Practical Multitone Architectures: Vladimir Stojanović

Practical multitone architectures
Lecture 4
Vladimir Stojanovi
6.973 Communication System Design Spring 2006
Massachusetts Institute of Technology
Cite as: Vladimir Stojanovic, course materials for 6.973 Communication System Design, Spring 2006.
MIT OpenCourseWare (http://ocw.mit.edu/), Massachusetts Institute of Technology.
Downloaded on [DD Month YYYY].
Loading with discrete information units
Chows algorithm (rounding of waterfilling results)

Levin-Campello algorithm (greedy optimization)
Each additional bit placed on subchannel that requires

the least incremental energy for its transport
Incremental energy en (bn ) n (bn ) n (bn )
e (b ) g
bn = log 2 1 +
n n n
bit increment
Efficiency of bit distribution
Cant move a bit from one channel to another and reduce the
total energy
6.973 Communication System Design
Levin-Campello (LC) algorithm- components
Efficientizing (EF) finding energy-efficient bit

distribution
Always replace a bit distribution with a more efficient

exhaustively search all single information unit changes
at each step
Smallest energy to add a new bit
Largest energy to subtract the bit
Subtract bits from costly sub-channels
and add to least costly sub-channels
Efficientizing example
1+0.9D-1 channel (Pe=10-6, gap=8.8dB, PAM/QAM)
PAM and single-sideband

QAM
bn
bn +1
Rate adaptive solution
Need also to satisfy the energy constraint
E-tightness (ET)
Cant add another bit without violating the E constraint
If energy constraint violated

subtract the most costly bit
If energy less than max add
the bit that costs the least to add
LC - Rate adaptive algorithm
ET example (continue from EF example)
Start from efficient distribution that blows up the energy

constraint
N = 8 1 = 8
x
Margin
N x
8
= 1.8 dB
5.32
Levin-Campello Rate Adaptive algorithm
Choose any bit distribution

Make it efficient using (EF algorithm)
Make it energy tight using (ET algorithm)
Dynamic rate adaptation
Change the loading when channel changes

LC is a natural candidate
Keep the ET bit distribution and perturb based on
channel changes
Bit is moved from channel n to m
Tightly coupled with channel and noise estimation
Will cover in later lectures

Channel partitioning
Divide the channel into a set of parallel,
independent channels
X2,k
2(t)
x
XN-1,k
XN,k
1(T-t)
1(t)
X1,k
.
.
.
2(T-t)
n(t)
x(t)
h(t) +
h*(T-t)
.
.
.
y(t)
N-1(T-t)
N(t)
N(-t)
t=T
Qk
(MIMO)
y1,k
y2,k
yN,k
n2,k
nN,k
x(t) =
y1,k
y2,k
yN-1,k
yN,k
(lower case y
indicates normalizing
||p|| factors included)
x(t)
h(t) * x(t) =
n1,k
X1,k
XN,k
t=T
t=T
N-1(t)
X2,k
t=T
N
n=1
N
k n=1
N
k n=1
xn,k . n(t - kT)
xn,k . n(t - kT) * h(t) =
xn . n(t)
xn,k . pn(t - kT)
k n=1
p
m(t) * p*n(-t)
qn,m (t) =
||pm|| . ||pn||
Qk = Q(kT)
Qk = Ik
Figure by MIT OpenCourseWare.
Generalized Nyquist criterion
No interference between symbols

No interference between sub-channels
Modal modulation
Transmission with eigen-functions
(-TH / 2, TH / 2)
r(t) = h(t) * h*(-t)

n . n (t) =
T/2
-T / 2 r(t - ) ()d n = 1,...,
[-(T - TH) / 2, (T + TH) / 2)]
X1,k
1(t)
1(T-t)
X2,k
2(t)
2(T-t)
.
.
.
x
XN-1,k
y(t) =
x(t)
h(t) +
h*(T-t)
y(t)
.
.
.
N-1(t)
N-1(T-t)
N(t)
N(-t)
XN,k
x(t) =
n(t)
N
n=1
N
t=T
t=T
t=T
t=T
y1,k
y2,k
Per-channel MAP is optimal

Problem is vector AWGN
yN-1,k
yN,k
k (t ) Channel eigenfunctions (channel dependent)
xnn(t)
y(t) =
~
xn . [r(t) * n (t)] + n(t)
N
n=1
(nxn) . n (t)
n=1
Convergence of multitone to modal modulation
Modal modulation is optimal for finite symbol time
Eigenfunctions hard to compute and channel dependent
Multitone converges to Modal modulation
As both N->inf and [-T/2, T/2]->(-inf,+inf)
Set of eigenvalues of any autocorrelation function is unique

This set determines the performance of MM through SNR
Eigen-functions are not unique
is also a valid eigen-function for inf symbol period
Corresponding eigenvalues are R(2n/T)
No ISI on any tone since symbol period is infinite
Each tone is AWGN channel
SBS detector is MAP optimal
If channel is periodic (does not exist in practice)
are then eigen-functions even on finite [-T/2,T/2]

Can use extra bandwidth in the design to make the channel look periodic
10
Finite symbol duration effects
ISI corrupts the neighboring symbol
Need to leave the guardband
Can try to create a more sinc-looking symbols in

time by filtering the tones in frequency domain
Need excess bandwidth for filter roll-off
11
Discrete time channel partitioning
Digital realization
X0
m0
X1
m1
1/T
x
+
XN-2
XN-1
D
A
C
1/T
n (t)
x(t)
(t)
(LPF)
mN-2
*(-t)
h (t)
y(t) (LPF)
A
D
C
y = Px + n
f1*
Y1
*
fN-2
YN-2
*
fN-1
yN-1
y0
Y0
mN-1
yN-2
f0*
p0
p1
py
p0
py-1
py
p0
p1
0
py
0
0
xN-1
x0
x-1
YN-1
nN-1
nN-1
n0
x-y
12
Vector coding
Creates a set of parallel channels by using SVD
singular-value decomposition of P
discrete transmitter waveforms

discrete matched filters
since F is unitary N is also AWGN
13
VC example
1+0.9D-1, N=8
Need one sample guardband
Etot=(N+1)*E_dim=(8+1)*1=9
SVD on
Gives singular values
Sub-channel SNRs
Waterfilling shows only 7 dimensions can be used
Sub-channel energies
SNRs are then
Total SNR
VC capacity
Would get 1.55bits/dim if N-> inf
14
DMT and OFDM
Discrete multitone (DMT) and Orthogonal

frequency Division Multiplexing (OFDM)
DMT used on slowly time-varying channels
OFDM uses same channel partitioning as DMT
Optimize bn and En per sub-channel

But uses same bn and En on all channels
Used on one-way broadcast channels
Forms of vector coding with added restrictions
In vector coding, M,F channel dependent

Make the channel circular and make M,F channel
independent - simplify hardware implementation
15
DMT/OFDM channel partitioning
Make channel look circular
Repeat the tail of the symbol at its beginning
CP
Add cyclic prefix to each transmitted symbol
DMT
xN ... xN 1 x0 ...
xN ... xN 1
VC
0...0
x0 ...
xN ... xN 1
SVD can be replaced by eigen-decomposition (spectral

factorization)
A discrete form of modal modulation
While SNRs are unique, many choices for M and F
16
IDFT and DFT as orthogonal transformations
DMT and OFDM use IDFT and DFT as eigen-vectors
IDFT
DFT
Proof
channel gain at tone n
qn is eigen-vector of P
17
DMT/OFDM implementation
X0
Y0
X1
.
.
.
IDFT
parallel
to serial
& cyclic
prefix
insert
1 = N+v
T
D
A
C
n(t)
x(t)
(t)
(LPF)
1 = N+v
h(t)
(-t)
y(t)
(LPF)
A
D
C
serial to
parallel
& cyclic
prefix
remover
Y1
DFT
XN-2
.
.
.
YN-2
XN-1
YN-1
Forces P to be cyclic matrix
XN-i = Xi if x(t)is real

Data rate penalty N /( N + )

18
One-tap frequency equalizer
Need to compensate for channel attenuation
To recover the original constellation distance

19
Rx/Tx Windowing
Rectangular window
Raised cosine 5%
Raised cosine 25%
Can overlap as long
as sum is constant
20
1+0.9D-1 DMT example
N=8
Waterfilling with Etot=8
Waste one unit on CP
lower than VC (8.1dB)
For N=16 quickly reaches max of 8.8dB
21
Complexity of DFE, VC, DMT
Example 1+0.9D-1 channel
MMSE-DFE, SNR=7.6 dB, 3ff, 1fb tap, 4 mac/sample
VC, N=8, SNR=8.1 dB, 7*8/9=6.2MAC/sample
DMT, N=8, SNR=7.6 dB, 8pt FFT/IFFT,

2.7MAC/sample
N=16, 3.8MAC/sample, SNR=8.8 dB

DFE needs 10FF taps, 1FB tap, SNR=8.4 dB, 11MAC/sample
22
Asymmetric digital subscriber line (ADSL)
video switch
12 Mbps
ADSL ADSL
Access ADSL
Mux
(DSLAM) ADSL
Split
POTS
(or ISDN)
Switch
1.5 Mbps
Split
ADSL
Telco CO
internet
138 kHz
Down
Down
frequency
1.1 MHz
psd
1/T=2.208 MHz (CP = 40 samples)

Each time domain symbol 2*256+40=552 samples
UP
Up
10 kHz
Symbol rate T=250us

Nd=256, 4.3125kHz wide
0 Hz
POTS
Hermitian symmetry creates real signal transmitted from 0-1.1MHz

First 2-3 tones near DC not used avoid interference with voice
Tone 256 also not used, 64 reserved for pilot
Nup=32, CP=5, each symbol 2*32+5=69 samples
Exactly 1/8 of downstream

23
802.11a Wireless LAN example
Up to 54Mbps symmetrically (<100m)

1-3 Tx power levels
Complex baseband
unlike ADSL which is real baseband
5.25 GHz
20 MHz
5.35 GHz
4 pilots
R (Mbps) constellation code rate
48 data tones
6
9
12
18
24
36
48
54
BPSK
BPSK
4QAM
4QAM
16QAM
16QAM
64QAM
64QAM
Broadcast channel cant optimize bit allocation
5.825 GHz
20 MHz
Symbol length = 80 samples, CP=16

Symbol rate 250kHz (T=4uS, T=50ns), CPguard=0.8us
not used
5.725 GHz
52 carriers total, each ~ 300 kHz
N=64 (-31 31) (so 128 dimensions)
5.15 GHz
800 mW
200 mW
40 mW
1/2
3/4
1/2
3/4
1/2
3/4
1/2
3/4
bn
bn
1/2
3/4
1
3/2
2
3/2
3
9/2
1/4
3/8
1/2
3/4
1
3/4
3/2
9/4
24
36
48
72
96
144
192
216
FCC demands flat spectrum so no energy-allocation

The only knob is data rate selection
24
High-level system view
Data link control layer (DLC)

Physical radio layer (PHY)
Guard
Viterbi
Demodulation
interval
decoder +
+
descrambler deinterleaver extraction +
FFT
Scrambler +
forward error
correction
(FEC) coder
Interleaver
+
modulation
IFFT +
guard
interval
insertion
A/D
D/A
Analog front-end
Analog
Digital
Synchronization
25
Transceiver architecture
DAC
Upconvert
AGC& ADC
LNA & Downconvert
Channel estimator
Remove prefix
Pilot insertion
Demapper
Windowing
Mapper
Deinterleaver
Synchronizer
Interleaver
Viterbi decoder
Cyclic prefix
Encoder
Descrambler
FFT/IFFT
Scrambler
[2]
26
Transmitter architecture detail
Data
MAC
Scrambler
Convolutional
Encoder
Signal
Mapper
Puncturer
IFFT
Interleaver
Cyclic
Extend
Preamble
Modulation
&
Upconvert
RF
D/A
Transmitter (Digital Components)
Transmitter (Analog Components)
27
Scrambling
Need to randomize incoming data

Enables a number of tracking algorithms in
the receiver
Provides flat spectrum in the given band

pseudo-random bit sequence (prbs) generator
What is the period of this pseudo-random sequence?
28
Interleaver
Protects the code from overload by burst errors

Block interleaver
Block size is the #of coded bits in OFDM symbol (NCBPS)

Two-step permutation
Adjacent coded bits mapped
Onto nonadjacent sub-carriers

Alternate between less and more significant bits in the
constellation avoid long runs of low reliability LSBs
29
Convolutional Encoder
Rate 1/2 convolutional encoder
Punctured to obtain 2/3 and 3/4 rate
Omit some of the coded bits

Punctured Coding (r = 3/4)
g0=1338
Source Data
Encoded Data
Input Data
Tb
Tb
Tb
Tb
Tb
Tb
A0 A1 A2 A3 A4 A5 A6 A7 A8
B0 B1 B2 B3 B4 B5 B6 B7 B8
Stolen Bit
Bit Stolen Data

A0 B0 A1 B2 A3 B3A4 B5 A6 B6 A7 B8
(sent/received data)
Bit Inserted Data
g1=1718
X0 X1 X2 X3 X4 X5 X6 X7 X8
Output Data A
A0 A1 A2 A3 A4 A5 A6 A7 A8
B0 B1 B2 B3 B4 B5 B6 B7 B8
Inserted Dummy Bit
Output Data B
Decoded Data
y0 y1 y2 y3 y4 y5 y6 y7 y8
64-state (constraint length K=7) code

Viterbi algorithm applied in the decoder
30
Signal mapper
BPSK, QPSK, 16-QAM, 64-QAM
Data divided into groups of (1,2,4,6) bits and mapped

to a constellation point (i.e. a complex number)
Gray-coded constellation mapings
Need the same average power for all mappings
Scale the output by KMOD

31
Pilot insertion and FFT/IFFT
Pilot insertion
Pilots BPSK, prbs modulated
FFT and IFFT shared
Just flip the Re and Im inputs
32
Spectral mask
Cannot use last 5 tones on each side

Does not use extra windowing
Power Spectral Density (dB)
Transmit Spectrum Mask

(not to scale)
-20 dBr
Typical Signal Spectrum

(an example)
-28 dBr
-40 dBr
-30
-20
-11 -9
fc
9 11
20
30
Frequency (MHz)
Transmit spectrum mask

33
Receiver architecture
I and Q from
analog
front-end
ADC
Remove
DC
of fset
FIR
FIR
Rx gain
to analog
front-end
Rotate
Frequency
lock
FFT
Channel
correct
Deinterleave
Viterbi
Rx data
to MAC
Channel
estimate
and
tracking
Autocorrelate
Signal
detect
and AGC
Symbol
timing
Pipeline
control
34
Pilot tracking and channel correction
OFDM packet structure

8 + 8 = 16 s
10 x 0.8 = 8 s
2 x 0.8 + 2 x 3.2 = 8.0 s
t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 GI2
T1
0.8 + 3.2 = 4.0 s 0.8 + 3.2 = 4.0 s 0.8 + 3.2 = 4.0 s
T2
GI SIGNAL GI Data 1 GI Data 2
Signal Detect,
Coarse Freq. Channel and Fine Frequency RATE
AGC, Diversity Offset Estimation
Offset Estimation
LENGTH
Selection
Timing Synchronize
SERVICE + DATA
DATA
to timing control
symbol
timing adjust
training symbol pilots

pilots
long2 training
data
angle adjust
magnitude
adjust
long1 training
from fft
pilot
tracking
core
fir
complex
inverse
pilot
magnitude
multiply
rotate
composite
channel correction
channel to de-interleaver
correction
multiply

35
Synchronizer
Processing
Data Path
crosscorrelator
NCO
Input
Data
Z-16
FFT
Output
Symbols
Z-49
IJ*
Moving
Average
JF(k)
IP
Plateau
Detector
Tracking Data Path
Moving
Average
(16)
Combine
Ig.1(.)
Jc(k)
36
Channel estimator
Equalizer
Sync. + FFT
X(i,k) =
Reference
Sign
Correction
Residual
Phase
Correction
Y(i,k)
H(i-D,k)
Soft
Demapper
Deinterleaver
Soft
Viterbi
X(i,k)
H(i-D,k)
Channel Estimator
Delay
Buffer
(D)
Estimation
Buffer
(Circular Buffer)
HREF (k)
Pilot Sign
Correction
Y(i-D,k)
Zero Forcing
Y(i-D,k)
H(i-D,k) =
X(i-D,k)
Mapper
Interleaver
Conv. Encoder
X(i-D,k)
37
First 802.11a chip
Image removed due to copyright restrictions.
[1]
38
References
[1] T.H. Meng, B. McFarland, D. Su and J. Thomson "Design and implementation of an all-CMOS 802.11a
wireless LAN chipset," Communications Magazine, IEEE vol. 41, no. 8 SN - 0163-6804, pp. 160-168, 2003
[2] M. Krstic, K. Maharatna, A. Troya, E. Grass and U. Jagdhold "Implementation of an IEEE 802.11a
compliant low-power baseband processor," Telecommunications in Modern Satellite, Cable and
Broadcasting Service, 2003. TELSIKS 2003. 6th International Conference on vol. 1, no. SN -, pp. 97-100
vol.1, 2003.
[3] J. Thomson, B. Baas, E.M. Cooper, J.M. Gilbert, G. Hsieh, P. Husted, A. Lokanathan, J.S. Kuskin, D.
McCracken, B. McFarland, T.H. Meng, D. Nakahira, S. Ng, M. Rattehalli, J.L. Smith, R. Subramanian, L.
Thon, Y.-H. Wang and R. Yu "An integrated 802.11a baseband and MAC processor," Solid-State Circuits
Conference, 2002. Digest of Technical Papers. ISSCC. 2002 IEEE International vol. 1, no. SN -, pp. 126
451 vol.1, 2002.
[4] E. Grass, K. Tittelbach-Helmrich, U. Jagdhold, A. Troya, G. Lippert, O. Kruger, J. Lehmann, K.
Maharatna, K.F. Dombrowski, N. Fiebig, R. Kraemer and P. Mahonen "On the single-chip implementation
of a Hiperlan/2 and IEEE 802.11a capable modem," Personal Communications, IEEE [see also IEEE
Wireless Communications] vol. 8, no. 6 SN - 1070-9916, pp. 48-57, 2001.
39

Practical Multitone Architectures: Vladimir Stojanović

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Practical Multitone Architectures: Vladimir Stojanović

Uploaded by

Copyright:

Available Formats

Practical multitone architectures

6.973 Communication System Design Spring 2006

Massachusetts Institute of Technology

Loading with discrete information units

Chows algorithm (rounding of waterfilling results)

Each additional bit placed on subchannel that requires

Efficiency of bit distribution

6.973 Communication System Design

Levin-Campello (LC) algorithm- components

Efficientizing (EF) finding energy-efficient bit

Always replace a bit distribution with a more efficient

6.973 Communication System Design

1+0.9D-1 channel (Pe=10-6, gap=8.8dB, PAM/QAM)

PAM and single-sideband

6.973 Communication System Design

Rate adaptive solution

Need also to satisfy the energy constraint

Cant add another bit without violating the E constraint

If energy constraint violated

6.973 Communication System Design

LC - Rate adaptive algorithm

ET example (continue from EF example)

Start from efficient distribution that blows up the energy

Levin-Campello Rate Adaptive algorithm

Choose any bit distribution

6.973 Communication System Design

Dynamic rate adaptation

Change the loading when channel changes

Bit is moved from channel n to m

Tightly coupled with channel and noise estimation

Will cover in later lectures

6.973 Communication System Design

Divide the channel into a set of parallel,

xn,k . n(t - kT)

xn,k . n(t - kT) * h(t) =

xn,k . pn(t - kT)

Figure by MIT OpenCourseWare.

Generalized Nyquist criterion

No interference between symbols

6.973 Communication System Design

Transmission with eigen-functions

r(t) = h(t) * h*(-t)

-T / 2 r(t - ) ()d n = 1,...,

[-(T - TH) / 2, (T + TH) / 2)]

Per-channel MAP is optimal

k (t ) Channel eigenfunctions (channel dependent)

6.973 Communication System Design

Convergence of multitone to modal modulation

Modal modulation is optimal for finite symbol time

Eigenfunctions hard to compute and channel dependent

Multitone converges to Modal modulation

As both N->inf and [-T/2, T/2]->(-inf,+inf)

Set of eigenvalues of any autocorrelation function is unique

If channel is periodic (does not exist in practice)

are then eigen-functions even on finite [-T/2,T/2]

6.973 Communication System Design

Finite symbol duration effects

ISI corrupts the neighboring symbol

Need to leave the guardband

Can try to create a more sinc-looking symbols in

Need excess bandwidth for filter roll-off

6.973 Communication System Design

Discrete time channel partitioning