). . . . . as well as Conventional and Square clipping. . . . . . RRCF with 128 taps and α=0. . . . . . . . . . . . . . .15. . . . . . . . . . AWGN=0.15. 2. RRCF with 128 taps. . . . . . . . . . BERF with varying IBO in HPA. . . . . . . . . 176 7. 179 7.15. . . . RRCF with 128 taps and α=0. and α=0.27: Simulated clip level vs. . . . as well as Conventional and Square clipping. . . . . Conventional and Square clipping vs. .33: Simulated 4 Sector clipping. . . . and α=0.37: Simulated Lucent clipping (1 iteration). . . . BERF with varying IBO in HPA 64 QAM symbols. . . . and α=0. . .15. and 4 iterations). . . . AWGN=0 . RRCF with 128 taps and α=0. Conventional and Square clipping vs.31: Simulated Lucent [101] clipping (1. . vs. . . AWGN=0 . . . . 2. . . 181 xiii . . . BERF with varying IBO in HPA. . . . . . . . . . BERF with a LPA. . BERF with varying IBO in HPA. and α=0. . . . . .15. . vs. 128 point IFFT (os=2). . .29: Simulated 3. . . . .36: Simulated Vector Subtraction (4 iterations). 3. AWGN=0 . . 173 7. . . 178 7. AWGN=0. BERF with varying IBO in HPA. . . vs. 64 QAM symbols.28: Simulated clip level vs. . RRCF with 128 taps. . . . . 64 QAM symbols. .15. . AWGN=0 . .35: Simulated Vector Subtraction (1 iteration). . . . vs. . BERF with varying IBO in HPA. . and 4 iterations). .
and α=0. AWGN=0. . . . . . . 128 taps. RRCF. . LPA). . . . . .40: PSD 0dB clipping in baseband with increasing amplifier backoffs (HPA=CL. 185 7.CL+4. .52: Clipping error vs. . . .CL+3. . 64 QAM symbols. . . . BERF with varying IBO in HPA. . .CL+3. Performance curves for LDA using Conventional clipping and Vector Subtraction with 2 xiv . . . . . . . CL+4. . . CL+1. LPA). . RRCF with 128 taps. . ACL=8dB.15. . . . .7. . . . . RRCF with 128 taps. . . . .39: Simulated Square clipping vs. . . RRCF. . . . . . .CL+3. . LPA). . vs. . . . . . . . alpha=0. CL+1. 189 7. . . .50: In band distortion for Conventional and Vector Subtraction (2 iter) with 64 and 128 taps in compensation filter. . . . . . . . .38: Simulated Lucent clipping (4 iterations). . . . .45: Detailed block diagram of the LDA. . . . RRCF. . . . . . .CL+2. . inband distortion. . . . . 128 taps. .42: PSD after 0dB clipping in baseband and receiver filtering. . . . Compensation filter has 64 and 128 taps. . . . .15. . . . LPA).41: PSD 5dB clipping in baseband with increasing amplifier backoffs (HPA=CL. . . . . . . . . . . . . . . . . . alpha=0. 195 7. .CL+2. 186 7. . . . 128 point IFFT (os=2). . . . . . . . . . . CL+1. . . . . . . .49: In band distortion for Conventional and Vector Subtraction (2 iter) with 64 and 128 taps in compensation filter. . ACL=5dB. .CL+4. .192 7. . . . . . 181 7. . . . CL+1. . .189 7. with increasing amplifier backoffs (HPA=CL. . AWGN=0. . . . Performance curves for LDA using Conventional clipping and Vector Subtraction with 2 iterations. . . . . . . . . . 128 taps. .CL+4.48: Block diagram of the simulation model used to evaluate LDA. . . 191 7. . . . . . . . . . . . . and α=0. . . . . . . 193 7. . . 128 taps. . . . . .15.15. .46: a) Amplitude of zero padded input to filter b) amplitude of filtered output c) LDA correction vectors. . . . . . . . . .15. . . . . . . . alpha=0. . . RRCF.44: Block diagram of the Level Detection Algorithm (LDA).186 7. . . . . .CL+3. . BERF with varying IBO in HPA.15. . . . 182 7. ACL=5dB. alpha=0. . . .43: PSD after 5dB clipping in baseband and receiver filtering. . 185 7. . .47: Vector representation of filtering with 3 active taps and the required correction vector to bring the output back to CL. .51: Clipping error vs. . .CL+2. . . . 193 7. . . . 190 7. .CL+2. . . . . . . . 64 QAM symbols. . . . with increasing amplifier backoffs (HPA=CL. . . . . 128 point IFFT (os=2). . inband distortion. . .
iterations. Compensation filter has 64 and 128 taps. . . . 195 xv . ACL=8dB.
Acronyms and Symbols ADC ACI ADSL AGC AM/AM AM/PM AWGN BER BPSK BSLM BW CCDF CCOFDM CDF CDMA CF CL COFDM CP CPRSPS CSS DAB DAC DAR DFT DMT DQSK DRL DSP DVB EDGE Analog to Digital Converter Adjacent Channel Interference Asymmetric Digital Subscriber Line Automatic Gain Control Amplitude Modulation to Amplitude Modulation Amplitude Modulation to Phase Modulation Additive White Gaussian Noise Bit Error Rate Binary Phase Shift Keying Blind Selected Mapping Bandwidth Complementary Cumulative Density Function Combined Coded OFDM Cumulative Density Function Code Division Multiple Access Crest Factor Clip Level Coded Orthogonal Division Multiplexing Cyclic Prefix Concatenated Pseudo Random Subblock Partition Scheme Cyclic Shifted Sequences Digital Audio Broadcasting Digital to Analog Conversion Decision Aided Reconstruction Discrete Fourier Transform Discrete MultiTone Differential Quadrature Phase Shift Keying Data Rate Loss Digital Signal Processing Digital Video Broadcasting Enhanced Data rates for Global Evolution xvi .
FEC FDM FFT FPGA GSM Forward Error Correction Frequency Division Multiplexing Fast Fourier Transform Field Programmable Gate Array Global System Mobile HDDIVINE High DefinitionDigital Video Narrowband Emission HIPERLAN2 HiPERformance Local Area Network version 2 HF HPA IBO ICI IDFT IEEE IFFT ISI LAN LDA LP LPA LUT MIMO MLD MMSE MSE MSR MQAM OBO OBR OLS PAP PAPR PDF PEP PICR High Frequency High Power Amplifier Input BackOff Inter Channel Interference Inverse Discrete Fourier Transform Institute of Electrical and Electronic Engineering Inverse Fast Fourier Transform Inter Symbol Interference Local Area Network Level Detection Algorithm Linear program Linear Power Amplifier Look Up Table Multiple Input Multiple Output Maximum Likelihood Detection Minimum Mean Squared Error Mean Squared Error Multiple Signal Representation Mary Quadrature Amplitude Modulation Output BackOff Out of Band Radiation Optimal Limited Search Peak to Average Power Peak to Average Power Ratio Power Density Function Peak Envelope Power PeakIntercarrierto Carrier Interference xvii .
PMEPR PRT PSD PSK P/S PTS QAM QCQP QPSK RF RMS RRCF RS SBC SC SER SES SL SLM SNR S/P SSPA STERNE SVD TCM TI TR TWTA VHDL VLSI WCDMA WLAN N Peak to Mean Envelope Power Ratio Peak Reduction Tones Power Spectral Density Phase Shift Keying Parallel to Serial Partial Transmit Sequences Quadrature Amplitude Modulation Quadratically Constrained Quadratic Program Quadrature Phase Shift Keying Radio Frequency Root Mean Square Root Raised Cosine Filter ReedSolomon Sub Block Coding Single Carrier Symbol Error Rate Suboptimal Exhaustive Search Soft Limiter Selected Mapping Signal to Noise Ratio Serial to Parallel Solid State Power Amplifier System de Television En Radiodiffusion NumeriquE Singular Value Decomposition Trellis Coded Modulation Time Inversion Tone Reservation Travelling Wave Tube Amplifier Visual Hardware Design Language Very Large Scale Integration Wide Band Code Division Multiplexing Wireless Local Area Network Number of subcarriers nth subcarrier in OFDM symbol xviii n .
n mth continuous OFDM symbol nth discrete sample of mth OFDM symbol Total occupied frequency bandwidth of OFDM symbol Frequency separation between subcarriers Total OFDM symbol duration Duration of one sample in OFDM symbol kth mapped transmit sample of mth OFDM symbol kth mapped received sample of mth OFDM symbol Length of Cyclic Prefix Number of taps in equalization algorithm Mean envelope power Average Power of an OFDM symbol Peak to Average Power Ratio of OFDM Crest Factor of OFDM mth passband OFDM symbol Probability density function of an OFDM symbol Transfer properties of HPA Minimum code distance Code distance Code rate 2nd order ReedMuller codes Phase rotation of PTS subblock Number of phase rotations in PTS Number of Subblocks Oversampling rate in IFFT Number of alternate SLM transmit symbols Cyclic shift in time domain for TI Autocorrelation of the input Cross correlation Autocorrelation of the output Roll off factor of pulse shaping filter xix W ∆f T Ts X m. k Ym . m) pm .k Ng TE P Pav ζ ζ CF xmPB Pζ (ζ ) n F [ρ] d min d(Ci.xm ( t ) xm . k W V os U ∂v Rxx Rxy Ryy α . Cj) Rc RM(2.
R SNRconv SNR3sec d Clip level SNR of conventional clipping SNR of 3 Sector clipping data stream xx .
promising raw data rates of between 6 and 54Mbps. One new modulation scheme which has received significant attention over the last few years is a form of multicarrier modulation called Orthogonal Frequency Division Multiplexing (OFDM). and maximum delay. New modulation schemes are required to transfer the large amounts of data which existing 3rd generation schemes such as Global System Mobile (GSM). the main advantage is OFDM’s immunity to frequency selective fading. OFDM has various properties that make it desirable over existing single carrier systems. g standard in the US. 1 . for that is where you and I are going to spend the rest of our lives. OFDM has been used for Digital Audio Broadcasting (DAB) and Digital Video Broadcasting (DVB) in Europe. Intro to "Plan 9 From Outer Space" (1958) With the advance of communications technology comes the demand for higher data rate services such as multimedia. its enhanced version Enhanced Data Rates for Global Evolution (EDGE)." Criswell. And remember my friend.Chapter 1: Introduction Chapter 1 Introduction "We are all interested in the future. future events such as these will affect you in the future. allowable Bit Error Rate (BER).11a. voice. and for Asymmetric Digital Subscriber Line (ADSL) high data rate wired links. and Wideband Code Division Multiple Access (WCDMA) cannot support. OFDM has also been standardized as the physical layer for the wireless networking standard ‘HIPERLAN2’ in Europe and as the IEEE 802. and data over both wired and wireless links. These new modulation schemes must be able to act over point to point links and in broadcast mode. support bidirectional communications. and be able to adapt to different requirements of individual services in terms of their data rate.
Perhaps the most serious problem is the reduced efficiency of the High Power Amplifier (HPA) which must cater for these low probability large peaks. As portable devices have a finite battery life it is important to find ways of reducing the PAPR allowing for a smaller more efficient HPA. Despite the many advantages of OFDM it still suffers from some limitations such as sensitivity to carrier frequency offset and a large Peak to Average Power Ratio (PAPR). reducing complexity greatly. The large PAPR is due to the superposition of N independent equally spaced subcarriers at the output of the Inverse Fast Fourier Transform (IFFT) in the transmitter. If the high PAPR is allowed to saturate the HPA out of band radiation is produced affecting adjacent channels and degrading the BER at the receiver. Other factors such as advances in silicon and Digital Signal Processing (DSP) allow the use of efficient Fourier transforms in the transmitter and receiver to perform the modulation. OFDM overcomes this problem by dividing the wideband channel into a series of narrowband channels which each experience flat fading. The thesis is structured as follows: 2 . 1. thereby increasing the occupied bandwidth. A large PAPR is a problem as it requires increased complexity in the wordlength at the output of the IFFT and the Digital to Analog Converter (DAC).Chapter 1: Introduction Single carrier systems can increase their data rate by shortening the symbol time. demodulation respectively.1 Structure of thesis This thesis analyses the principles of OFDM concentrating on the PAPR problem in OFDM. Therefore only 1 tap equalizers are required in the receiver. Wideband channels are sensitive to frequency selective fading which require complex equalizers in the receiver to recover the original signal. which in turn will mean a longer lasting battery life. Due to the orthogonality of the subcarriers the transmission bandwidth is used efficiently as the subcarriers are allowed to overlap each other and still be decoded at the receiver.
and frequency errors.Chapter 1: Introduction • Chapter 2 provides an initial overview of OFDM starting with a brief history of multicarrier networks and their evolution towards OFDM. Initially the family of coding techniques such as block codes. • Chapter 3 explores the issue of the PAPR in more detail starting with a mathematical definition of the PAPR. and Reed Muller codes are presented. and equalization. • Chapter 5 introduces several new alterations to PTS called Cyclic Shifted Sequences (CSS) and Time Inversion (TI). Golay complementary sequences. Problems with OFDM are also discussed such as synchronization. ShapiroRudin Sequences. Finally modified constellation techniques Tone Reservation (TR) and Tone Insertion (TI) are examined. carrier phase noise. OFDM in time varying channels is discussed. PTS produces alternative transmit signals by dividing the bit source into a V subblocks which each have an IFFT performed on them. The issue of the PAPR is also briefly presented. the use of a cyclic prefix. Subblocks are then rotated by a set phase rotation 3 . frequency selective fading. The multicarrier principle is explained mathematically encompassing the use of the Fourier transform and the principle of orthogonality. Non linearities are treated with a description of various models for the HPA. Multiple representation techniques such as SeLective Mapping (SLM). Distortionless techniques do not corrupt the data and encode it in such a way that it can be completely recovered at the receiver. Theoretical Cumulative Complementary Distribution Function (CCDF) results are compared to simulated CCDF results identifying the processes which influence large peaks such as the number of subcarriers and oversampling. • Chapter 4 begins the literature review for PAPR reduction techniques reviewing distortionless techniques. Partial Transmit Sequences (PTS) are reviewed with and without oversampling. and finally the effect of saturation of the HPA is analyzed in terms of the PSD and the BER. its advantages in terms of multipath propagation. The chapter concludes with a discussion on the applications of OFDM in society. which includes timing errors. cyclic codes.
Chapter 1: Introduction (which must be sent as side information to the receiver) and combined to produce a possible transmit symbol. This work is published in [3]. New low complexity clipping methods are introduced which avoid complex hardware operations while maintaining similar performance to conventional clipping. It is further shown that CSS and TI perform better after oversampling and filtering than PTS. Methods reviewed are pulse shaping (or windowing).1 Contributions The contribution that this research work has made to the wireless communications field is summarized as follows: 4 . and the HPA parameters. • This work is Chapter 6 continues the literature review for distorted PAPR reduction techniques which do not attempt to create a transmit signal with a low crest factor. 1. Results are • Chapter 7 introduces new low complexity clipping techniques starting with a comprehensive analysis of an OFDM transceiver with clipping at various points in the transmission chain and under other variable conditions such as the amount of oversampling in the IFFT. after a number of set phase rotations the transmit symbol with the lowest PAPR is chosen for transmission. published in [1. instead they take the output of the IFFT and then limit the amplitude of large samples which invariably causes distortion degrading the BER.1. pulse shaping taps and roll off. CSS and TI reduce complexity and improve performance of PTS by using time shifts of the data instead of phase rotations which can be combined with standard PTS to reduce complexity and in some cases allow for the removal of a whole IFFT operation without degrading performance. 2]. and clipping at every stage from the output of the IFFT to limited backoffs in the amplifier. analyzed in terms of their BER and affect on the PSD. The new clipping algorithms called Sector clipping and Vector Subtraction are then implemented in a new clip and filter algorithm which is much less susceptible to peak regrowth after baseband filtering.
and techniques of OFDM based wireless systems including a detailed analysis of PAPR reduction techniques are presented (Chapters 2. • • Several new low complexity clipping algorithms are proposed (Chapter 7). and 6). 3. principles. Implementation and analysis of a proposed clip and filter algorithm utilizing one of the new low latency clipping algorithms (Chapter 7).Chapter 1: Introduction • A thorough analysis of the theory. • A detailed analysis of the effect of clipping on an OFDM transceiver under various system conditions (Chapter 7). • A new method for producing Partial Transmit Sequences (PTS) signals and their performance under oversampling conditions is proposed (Chapter 5). 5 . 4.
and equalization. detailing the use of the Fourier transform. Section 2. Finally Section 2. however due to their implementational complexity and inefficient use of the frequency band they were restricted to military applications. Section 2. which includes timing errors.1 History of multicarrier networks Multicarrier networks such as Frequency Division Multiplexing (FDM) have been around since the late 1950’s [4]. frequency selective fading.3 explains how it is applied to OFDM.6 looks at applications of OFDM in society and discusses where this new communications technology will be used.5.4 explores OFDM in time varying channels describing its advantages in terms of multipath propagation. carrier phase noise.1 provides a brief history of multicarrier networks and their evolution towards OFDM. 2.7 summarizes the chapter with a brief recap of the chapter. and the importance of orthogonality. and frequency errors are discussed in Section 2. the use of a cyclic prefix. Limitations of OFDM.Chapter 2: Orthogonal Frequency Division Multiplexing Chapter 2 Orthogonal Frequency Division Multiplexing This chapter provides an initial overview of Orthogonal Frequency Division Multiplexing (OFDM). Section 2. such as synchronization. Section 2. A multicarrier system is basically a number of information bearing carriers 6 . Non linearities are also introduced as a major hindrance to a practical OFDM system in this section.2 explains the multicarrier principle and Section 2.
hence OFDM. This has led to the adoption of OFDM in many European standards. In the last 10 years more advances in practical OFDM systems have been made. particularly in Europe where various projects and prototypes were initiated such as DIgital VIdeo Narrowband Emission (HDDIVINE). frequency selective fading. and digital Terrestrial Television broadcasting (dTTb). System de Television En Radiodiffusion NumeriquE (STERNE). and the Peak to Average Power Ratio (PAPR). utilizing the frequency spectrum more efficiently.Chapter 2: Orthogonal Frequency Division Multiplexing transmitted in parallel. Chang [5] and Saltzberg [6] further developed FDM in the mid 60’s by introducing multiple carriers which overlap in the frequency domain without interfering with each other. In the 1970’s Weinstein and Ebert [7] used an Inverse Discrete Fourier Transform (IDFT) and Discrete Fourier Transform (DFT) to perform the modulation and demodulation respectively. However the complexity issue still remained. It has also been adopted as the physical layer modulation scheme for wireless networking standards such as Hiperlan2 in Europe and the Institute of Electrical and Electronic Engineers (IEEE) 802. exploiting the sinusoidal nature of the Fourier Transform and significantly reducing the complexity of an OFDM system. OFDM has progressed to the point where it has now been used for various communication applications such as Digital Audio Broadcasting (DAB) and Digital Video Broadcasting (DVB) in Europe. g standards in the United States.11a. 7 . Multicarrier systems in wireless applications are less susceptible to channel induced distortions than single carrier systems at corresponding data rates. However while OFDM successfully alleviates the problem of dispersive channels there are still some problems which need to be addressed such as time and frequency synchronization.
In OFDM the transmit signals are constructed in such a way that the frequency spectra of the individual subchannels are allowed to overlap thereby utilising the frequency spectrum much more efficiently.4.3 OFDM implementation of multicarrier modulation A more spectrally efficient implementation of the aforementioned multicarrier system is OFDM (Figure 2. The basic principle of multicarrier modulation is to divide the data stream. d.2 is 1 xm ( t ) = N N ∑ − N +1 2 2 X m . into N parallel data streams with a reduced data rate of d/N. Each subchannel will therefore experience flat fading reducing the equalization complexity in the receiver dramatically.wk ( t − mT ) 0>t >T (2.Chapter 2: Orthogonal Frequency Division Multiplexing 2.1) 8 .2 Multicarrier principle An early form of a multicarrier system is shown in Figure 2. The advantage of this structure over single carrier systems is that the extended symbol time (due to lower data rate) makes the signal less susceptible to effects of the channel such as multipath propagation which introduces Inter Symbol Interference (ISI). 2. thereby providing the same data rate as an equivalent single carrier system. This issue will be explored in more depth in Section 2. A disadvantage of the method shown in Figure 2. Mathematically the continuous time representation of the OFDM transmit signal depicted in Figure 2.k .e j 2π k ∆ft .1. Each low rate data steam is then modulated on a separate narrow band subcarrier and summed together for transmission.1 is the implementation complexity due to the large number of filter banks required in the transmitter and receiver as well as the inefficient use of the available frequency band [8].2). At the receiver a set of filter banks separate the wideband signal into the original narrowband subcarriers for demodulation. The spectra of the different carriers cannot overlap as this would introduce distortion degrading system performance.
1: Block diagram of a basic multicarrier system.Chapter 2: Orthogonal Frequency Division Multiplexing d/N b/s QAM d/N b/s Filter Filter Filter d0(t) d1(t) dN1(t) f0 RF QAM d/N b/s f1 fN1 QAM Transmitter f0 f1 Filter f0 f2 QAM f0 RF Filter f1 QAM f1 Filter fN1 QAM fN1 Receiver Figure 2. 9 .
Chapter 2: Orthogonal Frequency Division Multiplexing X0 f0 X1 Xn QAM R b/s Serial to parallel converter x (t ) f1 . . Occupied frequency band shown in between. 10 . . X N −1 . . RF Timelimited signal (block processing) N −1 x(t ) = Re ∑ X n e xp(− jω n t ) n =0 f N −1 Transmitter f0 f1 f2 ∫ f3 ˆ X0 f4 f0 RF ∫ ˆ X1 ˆ X N −1 Parallel to serial convertor QAM f1 ∫ fN1 Receiver Figure 2.2: Basic OFDM transmitter and receiver. .
A discrete time representation of (2. N T In the receiver an integrate and dump operation is performed over time T to recover the data.3 shows a block diagram of a basic OFDM system in the baseband utilising the IDFT.n = 1 N ∑ X m. To ensure the orthogonal relationship between subcarriers ∆f is set as W 1 = (W is the total bandwidth of the signal). e j 2π k ∆ft is the kth subcarrier. This equation describes exactly the IDFT operation.2) where ‘n’ are the discrete sampling points.2) N j 2π nk xm .k .e k =0 N −1 N = IDFT { X m . as first suggested in reference [7]. k is the mapped (QAM.3.1) can be obtained by sampling the continuous signal.Chapter 2: Orthogonal Frequency Division Multiplexing where X m. N is the number of subcarriers. the Inverse Fast Fourier Transform (IFFT) and Fast Fourier Transform (FFT) is used for the modulation. demodulation respectively. Under this condition (2. DFT pair. 2. Advances in silicon technology have made the production of the DFT more cost efficient [911]. Under the condition that W = N ∆f and ∆f = by its samples if sampled at t = 1 the signal can be determined T T .1) then becomes (2. and wk ( t − mT ) is a rectangular window applied to each subcarrier. and T is the total time of the transmit symbol.k } 0 ≤ n ≤ N −1 (2. etc) data to be transmitted on the kth subcarrier of the mth transmitted symbol.1 Use of Fourier Transform for modulation and demodulation In order to make multicarrier systems a more practical technology an IDFT and DFT are used for the baseband modulation and demodulation respectively. 11 . PSK. Figure 2. In hardware the more efficient form of the IDFT and DFT. ∆f is the frequency spacing between subcarriers. where N is set to be a power of 2. where the sinusoidal nature of the Fourier transform basic functions is exploited.
QAM.Ts. X0 to XN1.3 shows a baseband transceiver structure for OFDM utilising the Fourier transform for modulation and demodulation. Here the serial data stream is mapped to complex data symbols (PSK. Assuming the incoming complex data is random it follows that the IFFT is a set of N independent random complex sinusoids summed together.3: Basic OFDM transmitter and receiver pair utilizing Fourier transform. etc) with a symbol rate of 1 Ts . A Cyclic Prefix (CP). The samples. resulting in complex samples x0 to xN1. The data is then demultiplexed by a serial to parallel converter resulting in a block of N complex symbols. Data source Constellation mapping Serial to Parallel X0 X1 IFFT XN1 x0 x1 Parallel to Serial CP xN1 Channel + noise Data sink Constellation demapping Parallel to Serial Y0 Y1 FFT YN1 y0 y1 Serial to Parallel CP yN1 Complex data Binary data Figure 2.Chapter 2: Orthogonal Frequency Division Multiplexing Figure 2. x0 to xN −1 are then converted back into a serial data stream producing a baseband OFDM transmit symbol of length T=N.Ts. which is a copy of the last part of the samples is appended to the front of the serial data stream before Radio Frequency (RF) up conversion and 12 . The parallel samples are then passed through an N point IFFT (in this case no oversampling is assumed) with a rectangular window of length N.
ne n =0 N −1 − j 2π nk N ∑∑ X n =0 d =0 N −1 d =0 m.3) Ym .Chapter 2: Orthogonal Frequency Division Multiplexing transmission. d N δ [ d − k ] N d =0 = X m.4.Ts is symbol duration..4 shows the frequency response of a 5 carrier system where it is seen that because of the orthogonal relationship the maximum of a particular sample corresponds to a null in all other carriers.3) N 1 = N = ∑ X ∑e N −1 1 ∑ X m. Raised Cosine Filter) reduce 13 . the CP is removed prior to the FFT which reverses the effect of the IFFT.2.d N −1 n =0 e j 2π n ( d − k ) N (2. The N equally spaced subcarriers will be orthogonal if the frequency separation between subcarriers is ∆f = 1 N ⋅ Ts = 1 . The CP combats the disrupting effects of the channel which introduce Inter Symbol Interference (ISI) and is discussed in more detail in section 2.2 Orthogonality in OFDM One of the key advantages of OFDM is its efficient use of the frequency band as the subcarriers are allowed to overlap each other in the frequency domain. therefore eliminating the effects of interference. The complex symbols at the output of the FFT. and rectangular windowing of T the IFFT is performed. d N −1 N −1 j 2π n( d − k ) m. YN1 are then decoded and the original bit steam recovered.3.k = FFT { xm . Smoother window functions (eg. In the receiver the whole process is reversed to recover the transmitted data. Under these conditions the subcarriers will have a sinc waveform frequency response. Figure 2. Y0 . where N.k 2. Mathematically the demodulation process (assuming no CP and no channel impairments) using the FFT is (2.n } = = 1 N 1 N ∑ xm.
g.4: Frequency spectrum of 5 orthogonal subcarriers of an OFDM transmit signal 2.4 6 4 2 0 2 4 6 1/NTs Figure 2. k = l (2.4 0. 14 . NTs 0 ψ k ( t )ψ l∗ ( t ) dt = 0.4) Carrier Spectrum with no carrier offset 1 0.Chapter 2: Orthogonal Frequency Division Multiplexing the out of band emissions and Inter Carrier Interference (ICI) susceptibility to system imperfections (e. k ≠ l C . Mathematically. This section describes properties of the wireless channel and describes the advantages and disadvantages of OFDM in this environment.2 0 0.4) ∫ where C is a constant.Ts is described in reference [12] and expressed here as (2.8 0.6 0. orthogonality of two signals.2 0.4 OFDM transmission over time varying channels OFDM is being primarily deployed in the wireless environment. ψ k ( t ) and ψ l ( t ) over time period N. frequency offset) but they increase the symbol period.
2. The signal is reflected of buildings and mountains and other obstacles so that multiple delayed copies of the same transmitted signal arrive at the receiver affecting other symbols. but they will be even more affected by the rms delay spread and require more complex equalisation in the receiver.4.2 Use of a Cyclic Prefix In order to protect successive OFDM symbols from multipath a CP of length Ng is used which is a copy of the last part of the samples of a OFDM transmit block appended to the front before transmission as depicted in Figure 2. As the rms delay spread is a result of the physical channel it cannot be changed and systems must be designed to accommodate it. The transmitted signal does not only have a direct path to the receiver (in the case of line of sight). a measure of the delay is given by the rootmeansquare (rms) delay spread which is a measure of the delay experienced by a single pulse. This causes ISI which degrades the Bit Error Rate (BER).5. After cyclic shifting to get the samples back into the original order a FFT is performed to demodulate the data. Each subcarrier is modulated at a sufficiently low data rate so that it is not affected by the delay spread. The data rate in a single carrier system can be increased by shortening the symbol time of the transmitted pulses. At the receiver a window of N samples is chosen from the N+Ng length block for maximum power. The longer the delay of the paths the greater the ISI. 15 . electromagnetic signals travelling through this medium are fraught with disruptive and warping effects. Provided that the length of the CP is chosen so that it is longer than the longest expected delay path successive OFDM symbols will be free of ISI [13]. It is this effect which restricts single carrier systems from achieving high data rates.1 Multipath propagation The wireless channel is a harsh one. This phenomenon has prompted the use of multicarrier techniques where the transmitted bandwidth is divided into many narrow band channels which are then transmitted in parallel. The transmitted signal is therefore N+Ng samples.4. the rest of the repeated samples are discarded.Chapter 2: Orthogonal Frequency Division Multiplexing 2.
The CP (with repeated samples) retains the cyclic nature of the symbol by creating a periodic received signal for processing. eliminating ICI. A loss in the SNR of the received signal is also incurred due to the lost energy in the CP. and b) with cyclic prefix. by using a CP the delay will occur in CP so that N of the samples will be at full power.5) as the repeated samples are discarded in the receiver so it is important to keep the length of the CP as short as possible with respect to the rms delay spread. However there are techniques which use the CP for both frequency offset estimation and symbol synchronization [14]. 0 N1 time a) Original N point OFDM symbol time Ng Original N samples b) OFDM symbol with Cyclic Prefix (CP) Figure 2.Chapter 2: Orthogonal Frequency Division Multiplexing Obviously the use of a CP decreases the data rate by a factor of N N + Ng (2. Also when filtering the signal there is a delay before the filter is at full power.5: OFDM symbol a) without cyclic prefix. 16 .
2. Constant fading over the occupied bandwidth is known as frequency flat fading and is a much easier effect to correct in the receiver than frequency selective fading.15. Narrow pulses in time (such as high data rate single carrier transmission) occupy a wide frequency bandwidth. Under this condition the signal experiences multipath introducing ISI. If the channel has a constant magnitude and phase response over a bandwidth that is smaller than the bandwidth of the transmitted signal the channel creates frequency selective fading [16].3a ultra wideband standard. conversely pulses with a long duration (such as OFDM) occupy a relatively narrow frequency band. The received signal in the null sample positions is wrapped around and added to the samples at the start of the symbol to restore orthogonality and eliminate ICI.3 Frequency selective fading Multipath propagation as discussed in the previous section can be combated successfully through the use of a cyclic prefix. 17 . Frequency selective fading is the reciprocal effect of multipath propagation in the time domain and can be defined thus. Figure 2. Other forms of the CP have also been investigated. It was concluded that null samples have a detrimental effect through loss of orthogonality increasing ICI.6 compares a single and multicarrier signal in the time and corresponding frequency domain with equivalent data rates. This is one of the proposals for the new IEEE 802. in particular reference [15] examines the effect that using null values in the CP will have. Here we see that the many narrowband channels of the OFDM signal experience fading.4. however each subchannel has a constant gain within its own frequency band. this effect shows itself in the frequency domain where certain frequency components in the received spectrum have greater or less power than the transmitted spectrum.Chapter 2: Orthogonal Frequency Division Multiplexing More recent work has shown that it is possible to use null values for the CP provided additional processing is done at the receiver.
of the signal is [22]. This reduction in equalization complexity is a driving force for the use of OFDM. Equalization algorithms are usually implemented as tapped delay lines. one of the more popular methods utilize pilot tones [1719] which are certain (usually evenly spaced) subcarriers with a known amplitude and phase at the receiver.e. Several methods have been suggested. W. 18 . more taps) where the complexity is directly proportional the bandwidth of the signal.6) Signals which experience frequency selective fading such as single carrier systems require complex equalization (i. TE and the occupied bandwidth. Signals such as OFDM which experience frequency flat fading only require a 1 tap equalizer.4 Equalization In order for the receiver to correct the effect of fading equalization is performed in the receiver which is the process of measuring the channel response and using this information to correct the received signal.4. The relation between the number of taps required.Chapter 2: Orthogonal Frequency Division Multiplexing 2. namely time and frequency synchronization problems and non linearities. TE = ( BW ) 2 (2. The equalization technique depends on the modulation scheme and the channel properties. however the advantages are offset by some problems that are unique to OFDM. 21] which do not require pilot tones.5 Limitations in OFDM Previous sections have detailed the advantages of OFDM. 2. By measuring the difference between the received and the transmitted value a picture of the channel can be extrapolated. Other methods use blind estimation techniques [20.
1 Synchronization Both time and frequency synchronization are a major drawback in OFDM.Chapter 2: Orthogonal Frequency Division Multiplexing 1 2 time time Frequency fading envelope N NTs time 1 NTs freq OFDM time and frequency properties Frequency fading envelope time Ts freq 1 Ts Single carrier time and frequency properties Figure 2.7) 19 . If the timing mismatch is within the CP the demodulation produces a linear phase rotation at the output of the FFT which can be corrected with a channel estimator.5.1 Timing errors Timing synchronization is the process of finding the start of a symbol in the receiver.6: Time and frequency properties of single carrier and OFDM techniques. 2. 2. N −1 k =0 ɶ xb ( t ) = ∑ a [ k ] e j 2π k ∆ft + ( f c + k ∆f )δ t (2.1. the following sections detail the problem and provide a basic introduction into solutions for these problems.5.
Chapter 2: Orthogonal Frequency Division Multiplexing where j 2π f + n∆f δ t ˆ aδ t [ k ] = a [ k ] e ( c ) (2. 2. An alternative approach is to use pilot based methods [13] which uses certain carriers with a known amplitude and phase at the receiver. 2. 20 . A simplistic representation of this effect in the frequency domain is visualized in Figure 2. Power from adjacent subcarriers is also sampled as well. Doppler shifts. Several references [23.8) is the phase shift. No distinction can be made between phase rotations introduced by timing errors and carrier phase offset [14]. 24] have analyzed the effect of carrier phase noise on the performance of OFDM schemes. Frequency offset causes the received signal to not be sampled at the peak. Therefore a sufficient length of the CP needs to be chosen.5.2 Carrier phase noise Carrier phase noise is caused by a mismatch in the RF oscillators in the transmitter and receiver and manifests itself in the baseband as additional phase rotation and amplitude attenuation [13].3 Frequency errors Frequency offset errors are caused by mismatch between the RF oscillators. and phase noise introduced by non linear channels [14]. If the timing mismatch is not corrected additional interference (ISI) is generated.1. this means that the sample under consideration is not at maximum power. The effect of phase noise is more pronounced in differential detection schemes than coherent detection schemes [13].1. By analyzing the phase rotation and amplitude change.7.5. an estimate of the channel can be made.
21 . However this shortens the symbol time which increases the demands on timing synchronization.6 Amplitude 0.7: Effects of frequency offset ∆F: reduction of signal amplitude (star). Suggested solutions to frequency synchronization (like symbol synchronization) are based on pilot symbols and the cyclic prefix are used in reference [14] where it is noted that time and frequency synchronization are closely related.Chapter 2: Orthogonal Frequency Division Multiplexing ∆F 1 0. OFDM is more sensitive to frequency offset than single carrier systems due to the tight orthogonal packing of the subcarriers.4 0.2 0 −0. Frequency sensitivity can be made more robust by reducing the number of subcarriers within a set bandwidth thereby increasing the frequency distance between subcarriers. and ICI (circle). therefore a trade off must be made.2 −0.4 −4 −3 −2 −1 0 1 frequency (normalised) 2 3 4 Figure 2.8 0. Reference [14] concludes that to avoid severe degradation the frequency accuracy should be better than 2%.
and distorted techniques which deliberately reduce peaks but increase distortion and therefore the BER.2 Non linearities Another limiting aspect of multicarrier and OFDM modulation is the high instantaneous signal peak with respect to the signals average power. This problem is the area of research of this thesis and will be treated with more detail in the following chapters.Chapter 2: Orthogonal Frequency Division Multiplexing 2. These specially chosen bits allow powerful error correction codes in the receiver to reduce the BER. 2. Large peaks are due to the superposition of N random phase sine waves in the IFFT. 6. Hardware components such as the DigitaltoAnalog Converter (DAC). The more bits used for error correction the better the error correction properties. If the signal is allowed to go into saturation both in band noise which degrades the BER and out of band radiation introducing ICI will result. IFFT/FFT with limited word length and most importantly the High Power Amplifier (HPA) will be driven into saturation unless they are designed to operate over large dynamic ranges.6 Applications of OFDM The previous section detailed some of the problems with OFDM.6.1 COFDM Coded OFDM (COFDM) is a practical form of OFDM where redundant bits are inserted into the bit stream at the transmitter. it should be noted that depending on the application and medium different design issues take precedence. 5. Therefore many papers (refer to Chapters 4. however the useful data rate is decreased. 22 . This section identifies some of the current and future applications of OFDM. 2. and 7) have been published on ways to overcome the PAPR and can be divided into two methods: distortionless techniques which attempt to create a transmit signal with a low PAPR without affecting BER of the data.5. OFDM takes its place in the next generation of communication systems because of its high data rates and low complexity.
6. data associated with audio. The PAPR is a problem but as DAB only uses DQPSK modulation it is more impervious to noise generated through saturation of the amplifier.1 displays system parameters for DAB. Table 2.1: DAB parameters. In DAB between 192 and 1536 carriers are used with Differential Quadrature Phase Shift Keying (DQPSK).5GHz ≤3GHz 23 .Chapter 2: Orthogonal Frequency Division Multiplexing Types of error correction codes used for example DABOFDM are Trellis Coded Modulation (TCM) combined with frequency and time interleaving.2 Digital Audio Broadcasting Digital Audio Broadcasting (DAB) using OFDM has been standardized in Europe [25] and is the next step in evolution beyond FM radio broadcasting providing interference free transmssion. In practice all the following technologies use some form of COFDM. The standard for DAB is known as Eureka147 [26] and is a multiservice digital broadcasting method transmitting at around 1. Parameters I Application Modulation Total number of subcarriers OFDM symbol duration Guard interval Frequency range SFN DQPSK 1536 1246µS 246µS Mode II Terrestial DQPSK 384 312µS 62µS III Satellite DQPSK 192 156µS 31µS ≤375MHz ≤1. Table 2.536MHz band. The DAB data payload contains audio. and other optional data services. The very long symbol time means that large echo’s can be tolerated and that the redundancy due to the CP is not that great. 2.5Mbps in the 1. which allows the system to avoid channel estimation techniques. Large echoes are expected as the broadcasting is over large distances so that long delay paths will be present.
DVB adapts the baseband TV signal from the output of the MPEG2 [29] transport multiplexer to the terrestrial channel characteristics.6. this aspect is taken advantage of in DVB so that the data rate on a channel mirrors its quality. 16 QAM. The 8K mode encompasses the 2K mode as well as larger SFN’s. ¾ 1705 (2K mode) 303µS 75. DVB promises to deliver full multimedia in digital form in a broadcast format. Maximum spectral efficiency within the VHF and UHF bands is achieved by utilizing Single Frequency Network (SFN) operation. Table 2. and 2/ 8K mode. 2/3.11a use OFDM as the physical layer modulation scheme and operate in the unlicensed 5GHz frequency band. One of the many advantages of OFDM is that different mapping types can be used on different subcarriers.4 HiperLan2/802. 64 QAM Reed Solomon outer code Convolutional inner code Code Rates Total number of subcarriers OFDM symbol duration Guard interval Signal bandwidth ½.3 Digital Video Broadcasting Digital Video Broadcasting (DVB) [27.6. 28] is also using OFDM as the carrier modulation scheme. Table 2.2: DVB system parameters for 2K mode.62MHz 2. Hiperlan2 promises to deliver raw data rates of up to 56Mbps which puts them in the ballpark of wired LANs which have data rates of up to 100Mbps. The 2K mode is used for single transmitters and small SFN’s where the distance for transmission is limited. There are two modes defined in DVB: 1/ 2K mode. Wireless 24 .Chapter 2: Orthogonal Frequency Division Multiplexing 2.11a Wireless networking standards such as HIPERLAN2 and 802. Parameters Information data rate Modulation FEC code 530 Mbps Value QPSK.9µS 5.2 shows the system parameters for DVB in 2K mode [28].
3: HIPERLAN2 parameters Parameter Sampling Rate fs=1/T Useful symbol part duration TU Cyclic Prefix duration TCP Symbol Interval TS Number of data subcarriers NSD Number of pilot subcarriers NSP Total number of subcarriers NST Subcarrier spacing ∆f Spacing between two outmost subcarriers NST Value 20MHz 64 × T 3.2µS 16 × T 8×T 0. Tables 2.1Mbps in the downlink and 9.6µS (TU+TCP) 48 4 52 (NSD+NSP) 0.4: Data rates for HIPERLAN2 Modulation BPSK BPSK QPSK QPSK 16QAM 16QAM 64QAM Coding rate R 1/2 3/4 1/2 3/4 9/16 3/4 3/4 Nominal bit rate (Mbps) 6 9 12 18 27 36 54 (optional) 2.11a specifications and data rates respectively.11b wireless standard free of charge for customers and provides data rates up to 11Mbps [52]. while still supporting the standard telephone.25MHz (NST × ∆f) Table 2.6. Table 2.4µS 80 × T 72 × T 4.3 and 2.0µS (TU+TCP) 3.54Mbps to 6.8µS (mandatory) 0.6 to 192Kbps in the uplink over several kilometers of ordinary twisted pair telephone line.3125MHz (1/TU) 16.5 ADSL Asynchronous Digital Subscriber Lines (ADSL) utilize OFDM over wired links [30]. Data rates for ADSL standard [14] are 1. The unbalanced data rates make 25 .Chapter 2: Orthogonal Frequency Division Multiplexing LANs applications are for home and office networking over short distances (<50 metres) as well as community spaces such as Starbucks which operates the 802.4 list Hiperlan2/802.
This structure allows greater diversity when techniques such as Singular Value Decomposition (SVD) are used. Specifically the history of muticarrier networks was discussed and its evolution towards OFDM. This is one of the great advantages of OFDM over single 26 . OFDM transmission over wireless channels was discussed focusing on multipath propagation and the advantages of OFDM in this medium. since all data is transmitted both in the same frequency band and with separate spatial signatures. The introduction of the Fourier transform.7 Conclusion This chapter introduced fundamental properties of OFDM. Stationary channels like wireless links do not change over time. proportionally boosts the datatransmission speed by a factor equal to the number of transmitting antennas. and the efficient use of the frequency spectrum with orthogonally spaced subcarriers were shown to make OFDM a practical technology for the next generation of digital communications. 2. this technique utilizes spectrum very efficiently.6 MIMO OFDM Multiple In Multiple Out (MIMO) [31] OFDM combines OFDM with multiple antennas at the transmitter and receiver. The use of the cyclic prefix in OFDM was shown to reduce the equalizer complexity dramatically down to one tap per subcarrier.6. identifying its advantages and discussing its limitations. This process. 2. In addition. therefore a technique called bit loading is used. using the available bandwidth efficiently. Bit loading assigns a mapping type to subcarriers depending on its quality. Bit loading used in conjunction with OFDM over wired links is usually called Discrete MultiTone (DMT). called spatial multiplexing. advances in silicon technology.Chapter 2: Orthogonal Frequency Division Multiplexing ADSL particularly applicable to internet type applications where the downlink rate is typically much larger than the uplink rate.
27 .Chapter 2: Orthogonal Frequency Division Multiplexing carrier networks which require prohibitively high complexity equalizer structures as the carrier frequency is increased to cater for higher data rates. and frequency errors. Finally applications of OFDM were presented detailing where this new technology has manifested itself in society. linearities due to the Rayleigh distributed samples at the output of the IFFT were also briefly introduced and shown to have a degrading affect on the quality of OFDM systems. carrier phase noise. OFDM in DMT form is being used in wired networks for ADSL. DAB. synchronization errors and non linearities treated. with the two main disadvantages. Hiperlan2/802.11a for wireless networking potentially taking the place of large wired networks. Areas of application in the wireless field were shown to be COFDM. DVB where OFDM is used in a broadcast mode. Frequency selective fading was also introduced as a major advantage of OFDM where due to the long effective symbol time OFDM subcarriers experience flat fading. Limitations of OFDM were analyzed next. Synchronization errors were Non shown to include timing errors. Extensions of OFDM such as MIMO OFDM were also shown to expand the reach of OFDM systems.
Chapter 3: Peak to Average Power in OFDM
Chapter 3
Peak to Average Power in OFDM
Chapter 2 discussed fundamental principles of OFDM and showed how it is a practical technology for the next generation of high data rate communication systems. However several design issues need to be addressed, one of the most important being the Peak to Average Power Ratio (PAPR) of the highly fluctuating transmit signal envelope.
Due to the nature of the IFFT which, as described in Section 2.3, sums N sinusoids through superposition, some combinations of the sinusoids create large peaks. The drawback of a large dynamic range is that it places pressure on the design of components such as the word length of the IFFT/FFT pair, DAC and ADC, mixer stages, and most importantly the HPA which must be designed to handle irregularly occurring large peaks. Failure to design components with a sufficiently large linear range results in saturation of the HPA. Saturation creates both in band distortion, increasing the BER and out of band distortion, or spectral splatter, which causes ACI.
One obvious solution is to design the components to operate within large linear regions, however this is impractical as the components will be operating inefficiently and the cost becomes prohibitively high. This is especially apparent in the HPA where much of the cost and ~50% of the size of a transmitter lies.
This chapter provides a mathematical definition of the PAPR and identifies the processes which contribute to large peaks. 28 Specifically section 3.1 gives a
Chapter 3: Peak to Average Power in OFDM mathematical definition of the PAPR, section 3.2 provides a statistical analysis of PAPR and identifies contributing factors to large peaks. Section 3.3 introduces
oversampling of OFDM signals while section 3.4 investigates the effect of non linearities on OFDM, finally section 3.5 summarizes the chapter reiterating the main points of this section. Note that the terms subcarrier, tones, and N will be used interchangeably to signify the number of subcarriers in an OFDM symbol.
3.1 Peak to Average Power Ratio
The PAPR is the relation between the maximum power of a sample in a given OFDM transmit symbol divided by the average power of that OFDM symbol. The mean envelope power of the baseband expression (assuming same constellation on each subcarrier) is defined as (3.1)
N −1 k =0
P=
2 1 T 1 ∫t =0 xm ( t ) dt = N T
∑X
2 m ,k
(3.1)
where xm ( t ) is defined in (2.1), Xm,k are assumed to be complex Quadature Amplitude Modulated (QAM) data which are statistically independent, identically
2 distributed (i.i.d) random variables with 0 mean and variance σ 2 ≜ E X m ,k . The
average power is defined as (3.2)
2 Pav = E [ P ] = E xm ( t )
(3.2)
The PAPR can then be defined
ζ =
max 0≤t ≤T xm ( t ) Pav
2
(3.3)
where max 0≤t ≤T xm ( t ) 0≤t ≤T .
2
is the maximum instantaneous power within the period
29
Chapter 3: Peak to Average Power in OFDM Another definition for the PAPR is crest factor, defined in reference [32] as (3.4) max 0≤t ≤T xm ( t ) Pav
2
ξCF =
(3.4)
and results in a 3dB shift in results compared to (3.3). Throughout this thesis the definition of the PAPR given in (3.3) will used unless specified otherwise.
For passband transmission the OFDM symbol is modulated onto a carrier frequency, fc, xmPB = ℜ xm ( t ) e j 2π fct
{
}
(3.5)
= ℜ { xm ( t )} cos ( 2π f c t ) − j ℑ{ xm ( t )} sin ( 2π f c t ) = xmI ( t ) cos ( j 2π f c t ) − jxmQ ( t ) sin ( j 2π f c t )
The carrier frequency is usually much higher than the signal bandwidth, i.e. fc>>∆f, therefore the maximum of the passband signal is approximately equal to the baseband expression, i.e.
max xmPB ( t ) ≈ max xm ( t )
(3.6)
Most OFDM schemes usually employ QAM mapping for the modulation where
xmI ( t ) = xmQ ( t ) therefore
2 2
E xm ( t )
{
2
} = 2E { x
mI
(t )
2
} = 2E { x
mQ
(t )
2
} }
2
(3.7)
The average RF power of the passband signal can be derived as
E xmPB ( t )
{
2
} = E{ x = E{ x
mI
( t ) cos ( 2π f ct ) − jxmQ ( t ) sin ( 2π f ct ) ( t ) cos ( 2π f ct )
2
2
mI
}+ E{ x
mQ
( t ) sin ( 2π f ct )
}
30
Chapter 3: Peak to Average Power in OFDM
= = =
2 2 1 1 E xmI ( t ) + E xmQ ( t ) 2 2 2 1 E xm ( t ) 2
{ {
}
{
}
(3.8)
}
Pav
2
Substituting (3.8) back into (3.3) gives the PAPR in the passband.
ζ PB =
max 0≤t ≤T xmPB ( t ) E xmPB ( t )
2
{
2
}
max 0≤t ≤T xmPB ( t ) = Pav 2
2
(3.9)
The problem with OFDM is that theoretically the PAPR can be up to log2(N), which is huge. But as will be shown in the next section the general distribution of samples is much lower.
3.2 Statistical distribution of OFDM samples
Section 3.1 provided a worst case scenario or upper bound of the PAPR, but only a few combinations of input data sequences produce large peaks, therefore it is more pertinent to define the statistical distribution of the PAPR in OFDM. (2.2) describes a Nyquist sampled baseband OFDM symbol with N subcarriers, from the central limit theorem [33] the sum of these elements are zero mean complex random near Gaussian (provided N>64) distributed variables with variance, σ2, of ½. It then follows that the amplitude, an, of the OFDM symbol has a Rayleigh distribution [13] with a Probability Density Function (PDF) of
ζ −ζ pζ (ζ ) = 2 e σ
n
2
2σ 2
(3.10)
31
Comparing the theoretical and simulated results we see that the results only converge 32 .2 displays simulated and theoretical results of (3.3. As in Figure 3.1.13) with varying N. balance must be met between these design constraints.5.5.1 the probability that large peaks occur is very irregular. ζ 0 . Therefore.2 and 2. However a larger number of subcarriers leads to increased sensitivity to carrier and sampling frequency offsets as described in section 2. doubling the number of subcarriers results in a modest increase in the PAPR leading to the assumption that using a large number of subcarriers makes sense as this will allow for greater data throughput (provided the CP length is constant).12) Under the assumption of statistically independent samples the Complementary Cumulative Distribution Function (CCDF) can be found for the case where at least one sample in an OFDM symbol exceeds the magnitude.11) Figure 3. It is seen here that the probability of any given sample having a magnitude above 3dB decays rapidly. The probability that the magnitude of a sample is below a certain threshold. is given by the Cumulative Distribution Function (CDF) Pr {ζ ≤ ζ 0 } = ∫ pζ n (ζ ) ∂ζ = ∫ 2ζ e −ζ ∂ζ = 1 − e−ζ 0 2 ∞ ζ0 2 −∞ 0 (3.1. The simulation model passes N QPSK symbols through a N point IFFT.1 shows the simulated envelope (N=64) of a baseband OFDM system which as expected follows a Rayleigh distribution. ζ 0 .Chapter 3: Peak to Average Power in OFDM Substituting σ2= ½ pζ n (ζ ) = 2ζ e−ζ 2 (3. Pr maxζ > ζ 0 = 1 − Pr maxζ ≤ ζ 0 0≤ n < N 0≤ n < N = 1 − ( Pr {ζ ≤ ζ 0 } ) = 1 − 1 − e −ζ 0 N ( 2 ) N (3. the maximum sample of each OFDM symbol is stored and plotted.13) Figure 3.
2: Simulated (solid line) and theoretical (3.5 1 1.1 Probability 0. 10 0 10 1 N=32 N=64 Pr(ζ> ζ0 ) 10 2 N=128 N=256 10 3 10 4 2 3 4 5 6 7 8 9 10 11 12 . 30000 runs ζ0 (dB) Figure 3.08 0. volts 3 3.06 0.14 0.Chapter 3: Peak to Average Power in OFDM at N=128.12 0.5 Amplitude. dashed line) OFDM symbol CCDF for N=32. 33 . 128.13.1: Simulated envelope for OFDM system (N=64) normalized by average power. and 256 subcarriers. 64.5 2 2. the theoretical results for N<128 are slightly more pessimistic at higher PAPR levels.04 0. QPSK.5 4 Figure 3.02 0 0 0. 0.
2.4 where it is seen that changing the constellation has a minimum affect on the PAPR which is to be expected considering the Mary constellations are normalized to have the same average power. However null subcarriers have the advantageous side effect of increasing the resolution of the OFDM symbol due to the oversampling effect.Chapter 3: Peak to Average Power in OFDM To provide further insight into the total distribution of the PAPR Figure 3.3: Simulated OFDM sample CCDF for N=32. QPSK. This same principle applies to non active subcarriers which also do not influence the PAPR as the average power decreases in line with a reduction in the number of active subcarriers. 30000 runs. Here it is seen that the sample distribution is lower than the symbol distribution as is predicted by (3.13). The effect of various QAM mapping constellations is simulated in Figure 3.3 plots all simulated OFDM samples for the same case as Figure 3. and 256 subcarriers. 128. 34 . 64. 10 0 10 1 10 Pr(ζ> ζ0 ) 10 2 3 N=32 N=64 10 4 N=128 N=256 4 2 0 2 4 6 8 10 12 ζ0 (dB) Figure 3.
Oversampling the data in the IFFT increases the resolution of the OFDM symbol giving a closer approximation to the band limited signal after filtering. This section has shown that the PAPR per symbol is only a function of the length.3 Oversampling discrete OFDM symbols to find true (continuous) peaks Section 3. 30000 runs. N=64.2 provided an analysis of the PAPR for critically sampled baseband OFDM symbols.4: Simulated OFDM CCDF for M=4.Chapter 3: Peak to Average Power in OFDM 10 0 10 1 Pr(ζ> ζ0 ) 10 2 M=4 3 10 M=16 M=64 10 4 2 3 4 5 6 7 8 9 10 11 12 ζ0 (dB) Figure 3. N. However this analysis does not reveal the peak of the band limited OFDM signal.5 where the complex components of one OFDM symbol with no oversampling is overlaid with the same symbol oversampled at the IFFT by a factor of 8 (to approximate the continuous filtered signal). the growth in new peaks occurs in between the discretely sampled peaks. 3. This is best explained in Figure 3. The constellation type and number of active subcarriers have a negligible affect on the PAPR after modulation with the IFFT. of the IFFT. Note the parabolic trajectory from one discrete sample to another. and 64 constellation mapping. 35 . 16.
Some interesting observations that can be made viewing Figure 3.1 0 I channel 0. At P3 we see that the second largest peak in the oversampled case occurs at a position where no peak existed in the critically sampled symbol.5 are at P1 where the only sample of the critically sampled OFDM symbol is above 6dB.4 Figure 3.4 0.3 1 P 3 0.4 0.2 P 0. here it is seen that an oversampling factor of 8 is sufficient to represent the continuous signal and results in around 0.3 0.6. but after oversampling a peak is produced in between the 2 critical samples. These results show how the critically sampled OFDM symbol and its oversampled version can diverge greatly in PAPR.1 Q channel 0 0. 36 .1 0.Chapter 3: Peak to Average Power in OFDM 0. The solid circle represents the 6dB level with respect to the average power. At P2 it is seen that the two critical samples that make up its end points are well under 6dB.3 0.2 0.5: Simulated OFDM symbol with no oversampling (dashed) with it’s oversampled version (solid) overlaid on top.2 0.4 P 0.2 0.1 0. N=64. oversampling factors are 1 and 8.5dB increase in the PAPR. when oversampled the true peak grows slightly larger still. The CCDF for various oversampling rates at the IFFT is shown in Figure 3.3 2 0.
Various papers [3436] have been published which address the issue of oversampling. and 8. 4. Equation (3. 2. N=64. (3. Reference [37] suggests that adding a number of extra independent samples to (3.13) will give a closer approximation to the oversampled signal. QPSK.14) Therefore the independent assumption of (3.15) 37 .13) assumes that samples are mutually independent and uncorrelated. however Parsevals theorem states that N −1 ∑x m =0 2 m =N (3.6: Simulated OFDM CCDF for oversampling rates of 1. especially in the oversampling case where adjacent samples are highly correlated to each other. 15000 runs.Chapter 3: Peak to Average Power in OFDM 10 0 10 1 Pr(ζ> ζ0 ) os=1 os=2 10 2 os=4 os=8 10 3 2 3 4 5 6 7 8 9 10 11 ζ0 (dB) Figure 3.13) is not true.15) 2 Pr maxζ > ζ 0 = 1 − 1 − e−ζ 0 0≤ n < N ( ) αN (3.
17) which is the mean number of total peaks. As this method is numerically cumbersome to solve due to the double integration in (3.16) a simpler approximation of the distribution is developed which is as accurate as the previous method for a large number of N. and N p ( 0 ) ≈ 0. This is made under the assumption that 1) the complex components of the signal. 38 . The simpler method derives the peak distribution of the band limited Rayleigh process based on the level crossing rate approximation and then applies the result to the derivation of the distribution of the CF in OFDM signals. Reference [35] takes exception to this non theoretical approximation of the over sampled signal and states that the bound is not close to the theoretical bound for large numbers of N. Reference [35] developed a method for finding the exact peak distribution of band limited Rayleigh processes giving an expression for the CCDF of the PAPR as (3.8 gives a good approximation to oversampled signals. N>64) and 2) the peaks are statistically uncorrelated. Also a suitiably high level for a0 must be chosen well above 0 to make the assumption valid.64 N (3.16) Pr (ζ < ζ 0 ) N p ( 0) = (1 − Pr (ζ > ζ 0 ) ) N (ζ ) = 1 − p 0 N p ( 0) N p ( 0) N p ( 0) (3.16) where N p (ζ 0 ) = 4N 15π ∫ ∞ a0 u 2 ∫ ∞ 0 e − φ +1 u 2 ( ) 2 −5 4 e (φ −1) u 2 2 2 − 5π 2 (φ 2 − 1 uerfc ) 5 2 (φ 2 − 1 u dφ du ) which is the mean number of peaks above the level ζ 0 in one OFDM symbol. x(t) are ideally band limited Gaussian processes (i.e.Chapter 3: Peak to Average Power in OFDM where α=2.
18) FC ζ C > ζ 0 = Pr ζ < ζ 0 ζ > ζ 0 ( ) ( ) N p (ζ 0 ) (3.e. each positive crossing of the level ζ 0 has a single positive peak that is above the level ζ 0 . On the other hand (3. Figure 3.21) is an excellent bound for N=512 where the Gaussian assumption is true.15) and (3.19) An expression for the CDF.18) where N p (ζ 0 ) is the mean number of peaks above ζ 0 .20) ζ0 ≤ ζ0 The CCDF can then be expressed as 1 − FC (ζ 0 ) (3. The CDF of the CF is given as (3. FC (ζ 0 ) is then obtained: FC (ζ 0 ) ≈ FC ζ 0 C > ζ 0 ( ) π 3 Nζ 0e− ζ 0 2 2 ζ 0 e −ζ 0 1 − = −ζ 0 2 ζ 0e 0 for for ζ0 > ζ0 (3. and can be approximated for high ζ 0 by (3.15) is closer to the simulated results for N=64 as a small number of subcarriers does not have a completely Gaussian distribution. and 20000 OFDM symbols against (3.21) [35] suggests ζ 0 = π for QPSK modulation and a marginally lower value for 16 QAM.Chapter 3: Peak to Average Power in OFDM i.21). Here we see that the bound from (3. 39 .7 plots simulated CCDF of the PAPR for QPSK with N=64 and 512 with an oversampling factor of 16.19) N p (ζ 0 ) = π 3 N ζ 0 e −ζ 0 2 (3.
15) and (3.22) where k is the oversampling factor.Chapter 3: Peak to Average Power in OFDM 10 0 Eq (3. oversampling factor rate of 16. and ζ 0 is the clip value. only the Gaussian assumption for each sample is used. 40 . Figure 3. This method predicts a much higher distribution of the peaks than is seen in simulation and is therefore very pessimistic. Reference [36] finds bounds for the peak of the continuous envelope based on the maximum of the oversampled sequence. 512 with simulated results: QPSK.7: Theoretical OFDM CCDF from (3. Note that k must be > π 2 .8 plots the simulated results (QPSK. there is no assumption on the joint distribution of the samples.15) Eq (3. os=16) against the theoretical results of (3. N=64. 512. N is the number of subcarriers. The CCDF is given as Pr {PMEPR > ζ 0 } < kNe π2 − λ 1− 2 2k (3.22). Unlike [35] where the PMEPR is derived under the assumption that OFDM signals behave as band limited Gaussian processes.21) for N=64.21) Simulated 10 1 Pr(ζ> ζ0 ) 10 2 N=64 N=512 10 3 10 4 4 5 6 7 8 9 10 11 12 13 ζ0 (dB) Figure 3. This bound is used to derive a closed form expression for the upper bound of the CCDF in an uncoded OFDM system with large N. 20000 runs.
512 with simulated results: QPSK.8: Theoretical OFDM CCDF from (3. The upper bound on the CDF for QAM constellations is given as −4( M 2 −1) a0 2 3( 4 γ + M 2 )Ck 2 CL 2 F (ζ 0 ) ≤ min KLNe L . K > 2 (3.23) ζ0 ≥ 0 where CL = 1 π cos 2L L >1 (3. N=64.24) 41 . rather than just QPSK. Reference [38] extends the theory of [37]. 512.22) Simulated 10 1 N=64 Pr(ζ> ζ0 ) 10 2 N=512 10 3 10 4 4 6 8 10 12 14 16 18 20 ζ0 (dB) Figure 3. The theoretical results are useful at low probability regions where simulations are time consuming. 20000 runs. L >1. [35].22) for N=64. and [36] to find new upper bounds for different constellation types such as QAM and PSK. oversampling factor rate of 16.Chapter 3: Peak to Average Power in OFDM 10 0 (3. K ∈ℕ .
Interestingly the Gaussian approximation provides better results than the new bound for all N. M is the constellation type: M=2 (4 QAM). K > 2 l1 =1 l2 = 0 2 ζ0 ≥ 0 (3. 1 (16 QAM). L is the oversampling factor. 4 (16 QAM). For BPSK constellation the CDF is given by Nζ LN −1 − l1 l K −1 2 C1 N . For 16 QAM and 64 QAM the Gaussian bound again looks better. 2π . The CCDF results from Figure 3.5dB.26) where N 1 sin ( Nθ ) + cos ( ( N − 1) θ + 2α ) 2 2 sin (θ ) C1 ( N . however this is not true when PAPR reduction 42 . or 8 (64 QAM).6 imply that using an oversampling factor of 1 predicts the PAPR within 0. α ) = (3. L >1. K ∈ℕ .27) and B (ζ 0 ) = 1 2N k = ζ 0 N ∑ ( N − k ) / 2 N / CL keven N (3.Chapter 3: Peak to Average Power in OFDM 1 π cos K Ck = π 3 − cos K π 1 + cos K ( ) ( ) ( ) K >3 K ≥3 Keven (3. θ .13) show that for BPSK the new bound is tighter below 104 probability for large N. 2 2π C K 2 C L 2 2 K LN F (ζ 0 ) ≤ 4 B (ζ 0 ) + min ∑ ∑e L .28) Results comparing the new CDF calculation against the Gaussian model of (3.25) Kodd where ζ 0 is the clip level. or 5 (64 QAM). N is the number of subcarriers. γ =0 (4QAM).
as well as the same OFDM data set left unclipped at the IFFT and then interpolated and filtered with the RRCF. interpolated by 8 and then filtered with a 256 tap RRCF. 10 0 No clipping and filter 10 1 Pr(ζ> ζ0 ) Clip at 3dB and filter 10 2 10 3 2 3 4 5 6 7 8 9 10 11 12 ζ0 Figure 3. α=0. 43 .Chapter 3: Peak to Average Power in OFDM techniques are applied to the OFDM symbol [30]. the length of zero padding in the IFFT. The degree of peak regrowth is determined by the sequence of the filtered data. then filtered (dashed line). much worse than without clipping and filtering.9 shows the CCDF of a 64 point IFFT OFDM modulator clipped at 3dB. the value of the excess bandwidth. Peak regrowth after clipping and filtering is dramatic. No clipping. Figure 3.10.15. 256 tap RRCF. and the degree of clipping (harder the clipping the greater the regrowth). 64 point IFFT. of the RRCF (smaller α greater peak regrowth). Clipped at 3dB after IFFT.[39]. The null samples are introduced in the centre of the IFFT input to ensure that they occupy the outer samples in the frequency spectrum. α.9: Simulated OFDM CCDF QPSK. Oversampling means the introduction of null samples at the input of the IFFT as shown in Figure 3. then filtered (solid line).
11: Simulated OFDM CCDF after IFFT (os=1 and 2) and after filtering (os=1 and 2).10: Zero padding of the IFFT. 15000 runs. 44 . null carriers are set in the middle of the input 0 X(N×os)1 xN×os Figure 3.15. N=64. α=0. both of which are interpolated by a factor of 8 and filtered with a RRCF.11 shows the simulated CCDF for a non oversampled IFFT compared to a 2 times oversampled IFFT. 256 tap RRCF. 10 0 after filtering 64 pts after filtering 128 pts 10 1 Pr(ζ> ζ0 ) after IFFT 64 pts 10 2 after IFFT 128 pts 10 3 4 5 6 7 8 9 10 11 12 ζ0 (dB) Figure 3.Chapter 3: Peak to Average Power in OFDM X0 X1 X(N/2)1 X(N/2) 0 x0 x1 IFFT X(N/2)+1 X(N/2)2 XN1 XN Figure 3. QPSK.
this section describes common models for HPA’s which are used in wireless communications and the effect they have on the OFDM signal in terms of the Power Spectral Density (PSD) and the increase in the BER. Here we concentrate on the most common form of non linearity.10: U=1). and an increased BER at the receiver.4.1 describes commonly used HPA models. We can conclude that oversampling of the data at the IFFT is an effective way to counter the peak regrowth after filtering. 3.4. and finally section 3. Therefore it is pertinent to evaluate the performance of OFDM signals through different non linear devices. but what is of most importance is that the PAPR distribution after filtering has less peak regrowth and is below the critically sampled IFFT. distortion in the RF amplifier due to a limited linear range in the amplifier. 45 . Section 3.2 details the corruption of the frequency spectrum due to changing backoffs in the HPA. Oversampling beyond a factor of 2 provides a law of diminishing returns in terms of peak regrowth (see Figure 4.4 Effect of Non Linearity on OFDM The previous sections described the causes of large PAPR and the problems of non linearity. A balance must be met between allowable distortion and the linear region of an amplifier.1 Description of memoryless Non Linearity Non linearities provide the greatest obstacle to OFDM as a practical system due to their distorting effect on the quality of the system.Chapter 3: Peak to Average Power in OFDM This time we notice that the oversampled IFFT has a slightly higher PAPR distribution. Papers [4044] tend to focus on the distortion due to the RF amplifier stage as this is the most expensive component in a transmitter and takes up to 50% of the cost and space in a unit. This distortion causes spectral regrowth in the transmitter which can adversely affect adjacent frequency bands.4.4.3 describes the degradation in BER due to changing backoffs in the HPA. 3. however when operating near the saturation point it exhibits non linear behavior distorting the transmitted signal. The RF amplifier must be driven as close as possible to the maximum signal in the linear region to make it efficient. section 3.
Some commonly used models follow.Chapter 3: Peak to Average Power in OFDM A convenient form of expression for non linear devices is in polar coordinates reference [45].12. A is the clipping level of the amplifier. 46 . only amplitude distortion.29) Therefore the complex envelope of the output signal can be expressed as (3.30) where F [ρ] and ψ [ ρ ] represent the AM/AM and AM/PM conversion characteristics of the memoryless non linear amplifier respectively.31) There is no phase distortion in this model.29) x= xe arg ( x ) = ρ e jφ (3. if ρ < − A if − A ≤ ρ ≤ A if ρ > A (3. The input can be expressed as (3. Soft Limiter (SL) The Amplitude Modulation to Amplitude Modulation (AM/AM) and Amplitude Modulation to Phase Modulation (AM/PM) characteristics of a SL can be expressed as [43] − A F [ ρ ] = ρ A and ψ [ ρ ] = 0 . The AM/AM characteristics of a SL are plotted in Figure 3.30) j φ +ψ [ ρ ]) g ( x) = F [ρ ]e ( (3.
25 .12: AM/AM properties of a Soft Limiter (SL) Travelling Wave Tube Amplifier (TWTA) According to reference [40] the AM/AM and AM/PM functions are F [ρ] = ρ (1 + βa ρ 2 ) (3. 47 .Chapter 3: Peak to Average Power in OFDM A Output voltage A 0 A A 0 Input voltage Figure 3. αϕ = π (3.33) 12 .25 . and βϕ = 0.13.32) and αϕ ρ 2 ψ [ρ] = (1 + βϕ ρ 2 ) A common choice for the above parameters is β a = 0. The AM/AM characteristics of a TWTA are plotted in Figure 3.
35) The parameter P controls the smoothness of the transition from the linear region into the saturation region.13: AM/AM properties of a Traveling Wave Tube Amplifier (TWTA) Solid State Power Amplifier (SSPA) Probably the most common and practical model for amplifiers is the SSPA [40].6 1.Chapter 3: Peak to Average Power in OFDM 2 1.2 1.6 0.34) and ψ [ρ] = 0 (3.4 0.8 1 Input voltage 1.4 0. P=3 is a good 48 .2 0 0 0.2 TWT 1 0.8 2 Figure 3. The AM/AM and AM/PM transfer characteristics can be modeled as F [ρ] = ρ ρ 1 + A 2P 1 2P (3.8 Ideal 1.8 0.6 0.6 1. When P → ∞ the SSPA acts as a SL.2 0.4 1.4 Output voltage 1.
Chapter 3: Peak to Average Power in OFDM approximation of a practical amplifier.8 Input voltage 1 1.2 Ideal 1 Output voltage P=12 P=3 0. In all amplifiers discussed in this section ‘A’ represents the saturating amplitude of the amplifier.14: AM/AM properties of a Solid State Amplifier (SSPA) for different values of P. 1.2 0 0 0.6 P=1 0.2 1.36) OBO ≜ 10 log10 A2 POUT (3.6 1.37) 49 .4 1. The non linear distortion depends on the backoff of the amplifier and can be calculated as either the Input BackOff (IBO) or the Output BackOff (OBO).2 0.6 Figure 3. and ideal amplifier.6 0.4 0. and is defined as A IBO ≜ 10 log10 s PIN 2 (3.4 0.4 1. The AM/AM characteristics of a SSPA and an ideal amplifier are plotted in Figure 3.14.8 0.
Blocks of 4 QAM data were mapped to a 64 point IFFT for modulation. A RRCF was used for pulse shaping with 128 taps and an excess bandwidth of 0. The low rolloff factor of the filter results in a fast drop off of the spectral splatter outside the normalised FFT bandwidth. An amplifier backoff of 0dB results in noise power that is only 16dB lower than signal power. The frequency axis is normalized. With a backoff of 3dB the noise power is 19. With an infinite amplifier backoff the PSD matches exactly the PSD after transmit filtering.4.15.5dB below the signal power. A 6dB amplifier backoff results in a slight amount of spectral splatter 23dB below the signal power and an almost indistinguishable amount of in band distortion.Chapter 3: Peak to Average Power in OFDM where As is the amplifier input saturation voltage. The PSD is measured for each OFDM block then averaged over 2000 blocks to eliminate the effect of the rectangular time window [46]. This is achieved by 50 .2 Impact on Power Spectral Density The simulated spectrum for N=64 subcarriers is shown in Figure 3. the data is then interpolated by a factor of 8 and filtered. Changing the IBO of the SSPA results in inband distortion and spectral regrowth.15. and POUT is the average power at the output. A is the saturating amplitude at the output. OFDM standards such as Hiperlan2 and 802.11a require adjacent OFDM carriers to be closely packed in the frequency domain with overlapping spectrums. 3. A SSPA with P=3 is used as the amplifier model. PIN is the average power at the input. or splatter outside the normalized frequency bandwidth of the filter.
setting null subcarriers at the edge of the spectrum. This successfully reduces the out of band distortion but the BER due to the in band distortion remains.5dB 23dB 15 HPA=3dB Absolute power (dB) 20 25 HPA=0dB 30 35 40 45 50 4 3 2 1 0 Hz 1 2 3 4 Figure 3. Two observed effects of clipping (whether due to the amplifier non lineraity or baseband clipping) are a Gaussian like spreading of the 51 . Filtering after the SSPA to reduce spectral splattering is complex. resulting in an increase in the BER when the data is decoded. A SSPA (HPA in figure) with P=3 and various backoffs is used. therefore reference [46] proposed clipping in the baseband after the IFFT followed by filtering. 7. These issues are explored in more depth in chapters 5. Also peak regrowth after filtering remains a problem. RRCF with excess bandwidth of 0.15: PSD of 64 subcarrier OFDM signal with 64 point IFFT.3 Impact on Bit Error Rate The inband distortion due to non linear amplification at the transmitter results in ISI when filtered with the matched filter at the receiver.4. easing the design constraints on the filters.15. where the rolloff of the filters occurs in the unused subcarriers.Chapter 3: Peak to Average Power in OFDM 0 LPA 5 10 HPA=6dB 16dB 19. 3.
P=3 after non linear amplification a) 4 QAM. Figure 3. Figure 3. Indeed 4 QAM has a very acceptable BER performance without any coding or PAPR correction. Mapping types 16 and 64 QAM are much more susceptible to clipping in the amplifier. 4 QAM is very impervious to clipping noise as there is a greater Euclidean distance between constellation points.17 could be improved with Automatic Gain Control (AGC) in the receiver which would expand the constellation to the correct size before decoding [48]. 52 . shown in Figure 3. The BER of Figure 3. c) 16 QAM 6dB IBO. and d) 16 QAM 0dB IBO. Therefore the BER is nearly equal on all subcarriers. b) 4 QAM.16: Signal constellation at the output of the SSPA.16 where the two constellation types (4 and 16 QAM) are normalized to have the same average power.17 shows the BER due to a non linearity in the HPA for 3 Mary constellation types. this is because each sample that is clipped is a conglomeration of all the input samples to the IFFT. 6dB IBO. The noise due to clipping is evenly spread across all subcarriers [47].Chapter 3: Peak to Average Power in OFDM decoded constellation points and a compaction of the whole received constellation. 0dB IBO.
3. These results were supported analytically by various papers which analysed the issue of oversampling. and 64 QAM constellations.17: BER of 64 subcarrier OFDM signal with 64 point IFFT. it was shown that oversampling the IFFT increased the CCDF by ~0. Oversampling of the IFFT was also introduced as an important issue.5dB.Chapter 3: Peak to Average Power in OFDM 10 0 4 QAM 16 QAM 64 QAM 10 1 10 2 BER 10 3 10 4 10 5 10 6 4 2 0 2 HPA backoff (dB) 4 6 8 Figure 3. It was also shown that the general distribution of samples is quite low and that the transmit envelope follows a Rayleigh distribution.15 for 4. 16. A SSPA (HPA on xaxis) with P=3 and various backoffs is used. Next the stochastic distribution of OFDM samples was shown through analytical and simulated means to only be a function of the number of subcarriers.5 Conclusion This chapter defined the problem of PAPR in OFDM beginning with a mathematical analysis of the PAPR in both the baseband and passband. 53 . RRCF with excess bandwidth of 0.
filtered CCDF. It was shown that reducing the IBO of the SSPA causes both spectral spreading. it was shown through simulation that filtering has a dramatic affect on the CCDF when combined with PAPR reduction techniques such as clipping.Chapter 3: Peak to Average Power in OFDM Furthermore. The peak regrowth is much worse than when no PAPR reduction techniques are used. and in band distortion corrupting the BER at the receiver. described mathematically. Oversampling at the IFFT was shown to both increase the discrete CCDF and reduce the continuous. Non linearities were also examined with models of the most common form of non linearity. the SSPA. 54 . Higher order constellations were shown to be more susceptible to lower backoffs in the SSPA. affecting the adjacent channels. The impact of non linearity on the PSD and BER was simulated.
and finally Section 4. each solution has advantages and disadvantages in terms of PAPR reduction.2 elaborates on Multiple Signal Representation (MSR) and phase rotating techniques. This chapter reviews distortionless PAPR reduction techniques. Section 4. however things can be done to lessen the effect of the introduced distortion. The other side of PAPR reduction are distortion introducing techniques.4. and complexity. 55 . Distortionless techniques do not corrupt the data and encode it in such a way that it can be completely recovered at the receiver. distortion of data. These techniques will be explored in Chapter 6.1 explains the family of coding techniques for PAPR reduction. however they are usually more complex.3 details modified constellation techniques. Section 4. these techniques deliberately attenuate the envelope of the signal corrupting the BER as described in Section 3. Specifically Section 4.4 summarizes the chapter and provides a comparison in terms of complexity and performance of the detailed techniques. This chapter and chapter 6 provide an analysis of solutions to the PAPR problem.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques Chapter 4 Peak to Average Power Solutions Distortionless Techniques The previous chapter outlined the disruptive effects of an uncontrolled OFDM signal envelope on system performance.
2) 56 . Bit source k PAPR coding n+k Serial To parallel IFFT n Figure 4. C j ) { } (4. C j ) = ∑ Ci .k) code. There are 2 types of error detection and correction codes.l ⊕ C j .2) d min = Min d ( Ci .1 is a block diagram showing where coding for PAPR reduction is located in an OFDM transmitter. The basic premise of coding is to insert redundant bits into the data stream which can be used for error correction at the receiver.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques 4. Their application to PAPR reduction is in creating sequences of bits which will exhibit low PAPR after the IFFT. Most papers relate to the block coding family for PAPR reduction. The block code is referred to as an (n. (4. therefore (nk) redundant non information bits are added to the k information bits [16].l ( mod − q ) N l =1 (4.1) where d is the distance of the codeword and q is the number of possible values of Cj and Cl.1 Coding techniques Many early papers considered how standard coding techniques could be applied to OFDM. block codes and convolutional codes. During the encoding process k information bits are encoded into n code d bits. The smallest distance dmin is the minimum distance for a given set. Figure 4. and the rate of the code as Rc=k/n. (4.1) d ( Ci .1: Block diagram of OFDM transmitter showing PAPR coding The ability of a code to correct errors is a function of the code distance.
48dB.4) block code.1.01dB. a reduction of 3.03dB. However the increase in bandwidth is small and is offset by the high spectral efficiency of OFDM. a (3. They identify the code as an odd parity code and state that the PEP is reduced from 6.45dB.1 Block Codes The first paper to apply coding techniques for PAPR reduction in OFDM was reference [49]. the resultant PAPR is 4.8) rate code. i. ie (3.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques Different codes exhibit different degrees of error correction ability. a (7. Types of block codes are Hamming.54dB. and ReedSolomon.e. some of which are used for PAPR reduction.02dB to 2. as is the reduction in energy per transmitted bit which is offset by the possible error detection/correction potential of block coding. An 8 subcarrier system with BPSK constellation is also evaluated to illustrate how block coding can be traded of against PAPR reduction. so that the set of allowable code words does not create excessive envelope spikes. Practical OFDM systems employ at least 64 subcarriers and higher order mapping types which make the complexity of a block coding scheme 57 . The basic premise of this paper was to determine which combinations of data at the IFFT input produced large peaks at the output and to avoid transmitting these sequences by adding redundant bits to the input. Golay. With no block coding the PAPR is 9. If half the code words are allowed.4) rate code the PAPR is 3. CWposs. which is the number of non zero elements in the codeword. The PAPR reduction comes at the cost of an increase of bandwidth for the same data rate and a reduction of the energy per transmitted bit for the same transmit power. These codes are explored in greater depth in later papers. Initially a simple (impractical) OFDM system with 4 subcarriers and BPSK modulation is considered. No results on the minimum distance of the code are given but the authors indicate that a large number of the codes found are Golay complementary sequences. The permissible number of codewords CWperm is traded off against the total possible number of codewords. 4. Another important property of codes is the weight of the code. If a quarter of the code words are allowed. A 3 bit data word is mapped on to a 4 bit code word.
For an 8 subcarrier BPSK OFDM signal based on Golay complementary sequences a (5. The authors review a set of papers published in the 80’s which addressed a similar problem for multitone test signals with low PAPR. i.e. Selection of code words that also offer error detection/correction properties. for the 4 subcarrier BPSK case ‘1000’ will give the same PAPR as ‘0001’. techniques more practical • • • 3 areas are identified to make coding Selection of suitable code words for any number of carriers and any Mary mapping type.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques grow exponentially as the coding must be done in combinational logic or a Look Up Table (LUT). Selection of code words that allow efficient implementation of the coding/decoding. Larger code sets can be found by rewriting the definition of an OFDM signal (4. but they are amenable to mathematical encoding and decoding and the authors note that they may have error correction properties.01dB. Analysis of the code words that exhibited low PAPR’s for 4 and 8 subcarriers revealed them to be Golay complementary sequences. finite sequences whose aperiodic autocorrelations sum to zero for all nonzero displacements”. Aware of the limitations of the previous paper Jones and Wilkinson extended the block coding principle in reference [50]. Although certain parameters differ much of this work can be applied to the OFDM case.3) 58 . however they did not give the optimum minimum solution. The reverse of a code word will result in the same PAPR. Codes such as these are derived from ShapiroRudin sequences which are a subset of Golay complementary sequences. 4 and 8 subcarrier systems are evaluated using QPSK mapping.8) rate code can give a PAPR of 3. They prove that the code rate is unaffected by changing the number of carriers and mapping types. A definition of Golay complementary sequences is given in the paper as “a pair of equally long.
These phase shifts that calculated for various coding rates in the Hiperlan2 [53] standard where there are 48 information bearing subcarriers. The authors of [49. The principle of these sequences can be extended to give polyphase sequences suitable for multilevel phase modulation. The BER in a non linear channel for CCOFDM is compared to the COFDM channel and significant improvement in the BER is seen in a highly non linear channel.09dB is reported.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques s ( t ) = ∑ dn (t ) e n =1 n=N j ( 2π f n t +φn ) (4. Reference [52] expanded on the work of reference [51] by developing an algorithm to compute the phases that minimize the PAPR for larger sets of data of practical interest. For ½ rate BPSK with 90° phase shifts a reduction of 4. CCOFDM attempts to exploit the error detection/correction properties presented in reference [50] while still maintaining the PAPR suppression. The search for better sets of code words with error detection/correction properties is addressed in later papers.4) n= N / 2 x (t ) = ∑ n =1 d n ( t ) e+ j ( 2 n −1)π f st + d − n ( t ) e− j( 2 n −1)π f st (4.4) This form gives a term in the summation as 2 carriers that are equidistant from and on either side of the centre frequency of the complex envelope representation. Polyphase weighting codes are applied to the encoded data (chosen from a low PAPR set) which are known at the receiver. the work is limited to 8 subcarriers. [51]. Thus they can be compensated for without affecting the distance properties of the code. Still.3) as (4. If the dn data words are chosen so that the resultant vector lie on one of two orthogonal lines. 50] presented another paper. The authors found sets of phase values (calculated offline) which are known at both the transmitter and receiver that reduce the PAPR without affecting the error correction properties. in which they describe Combined Coded OFDM (CCOFDM). then the PAPR is less than or equal to N/2. The reduction 59 .
8 PSK. Also where the envelope power is below the threshold these subcarriers are increased to obtain equality.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques can be improved slightly by using smaller phase shifts such as 8PSK and 16PSK. ReedSolomon codes are examined with parameters N=16 and M=4. They are grouped into equivalence classes where messages which have the same PAPR are in the same equivalence class. They show that further redundancy provides little benefit in terms of PAPR reduction. The PAPR is reduced significantly however this comes at the cost of a reduction in the SNR which becomes more pronounced as the number of subcarriers is increased. Reference [57] presented an idea again based on reference [49] where vectors or messages of data which exhibit a high PAPR are attenuated. Going back in time another early paper [55] extended the work in reference [49] by identifying Quadature Phase Shift Keying (QPSK) message structures which produce high PAPR. This concept is used to create code sets with low PAPR properties. This allows the PAPR to be reduced without affecting the net bit rate. analyzed with different coding rates. then the envelope remains unchanged. 60 . 56] to provide bounds for the PEP of OFDM using basic coding techniques as described in reference [49] for up to 16 subcarriers. The amplitude of subcarriers which are above a given threshold are uniformly reduced to achieve equality between the maximum of the envelope power and the threshold level. It is stated that if sub carriers in the same group (where a group is a set of subcarrier frequencies sharing some relation) are phase shifted by the same amount. Results are given for a BPSK and QPSK with less than 20 subcarriers. The authors elaborated on the results given in [52] in [54] where the computation of the PAPR is given in a rigorous mathematical proof. Reference [58] is another early paper which draws a link between the number of subcarriers and the mapping type used (4 Phase Shift Keying (PSK). Up to N=5 subcarriers are They show that that a small amount of redundancy can significantly reduce the PAPR. They prove through analytic and simulated results that only 4 bits of redundancy are required to reduce the PEP to within 10% of its optimum value as the number of subcarriers is increased. The authors of reference [55] developed their results in references [55. etc).
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques They also note that the constraints on the code can be loosened up if at least some variation in the envelope is allowed.6) and an is the input data sequence to the IFFT. 4.6) shows that binary or polyphase sequences with low out of phase aperiodic correlation values can be used to construct low PAPR signals.6) is useful as it is a much less complex process to calculate than (3. The problem remains to find sequences that reduce the PAPR.3) can be bounded as (4. Duals of primitive BCH codes are identified as good error correction codes. (4. Reference [60] presented an interesting paper which clarified the relation between the PAPR and the out of phase aperiodic autocorrelation of the message or data sequence.5) N −1 k =1 ζ ≤ 1+ 2 N ∑ ρ (k ) (4.3 Cyclic Codes Reference [61] developed a simple method based on ¾ rate code rate cyclic coding which can reduce the PAPR by 3dB when the number of subcarriers is a multiple of 4. (4. which could allow large reductions in the PAPR. 4. but no concrete level on the PAPR has been proven. The PAPR as defined in (3.2 Bounds on PAPR Reference [52] provided an efficient computational method for finding offsets which can be used with coding in OFDM. Reference [59] addressed this issue by providing bounds for the PAPR for different error correction coding schemes coined trace codes. 61 .1.3). …N1 (4.5) where N −k n =1 * ρ ( k ) = ∑ an + k an for k=0.1.
By dividing the OFDM frame into subblocks larger number of subcarriers can be used while still maintaining a reduced PAPR. which is a problem with other earlier coding schemes. In [64] a ½ rate code is employed for QPSK signals with 16 information carriers giving 32 subcarriers in total. The authors also state that Newman phases which vary quadratically exhibit a lower CF than ShapiroRudin sequences which vary linearly for all N.6dB using Newman phases for an arbitrarily large number of subcarriers. Reference [63] stated that a multitone signal can achieve a CF under 6dB for ShapiroRudin phases and around 4.1. The method is simple to implement. Reference [63] closes the paper with an open question as to whether the CF can be reduced lower than 3dB. 4 bit input messages are concatenated to 4 bit codewords which are determined according to the message data.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques The phase of every 4th subcarrier is calculated so as to minimize the amplitude giving a ¾ rate code. as well as identifying some very pertinent parameters in the makeup of multitone signals. A simple digital circuit using eight two input XOR gates can be used for the encoder. Reference [62] achieves the same results as [61] but with reduced complexity.4 ShapiroRudin codes A very early paper [63] applied ShapiroRudin and Newmen phases to multitone frequency response testing. The idea is developed to optimize the positions of the odd parity checking bits for further PAPR reduction at the cost of the introduction of side information to inform the receiver of the positions of odd parity checking bits. It was also noted that the set of tones needs to be a power of 2 (which is the case in all OFDM standards) in order for the codes to be optimum. 62 . 4. The authors also introduced Sub Block Coding (SBC) where systems with a large number of subcarriers are divided into subblocks with the last bit of each subblock altered according to the method described in [49]. Eleven years later reference [64] revisited the application of ShapiroRudin sequences to reduce the PAPR in OFDM. Many definitions which were to become convention in later papers as far as the statement of the problem of large PAPR. Although the application was not for multicarrier applications the theory still holds with multicarrier signals.
Therefore the authors developed a more suitable coding method by recognizing the relation between Golay complementary sequences and 63 . The author notes that the scheme may be unfeasible for a large number of subcarriers as the length of the codes is the same as the number of subcarriers.1. An algorithm was developed where certain subsets of codes up to length 16 have a minimum distance of half the code length and have a PAPR of 3dB.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques Results taken from a limited subset of possible messages suggest that the coded signal will reduce the PAPR to ¼ of the original uncoded message.5 Golay complementary codes Golay codes are linear binary block codes which are the only non trivial example of a perfect code. However the new decoding algorithm has 3dB worse performance than the optimum maximum likelihood detection. In reference [65] the use of Golay complementary codes is examined.6 ReedMuller Codes Golay complementary sequences were further developed for PAPR reduction in reference [66] where again the PMEPR is found to be at most 2 (3dB) when the data sequence is constrained to be a member of a Golay complementary pair. However reference [65] proposed ways to nullify this effect at the cost of the PAPR reduction made and error correction capability by breaking the total number of subchannels into smaller groups and applying a complementary code to each group.1. 4. As every codeword lies within distance 3 of any codeword they can be used in conjunction with maximum likelihood detection [16] for decoding which is not overly complex to implement. Existing Forward Error Correction (FEC) codes are incorporated into PAPR techniques with a new decoding algorithm developed which utilizes the efficient inverse WalshHadamard transform. 4. However Golay pairs have a high overhead in terms of redundancy and may not be a practical coding solution for OFDM. Golay sequences were first recognized to have good PAPR properties for application in OFDM in reference [50].
Results indicated that increasing the code length improves error correction capabilities for high SNR values.m). and it was shown that to maintain the same BER the code length for 8PSK must be increased. PMEPR reduction made. However a limitation which the authors note is that the codes are limited to 32 subcarriers where the resulting code rate will be high. The number of subcarriers is chosen such that N=2m. m. As the code rate decreases performance is improved. however this gain drops to 4dB for the simulated results. The complexity and performance of their decoding algorithm is compared to the standard maximum 64 . QPSK and 8PSK are compared in terms of their BER. RM(2. Basically this paper formalized the results given in reference [66] and extended the results for larger sets of variables such as an increase in the PSK mapping type. but at low values of SNR the uncoded system has better performance. The authors of reference [66] advanced their work further in reference [67] providing many mathematical proofs for Golay sequences. It was shown that at 40dB ACI the new coding scheme had around a 12dB gain in IBO over an uncoded system in experimental results. i. code rate.m) allow the code rate to be improved by allowing a slight increase in the PMEPR. and 8 PSK mapping types.m) further developed the theory of reference [65] by allowing a trade off to be made between the Hamming distance.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques second order ReedMuller codes (RM(2. and different code lengths. Reference [69] also uses ReedMuller codes in a simulation environment to test their performance in the presence of AWGN to determine the BER with QPSK. the length of the code. in the binary case Golay sequences occur as cosets of the first order ReedMuller code within the second order ReedMuller code. Standard decoders can be used in the receiver for the RM(2. decoding using fast Hadamard transforms. RM(2. Reference [68] examined the ReedMuller coding scheme presented in reference [67] through simulation of an end to end system with various non linearity’s and 16 subcarriers.m)). Reference [70] also enlarged on the work of [67] by developing new decoding algorithms with generalized fast Hadamard transforms. ReedMuller codes.e. and the number of phases allowed in the PSK mapping type.
The authors point out the inherent problem with coding schemes i. The performance is evaluated in an AWGN channel with PSK modulation and 8 and 16 subcarriers.xiN −1 ) where the elements of xi are in Z4.. 2. 50] by finding codes that simultaneously reduce both the PAPR and the ICI which is introduced by frequency offset between the transmitter and receiver. Any QPSK sequence a = ( a0 a1.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques likelihood method.. Suboptimal algorithms are also presented with reduced complexity which were shown to have minimal degradation over optimal methods. a balance 65 . where xi ∈ Z 4 = {0. As minimum PICR and PAPR do not occur in the same code word.7) where the QPSK constellations can be expressed as j xi .. the decrease in coding rate with the increase in the number of subcarriers but state as others have that the PAPR reduction gained can be traded off for reduction in complexity and code length. Most practical OFDM standards use QAM mapping [53].e. This is achieved by using shift and rotation operations as defined in (4.5) block codes utilizing 8 and 16 PSK show that with partial order decoding almost optimum results are achieved compared to the union bound. Another paper to address decoding issues is [71] which used soft decision decoding methods for block codes.3} .. They define a new measurement term. BERs for (8. Reference [73] extended on the coding theory of references [49. Reference [72] addresses this limitation by combining QPSK constellations to form any Mary QAM constellation.aN −1 ) can be associated with another sequence xi = ( xi0 xi1. Complexity compared to MLD methods is greatly reduced at the expense of a minor error increase. A limitation of references [67. 68] is that PSK mapping types are used.1.7) 2 xi1 π j QAM M = ∑ ( 2i1 ) j exp 2 4 ii = 0 n −1 2 ( ) (4. PeakIntercarriertoCarrier Interference (PICR) to quantify their results. Golay and ReedMuller sequences can now be created using the theory developed in reference [67].4) and (16. The new block coding method coined Ordered Statistic Codewords (OSD).
k .k .k ∈ [ 0. The transmit signal with the lowest PAPR is chosen for transmission. X m.2) and is reproduced in (4. i. It is also worth noting that none of the papers referenced consider the effect of over sampling. 4. Simulations performed on a rudimentary system with 16 subcarriers and BPSK modulation in AWGN show that the PAPR is reduced by 3dB with a decrease in the PICR.2. Their performance would be largely negated in a practical communications system due to the interleaving stage which follows the coding. 4.e N where 0 ≤ n ≤ N − 1 (4. The basic premise of MSR is to produce a set of alternative transmit signals seeded from the same data source.e. Coding techniques while popular in early OFDM papers have since fallen out of favour.8) The signal to be transmitted is broken up into several sub blocks. The original signal is given by (2. φm .1 Partial Transmit Signals Reference [74] first proposed Partial Transmit Sequences (PTS).Chapter 4: Peak to Average Power Solutions – Distortionless Techniques is found where both are at an acceptable level.k Next a constant phase rotation. X m = ∑ X m . pm . 2π ) 1<k<V is 66 . The authors note that finding codes for higher mapping types and more subcarriers is difficult. PTS generates a signal with a low PAPR through the addition of appropriately phase rotated signal parts. All subcarrier positions which are occupied in another block are set to 0. Various techniques are used to encode the alternative sets of transmit signals. of length N/V (where N is the number of sub carriers and V is the number of sub blocks).2 Multiple Signal Representation Multiple Signal Representation (MSR) techniques are another distortionless method for PAPR reduction. which are encoded in such a way so that they will have different PAPR properties.k = e k =1 V + jφm .8) N −1 k =0 j 2π nk m xm = 1 N ∑X .
xm .V } = arg min max ∑ pm .2 shows a block diagram of a PTS transmitter.9) ɶ The information in X m is the same as X m but with an added phase rotation. k . which must be known at the receiver.k k =1 V (4.12) ɶ ɶ xm = ∑ pm .k } = ∑ pm. to produce another alternative transmit signal. An IFFT is performed on each subblock which are then all summed together to create a possible transmit symbol.IFFT { X m .k .Chapter 4: Peak to Average Power Solutions – Distortionless Techniques performed on each subblock except for the first one which is kept constant. is obviously the one where the PAPR is minimized.xm .10) The process is repeated again with a different phase rotation.k { 0 ≤ n < N −1 k =1 ɶ ɶ { pm .k . Therefore the actual transmit signal is given as (4.10) ɶ xm = ∑ pm .1 … pm . pm. The optimum parameters for the transmit symbol are (4. pm.k k =1 V (4.k .V } (4.1… pm .xm . giving (4. (4.k .11) ɶ The optimum phase angle.9) ɶ X m = ∑ pm . X m .k k =1 k =1 V V (4. k . 67 . k .12) Figure 4.11) V ɶ ɶ pm .
In this case only the block partitioning need to be known at the receiver and one subcarrier in each subblock must be left unmodulated as a reference carrier. ± j} .2 ɶ xm + X m. i. If each phase rotation is chosen from a set of W admissible angles then the required number of bits for side information is. 68 .2: Block diagram of PAPR reduction using the PTS approach. Reference [75] noted that explicit side information can be avoided if differential encoding is used for the modulation across the subcarriers within each subblock.2 IFFT IFFT xm.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques X m.1 Bit source Coding Mapping Sub block partitioning xm. PTS requires side information to be sent to the receiver to inform it of the phase rotation used so the data can be decoded.e.V IFFT xm.1 X m.V ɶ pm. In order to reduce complexity the phase angles should be restricted to {±1. W=4. this allows multiplications to be performed with sign changes.V Peak value optimization Side Information (if needed) Figure 4. Simulations shown in Chapter 5 reveal that increasing the number of allowed phase angles has a minimum impact of PAPR reduction.2 ɶ pm. Rap=(V1)log2W bits per OFDM symbol. Reference [74] noted that the number of angles should be kept low to keep the side information to a minimum.
It should also be noted that the data can be divided into sub blocks in different ways as noted in reference [76] and shown in Figure 4. Also computationally efficient IFFT’s can be used to exploit the number of zero’s in the PTS subblocks. Results from reference [74] indicate that pseudorandom partitioning is 0. 69 .Chapter 4: Peak to Average Power Solutions – Distortionless Techniques PTS is flexible as the number of blocks and phase rotations can be increased providing more alternative transmit signals to choose from. Different PTS sub block structures have varying performance with pseudo random having the best and interleaving having the worst. It should be noted that no guaranteed level of PAPR reduction can be provided with MSR techniques. Also.1. The simulation model is described in Figure 5.) and pseudo random PTS is the most complex (random sequences require more hardware complexity to implement in this case).5 show CCDF results with varying factors of V and W. especially with an increase in V and W.4 and 4. all they can do is reduce the probability of large peaks. The disadvantage of this scheme is the complexity.5 to 0.3. a large amount of memory is required to store the alternative transmit signals (if check performed in parallel) in order to compare them to find the one with the lowest peak value. Alternatively the optimisation can be performed in an iterative fashion where the current best transit signal is stored until a better one is found. Of course there is a trade off with complexity. interleaved sub blocks are the least complex PTS structure to implement (The size of the IFFT’s can be halved by interleaving the input data to the IFFT and performing the last stage of the IFFT operation at the IFFT output [77].9dB better than adjacent partitioning. N is set at 64. at the cost of increased latency. Figures 4.
Chapter 4: Peak to Average Power Solutions – Distortionless Techniques X (0 ) n X (1 ) n X (2 ) N 1 n (a ) X (0 ) n X (1 ) n X (2 ) N 1 n (b ) X (0 ) n X (1 ) n X (2 ) N 1 n (c ) Figure 4.3: An example of the 3 main PTS structures: (a) Interleaved (b) Adjacent (c) Pseudorandom. 70 .
4: Simulated CCDF for PTSOFDM with V=2 and varying W. adjacent subblock partitioning. W=8 10 1 Pr(ζ > ζ0 ) 10 2 10 3 10 4 2 3 4 5 6 7 8 9 10 11 12 ζo (dB) Figure 4. W=4 V=2.5: Simulated CCDF for PTSOFDM with W=4 and varying V.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques 10 0 Uncoded V=2. adjacent subblock partitioning. W=4 V=3. 71 . N=64. W=4 V=4. N=64. 10 0 Uncoded V=2. W=4 10 1 Pr(ζ > ζ0 ) 10 2 10 3 10 4 2 3 4 5 6 7 8 9 10 11 12 ζ0 (dB) Figure 4.
The new method coined Concatenated Pseudo Random Subblock Partition Scheme (CPR SPS) divides the OFDM symbol into multiple disjoint subblocks and assigns signals randomly in each subblock. the discrete PAPR. to explore the PTS approach looked at alternative ways to create subblocks.10) A reduction of 2. Oversampling of PTS is explored in more depth in Chapter 5. However this process is complex. Another paper.5dB can be achieved when using the optimisation criteria of (4.3 oversampling at the IFFT is important to get a true reflection of PAPR reduction.10) at Pr(ζ>ζ0)=105.1… pm .2. The partial sub band is then 72 . peak regrowth becomes more pronounced with a sharper rolloff. for QPSK the calculation of λk can be achieved with less multiplications and replaced with integer additions giving a total complexity of O(4V1N2). Reference [34] first identified this issue and showed that for a PTS system with V=4. oversampled signal is only 1dB better than the uncoded case.2 Oversampling PTS Early on in the development of PTS a problem was identified which skewed the true gains of the proposed technique.e. λk . W=4 the PAPR reduction gained using the True Peak Factor (TPF) i.1 … pm .k .V } = arg min ∑ λk { ɶ ɶ { pm .e. i. There is a 3dB difference in the reduction between TPF and Lower Peak Factor (LPF). of the filter. [78]. α. N −1 ɶ ɶ pm . As already mentioned in section 3. In reference [34] it is shown that the true peak will move away from the discrete sampled points. Reference [34] goes on to suggest a new optimisation technique based on the aperiodic autocorrelation of xm .Chapter 4: Peak to Average Power Solutions – Distortionless Techniques 4. Reference [39] suggests that oversampling the IFFT by a factor of 2 (zero padding) will reduce the peak regrowth affects after filtering. Reference [39] acknowledges the limitations of discrete sampled PTS and notes that the mismatch between discrete and continuous CCDF’s occurs due to DSP filtering after the IFFT.V } k =1 (4.
Results indicate that CPR SPS can achieve similar results to the optimum pseudo random block allocation method but with a decrease in complexity. This method is coined Suboptimal Exhaustive Search (SES) algorithm and exhibits good performance with a minimal number of trials. especially when efficient FFT structures such as the CooleyTukey algorithm are used. partia l sub ban d d up lica tio n X (0 ) n X (1) n X (2) N ’1 N 1 n Figure 4. For a significant reduction in complexity C=16 has similar performance to the adjacent subblock allocation method. i. 80] also examined the PTS complexity reduction. the number of subblocks.6 for 3 subcarriers. C=1 is equivalent to pseudo random PTS.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques duplicated and assigned to the rest of the subbands repetitively to form a complete sub block. The basic principle of these papers is to produce a suboptimal iterative algorithm which uses a similar structure to PTS but with reduced complexity. C is the concatenation factor.e. References [79. One of the problems of PTS is that in order to work out which 73 .6: Generation of subblocks for PTS using Concatenated Pseudo Random Sub Block Partition Scheme (CPR SPS). Best results are obtained with low values of C (2 or 4) which results in a marginal decrease in complexity compared to pseudo random PTS. This procedure is depicted in Figure 4.
1). ±j] provides a minor improvement for higher values of V and up to a 1dB improvement at low values of V.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques transmit signal has the largest peak. However the optimisation process has been reduced to N2 additions. Note that in the simulations an oversampling factor of 4 is used.9) then the phase optimization values are restricted to pm . The SES algorithm works as follows: 1/ Assume pm. and N=256 subcarriers exhibited around a 1dB degradation compared to standard PTS with adjacent subblock partitioning at Pr(ζ>ζ0)=103. a considerable saving in terms of hardware operations. no optimisation). The new method called Optimal Limited Search (OLS) also 74 .k=1 for all k and compute the PAPR of the combined signal (ie. V. Various parameters were varied in the simulation such as a) the number of sub blocks. Simulations for V=16 sub blocks. V. The process of finding large peaks contributes significantly to the complexity of PTS. Reference [81] again looked at developing optimum phase factors for PTS with reduced complexity. If the new PAP is lower than the previous value use b1 as part of the final phase sequence. However as V is increased the performance improvement becomes less pronounced. b) number of allowed phases. If the time domain signal is given as (4.k = [ ±1] . pm. otherwise use the original value for pm. The SES algorithm has similar performance to traditional PTS with optimum phase selection (within 1dB of optimum solution at Pr(ζ>ζ0)=104) but with reduced complexity.k=1) and recompute the resulting PAP for the first value in each sub block. It was concluded that the performance improves with an increase in V. invert the first phase factor (pm. 2/ Next.k for optimization and c) data constellation size. every sample in each possible signal has to be checked to work out its peak value. which is sufficient to catch all peaks. The data constellation size has a minimal effect on performance. Increasing the number of phases to four [±1.1 (ie. However complexity increases substantially with an increase in V. 3/ The algorithm continues until all V points have been given this treatment.
14) nearly always yields the maximum amount of amplitude cancellation for the ith signal sample.. AMLN e jφM (4. .14) The minimum amplitude sum for one IFFT point is then given by (4. Next all LN solutions are calculated and the one that minimises the maximum signal samples is chosen.15) The phase selection (4.15) Si (φ ) = Ar1i − Ar 2i + Ar 3i − ... The output of the PTS algorithm can be written as (4..... A1LN A21 A22 . A2 LN .7 shows a block diagram of an OLSPTS transmitter.. L.11) A11 A S = 12 . .11) where S = S1 (φ ) ....... AM 2 e jφ2 .. Figure 4. (4. > ArMi (4..13) then chooses rli φrl = π − ∠Arli −∠A l = 1. l = 2..Chapter 4: Peak to Average Power Solutions – Distortionless Techniques addresses the oversampling issue by including an oversampling factor..... (4... 75 .. AM 1 e jφ1 .. S LN (φ ) T (4. .. 4....3.....13) Ar1i > Ar 2i > . OLS sorts same subcarrier positions over V IFFTs in order of magnitude as depicted in (4..12) contains the optimized signal samples.
OLS out performs the methods of references [79. N=256. 80] for larger values of V (>8).554.432 1. Table 4.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques A11 X1 S/P IFFT A12 A13 S1(Φ) Data source X A1LN S2(Φ) S3(Φ) Phase optimisation Av1 Av2 S/P XV AvLN IFFT Av3 SLN(Φ) Figure 4. also the number of iterations is independent of V. The complexity of the 3 preceding methods is summarized in Table 4. The authors state that OPPTS can reduce the PAPR within 0. The effect of quantization and the PSD is also examined for OLS. W=2 V=2 OBPS (Huber et al) 33.52*1070 33. 76 .40.048.576 Reference [82] developed another scheme to find optimal phases based on orthogonal vectors called Orthogonal Projectionbased PTS (OPPTS). Simulations (QPSK N=256) show OLS to be slightly better than the OBPS of reference [74] for V>4 and much better for V<4.576 V=16 5.432 SES (Cimini.048. et al) 2048 OLS (Tellambura et al) 1.1: Number of complex multiplications and magnitude operations required.8dB (using adjacent partitioning) of OBPS using pseudo random partitioning and is suitable for a large number of subcarriers. L=4.1. Quantization is shown to have a negligible effect on performance and the out of band radiation is also shown to be better than the OBPS case.554.7: Block diagram of OLSPTS transmitter.
The idea stems from the fact that as the PAPR is determined by the sequence of the transmit data vectors. and the transmit symbol with the lowest PAPR. P ( ) = P0u . The basic idea of SLM is to produce U alternative transmit sequences seeded from the same data source and then to select the transmit signal exhibiting the lowest PAPR. PD −)1 .1 ≤ u < U (4. is multiplied subcarrier wise with each one of the U vectors. Xm.k = X m . pseudo random fixed vectors are u (u generated.8.1 ≤ u < U . Note that no oversampling is used in OPPTS. ( resulting in a set of U different possible transmit symbols. xm ) = IFFT X m ) . ϕk( ) ∈ [ 0. multiplying the data vectors by some random phase will change the PAPR properties after the IFFT. Mathematically.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques The complexity of OPPTS is better than OBPS when a large number of subcarriers and subblocks are used (N=128. xm .… . all U possible transmit vectors are transferred to the to the time domain via the (u (u ɶ IFFT. 4. P ( ) . Xm. X m ) as depicted in (4.k .3 Selective Mapping In reference [83] another multiple signal representation method is presented called SeLected Mapping (SLM). The complexity of OPPTS is also slightly lower than SES for high N and V. 77 . V<=5). a set of U markedly different. 0 ≤ k < N .16) Next. Pk( ) = e+ϕk . The optimum phases are chosen from a set of 32 selected phases. is { } chosen for transmission. u u 0 ≤ k < N .2. but improvement using OPPTS reduces with an increase in V. A suboptimum solution for OPPTS approach to find near optimum phases with negligible affect on performance is also presented. u u u (u ) The data. W=4. A SLM transmitter block structure is depicted in Figure 4.Pk( ) . 2π ) .16) u ( ) X m .
At Pr(ζ>ζ0)=104 the PAPR is reduced from 2 to 4. Again oversampling is seen to have a negligible affect on the PAPR. Simulations performed with N=128 and 4PSK mapping indicate that at Pr(ζ>ζ0)=104 . 2. 78 . the purpose of this plot is to show that oversampling still has a minimal affect at higher values of U.6.8: Block diagram of an SLM OFDM transmitter.5dB.e. An oversampling rate of 4 is sufficient. and 4. i. and U=1 to 32. and 8. Here it is seen that oversampling has a minimal affect on the PAPR. that oversampling and filtering does not increase the PAPR dramatically as it does in PTS (shown in Chapter 5).5dB for U=1 to U=32 respectively. and 8. Figure 4. with U=4 a PAPR reduction of ~3dB is gained compared to the uncoded case.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques ( Xm) 1 IFFT ( xm ) 1 Bit source Coding Mapping Serial to parrallel Xm p( 2) ( Xm ) 2 IFFT ( xm ) 2 p( U) Select transmit symbol with lowest PAR ɶ xm ( Xm U) IFFT ( xm U) Side information Figure 4. ± j} avoiding complex multiplications.9 shows the CCDF of SLM with N=64. only increasing it by ~0.10 shows the CCDF of SLM with N=64 and U=1 (uncoded) 3. This is one of the main advantages of SLM. This is due to all the alternative transmit signals being uncorrelated as shown in Figure 3. The oversampling rate is set at 1. os=1. Complexity reduction can also be achieved by restricting the random generated data ( to Pm ) ∈ {±1. u Figure 4.
W (PTS) and U (SLM). In terms of redundancy (side information) SLM with U=1. In reference [74] SLM is compared to PTS for various combinations of V.5 outperforms PTS using pseudorandom subblock partitioning (W=4.5) at the cost of greater system complexity.. As loss of this information (in a fading channel) means the complete loss of the transmit symbol channel coding is required to ensure correct recovery of the data at the receiver. N=64. increasing redundancy further..Chapter 4: Peak to Average Power Solutions – Distortionless Techniques Side information is also an important issue in SLM as the receiver needs to be informed which vector. 10 0 10 1 U=1 U=2 U=3 Pr(ζ> ζ0 ) 10 2 U=4 U=8 U=16 10 3 U=32 10 4 2 3 4 5 6 7 8 9 10 11 12 ζ0 (dB) Figure 4. Log2(U) bits are required to send this u information increasing redundancy. V=1. os=1. 4 alternative signals produced with PTS require 3 IFFT’s while the same performance with SLM can only be achieved with 4 IFFT’s. P ( ) .9: Simulated CCDF for SLMOFDM for varying values of U. 79 . In terms of performance PTS has better PAPR reduction if the number of IFFT’s is fixed as PTS can vary W with no additional IFFT’s. was used.
The scrambled data is then processed as usual.10: Simulated CCDF for SLMOFDM. The receiver structure complexity is hardly 80 . b( ) .11b is a variation where the linearity of the scrambler is exploited. U=1. 2. of length log2U as a prefix to the data sequence as shown in Figure 4. and 4. here only ⌣0 a single codeword q ( ) needs to be generated. The issue of side information for SLM is further explored in [84] where the need for explicitly sent side information is avoided. The process is repeated with the other different U labels to produce U alternative transmit sequences as in standard SLM. The data with prefix is then fed into a scrambler polynomial as shown in Figure 4.12.11a. and 8. U different subcarrier vectors can then ⌣0 be generated by applying U different vector mappings to q ( ) . and 8. Oversampling rates are 1. The new technique employs a u scrambling sequence and inserts U different ‘labels’.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques 10 0 os=1 os=2 os=4 os=8 10 1 Pr(ζ> ζ0 ) 10 2 U=1 U=3 3 10 U=8 10 4 3 4 5 6 7 8 9 10 11 12 ζ0 (dB) Figure 4. Figure 4. The advantage of this structure is that uth calculated vector mapping can be calculated once from the label b( u) and can be stored in memory. 3. The labels drive the scrambler into one of U different states before scrambling the data itself.
At Pr(ζ>ζ0)=105 the new SLM algorithm can 81 . results were shown after the transmit filter.12: Scrambler polynomial for new SLM technique. (a) A( 0) 0 Label inserter Scrambler Mapper(0) S/P IFFT P/S a( 0) ⌣0 q( ) A( Mapper(UU −1) Select ɶ a RF up conv. q( u) ⌣u q( ) Delay Delay Delay Delay Figure 4. S/P IFFT P/S a( U −1) (b) Figure 4. The data was then interpolated by a factor of 8 and a RRCF with a rolloff of 0. only the label needs to be removed after descrambling. a) serial form.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques increased at all. Simulations were performed with N=256 carriers. b( 0) b( U −1) b( q u) Label inserter Scrambler ⌣u q( ) Mapper A( u) a( S/P IFFT 0) P/S Select ɶ a a (u ) a( U −1) RF up conv. 16 QAM mapping is used.12 was used. The IFFT was oversampled by a factor of 2 yielding more accurate results. b) parallel form. The complexity of the new scheme is increased by around U compared to standard SLM. 219 of which are active (to decrease the design constraints of the transmit filter).11: SLM transmitter block diagram employing technique to avoid explicit transmission of side information.
In order for the method to work cn e jφn ∈ Q .13: a) Receiver structure of the proposed SLM system. where Q is the u constellation mapping type.2dB.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques reduce the PAPR by 1. but the effect is negligible as the BER is degraded by 0. The impact on the BER is also analyzed where perfect knowledge of the channel and noise is assumed.5dB for U=4 and 8 respectively. Figure 4. Reference [85] again looks at the issue of side information in SLM and proposes a variation of reference [84] called Blind SLM (BSLM) which does not require the labels to be sent with the data for descrambling in the receiver. The U sets pseudo scrambling noise vectors are restricted to a known set at the receiver and all U sets are sufficiently different. y Y FFT ⌣ ˆ q ˆ q RF down conversi on S/P P/S Demapper Descramble r Label dumper ˆ q a) ˆ q ⌣ ˆ q Delay Delay Delay Delay b) Figure 4. At the receiver the decision metric is (4. The descrambler can multiply errors in receiver.17) 82 . b) Descrambler polynomial at the receiver.13 shows the receiver structure of the new SLM technique. PSD plots showed that between 1 and 2dB can be saved in power backoff of the LPA with 4 bits redundancy per OFDM symbol.8 and 2.
PU1. P2. X is the transmit data before the IFFT. . C km .17) is performed for P1..14. BSLM slightly outperforms the standard SLM approach. However there is a slight degradation when the new algorithm is used in a fading channel. c1. k ∉ ℜC (4.u∈{0.1.18) where Q is the IFFT matrix. Simulations of the BER revealed that the new technique has almost the same performance as SLM with perfect side information available with a infinite backoff in the LPA. and C are the PRT’s.3 Tone Reservation/Injection Tone Reservation (TR) and Tone Injection (TI) were first introduced in reference [86] and further detailed in reference [30].cN −1.3.….. A suboptimal metric is also presented to reduce complexity in the receiver. When a SL is employed with varying backoffs.U −1} ˆ ˆ min ∑re n u − jφnˆ ˆ − H n cn (4. 4. 4.. These methods use an iterative algorithm to provide increasingly better CCDF results.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques N −1 2 D= c0. the global minimumdistance solution provides the best estimate for the transmitted data. ∈C ˆ ˆ ˆ n =0 Pu . are set aside for PAPR reduction as shown in the transceiver block diagram in Figure 4.. k ∈ ℜ X km + C km = m X k . The signal plus Peak Reduction Tones (PRTs) are represented in (4.1 Tone Reservation In TR subcarriers..19) 83 . called Peak Reduction Tones (PRT’s).18) x + c = Q( X + C ) (4.
where C k = 0 . i. 84 .e.. which introduces redundancy..21) Vector c is computed so that the maximum peak value is reduced as (4.21) PAR{x + c} = ε x [ ]/ N 2 ^ ^ x+c 2 (4.. i L } and Coefficients for Ck can be found with a reduced complexity iterative algorithm which achieves a solution close to the optimum in a few steps to reduce the PAPR. which can also be solved as a Linear Program (LP)..22) ∞ Solving x + c 2 is a convex problem. When the terms are expanded out it can be solved as a Quadratically Constrained Quadratic Program (QCQP). The new transmit signal has PAPR defined as (4.. k ∈ {i1 .Chapter 4: Peak to Average Power Solutions – Distortionless Techniques X0 0 X2 IFFT N P/S x(t) 0 C1 X0 N S/P 0 C1 0 IFFT FFT X2 CN1 N P/S c(t) CN1 Figure 4.... k ∉ {i1 . This method is distortionless as the data lies in disjoint frequency bins. The complexity of the LP is O(NlogN).22) min x + c c ∞ = min x + Q C ^ C (4.. X km C km = 0 .. where the optimum solution lies at the bottom of the parabola. i L } X k = 0 .14: Block diagram of a Tone Reservation (TR) OFDM transceiver.
1dB away from optimum solution. As with PTS it was found that random positions of PRC gave the best peak reduction results of 6. i. i. For example with L/N= 5%. The performance of TR is influenced by the position of the peak reduction tones. 1 iteration gives a 2dB reduction at Pr(ζ>ζ0)=105 .e.23) ∑ DRL = ∑ L k =1 ik N −1 k =0 k b (4.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques L/N is the ratio of redundancy where L is the number of PRT’s and N is the total number of data tones or subcarriers. QPSK as opposed to 16 QAM subchannels.e.8dB reduction at Pr(ζ>ζ0)=105 .23) b Simulation results indicated that for a L/N ratio of 5%. Basically an increase in the number of PRT’s improves the PAPR reduction made. All simulations are based on N=512 subcarriers.2dB at Pr(ζ>ζ0)=105 probability. To reduce the bit redundancy. which is 2. The Data Rate Loss (DRL) can be given as (4. 85 .5dB away from optimum solution.4dB. after 40 iterations the new peak value is still 0. Adjacent PRTs on the other hand only had a peak reduction of 3. PRTs can be assigned to subcarriers which have a small number of bits assigned to them (in the case of ADSL). 5 iterations provides a 3. The complexity can be reduced using general purpose iterative algorithms which find suboptimal solutions such as the gradient algorithm. the clip probability is reduced from 15dB to 9dB at Pr(ζ>ζ0)=105. These results are for the optimum solution where the complexity is quite high. The number of iterations required to achieve various peak reduction values for random PRT also has a pronounced affect on the performance of TR. For a L/N ratio of 20% the clip probability is reduced from 15dB to 5dB at Pr(ζ>ζ0)=105. Note that law of diminishing returns applies in that the more iterations performed the less the PAPR reduction achieved.
Instead of solving for c (c=QC) [30] solves for the gradient of c giving (4.24) Including the PRT’s in the transmitted time signal x+c. to reduce the hardware requirements while still providing a good approximation to the optimum solution.25) In order to maximize the SCR the denominator is minimized. simple iterative algorithms were produced which achieved a solution close to the optimum in a few steps. By taking the gradient of the clipping noise Mean Square Error (MSE).26) ( ( ( where α nk ) = sign x n + c nk ) x n + c nk ) − A . 86 . circularly shifted replicas of the vector p 0 to ˆ ˆ0 cancel large peaks ( p 0 = Qq row ). the Signal to Clipping noise Ratio (SCR) is given by (4.24) N −1 n =0 2 2 x − clip A ( x ) = ∑ ( x n − clip A ( x n )) (4.26) c ( k +1) = c ( k ) − µ n ∑α n( k ) Q qrow xi + ci( k ) >A ^ ^ (4.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques Gradient algorithm for computing c Reference [86] uses gradient algorithms with reduced complexity. The clipping noise of the gradient algorithm is defined as (4.26). Basically the algorithm searches for the largest terms in x + c (k ) and subtracts scaled. The complexity of this algorithm is O(N) per iteration compared to O(NlogN) in the optimum case. each iteration requires an IFFT to be performed. The gradient algorithm searches for the largest terms in x+c and subtracts scaled replicas of them to cancel large peaks.25) SCR = x 2 2 x + c − clip A ( x + c ) (4. ( )( ) As is seen in (4.
15: Block diagram of a Tone Injection (TI) OFDM transceiver.28) where p and q are any integer values and D is a positive real number known at the receiver.15.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques 4. as well as the average power of the transmit symbol. This process effectively increases the size of the constellation. The data can then be easily decoded correctly with a modulo D operation in the receiver as shown in the block diagram in Figure 4. Reference [30] indicated that careful selection ˆ A can reduce the PAPR by over 5dB with only a 2% increases in the average power.3. If Xk = A = dk 2 + j 3d k 2 (4. 87 . X0 X1 X2 IFFT N P/S x(t) XN1 X0+C0 N S/P C0 C1 C2 IFFT X1+C1 X2+C2 X0 FFT XN1+CN1 Mod D X1 X2 XN1 N P/S c(t) CN1 Figure 4.2 Tone Injection Another method developed in reference [86] called TI maps data that cause large peaks to new positions which will not produce peaks when the IFFT is performed.27) and A is changed so that ˆ A = A + pD + jqD (4.
16: Example of possible expansions of constellation in TI for 16 QAM.29) Q Q Q ⊗ ⊗ ⊗ I ⊗ ⊗ ⊗ I ⊗ ⊗ ⊗ I D Figure 4.30) 88 .Chapter 4: Peak to Average Power Solutions – Distortionless Techniques The possible expansions of the selected tone for a 16QAM constellation are shown in Figure 4. The new transmitted signal can be given as (4. ˆ xn = 1 ∑ (X N k =0 N −1 k + p k Dk + jq k Dk )e j 2πkn / N (4.29) The maximum peak reduction per tone shift is (4.16.30) δ = 6 M k20 2ρ M 2 k0 −1 N (4.
29) is updated. Coding techniques for PAPR reduction were first discussed in Section 4. Each iteration decreases the PAPR by around 1 dB up to a max of 6dB of reduction. n0. Assuming L duplicate points per constellation. After the maximum sample. The peak reduction factor δ decreases as N increases.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques where M k is the number of levels per dimension. if K dimensions are to be modified. 4. these codewords ensure that the PAPR after the IFFT is kept low. the tone. k0. is found (4. is located. a new value for k0 that reduces as many peaks of possible is found. Low complexity iterative algorithms are used to find approximations close to the optimum solution. The algorithm starts with the original multicarrier symbol (pk=0. The complexity in TI comes from finding values of pk and qk which produce low PAPR which in turn require (as in TR) the solving of an integer programming problem. This procedure can be repeated several times until the desired PAPR is met or the maximum of iterations is reached. The maximum peak reduction factor (assuming normalized data) is K∂ . These codes can be 89 . where L is the K oversampling factor.4 Conclusion This chapter introduced distortionless techniques for PAPR reduction in OFDM. thereby separating the original constellation point from the expanded one. If more ˆ than 1 value of xn is large. where K is the number of real/imaginary dimensions. qk=0). which has exponential complexity. Distortionless techniques have the advantage of not corrupting the data thereby maintaining a low BER but come at the cost of increased complexity and bandwidth. the algorithm must search over ( NL ) combinations for vectors pk and qk. so to keep the peak reduction factor constant K must be increased. There is no redundancy in TI as the receiver uses a modulo D operation to decode information.1 where redundant bits are added to the bit stream before the IFFT. Properly chosen.
reduce the amplitude peaks in a uniform manner to lower PAPR’s. Various techniques are used to encode the alternative sets of transmit signals. Still complexity remains a restrictive issue in coding.e. The number of alternative transmit symbols that can be produced for 1 OFDM symbol is WV1 for PTS and U for SLM. Both PTS and SLM require side information to be sent with the chosen transmit signal. An advantage of MSR over coding techniques is that the reductions are independent of the constellation type. where a set of alternative transmit signals are seeded from the same data source. Oversampling at 90 . The transmit signal with the lowest PAPR is chosen for transmission. second order Reed Muller codes were found to have excellent PAPR properties restricting the PAPR to 3dB. Golay codes and their subset. ShapiroRudin Sequences.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques combined with existing COFDM to reduce the redundancy and complexity inherent in coding.2 introduced MSR methods. and ReedMuller codes. only a shift to lower PAPR as shown in the CCDF curves. Section 5. reductions in the PAPR also become less pronounced with larger numbers of alternative transmit symbols. which are constructed in such a way so that they will have different PAPR properties. However unlike coding no specific level of PAPR can be guaranteed with MSR methods. Various codewords were presented such as cyclic codes. Golay Complementary codes. A disadvantage of coding is that the complexity becomes prohibitively high with an increase in the number of subcarriers (>32). The effect of MSR OFDM is to shift the CCDF curve from the right to left. This information must be protected as loss of the side information means loss of the whole transmit symbol. The redundancy required for PTS is (V1)log2W while SLM requires log2U bits per OFDM symbol. The effect of oversampling PTS is examined in the next chapter where filtering is shown to degrade the gains made with PTS by ‘regrowing’ peaks. and only marginally affected by the number of subcarriers. Increasing the number of alternative transmit symbols reduces the PAPR but also increases the complexity. Variations of SLM and PTS have also been presented which alleviate the need for explicit side information. This reduction could be traded off with reductions in complexity and the code length. i.
Generally SLM outperforms PTS in terms of reduction versus redundancy. both methods prevent distortion by reducing the PAPR before the HPA. Suboptimal solutions to the PAPR minimization problem were presented which successfully reduced the complexity without sacrificing the amount of PAPR reduction to a great degree. Again suboptimal solutions were proposed with similar performance to the optimal case. SLM is much more robust to filtering due to the greater independence of the alternative transmit symbols.Chapter 4: Peak to Average Power Solutions – Distortionless Techniques the IFFT is required to maintain the reductions gained. but PTS is much better in terms of PAPR reduction versus additional system complexity. Section 4. 91 .3 described TR and TI for PAPR reduction. In TI a similar method to TR is presented where most of the complexity is in the transmitter with some additional complexity in the receiver which is a simple modulo operation of the demodulated complex vectors. In TR the additional complexity is only in the transmitter.
Increasing the 92 .5 compares the complexity of standard PTS and the new techniques and Section 5. Section 5.4 provides results for the rotated PTS as well as CSS and TI.2.1 describes Cyclic Shifted Sequences (CSS) where shifts in the data are used in place of phase rotations.2.2 looks at a combination of traditional PTS using phase rotations and CSS to produce alternative transmit signals.3 describes Time Inversion (TI) where the data sequence is reversed to provide an alternative transmit signal. Section 5.6 closes the chapter with a conclusion.1 PTS can provide promising reductions in the PAPR.1 PTS subblock creation As was shown in section 4.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS Chapter 5 Cyclic Shifted Sequences and Time Inversion of PTS This chapter introduces new techniques to produce alternative transmit signals in PTS.2.3 analyses the effect of filtering on traditional phase Section 5.1 reviews standard PTS subblock creation methods. However issues such as complexity and filtering after PTS limit the techniques usefulness.2. the number of IFFT’s. and W. Two new techniques for subblock creation in PTS are described in section 5. Section 5. Section 5. aforementioned techniques when oversampling of the IFFT is done before filtering. Section 5. 5.2. In traditional PTS alternative subblocks are created by increasing V. the number of allowable phase rotations. Section 5.
2. 5. Data source Mary Mapping PTS /CSS /TI Measure CCDF Figure 5. note that the CCDF is measured after the nyquist sampled IFFT.1 Cyclic Shifted Sequences As detailed in Section 4.1 PTS produces alternative signals by breaking up the transmit bit stream and phase rotating whole parts before performing the IFFT. This limitation in PTS is the impetus for finding alternatives to phase rotations in order to create unique transmit sequences. This advantage is combined with a reduction in complexity of the new algorithms. If the number of phase 93 .2 New techniques for PTS subblock creation Of all the MSR techniques PTS suffers the most after filtering [39] as the alternative sequences are not necessarily independent of each other as in the case of SLM. Figure 5. Note that all simulations use adjacent block partitioning.2.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS number of phase rotations provides increasingly less PAPR reduction for a set V.1: Block diagram showing the simulation model of Section 5. In other words for phase rotations of 90. These phase rotations can be kept trivial if W is restricted to 4 or less. 180.2.1 shows a block diagram of the simulation model. and 270 degrees only the sign of the I or Q value needs to be changed. In addition the hardware operations are non trivial for W>4. 5. making them trivial hardware operations. The following sections propose new techniques to create alternative sequences for PTS and shows through simulation that the peak regrowth of the proposed techniques after filtering is not as severe as in traditional PTS.
5dB compared to the uncoded conventional case for CSS: V=2. No oversampling is performed. and 5. CSS. CSS: V=2.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS rotations.2dB over PTS: V=4. especially for V=4. The performance of CSS at lower probability regions should be increasingly better than PTS due to the steeper slope of the curve. S=8. also the complexity of CSS when W>4 is reduced. 4. The system represents a PTS/CSS/TI system with V=2 subblocks and W/S=4 rotations/shifts. W=8. Here we see that CSS outperforms PTS in all cases. then the rotation becomes non trivial as the I and Q values have to be manipulated requiring a more complex circuit. Figure 5. Mathematically CSS can be expressed as (5. Unless otherwise stated all simulations use N=64 subcarriers with QPSK mapping.3 compares the CCDF of CSS to PTS where the phase shifts are replaced with cyclic shifts. they are equivalent in all other aspects. Figure 5.1) Where δ(v) is a cyclic shift in the time domain. S=8 performs 0. CSS: V=4.4dB better at Pr(ζ>ζ0)=104 than PTS: V=2.2 shows a block diagram of a PTS system depicting PTS. is greater than four. the number of positions that can be shifted to provide alternative transmit signals can be up to N/2 (before identical signals are produced) where N is the number of subcarriers. CSS: V=4. W. CSS has a PAPR reduction of 3dB.1): a n = ∑ a n +δ ( v ) v =1 ~ V ~ (v) (5. S=4 has a minor improvement of 0. 94 . It seems that with CSS not only is complexity reduced but the independence of the generated alternative signals is greater thereby giving better PAPR reduction.5dB.6dB better performance than PTS: V=4. S=8. S=8 at Pr(ζ>ζ0)=104. W=4 at Pr(ζ>ζ0)=104. and V=4. S=8 has 0. CSS does not suffer from this limitation. S=8 respectively at Pr(ζ>ζ0)=104. V=4. and TI which is introduced in the next section. S=4.
B. B.0.A.B. B Correction A. B.2: Block diagram of PTS.B. For a CSS structured OFDM transmitter with V=2 and 64 subcarriers.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS Select Store v=0 v=1 IFFT Store 0.A.A.0.A’.A. A.0 Modify v=2 v=3 A. B. The effect of double shifts has also been analysed 95 .0. B b) Receiver Figure 5.0.B IFFT Store A. Increasing the number of subcarriers in CSS allows for more alternative transmit signals to be produced with trivial operations. with V=3 and W=4 only 16 alternative transmit signals with trivial phase rotations can be produced. B.A.B PTS a = ae ' jθv Correction CSS a ' n = an + m a ' n = a− n a) Transmitter TI FFT A’.B. 32 alternative transmit signals can be created. CSS and TI transceiver. In a PTS transmitter.B. A.A’ B. A.0. B.A’.A.B.0.
W=4. V=4. 5. W=8.e jφ (v) (5. S=4. The PTS signal can be expressed as (5. with no oversampling in IFFT. N=64.2. V=4. S=8. This comes at the cost of reducing the possible number of transmit signals that can be constructed. V=4. CSS (V=2.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS where it was found that large shifts produce better results. (os=1). S=8) and uncoded OFDM. adjacent partitioning. V=4.2) Adding cyclic shifts to PTS gives (5.3: Simulated CCDF for PTS (V=2.2 PTS with CSS Another variation of CSS combining the trivial phase rotations (W=4) in standard PTS with CSS shifts to create more alternative transmit sequences was first presented in [1]. 10 0 Uncoded CSS PTS 10 1 Pr(ζ> ζ0 ) 10 2 V=4 S/W=8 10 3 V=4 S/W=4 V=2 S/W=8 10 4 3 4 5 6 7 8 9 10 11 12 ζ0 (dB) Figure 5.2) an = ~ V ~ ( v) v =1 ∑an . W=8).3) 96 .
S=4 additional signals can be produced by cyclic shifting with trivial operations in the transmitter and receiver.1.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS an = ~ V ~ ( v) v =1 ∑ a n +δ (v) . reducing complexity significantly. This linear phase component must be removed prior to demodulation and requires one complex multiplication for each data symbol. The number of alternative transmit signals is now WS. W=8. Therefore for 4 corresponds to a signal shift of δ ( v ) = each of the W different phase rotations.5dB and ~3. W=4.or 3). PTS/CSS: V=2. The reduction in PAPR from the uncoded case is ~2.2.5dB for (PTS: V=2. There is a negligible performance difference in the two methods. 97 . W=4. PTS/CSS: V=2. For 16 alternative transmit sequences (PTS: V=3. W=4. This 2 N. W=4. S=2) and (PTS: V=3. The advantage here is that for 8 alternative transmit sequences (PTS: V=2. Cyclic shifting generates a linear phase shift at the output of the FFT in the receiver. This process can be made trivial if the linear phase shift is constrained to gπ radians per frequency bin (g=0. PTS/CSS: V=2. W=4.4 shows CCDF plots comparing standard PTS to PTS/CSS.2. W=8. S=4) respectively at Pr(ζ>ζ0)=104. S=4) a whole IFFT can be avoided. or 3). PTS/CSS: V=2.e jφ (v) (5.g samples (g=0.1.3) Where δ(v) is a cyclic shift in the time domain. S=2) 4 complex multiplications per sample are avoided in the transmitter. W=4. Figure 5. where S is the number of cyclic shifting options.
TI involves reversing the output sequence of the IFFT sub block before addition and transmission of the sub blocks.4) where i takes the value 0 or 1 for a time inversion of the Vth sub block in the time domain. The effect that this has on the demodulated subchannels can be reversed by 98 . V=2. TI of a subblock must be accompanied by conjugation of its elements to stop the information from jumping into image subchannels after the demodulation process. Mathematically TI can be expressed as (5. 5. (os=1). W=8. and uncoded OFDM. S=2. W=4. S=2 10 4 3 4 5 6 7 8 9 10 11 12 ζ0 (dB) Figure 5.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS 10 0 Uncoded PTS/CSS PTS 10 1 Pr(ζ> ζ0 ) 10 2 10 3 V=3 W=4 V=2 W=4. with no oversampling in IFFT. W=4).2. S=4). V=3. adjacent partitioning.3 Time Inversion Another new proposed PAPR technique based on a variation of PTS first presented in [2] is Time Inversion (TI). a block diagram description of which is shown in Figure 5. W=4. N=64. S=4 V=2 W=8 V=2 W=4.4: Simulated CCDF for PTS (V=2.2.e jφ ) (v) (5. PTS/CSS (V=2.4) a n = ∑ conj v =1 ~ V ~ i( v ) (v) (a ( −1)i ( v ) ( n +δ ( v ) ) .
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS taking the conjugate of the output after the FFT process in the receiver circuit. No extra multiply operations are required in the transmitter or receiver.
Simulation results for PTS with TI are shown in Figure 5.5 where it is seen that TI has slightly worse performance than PTS with an equivalent number of alternative transmit sequences. TI has ~0.5dB degradation compared to PTS at Pr(ζ>ζ0)=104. TI reduces the PAPR by ~2.2dB over the uncoded case while PTS has a reduction of ~2.7dB at Pr(ζ>ζ0)=104.
10
0
Uncoded PTS/TI: V=2, W=4, S=2 PTS: V=2, W=8
10
1
Pr(ζ> ζ0 )
10
2
10
3
10
4
3
4
5
6
7
8
9
10
11
12
ζ0 (dB)
Figure 5.5: Simulated CCDF for PTS (V=2, W=8), PTS/TI (V=2, W=4, S=2) and uncoded OFDM. N=64, adjacent partitioning, with no oversampling in the IFFT, (os=1).
5.3 Filtering new techniques
As noted in [34, 39] the reductions made with PTS are drastically reduced when passed through a pulse shaping filter due to peak regrowth. Furthermore as shown in Figure 3.10 peak regrowth after some form of PAPR reduction is more severe than when no PAPR reduction is performed. Certain PAPR reduction schemes such as 99
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS PTS and clipping display greater peak regrowth after interpolation and filtering than coding and SLM (Figure 4.11). This section analyses the effect of interpolation and pulse shaping filtering on CSS and TI, comparing them to standard PTS through simulation.
Simulations performed in this section use N=64 subcarriers with no oversampling in the IFFT (i.e. os=1) and adjacent partitioning. As shown in reference [39] the
partitioning type has a minimal effect on the peak regrowth after filtering. A RCF is used with a rolloff factor of 0.15 and 128 filter taps with a normalized sampling rate. The data is interpolated by a factor of 8 before filtering. The simulation model is represented in Figure 5.6. IFFT size = N×os
PTS /CSS /TI
(Figure 5.2)
Data source
Mary mapping
8
RCF
Measure CCDF
Measure CCDF
Figure 5.6: Block diagram showing the simulation model of Section 5.2.
In Figure 5.7 the discrete CCDF after the IFFT is plotted with the CCDF after filtering for PTS and CSS. Both PTS and CSS have 16 alternative transmit symbols. CSS has a peak regrowth of ~3.5dB from the discrete level and PTS has a peak regrowth of ~4dB at Pr(ζ>ζ0)=104. This is contrasted with only 1dB of peak regrowth in the uncoded case at Pr(ζ>ζ0)=104. So the PAPR reduction after filtering is only ~1.5dB and ~2dB for PTS and CSS respectively at Pr(ζ>ζ0)=104. It is also worth noting that CSS has a steeper rolloff than PTS leading to the increasingly better performance than PTS at low probability levels.
100
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
10
0
discrete filtered
10
1
Pr(ζ> ζ0 )
10
2
Uncoded
10
3
CSS
PTS
4
10
3
4
5
6
7
8
9
10
11
12
13
ζ0 (dB)
Figure 5.7: Simulated CCDF of discrete and filtered PTS (V=4, W=4), CSS (V=4, S=4), and uncoded OFDM. N=64, adjacent partitioning, with no oversampling in IFFT (os=1), interpolated by 8, filtered with RCF (α=0.15).
Figure 5.8 maintains the same number of IFFT’s (V=4) as Figure 5.7 while increasing the number of phase rotations/shifts from 4 to 8. The peak regrowth of CSS above the discrete level after filtering is 4.3dB while PTS peak regrowth is 4.1dB at Pr(ζ>ζ0)=104. The peak power reduction after filtering is only ~1.7dB and ~2.2dB for PTS and CSS. Comparing these results with Figure 5.7 it is seen that increasing the number of phase rotations/shifts to 8 provides only a further 0.2dB PAPR reduction for both PTS and CSS after filtering.
101
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS
10
0
discrete filtered
10
1
Pr(ζ> ζ0 )
10
2
Uncoded
10
3
CSS
PTS
4
10
3
4
5
6
7
8
9
10
11
12
13
ζ0 (dB)
Figure 5.8: Simulated CCDF of discrete and filtered PTS (V=4, W=8), CSS (V=4, S=8), and uncoded OFDM. N=64, adjacent partitioning, with no oversampling in IFFT (os=1), interpolated by 8, filtered with RCF (α=0.15).
Figure 5.9 compares combined PTS/CSS and PTS/TI with standard PTS, with 8 alternative transmit signals. Both PTS/CSS and PTS/TI have very similar
performance after filtering with a peak regrowth of ~2.5dB and 2.1dB respectively from the discrete level, while standard PTS is ~0.4dB worse with a peak regrowth of ~3dB at Pr(ζ>ζ0)=104 . Note that again the slope at Pr(ζ>ζ0)=104 is slightly steeper for the new techniques after filtering implying further PAPR improvement over PTS at lower probability levels. The PAPR reduction after filtering is very poor in all cases, only ~1dB for the new techniques and 0.6dB for standard PTS at Pr(ζ>ζ0)=104.
102
As stated earlier the advantage of PTS/CSS in this case is the removal of 1 IFFT operation per transmit symbol. filtered with RCF (α=0.15).8dB and ~1. N=64. S=2). The peak regrowth is ~2. In Figure 5. W=4.9: Simulated CCDF of discrete and filtered PTS (V=2.10 the CCDF of PTS and PTS/CSS is compared. and uncoded OFDM. where 16 alternative transmit sequences produced. W=4.1dB above the discrete level for PTS and PTS/CSS respectively at Pr(ζ>ζ0)=104. 103 .8dB and ~3. PTS/CSS (V=2.5dB gained in PAPR reduction at Pr(ζ>ζ0)=104 for PTS and PTS/CSS respectively. W=8). with no oversampling in IFFT (os=1). adjacent partitioning. and PTS/TI (V=2. S=2).Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS 10 0 discrete filtered 10 1 Pr(ζ> ζ0 ) 10 2 Uncoded PTS/CSS 10 3 PTS PTS/TI 4 10 3 4 5 6 7 8 9 10 11 12 13 ζ0 (dB) Figure 5. Again the PAPR reduction after filtering is heavily penalized with only ~1. interpolated by 8.
The data is interpolated by 8 before filtering with a RCF with 128 taps and a rolloff factor of 0. and uncoded OFDM. adjacent partitioning.12 shows PTS/CSS (V=2. The simulation model here is the same as Figure 5. S=4) and PTS (V=3.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS 10 0 discrete filtered 10 1 Pr(ζ> ζ0 ) Uncoded 10 2 PTS 10 3 PTS/CSS 10 4 3 4 5 6 7 8 9 10 11 12 13 ζ0 (dB) Figure 5. while reducing the PAPR after filtering. i. Oversampling increases the PAPR in the discrete domain.11 and 5.10: Simulated CCDF of discrete and filtered PTS (V=3. W=4. S=4).15). N=64.4 Oversampling new techniques Oversampling [39] is required to improve the performance of PTS techniques.e. W=4). Again N=64 subcarriers. 5. 4. with 16 alternative transmit signals.6 except that oversampling is performed at the IFFT before filtering. interpolated by 8. W=4) under oversampled and filtered conditions respectively. It is seen that oversampling by 2 brings the discrete and filtered CCDF curves to 104 . PTS/CSS (V=2. and adjacent partitioning is used. 2. Figure 5. W=4. filtered with RCF (α=0. os=1.15. or 8. Oversampling the IFFT increases the convergence between the discrete PAPR and the filtered PAPR. with no oversampling in the IFFT (os=1).
The PAPR reduction for PTS/CSS and PTS after filtering is 3.1dB) than PTS (~0. N=64. 2. 10 0 discrete filtered 10 1 Uncoded Pr(ζ> ζ0 ) 10 2 os=1 os=2 os=4 os=8 os=1 os=2 os=4 os=8 10 3 10 4 3 4 5 6 7 8 9 10 11 12 13 ζ0 (dB) Figure 5. S=4). Discrete oversampled curves (dashed) move from left to right and oversampled (solid) filtered curves move from right to left. 4. and Uncoded OFDM.3dB) at Pr(ζ>ζ0)=104. interpolated by 8. 2.15). 105 . all further simulations will use oversampling rates of os=1. As oversampling the IFFT by a factor of 8 provides minimal improvement over 4.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS within 1dB of each other at Pr(ζ>ζ0)=104. and filtered with RCF (α=0. with oversampling rates in the IFFT of 1. W=4. Furthermore oversampling by a factor of 8 brings them almost to convergence with PTS/CSS slightly closer (~0. and 8. adjacent partitioning.3dB and 3dB respectively at Pr(ζ>ζ0)=104. and 4.11: Simulated CCDF for PTS/CSS (V=2.
with an oversampling factor of 2 PTS has a PAPR of 10. S=2). and filtered with RCF (α=0. Oversampling by a factor of 2 brings the discrete and filtered CCDF curves to within 0. 4. and Uncoded OFDM.2dB). W=4.15 compare PTS (V=2. 106 . 2. PTS/CSS. adjacent partitioning. with 8 alternative transmit signals.2dB at Pr(ζ>ζ0)=104 . interpolated by 8. and PTS/TI after filtering (os=4) is 2. with oversampling rates in the IFFT of 1.14. Oversampling by a factor of 4 brings them almost to convergence (0. Figures 5. W=8). 2.2dB.5dB. and 5. and PTS/TI (V=2. while PTS/CSS and PTS/TI has a PAPR of 9. For example. The PAPR reduction for PTS. S=2) under oversampled and filtered conditions respectively. W=4. and 2. PTS/CSS and PTS/TI have slightly better performance at any given oversampling rate. W=4).9dB and 10dB respectively at Pr(ζ>ζ0)=104.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS 10 0 discrete filtered 10 1 Uncoded Pr(ζ> ζ0 ) 10 2 os=1 os=2 os=4 10 3 os=1 os=2 os=4 os=8 os=8 10 4 3 4 5 6 7 8 9 10 11 12 13 ζ0 (dB) Figure 5. N=64. Discrete oversampled (dashed) curves move from left to right and oversampled (solid) filtered curves move from right to left. and 8.12: Simulated CCDF for PTS (V=3.8dB of each other at Pr(ζ>ζ0)=104.13. PTS/CSS (V=2. 5.3dB respectively at Pr(ζ>ζ0)=104.15).
and Uncoded OFDM. 2. with oversampling rates in the IFFT of 1.15). and filtered with RCF (α=0. and 4.14: Simulated CCDF for PTS/CSS (V=2.13: Simulated CCDF for PTS (V=2. 2. interpolated by 8.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS 10 0 discrete filtered 10 1 Uncoded Pr(ζ> ζ0 ) 10 2 os=1 os=2 os=4 os=1 os=2 os=4 10 3 10 4 3 4 5 6 7 8 9 10 11 12 13 ζ0 (dB) Figure 5. and 4. N=64. and filtered with RCF (α=0. W=4. interpolated by 8. with oversampling rates in the IFFT of 1. and Uncoded OFDM. 10 0 discrete filtered 10 1 Uncoded Pr(ζ> ζ0 ) 10 2 os=1 os=2 10 3 os=1 os=2 os=4 os=4 10 4 3 4 5 6 7 8 9 10 11 12 13 ζ0 (dB) Figure 5. Discrete oversampled curves (dashed) move from left to right and filtered (solid) oversampled curves move from right to left. adjacent partitioning. N=64. adjacent partitioning. S=2). Discrete oversampled curves (dashed) move from left to right and filtered (solid) oversampled curves move from right to left.15). W=8). 107 .
W=8) requires 256 multiplications. W=4. A PTS system (N=64. interpolated by 8. and 528 comparisons as well as the removal of a whole IFFT operation. V=2. adjacent partitioning.5 Complexity evaluation The performance in terms of the CCDF and PSD of CSS. The equivalent PTS/CSS system (N=64. and Uncoded OFDM.15: Simulated CCDF for PTS/TI (V=2. 2. S=2). and S are chosen. with oversampling rates in the IFFT of 1.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS 10 0 discrete filtered 10 1 Uncoded Pr(ζ> ζ0 ) 10 2 os=1 os=2 10 3 os=1 os=2 os=4 os=4 10 4 3 4 5 6 7 8 9 10 11 12 13 ζ0 (dB) Figure 5. 5. W.15). and 178 compares are required. 170 squaring operations. and filtered with RCF (α=0. Table 5. N=64. half of which 108 . For example with a PTS system (N=64. V=2. PTS/CSS. W=4) 85 non complex multiplications. W=4. and PTS/TI has been established. V=3. Discrete oversampled curves (dashed) move from left to right and filtered (solid) oversampled curves move from right to left. and 4. S=2) requires no multiplications in exchange for an increase in 512 squaring operations.1 compares them in terms of hardware operations where it is seen that multiplications are avoided when appropriate values of V.
Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS Table 5. 256 squaring operations. S=4) but with the added advantage of the removal of one whole IFFT operation. PTS (V=3. The resultant reduction in PAPR after filtering was between 1 and 2dB for various combinations of V. 5. In the simulations a RCF with 128 taps and a roll off factor of 0. W=4. In section 5. W=4. and S with the new techniques having up to 109 .3 where it was shown that the gains made with PTS techniques are dramatically affected when passed through a pulse shaping filter. Combining PTS with CSS and TI was shown to provide similar performance to standard PTS but with the advantage of a reduction in complexity. S=2) requires no multiplications.1: Complexity comparison of various techniques (adjacent partitioning) Type PTS  Comments V blocks W rotations  Operations (N/V)W multiplications (trivial if W<4) (N/V)W(V1) squaring (N/V+1)W(V1) comparisons V*(N/V) IFFT’S (N/V)WS(V1) squaring (N/V+1)WS(V1) comparisons V*(N/V) IFFT’S (N/V)TW(V1) squaring (N/V+1)TW(V1) comparisons V*(N/V) IFFT’S PTS/CSS  V blocks W rotations S shifts V blocks W rotations T(=2) inversion  PTS/TI are complex.2 it was shown through simulation of the CCDF that CSS has better performance (up to 1dB) than PTS for the same number of alternative signals over various combinations of V and W. and 264 comparisons.6 Conclusion This chapter introduced new techniques for the creation of PTS OFDM signals. The equivalent PTS/CSS system (V=2. The performance after filtering was examined in Section 5. W.15 was used. W=4) was shown to have the same performance as PTS/CSS (V=2. and has the same number of squaring and comparison operations as the PTS system.
The effect of oversampling on PTS and the new techniques was also examined. The avoidance of multiplications and in some cases IFFT operations. The only exception is PTS/CSS (V=2. 110 . W=4) for the same number of alternative signals.7dB better PAPR for all combinations. W=4. S=4) where the performance was slightly worse than PTS (V=3.Chapter 5: Cyclic Shifted Sequences and Time Inversion of PTS 0. Oversampling by a factor of 8 (implying IFFT sizes of N×8)was necessary to bring both curves almost to convergence. This small degradation is a trade off for removing 1 of the IFFT’s. as well as a modest performance gain combine to make the new techniques viable alternatives to standard PTS OFDM. The effect of oversampling was to increase the PAPR of the discrete CCDF while reducing the CCDF of the filtered CCDF. However in a practical system an oversampling factor of 2 was shown to be sufficient bringing the discrete and filtered CCDF curves to within 1dB of each other. where it was seen that an oversampling factor of 2 was sufficient to bring the discrete PAPR to within 1dB of the filtered PAPR. Two publications resulted from this chapter.
111 . clipping and quantization.1 describes the most common form of clipping in the baseband and reviews papers which seek to quantify the effect of clipping on the BER and PSD under different OFDM system conditions. These methods include approaches such as pulse shaping (or windowing). instead they take the output of the IFFT and then limit the amplitude of large samples which invariably causes distortion.3 analyses windowing and pulse shaping techniques. and clipping at every stage from the output of the IFFT to limited backoffs in the amplifier. distortion introducing techniques. These methods do not attempt to create a transmit signal with a low crest factor.2 looks at amplifier limiting where the unconstrained OFDM signal is allowed to saturate the amplifier introducing spectral regrowth. the disadvantages are the inband distortion (increasing BER) and spectral splatter.1 looks at the relation between Section 6.1. Section 6. affecting adjacent channels by increasing out of band distortion. The advantages of clipping are a reduction in complexity and ease of implementation.Chapter 6: Peak to Average Power Solutions – Distorted Techniques Chapter 6 Peak to Average Power Solutions Distorted Techniques This chapter details another approach to PAPR reduction. Section 6. Section 6.
6dB of peak regrowth. The unclipped signal at the same percentile had a CF of 13dB.4) it was shown 112 . of 32 samples.6) with no filtering displayed both in band distortion and spectral splatter. Results also indicated that the peak regrowth after filtering was significant. As direct clipping of the samples will cause all the noise to fall in band the OFDM symbol is oversampled by a factor of 8 before clipping. The authors note that clipping causes both inband distortion.4 the out of band noise was only 16dB lower than the signal power.1) where the clipping level A is determined as CL = A . a stopband of 40dB.999 percentile.8 to 1. A if x < − A if − A ≤ x ≤ A if x > A (6. Finally the BER was examined after clipping and filtering in an AWGN channel. Filtering is performed with an equiripple bandpass Finite Impulse Response (FIR) filter with 103 taps. Filtering after clipping was noted to reduce the spectral splatter but at the cost of peak regrowth. This demonstrated the need for filtering to suppress the sidelobes. QPSK modulation was applied to the data samples before modulation with an IFFT.4 (3dB) the clipped and filtered signal had a CF of ~9dB at the 99. The complex baseband samples were modulated up to a carrier frequency 1/4 of the sampling frequency in order to reduce the complexity of the simulation model. x. where σ is the rms level of σ the OFDM signal. For reasonable clipping levels (CL>1. according to (6. for a CL=1. The real valued bandpass samples. Tg. When the filter was applied the sidelobes were suppressed to ~50dB below the signal power for the same CL.e. The simulation model used N=128 subcarriers with a guard interval.1 Clipping in the Baseband Reference [46] is an early paper looking at the effect of clipping and filtering on OFDM. A. and 1dB ripple in the passband. For a CL=1. PSD results for clipping (CL=0. deteriorating the BER and out of band noise which reduces the spectral efficiency. are then clipped at amplitude. For example. y = x.1) − A.Chapter 6: Peak to Average Power Solutions – Distorted Techniques 6. i.
Clipping is performed according to the rule set in (6.5 for square clippers and Aclip σ = 2. Another reference [88] to analyze the performance of Cartesian clipping developed analytical expressions for the PSD and SER for any Mary constellation with various clipping backoffs.0 for magnitude clippers. and included the effects of AWGN as well as clipping noise. However it should be noted as shown in Figure 3. The analytical results were compared to a simulated system with N=128 subcarriers and QPSK modulation. Square. Reference [87] presented a theoretical analysis of both cartesian and envelope clipping for various oversampling rates.e.Chapter 6: Peak to Average Power Solutions – Distorted Techniques that less than 1dB of degradation at the 102 BER level is encountered.2) Under the assumption of a Gaussian like nature of the samples (i. Optimum clipping levels were shown to be Aclip σ = 2. z > z0 (6. An oversampling factor of 2 was shown to be enough to have the same performance as an infinitely oversampled signal.2) 113 . z < z0 . As the magnitude is not required this method is much less complex to implement than envelope clipping. A moderate deterioration was seen when the oversampling factor was set at 1 for both square and envelope clipping. unlike higher modulation types is inherently impervious to clipping. Cartesian clipping is where the I and Q values of the complex sample are clipped independently. − z0 < z < + z0 .2) − y0 y = g ( z ) = ( y0 z0 ) z + y 0 . or Cartesian clipping was shown to have worse performance by around 1dB than envelope clipping (as expected). Considering z as the I or Q component of the complex OFDM signal the output of the Cartesian clipper is expressed as (6. high number of subcarriers) the impact of non linear distortion was analytically derived .17 that QPSK. however it introduces more distortion than envelope clipping at a set clipping level.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques Clipping is defined as x ibo = 20 log10 0 σ (6.5) and α = y0 x0 1 − 2Q . The Gaussian assumption only holds for hard clipping. Increasing the number of subcarriers (which increases the CF) distributes the clipping noise over more subcarriers improving the BER. σ is the variance of the x0 σ input signal. A FFT was used to demodulate the data.3) The SER was found to be SER = ps = 1 − (1 − pe ) 2 (6. and M is the Mary constellation mapping type. namely • • The BER is not uniform across all the subcarriers. • The authors note the limitation of a Gaussian assumption of noise as identified in reference [89]. especially in higher order Mary constellation (M>64). Many interesting aspects of clipping were revealed analytically and supported through simulation. A simple simulation model was created where an IFFT was used to modulate the complex samples. some subcarriers have slightly worse performance. 114 . the noise tends to have an impulsive distribution. which were then clipped using the Cartesian method.4) where pe = 2 M −1 3 Pr r > ασ M −1 M (6. at higher IBO backoffs. Q is the Gaussian error function.
L is the constellation type (eg: L=2 is 4 QAM.7) where N is the number of subcarriers. Clipping is performed as (6. the clipping level is set higher so that a low probability of error is maintained making clipping a rare event.Q 2 L 8 ( L − 1) (6. (assuming a constant constellation size on all subcarriers). In practice. The probability of symbol error for a discretely sampled signal is given as (6. The analysis treats clipping noise differently to the standard AWGN assumption which is sufficient if the clipping level is set low enough so that there are a number of clips per symbol.6) −l y = h ( x) = x l x≤l x <l x≥l (6. and Q(. Under this condition the clipping noise is of an impulsive nature as identified in reference [88] leading to a different type of error mechanism.) is the Gaussian error function. An expression for the BER is given as (6.Chapter 6: Peak to Average Power Solutions – Distorted Techniques Reference [89] also analytically examined clipping in the baseband focusing on magnitude clipping and derived an expression for the SER and BER versus the clip level. and l is the clipping level.8) 115 . Reference [89] identified from the distortion spectrum analysis that the probability of error varies across the subcarriers with the lower subcarriers dominating the errors.6) where x(t) is a continuous time baseband multicarrier signal. L=8 is 64 QAM). as well as performance in both AWGN and Rayleigh fading channels. y(t) is the clipped output.7) 1 3 2 8 N ( L − 1) 3π l Pr(error ) = Q(l ). Bounds on the probability of error due to clipping are derived for both the transmitter and the receiver.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques 1 3 2 4 N ( L − 1) 3π l Prb (error ) = Q(l ).1 it is seen that even with the pessimistic assumption of (6. The AWGN model is appropriate for hard clipping levels but it underestimates the error probability by several magnitudes at higher clipping levels.1 exaggerates the difference between the 2 equations. The advantage of (6.8) The AWGN approach is also presented for comparison and is given as (6. This is because instead of being spread uniformly over time. as is assumed in (6.9) where l2 σ = − l exp− + 2 (1 + l 2 ) Q ( l ) 2 π 2 c 2 (6. 116 . Comparing the two analytical methods in Figure 6.Q 2 L log 2 L 8 ( L − 1) (6.7). and will therefore not contribute to the in band distortion.9) the error probability is much lower than (6.9). A probability of error of 106 is probably sufficient.9) Pr(error ) = 4 L −1 3 Q σ L2 − 1 L c (6. which leads to a greater error probability.7) is that the error probabilities for high clipping levels can be calculated analytically avoiding laborious simulations times which would be required to get accurate results at these levels.10) Reference [89] noted that the Gaussian model is pessimistic as it does not take into account that some of the noise power will fall out of band. It should be noted that the yaxes in Figure 6. clipping noise is actually concentrated in time as impulses.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques
10
0
Pe vs clip level using AWGN approachMetsgadh AWGN model Impulsive noise model
10
2
10
4
10 Probability of error
6
10
8
10
10
10
12
10
14
10
16
2
2.2
2.4
2.6
2.8
3 Clip level
3.2
3.4
3.6
3.8
4
Figure 6.1: Analytical symbol error probability from (6.7) and (6.9) for 64 QAM and N=64.
The affect of clipping in the presence of channel impairments is also analyzed both through analytical methods and simulation to observe the effect at the receiver. The simulation model uses the Hiperlan2 [52] specification with a guard interval length of 16 samples, perfect synchronization and a 1 tap equalizer. Analytical analysis
revealed that the error probability is further degraded in the presence of channel fades, together with clipping at the receiver. The BER performance was degraded by 1 to 1.5dB on the lower subcarriers. The analytical model is within an order of magnitude of the simulated results.
Reference [90] presented another analytical derivation of the SER resulting from clipping in the baseband extending on the impulsive nature of noise at higher clip levels first presented in reference [89]. Reference [90] claims that the SER curves of reference [89] are too pessimistic and that their claim that approximations become tight in the higher OBO region is unsubstantiated.
117
Chapter 6: Peak to Average Power Solutions – Distorted Techniques Reference [90] proposed a different technique to limit the upper bound by using the Chernoff bounding technique. The expression for the average SER due to clipping is given as (6.11)
SER ( λ ) ≤ 4 min e − d ε >0
Nε
(
b ( λ ) + ∫ J 0 ( ε r ) p ( r + λ ) dr
0
∞
)
N
(6.11)
where J 0 ( x ) := 1
2π ∫
2π
0
e x cosϕ dϕ is the modified Bessel function of the first kind.
Unfortunately no closed form solution is provided, however the computational effort is much less compared to long simulation times required for smooth curves below 106 probability. Due to the Gaussian assumption of the non linearity the results are only within an order of magnitude of simulated results provided N>256 and the clip level,
λ , is greater than 7dB.
Further expressions for the SER in AWGN and Rayleigh fading channels using the previously described Chernoff method are provided. Oversampling at both the
transmitter and receiver is also treated, resulting in the realization that the side lobes generated by the transmitting non linearity are suppressed at the receiver by the filter. Furthermore, reference [90] claimed that the average SER is the same for the nyquist and oversampling case, only the out of band radiation is reduced by oversampling. An interesting claim which is not supported by simulations of the BER in the next chapter. The asymptotic behavior is treated and it was found that for N → ∞ ,
λN → ∞ the Gaussian model matches the Chernoff bound of this paper in that the
SER → 0 .
Reference [91] looked at the out of band radiation produced by clipping and presented correcting functions which limit the signal while avoiding out of band radiation, and keeping the in band interference to a minimum. This is relevant when oversampling is performed at the IFFT as out of band radiation is created by the clipping process. Filtering after clipping reduces the out of band radiation but regrows previously clipped peaks.
118
Chapter 6: Peak to Average Power Solutions – Distorted Techniques Two correction functions were suggested, the first is a Gaussian correcting function k(t) which is an additive correction of the OFDM signal. If the signal exceeds the amplitude threshold A0 at times tn, then the corrected signal is (6.12)
c (t ) = s (t ) + k (t )
(6.12)
where k ( t ) = ∑ An g ( t − tn ) , g ( t ) = e − t
n
2
2σ 2
, and An = − s ( tn ) − A0
(
) s (( t )) .
s tn
n
The
correcting function must be normalized so that g(0)=1, which limits the signal s(t) to A0 at the positions tn. However, the correction function may cause peaks in other positions, but this consequence is shown to have a minor effect. Other functions for g(t) are developed which cause no out of band interference and keep the in band interference to a minimum. A Gaussian function is defined as (6.13)
N −1 k =0
g ( t ) = ∑ Gk e j 2π k ∆ft g ( 0 ) = ∑ Gk = 1
k =0 N −1
(6.13)
And a sinc correcting function is defined as (6.14)
N −1 k =0
g (t ) =
1 N
∑e
j 2π k ∆ft
≈ sin c (π Bt ) e jπ Bt
(6.14)
The correcting function (6.14) can correct an amplitude peak in an OFDM signal with minimal in band distortion and no out of band radiation. Note that if the signal is not oversampled then the correction scheme is the same as normal clipping.
Simulations with the correcting functions were performed with N=128 subcarriers and an oversampling rate of 4. The signal is corrected with k(t) and any peak regrowth after the correction is clipped at A0. An IBO level of 4dB was used and the algorithm was tested in both an AWGN and fading environment. It was shown that the
119
Chapter 6: Peak to Average Power Solutions – Distorted Techniques correction functions introduce more inband noise than standard clipping, however the slight increase in inband noise is offset by the alleviation of out of band interference. For example, at an IBO of 4dB standard clipping has a Signal to Interference Ratio (SIR) of 21dB, Gaussian 13dB, and Sinc 16dB. The BER vs IBO in an AWGN channel with SNR of 18dB shows a magnitude greater degradation for the sinc correction function at 4dB IBO, but in the fading channel the BER degradation is negligible for the sinc case and only marginally worse for the Gaussian case.
Clipping and filtering issues were addressed in reference [92] where an oversampled (LN where L is the oversampling factor) IFFT zero padded in the middle is used to modulate the OFDM symbol. The resultant samples are then clipped by a SL in the normal way, as this results in out of band radiation as described section 3.4.2 the data is filtered by a FFT/IFFT pair of size LN. The filter passes the wanted in band samples while nulling out the out of band components. The advantage of this
technique is twofold, firstly out of band radiation is greatly attenuated and secondly, by oversampling peak regrowth after filtering is greatly reduced (refer to Figure 3.11).
Simulations were performed with 4QAM data and N=64 subcarriers. CCDF results for a CR of 6dB compared the new algorithm (L=2, N=64) with standard non oversampled clipping (L=1, N=64) where it is seen that the L=2 case has 1dB less peak regrowth than the L=1 case at 105 probability region. It was noted that
increasing the oversampling rate (L>2) at the IFFT provided minimal further improvement. Out of band radiation is also analyzed in the form of the PSD where a perfectly linear amplifier with a CR 1dB higher than the baseband clipping level is used after modulation with a carrier frequency. Out of band radiation was reduced down to 65dB using the new technique (L=2), compared to 55dB for clipping before interpolation (L=1) and 45dB with no clipping before amplification. In band distortion is also an affect of clipping, reference [92] stated that clipping adds a noise like component and a reduction in the constellation size (refer to Figure 3.16) which can be corrected at the receiver with AGC. Also as the noise from clipping is created at the transmitter it will lessen its effect in a fading channel. These properties will improve the BER. This technique can be implemented in existing OFDM systems and requires no redesign at the receiver only replacing the IFFT at the transmitter.
120
and that it worked better with higher order constellations with a clip level >4dB. and although they are still distorted decisions made on the new symbols are much less affected by clipping noise.1. The algorithm can predict false clipped peaks when the clipping level is set too low (<2dB) worsening performance more than standard clipping.15) 121 . The problem with the methods of references [92.Chapter 6: Peak to Average Power Solutions – Distorted Techniques Another technique to lessen the effects of clipping [93] called Decision Aided Reconstruction (DAR) reduces clipping noise with an algorithm in the baseband at the receiver. It was also noted that a slight further improvement is seen as N is increased. An expression to calculate the number of bits that can be saved is (6. The number of iterations required for good performance was shown to be around 3.16).1 Quantisation and Clipping An early paper to investigate the relation between clipping and quantization in DMT transceivers is reference [94] where an analytical expression was developed to find the minimum number of bits required in the A/D. Using the FFT/IFFT pair to make decisions in the frequency domain regrows samples that were clipped at the transmitter. D/A operation without changing the SNR. For example in a 64QAM OFDM system with a CL of 5dB in an AWGN channel the improvement was quite dramatic being only ~1dB lower than the theoretical lower bound. Like the method of [92] it uses a FFT/IFFT pair in an iterative algorithm at the receiver to try to estimate which samples were clipped at the transmitter. However much of the gain from DAR may be achieved by simply correcting the constellation shrinkage which clipping causes (refer to Figure 3. 6. Through simulation reference [92] found that DAR worked best when a small number of samples were clipped. When the clipping noise is large compared to the AWGN in the channel. performance is limited by the clipping noise. Due to the Rayleigh nature of the envelope of the DMT envelope 2 to 3 bits can be saved in the A/D. D/A converters when the signal is clipped to a predefined level. 93] is the latency and complexity of performing extra FFT/IFFT operations.
Aclip. the QAM constellation size L2. ( 3N ) . R2 is the number of bits required for the A/DD/A with clipping to keep the same SNR as when no clipping is performed. Otherwise the symbol is sent on to further processing unmolested. the probability of the new symbol requiring clipping at Aclip will be reduced. Clipping and quantization are further explored in reference [95] for DMT based transceivers where a improved clipping technique allows for up to an 8dB improvement in the SNR over standard clipping. The overall probability of clipping for the 2 pass method described above is (6. and µ= Amax σ is the clip level. 2 L +1 is a parameter set by the number of subcarriers.log 2 2 ( 8 π ) .17) 122 .Pclip / 2 = Pclip (6. By careful selection of the phasor rotation. if a sample is above the clipping threshold. then the phase of each QAM modulated carrier is changed by means of a fixed phasor rotation.16) 2 PClip / Total = Pclip /1.Chapter 6: Peak to Average Power Solutions – Distorted Techniques 2 ν − 3 1 ∆ = . The new method analyzed the samples after the IFFT. however the authors claim that different mapping types on subchannels would have a minimum effect.ν = 1+ 2 L −1 .22 R1. It was assumed in the paper that all subcarriers have the same constellation type.e µ2 −µ 2 2 (6.µ −3 . and a new DMT symbol is generated by the IFFT.16) which is determined to be PClip / Total µ = 1 − erf 2 N 2 2 (6.. R1 is the number of bits required for the A/DD/A when no clipping is performed. (L=4 equates to 16 QAM). Amax is the clip level in volts and σ is the rms voltage of the DMT transmit symbol.15) where ∆ = R1 − R2 is the number of bits saved. N.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques where µ = AClip / σ .5 3 3.2.2: Probability of clipping DMT signal as a function of µ for p=1. 123 . N=64 subcarriers.3. p. used (if any) of log2p bits per transmitted symbol. Recalculating the IFFT can create a bottleneck slowing down system performance.5 A clip /sigma 4 4. As seen in Figure 6. the clipping probability drops with an increase in the number of passes.5 5 Figure 6. 10 0 10 2 p=1 10 4 p=2 10 P clip /p 10 8 6 p=3 10 10 10 12 10 14 2.2. For 16 QAM. but does not necessarily require a factor 2 increase of the IFFT as not every symbol will require clipping. At µ =4 the 2 pass and 3 pass methods reduce the probability of clipping down to ~104 and ~106 respectively. b. Side information is also required to inform the receiver of the number of passes. especially at higher levels of Aclip. Quantization effects are also examined in terms of noise from the DAC and ADC in the transmitter and receiver respectively for various wordlengths.
However as noted in reference [94] the wordlength can be reduced with minimal affect on the BER. An 8 bit wordlength is also recommended with an extra 2 bits for the receiver ADC to compensate for peak regrowth affects after transmit filtering (upsampled by 4 before filtering) and imperfect AGC in the receiver. In a nutshell. Also clipping at the IFFT output increases the resolution giving a better average signal/quantization noise power ratio.7 with p=1 and µ ≥ 3. For 16bit arithmetic the headroom was shown to be around 15dB. lowering the clipping level increases the clipping noise while at the same time reducing the quantization noise. Results from simulation revealed that performance for fixed point FFT’s is improved when overflow is allowed to occur with low probability.4 if no attempt is made to control the peak excursions of an OFDM symbol the HPA will saturate causing spectral regrowth and an increase in the 124 . µ ≥ 3.Chapter 6: Peak to Average Power Solutions – Distorted Techniques N=256. Clipping is performed on the I and Q outputs of the IFFT. Reference [97] examined the effect of rounding and saturation in fixedpoint DSP implementation of the IFFT and FFT where optimum tradeoffs are found between saturation and rounding.4 with p=2 an improvement of 3 and 8dB can be achieved. Another reference [96] analyzed quantization effects on OFDM used a simulation model with Hiperlan2 [53] specifications. of course this is at the expense of clipping noise. which increases the BER. As the wordlength at the IFFT output is decreased. the power consumption and complexity of the DAC/ADC decreases at the expense of quantization noise. Doubling the size of the FFT resulted in an improvement in the new scaling method of 3dB. The distribution of error was shown to depend on the ratio of the maximum quantization level to the RMS power of the random variable.2 Amplifier non linearities As shown in Section 3. b=12. 6. Results from reference [96] indicated that the optimum clipping level for wordlengths between 6 and 9 bits occurs at around 4σ (slightly lower for smaller wordlengths).
Again the number of subcarriers was shown to have a negligible effect on the BER. The number of subcarriers has a negligible affect on OFDM due to the Gaussian distribution of samples. It was shown that the SER versus SNR is almost the same for both systems with OFDM having slightly better performance due to the frequency guard interval used in OFDM which reduces the equivalent noise bandwidth. The simulated results were shown to be in good agreement with theoretical results.Chapter 6: Peak to Average Power Solutions – Distorted Techniques BER at the receiver. The simulation model used N=2K and 8K (DVB. and that only 3rd order distortion will affect the system. DVA) subcarriers with QPSK and 16 QAM mapping to produce SER versus clipping and AWGN plots in a Monte Carlo simulation. Reference [98] focused on simulation comparing Single Carrier (SC) and OFDM systems with clipping in the baseband and RF amplifier non linearities. An early reference [42] to look at the effect of a non linearity on QPSK OFDM compared analytically derived and simulated results for the BER versus SNR. This section reviews papers which analyze the effect of an uncontrolled OFDM symbol on the HPA.27Vo . by V1dB = 0. The non linearity is set at the 1dB compression point of the in band output and is related to the output signal. Vo. with 1dB backoff show that even with a 1dB backoff the BER is within 1 order of magnitude of the linear amplifier at SNR=10dB. Specifically they assumed that the intermodulation products were generated by non linearities in the receiver IF module. The performance of the SC and OFDM system converges at a baseband clipping of BO=6dB for OFDM when 16 125 . The BER is calculated under the assumption that the intermodulation products cause an additive Gaussian interference and that the BER is approximately equal on all subcarriers. These results are stated to be applicable to the transmitter amplifier. The simulated results for the amplifier which is linear up to 3dB. The baseband clipping effect was simulated with the SNR set at 16dB for 16 QAM 2 and the backoff was defined as BO = A 2σ 2 .
At 12dB IBO the non equalized system lost so much SNR that its error floor became 103 . the HPA output can then be expressed as (6. a backoff of 3dB was only required for QPSK due to the larger Euclidean distance of QPSK constellation points. however the non equalized OFDM model lost a further 7dB of SNR.4. The phase distortion was considered to be linear.1) is stated to cause two effects on the detected samples: • • constellation warping (amplitude and phase distortion) non linear distortion which generates a Gaussian spread like cluster of received values around each constellation point (refer to Figure 3.Chapter 6: Peak to Average Power Solutions – Distorted Techniques QAM mapping was used.18) y ( t ) = f A ( t ) e j θ ( t ) + g A( t ) ( ) (6. Amplifier non linearity (using TWT as described in Section 3. where the signal is of the form A (t ) e jθ ( t ) .18) Where f [ A] is the AM/AM characteristic of the amplifier and g [ A] is the AM/PM characteristic of the HPA. Performance for 16 QAM OFDM under different amplifier backoffs is almost identical to the baseband clipping effects with a 6dB BO required to bring performance in line with its equivalent SC counterpart.16) For 16 QAM OFDM system it was shown that at high IBO (25dB) equalization had no effect on the BER. Reference [40] both analytically derived and simulated an OFDM system in a AWGN channel with and without equalization (1 tap) to produce BER versus SNR plots. As the IBO is reduced (14dB) the equalized BER degraded only a little. 126 . For all cases a good agreement is shown between analytical and simulated results. QPSK was again shown to be more robust to amplifier non linearity. For the RF amplifier the low pass equivalent was used. It was also shown that the degradation increases as the third order interception point approaches the 1dB compression point.
a result supported by reference [88]. Phase noise is caused by the oscillators in the RF stage and becomes a more dominant source of noise at higher carrier frequencies (up to 40GHz). The theoretical expressions were supported by simulations using SSPA and TWTA with different values of phase noise. 127 In . It was found through analytical means that the variance is approximately equal on all subcarriers with the middle subcarriers experiencing the most noise. However a good agreement is seen between the analytical and simulated curves. Under the assumption of modeling the phase and amplifier distortions as additive Gaussian noise. Results indicated that while QPSK is rather impervious to both amplifier non linearity and phase noise. The simulation results for 16 QAM were found to better fit the analytical results than QPSK. this means that the SER is similar on all subcarriers. The data was demodulated with an FFT and the The difference was transmit symbols were compared to the received samples. the performance of 16 and 64 QAM is greatly diminished in terms of the BER even with a large OBO in the amplifier. Simulated results are compared to analytical results for QPSK and 16 QAM OFDM. The simulation model used an equivalent low pass representation to avoid RF up and down conversion. the tails in both cases fall of quicker in the simulated case because they are not exactly Gaussian. Initially the spectrum of the distorted OFDM signal is examined.Chapter 6: Peak to Average Power Solutions – Distorted Techniques Reference [47] provided an analytical analysis of smooth non linear distortion on the SER as a function of third order distortion (the most dominant distortion for smooth non linearities) in a memoryless non linear power amplifier. Reference [41] extended on earlier work presented in reference [40] to theoretically analyze the effect of non linear amplifiers in conjunction with phase noise on MQAM OFDM. A SSPA with p=2 introduced 4th order distortions which further diminished the performance of 16 and 64 QAM. the computed variances are used to get an estimate of the BER in an AWGN channel. N=1024 subcarriers were processed by an IFFT at which point the non linearity was performed. which is then used to determine the detection error for the matched filter detection. squared and stored and then averaged to find the noise variance. The analysis was performed on a matched filter pair.
Chapter 6: Peak to Average Power Solutions – Distorted Techniques general it was shown that the joint effects of amplifier non linearity and phase noise have two major effects. Amplifier non linearity generates a uniform amplitude
attenuation and phase rotation which can be corrected at the receiver by AGC. Phase noise introduced a constant phase rotation within 1 OFDM symbol which can be estimated and corrected using the pilot tones. The second effect is constellation clustering due to the interference produced by the HPA to the ICI caused by the phase noise.
Finally it was shown that the phase noise impairment was dependant on the relationship between the phase noise rate and the OFDM symbol period. It was shown that the phase noise can become a limiting factor if a large number of subcarriers and a high frequency carrier are used.
6.3 Windowing
Windowing or pulse shaping are similar to clipping in that they attenuate large peaks. However in windowing a corrective window is multiplied with the data so that not only the peak cancelled but surrounding samples are also affected. The advantage of this process is to keep the OBR lower than in standard clipping. Windows should be as narrowband as possible in the frequency spectrum domain, so as to have good OBR properties. However narrowband windows have the reciprocal affect of being long in the time domain which implies many signal samples being affected, which increases the BER.
Reference [99] uses window types such as such as cosine, Kaiser, and Hamming, comparing their use to standard clipping in terms of the frequency spectrum and BER. The simulation model uses Hiperlan2 [53] specifications with a ½ rate convolutional code and 16 QAM where it is shown that clipping the signal at 5dB has a minor affect on the BER with a 0.2dB loss in SNR. Windowing is shown to have almost identical affect on the BER above 5dB clipping but interestingly has worse performance at harder clipping levels.
128
Chapter 6: Peak to Average Power Solutions – Distorted Techniques To simulate the affect of the required backoff in the HPA a SSPA with p=3 as described in Section 3.4.1 is used. In order to keep the OBR to below 30dB for 64 subcarriers a backoff of 6.3dB was required, which could be reduced by 1 dB to 5.3dB when peak windowing is used. When 256 subcarriers are used the backoff of 6.3dB can be reduced by 0.8dB to 5.5dB with peak windowing, showing that windowing is independent of N.
A later reference [100] uses broadband pulse shaping on individual subcarriers as a way to reduce the PAPR. By making the cross correlation between samples in the same block close an OFDM signal with a low PAPR can be created. The new OFDM signal is given as
x ( t ) = ∑ X n ( m ) pm ( t ) e j 2π m T
nT ≤ t ≤ ( n + 1) T
(6.19)
Where X n ( m ) is the modulated data symbol of subcarrier m, T is the duration of the OFDM block, and the waveform pm ( t ) is a pulse shape of duration T, on subcarrier m which has a bandwidth less than or equal to the bandwidth of the OFDM signal x(t).
Unique RRC waveforms are multiplied with each sample, which are cyclic shifts of each other within the same time interval 0<=t<T. As each RRC pulse is seeded from the same source they are easy to create, the RRC pulse for each subcarrier is defined as
N + L −1 k =− L k −m t N
pm ( t ) =
∑
C (k )e
− j 2π
mk N
e
− j 2π
0≤t <T
(6.20)
where
C (k ) =
k − j 2π t 1 T 1 m p ( t ) e T dt = P ∫0 T T T
(6.21)
129
Chapter 6: Peak to Average Power Solutions – Distorted Techniques is the Fourier series of p(t) and
prc ( t − T 2 ) , p (t ) = 0,
0≤t <T elsewhere
(6.22)
where prc ( t − T / 2 ) are the samples of the time domain RRC pulse. When the rolloff factor is increased to 0.5 the reduction in the CCDF at Pr(ζ>ζ0)=103 is around 6.5dB. However the effect of oversampling and the frequency spectrum is not treated. I anticipate that the spectrum is at least 50% larger than normal OFDM because the spectrum of the broadband pulse would have to be convolved with the basic OFDM linear spectrum.
Note that reference [100] is actually mathematically equivalent to clipping and filtering with a linear time invariant filter.
6.4 Conclusion
This chapter presented distorted techniques for the reduction of PAPR in OFDM. Clipping in the baseband was first introduced as this is the simplest and most widely examined area in distorted PAPR reduction techniques. Clipping was shown both analytically and through simulated means to have a minor effect on the BER when QPSK modulation was used due to the large Euclidean distance between constellation points. Higher order mapping types such as 16 and 64 QAM were much more susceptible to clipping. The BER on individual subcarriers was also treated where it was shown that the probability of error was almost equal on all subcarriers. Increasing the number of subcarriers was shown to have a beneficial affect as the noise introduced by clipping would be spread over more subcarriers. The Gaussian like assumption of the noise which is assumed in most analysis of clipping noise was also shown to be unsubstantiated resulting in optimistic error probabilities. Clipping noise was shown to have an impulsive nature at higher clip levels resulting in a much smoother decay in error probability as the clip level was increased.
130
Chapter 6: Peak to Average Power Solutions – Distorted Techniques Clipping at the nyquist rate (no oversampling) does not cause out of band radiation as all the noise was shown to fall in band. Oversampling before clipping was shown [46] to produce less in band noise but to increase the out of band noise, requiring filtering. Quantization in hardware and clipping was also treated where it was shown that some bits could be saved after clipping, due to the Rayleigh distribution of samples without degrading system performance. This also improves the resolution of the clipped samples. Furthermore it was shown that several steps could be taken to mitigate the errors caused by clipping. As clipping noise caused shrinking in the constellation size as well as a Gaussian like spreading the AGC in the receiver could be used to correct this depending on the clip level. Also, the noise due to clipping will also experience fading along with the signal further lessening its effect.
Early windowing and pulse shaping techniques displayed little improvement and sometimes a further degradation in the BER while attempting to reduce the OBR. Later work in this area produced a markedly greater improvement in the PAPR by the selection of appropriate pulse shapes which were applied to individual subcarriers rather than the whole transmit signal.
131
Chapter 7: Reduced Complexity Clipping Algorithms Chapter 7 Reduced Algorithms Complexity Clipping The previous chapter covered distorted techniques for PAPR reduction. This chapter presents new low complexity clipping techniques which avoid complex hardware operations while maintaining similar performance to conventional clipping.2 details a new technique coined Sector Clipping and provides theoretical and simulated analysis of the new method. Section 1 In this work the BER will be plotted against clipping level with the noise set to zero. Section 7.1 describes conventional clipping and a simulation model is developed to quantify the effects that various transceiver components have on the BER1. although clipping is less complex in terms of hardware operations than distortionless techniques. The BER’s in these plots therefore represent the ‘error floor’ of the more commonly used BER vs. estimates of the magnitude still need to be made in order to decide whether a sample needs to be clipped or not and multiplications have to be made to correct the signal.3 presents another new technique which is similar to the CORDIC algorithm but with reduced complexity called Vector Subtraction. 132 . Section 7. Also. SNR plots. All the errors are therefore caused by clipping noise. The new clipping algorithms are then implemented in a new clip and filter algorithm which is much less susceptible to peak regrowth after baseband filtering. As described clipping suffers after filtering as clipped peaks can regrow resulting in saturation of the HPA. Section 7. This model is then used to test the new clipping algorithms.
the BER. For comparison Figure 7.35. The OFDM transceiver system is shown in Figure 7.1 and 7.4. 133 . and 64 Mary QAM symbols are modulated with a 64 point IFFT and then pulse shape filtered with a Root Raised Cosine Filter (RRCF) with α=0. Note that the results in Sections 7.3. and 64 QAM.4 compares the new and existing clipping methods in terms of their baseband Clipping Level (CL) vs. 4. BER for 4. 16.1 shows the BER with Additive White Gaussian Noise (AWGN) in the channel and the CL set at 3.1 show the Bit Error Rate Floor (BERF).Chapter 7: Reduced Complexity Clipping Algorithms 7.e. and 128 taps. 10 0 64 QAM 1 No clipping CL=5dB CL=3dB 10 16 QAM 2 10 BER 10 3 QPSK 10 4 0 5 10 15 AWGN (dB) 20 25 30 Figure 7.1: Average noise in the channel vs. The error floors for 64 QAM can be cross referenced with figure 7. Section 7. 16. 5dB and no clipping. Finally section 7.4.6 concludes the chapter with a review of the advantages of the proposed clipping techniques.5 implements the new algorithms in a new clip and filter algorithm which is less susceptible to peak regrowth. It can be observed that 16 and 64 QAM mapping is much more susceptible to clipping noise than 4 QAM. i. the BER due to clipping as that is the focus of this chapter.
A block diagram of the simulation model used is shown in Figure 7. Division and square roots) and therefore take a number of clock cycles.1) In order to analyze the performance of conventional clipping it is useful to quantify the effects of the different system components involved. the mapping type. This method is the most hardware intensive of all methods described in the following sections. i. Some of the operations require iterative techniques when implemented in fixed point processes (e. Clipping in this way only introduces amplitude distortion. Other methods require vast LUT’s which consume chip area or memory space. and IBO of the HPA.2: IQ diagram showing conventional clipping region.e. Mathematically conventional clipping can be described as x '' = r rclip e ∡( r ) r ≤ rclip r > rclip (7. the phase is unaffected. 134 .Chapter 7: Reduced Complexity Clipping Algorithms 7. The vector r is reduced to rclip.2. q i x’’ rclip r Figure 7.g. filter rolloff factor.1 Conventional clipping Conventional clipping is defined here as any hardware method which reduces the amplitude of the signal to a predefined level in line with the origin as shown in Figure 7. number of filter taps. IFFT size.3.
A SSPA (as described in Section 3. A comparison between the gray encoded transmitted bits and the received bits is made to determine the BER. Finally. a simple Least Square (LS) algorithm is used to make decisions on the decoded data. After clipping to a predefined level relative to the mean power of the transmit symbol a cyclic prefix can be added to the data. The cyclic prefix (if used) is removed prior to demodulation with the FFT. The data is then interpolated by a factor of 8 before being filtered by a matched Root Raised Cosine Filter (RRCF). after modulation the data is converted back into serial form and sent to the baseband clipping block. Again note that the data is buffered at the input to the FFT until all N samples are ready. Mary mapping IFFT Clip block CP Interp by 8. AWGN and multipath delayed versions of the signal can then be added in the channel.3 Block diagram of simulation model used for clipping models. At the receiver side the process is reversed.1) models the HPA.4.Chapter 7: Reduced Complexity Clipping Algorithms Data gen. The baseband Monte Carlo simulation model of Figure 7.2.3 randomly generates Mary mapped data and then modulates the signal with the IFFT. In the results that follow the conventional baseband clipping algorithm is used as described in Figure 7. Note that 64 QAM mapping is used unless otherwise stated 135 . Note the data is buffered at the input to the IFFT so that N samples are fed into the IFFT. RRC Filter SSPA BER calc AWGN Data decision Demap FFT Remove CP Deci. by 8 Matched RRC Filter Figure 7. a matched RRCF filters the received samples which are then decimated by 8 to retrieve the transmitted samples.
35 alpha=0. the first is that when the easier roll off or excess bandwidth of α=0. AWGN=0. no channel impairments.15 and 0.000 OFDM symbols are transmitted. 64 QAM symbols. Two points can be made reviewing Figure 7.35.4: Baseband clip level vs.15) the BER is substantially affected by the number of taps. 10 0 10 1 10 2 10 BER 3 10 4 10 5 10 6 alpha=0. 128 and 256. The second point is that when a tighter roll off factor is used (α=0. the BERF for varying RRCF parameters. Effect of filter on the BER Figure 7.4). There are no other sources of distortion or noise. and the roll off factor or excess bandwidth is set at 0. In the following simulations 10. 64 filter taps has a BER almost 2 magnitudes worse than 128 taps at CL=7dB.4 shows the baseband clip level vs.35 is used the number of taps has little effect on the BER. For 64 QAM this means around 3 million bits are transmitted. This is due to the inband amplitude 136 .15 64 filter taps 128 filter taps 256 filter taps 10 7 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. the BER for a RRCF with different numbers of coefficient taps and roll off factors (named alpha in figure 7.Chapter 7: Reduced Complexity Clipping Algorithms since the more traditional QPSK has such a high tolerance to clipping excessive simulation times are required to get BER plots. The number of taps is set at 64.4. LPA. 64 point IFFT. each symbol has 52 information bearing subcarriers.
The extreme subcarriers will then be effected by additional attenuation which will reduce the noise margin in the receiver decision (slicer).5a and b).15 for the same number of filter taps. 64 filter taps in RRCF.Chapter 7: Reduced Complexity Clipping Algorithms distortion (ripple) created when insufficient taps are used with a small excess bandwidth (refer to Figures 7.15) so that the filter skirts occur in the null subcarriers.5a: Demapped constellation.15 will be used unless stated otherwise. The 0. This time the IFFT size is set to 64 and 128 while maintaining the same 137 . α=0. The Hiperlan2/802. performance than α=0.15 roll off curves have a steeper transition band and avoid the problem of the extreme subcarriers. However passband ripple will be introduced when the number of taps is low (taps=64). This creates a linear spreading of the In all cases α=0. BERF with 64 QAM mapping. 128 filter taps in RRCF. Effect of HPA amplifier backoff and IFFT on the BERF Figure 7.5b: Demapped constellation. Figure 7.15.35 has worse demodulated samples away from the origin.35 roll off factor curves will have a wide transition bandwidth in the frequency domain.11a physical layer specification requires a tight roll off. Figure 7.7 again shows the clip level vs.15. α=0. M=64. with no clipping or channel impairments. In all further simulations 128 filter taps with a roll off of α=0. M=64. with no clipping or channel impairments. The 0. (α≈0.6 and Figure 7.
With no extra IBO in the HPA (IBO=CL) the BERF is between 102 and 103 at CL=6dB while with a LPA the BERF is below 104 at 6dB baseband clipping.e. degrading the BERF.5 order of magnitude improvement. This allows us to see the effect of peak regrowth on the BERF with and without oversampling. Some of the clipping noise falls into the null bins and is subsequently filtered away while os=1 systems would cause this noise to fold back into the inband subcarriers. In order to further highlight the effect of oversampling. In this case the baseband clip level is set at 5dB. showing the peak regrowth under critically and oversampled conditions. In Figure 7.6 (os=1) even with an IBO of 4dB above the baseband clipping level (IBO=CL+4dB) peak regrowth still causes a small amount of saturation in the HPA. The other variation is the inclusion of a HPA with IBO.6) while the 128 point IFFT has a BERF just above 103 at the same clip level. an improvement of half a magnitude.7 that for the curve IBO=CL+4 to the LPA curve (IBO=CL+∞) there are no errors occurring in over 3 million transmitted bits at 7dB clipping level. The HPA follows the Rapp model of (3.Chapter 7: Reduced Complexity Clipping Algorithms number of information bearing subcarriers.6 and 7. In Figure 7. Results not shown here indicate that the harder the clip level the more extreme the peak regrowth after filtering.7. the data is therefore critically sampled (almost) and oversampled by a factor of 2 respectively. a 1. 138 . Peak regrowth is 3dB (2dB less) for the oversampled case (os=2) at Pr(ζ>ζ0)=104. For no extra IBO in the HPA the 64 point IFFT has a BERF between 102 and 103 at 6dB baseband clipping (Figure 7.7 (os=2) the performance is better for all IBO across the board. Note in Figure 7. i. peak regrowth for the critically sampled case is extreme with almost 5dB peak regrowth at Pr(ζ>ζ0)=104.34) with p=3. the CCDF in Figure 7.8 is also shown for the case described in Figures 7.
64 QAM symbols. RRCF with 128 taps and α=0. 128 point IFFT (os=2).15.15.Chapter 7: Reduced Complexity Clipping Algorithms 10 0 10 1 IBO=CL IBO=CL+2 IBO=CL+4 LPA 10 2 BER 10 3 10 4 10 5 10 6 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. p=3 in HPA. 64 point IFFT (os=1). 139 . p=3 in HPA. the BERF with varying IBO in HPA.7: Baseband clip level vs. 10 0 10 1 IBO=CL IBO=CL+2 IBO=CL+4 LPA 10 2 BER 10 3 10 4 10 5 10 6 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. AWGN=0. the BERF with varying IBO in HPA. 64 QAM symbols. AWGN=0. RRCF with 128 taps and α=0.6: Baseband clip level vs.
9 shows the effect of changing the mapping constellation M with an oversampling factor of 2 in conjunction with a LPA. 16 QAM also has rather robust performance in the presence of clipping with no errors being detected above 4dB clipping at 105 probability. it is extremely impervious to clipping with no detected errors at 0dB and above. Figure 7. 140 . Effect of constellation size on the BERF.8: Simulated CCDF clipped in baseband at 5dB.Chapter 7: Reduced Complexity Clipping Algorithms 10 0 64 point IFFT 128 point IFFT 10 1 After Filtering Pr(ζ> ζ0 ) After HPA 10 2 After Baseband Clipping 10 3 10 4 3 4 5 6 7 8 9 10 ζ0 (dB) Figure 7. Although 4 QAM was simulated. 64 QAM clipping has performance around 2 magnitudes worse at equivalent clipping levels to 16 QAM. IBO in HPA set to 8dB for 64 and 128 IFFT.
10 demonstrates the effect of changing p on the BERF.1.9: Baseband clip level vs. AWGN=0.Chapter 7: Reduced Complexity Clipping Algorithms 10 0 16 QAM 64 QAM 10 1 10 BER 2 10 3 10 4 10 5 0 1 2 3 Clip Level (dB) 4 5 6 7 Figure 7. 141 . the BERF with varying Mary constellations.4. 128 point IFFT (os=2). As p=3 is a practical value used in many designs [40] it will be used in all further simulations. Varying the value of p in the SSPA controls the input to output curve of the amplifier as shown in Figure 3. Here it is seen that p=1 has an extreme effect on the BERF. RRCF with 128 taps and α=0.14. Effect of changing P in HPA The model used to simulate the HPA is the SSPA described in Section 3.15. the saturation level of the SSPA is set equal to the baseband clipping level. P=3 causes a magnitude of degradation over the absolutely linear region of p=1000. Figure 7.
15. and mapping type were simulated to see their effect on the baseband CL vs. The effects of oversampling the IFFT. 142 . AWGN=0.Chapter 7: Reduced Complexity Clipping Algorithms 10 0 10 1 p=1 p=3 p=5 p=1000 10 2 BER 10 3 10 4 10 5 10 6 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. This section provided a description and simulations of an OFDM system with baseband clipping. It can be concluded that a good set of parameters for further analysis of an OFDM system are: • • • Oversampling factor of 2 in the IFFT. HPA backoff set equal to baseband clipping level. the BERF. 128 point IFFT (os=2). RRCF with 128 taps and α=0. A SSPA with p=3. RRCF with 128 filter taps and a roll off factor of 0. filter parameters. HPA parameters.10: Baseband clip level vs the BERF with varying p in the SSPA.15.
Square clipping can be construed as a DAC with limited word length. The number of sectors can vary from 2 (square clipping) to 5 or more.11 shows the I Q plane of Sector clipping with the clipping regions clearly identified. As there are 3 unique sectors in Figure 7. Sector clipping with 3 sectors.11. 143 . As seen in Figure 7.Chapter 7: Reduced Complexity Clipping Algorithms 7.2 New Sector Clipping method A new technique for clipping developed by the author which reduces hardware complexity is Sector clipping [3] Sector clipping avoids magnitude estimates which require hardware multiplications to perform corrective scaling.11 this structure is known as 3 Sector clipping. The symmetry of Sector clipping can be exploited to further reduce hardware complexity.2. and so divides the clipping region into different ‘sectors’. the clipping operation is then performed on the new value. Note that for the theoretical analysis. Note that increasing the number of decisions beyond 5 Sector will clipping increase the complexity as much as multiplications. and Square clipping.1 Theoretical Analysis of Clipping Techniques In order to compare the performance of the various clipping techniques the Clip Level vs. no oversampling. Sign bits and the relative size of the real and imaginary components can then be used to extrapolate the original position of the clipped sample. The new method requires only comparators and can be implemented in hardware as either an iterative (to reduce complexity) or parallel structure (to increase speed). SNR is mathematically derived for 3 cases: Conventional clipping. Figure 7. Sector clipping not only introduces extra amplitude distortion over conventional clipping. the data can be ‘folded’ into the first octant by removing the sign bits (making it positive) and making the largest value of the complex signal the real component. but also phase distortion as data outside the clipping regions is not reduced in line with the origin. although no discernable improvement in performance is seen above this number. The decision to clip is based on the I and Q values in conjunction with comparisons between them. filtering. 7. or amplifier is assumed.
1. this relation between the input.2. Mathematically.Chapter 7: Reduced Complexity Clipping Algorithms q θ i rclip Figure 7.11: I Q diagram showing different sector clipping regions and the direction of data reduction for 3 Sector Clipping.11 and can be used to find an expression for the SNR. and the output.1 SNR Analysis The relation between the input and output of the clipping operation is expressed pictorially in Figure 7. x.12: Input output relationship of clipping operation. y can be expressed as Bussgang’s theorem 144 . 7. x Amp y n Noise Figure 7.
Chapter 7: Reduced Complexity Clipping Algorithms y = αx + n (7.4) and (7. is uncorrelated.5) and then the correlation functions yields E [nn] = E[( y − αx )( y − αx )] R nn = R yy − 2αR xy + α 2 R xx Equating (7.2) gives E [ yy ] = E[(αx + n )(αx + n )] R yy = α 2 R xx + 2αE[ xn] + Rnn (7.3) can be rearranged and solved for the noise power. R nn = R yy − α 2 R xx (7. n.4) Rearranging (7.7) α= R xy R xx 145 . Ryy is the autocorrelation of the output signal giving the output power. As x is assumed to be uncorrelated to the noise the term 2α E[ xn] can be removed and (7.6) and solving for α: (7.2) n = y − αx (7. y is the output signal and α is chosen so that the input signal and the noise.6) R yy − α 2 R xx = R yy − 2αR xy + α 2 R xx 0 = 2α 2 R xx − 2αR xy (7. Finding the 2nd moment of y (7.5) Taking the expected moments of (7.3) where Rxx is the autocorrelation of the input signal giving the input power. Rnn.2) where x in the input signal.
Chapter 7: Reduced Complexity Clipping Algorithms The signal. 7.8) and the noise is given in (7. f (r )∂r ∞ rclip (7. autocorrelation of the output and cross correlation for conventional clipping are respectively given below. R xx ( 0) = ∫ r 2 .10). S is equal to: S = α 2 R xx (7.9) To reduce mathematical complexity substitutions are made to factor out Rxx.11) R yy ( 0) = ∫ rclip 0 2 r 2 .10) The expressions for the correlation functions need to be defined for each of the clipping techniques. f (r )∂r 0 ∞ (7.2. The new SNR is given by (7.12) R xy ( 0) = ∫ rclip 0 r 2 .1. f (r )∂r ∞ rclip (7.2 Conventional clipping The definitions for the autocorrelation of the input. α 2 R xx S = N R yy − α 2 R xx (7.rclip . α2 S = N R xy −α 2 R xx (7. f (r )∂r + ∫ r.9).13) 146 . which are solved to find a closed form solution. f (r )∂r + ∫ rclip .4). the SNR is then (7.
rclip is the clipping level. σ 2 .e. and f(r) is the probability distribution of the data.16) R yy ( 0) = 4σ 2 ∫ rclip 0 2 R 3 .15a.16).15c) Substituting (7.15b) ∂R = σ 2 (7. f (R )∂R + 4σ 2 ∫ ∞ Rclip R 2 . r. R xx ( 0) = 4σ 2 ∫ R 3 .18). which is assumed to be Rayleigh distributed.17) R xy ( 0) = 4σ 2 ∫ Rclip 0 R 3 .11). f (R )∂R ∞ 0 (7. and (7. In order to reduce the mathematical complexity of the correlation functions they are normalized by the average input voltage. Rxy. f (R )∂R ∞ rclip (7. R= r σ 2 rclip (7.R.18) 147 . the magnitude of the data at the output of the IFFT). and (7.Chapter 7: Reduced Complexity Clipping Algorithms where r is the input data (i.13) yields the normalized expressions for Rxx. c) into (7. and Ryy given in (7. b. (7. f (R )∂R (7. f (R )∂R + 4σ 2 ∫ Rclip . (7.14) Note that for the Rayleigh assumption of the distribution of ‘r’ to hold the number of subcarriers is assumed to be greater or equal to 64 [39].17). −r 2 f (r ) = r σ2 e 2σ 2 (7.12).Rclip .15a) Rclip = σ 2 ∂r (7.
16).22) α= Rxy ( 0) Rxx( 0 ) = −e − Rclip 2 +1+ π 2 Rclip − π 2 Rclip .24) SNRconv = ( −e − Rclip 2 − Rclip 2 π π +1+ Rclip − Rclip . and (7. These are substituted into (7. 148 .19) The normalised clip level is also required (7.23).18) are shown in (7.20) The evaluated correlation expressions of (7.24) 2 Rxy ( 0 ) π π −R 2 +1 − α = = −e clip + 1 + Rclip − Rclip .22).23) The SNR vs.10) to calculate the SNR at various clipping levels.21).Chapter 7: Reduced Complexity Clipping Algorithms where f(R) is the normalized Rayleigh probability distribution of the data. clipping level for conventional clipping reduces to (7.erf ( Rclip ) (7. f (R ) = e − R 2 (7.16. R.21) +1 (7.erf ( Rclip ) −e 2 2 (7.20) r dBclip = 20 log10 σ 2 dBclip = 20 log10 (R ) dBclip R = 10 20 (7. conventional clip level is shown in Figure 7.erf ( Rclip ) Rxx( 0) 2 2 2 ) A plot of the theoretical SNR vs. and (7. Rxx( 0 ) = 2σ 2 Ryy ( 0 ) Rxx( 0 ) = −e − Rclip 2 (7. (7.17). (7.
Rxy. y ) = 1 2πσ 2 − x2 − y2 (7.3 Sector Clipping In sector clipping the reduction in hardware complexity comes at a cost of introducing extra amplitude distortion and phase distortion as the components of the complex signal are not attenuated by the same scaling factor.Chapter 7: Reduced Complexity Clipping Algorithms 7.25).26a) (7. c) of the conventional clipping method.13 showing the vectors of an unclipped signal and the resultant clipped sample. 1 1 2 2 f ( x.26b) (7. A way to realize the transition from polar to Cartesian is to recognize that the joint probability of two independent.2.1.26c) l2 = rclip cos θ 149 . zero mean. y ) = e 2σ × e 2σ σ 2π σ 2π f ( x. This leads to more complex equations for Rxx. Gaussiandistributed variables. and Ryy as the data must be represented in terms of its Cartesian coordinates. x and y.26a. The clipping regions for 3 sector clipping are shown in Figure 7. quadrature shifted.25) − x2 − y2 e 2σ 2 3 sector clipping levels are defined relative to rclip (7. l0 = rclip sin θ l1 = rclip 2 (7. b. with the same σ create a Rayleigh distribution (7.
The limits of the integrals of each part (7.27.27) 0 ∫ (x ∞ 2 0 + y 2 ) . Due to symmetry.28) where the symmetry of the clipping regions is exploited. f ( x. 2.13: I Q diagram of 1st quadrant of a 3 Sector clipping system showing the vector of an unclipped and clipped sample. only the first quadrant is used ∞ Rxx (0) = ∫ (7. y ) ∂x∂y The autocorrelation of the output equations are set out in (7.13)) perform the attenuation as evidenced by the limits.28) (1. 150 .28) do nothing to the data as they are inside the clipping region.28) dictate what will happen at the output. and 3 in Figure 7.28. The last 3 parts of (7. The new sector clipping correlation equations for Rxx. 7. 7.Chapter 7: Reduced Complexity Clipping Algorithms Q 1 l2 l1 Clipped vector l0 1 I l0 l1 l2 2 2 3 Original vector Figure 7. Ryy. Hence the first 2 parts of (7. and Rxy are listed (7.29) respectively below.
y )∂x∂y l1 0 l2 l0 +8∫ ∫ ( xl2 + y 2 ) .30a) Y= y σ 2 ∂y (7. f ( x.29) show the translation of the input sample to the output.29) +8∫ ∫ ( xl1 + y 2 ) . y )∂x∂y l1 0 l2 l0 +8∫ ∫ l2 + y 2 . f ( x.30c) ∂Y = σ 2 (7.28) where the first 2 parts are inside the clipping regions.Chapter 7: Reduced Complexity Clipping Algorithms Ryy (0) = 4 ∫ ∫ ( x 2 + y 2 ) . Rxy (0) = 4 ∫ ∫ ( x 2 + y 2 ) . f ( x. The last 3 parts of (7.29) follows the same form as (7. y )∂x∂y 0 0 l1 l1 +8∫ ∫ ( x 2 + y 2 ) . f ( x. y )∂x∂y l1 l1 ∞∞ As in the normalized conventional clipping method. f ( x. y )∂x∂y 2 l2 0 ∞ l0 ( ( ) (7. f ( x.30d) 151 . y )∂x∂y l1 l0 ∞ l1 +4 ∫ ∫ ( xl1 + yl1 ) . f ( x. f ( x. f ( x. f ( x.13 is exploited. y )∂x∂y l2 0 ∞ l0 (7. y )∂x∂y 2 2 l1 l1 ( ) The cross correlation (7. substitutions are made to remove σ and reduce the complexity of the correlation functions.30b) ∂X = σ 2 (7. Again the symmetry of Figure 7. y )∂x∂y 0 0 l1 l1 +8∫ ∫ ( x 2 + y 2 ) . X = x σ 2 ∂x (7. y )∂x∂y 2 l1 l0 ∞∞ ∞ l1 ) +4 ∫ ∫ l1 + l1 .28) +8∫ ∫ l1 + y 2 .
33) L2 ∞ ∫ (L L0 0 2 2 + Y 2 ). Y )∂X ∂Y L1 1 1 The normalized equation for Ryy(0) is Ryy ( 0 ) Rxx( 0 ) +8 ∫ +8 ∫ +8 ∫ +8 ∫ L2 L1 ∞ = 4∫ L0 L1 0 ∫ (X L1 0 2 2 + Y 2 ). f ( X .31) The normalized equation for α is α= +8 ∫ +8 ∫ +8 ∫ Rxy ( 0 ) Rxx( 0) L2 L0 L1 ∞ 0 = 4∫ 2 L1 0 ∫ (X L1 0 2 + Y 2 ).30g) Substituting (7.30) into (7. (7. Y )∂X ∂Y Where f(X.30e) L1 = σ 2 (7.30f) L2 = σ 2 (7. f ( X . Y )∂X ∂Y L1 ∞ ∫ (L L1 L0 2 1 L1 ∫ ∞ L1 2 L1 . Y )∂X ∂Y ∫ (X 0 + Y 2 ).27). Y )∂X ∂Y ∫ (X L0 0 L1 + Y 2 ). The normalized equation for Rxx(0) is ∞ Rxx( 0 ) = 2σ 2 ∫ −∞ −∞ ∫ (X ∞ 2 + Y 2 ). f ( X . Y )∂X ∂Y (7. Y )∂X ∂Y + Y 2 ). f ( X .29) yields the simplified expressions for the 3 sector clipping correlation functions. Y )∂X ∂Y 1 +4 ∫ L1 ∫ ( XL + YL ). f ( X . f ( X . f ( X . Y )∂X ∂Y 2 L2 ∞ ∫ ( XL L0 ∞ 2 (7.Y) is given by (7. Y )∂X ∂Y + Y 2 ).32) L1 ∞ ∫ ( XL + Y ).28).34) e− X −Y 2 f ( X . f ( X . Y ) ∂X ∂Y (7. f ( X . and (7.34) 152 .Y ) = 2 π (7. f ( X . f ( X .Chapter 7: Reduced Complexity Clipping Algorithms l0 l1 l2 L0 = σ 2 (7.
33) are shown in (7. Rxx( 0 ) = 2σ 2 (7. (7.10) to calculate the SNR at various clipping levels.38) 153 . L0 e 0 − L1e 1 + 2 2 π 2 2 2 − L1 2L L e ( erf ( L1 ) − erf ( L0 ) ) + 1 e − L1 (1 − erf ( L1 ) ) + 1 π π Which after expansion and reduction becomes α= Rxy ( 0 ) Rxx( 0) = erf ( L2 ) erf ( L0 ) − erf ( L1 ) erf ( L0 ) + erf ( L1 ) (7. and (7. (7. and (7.38).36) α was calculated to be (7.36).32).40).Chapter 7: Reduced Complexity Clipping Algorithms The normalized clipping levels are r dBclip = 20 log10 σ 2 dBclip = 20 log10 (R ) dBclip R = 10 20 (7. These are substituted into (7.31).37) 2 −L erf ( L1 ) = 2erf ( L1 ) 1 e− L1 + 2 π α= Rxy ( 0 ) Rxx( 0) 2 −L erf ( L0 ) +2 ( erf ( L2 ) − erf ( L1 ) ) .37) −L 2 2 π π −L 2 erf ( L1 ) − erf ( L0 ) (1 − erf ( L1 ) ) . 0 e − L0 + 2 π + + 2 L2 2 2 erf ( L0 ) −L e − L2 erf ( L0 ) + 2 1 − erf ( L2 ) 0 e − L0 + 2 π π (7.35) The evaluated correlation expressions of (7.
41). 154 .39) Ryy ( 0 ) Rxx( 0 ) 2 −L erf ( L1 ) = 2erf ( L1 ) 1 e − L1 + 2 π 2 −L erf ( L0 ) +2 ( erf ( L2 ) − erf ( L1 ) ) 0 e− L0 + 2 π 2 −L erf ( L2 ) L1 − L12 erf ( L1 ) +2erf ( L0 ) 2 e − L2 + + e − 2 2 π π 2 −L erf ( L0 ) 2 +2 L2 erf ( L0 ) (1 − erf ( L2 ) ) + 2 (1 − erf ( L2 ) ) 0 e− L0 + 2 π (7.40) 2 2 L −L = 2erf ( L0 ) 2 e − L2 + 1 e − L1 π π Ryy ( 0 ) Rxx( 0 ) + erf ( L0 ) erf ( L2 ) − erf ( L1 ) 2 2L 2 +2 L2 erf ( L0 ) 1 − erf ( L2 ) − 1 e − L1 + erf ( L1 ) +2 L1 erf ( L1 ) erf ( L0 ) − 1 + 2 L1 1 − erf ( L0 ) 2 2 (7.40) π The closed form solution is shown in (7. was derived to be (7.39) 2 −L erf ( L1 ) L0 − L02 erf ( L0 ) +2 (1 − erf ( L1 ) ) 1 e − L1 + + e − 2 2 π π 2 2 2 +2 L1 (1 − erf ( L1 ) ) ( erf ( L1 ) − erf ( L0 ) ) + 2 L1 (1 − erf ( L1 ) ) Which after expansion and reduction becomes (7.Chapter 7: Reduced Complexity Clipping Algorithms Ryy (0) Rxx (0) The normalized autocorrelation of the output.
Table 7.14 is derived and shown in Table 7. A plot of the theoretical SNR vs. where it is seen that the LUT requires a 8 bit input and a (N×2)+2 bit output. A flowchart detailing the decision matrix in the algorithm is shown in Figure 7. Further more when reviewing the truth table in Table 7.1: Truth Table for 3 Sector clipping Input Iin 0 X X X 1 1 L0 X X X 0 1 L1 1 X 1 X 0 L2 X 1 X X X 0 X 1 1 X X L0 X 0 1 X X Output Qin L1 1 X 0 X 1 L2 X X X 1 X L1 L2 L1 Iin Iin Iin L1 Qin Qin L2 L1 Qin Iclip Qclip Ctrl Ctrl I 1 1 1 0 0 0 Q 1 0 0 1 1 0 For all other combinations data is passed through 155 .14.1 the inputs for ‘0’ and L2 can replaced with logic 0 and 1 respectively reducing the number of inputs to the LUT to 4 and by implication the size of the LUT. The latency of the structure in Figure 7.14 is low as the level comparisons are made in parallel and fed into the LUT. From this the LUT in Figure 7.Chapter 7: Reduced Complexity Clipping Algorithms SNR3 Sec = ( erf ( L ) erf ( L ) − erf ( L ) erf ( L ) + erf ( L ) ) 2 0 1 0 1 2 − L2 − L22 L1 − L12 e + e 2erf ( L0 ) π π + erf ( L ) erf ( L ) − erf ( L ) 0 2 1 erf ( L ) erf ( L ) 2 2 0 +2 L2 2 erf ( L0 ) 1 − erf ( L2 ) − −erf ( L1 ) erf ( L0 ) 2 L1 − L12 e + erf ( L1 ) − +erf ( L1 ) π +2 L12 erf ( L1 ) erf ( L0 ) − 1 +2 L12 1 − erf ( L0 ) (7. Note that the sign bits are removed in Iin and Qin and attached back on at Iout and Qout.1.15. 3 Sector clipping is shown in Figure 7.41) A block diagram showing 3 Sector clipping implemented with a LUT is shown in Figure 7.16.
14: Block diagram of 3 Sector clipping. 156 .Chapter 7: Reduced Complexity Clipping Algorithms 0 N N L0 Iin L1 N Iout Iclip Control I L2 LUT Control Q N 0 Qclip Qin N L0 Qout N L1 L2 Figure 7.
14 (3 Sector clipping). 157 .15: Flowchart for the LUT in Figure 7.Chapter 7: Reduced Complexity Clipping Algorithms I+j×Q Store sign bits yes Abs(I)≥L1 no yes Abs(Q)≥0 no yes Abs(Q)>L0 no yes Abs(I)≥0 no yes Abs(I)>L0 no Abs(I)+j×abs(Q) Abs(I)≤L1 no Abs(I)≤L0 no Abs(Q) ≤L1 no Abs(Q)≤L0 no Abs(Q)≥L1 no yes L1+j×L1 yes Abs(I)≥L2 no yes Abs(I)≥L1 no yes Abs(Q)≥L2 no yes Abs(Q)≥L1 no yes L2+j×abs(Q) yes L1+j×abs(Q) yes Abs(I)+j×L2 yes Abs(I)+j×L1 Reattach sign bits Figure 7.
That is by setting both I0 and I2 equal to I1 in Figure 7. making it only marginally more complex than square clipping.1. clip level are shown in Figure 7. This makes sense as harder clip levels reduce the size of regions 1 and 2 in Figure 7.Chapter 7: Reduced Complexity Clipping Algorithms 7.13. making the corner sector the dominant clip region. the noise is measured after clipping. Generally 3 Sector clipping suffers a 1dB penalty in SNR compared to conventional clipping levels above 0dB.5 Theoretical Results The theoretical results for Conventional. Therefore 3 Sector and Square clipping share the same lower SNR bound of 2.4 Square Clipping Square clipping is the least complex form of clipping and is shown in the following theoretical results as a lower bound on the presented clipping techniques.2. oversampling. 7.32) and (7.1.16 together with the equivalent simulated results. 3 Sector. just as in Square clipping. The theoretical SNR for Square clipping can easily be found by using the same method as for 3 Sector clipping and changing the limits on (7. and Square clipping SNR vs. Note that for the simulated results no filtering. The equations are simpler to derive and can be construed as 2 sector clipping which is in fact Cartesian clipping. 3 Sector clipping has performance in the middle. Also of note it is seen that sector clipping performance approaches the same SNR as Square clipping at clip levels below 0dB.33) where appropriate.439dB. Conventional clipping has a lower SNR bound of 5. but with greatly reduced complexity requiring only a few comparators and a simple LUT. 158 . In other words in order for 3 sector clipping to achieve the same SNR as Conventional clipping the clip level needs to be set 1dB higher than the Conventional case. As expected Conventional clipping has the best performance and Square clipping the worst.13.2.634dB. or amplification is performed.
simulated (solid). At higher levels of clipping the simulated results have slightly better SNR. 159 .16: Theoretical Clip Level vs.e. θ (shown in Figure 7. The choice of θ has no impact on harder clipping levels. L 2 for all practical clipping levels as shown in Figure 7.16). Simulated OFDM symbols have lower PDF values in the tail of the distribution.44dB 0 10 8 6 4 2 0 Clip Level (dB) 2 4 6 8 10 Figure 7.5°. Figure 7. SNR for Conventional (Standard). this is due to the theoretical Rayleigh distribution assumption of the signal not holding at higher amplitude levels [89].11) vs.5 ) . i.16) to 10dB (at the top of Figure 7. SNR for clip levels ranging from 10dB (at the bottom of Figure 7. Theoretical (dashed). It is seen that the optimum angle for θ is 27. and Square Clipping. but it does affect SNR performance at weaker levels L (CL>5dB). The angle θ is varied from 20° to 40°. 0 = tan ( 27.17. Comparing the simulated curves with the theoretical ones it is seen that they are well matched at clip levels below 6dB. 3 Sector.Chapter 7: Reduced Complexity Clipping Algorithms 70 Theoretical Simulated Conventional 3 Sector Square 60 50 40 SNR (dB) 30 20 SNRlimit=5.17 plots the theoretical 3 sector clipping angle.64dB 10 SNRlimit=2.
the SNR for 3 sector clipping.2.41). Figure 7. Increasing the numbers of sectors from 3 to 4 improves the SNR by ~2dB at a clipping level of 6dB.19 depicts simulated results for Sector Clipping with 3. 5 sectors provides a further 0. Figure 7. 4.11) vs. the SNR difference between the schemes increases with clipping level.2 Extensions of Sector Clipping 3 Sector clipping can be extended to incorporate more sectors thereby improving performance. 160 .18 shows the I Q plot of the first quadrant of 4 and 5 Sector clipping.17: Clipping angle. θ (of Figure 7. and 5 sectors as well as Conventional and Square clipping for comparison.Chapter 7: Reduced Complexity Clipping Algorithms 50 CL=10dB 45 40 35 30 SNR (dB) 25 20 15 10 5 CL=10dB 0 20 22 24 26 28 30 32 34 36 38 40 θ (°) Figure 7. 7.5dB gain in SNR at the same clipping level. based on (7. Generally. The complexity increase for 4 and 5 sectors is minimal over 3 sector clipping requiring only a few extra comparators and a doubling of the LUT size.
75°. β=22.18: I Q diagram showing the 1st quadrant Sector clipping regions of a) 4 Sector clipping and b) 5 sector clipping. The new method is very simple to implement requiring only a few extra comparators compared to square clipping. i. for 4 sectors: θ=28. and α=33. Results not shown here indicate that increasing the number of sectors above 5 shows no discernable improvement in the SNR.13° and β=39. The degradation in SNR for Sector clipping from optimal Conventional clipping is ~3dB and the improvement in SNR over square clipping is ~4dB at a clip level of 6dB. however its performance is closer to conventional clipping where traditionally more complex circuits are required.Chapter 7: Reduced Complexity Clipping Algorithms (a) (b) Figure 7.25°.e.5°. The choice of angles for 4 and 5 sector clipping has been found to be optimum when the angles which determine the sectors are equally spaced. and 5 sectors: θ=11.75°. This section detailed a new method for clipping OFDM symbols called Sector Clipping. 161 .
7.19: Simulated clip level vs.5µ standard cell and port library.3 Hardware Implementation 3 Sector and Square clipping was implemented in digital form via Visual Hardware Design Language (VHDL). 3. where routing and cell placement was performed with a 0.2. therefore the algorithm was only proved through hardware simulation. and 5 Sector Clipping. etc. SNR for Conventional. The two algorithms were then simulated using ‘Synopsys’ to ensure proper operation. 162 . this was not done.Chapter 7: Reduced Complexity Clipping Algorithms 70 Conventional 5 Sector 4 Sector 3 Sector Square 60 50 40 SNR (dB) 30 20 10 0 10 8 6 4 2 0 Clip Level (dB) 2 4 6 8 10 Figure 7.21 depicts the layout of the 3 sector clipping algorithm with N=8 bit inputs.) code and simulated again in ‘Synopsys’. 4. Square. Next the algorithms were compiled into ‘Verilog’ (VHDL code expressed in terms of gates. The ‘Verilog’ code was then exported into the ‘Cadence’ silicon design package. flip flops. However. ‘Silicon Ensemble’. Figure 7.20. A gain table was included to allow 16 different Clipping levels (CL) set from 3dB to 12dB. After routing a ‘gds2’ file was produced which could then be sent to the foundry for production. The design flow is summarized in Figure 7.
5µ standard cell library Standard port library Foundry Figure 7. 163 . 8 Q shift reg.Chapter 7: Reduced Complexity Clipping Algorithms VHDL Simulate (Synopsys) Verilog testing Silicon Ensemble (Cadence) Gds2 file Cmos 0. Iin eri Qin eri Input data 8 8 7 CL 4 eri 7 Gain table Sector Clipper (Fig 7.14) Qin eri Iout eri 7 Qout erial Figure 7.20: Design flow for silicon implementation of 3 sector clipping. Iserial in erial Qserial in erial 8 I shift reg.21: Block diagram of 3 Sector clipping implemented in VHDL.
21) Hierarchy level Report Total cell area Square clipping 13593.21 without the input and output buffers. In practice however the buffers are not necessary as only the clipping algorithm is implemented as a block in the whole OFDM transmitter. The algorithm only takes up 4.08 dbu 8.55 dbu 35.16 dbu 11. In order to save the number of input/output pins required.Chapter 7: Reduced Complexity Clipping Algorithms Due to the clipping operation the Most Significant Bit (MSB) on each of the I and Q channels can be removed at the output. Table 7.1893mW System with buffer Total dynamic power Total cell area System Total dynamic power Total cell area Sector Clipper Total dynamic power Total cell area Gain table Total dynamic power Total cell area I and Q shift registers Total dynamic power ‘System’ again refers to Figure 7.3737mW 37.70 dbu 4.62% for 164 .3874mW 591.37 dbu 6. Table 7.2 shows a summary of various ‘Synopsys’ reports for the power consumption. ‘System with buffer’ refers to the whole system shown in Figure 7. delay path.36 dbu 0.1893mW 3 sector clipping 13827.2: ‘Synopsys’ reports for Square and 3 sector clipping.1800 dbu 17. and cell area of Square and 3 Sector clipping.27% of the total cell area for 3 Sector clipping and only 2. (Refer to Figure 7.38 dbu 8.268mW 357.8317mW 62.7933mW 131. The cell area is greatly reduced in this case as the buffers take up a lot of space. the I and Q input data is fed serially (Iin serial and Qin serial) into shift registers which then feed 8 bit words (Iin and Qin) into the 3 Sector clipping algorithm.4682mW 37.21 with the input and output buffers included.52 dbu 23.5455mW 385.36 dbu 0. The buffers are necessary when the system is implemented as a stand alone device in silicon. The input data word is stored and output in parallel fashion for comparison to the clipped sample.5896mW 110.
Chapter 7: Reduced Complexity Clipping Algorithms Square clipping.70 CORNERSITE_495 Rows CORNERSITE_495 Cells 100.62% for 3 Sector clipping.37 dbu for 3 Sector and Square clipping respectively. the power consumption is 8. The shift registers are the same in both systems. Comparing the number of ‘CORE Rows’ and ‘CORE Cells’ for Square and 3 Sector clipping algorithms it is seen that Square clipping requires 22 rows and 148 cells while 3 Sector clipping requires 26 rows and 319 cells.8317mW for 3 Sector and Square clipping respectively. As is seen in the Application Specific Integrated Circuit (ASIC) view of 3 Sector clipping Figure 7. Table 7. The clipping algorithms themselves (‘Sector Clipper’ in Table 7.00 IOPADSITE_495 Rows IOPADSITE_495 Cells 100.38 dbu and 110.2) have an area of 385.4 show the ‘Cadence’ reports for Square and 3 Sector clipping respectively.3 and 7. A 0. The area of utilization (occupied chip area) is 74. The 3 Sector Gain Table requires 50% more cells than Square clipping and the power consumption is doubled due to the extra clip levels required for sector clipping.00 Number 22 148 4 4 4 28 Length 1205160 237380 198000 198000 420000 420000 Area 3012900000 593450000 9801000000 9801000000 20790000000 20790000000 Area of chip: 41699680000 (square DBU) Area required for all cells: 31184450000 (square DBU) Area utilization of all cells: 74.3: ‘Cadence’ area utilization report on Square clipping Type %_Row_Space CORE Rows CORE Cells 19.78% 165 .22 the size of the chip is determined by the input/output pads which are abutted to make the area as small as possible.5µ process with 3 metal layers was used. Table 7.7933mW and 6.78% in Square clipping and 75. This made the wire routing process easy as there was a lot of room to work with. a small difference.
22: Application Specific Integrated Circuit (ASIC) view of 3sector clipping algorithm implemented in Silicon using ‘Cadence Silicon Ensemble’.5µ process. 0.Chapter 7: Reduced Complexity Clipping Algorithms Table 7.00 IOPADSITE_495 Rows IOPADSITE_495 Cells 100. 166 . 3 metal layers.62% Figure 7.00 Number 26 319 4 4 4 28 Length 1693120 390060 198000 198000 420000 420000 Area 4232800000 975150000 9801000000 9801000000 20790000000 20790000000 Area of chip: 41744615000 (square DBU) Area required for all cells: 31566150000 (square DBU) Area utilization of all cells: 75.04 CORNERSITE_495 Rows CORNERSITE_495 Cells 100.4: ‘Cadence’ area utilization report on 3 Sector clipping Type %_Row_Space CORE Rows CORE Cells 23.
to clip samples. x . of fixed phase angles.. and the magnitude of the real part. K.44) 167 ..min xi . towards the real axis.. x = xi + jxq is folded into the first octant to give x ' = max xi .43) where lk = {1.42) x’ is then rotated by a number.3 New Vector Subtraction clipping method This section describes another new method called Vector Subtraction which is an enhancement of an existing algorithm [101] called the Lucent algorithm here.e − jθk θk ( ( )) (7.5. is the closest approximation to the actual magnitude. θk. 2 K − 1} . Vector Subtraction reduces complexity of the Lucent algorithm by removing the need for divisions. xq + j. The Lucent algorithm works as follows.3.Chapter 7: Reduced Complexity Clipping Algorithms 7. which are complex operations in hardware and add significantly to the complexity of the original algorithm.3.1 Lucent Algorithm The Lucent algorithm produces good estimates of the magnitude in K iterations. The phase angle which returns the largest real part gives the closest approximation to the ˆ direction (phase) of the vector x’. 7. For all K values of θk the magnitude estimate is given by ˆ x = max Re x ' . first the complex sample.. xq ( ) ( ) (7. The K phase values are spaced in the octant as θ k = π lk 16 K (7.
7.2 Vector Subtraction ˆ The process of finding the magnitude estimate. xclip . are multiplied by a scaling factor (which reintroduces complexity that was mitigated by the iterative magnitude estimator).max giving the correction vectors yi and yq. os. x .46a) The correction vectors are octant adjusted prior to subtraction from the original signal.45) ˆ The only error in this technique is a slight under estimate of x which reduces as K increases. 168 . The scheme has many similarities to the CORDIC [102] method. x’. However the phase of the main signal is not known therefore the best estimate of the phase is used.m (7. but gives better estimates of the magnitude at low values of K. First.46) and then subtracted from the signal. the clipped output signal is given by xclip xout = x ( ˆ x x ) ˆ x > xclip ˆ x ≤ xclip (7. The overshoot.Chapter 7: Reduced Complexity Clipping Algorithms After the magnitude has been estimated. is calculated ˆ os = x − xclip (7. the overshoot. The scaling operation requires a division which is a complex hardware operation and can be avoided by using the new Vector Subtraction technique described below. os.44). this is θk. is rotated by e jθ k . is the same as [101] described in (7. x. y = yi + jyq = os e jθ k . but the scaling operation is replaced by a subtraction.max.3. any samples exceeding the clipping threshold.
23: Block diagram of new Vector Subtraction scaling operation.24 shows the I Q diagram for Vector Subtraction with K=2 iterations corresponding to phases of π 16 and 3π 16 . x’’. has an additional phase error compared to the Lucent method. Note that the clipped value.+) ' xq' Figure 7. Here it is seen that a slight underestimate in the magnitude means that samples are not always clipped back to the desired level.48) The addition/subtraction of yi and yq can be extrapolated from the maximum and minimum values of xi. yi ) (7.7dB and 1.23. 169 . Figure 7. xi xq (.47) (7.max jθk . It has both amplitude and phase error compared to the ideal clipped value. k ) Im e ( ( jθk . The vector x’ is closest to the π 16 vector. This is a function of the number of iterations in the algorithm with 1 iteration underestimating the magnitude by 0.max ) ) yq (..25 and 7. This problem is exacerbated in Vector Subtraction where the new scaling operation magnifies the error in the magnitude estimate.+) ˆ x Re e ˆ x > xclip ˆ os = x − xclip yi xi'' e jθ(1. A block diagram of Vector Subtraction is shown in Figure 7. xq and the sign bits of the original data. yq ) ' xq' = xq ± ( yq . Figure 7.26 show the CCDF of the Lucent algorithm [101] and the new Vector Subtraction variation respectively where 1000 OFDM symbols (N=64) are clipped at 5dB.2dB to pass through in the Lucent algorithm and Vector Subtraction respectively.Chapter 7: Reduced Complexity Clipping Algorithms xi'' = xi ± ( yi .
This is not an issue with the Lucent patent where the magnitude estimate error is constant irrespective of clip level. Alternatively more than 1 iteration can be used making the error in the magnitude estimate small.24: IQ plane for the new Vector Subtraction method showing vector. 170 . The upshot of this is that the clip level will need to be backed off to avoid saturation of the amplifier. x’ being clipped to x”.max I Figure 7.Chapter 7: Reduced Complexity Clipping Algorithms Q x’’ yq yi x’ θk. This small error can normally be neglected as filtering will cause substantial peak regrowth anyway. The amount of error in the magnitude estimate increases with the decrease in clipping level for Vector Subtraction with clipping under 3dB requiring a prohibitive amount of backoff.
Chapter 7: Reduced Complexity Clipping Algorithms 10 0 10 1 Unclipped Pr(ζ> ζ0 ) 10 2 1 iteration 2 iterations 3 iterations 4 iterations 10 3 4 5 6 7 8 9 10 11 ζ0 (dB) Figure 7.26: Simulated CCDF for Vector Subtraction for various iterations clipped at 5dB showing the leakage of under clipped samples. ζ0 (dB) 171 . 10 0 Unclipped 10 1 Pr(ζ> ζ0 ) 1 iteration 10 2 2 iterations 3 iterations 4 iterations 10 3 4 5 6 7 8 9 10 11 Figure 7.25: Simulated CCDF for Lucent patent [101] for various iterations clipped at 5dB showing the leakage of under clipped samples.
3 iterations Lucent . 70 Conventional clipping Lucent .26. 172 . The reason for this is the increasing amount of back off required at harder clip levels to avoid saturation of the amplifier.4 iterations Lucent .7.1 iteration Square clipping 60 50 40 SNR (dB) 30 20 10 0 0 1 2 3 4 5 Clip Level (dB) 6 7 8 9 10 Figure 7. Note that the clip level is adjusted depending on the iteration and clipping method according to the CCDF results of Figures 7. for example the Lucent patent with 1 iteration requires the clip level to be set to CL0.28 for the Lucent algorithm and Vector Subtraction method respectively. 1 iteration results in a 5dB degradation from the optimum clipping method for the Lucent algorithm while 1 iteration in Vector Subtraction results in a more serious degradation across the board. SNR for Lucent clipping technique with varying iterations.25 and 7. This ensures that the samples will not saturate the HPA if it were present. making it impractical.Chapter 7: Reduced Complexity Clipping Algorithms The simulated clip level vs. SNR results are shown in Figure 7.27: Simulated clip level vs. 3. When 2. both the Lucent patent and Vector subtraction have very similar performance to the conventional clipping method. In fact its performance goes below that of Square clipping for clip levels below 4dB.2 iterations Lucent .27 and 7. as well as Conventional and Square clipping. or 4 iterations are used.
Conventional and Square clipping are also shown as a reference. and 5 Sector clipping where it is seen that increasing the number of sectors to 4 provides an improvement of 1 and a half magnitudes over Square clipping at 4dB clipping. 7.2 Vector Subtraction .29 shows the baseband clipping vs.28: Simulated clip level vs.4.1 analyses their performance in the OFDM system described in Section 7.15.1 Square Clipping iterations iterations iterations iteration 60 50 40 SNR (dB) 30 20 10 0 0 1 2 3 4 5 Clip Level (dB) 6 7 8 9 10 Figure 7.3 Vector Subtraction .4 Vector Subtraction .1 BERF Figure 7. 4.4.Chapter 7: Reduced Complexity Clipping Algorithms 70 Conventional Clipping Vector Subtraction . SNR for Vector Subtraction clipping with varying iterations.2 compares them in terms of their complexity by comparing hardware operations. the oversampling rate in the IFFT is set to 2. BERF for 3. A LPA is used initially and later simulations use a HPA (P=3) with increasing backoffs. 7. The mapping type used in the following simulations is 64 QAM.4. the number of taps in the RRCF is 128 with a roll off factor of 0. Section 7. Increasing the number of sectors to 5 provides a more modest decrease in the BERF beyond 4 Sector 173 .1.4 Comparison of new and existing clipping methods The previous sections described new low complexity clipping algorithms and evaluated them both theoretically and through simulation. as well as Conventional and Square clipping. Section 7.
BERF with a LPA. AWGN=0. however curiously 1 iteration has the best performance while conventional clipping has the worst. Conventional and Square clipping vs. At 5 dB clipping Conventional. 174 . This can be explained by reviewing Figure 7. and 5 Sector. Performance is nearly the same for all methods. 10 0 10 1 Conventional 5 sectors 4 sectors 3 sectors Square 10 BER 10 2 3 10 4 10 5 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. 3. and 4 iterations. and 5 Sector clipping have a BERF below 106. Figure 7.30 shows baseband clipping vs. 4. 5 sector clipping is 1 magnitude worse than conventional clipping at 4 dB clipping.15. This will be explored further later in this section. Note that Sector clipping over clips the data therefore performance after a HPA with a limited backoff may be better due to the extra regrowth allowed. 4. RRCF with 128 taps.Chapter 7: Reduced Complexity Clipping Algorithms clipping at 4 dB clipping. 2.25 where it seen that Vector Subtraction under clips some samples which would lead to saturation of the amplifier had it not been a LPA. BERF for Vector Subtraction with 1. and α=0. 128 point IFFT (os=2). 64 QAM symbols.29: Simulated 3.
and α=0. RRCF with 128 taps. 175 . Vector Subtraction has an almost identical performance to the Lucent method except for the trivial case of 1 iteration. AWGN=0. 128 point IFFT (os=2). BERF for the Lucent algorithm [101] where it seen that the performance of all iterations is almost the same as Conventional clipping. Again 1 iteration has slightly better performance than the other methods due to the under clipping of some samples as seen in Figure 7. BERF with a LPA. and 4 iterations). 2. 64 QAM symbols. Figure 7.Chapter 7: Reduced Complexity Clipping Algorithms 10 0 10 1 Conventional 4 iterations 3 iterations 2 iterations 1 iteration Square 10 BER 10 2 3 10 4 10 5 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. Conventional and Square clipping vs.31 shows baseband clipping vs. 3.25.15.30: Simulated Vector Subtraction (1.
33. BERF with a LPA. 64 QAM symbols. 4. AWGN=0. 7. and α=0. and 7. This section simulated the performance of the new and existing clipping algorithms in an OFDM environment with a LPA. 176 . with increasing backoff in the HPA. BERF for 3. RRCF with 128 taps. The rest of this section uses a HPA with different IBO’s relative to the baseband clipping level.34 show baseband clipping vs. and 4 iterations). Note that Sector Clipping overclips the signal in some instances while Vector Subtraction and the Lucent algorithm under clip the signal which explains the higher BERF’s of the Sector clipping when the HPA is taken into account.32. Changing HPA backoff Figures 7. Increasing the IBO in the HPA improves the BERF in all cases.31: Simulated Lucent [101] clipping (1. 128 point IFFT (os=2). Conventional and Square clipping vs.15. 2. 3.Chapter 7: Reduced Complexity Clipping Algorithms 10 0 10 1 Conventional 4 iterations 3 iterations 2 iterations 1 iteration Square 10 BER 10 2 3 10 4 10 5 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. and 5 Sector clipping. respectively.
8dB 10 6 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. vs.Chapter 7: Reduced Complexity Clipping Algorithms 10 0 10 1 LPA IBO=CL+2 IBO=CL+1 IBO=CL 10 2 10 BER 3 0.33: Simulated 4 Sector clipping. BERF with varying IBO in HPA. 64 QAM symbols. RRCF with 128 taps and α=0. BERF with varying IBO in HPA.15. 128 point IFFT (os=2). RRCF with 128 taps and α=0. 177 . AWGN=0. 64 QAM symbols. vs.15. 10 0 10 1 LPA IBO=CL+2 IBO=CL+1 IBO=CL 10 2 BER 10 3 10 4 10 5 0. AWGN=0. 128 point IFFT (os=2).32: Simulated 3 Sector clipping.8dB 10 4 10 5 10 6 10 7 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7.
8dB 10 6 0 1 2 3 4 Clip level (dB) 5 6 7 8 Figure 7. 64 QAM symbols. IBO=7dB). BERF with varying IBO in HPA.34 show that this additional distortion is equivalent to a 0. AWGN=0. Comparing Figures 7. The nonlinear characteristics of a HPA adds another source of distortion to the transmitted signal. 5 Sector clipping is just under 1 magnitude better in terms of the BERF than 3 Sector clipping at 5dB baseband clipping with an IBO of 2dB above the baseband clip level (i. Figures 7.32 and 7.Chapter 7: Reduced Complexity Clipping Algorithms 10 0 10 1 LPA IBO=CL+2 IBO=CL+1 IBO=CL 10 2 BER 10 3 10 4 10 5 0.15.34: Simulated 5 Sector clipping. vs. 128 point IFFT (os=2).34.32 to 7.e.8dB change in the CL (at BERF=104) when the HPA saturation level is equal to the 178 . RRCF with 128 taps and α=0. 1 dB of additional baseband backoff is required for 3 and 4 Sector and slightly more for 5 Sector clipping to maintain the same BERF as an LPA at a BERF=104 with no additional IBO in the HPA. At an IBO=CL+2 the BERF is within 1 magnitude of the LPA at 5dB clipping for all methods and within half a magnitude at 4dB clipping.
then the additional distortion is reduced. as the amplifier is backed off.15. vs.35: Simulated Vector Subtraction (1 iteration).2dB 10 5 10 6 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. AWGN=0. 64 QAM symbols.15.5dB 10 5 10 6 10 7 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. BERF with varying IBO in HPA.36: Simulated Vector Subtraction (4 iterations). 128 point IFFT (os=2). 10 0 10 1 LPA IBO=CL+2 IBO=CL+1 IBO=CL 10 2 BER 10 3 10 4 1. AWGN=0. 64 QAM symbols. clipping level (IBO=CL). RRCF with 128 taps and α=0. vs. IBO=CL+2. 179 . BERF with varying IBO in HPA. 128 point IFFT (os=2). RRCF with 128 taps and α=0.Chapter 7: Reduced Complexity Clipping Algorithms 10 0 10 1 LPA IBO=CL+2 IBO=CL+1 IBO=CL 10 2 10 BER 10 3 4 1. Obviously.
37 and 7. At an amplifier backoff of HPA=CL+2 the BERF is within 1 magnitude of LPA performance at 5dB clipping for all methods and within half a magnitude at 4dB clip level.5. BERF for the Lucent patent with 1 and 4 iterations respectively. This is due to the underestimation of the magnitude reducing the clipping distortion. As with the 2 previous methods increasing the IBO improves the BERF.39. Increasing the IBO improves the BER in all cases. 1. and increasing backoff in the HPA. There is little difference between 1 and 4 iterations when HPA=CL but this increases with the larger IBO in the HPA. The For comparison Square clipping with different backoffs in the HPA is plotted in Figure 7. Under the same conditions the difference is under 2 magnitudes for Figure 7.36 show baseband clipping vs.37 (1 iteration) at a baseband clip level of 6dB there is a difference of 1.2. Here it is seen that like Sector clipping Square clipping is more robust against a reduction in the amplifier backoff with only half a magnitude in difference between a LPA and a HPA with no additional IBO (IBO=CL). with increasing backoff in the HPA.38 show baseband clipping vs.Chapter 7: Reduced Complexity Clipping Algorithms Figures 7.5 magnitudes between HPA=CL and a LPA. Figures 7.38 (4 iterations). and 1.5 magnitudes 180 . However 1 iteration out performs 4 iterations for all HPA backoffs. Comparing the 3 methods it is interesting to note that Sector clipping has the least increase in BERF from IBO=CL to LPA with a difference of 1. In Figure 7. with an additional backoff of 2dB in the HPA (HPA=CL+2) 1 iteration outperforms 4 iterations by 0.2 dB at a BERF=104.35 and 7. When no extra backoff is allowed in the HPA above the baseband clip level the performance of 1 and 4 iterations is almost the same. performance of 4 iterations is slightly worse for IBO=CL+1 and IBO=CL+2. When no extra backoff is allowed in the HPA both 1 and 4 iterations have approximately the same performance. BERF for Vector Subtraction with 1 and 4 iterations.
BERF with varying IBO in HPA. and α=0. AWGN=0. 128 point IFFT (os=2).38: Simulated Lucent clipping (4 iterations). 64 QAM symbols. AWGN=0. RRCF with 128 taps. vs.2dB 10 5 10 6 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. 64 QAM symbols.4dB 10 5 10 6 10 7 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7. and α=0. BERF with varying IBO in HPA.15. 181 .Chapter 7: Reduced Complexity Clipping Algorithms 10 0 10 1 LPA IBO=CL+2 IBO=CL+1 IBO=CL 10 2 10 BER 3 10 4 1. vs.15. 10 0 10 1 LPA IBO=CL+2 IBO=CL+1 IBO=CL 10 2 BER 10 3 10 4 1. 128 point IFFT (os=2).37: Simulated Lucent clipping (1 iteration). RRCF with 128 taps.
1dB increase for 5 Sector clipping. BERF with varying IBO in HPA.Chapter 7: Reduced Complexity Clipping Algorithms for 3. For Vector Subtraction and the Lucent clipping method the increase required is around 1. RRCF with 128 taps. and 5 Sector clipping respectively. AWGN=0.4dB to maintain the same BERF. 10 0 LPA IBO=CL+2 IBO=CL+1 IBO=CL 10 1 BER 10 2 10 3 0.4dB 10 4 0 1 2 3 4 Clip Level (dB) 5 6 7 8 Figure 7.2dB increase in clipping level and Square clipping only requires an extra 0. Conventional (Figure 7. 4. and α=0. Reviewing the results in terms of performance at a set HPA backoff it is seen that for 3 Sector clipping there is an additional 0.5 compares the baseband clip level for the aforementioned clipping techniques required to maintain a BER of 104 for different backoffs in the HPA. and a 1.3 dB for 4 iterations.2 and 2 magnitudes. 128 point IFFT (os=2). For 3 and 4 Sector clipping decreasing the difference from a LPA to no additional backoff in the HPA (IBO=CL) comes at a cost of around an additional 1dB increase in the baseband clip level to maintain the same BERF=104.4dB of extra baseband clipping backoff 182 . 64 QAM symbols. Table 7.5dB for 1 iteration and 1. Vector Subtraction has a difference of 2.5 and 2 magnitudes for 1 and 4 iterations respectively.39: Simulated Square clipping vs.6) clipping requires a 1. while the Lucent algorithm has a difference of 2.15.
2 dB 7 dB 7 dB 6.6 dB Conventional 6.6dB required under LPA conditions.1 dB Vector subtraction 1 iteration (Fig 7.34) 7.8 dB Conventional (Fig 7.38) 7 dB 7 dB 6. Table 7.7 dB 6. under LPA conditions). Note that for 3 and 4 Sector clipping an extra backoff of 2dB in the HPA is sufficient to provide near optimal results (i.39) 8.5: Baseband clip level required to maintain a BER=104 at varying IBO in HPA.2 dB 5.9 dB 6.4 dB 6.32) 4 sectors (Fig 7.9 dB 6.8 dB Square (Fig 7.5 dB 7.33) 5 sectors (Fig 7.Chapter 7: Reduced Complexity Clipping Algorithms required than for 5 Sector clipping with no additional backoff in the HPA.6 dB Lucent 6 dB 6. however for 5 Sector clipping an extra backoff of 3dB is required to approach the LPA results.3 dB 6.7) 7 dB Square 6.8 dB 1 iteration (Fig 7.1 dB 8 dB 183 .3 dB 7.36) 7 dB 7 dB 6.e.2 dB 5.6 dB 5. Baseband clip level required for BER=104 at relevant HPA backoff HPA backoff CL CL+1 CL+2 LPA Sector Clipping 3 sectors (Fig 7.37) 4 iterations (Fig 7.6 dB 6.4 dB 6.5 dB 5. and an additional 0. interestingly 1 iteration slightly outperforms 4 iterations.6 dB 6.1 dB 6. Performance of Vector Subtraction and the Lucent method for both 1 and 4 iterations have similar performance at a set amplifier backoff.2 dB 8.35) 4 iterations (Fig 7.4 dB 8. Underestimation of the amplitude must therefore dominate yje phase error in the two schemes.2 dB 5.
4. CL+2. Here it is seen that there is much less distortion.Chapter 7: Reduced Complexity Clipping Algorithms 7. Increasing the IBO by 1dB (IBO=CL+1) results in spectral splatter 20dB below the signal power and an inband distortion of 2. CL+4. 184 . hence the greater ACI and in band distortion. With no additional backoff in the HPA (IBO=CL) the ACI is only 18dB below the signal power and the inband distortion is 3dB below the LPA case. With IBO=CL+2 the ACI is 22dB below the signal power and the inband distortion is 1.15.2 PSD results The PSD after the HPA and receiver filtering is shown in the following figures. 128 filter taps and the HPA is a SSPA with p=3. i.41 the baseband clip level is set at 0dB and 5dB respectively and the amplifier backoff is set at increasing levels above this clip level (IBO=CL. Matched RRCF.40 and 7. The out of band distortion will lead to interference with adjacent channels. An IBO level set at CL+3 results in 23dB ACI and 1.5dB. There is only 0. In Figures 7.2 the PSD is measured for each OFDM block and then averaged over many blocks.5dB difference in inband distortion between no additional IBO in the HPA and a LPA. The inband distortion leads to a worse BERF as the difference between signal power and noise is reduced.4. The filter used is the same as the previous simulations. CL+1. i.41 the baseband clip level is set at 5dB and the amplifier backoff is set at increasing levels above the clip level as in Figure 7. Baseband clipping at 0dB results in a large amount of both in band and out of band distortion (ACI). These PSD results of 7.5dB. and the ACI is 24dB for the IBO=CL case. and a LPA). and for higher backoffs is almost non existent.e.3dB inband distortion. α=0.40 and 7. In Figure 7. SNR is reduced. As in Section 3.40.e.41 show that harder clipping in the baseband results in greater peak regrowth which leads to heavier saturation of the amplifier. CL+3. Even with a backoff of 4dB in the amplifier (IBO=CL+4) 24dB in ACI is present and the inband distortion is 1dB below the LPA case. both in band and out of band.
LPA). RRCF.Chapter 7: Reduced Complexity Clipping Algorithms 0 5 10 15 Absolute Power (dB) 20 LPA 25 HPA=CL 30 35 40 45 50 4 3 2 1 0 1 2 3 4 Normalized Frequency (Hz) Figure 7.CL+2.15.CL+2. CL+1. 0dB clipping in baseband with increasing amplifier backoffs (HPA=CL.CL+3. LPA). alpha=0. CL+1.40: PSD. 128 taps.15. 128 taps. 5dB clipping in baseband with increasing amplifier backoffs (HPA=CL. alpha=0. 185 .41: PSD.CL+4.CL+4.CL+3. 0 5 10 15 Absolute Power (dB) HPA=CL LPA 25 20 30 35 40 45 4 3 2 1 0 1 Normalized Frequency (Hz) 2 3 4 Figure 7. RRCF.
CL+3. 186 . alpha=0. RRCF. with increasing amplifier backoffs (HPA=CL. CL+1.CL+3. with increasing amplifier backoffs (HPA=CL.Chapter 7: Reduced Complexity Clipping Algorithms 5 10 15 20 Absolute Power (dB) HPA=CL 25 LPA 30 35 40 45 50 4 3 2 1 0 1 2 3 4 Normalized Frequency (Hz) Figure 7. after 5dB clipping in baseband and receiver filtering.CL+4. LPA). after 0dB clipping in baseband and receiver filtering.15. LPA).15. RRCF.43 : PSD.CL+2.CL+4. 128 taps. 128 taps. CL+1. 0 5 10 15 Absolute Power (dB) 20 HPA=CL 25 LPA 30 35 40 45 50 4 3 2 1 0 1 2 3 4 Normalized Frequency (Hz) Figure 7.42: PSD. alpha=0.CL+2.
This section introduces a new adaptive clip and filter algorithm. both of which are important in LDA. x(n). 7. However LDA as presented in [103] requires the use of a conventional clipper which adds significantly to the overall complexity and latency of the algorithm. The suitably delayed correction vector is subtracted from the signal. LDA uses an extra matched filter before the standard pulse shaping filter to predict the response of the signal from which the amount of compensation required can be calculated. Here as expected the out of band distortion is mitigated but the in band distortion remains.42 and 7. In this section Vector Subtraction with 2 iterations is used in place of the conventional clipper in LDA and their performance is compared through simulation. This effect can be compensated by putting the clipper before the filter as done earlier in this chapter. and on to the 187 . However.46a) and then fed to 2 identical filters. v(n). The input data is interpolated to form the signal. A block diagram showing the LDA algorithm is shown in Figure 7.43 show the PSD after receiver filtering for 0dB and 5dB baseband clipping respectively. Latency is an important issue in LDA.45. Level Detection Algorithm (LDA) first presented in [103] which overclips the signal at certain times avoiding the peak regrowth issue and sparing the amplifier from saturation. (Figure 7. The input data is modeled as a complex Gaussian process which describes accurately either an OFDM (N>64) or CDMA distribution of data where the algorithm could be implemented. therefore Vector Subtraction is very useful in this algorithm. as shown clipping and filtering can regrow peaks causing saturation of the amplifier resulting in an increase in the BERF. The first filter is used for peak detection and for generation of the correction vector.Chapter 7: Reduced Complexity Clipping Algorithms Figures 7.44 with a more detailed view of the filtering operation shown in Figure 7. Vector Subtraction has the advantage of providing good estimates of the error magnitude with low latency.5 New adaptive clipping method Clipping after interpolation and filtering will ensure that the amplifier does not saturate eliminating AM to PM distortions in the amplifier. and low complexity. before being passed to the second filter. but results in ACI affecting adjacent channels. x.
exceeds the CL as shown in (Fig 7.46b) the correction vector.50) where h0 is the central filter tap value. α(n) will be α ( n ) = h0v ( n ) (7. When the filter output magnitude.Chapter 7: Reduced Complexity Clipping Algorithms rest of the transmission chain.49) and(7. This reduces the potential for additional peaks by including the latter half of the smeared v(n) waveform in the output estimate y(n). The correction required at the filter output to stop saturation at the amplifier is α(n) which is in phase with y(n) and has amplitude α ( n ) = y ( n ) − CL (7.51) 188 . n.47 shows the I Q diagram for LDA showing how the vector. is calculated based on the filter input samples. y(n).46c) show the positions where the correction takes place at n1. in order to compensate for the group delay of the filter the subtraction point is set at the centre of the first filters delay line. Combining (7.49) When the correction vector v(n) is added to the centre of the filter delay line. v(n). v(n) is subtracted from the second filter input. The clipping processor compares the amplitude of the first filter output with the Clipping Level (CL) threshold to detect peaks. and n+1. Figure 7. The correction vectors (Figure 7. The filter used is a linear phase RRCF.50) gives v(n) as CL y ( n ) v ( n ) = 1 − y ( n ) h0 (7.
45: Detailed block diagram of the LDA. A limitation with LDA is that the correction vector smears into other parts of the signal.Chapter 7: Reduced Complexity Clipping Algorithms I/Q I/P 8 x delay v(n) 1st filter y Clipping processor 2nd filter I/Q O/P CL Figure 7. the saturation level of the amplifier.44: Block diagram of system with Level Detection Algorithm (LDA). The lower CL reduces the chance of regrown peaks saturating the amplifier. 1st filter z1 x(n) h0 z1 h1 x(n1) 2nd filter RRCF + v(n) Clipping processor y(n) CL Figure 7. which can introduce new peaks where none existed before. This problem is especially apparent when a number of peaks appear successively. To combat this problem the baseband clip level CL is set lower than the Amplifier Clip Level (ACL). 189 .
The CE is the noise power of the clipped part in the amplifier.Chapter 7: Reduced Complexity Clipping Algorithms x(n) y(n) CL+α(n) CL time n+1∆ n n+1 n+1+∆ n time (a) v(n) (b) time n1 n n+1 (c) Figure 7. The MSE introduces in band distortion that interferes with the desired signal.46: a) Amplitude of zero padded input to filter b) amplitude of filtered output c) LDA correction vectors. The MSE distortion introduced by the compensation filters is filtered by the second filter so no extra ACI is generated. and is wideband and spread over many channels causing ACI. 190 . In order to compare the performance of the LDA 2 parameters are used: the Clipping Error (CE) produced by the amplifier and the Mean Square Error (MSE) introduced by the compensation vector.
A matched RRCF simulates the receiver filter and a decimator follows to sample the received data symbols. The measured clipped power is the average power of the difference between the amplifier input and output signals. A block diagram of the simulation model used to evaluate LDA is shown in Figure 7. 191 . The HPA amplifier used in this case is just a linear limiter which saturates at ACL while the baseband clipper saturates at CL. the MSE is the average power of the difference between the original data samples and the received ones. Similarly.Chapter 7: Reduced Complexity Clipping Algorithms Q x(n+1+∆)h(1+∆) y(n) x(n+1)h0 α(n)=v(n)h0 Correction vector x(n+1∆)h(1∆) I CL Figure 7. The interpolation factor is 8 and the RRCF has a roll off factor of 0.47: Vector representation of filtering with 3 active taps and the required correction vector to bring the output back to CL.2.48.
45) Amp Clipping Error (CE) Transmit signal power Measure clip noise RRCF RRCF In band distortion (MSE) Receive signal power Measure noise 8 RRCF Delay Figure 7. Figures 7.48: Block diagram of the simulation model used to evaluate LDA.50 the ACL is set at 8dB and the CL is varied from 7 to 12dB. both of which are implemented in a 64 tap and 128 tap filter.49 and 7.49. The ACL is set at 5dB and CL is varied from 2 to 9dB in Figure 7. Curves are plotted for Vector Subtraction with 2 iterations and Conventional clipping.50 plot the simulated CL versus the inband distortion at the receiver.Chapter 7: Reduced Complexity Clipping Algorithms CL Random data source LDA ACL 8 (Figure7. and in Figure 7. 192 .
5 10 SCL to input power 10.64 taps Vector Sub .2 iter .5 8 8.128 taps Inband distortion to Rx signal 7 7.5 9 9.2 iter .128 taps Vector Sub .64 taps Vector Sub .2 iter .64 taps Conventional .128 taps Vector Sub .50: In band distortion for Conventional and Vector Subtraction (2 iter) with 64 and 128 taps in compensation filter.2 iter . 193 .Chapter 7: Reduced Complexity Clipping Algorithms 10 12 14 16 18 20 22 24 26 Conventional .64 taps Conventional .5 12 Figure 7. ACL=8dB. 28 30 32 34 36 38 40 42 44 Conventional .5 11 11.128 taps Inband distortion to Rx signal 2 3 4 5 6 SCL to input power 7 8 9 Figure 7.49: In band distortion for Conventional and Vector Subtraction (2 iter) with 64 and 128 taps in compensation filter. ACL=5dB.
52 show noise produced when the CL is set smaller than the ACL. The results are obtained by varying the CL and recording the two error sources (CE and MSE). The amplifier CE is low enough for LDA with Vector Subtraction to meet the ACI specifications of most standards. The noise produced at the bottom of the curves is due to peak regrowth after baseband clipping saturating the amplifier. 194 . At the bottom of the curves where the noise is due to peak regrowth there is a divergence in performance between Conventional and Vector Subtraction with Conventional clipping out performing Vector Subtraction by around 6dB from 55dB to 61dB in terms of the clipping noise. Little variation is seen in performance between Conventional and Vector Subtraction with the 64 tap filter having slightly worse performance in the conventional case.Chapter 7: Reduced Complexity Clipping Algorithms In all cases when the CL is set below the ACL more inband noise is produced as the noise is dominated by the baseband clipping (CL). In the ACL=5dB case it is seen that at the top of the curves there is no discernable difference in performance between Conventional and Vector Subtraction.52 show the simulated results for the Clipping Error (CE) versus the inband distortion. When the ACL is increased to 8dB the performance improves across the board as expected. When the CL is set above the ACL the clipping noise is dominated by the amplifier and levels off at CL=9dB for ACL=5dB and at CL=10dB ACL=8dB. or Mean Squared Error (MSE) for a constant ACL. the number of taps also has a negligible effect on performance. with Vector Subtraction displaying 64dB of noise and Conventional clipping displaying no noise below 65dB.51 and 7.51 and 7. as a result clipping is mostly performed by the baseband clipper rather than the amplifier. Figures 7. all the clipping is performed by the amplifier and the most noise is produced as is evidenced by the top of the curves in Figures 7.51 and 7.52. The bottom part of the curves in Figures 7. Again Conventional clipping outperforms Vector Subtraction by at least 1dB. The number of taps in the filter has a negligible effect. When the ACL=CL=5dB the inband distortion is 20dB below the received signal and when ACL=CL=8dB the inband distortion is around 35dB below the receive signal. When the CL is set larger than the ACL.
2 iter .128 Vector Sub .128 Vector Sub .64 Vector Sub .2 iter .64 Conventional .64 Vector Sub . inband distortion. Performance curves for LDA using Conventional clipping and Vector Subtraction with 2 iterations. 35 Compensation filter has 64 and 128 taps. Performance curves for LDA using Conventional clipping and Vector Subtraction with 2 iterations. 40 CL > ACL Conventional .2 iter 128 taps 45 CL = ACL = 8dB 50 CL < ACL 55 60 65 44 42 40 38 36 34 32 Inband distortion to Rx signal 30 28 Figure 7. 195 .2 iter .Chapter 7: Reduced Complexity Clipping Algorithms 20 25 30 35 CL > ACL Conventional .52: Clipping error vs.64 Conventional . inband distortion. ACL=5dB.128 taps CL = ACL = 5dB 40 45 CL < ACL 50 55 60 65 70 26 24 22 20 18 16 14 12 10 Figure 7. Compensation filter has 64 and 128 taps.51: Clipping error vs. ACL=8dB.
and a SSPA with p=3. oversampling beyond this point provided little discernable improvement.Chapter 7: Reduced Complexity Clipping Algorithms 7. Initially a Monte Carlo OFDM transceiver simulation model was used to establish the basic parameters on which the new clipping algorithms were tested. Good parameters for further simulation were found to be oversampling the IFFT by a factor 2. This was shown to be untrue at higher clip levels where the clipping noise is more impulsive in nature. therefore making hard clipping a viable alternative at lower data rates. hence the divergence in results.15. the amount of oversampling in the IFFT. Theoretical analysis of Conventional. 3 metal layer process. Reports generated by Synopsys and cadence proved that the complexity increase of the Sector clipping algorithm was small compared to Square clipping with the advantage of having much better SNR properties. In this chapter the BERF was plotted against the clipping level with the noise set to zero. Various parameters such as the mapping type. the number of taps in the pulse shaping filter. Sector. The peak regrowth could be countered to a certain degree by oversampling the IFFT by a factor of 2. QPSK mapping was shown to be extremely robust to clipping.5µ. and the HPA backoff were varied to quantify their affect on the BERF. 4 and 5 Sector clipping provided a marginal improvement over 3 sector in terms of the BERF with a minor increase in complexity. The effect of filtering after clipping was also studied. Sector clipping was expanded to include more sectors thereby improving performance. 196 . At higher clip levels the theoretical and simulated results diverged as the theoretical SNR used the assumption that the clipping noise was Gaussian in nature. this meant that all the errors produced were due to clipping noise alone. it was shown through simulation how peak regrowth was more severe the harder the signal is clipped in the baseband. In Section 7. Both 3 Sector and square clipping were implemented in silicon using a 0. and Square clipping SNR was shown to be in good agreement with the simulated results.2 Sector clipping was introduced as a low complexity alternative to conventional clipping. the baseband clip level. a RRCF with 128 filter taps for a rolloff factor of 0.6 Conclusion This chapter presented new low complexity clipping methods for OFDM.
The variation called Vector Subtraction further reduced complexity by the removal of a complex multiplication for the price of 2 extra additions and a comparison operation. In this situation the HPA must have a peak power some 6dB above the average signal power. Section 7.Chapter 7: Reduced Complexity Clipping Algorithms Section 7. an IFFT with an oversampling factor of 2. Vector Subtraction. Square. The PSD results for conventional clipping at 0dB and 5dB at various HPA backoffs was shown. It was shown that while the Vector Subtraction and Lucent algorithms have better performance across the board.8 to 4dB above the average signal power with a HPA with backoff 2dB above the baseband clip level is recommended. a pulse shaping RRCF with 128 taps and α=0. Finally in Section 7. While the inband distortion remains in both clipping modes the ACI present at 0dB is up to 17dB below the signal power while ACI is almost non existent at 5dB baseband clipping. Its performance was found to be as good as conventional clipping with only 2 iterations. A low latency magnitude estimate was therefore required. 197 .4 compared simulated BERF results of the Sector. An iterative method similar to the CORDIC algorithm to calculate the magnitude of a complex sample. Sector clipping was more tolerant to harder backoff levels in the HPA. As most WCDMA and OFDM standards require an ACI of 20dB below the signal power it was shown that even with hard clipping the ACI specifications are met.5 Vector Subtraction with 2 iterations was implemented into an existing adaptive clipping system which required the clipping operation to take place within the filter. and Conventional clipping under various backoffs in the HPA. Simulated results showed that Vector Subtraction had similar performance to conventional clipping in this algorithm but with less latency and complexity. For a baseband clipping system with 16 QAM mapping.15.3 introduced another low complexity enhancement to an existing Lucent algorithm. Lucent. to maintain a BER of 104 clipping at 3.
however later papers identified promising code sets such as second order ReedMuller codes. and Tone Reservation /Insertion. Tone Insertion/variation used peak reduction carriers which introduced redundancy and required some additional processing at the receiver. The number of subcarriers. Initially Chapter 2 introduced the theory and principles behind OFDM and detailed scenarios where it is used. Chapter 4 began the literature review with an analysis of non distorted PAPR reduction techniques which included coding techniques. These methods are complex and the amount of PAPR is not guaranteed. PTS and SLM reduced the PAPR by producing a series of alternative transmit signals seeded from the same data source which are altered before the IFFT process so that they will have different PAPR properties. Coding introduced the most redundancy and became extremely complex at a higher number of subcarriers (N>64). CSS and TI were shown to be less complex than PTS when the number of phase rotations was greater than 4. The waveform with the lowest PAR is chosen for transmission. Chapter 3 identified contributing factors to the PAPR. and to a lesser degree the mapping type were also shown to contribute to the PAPR. Simulations of an OFDM system revealed how uncontrolled large peaks will saturate the HPA creating ACI and an increased BER at the receiver. called Cyclic Shifted Sequences (CSS) and Time Inversion (TI). 198 . SLM.Chapter 8: Conclusion Chapter 8 Conclusion This thesis analysed and proposed new methods to deal with the PAPR in OFDM. these were that peaks are a function of the IFFT operation. PTS. The general distribution of OFDM samples was shown to have a Rayleigh distribution. Chapter 5 proposed two new low complexity variations to PTS. where in phase waveforms add to create a large peak.
Chapter 6 picked up the literature review again from Chapter 4 for distorted PAPR solutions. and TI bringing the discrete and filtered CCDF to within 1dB of each other. These papers revealed that a backoff of around 6dB is sufficient to maintain a respectable error rate for 4 and 16 QAM mapping. CSS and TI where shown to display less peak regrowth after pulse shaping filtering. The gains in reduction in ACI were shown to be minimal when an acceptable amount of backoff in the HPA was used. The complexity was further reduced by the removal of a multiplication at the expense of 2 additions and comparisons. Chapter 7 presented the next set of new PAPR solutions where a series of low complexity and low latency clipping algorithms were proposed. It was implemented in silicon and shown to have negligible extra complexity to square clipping but with much better performance. This chapter started with a detailed Monte Carlo simulation of an OFDM transceiver were different stages of the transmission chain were modified to ascertain their effect on the BERF. Vector Subtraction was shown to have identical performance to conventional clipping with only 2 iterations. The rule base was extended up to 5 sectors where performance was close to conventional clipping. Finally Vector Subtraction with 2 iterations was implemented in a previously 199 . usually in the baseband so that the HPA would not saturate. Sector clipping uses a rule base to perform the clipping operation.Chapter 8: Conclusion Combining shifts of CSS and TI with non complex phase rotations of standard PTS allowed a whole IFFT operation to be removed at the expense of some extra non complex operations. Other papers analysed the affect on the amplifier of saturation and the resultant in band and out of band distortion. Oversampling at the IFFT by a factor of 2 was shown to improve the performance of PTS. CSS. clipping is a very rare occurrence affecting the BER negligibly. These methods provided reduction in the CCDF of between 2 and 3dB at Pr(ζ>ζ0)=104. The second method used a variation of a CORDIC like magnitude estimator and was called Vector Subtraction. Distorted PAPR methods were defined here as methods which intentionally limit the excess peaks at the transmitter. The first method. Windowing was also examined which is the process of using a pulse shape to clip the signal and surrounding samples to give better spectrum properties. The motivation behind clipping as a solution to the PAPR was that with an acceptable HPA backoff of around 6dB.
Further enhancements of Sector clipping can be made where the error introduced by the phase distortion could be minimised so that only the amplitude distortion remains.1 Future Work The PAPR problem in OFDM is still an ongoing issue. especially for portable devices where the need to minimise the power amplifier linear range is paramount. 8. The LDA algorithm with Vector Subtraction could be implemented in a complete OFDM transceiver and then implemented in hardware via a FPGA or silicon to ascertain whether the latency requirements can be met with the wider bandwidth systems now being proposed (40 MHz for 802. Code sets for larger numbers of subcarriers are an open ended problem and there is much ongoing research in this area. A hybrid system utilizing clipping techniques could also be added as a last stage so as to have an upper bound for the PTS signal. The PTS/CSS/TI methods developed in this thesis to reduce the PAPR can be combined with other PTS techniques such as adaptive PTS and variations of the blind SLM techniques to further reduce complexity and the peak power. Analysis of the proposed algorithms in a MIMO OFDM system is an area that has gained a lot of focus recently and analysis of the performance of the proposed techniques in such an environment would be valid.11n. and up to 100 MHz for 4G LTE).Chapter 8: Conclusion proposed clip and filter algorithm where its low latency was important as the magnitude had to be found quickly as it was operating in a feedback loop on filter samples. All of the clipping techniques detailed in this thesis could be combined with coding techniques to further improve the BER performance. 200 .
"Orthogonal Frequency Division Multiplexing for Wireless Channels. Mar 16 2000. 656661. "Efficient FFT algorithm for OFDM modulation. A. Nov 2000. "Fast Fourier Transform for high speed OFDM wireless multimedia system." in Circuits and Systems. and H. D. TENCON." IEEE Transactions on Communication Techniques. Single and Multicarrier Quadrature Amplitude Modulation. Hill. M. vol. "Reducing the peaktoaverage power ratio in OFDM by cyclically shifting partial transmit sequences. Cimini Jr. J. van de Beek. Jo." Division of [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] 201 .2. Hill." in School of Computer Science and Software Engineering Melbourne: Monash University. L. Heald. "Data transmission by frequency division multiplexing using the discrete fourier transform. 48. 2001. Park. vol. H." in IEEE Global Telecommunications Conference GLOBECOM. R. Kim..796. 1. Tech. pp. Mikhael. vol. Doelz. 560561." in PIMRC. M. T. Australia. T. Edfors. pp. Chichester: Wiley. Hanzo." IEEE Transactions on Signal Processing. 257. A. Park." in Electrical and Electronic Technology.. Sadat and W. O. Chang. 938942 vol. L. 2001. Ebert. Chang. W. L. "Performance of an efficient parallel data transmission system. Faulkner. 82.. Y. and J. T. Keller.. Jung. 2000. J. Faulkner. M. Proceedings of IEEE Region 10 International Conference on. "Efficient VLSI architectures for fast computation of the discrete fourier transform and its inverse. pp. vol." Electronics Letters. "Comparison of Low complexity clipping algorithms for OFDM. 2001. "Cyclic Shifting and Time Inversion of Partial Transmit Sequences to Reduce the PeaktoAverage Power Ratio in OFDM.7751. J. L.. 628634. p. E. May 1957. Webb. Sydney. 2001. pp. M. D.." Proc IRE. and J. MWSCAS 2001. S. and D.2. Weinstein and P.J. Spectrum and Coding). COM19. R. Singh. vol. Kim. Sjöberg. Portugal. 1998. Y." in PIMRC. "Binary Data Transmission Techniques for Linear Systems. C. Saltzberg. D. Wang. and Y. L. Singh. Jayalath. G. L. COM15. pp. J. Landström. Tak. G. C. October 1971. vol. 2002. 2000. "An Introduction to Orthogonal Frequency Division Multiplexing. p. pp. B. 45. "Synthesis of Band Limited Orthogonal Signals for Multichannel Data Transmission. Dec 1966.. pp. Dec 1967 1967. 2002. Chang. W. Martin. 36. "OFDM for Wireless Broadband Communications (Peak Power Reduction. Ye." IEEE Transactions on Communication Technologies. 676678 vol. Sandell. Hill. Proceedings of the 44th IEEE 2001 Midwest Symposium on.Bibliography Bibliography [1] G. S. 32063216. London. B. 45." Bell Syst.
"Design of Wireless Communication Systems . pp. pp. IEEE VTS 54th. pp. The 2001 IEEE International Symposium on. 1999.. J. June 1995. Tufvesson. framing structure. Faulkner." Melbourne. "OFDM: Overview. and Y. A." Broadcasting. J. 7. S. Bahai. A. Channel Estimation and MultiCarrier Systems. "The DVB terrestial (DVBT) specification and its implementation in a practical modem. Armstrong. J. Moeneclaey. H. M. "Twodimensional pilotsymbolaided channel estimation for OFDM systems with transmitter diversity. Jeon. 50. pp. W. 30653073. [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] 202 . pp. Apr 2002. Nov 1994. VTC 2001 Fall. R." ISO/IEC 13818. 43. "On the capacity of OFDMbased spatial multiplexing systems. Nov 1999. "The COFDM modulation system: The heart of digital audio broadcasting. "A comparison of pilotaided channel estimation methods for OFDM systems. and A. Paik. P. 664668 vol." IEEE Transactions on Communications. Cimini. T." in Vehicular Technology Conference. 49." IEEE Transactions on Broadcasting. M. 2002. and A. Armada. 1996. Gesbert. pp. Luleå University of Technology. TelladoMourelo. "Understanding the effects of phase noise in orthogonal frequency division multiplexing (OFDM). J. F. T. Wireless Communications. 47." in Department of Applied Electronics Lund: Lund University. E85B. and M. vol. "On the use of a cyclic extension in OFDM. Cho. Rappaport. Paulraj. 1996. 153159." in Circuits and Systems. Dec 2001. 4. G. Bolckei.generic coding of moving pictures and associated audio information: Part 2 video. 570573 vol. 2002." Communications. G. 255260." in Department of Electrical Engineering: Stanford University. Ergen. Jun 2001. pp. K." International Broadcasting Convention. 840844. 48. pp. Research Report TULEA 1996:16 1996." Electronic and Communication Engineering Journal. M. "Peak to Average Power Reduction for Multicarrier Modulation.Bibliography Signal Processing. IEEE Transactions on. ISO/IEC. vol. vol. "Digital Video Broadcasting (DVB). and S. vol. vol. S. 225234. 2000. 191193. vol. pp. 2001." IEICE Transactions on Communications. H. 2002. K. Shelswell. and modulation for digital terrestial television (DVBT). D. ISCAS 2001. M. S. 223229. Pollet. channel coding. "Blind frequency offset estimation for PCCOFDM with symbols overlapped in the time domain. Coleri. ETSI. 2001. 2001. Mengali. pp. Puri. portable and fixed receivers. New Jersey: Prentice Hall. vol." IEEE Transactions on Signal Processing. Shentu and J. 2000.Issues on Synchronization. "Information technology ." European Telecommunications Standard Institute ETSI EN 300 744. L. Wilson. Morrison. Morelli and U. 1995. Stott.2. ETSI. 2001. "Channel estimation techniques based on pilot arrangement in OFDM systems. Vanbladel. Jr." ETSI final draft ETS 300 401. J. "Ber sensitivity of ofdm systems to carrier frequency offset and wiener phase noise. "Radio Broadcast Systems: Digital Audio Broadcasting (DAB) to mobile. IEEE Transactions on. 127135. Principles and Practice.
47. pp. Mazzenga. 3739. 44. Tellambura. IEEE Transactions on. Khalaj. 1997. G. 488494. Fischer. Müller. 49." IEEE Transactions on Communications. Boston . vol. 51. G. G. Zhou and J. "Upper bounds on the statistical distribution of the crestfactor in ofdm transmission. "Central Limit Theorem. 1998. G. A. and J. vol. Imai. pp. "On the degradation of OFDM signals due to peak clipping in optimally predistorted power amplifiers. 142151. May 1998. Armada. and W. "Effects of clipping and filtering in the performance of OFDM. Ochiai and H." IEEE Transactions on Communications. 2003. vol. A. Boche. 52." IEEE Transactions on Vehicular Technology." in Reihe Kommunikations. 2000. E. Serdijn. pp." in Globecom. Cimini. pp." Communications Letters. X. Midrio. pp. 49. 443447. C.und Informationstechnik Nürnberg: Universität Erlangen. 2000. Sep 2001. H. 2. vol. Mar 2002. H. pp. S. vol. "Ofdm performance in amplifier nonlinearity. and J. pp." IEEE Communications Letters. vol. Santella and F." Information Theory. Pupolin." Communications. C. v. R. T.htm. 3. Pupolin." IEEE Transactions on Consumer Electronics. Huber. 169170. pp." Annales des Telecommunications Annals of Telecommunications. 34. M. E.Bibliography [32] S. L. pp. H. http://www. Annis. "Predicting spectral regrowth of nonlinear power amplifiers. 49. Merchan. C. 2003. London: Artech House. van den Bos. and S. 7278. Garcia. M. 282289. 15101514. S. C. GharaviAlkhansari." IEEE Transactions on Communications." IEEE Transactions on Broadcasting. vol." Communications. Bauml. Kenney. 1998. 50. vol. Sharif. J. vol. 50. 1999. Li and L. "A hybrid analyticalsimulation procedure for performance evaluation in mqamofdm schemes in presence of nonlinear distortions. Costa and S. "On the peaktoaverage power of ofdm signals based on oversampling. OFDM for wireless multimedia communications. L. MüllerWeinfurtner. Feb 2001. IEEE Transactions on. 131133. 462472. M. 1997. "Phase optimisation criterion for reducing peaktoaverage power ratio in ofdm." 2004. "On the distribution of the peaktoaverage power ratio in OFDM signals. pp. Liu. "MQAMOFDM system performance in the presence of a nonlinear amplifier and phase noise. [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] 203 . "Impact of amplifier nonlinearities on OFDM transmission system performance. L. H. Kouwenhoven. "Effect of smooth nonlinear distortion on OFDM symbol error rate. vol.com/central_limit_theorem. 106114. "Ofdm with reduced peaktoaverage power ratio by multiple signal representation. and Synchronization. IEEE Transactions on. H. Nee. vol. M." Electronics Letters. pp. M. "The effect of nonlinearity on a qpskofdmqam signal. 2002. Wunder and H. "OFDM for Wireless Communications: Nyquist Windowing. Friese. Costa. 5867. Sydney. 1998. vol.statisticalengineering. pp. 718722. R. R. vol. 43. B. PeakPower Reduction. and B. W. Australia. F. S. 1998. pp.
." in IEEE 46th VTC." Hewlett Packard Laboratories. 1996. M. Wilkinson and A. 11311132. V." in Radio Vetenskap och Kommunikation '96 (RVK). Orriss. pp. 8795. pp 680684.. 1996. C." in Vehicular Technology Conference.Bibliography [49] A. T. "Upper bound on peak factor on nmultiple carriers. Barton. R. 1998. 1997. Luleå. USA. 432433. Shepherd. 1995. Atlanta. February 1996." in Wireless Communications Netherlands: Kluwer Academic Publishers.1. Jafarkhani." IEEE Transactions on Communications. Barton. E. 33. Zhang.. Chicago. Wulich. Jones. IEEE Transactions on." in IEEE VTC. J. 31. 16081609. USA. "Coding to control the signal waveform in Mary PSK multicarrier communications. pp 658662. E. 1999. Physical (PHY) layer. S. GLOBECOM '96." Electronic Letters. 30. Y. K. [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] [62] [63] [64] [65] 204 . pp. 1986. pp." in Global Telecommunications Conference. October 1998. Bristol HPL98176. "OFDM peak power reduction by subblockcoding and its extended versions. K. 2000. NJ. USA." Communications. 33. Shepherd. T. pp. G. C. "On the computation and reduction of the peaktoaverage power ratio in multicarrier communications. 1996. pp. E. 695699 vol. pp. vol. Jones. 32.1. E. pp. A. W.Y. V. and S. pp. J. P. 1995. C. V.. Barton." Electronics Letters.K.1. Shepherd." Electronic Letters. "Combined Coding for Error Control and Increased Robusteness to System Nonlinearities in OFDM. "Asymptotic limits in peak envelope power reduction by redundant coding in orthogonal frequencydivision multiplex modulation. 20982099. S. "OFDM codes for peaktoaverage power reduction and error correction." Electronics Letters. L. 1995. "The Distribution of Peak Factor in QPSK MultiCarrier Modulation. vol. vol.. "Minimisation of the Peak to Mean Envelope Power Ratio of Multicarrier Transmission Schemes by Block Coding. A.J. Y. "Peak Factor in Orthogonal Multicarrier Modulation with Variable Levels. Sweden. A. "Multitone signals with low crest factor. Barton. pp 904908. Tarokh and H." in VTC.W. S. and S. vol. pp. "An Algorithm for Reducing the Peak to Average Power Ratio in a Multicarrier Communication System. 18591861. pp. vol. "Block coding scheme for reduction of peak to mean envelope power ratio of multicarrier transmission schemes. Piscataway. and S. pp. September 1996. WyattMillington. 'Communications: The Key to Global Prosperity. 3744." European Telecommunications Standards Institute ETSI TS 101 475 v1." in Electronics Letters. 510. vol. Boyd. Paterson." Circuits and Systems. 48. 32. Jafarkhani.W. pp. Tellambura. 2000. J. J. 46. D. "Broadband Radio Access Networks (BRAN). Wulich. 1999. Wilkinson. pp. E. A. D. vol. vol. Chouinard. "Bounds on Envelope Power of Trace Codes for OFDM. J. "Reduction of Peak to Mean Ratio of Multicarrier Modulation using Cyclic Coding. "Simple coding scheme to reduce peak factor in QPSK multicarrier modulation.. J. IEEE Transactions on. HIPERLAN Type 2. Nilsson. 1999 IEEE 49th.1. 1996. VanEetvelt. Tarokh and H. Zhang. ETSI. 740744 vol. K. P. van Nee. 1994. S. pp. S. J. Jones. 10181022.
" IEEE Communications Letters. pp. Huber. "Construction of ofdm m qam sequences with low peaktoaverage power ratio. Ulm. K. Golay complementary sequences. vol. Phoenix." in ICC'99. Jedwab." Communications. Gil." in Vehicular Technology Conference. "Efficient maximum likelihood decoding of peak power limiting codes for OFDM.Bibliography [66] J. J." in Vehicular Technology Conference. 267268. and ReedMuller codes. 1999. pp. 465469 vol. 23972417." IEEE Transactions on Information Theory. "On decoding of block codes with peakpower reduction in OFDM systems. "Ofdm with reduced peaktoaverage power ratio by optimum combination of partial transmit sequences. Jayalath. E. Germany WG3 temporary document 25. Sollenberger. Sathananthan and C.. pp. Helsinki. 2002. S. pp. Davis and J. Huber. 20812084 vol. and J. S. 316318. Sollenberger. J." in Globecom '97." ETSI BRAN .R. 1998. 1999. pp. 16 September 1997." Electronics Letters. 4. 2528. "Peaktoaverage power ratio reduction of an OFDM signal using partial transmit sequences. 368369. United Kingdom." Communications Letters. 949952. B. 45. 47. "An orthogonal projectionbased approach for PAR reduction in OFDM. 2001. R. "A comparison of Peak Power Reduction Schemes for OFDM. 1997.1. 48th IEEE. "Improving Peak to Average power Ratio (PAR) and Probability of Error in OFDMbased WLAN. pp. H. A. "A novel subblock partition scheme for partial transmit sequence OFDM. Pottie. 33. Jones and T. "PAR reduction of an OFDM signal using partial transmit sequences. Cimini and N. L. pp. 4. J. May 2002. H. vol.European Telecommunications Standards Institute Broadband Radio Access Network. "A novel peak power reduction scheme for OFDM. 169171. Armada. G. Wilkinson. A. 51. Tellambura and A. A. H. G.3. Davis and J. I. "Peaktoaverage power ratio reduction of an OFDM signal using partial transmit sequences. 2000. 1999. J. pp. R. vol. M." IEEE Transactions on Broadcasting. B. K. Grant and R. Müller and J. Kim. 10901094. 33. VTC 98. van Nee. 1998." Electronics Letters. J. Kang. 6. pp. IEEE VTS 54th. "Peaktomean power control in OFDM. 2003. C. Joo. 45. S. USA. Chen and G. pp. pp." IEEE Transactions on Communications. Imai. H. pp. L. [67] [68] [69] [70] [71] [72] [73] [74] [75] [76] [77] [78] [79] [80] [81] [82] 205 . 1999. 2001. vol. VTC 2001 Fall. 1997. vol. "Peaktomean power control and error correction for ofdm transmission using golay sequences and reedmuller codes. G. Ochiai. pp. Mar 2000. Huber. H. Müller and J. Finland. S. A. Garcia. vol." in IST Summit. Jedwab. vol. IEEE Transactions on. 226228. S. D." in IEEE Personal Indoor Mobile Radio Communications. "Coding to reduce both par and picr of an ofdm signal. Tarokh. "Performance of ReedMuller codes and a maximumlikelihood decoding algorithm for OFDM. Tellambura. S. 1997. V. A. C. Müller and J. B. B. Huber. 2001. 6. J. Müller and J. 8688. Fossorier." IEEE Communications Letters. A. vol. Cimini and N." Communications Letters. and H. B. 333338. P. H. 1997. "OFDM with Reduced Crest Factor. S. and E. L. vol. vol.
"Peaktoaverage power ratio reduction of OFDM signals using broadband pulse shaping. October 1996. Bolsens. [99] R. [91] T. and J. 13541355."?.?. vol. "Clipping noise mitigation for OFDM by decisionaided reconstruction. pp. A. Ottawa. Come. Nikopour and S. "Reducing the Peak to Average Power Ratio in OFDM Radio Transmission Systems." in Microwave Journal. T. vol. [89] A. [90] G. Garcia." Broadcasting. Canada. Spruyt. [92] J." in Vehicular Technology Conference. 246247. Fischer. Mestdagh. J.. pp." IEEE Communications Letters." Electronic Letters. [94] D. and I. 5. Bauml. 1998. 1999. G. "Effects of Cartesian Clipping Noise on the Performance of Orthogonal Frequency Division Multiplexing System: A Theoretical Approach. "Reducing the peaktoaverage power ratio of multicarrier modulation by selected mapping. 1996. Goldsmith. P. 20722076. Canada. pp. Wunder and H. Saltzberg. 889893. 1998. pp. Kim and G. 46. [83] 206 ." in IEEE VTC '02. Feb 28 2002. R. pp." in Globecom 1998.Bibliography R. Breiling. Jayalath. and J. [87] R. pp. M. Australia. "Reducing the PeaktoAverage Power Ratio of OFDM. Boche." IEEE Journal on Selected Areas in Communications. F. [93] D. vol. "Peak Power Reduction of Multicarrier Transmission. Singh. May and H. G. 219222. [88] H. [84] H." in The Embedded Signal Processing Conference (GSPx 2004). 12541258. and B. 22nd July 1993. pp." Electronic Letters. IEEE Transactions on. "The Impact of Nonlinear Devices on the Symbol Error Rate in Broadband OFDM Transmission. and J. 20562057. P. Ness. "A Method to Reduce the Probability of Clipping in DMTBased Transceivers. H. MullerWeinfurtner. vol. VTC 2002Fall. V. M. Wild. S. 32. 24742478. Slimane. "SLM peakpower reduction without explicit side information. S. and B. "Effect of Rouunding and Saturation in FizedPoint DSP Implementation of IFFT and FFT for OFDM Applications. Huber. B. 1999. Slaviero. [86] J. vol. Santa Clara. pp. Armstrong. 239241. v. J. pp. "A new approach for evaluating clipping distortion in multicarrier systems. Eberle. 2002. Mestdagh and P. Stuber. D. W. L. 126140. W. 2002. R. 2004. Gusmao. 3. vol. pp. Donnay. J. [96] B. 2001. vol. Jun 2002. "Peaktoaverage power reduction for OFDM by repeated clipping and frequency domain filtering. S. Armstrong. Spruyt. vol. L. H. 44. Huber." in IEEE GLOBECOM 2002. 1998. S. Tellado and J. pp. ?. Jamali. P." Communications Letters. Rohling. 10371046. S. 2001. R. Brewer. Armada. G. M. 106114. CA. Suraweera. [95] D. "A blind slm receiver for parreduced ofdm. "Impact of FrontEnd NonIdealities on Bit Error Rate of WLANOFDM Transceivers. Proceedings. and R. Wambacq. USA. "OFDM performance in amplifier nonlinearity. B. d. Sydney. 2002. pp. d. Bahai. P." in VTC '98. 1998. pp." Electronics Letters. Perre. Nee and A. 29. C. H. L. M. 2002 IEEE 56th. Dinis and A. [100] S. A. R." in Vehicular Technology Conference." IEEE Transactions on Communications. 44. [85] A. 2002. Biran. [98] S. H. 12341238. pp. A. "On the performance evaluation of OFDM transmission using clipping techniques." in VTC'99. Cioffi. Merchan. Engels. 38. 20. 2002. "Effect of Amplitude Clipping in DMTADSL Transceivers. [97] J. vol. B.
"System and method to reduce the peaktoaverage power ratio in DSCDMA transmitter. pp. Berangi and M." IRE Trans. "The CORDIC Trigonmetric Computing Technique. Iran. E. vol. "Peak Power Reduction for MultiCode CDMA and Critically Sampled Complex Gaussian Signals. 2001. [102] J. Electronic Computing. S." in IST 2001. Sept 1959. Creighton. 207 .Bibliography [101] R. Volder. 330334. 1999. Faulkner. [103] R." in European Patent Application Great Britain: Lucent Technologies (UK). 8.
This action might not be possible to undo. Are you sure you want to continue?