Contents

3 Equalization 149
3.1 Intersymbol Interference and Receivers for Successive Message Transmission . . . . . . . . 151
3.1.1 Transmission of Successive Messages . . . . . . . . . . . . . . . . . . . . . . . . . . 151
3.1.2 Bandlimited Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
3.1.3 The ISI-Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
3.2 Basics of the Receiver-generated Equivalent AWGN . . . . . . . . . . . . . . . . . . . . . 164
3.2.1 Receiver Signal-to-Noise Ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
3.2.2 Receiver Biases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
3.2.3 The Matched-Filter Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
3.3 Nyquist’s Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
3.3.1 Vestigial Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
3.3.2 Raised Cosine Pulses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
3.3.3 Square-Root Splitting of the Nyquist Pulse . . . . . . . . . . . . . . . . . . . . . . 174
3.4 Linear Zero-Forcing Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
3.4.1 Performance Analysis of the ZFE . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
3.4.2 Noise Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
3.5 Minimum Mean-Square Error Linear Equalization . . . . . . . . . . . . . . . . . . . . . . 186
3.5.1 Optimization of the Linear Equalizer . . . . . . . . . . . . . . . . . . . . . . . . . . 186
3.5.2 Performance of the MMSE-LE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
3.5.3 Examples Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
3.5.4 Fractionally Spaced Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
3.6 Decision Feedback Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
3.6.1 Minimum-Mean-Square-Error Decision Feedback Equalizer (MMSE-DFE) . . . . . 198
3.6.2 Performance Analysis of the MMSE-DFE . . . . . . . . . . . . . . . . . . . . . . . 200
3.6.3 Zero-Forcing DFE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
3.6.4 Examples Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
3.7 Finite Length Equalizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
3.7.1 FIR MMSE-LE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
3.7.2 FIR ZFE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
3.7.3 example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
3.7.4 FIR MMSE-DFE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
3.7.5 An Alternative Approach to the DFE . . . . . . . . . . . . . . . . . . . . . . . . . 218
3.7.6 The Stanford DFE Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
3.7.7 Error Propagation in the DFE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
3.7.8 Look-Ahead . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
3.8 Precoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
3.8.1 The Tomlinson Precoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
3.8.2 Partial Response Channel Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
3.8.3 Classes of Partial Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
3.8.4 Simple Precoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
3.8.5 General Precoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
3.8.6 Quadrature PR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
147
3.9 Diversity Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
3.9.1 Multiple Received Signals and the RAKE . . . . . . . . . . . . . . . . . . . . . . . 247
3.9.2 Infinite-length MMSE Equalization Structures . . . . . . . . . . . . . . . . . . . . 249
3.9.3 Finite-length Multidimensional Equalizers . . . . . . . . . . . . . . . . . . . . . . . 251
3.9.4 DFE RAKE Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
3.9.5 Multichannel Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
Exercises - Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
A Useful Results in Linear Minimum Mean-Square Estimation 279
A.1 The Orthogonality Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
A.2 Spectral Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
A.2.1 Cholesky Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
B Equalization for Partial Response 286
B.1 Controlled ISI with the DFE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
B.1.1 ZF-DFE and the Optimum Sequence Detector . . . . . . . . . . . . . . . . . . . . 286
B.1.2 MMSE-DFE and sequence detection . . . . . . . . . . . . . . . . . . . . . . . . . . 288
B.2 Equalization with Fixed Partial Response B(D) . . . . . . . . . . . . . . . . . . . . . . . . 288
B.2.1 The Partial Response Linear Equalization Case . . . . . . . . . . . . . . . . . . . . 288
B.2.2 The Partial-Response Decision Feedback Equalizer . . . . . . . . . . . . . . . . . . 289
C The Matrix Inversion Lemma: 291
148
Chapter 3
Equalization
The main focus of Chapter 1 was a single use of the channel, signal set, and detector to transmit one
of M messages, commonly referred to as “one-shot” transmission. In practice, the transmission system
sends a sequence of messages, one after another, in which case the analysis of Chapter 1 applies only
if the channel is memoryless: that is, for one-shot analysis to apply, these successive transmissions
must not interfere with one another. In practice, successive transmissions do often interfere with one
another, especially as they are spaced more closely together to increase the data transmission rate.
The interference between successive transmissions is called intersymbol interference (ISI). ISI can
severely complicate the implementation of an optimum detector.
Figure 3.1 illustrates a receiver for detection of a succession of messages. The matched filter outputs
are processed by the receiver, which outputs samples, z
k
, that estimate the symbol transmitted at time
k, ˆ x
k
. Each receiver output sample is the input to the same one-shot detector that would be used on an
AWGN channel without ISI. This symbol-by-symbol (SBS) detector, while optimum for the AWGN
channel, will not be a maximum-likelihood estimator for the sequence of messages. Nonetheless, if the
receiver is well-designed, the combination of receiver and detector may work nearly as well as an optimum
detector with far less complexity. The objective of the receiver will be to improve the performance of
this simple SBS detector. More sophisticated and complex sequence detectors are the topics of Chapters
4, 5 and 9.
Equalization methods are used by communication engineers to mitigate the effects of the intersym-
bol interference. An equalizer is essentially the content of Figure 3.1’s receiver box. This chapter studies
both intersymbol interference and several equalization methods, which amount to different structures
for the receiver box. The methods presented in this chapter are not optimal for detection, but rather are
widely used sub-optimal cost-effective methods that reduce the ISI. These equalization methods try to
convert a bandlimited channel with ISI into one that appears memoryless, hopefully synthesizing a new
AWGN-like channel at the receiver output. The designer can then analyze the resulting memoryless,
equalized channel using the methods of Chapter 1. With an appropriate choice of transmit signals,
one of the methods in this chapter - the Decision Feedback Equalizer of Section 3.6 can be generalized
into a canonical receiver (effectively achieves the highest possible transmission rate even though not an
optimum receiver), which is discussed further in Chapter 5 and occurs only when special conditions are
met.
Section 3.1 models linear intersymbol interference between successive transmissions, thereby both
illustrating and measuring the problem. In practice, as shown by a simple example, distortion from
overlapping symbols can be unacceptable, suggesting that some corrective action must be taken. Section
3.1 also refines the concept of signal-to-noise ratio, which is the method used in this text to quantify
receiver performance. The SNR concept will be used consistently throughout the remainder of this text
as a quick and accurate means of quantifying transmission performance, as opposed to probability of
error, which can be more difficult to compute, especially for suboptimum designs. As Figure 3.1 shows,
the objective of the receiver will be to convert the channel into an equivalent AWGN at each time k,
independent of all other times k. An AWGN detector may then be applied to the derived channel, and
performance computed readily using the gap approximation or other known formulae of Chapters 1 and
149
Figure 3.1: The band-limited channel with receiver and SBS detector.
2 with the SNR of the derived AWGN channel. There may be loss of optimality in creating such an
equivalent AWGN, which will be measured by the SNR of the equivalent AWGN with respect to the best
value that might be expected otherwise for an optimum detector. Section 3.3 discusses some desired
types of channel responses that exhibit no intersymbol interference, specifically introducing the Nyquist
Criterion for a linear channel (equalized or otherwise) to be free of intersymbol interference. Section 3.4
illustrates the basic concept of equalization through the zero-forcing equalizer (ZFE), which is simple to
understand but often of limited effectiveness. The more widely used and higher performance, minimum
mean square error linear (MMSE-LE) and decision-feedback equalizers (MMSE-DFE) are discussed in
Sections 3.5 and 3.6. Section 3.7 discusses the design of finite-length equalizers. Section 3.8 discusses
precoding, a method for eliminating error propagation in decision feedback equalizers and the related
concept of partial-response channels and precoding. Section 3.9 generalizes the equalization concepts to
systems that have one input, but several outputs, such as wireless transmission systems with multiple
receive antennas called “diversity.”
The area of transmit optimization for scalar equalization, is yet a higher performance form of equal-
ization for any channel with linear ISI, as discussed in Chapters 4 and 5.
150
3.1 Intersymbol Interference and Receivers for Successive Mes-
sage Transmission
Intersymbol interference is a common practical impairment found in many transmission and storage
systems, including voiceband modems, digital subscriber loop data transmission, storage disks, digital
mobile radio channels, digital microwave channels, and even fiber-optic (where dispersion-limited) cables.
This section introduces a model for intersymbol interference. This section then continues and revisits
the equivalent AWGN of Figure 3.1 in view of various receiver corrective actions for ISI.
3.1.1 Transmission of Successive Messages
Most communication systems re-use the channel to transmit several messages in succession. From Section
1.1, the message transmissions are separated by T units in time, where T is called the symbol period,
and 1/T is called the symbol rate.
1
The data rate of Chapter 1 for a communication system that
sends one of M possible messages every T time units is
R

=
log
2
(M)
T
=
b
T
. (3.1)
To increase the data rate in a design, either b can be increased (which requires more signal energy to
maintain P
e
) or T can be decreased. Decreasing T narrows the time between message transmissions and
thus increases intersymbol interference on any band-limited channel.
The transmitted signal x(t) corresponding to K successive transmissions is
x(t) =
K−1

k=0
x
k
(t − kT) . (3.2)
Equation (3.2) slightly abuses previous notation in that the subscript k on x
k
(t −kT) refers to the index
associated with the k
th
successive transmission. The K successive transmissions could be considered an
aggregate or “block” symbol, x(t), conveying one of M
K
possible messages. The receiver could attempt
to implement MAP or ML detection for this new transmission system with M
K
messages. A Gram-
Schmidt decomposition on the set of M
K
signals would then be performed and an optimum detector
designed accordingly. Such an approach has complexity that grows exponentially (in proportion to M
K
)
with the block message length K. That is, the optimal detector might need M
K
matched filters, one for
each possible transmitted block symbol. As K →∞, the complexity can become too large for practical
implementation. Chapter 9 addresses such “sequence detectors” in detail, and it may be possible to
compute the `a posteriori probability function with less than exponentially growing complexity.
An alternative (suboptimal) receiver can detect each of the successive K messages independently.
Such detection is called symbol-by-symbol (SBS) detection. Figure 3.2 contrasts the SBS detector
with the block detector of Chapter 1. The bank of matched filters, presumably found by Gram-Schmitt
decomposition of the set of (noiseless) channel output waveforms (of which it can be shown K dimen-
sions are sufficient only if N = 1, complex or real), precedes a block detector that determines the
K-dimensional vector symbol transmitted. The complexity would become large or infinite as K becomes
large or infinite for the block detector. The lower system in Figure 3.2 has a single matched filter to the
channel, with output sampled K times, followed by a receiver and an SBS detector. The later system
has fixed (and lower) complexity per symbol/sample, but may not be optimum. Interference between
successive transmissions, or intersymbol interference (ISI), can degrade the performance of symbol-by-
symbol detection. This performance degradation increases as T decreases (or the symbol rate increases)
in most communication channels. The designer mathematically analyzes ISI by rewriting (3.2) as
x(t) =
K−1

k=0
N

n=1
x
kn
ϕ
n
(t −kT) , (3.3)
1
The symbol rate is sometimes also called the “baud rate,” although abuse of the term baud (by equating it with data
rate even when M = 2) has rendered the term archaic among communication engineers, and the term “baud” usually now
only appears in trade journals and advertisements.
151
Figure 3.2: Comparison of Block and SBS detectors for successive transmission of K messages.
where the transmissions x
k
(t) are decomposed using a common orthonormal basis set ¦ϕ
n
(t)¦. In (3.3),
ϕ
n
(t − kT) and ϕ
m
(t − lT) may be non-orthogonal when k ,= l. In some cases, translates of the basis
functions are orthogonal. For instance, in QAM, the two bandlimited basis functions
ϕ
1
(t) =
_
2
T
cos
_
mπt
T
_
sinc
_
t
T
_
(3.4)
ϕ
2
(t) = −
_
2
T
sin
_
mπt
T
_
sinc
_
t
T
_
, (3.5)
or from Chapter 2, the baseband equivalent
ϕ(t) =
1

T
sinc
_
t
T
_
. (3.6)
(with m a positive integer) are orthogonal for all integer-multiple-of-T time translations. In this case,
the successive transmissions, when sampled at time instants kT, are free of ISI, and transmission is
equivalent to a succession of “one-shot” uses of the channel. In this case symbol-by-symbol detection
is optimal, and the MAP detector for the entire block of messages is the same as a MAP detector used
separately for each of the K independent transmissions. Signal sets for data transmission are usually
designed to be orthogonal for any translation by an integer multiple of symbol periods. Most linear
AWGN channels, however, are more accurately modeled by a filtered AWGN channel as discussed in
Section 1.7 and Chapter 2. The filtering of the channel alters the basis functions so that at the channel
output the filtered basis functions are no longer orthogonal. The channel thus introduces ISI.
3.1.2 Bandlimited Channels
The bandlimited linear ISI channel, shown in Figure 3.3, is the same as the filtered AWGN channel
discussed in Section 1.7. This channel is used, however, for successive transmission of data symbols.
152
Figure 3.3: The bandlimited channel, and equivalent forms with pulse response.
153
The (noise-free) channel output, x
p
(t), in Figure 3.3 is given by
x
p
(t) =
K−1

k=0
N

n=1
x
kn
ϕ
n
(t −kT) ∗ h(t) (3.7)
=
K−1

k=0
N

n=1
x
kn
p
n
(t − kT) (3.8)
where p
n
(t)

= ϕ
n
(t) ∗ h(t). When h(t) ,= δ(t), the functions p
n
(t − kT) do not necessarily form an
orthonormal basis, nor are they even necessarily orthogonal. An optimum (MAP) detector would need
to search a signal set of size M
K
, which is often too complex for implementation as K gets large. When
N = 1 (or N = 2 with complex signals), there is only one pulse response p(t).
Equalization methods apply a processor, the “equalizer”, to the channel output to try to convert
¦p
n
(t − kT)¦ to an orthogonal set of functions. Symbol-by-symbol detection can then be used on the
equalized channel output. Further discussion of such equalization filters is deferred to Section 3.4. The
remainder of this chapter also presumes that the channel-input symbol sequence x
k
is independent and
identically distributed at each point in time. This presumption will be relaxed in later Chapters.
While a very general theory of ISI could be undertaken for any N, such a theory would unnecessarily
complicate the present development.
2
This chapter handles the ISI-equalizer case for N = 1. Using
the baseband-equivalent systems of Chapter 2, this chapter’s (Chapter 3’s) analysis will also apply to
quadrature modulated systems modeled as complex (equivalent to two-dimensional real) channels. In
this way, the developed theory of ISI and equalization will apply equally well to any one-dimensional
(e.g. PAM) or two-dimensional (e.g. QAM or hexagonal) constellation. This was the main motivation
for the introduction of bandpass analysis in Chapter 2.
The pulse response for the transmitter/channel fundamentally quantifies ISI:
Definition 3.1.1 (Pulse Response) The pulse response of a bandlimited channel is de-
fined by
p(t) = ϕ(t) ∗ h(t) . (3.9)
For the complex QAM case, p(t), ϕ(t), and h(t) can be complex time functions.
The one-dimensional noiseless channel output x
p
(t) is
x
p
(t) =
K−1

k=0
x
k
ϕ(t − kT) ∗ h(t) (3.10)
=
K−1

k=0
x
k
p(t −kT) . (3.11)
The signal in (3.11) is real for a one-dimensional system and complex for a baseband equivalent quadra-
ture modulated system. The pulse response energy |p|
2
is not necessarily equal to 1, and this text
introduces the normalized pulse response:
ϕ
p
(t)

=
p(t)
|p|
, (3.12)
where
|p|
2
=
_

−∞
p(t)p

(t)dt = ¸p(t), p(t)) . (3.13)
The subscript p on ϕ
p
(t) indicates that ϕ
p
(t) is a normalized version of p(t). Using (3.12), Equation
(3.11) becomes
x
p
(t) =
K−1

k=0
x
p,k
ϕ
p
(t − kT) , (3.14)
2
Chapter 4 considers multidimensional signals and intersymbol interference, while Section 3.7 considers diversity re-
ceivers that may have several observations a single or multiple channel output(s).
154
Figure 3.4: Illustration of intersymbol interference for p(t) =
1
1+t
4
with T = 1.
where
x
p,k

= x
k
|p| . (3.15)
x
p,k
absorbs the channel gain/attenuation |p| into the definition of the input symbol, and thus has
energy c
p
= E
_
[x
p,k
[
2
¸
= c
x
|p|
2
. While the functions ϕ
p
(t − kT) are normalized, they are not
necessarily orthogonal, so symbol-by-symbol detection is not necessarily optimal for the signal in (3.14).
EXAMPLE 3.1.1 (Intersymbol interference and the pulse response) As an exam-
ple of intersymbol interference, consider the pulse response p(t) =
1
1+t
4
and two successive
transmissions of opposite polarity (−1 followed by +1) through the corresponding channel.
Figure 3.4 illustrates the two isolated pulses with correct polarity and also the waveform
corresponding to the two transmissions separated by 1 unit in time. Clearly the peaks of the
pulses have been displaced in time and also significantly reduced in amplitude. Higher trans-
mission rates would force successive transmissions to be closer and closer together. Figure
3.5 illustrates the resultant sum of the two waveforms for spacings of 1 unit in time, .5 units
in time, and .1 units in time. Clearly, ISI has the effect of severely reducing pulse strength,
thereby reducing immunity to noise.
EXAMPLE 3.1.2 (Pulse Response Orthogonality - Modified Duobinary) A PAM
modulated signal using rectangular pulses is
ϕ(t) =
1

T
(u(t) − u(t −T)) . (3.16)
The channel introduces ISI, for example, according to
h(t) = δ(t) + δ(t − T). (3.17)
155
Figure 3.5: ISI with increasing symbol rate.
156
The resulting pulse response is
p(t) =
1

T
(u(t) −u(t − 2T)) (3.18)
and the normalized pulse response is
ϕ
p
(t) =
1

2T
(u(t) −u(t −2T)). (3.19)
The pulse-response translates ϕ
p
(t) and ϕ
p
(t −T) are not orthogonal, even though ϕ(t) and
ϕ(t − T) were originally orthonormal.
Noise Equivalent Pulse Response
Figure 3.6 models a channel with additive Gaussian noise that is not white, which often occurs in practice.
The power spectral density of the noise is
N0
2
S
n
(f). When S
n
(f) ,= 0 (noise is never exactly zero at
any frequency in practice), the noise psd has an invertible square root as in Section 1.7. The invertible
square-root can be realized as a filter in the beginning of a receiver. Since this filter is invertible, by
the reversibility theorem of Chapter 1, no information is lost. The designer can then construe this filter
as being pushed back into, and thus a part of, the channel as shown in the lower part of Figure 3.6.
The noise equivalent pulse response then has Fourier Transform P(f)/S
1/2
n
(f) for an equivalent filtered-
AWGN channel. The concept of noise equivalence allows an analysis for AWGN to be valid (using the
equivalent pulse response instead of the original pulse response). Then also “colored noise” is equivalent
in its effect to ISI, and furthermore the compensating equalizers that are developed later in this chapter
can also be very useful on channels that originally have no ISI, but that do have “colored noise.” An
AWGN channel with a notch in H(f) at some frequency is thus equivalent to a “flat channel” with
H(f) = 1, but with narrow-band Gaussian noise at the same frequency as the notch.
3.1.3 The ISI-Channel Model
A model for linear ISI channels is shown in Figure 3.7. In this model, x
k
is scaled by |p| to form x
p,k
so that c
xp
= c
x
|p|
2
. The additive noise is white Gaussian, although correlated Gaussian noise can
be included by transforming the correlated-Gaussian-noise channel into an equivalent white Gaussian
noise channel using the methods in the previous subsection and illustrated in Figure 3.6. The channel
output y
p
(t) is passed through a matched filter ϕ

p
(−t) to generate y(t). Then, y(t) is sampled at the
symbol rate and subsequently processed by a discrete time receiver. The following theorem illustrates
that there is no loss in performance that is incurred via the matched-filter/sampler combination.
Theorem 3.1.1 (ISI-Channel Model Sufficiency) The discrete-time signal samples y
k
=
y(kT) in Figure 3.7 are sufficient to represent the continuous-time ISI-model channel output
y(t), if 0 < |p| < ∞. (i.e., a receiver with minimum P
e
can be designed that uses only the
samples y
k
).
Sketch of Proof:
Define
ϕ
p,k
(t)

= ϕ
p
(t − kT) , (3.20)
where ¦ϕ
p,k
(t)¦
k∈(−∞,∞)
is a linearly independent set of functions. The set ¦ϕ
p,k
(t)¦
k∈(−∞,∞)
is related to a set of orthogonal basis functions ¦φ
p,k
(t)¦
k∈(−∞,∞)
by an invertible transfor-
mation Γ (use Gram-Schmidt an infinite number of times). The transformation and its
inverse are written
¦φ
p,k
(t)¦
k∈(−∞,∞)
= Γ(¦ϕ
p,k
(t)¦
k∈(−∞,∞)
) (3.21)
¦ϕ
p,k
(t)¦
k∈(−∞,∞)
= Γ
−1
(¦φ
p,k
(t)¦
k∈(−∞,∞)
) , (3.22)
where Γ is the invertible transformation. In Figure 3.8, the transformation outputs are
the filter samples y(kT). The infinite set of filters ¦φ

p,k
(−t)¦
k∈(−∞,∞)
followed by Γ
−1
is
157
Figure 3.6: White Noise Equivalent Channel.
Figure 3.7: The ISI-Channel model.
158
Figure 3.8: Equivalent diagram of ISI-channel model matched-filter/sampler
equivalent to an infinite set of matched filters to
_
ϕ

p,k
(−t)
_
k∈(−∞,∞)
. By (3.20) this last set
is equivalent to a single matched filter ϕ

p
(−t), whose output is sampled at t = kT to produce
y(kT). Since the set ¦φ
p,k
(t)¦
k∈(−∞,∞)
is orthonormal, the set of sampled filter outputs in
Figure 3.8 are sufficient to represent y
p
(t). Since Γ
−1
is invertible (inverse is Γ), then by the
theorem of reversibility in Chapter 1, the sampled matched filter output y(kT) is a sufficient
representation of the ISI-channel output y
p
(t). QED.
Referring to Figure 3.7,
y(t) =

k
|p| x
k
q(t −kT) + n
p
(t) ∗ ϕ

p
(−t) , (3.23)
where
q(t)

= ϕ
p
(t) ∗ ϕ

p
(−t) =
p(t) ∗ p

(−t)
|p|
2
. (3.24)
The deterministic autocorrelation function q(t) is Hermitian (q

(−t) = q(t)). Also, q(0) = 1, so the
symbol x
k
passes at time kT to the output with amplitude scaling |p|. The function q(t) can also exhibit
ISI, as illustrated in Figure 3.9. The plotted q(t) corresponds to q
k
= [−.1159 .2029 1 .2029 −.1159] or,
equivalently, to the channel p(t) =
_
1
T
(sinc(t/T) + .25 sinc((t −T)/T) −.125 sinc((t −2T)/T)), or
159
Figure 3.9: ISI in q(t)
P(D) =
1

T
(1 + .5D)(1 − .25D) (the notation P(D) is defined in Appendix A.2). (The values for q
k
can be confirmed by convolving p(t) with its time reverse, normalizing, and sampling.) For notational
brevity, let y
k

= y(kT), q
k

= q(kT), n
k

= n(kT) where n(t)

= n
p
(t) ∗ ϕ

p
(−t). Thus
y
k
= |p| x
k
. ¸¸ .
scaled input (desired)
+ n
k
.¸¸.
noise
+|p|

m=k
x
m
q
k−m
. ¸¸ .
ISI
. (3.25)
The output y
k
consists of the scaled input, noise and ISI. The scaled input is the desired information-
bearing signal. The ISI and noise are unwanted signals that act to distort the information being trans-
mitted. The ISI represents a new distortion component not previously considered in the analysis of
Chapters 1 and 2 for a suboptimum SBS detector. This SBS detector is the same detector as in Chap-
ters 1 and 2, except used under the (false) assumption that the ISI is just additional AWGN. Such a
receiver can be decidedly suboptimum when the ISI is nonzero.
Using D-transform notation, (3.25) becomes
Y (D) = X(D) |p| Q(D) +N(D) (3.26)
where Y (D)

=


k=−∞
y
k
D
k
. If the receiver uses symbol-by-symbol detection on the sampled output
y
k
, then the noise sample n
k
of (one-dimensional) variance
N0
2
at the matched-filter output combines
with ISI from the sample times mT (m ,= k) in corrupting |p| x
k
.
There are two common measures of ISI distortion. The first is Peak Distortion, which only has
meaning for real-valued q(t):
3
Definition 3.1.2 (Peak Distortion Criterion) If [x
max
[ is the maximum value for [x
k
[,
then the peak distortion is:
T
p

= [x[
max
|p|

m=0
[q
m
[ . (3.27)
3
In the real case, the magnitudes correspond to actual values. However, for complex-valued terms, the ISI is charac-
terized by both its magnitude and phase. So, addition of the magnitudes of the symbols ignores the phase components,
which may significantly change the ISI term.
160
For q(t) in Figure 3.9 with x
max
= 3, T
p
= 3|p|(.1159+.2029+.2029+.1159) ≈ 3

1.078.6376 ≈ 1.99.
The peak distortion represents a worst-case loss in minimum distance between signal points in the
signal constellation for x
k
, or equivalently
P
e
≤ N
e
Q
_
_
|p|
d
min
2
− T
p
σ
_
_
, (3.28)
for symbol-by-symbol detection. Consider two matched filter outputs y
k
and y

k
that the receiver at-
tempts to distinguish by suboptimally using symbol-by-symbol detection. These outputs are generated
by two different sequences ¦x
k
¦ and ¦x

k
¦. Without loss of generality, assume y
k
> y

k
, and consider the
difference
y
k
−y

k
= |p|
_
_
(x
k
− x

k
) +

m=k
(x
m
− x

m
)q
k−m
_
_
+ ˜ n (3.29)
The summation term inside the brackets in (3.29) represents the change in distance between y
k
and y

k
caused by ISI. Without ISI the distance is
y
k
−y

k
≥ |p| d
min
, (3.30)
while with ISI the distance can decrease to
y
k
−y

k
≥ |p|
_
_
d
min
−2[x[
max

m=0
[q
m
[
_
_
. (3.31)
Implicitly, the distance interpretation in (3.31) assumes 2T
p
≤ |p|d
min
.
4
While peak distortion represents the worst-case ISI, this worst case might not occur very often in
practice. For instance, with an input alphabet size M = 4 and a q(t) that spans 15 symbol periods,
the probability of occurrence of the worst-case value (worst level occurring in all 14 ISI contributors) is
4
−14
= 3.710
−9
, well below typical channel P
e
’s in data transmission. Nevertheless, there may be other
ISI patterns of nearly just as bad interference that can also occur. Rather than separately compute each
possible combination’s reduction of minimum distance, its probability of occurrence, and the resulting
error probability, data transmission engineers more often use a measure of ISI called Mean-Square
Distortion (valid for 1 or 2 dimensions):
Definition 3.1.3 (Mean-Square Distortion) The Mean-Square Distortion is defined
by:
T
ms

= E
_
_
_
[

m=k
x
p,m
q
k−m
[
2
_
_
_
(3.32)
= c
x
|p|
2

m=0
[q
m
[
2
, (3.33)
where (3.33) is valid when the successive data symbols are independent and identically dis-
tributed with zero mean.
In the example of Figure 3.9, the mean-square distortion (with c
x
= 5) is T
ms
= 5|p|
2
(.1159
2
+.2029
2
+
.2029
2
+.1159
2
) ≈ 5(1.078).109 ≈ .588. The fact

.588 = .767 < 1.99 illustrates that T
ms
≤ T
2
p
. (The
proof of this fact is left as an exercise to the reader.)
The mean-square distortion criterion assumes (erroneously
5
) that T
ms
is the variance of an uncor-
related Gaussian noise that is added to n
k
. With this assumption, P
e
is approximated by
P
e
≈ N
e
Q
_
|p|d
min
2
_
σ
2
+
¯
T
ms
_
. (3.34)
4
On channels for which 2Dp ≥ pd
min
, the worst-case ISI occurs when |2Dp −pd
min
| is maximum.
5
This assumptionis only true when x
k
is Gaussian. In very well-designeddata transmissionsystems, x
k
is approximately
i.i.d. and Gaussian, see Chapter 6, so that this approximation of Gaussian ISI becomes accurate.
161
One way to visualize ISI is through the “eye diagram”, some examples of which are shown in Figures
3.10 and 3.11. The eye diagram is similar to what would be observed on an oscilloscope, when the trigger
is synchronized to the symbol rate. The eye diagram is produced by overlaying several successive symbol
intervals of the modulated and filtered continuous-time waveform (except Figures 3.10 and 3.11 do not
include noise). The Lorentzian pulse response p(t) = 1/(1 + (3t/T)
2
) is used in both plots. For binary
transmission on this channel, there is a significant opening in the eye in the center of the plot in Figure
3.10. With 4-level PAM transmission, the openings are much smaller, leading to less noise immunity.
The ISI causes the spread among the path traces; more ISI results in a narrower eye opening. Clearly
increasing M reduces the eye opening.
162
Figure 3.10: Binary eye diagram for a Lorentzian pulse response.
Figure 3.11: 4-Level eye diagram for a Lorentzian pulse response.
163
Figure 3.12: Use of possibly suboptimal receiver to approximate/create an equivalent AWGN.
3.2 Basics of the Receiver-generated Equivalent AWGN
Figure 3.12 focuses upon the receiver and specifically the device shown generally as R. When channels
have ISI, such a receiving device is inserted at the sampler output. The purpose of the receiver is to
attempt to convert the channel into an equivalent AWGN that is also shown below the dashed line. Such
an AWGN is not always exactly achieved, but nonetheless any deviation between the receiver output z
k
and the channel input symbol x
k
is viewed as additive white Gaussian noise. An SNR, as in Subsection
3.2.1 can be used then to analyze the performance of the symbol-by-symbol detector that follows the
receiver R. Usually, smaller deviation from the transmitted symbol means better performance, although
not exactly so as Subsection 3.2.2 discusses. Subsection 3.2.3 finishes this section with a discussion of
the highest possible SNR that a designer could expect for any filtered AWGN channel, the so-called
“matched-filter-bound” SNR, SNR
MFB
. This section shall not be specific as to the content of the box
shown as R, but later sections will allow both linear and slightly nonlinear structures that may often be
good choices because their performance can be close to SNR
MFB
.
3.2.1 Receiver Signal-to-Noise Ratio
Definition 3.2.1 (Receiver SNR) The receiver SNR, SNR
R
for any receiver R with
(pre-decision) output z
k
, and decision regions based on x
k
(see Figure 3.12) is
SNR
R

=
c
x
E[e
k
[
2
, (3.35)
where e
k

= x
k
−z
k
is the receiver error. The denominator of (3.35) is the mean-square
error MSE=E[e
k
[
2
. When E[z
k
[x
k
] = x
k
, the receiver is unbiased (otherwise biased) with
respect to the decision regions for x
k
.
The concept of a receiver SNR facilitates evaluation of the performance of data transmission systems
with various compensation methods (i.e. equalizers) for ISI. Use of SNR as a performance measure
builds upon the simplifications of considering mean-square distortion, that is both noise and ISI are
jointly considered in a single measure. The two right-most terms in (3.25) have normalized mean-square
value σ
2
+
¯
T
ms
. The SNR for the matched filter output y
k
in Figure 3.12 is the ratio of channel output
sample energy
¯
c
x
|p|
2
to the mean-square distortion σ
2
+
¯
T
ms
. This SNR is often directly related to
probability of error and is a function of both the receiver and the decision regions for the SBS detector.
This text uses SNR consistently, replacing probability of error as a measure of comparative performance.
SNR is easier to compute than P
e
, independent of M at constant
¯
c
x
, and a generally good measure
164
Figure 3.13: Receiver SNR concept.
of performance: higher SNR means lower probability of error. The probability of error is difficult to
compute exactly because the distribution of the ISI-plus-noise is not known or is difficult to compute.
The SNR is easier to compute and this text assumes that the insertion of the appropriately scaled SNR
(see Chapter 1 - Sections 1.4 - 1.6)) into the argument of the Q-function approximates the probability of
error for the suboptimum SBS detector. Even when this insertion into the Q-function is not sufficiently
accurate, comparison of SNR’s for different receivers usually relates which receiver is better.
3.2.2 Receiver Biases
Figure 3.13 illustrates a receiver that somehow has tried to reduce the combination of ISI and noise.
Any time-invariant receiver’s output samples, z
k
, satisfy
z
k
= α (x
k
+ u
k
) (3.36)
where α is some positive scale factor that may have been introduced by the receiver and u
k
is an
uncorrelated distortion
u
k
=

m=0
r
m
x
k−m
+

m
f
m
n
k−m
. (3.37)
The coefficients for residual intersymbol interference r
m
and the coefficients of the filtered noise f
k
will
depend on the receiver and generally determine the level of mean-square distortion. The uncorrelated
distortion has no remnant of the current symbol being decided by the SBS detector, so that E[u
k
/x
k
] = 0.
However, the receiver may have found that by scaling (reducing) the x
k
component in z
k
by α that
the SNR improves (small signal loss in exchange for larger uncorrelated distortion reduction). When
E[z
k
/x
k
] = α x
k
,= x
k
, the decision regions in the SBS detector are “biased.” Removal of the bias is
easily achieved by scaling by 1/α as also in Figure 3.13. If the distortion is assumed to be Gaussian
noise, as is the assumption with the SBS detector, then removal of bias by scaling by 1/α improves the
probability of error of such a detector as in Chapter 1. (Even when the noise is not Gaussian as is the
case with the ISI component, scaling the signal correctly improves the probability of error on the average
if the input constellation has zero mean.)
The following theorem relates the SNR’s of the unbiased and biased decision rules for any receiver
R:
Theorem 3.2.1 (Unconstrained and Unbiased Receivers) Given an unbiased receiver
R for a decision rule based on a signal constellation corresponding to x
k
, the maximum
unconstrained SNR corresponding to that same receiver with any biased decision rule is
SNR
R
= SNR
R,U
+ 1 , (3.38)
165
where SNR
R,U
is the SNR using the unbiased decision rule.
Proof: From Figure 3.13, the SNR after scaling is easily
SNR
R,U
=
¯
c
x
¯ σ
2
u
. (3.39)
The maximum SNR for the biased signal z
k
prior to the scaling occurs when α is chosen to maximize
the unconstrained SNR
SNR
R
=
c
x
[α[
2
σ
2
u
+ [1 − α[
2
c
x
. (3.40)
Allowing for complex α with phase θ and magnitude [α[, the SNR maximization over alpha is equivalent
to minimizing
1 −2[α[cos(θ) +[α[
2
(1 +
1
SNR
R,U
) . (3.41)
Clearly θ = 0 for a minimum and differentiating with respect to [α[ yields −2 + 2[α[(1 +
1
SNRR,U
) = 0
or α
opt
= 1/(1 + (SNR
R,U
)
−1
). Substitution of this value into the expression for SNR
R
finds
SNR
R
= SNR
R,U
+ 1 . (3.42)
Thus, a receiver R and a corresponding SBS detector that have zero bias will not correspond to a max-
imum SNR – the SNR can be improved by scaling (reducing) the receiver output by α
opt
. Conversely,
a receiver designed for maximum SNR can be altered slightly through simple output scaling by 1/α
opt
to a related receiver that has no bias and has SNR thereby reduced to SNR
R,U
= SNR
R
− 1. QED.
To illustrate the relationship of unbiased and biased receiver SNRs, suppose an ISI-free AWGN
channel has an SNR=10 with c
x
= 1 and σ
2
u
=
N0
2
= .1. Then, a receiver could scale the channel output
by α = 10/11. The resultant new error signal is e
k
= x
k
(1 −
10
11
) −
10
11
n
k
, which has MSE=E[[e
k
[
2
] =
1
121
+
100
121
(.1) =
1
11
, and SNR=11. Clearly, the biased SNR is equal to the unbiased SNR plus 1. The
scaling has done nothing to improve the system, and the appearance of an improved SNR is an artifact
of the SNR definition, which allows noise to be scaled down without taking into account the fact that
actual signal power after scaling has also been reduced. Removing the bias corresponds to using the
actual signal power, and the corresponding performance-characterizing SNR can always be found by
subtracting 1 from the biased SNR. A natural question is then “Why compute the biased SNR?” The
answer is that the biased receiver corresponds directly to minimizing the mean-square distortion, and
the SNR for the “MMSE” case will often be easier to compute. Figure 3.27 in Section 3.5 illustrates
the usual situation of removing a bias (and consequently reducing SNR, but not improving P
e
since the
SBS detector works best when there is no bias) from a receiver that minimizes mean-square distortion
(or error) to get an unbiased decision. The bias from a receiver that maximizes SNR by equivalently
minimizing mean-square error can then be removed by simple scaling and the resultant more accurate
SNR is thus found for the unbiased receiver by subtracting 1 from the more easily computed biased
receiver. This concept will be very useful in evaluating equalizer performance in later sections of this
chapter, and is formalized in Theorem 3.2.2 below.
Theorem 3.2.2 (Unbiased MMSE Receiver Theorem) Let ¹ be any allowed class of
receivers R producing outputs z
k
, and let R
opt
be the receiver that achieves the maximum
signal-to-noise ratio SNR(R
opt
) over all R ∈ ¹ with an unconstrained decision rule. Then
the receiver that achieves the maximum SNR with an unbiased decision rule is also R
opt
, and
max
R∈R
SNR
R,U
= SNR(R
opt
) −1. (3.43)
Proof. From Theorem 3.2.1, for any R ∈ ¹, the relation between the signal-to-noise ratios of unbiased
and unconstrained decision rules is SNR
R,U
= SNR
R
− 1, so
max
R∈R
[SNR
R,U
] = max
R∈R
[SNR
R
] − 1 = SNR
Ropt
−1 . (3.44)
166
QED.
This theorem implies that the optimum unbiased receiver and the optimum biased receiver settings
are identical except for any scaling to remove bias; only the SNR measures are different. For any SBS
detector, SNR
R,U
is the SNR that corresponds to best P
e
. The quantity SNR
R,U
+ 1 is artificially high
because of the bias inherent in the general SNR definition.
3.2.3 The Matched-Filter Bound
The Matched-Filter Bound (MFB), also called the “one-shot” bound, specifies an upper SNR limit
on the performance of data transmission systems with ISI.
Lemma 3.2.1 (Matched-Filter Bound SNR) The SNR
MFB
is the SNR that character-
izes the best achievable performance for a given pulse response p(t) and signal constellation
(on an AWGN channel) if the channel is used to transmit only one message. This SNR is
SNR
MFB
=
¯
c
x
|p|
2
N0
2
(3.45)
MFB denotes the square of the argument to the Q-function that arises in the equivalent
“one-shot” analysis of the channel.
Proof: Given a channel with pulse response p(t) and isolated input x
0
, the maximum output sample
of the matched filter is |p| x
0
. The normalized average energy of this sample is |p|
2 ¯
c
x
, while the
corresponding noise sample energy is
N0
2
, SNR
MFB
=
¯
E
x
p
2
N
0
2
. QED.
The probability of error, measured after the matched filter and prior to the symbol-by-symbol detector,
satisfies P
e
≥ N
e
Q(

MFB). When
¯
c
x
equals (d
2
min
/4)/κ, then MFB equals SNR
MFB
κ. In effect the
MFB forces no ISI by disallowing preceding or successive transmitted symbols. An optimum detector is
used for this “one-shot” case. The performance is tacitly a function of the transmitter basis functions,
implying performance is also a function of the symbol rate 1/T. No other (for the same input constel-
lation) receiver for continuous transmission could have better performance, if x
k
is an i.i.d. sequence,
since the sequence must incur some level of ISI. The possibility of correlating the input sequence ¦x
k
¦
to take advantage of the channel correlation will be considered in Chapters 4 and 5.
The following example illustrates computation of the MFB for several cases of practical interest:
EXAMPLE 3.2.1 (Binary PAM) For binary PAM,
x
p
(t) =

k
x
k
p(t − kT) , (3.46)
where x
k
= ±

c
x
. The minimum distance at the matched-filter output is |p| d
min
=
|p| d = 2 |p|

c
x
, so c
x
=
d
2
min
4
and κ = 1. Then,
MFB = SNR
MFB
. (3.47)
Thus for a binary PAM channel, the MFB (in dB) is just the “channel-output” SNR,
SNR
MFB
. If the transmitter symbols x
k
are equal to ±1 (c
x
= 1), then
MFB =
|p|
2
σ
2
, (3.48)
where, again, σ
2
=
N0
2
. The binary-PAM P
e
is then bounded by
P
e
≥ Q(
_
SNR
MFB
) . (3.49)
167
EXAMPLE 3.2.2 (M-ary PAM) For M-ary PAM, x
k
= ±
d
2
, ±
3d
2
, ..., ±
(M−1)d
2
and
d
2
=
_
3c
x
M
2
−1
, (3.50)
so κ = 3/(M
2
− 1). Thus,
MFB =
3
M
2
−1
SNR
MFB
, (3.51)
for M ≥ 2. If the transmitter symbols x
k
are equal to ±1, ±3, , ..., ±(M −1), then
MFB =
|p|
2
σ
2
. (3.52)
Equation (3.52) is the same result as (3.48), which should be expected since the minimum
distance is the same at the transmitter, and thus also at the channel output, for both (3.48)
and (3.52). The M’ary-PAM P
e
is then bounded by
P
e
≥ 2(1 −1/M) Q(
_
3 SNR
MFB
M
2
−1
) . (3.53)
EXAMPLE 3.2.3 (QPSK) For QPSK, x
k
= ±
d
2
±
d
2
, and d = 2
_
¯
c
x
, so κ = 1. Thus
MFB = SNR
MFB
. (3.54)
Thus, for a QPSK (or 4SQ QAM) channel, MFB (in dB) equals the channel output SNR. If
the transmitter symbols x
k
are ±1 ±, then
MFB =
|p|
2
σ
2
. (3.55)
The best QPSK P
e
is then approximated by
¯
P
e
≈ Q(
_
SNR
MFB
) . (3.56)
EXAMPLE 3.2.4 (M-ary QAM Square) For M-ary QAM, '¦x
k
¦ = ±
d
2
, ±
3d
2
, ..., ±
(

M−1)d
2
,
·¦x
k
¦ = ±
d
2
, ±
3d
2
, ..., ±
(

M−1)d
2
, (recall that ' and · denote real and imaginary parts,
respectively) and
d
2
=
¸
3
¯
c
x
M −1
, (3.57)
so κ = 3/(M −1). Thus
MFB =
3
M − 1
SNR
MFB
, (3.58)
for M ≥ 4. If the real and imaginary components of the transmitter symbols x
k
equal
±1, ±3, , ..., ±(

M − 1), then
MFB =
|p|
2
σ
2
. (3.59)
The best M’ary QAM P
e
is then approximated by
¯
P
e
≈ 2(1 − 1/

M) Q(
_
3 SNR
MFB
M − 1
) . (3.60)
168
In general for square QAM constellations,
MFB =
3
4
¯
b
−1
SNR
MFB
. (3.61)
For the QAM Cross constellations,
MFB =
2
¯
E
x
p
2
31
48
M−
2
3
σ
2
=
96
31 4
¯
b
− 32
SNR
MFB
. (3.62)
For the suboptimum receivers to come in later sections, SNR
U
≤ SNR
MFB
. As SNR
U
→SNR
MFB
,
then the receiver is approaching the bound on performance. It is not always possible to design a receiver
that attains SNR
MFB
, even with infinite complexity, unless one allows co-design of the input symbols
x
k
in a channel-dependent way (see Chapters 4 and 5). The loss with respect to matched filter bound
will be determined for any receiver by SNR
u
/SNR
MFB
≤ 1, in effect determining a loss in signal power
because successive transmissions interfere with one another – it may well be that the loss in signal power
is an acceptable exchange for a higher rate of transmission.
169
3.3 Nyquist’s Criterion
Nyquist’s Criterion specifies the conditions on q(t) = ϕ
p
(t) ∗ ϕ

p
(−t) for an ISI-free channel on which a
symbol-by-symbol detector is optimal. This section first reviews some fundamental relationships between
q(t) and its samples q
k
= q(kT) in the frequency domain.
q(kT) =
1

_

−∞
Q(ω)e
ωkT
dω (3.63)
=
1

n=−∞
_
(2n+1)π
T
(2n−1)π
T
Q(ω)e
ωkT
dω (3.64)
=
1

n=−∞
_ π
T

π
T
Q(ω +
2πn
T
)e
(ω+
2πn
T
)kT
dω (3.65)
=
1

_ π
T

π
T
Q
eq
(ω)e
ωkT
dω , (3.66)
where Q
eq
(ω), the equivalent frequency response becomes
Q
eq
(ω)

=

n=−∞
Q(ω +
2πn
T
) . (3.67)
The function Q
eq
(ω) is periodic in ω with period

T
. This function is also known as the folded or aliased
spectrum of Q(ω) because the sampling process causes the frequency response outside of the fundamental
interval (−
π
T
,
π
T
) to be added (i.e. “folded in”). Writing the Fourier Transform of the sequence q
k
as
Q(e
−ωT
) =


k=−∞
q
k
e
−ωkT
leads to
1
T
Q
eq
(ω) = Q(e
−ωT
)

=

k=−∞
q
k
e
−ωkT
, (3.68)
a well-known relation between the discrete-time and continuous-time representations of any waveform
in digital signal processing.
It is now straightforward to specify Nyquist’s Criterion:
Theorem 3.3.1 (Nyquist’s Criterion) A channel specified by pulse response p(t) (and
resulting in q(t) =
p(t)∗p

(−t)
p
2
) is ISI-free if and only if
Q(e
−ωT
) =
1
T

n=−∞
Q(ω +
2πn
T
) = 1 . (3.69)
Proof:
By definition the channel is ISI-free if and only if q
k
= 0 for all k ,= 0 (recall q
0
= 1 by
definition). The proof follows directly by substitution of q
k
= δ
k
into (3.68). QED.
Functions that satisfy (3.69) are called “Nyquist pulses.” One function that satisfies Nyquist’s
Criterion is
q(t) = sinc
_
t
T
_
, (3.70)
which corresponds to normalized pulse response
ϕ
p
(t) =
1

T
sinc
_
t
T
_
. (3.71)
The function q(kT) = sinc(k) = δ
k
satisfies the ISI-free condition. One feature of sinc(t/T) is that it has
minimum bandwidth for no ISI. No other function has this same minimum bandwidth and also satisfies
170
Figure 3.14: The sinc(t/T) function.
the Nyquist Criterion. (Proof is left as an exercise to the reader.) The sinc(t/T) function is plotted in
Figure 3.14 for −20T ≤ t ≤ 20T.
The frequency
1
2T
(1/2 the symbol rate) is often construed as the maximum frequency of a sampled
signal that can be represented by samples at the sampling rate. In terms of positive-frequencies, 1/2T
represents a minimum bandwidth necessary to satisfy the Nyquist Criterion, and thus has a special name
in data transmission:
Definition 3.3.1 (Nyquist Frequency) The frequency ω =
π
T
or f =
1
2T
is called the
Nyquist frequency.
6
3.3.1 Vestigial Symmetry
In addition to the sinc(t/T) Nyquist pulse, data-transmission engineers use responses with up to twice
the minimum bandwidth. For all these pulses, Q(ω) = 0 for [ω[ >

T
. These wider bandwidth responses
will provide more immunity to timing errors in sampling as follows: The sinc(t/T) function decays in
amplitude only linearly with time. Thus, any sampling-phase error in the sampling process of Figure 3.7
introduces residual ISI with amplitude that only decays linearly with time. In fact for q(t) = sinc(t/T),
the ISI term

k=0
q(τ +kT) with a sampling timing error of τ ,= 0 is not absolutely summable, resulting
in infinite peak distortion. The envelope of the time domain response decays more rapidly if the frequency
response is smooth (i.e. continuously differentiable). To meet this smoothness condition and also satisfy
Nyquist’s Criterion, the response must occupy a larger than minimum bandwidth that is between 1/2T
and 1/T. A q(t) with higher bandwidth can exhibit significantly faster decay as [t[ increases, thus
reducing sensitivity to timing phase errors. Of course, any increase in bandwidth should be as small as
possible, while still meeting other system requirements. The percent excess bandwidth
7
, or percent
roll-off, is a measure of the extra bandwidth.
Definition 3.3.2 (Percent Excess Bandwidth) The percent excess bandwidth α is
determined from a strictly bandlimited Q(ω) by finding the highest frequency in Q(ω) for
6
This text distinguishes the Nyquist Frequency from the Nyquist Rate in sampling theory, where the latter is twice the
highest frequency of a signal to be sampled and is not the same as the Nyquist Frequency here.
7
The quantity alpha used here is not a bias factor, and similarly Q(·) is a measure of ISI and not the integral of a
unit-variance Gaussian function – uses should be clear to reasonable readers who’ll understand that sometimes symbols
are re-used in obviously differenct contexts.
171
Figure 3.15: The sinc
2
(t/T) Function – time-domain
which there is nonzero energy transfer. That is
Q(ω) =
_
nonzero [ω[ ≤ (1 +α)
π
T
0 [ω[ > (1 +α)
π
T
.
. (3.72)
Thus, if α = .15 (a typical value), the pulse q(t) is said to have “15% excess bandwidth.” Usually, data
transmission systems have 0 ≤ α ≤ 1. In this case in equation (3.69), only the terms n = −1, 0, +1
contribute to the folded spectrum and the Nyquist Criterion becomes
1 = Q(e
−ωT
) (3.73)
=
1
T
_
Q
_
ω +

T
_
+Q(ω) +Q
_
ω −

T
__

π
T
≤ ω ≤
π
T
. (3.74)
Further, recalling that q(t) = ϕ
p
(t) ∗ ϕ

p
(−t) is Hermitian and has the properties of an autocorrelation
function, then Q(ω) is real and Q(ω) ≥ 0. For the region 0 ≤ ω ≤
π
T
(for real signals), (3.74) reduces to
1 = Q(e
−ωT
) (3.75)
=
1
T
_
Q(ω) +Q
_
ω −

T
__
. (3.76)
For complex signals, the negative frequency region (−
π
T
≤ ω ≤ 0) should also have
1 = Q(e
−ωT
) (3.77)
=
1
T
_
Q(ω) +Q
_
ω +

T
__
. (3.78)
Any Q(ω) satisfying (3.76) (and (3.78) in the complex case) is said to be vestigially symmetric
with respect to the Nyquist Frequency. An example of a vestigially symmetric response with 100% excess
bandwidth is q(t) = sinc
2
(t/T), which is shown in Figures 3.15 and 3.16.
3.3.2 Raised Cosine Pulses
The most widely used set of functions that satisfy the Nyquist Criterion are the raised-cosine pulse
shapes:
172
Figure 3.16: The sinc
2
(t/T) function – frequency-domain
Figure 3.17: Raised cosine pulse shapes – time-domain
173
Figure 3.18: Raised Cosine Pulse Shapes – frequency-domain
Definition 3.3.3 (Raised-Cosine Pulse Shapes) The raised cosine family of pulse shapes
(indexed by 0 ≤ α ≤ 1) is given by
q(t) = sinc
_
t
T
_

_
cos
_
απt
T
_
1 −
_
2αt
T
_
2
_
, (3.79)
and have Fourier Transforms
Q(ω) =
_
_
_
T [ω[ ≤
π
T
(1 −α)
T
2
_
1 −sin
_
T

([ω[ −
π
T
)

π
T
(1 −α) ≤ [ω[ ≤
π
T
(1 +α)
0
π
T
(1 +α) ≤ [ω[
. (3.80)
The raised cosine pulse shapes are shown in Figure 3.17 (time-domain) and Figure 3.18 (frequency-
domain) for α = 0, .5, and 1. When α = 0, the raised cosine reduces to a sinc function, which decays
asymptotically as
1
t
for t →∞, while for α ,= 0, the function decays as
1
t
3
for t →∞.
3.3.3 Square-Root Splitting of the Nyquist Pulse
The optimum ISI-free transmission system that this section has described has transmit-and-channel
filtering ϕ
p
(t) and receiver matched filter ϕ

p
(−t) such that q(t) = ϕ
p
(t) ∗ ϕ

p
(−t) or equivalently
Φ
p
(ω) = Q
1/2
(ω) (3.81)
so that the matched filter and transmit/channel filters are “square-roots” of the Nyquist pulse. When
h(t) = δ(t) or equivalently ϕ
p
(t) = ϕ(t) , the transmit filter (and the receiver matched filters) are square-
roots of a Nyquist pulse shape. Such square-root transmit filtering is often used even when ϕ
p
(t) ,= ϕ(t)
and the pulse response is the convolution of h(t) with the square-root filter.
The raised-cosine pulse shapes are then often used in square-root form, which is
_
Q(ω) =
_
¸
_
¸
_

T [ω[ ≤
π
T
(1 − α)
_
T
2
_
1 − sin
_
T

([ω[ −
π
T
)

1/2
π
T
(1 −α) ≤ [ω[ ≤
π
T
(1 + α)
0
π
T
(1 +α) ≤ [ω[
. (3.82)
174
This expression can be inverted to the time-domain via use of the identity sin
2
(θ) = .5(1 −cos(2θ)), as
in the exercises, to obtain
ϕ
p
(t) =

π

T

cos
_
[1 + α]
πt
T
_
+
T·sin([1−α]
πt
T
)
4αt
1 −
_
4αt
T
_
2
. (3.83)
175
3.4 Linear Zero-Forcing Equalization
This section examines the Zero-Forcing Equalizer (ZFE), which is the easiest type of equalizer to
analyze and understand, but has inferior performance to some other equalizers to be introduced in later
sections. The ISI model in Figure 3.7 is used to describe the zero-forcing version of the discrete-time
receiver. The ZFE is often a first non-trivial attempt at a receiver R in Figure 3.12 of Section 3.2.
The ZFE sets R equal to a linear time-invariant filter with discrete impulse response w
k
that acts
on y
k
to produce z
k
, which is an estimate of x
k
. Ideally, for symbol-by-symbol detection to be optimal,
q
k
= δ
k
by Nyquist’s Criterion. The equalizer tries to restore this Nyquist Pulse character to the
channel. In so doing, the ZFE ignores the noise and shapes the signal y
k
so that it is free of ISI. From
the ISI-channel model of Section 3.1 and Figure 3.7,
y
k
= |p| x
k
∗ q
k
+ n
k
. (3.84)
In the ZFE case, n
k
is initially viewed as being zero. The D-transform of y
k
is (See Appendix A.2)
Y (D) = |p| X(D) Q(D) . (3.85)
The ZFE output, z
k
, has Transform
Z(D) = W(D) Y (D) = W(D) Q(D) |p| X(D) , (3.86)
and will be free of ISI if Z(D) = X(D), leaving the ZFE filter characteristic:
Definition 3.4.1 (Zero-Forcing Equalizer) The ZFE transfer characteristic is
W(D) =
1
Q(D) |p|
. (3.87)
This discrete-time filter processes the discrete-time sequence corresponding to the matched-filter output.
The ZFE is so named because ISI is “forced to zero” at all sampling instants kT except k = 0. The
receiver uses symbol-by-symbol detection, based on decision regions for the constellation defined by x
k
,
on the output of the ZFE.
3.4.1 Performance Analysis of the ZFE
The variance of the noise, which is not zero in practice, at the output of the ZFE is important in
determining performance, even if ignored in the ZFE design of Equation (3.87). As this noise is produced
by a linear filter acting on the discrete Gaussian noise process n
k
, it is also Gaussian. The designer can
compute the discrete autocorrelation function (the bar denotes normalized to one dimension, so N = 2
for the complex QAM case) for the noise n
k
as
¯ r
nn,k
= E
_
n
l
n

l−k
¸
/N (3.88)
=
_

−∞
_

−∞
E
_
n
p
(t)n

p
(s)
¸
ϕ

p
(t − lT)ϕ
p
(s − (l − k)T)dtds (3.89)
=
_

−∞
_

−∞
^
0
2
δ(t −s)ϕ

p
(t −lT)ϕ
p
(s −(l −k)T)dtds (3.90)
=
^
0
2
_

−∞
ϕ

p
(t −lT)ϕ
p
(t −(l −k)T)dt (3.91)
=
^
0
2
_

−∞
ϕ

p
(u)ϕ
p
(u + kT)du (letting u = t −lT) (3.92)
=
^
0
2
q

−k
=
^
0
2
q
k
, (3.93)
or more simply
¯
R
nn
(D) =
^
0
2
Q(D) , (3.94)
176
where the analysis uses the substitution u = t −lT in going from (3.91) to (3.92), and assumes baseband
equivalent signals for the complex case. The complex baseband-equivalent noise, has (one-dimensional)
sample variance
N0
2
at the normalized matched filter output. The noise at the output of the ZFE,
n
ZFE,k
, has autocorrelation function
¯ r
ZFE,k
= ¯ r
nn,k
∗ w
k
∗ w

−k
. (3.95)
The D-Transform of ¯ r
ZFE,k
is then
¯
R
ZFE
(D) =
^
0
2
Q(D)
|p|
2
Q
2
(D)
=
N0
2
|p|
2
Q(D)
. (3.96)
The power spectral density of the noise samples n
ZFE,k
is then
¯
R
ZFE
(e
−ωT
). The (per-dimensional)
variance of the noise samples at the output of the ZFE is the (per-dimensional) mean-square error
between the desired x
k
and the ZFE output z
k
. Since
z
k
= x
k
+ n
ZFE,k
, (3.97)
then σ
2
ZFE
is computed as
σ
2
ZFE
=
T

_ π
T

π
T
¯
R
ZFE
(e
−ωT
)dω =
T

_ π
T

π
T
N0
2
|p|
2
Q(e
−ωT
)
dω =
N0
2
|p|
2
γ
ZFE
=
N0
2
|p|
w
0
, (3.98)
where
γ
ZFE
=
T

_ π
T

π
T
1
Q(e
−ωT
)
dω = w
0
|p| . (3.99)
The center tap of a linear equalizer is w
0
, or if the ZFE is explicity used, then w
ZFE,0
. This tap is
easily shown to always have the largest magnitude for a ZFE and thus spotted readily when plotting the
impulse response of any linear equalizer. From (3.97) and the fact that E[n
ZFE,k
] = 0, E[z
k
/x
k
] = x
k
,
so that the ZFE is unbiased for a detection rule based on x
k
. The SNR at the output of the ZFE is
SNR
ZFE
=
¯
c
x
σ
2
ZFE
= SNR
MFB

1
γ
ZFE
. (3.100)
Computation of d
min
over the signal set corresponding to x
k
leads to a relation between
¯
c
x
, d
min
, and
M, and the NNUB on error probability for a symbol-by-symbol detector at the ZFE output is (please,
do not confuse the Q function with the channel autocorrelation Q(D))
P
ZFE,e
≈ N
e
Q
_
d
min

ZFE
_
. (3.101)
Since symbol-by-symbol detection can never have performance that exceeds the MFB,
σ
2
≤ σ
2
ZFE
|p|
2
= σ
2
T

_ π
T

π
T

Q(e
−ωT
)
= σ
2
γ
ZFE
(3.102)
SNR
MFB
≥ SNR
ZFE
(3.103)
γ
ZFE
≥ 1 , (3.104)
with equality if and only if Q(D) = 1.
Finally, the ZFE equalization loss, 1/γ
ZFE
, defines the SNR reduction from the MFB:
Definition 3.4.2 (ZFE Equalization Loss) The ZFE Equalization Loss, γ
ZFE
in deci-
bels is
γ
ZFE

= 10 log
10
_
SNR
MFB
SNR
ZFE
_
= 10 log
10
|p|
2
σ
2
ZFE
σ
2
= 10 log
10
|p| w
ZFE,0
, (3.105)
and is a measure of the loss (in dB) with respect to the MFB for the ZFE. The sign of the
loss is often ignored since it is always a negative (or 0) number and only its magnitude is of
concern.
Equation (3.105) always thus provides a non-negative number.
177
Figure 3.19: Noise Enhancement
3.4.2 Noise Enhancement
The design of W(D) for the ZFE ignored the effects of noise. This oversight can lead to noise enhance-
ment, a σ
2
ZFE
that is unacceptably large, and the consequent poor performance.
Figure 3.19 hypothetically illustrates an example of a lowpass channel with a notch at the Nyquist
Frequency, that is, Q(e
−ωT
) is zero at ω =
π
T
. Since W(e
−ωT
) = 1/(|p| Q(e
−ωT
)), then any noise
energy near the Nyquist Frequency is enhanced (increased in power or energy), in this case so much
that σ
2
ZFE
→∞. Even when there is no channel notch at any frequency, σ
2
ZFE
can be finite, but large,
leading to unacceptable performance degradation.
In actual computation of examples, the author has found that the table in Table 3.1 is useful in
recalling basic digital signal processing equivalences related to scale factors between continuous time
convolution and discrete-time convolution.
The following example clearly illustrates the noise-enhancement effect:
EXAMPLE 3.4.1 (1 +.9D
−1
- ZFE) Suppose the pulse response of a (strictly bandlim-
ited) channel used with binary PAM is
P(ω) =
_ √
T
_
1 +.9e
ωT
_
[ω[ ≤
π
T
0 [ω[ >
π
T
. (3.106)
Then |p|
2
is computed as
|p|
2
=
1

_ π
T

π
T
[P(ω)[
2
dω (3.107)
=
1

_ π
T

π
T
T [1.81 + 1.8 cos(ωT)] dω (3.108)
=
T

1.81

T
= 1.81 = 1
2
+.9
2
. (3.109)
The magnitude of the Fourier transform of the pulse response appears in Figure 3.20. Thus,
with Φ
p
(ω) as the Fourier transform of ¦ϕ
p
(t)¦,
Φ
p
(ω) =
_ _
T
1.81
_
1 +.9e
ωT
_
[ω[ ≤
π
T
0 [ω[ >
π
T
. (3.110)
178
Table 3.1: Conversion factors of T.
179
Figure 3.20: Magnitude of Fourier transform of (sampled) pulse response.
Then, (using the bandlimited property of P(ω) in this example)
Q(e
−ωT
) =
1
T

p
(ω)[
2
(3.111)
=
1
1.81
[1 + .9e
ωT
[
2
(3.112)
=
1.81 + 1.8 cos(ωT)
1.81
(3.113)
Q(D) =
.9D
−1
+ 1.81 + .9D
1.81
. (3.114)
Then
W(D) =

1.81D
.9 + 1.81D+ .9D
2
. (3.115)
The magnitude of the Fourier transform of the ZFE response appears in Figure 3.21. The
time-domain sample response of the equalizer is shown in Figure 3.22.
Computation of σ
2
ZFE
allows performance evaluation,
8
σ
2
ZFE
=
T

_ π
T

π
T
N0
2
Q(e
−ωT
) |p|
2
dω (3.116)
=
T

_ π
T

π
T
(1.81/1.81)
N0
2
1.81 + 1.8 cos(ωT)
dω (3.117)
=
^
0
2
1

_
π
−π
1
1.81 + 1.8 cos u
du (3.118)
8
From a table of integrals
_
1
a+b cos u
du =
2

a
2
−b
2
Tan
−1
_

a
2
−b
2
tan(
u
2
)
a+b
_
.
180
Figure 3.21: ZFE magnitude for example.
Figure 3.22: ZFE time-domain response.
181
= (
^
0
2
)
2

_
2

1.81
2
− 1.8
2
Tan
−1

1.81
2
− 1.8
2
tan
u
2
1.81 + 1.8
¸
¸
¸
¸
¸
π
0
(3.119)
= (
^
0
2
)
4


1.81
2
− 1.8
2

π
2
(3.120)
= (
^
0
2
)
1

1.81
2
− 1.8
2
(3.121)
=
^
0
2
(5.26) . (3.122)
The ZFE-output SNR is
SNR
ZFE
=
¯
c
x
5.26
N0
2
. (3.123)
This SNR
ZFE
is also the argument of the Q-function for a binary detector at the ZFE output.
Assuming the two transmitted levels are x
k
= ±1, then we get
MFB =
|p|
2
σ
2
=
1.81
σ
2
, (3.124)
leaving
γ
ZFE
= 10 log
10
(1.81 5.26) ≈ 9.8dB (3.125)
for any value of the noise variance. This is very poor performance on this channel. No matter
what b is used, the loss is almost 10 dB from best performance. Thus, noise enhancement
can be a serious problem. Chapter 9 demonstrates a means by which to ensure that there
is no loss with respect to the matched filter bound for this channel. With alteration of the
signal constellation, Chapter 10 describes a means that ensures an error rate of 10
−5
on this
channel, with no information rate loss, which is far below the error rate achievable even when
the MFB is attained. thus, there are good solutions, but the simple concept of a ZFE is not
a good solution.
Another example illustrates the generalization of the above procedure when the pulse response is complex
(corresponding to a QAM channel):
EXAMPLE 3.4.2 (QAM: −.5D
−1
+ (1 +.25) −.5D Channel) Given a baseband equiv-
alent channel
p(t) =
1

T
_

1
2
sinc
_
t + T
T
_
+
_
1 +

4
_
sinc
_
t
T
_


2
sinc
_
t −T
T
__
, (3.126)
the discrete-time channel samples are
p
k
=
1

T
_

1
2
,
_
1 +

4
_
, −

2
_
. (3.127)
This channel has the transfer function of Figure 3.23. The pulse response norm (squared) is
|p|
2
=
T
T
(.25 + 1 +.0625 + .25) = 1.5625 . (3.128)
Then q
k
is given by
q
k
= q(kT) = T
_
ϕ
p,k
∗ ϕ

p,−k
_
=
1
1.5625
_


4
,
5
8
(−1 +) , 1.5625 , −
5
8
(1 +) ,

4
_
(3.129)
or
Q(D) =
1
1.5625
_


4
D
−2
+
5
8
(−1 +)D
−1
+ 1.5625 −
5
8
(1 +)D +

4
D
2
_
(3.130)
182
Figure 3.23: Fourier transform magnitude for pulse response for complex channel example.
Then, Q(D) factors into
Q(D) =
1
1.5625
_
(1 −.5D)
_
1 − .5D
−1
_ _
1 + .5D
−1
_
(1 − .5D)
_
, (3.131)
and
W
ZFE
(D) =
1
Q(D)|p|
(3.132)
or
W
ZFE
(D) =

1.5625
(−.25D
−2
+ .625(−1 +)D
−1
+ 1.5625 −.625(1 + )D +.25D
2
)
. (3.133)
The Fourier transform of the ZFE is in Figure 3.24. The real and imaginary parts of the
equalizer response in the time domain are shown in Figure 3.25 Then
σ
2
ZFE
=
T

_ π
T

π
T
N0
2
|p|
2
Q(e
−ωT
)
dω , (3.134)
or
σ
2
ZFE
=
^
0
2
w
0
|p|
, (3.135)
where w
0
is the zero (center) coefficient of the ZFE. This coefficient can be extracted from
W
ZFE
(D) =

1.5625
_
D
2
(−4)
¦(D + 2) (D− 2) (D −.5) (D + .5)¦
_
=

1.5625
_
A
D− 2
+
B
D + 2
+ terms not contributing to w
0
_
,
where A and B are coefficients in the partial-fraction expansion (the residuez and residue
commands in Matlab can be very useful here):
A =
4(−4)
(2 + 2)(2 + .5)(1.5)
= −1.5686 −.9412 = 1.8293

− 2.601 (3.136)
183
Figure 3.24: Fourier transform magnitude of the ZFE for the complex baseband channel example.
Figure 3.25: Time-domain equalizer coefficients for complex channel.
184
B =
−4(−4)
(−2 − .5)(−2 − 2)(−1.5)
= .9412 +1.5685 = 1.8293

1.030 . (3.137)
Finally
W
ZFE
(D) =

1.5625
_
A(−.5)
1 −.5D
+
B(−.5)
(1 −.5D)
+ terms
_
, (3.138)
and
w
0
=

1.5625[(−.5)(−1.5686 − .9412) −.5(.9412 + 1.5686)] (3.139)
=

1.5625(1.57) = 1.96 (3.140)
The ZFE loss can be shown to be
γ
ZFE
= 10 log
10
(w
0
|p|) = 3.9 dB , (3.141)
which is better than the last channel because the frequency spectrum is not as near zero
in this complex example as it was earlier on the PAM example. Nevertheless, considerably
better performance is also possible on this complex channel, but not with the ZFE.
To compute P
e
for 4QAM, the designer calculates
¯
P
e
≈ Q
_
_
SNR
MFB
−3.9 dB
_
. (3.142)
If SNR
MFB
= 10dB, then
¯
P
e
≈ 2.2 10
−2
. If SNR
MFB
= 17.4dB, then
¯
P
e
≈ 1.0 10
−6
. If
SNR
MFB
= 23.4dB for 16QAM, then
¯
P
e
≈ 2.0 10
−5
.
185
3.5 Minimum Mean-Square Error Linear Equalization
The Minimum Mean-Square Error Linear Equalizer (MMSE-LE) balances a reduction in ISI
with noise enhancement. The MMSE-LE always performs as well as, or better than, the ZFE and is of
the same complexity of implementation. Nevertheless, it is slightly more complicated to describe and
analyze than is the ZFE. The MMSE-LE uses a linear time-invariant filter w
k
for R, but the choice of
filter impulse response w
k
is different than the ZFE.
The MMSE-LE is a linear filter w
k
that acts on y
k
to form an output sequence z
k
that is the best
MMSE estimate of x
k
. That is, the filter w
k
minimizes the Mean Square Error (MSE):
Definition 3.5.1 (Mean Square Error (for the LE)) The LE error signal is given by
e
k
= x
k
−w
k
∗ y
k
= x
k
− z
k
. (3.143)
The Minimum Mean Square Error (MMSE) for the linear equalizer is defined by
σ
2
MMSE−LE

= min
wk
E
_
[x
k
− z
k
[
2
¸
. (3.144)
The MSE criteria for filter design does not ignore noise enhancement because the optimization of this
filter compromises between eliminating ISI and increasing noise power. Instead, the filter output is as
close as possible, in the Minimum MSE sense, to the data symbol x
k
.
3.5.1 Optimization of the Linear Equalizer
Using D-transforms,
E(D) = X(D) − W(D)Y (D) (3.145)
By the orthogonality principle in Appendix A, at any time k, the error sample e
k
must be uncorrelated
with any equalizer input signal y
m
. Succinctly,
E
_
E(D)Y

(D
−∗
)
¸
= 0 . (3.146)
Evaluating (3.146), using (3.145), yields
0 =
¯
R
xy
(D) − W(D)
¯
R
yy
(D) , (3.147)
where (N = 1 for PAM, N = 2 for Quadrature Modulation)
9
¯
R
xy
(D) = E
_
X(D)Y

(D
−∗
)
¸
/N = |p|Q(D)
¯
c
x
¯
R
yy
(D) = E
_
Y (D)Y

(D
−∗
)
¸
/N = |p|
2
Q
2
(D)
¯
c
x
+
^
0
2
Q(D) = Q(D)
_
|p|
2
Q(D)
¯
c
x
+
^
0
2
_
.
Then the MMSE-LE becomes
W(D) =
¯
R
xy
(D)
¯
R
yy
(D)
=
1
|p| (Q(D) + 1/SNR
MFB
)
. (3.148)
The MMSE-LE differs from the ZFE only in the additive positive term in the denominator of (3.148).
The transfer function for the equalizer W(e
−ωT
) is also real and positive for all finite signal-to-noise
ratios. This small positive term prevents the denominator from ever becoming zero, and thus makes
the MMSE-LE well defined even when the channel (or pulse response) is zero for some frequencies or
frequency bands. Also W(D) = W

(D
−∗
). Figure 3.26 repeats Figure 3.19 with addition of the MMSE-
LE transfer characteristic. The MMSE-LE transfer function has magnitude
¯
c
x

2
at ω =
π
T
, while the
ZFE becomes infinite at this same frequency. This MMSE-LE leads to better performance, as the next
subsection computes.
9
The expression Rxx(D)

= E
_
X(D)X

(D
−∗
)
¸
is used in a symbolic sense, since the terms of X(D)X

(D
−∗
) are of
the form

k
x
k
x

k−j
, so that we are implying the additional operation lim
K→∞
[1/(2K + 1)]

−K≤k≤K
on the sum in
such terms. This is permissable for stationary (and ergodic) discrete-time sequences.
186
Figure 3.26: Example of MMSE-LE versus ZFE.
3.5.2 Performance of the MMSE-LE
The MMSE is the time-0 coefficient of the error autocorrelation sequence
¯
R
ee
(D) = E
_
E(D)E

(D
−∗
)
¸
/N (3.149)
=
¯
c
x
−W

(D
−∗
)
¯
R
xy
(D) − W(D)
¯
R

xy
(D
−∗
) +W(D)
¯
R
yy
(D)W

(D
−∗
) (3.150)
=
¯
c
x
−W(D)
¯
R
yy
(D)W

(D
−∗
) (3.151)
=
¯
c
x

Q(D)
_
|p|
2
Q(D)
¯
c
x
+
N0
2
_
|p|
2
(Q(D) + 1/SNR
MFB
)
2
(3.152)
=
¯
c
x

¯
c
x
Q(D)
(Q(D) + 1/SNR
MFB
)
(3.153)
=
N0
2
|p|
2
(Q(D) + 1/SNR
MFB
)
(3.154)
The third equality follows from
W(D)
¯
R
yy
(D)W

(D
−∗
) = W(D)
¯
R
yy
(D)
¯
R
yy
(D)
−1
¯
R

xy
(D
−∗
)
= W(D)
¯
R

xy
(D
−∗
) ,
and that
_
W(D)
¯
R
yy
(D)W

(D
−∗
)
_

= W(D)
¯
R
yy
(D)W

(D
−∗
). The MMSE then becomes
σ
2
MMSE−LE
=
T

_ π
T

π
T
¯
R
ee
(e
−ωT
)dω =
T

_ π
T

π
T
N0
2

|p|
2
(Q(e
−ωT
) + 1/SNR
MFB
)
. (3.155)
By recognizing that W(e
−ωT
) is multiplied by the constant
N0
2
/|p| in (3.155), then
σ
2
MMSE−LE
= w
0
N0
2
|p|
. (3.156)
From comparison of (3.155) and (3.98), that
σ
2
MMSE−LE
≤ σ
2
ZFE
, (3.157)
with equality if and only if SNR
MFB
→∞. Furthermore, since the ZFE is unbiased, and SNR
MMSE−LE
=
SNR
MMSE−LE,U
+1 and SNR
MMSE−LE,U
are the maximum-SNR corresponding to unconstrained and
unbiased linear equalizers, respectively,
SNR
ZFE
≤ SNR
MMSE−LE,U
=
¯
c
x
σ
2
MMSE−LE
−1 ≤ SNR
MFB
. (3.158)
187
One confirms that the MMSE-LE is a biased receiver by writing the following expression for the
equalizer output
Z(D) = W(D)Y (D) (3.159)
=
1
|p| (Q(D) + 1/SNR
MFB
)
(Q(D)|p|X(D) + N(D)) (3.160)
= X(D) −
1/SNR
MFB
Q(D) + 1/SNR
MFB
X(D) +
N(D)
|p| (Q(D) + 1/SNR
MFB
)
, (3.161)
for which the x
k
-dependent residual ISI term contains a component
signal-basis term = −1/SNR
MFB
w
0
|p| x
k
(3.162)
= −1/SNR
MFB

σ
2
MMSE−LE
|p|
2
N0
2
x
k
(3.163)
= −
σ
2
MMSE−LE
¯
c
x
x
k
(3.164)
= −
1
SNR
MMSE−LE
x
k
. (3.165)
(3.166)
So z
k
=
_
1 −
1
SNRMMSE−LE
_
x
k
−e

k
where e

k
is the error for unbiased detection and R. The optimum
unbiased receiver with decision regions scaled by 1 −
1
SNRMMSE−LE
(see Section 3.2.1) has the signal
energy given by
_
1 −
1
SNR
MMSE−LE
_
2
¯
c
x
. (3.167)
A new error for the scaled decision regions is e

k
= (1 − 1/SNR
MMSE−LE
) x
k
−z
k
= e
k

1
SNRMMSE−LE
x
k
,
which is also the old error with the x
k
dependent term removed. Since e

k
and x
k
are then independent,
then
σ
2
e
= σ
2
MMSE−LE
= σ
2
e
+
_
1
SNR
MMSE−LE
_
2
¯
c
x
, (3.168)
leaving
σ
2
e
= σ
2
MMSE−LE

_
1
SNR
MMSE−LE
_
2
¯
c
x
=
SNR
2
MMSE−LE
σ
2
MMSE−LE

¯
c
x
SNR
2
MMSE−LE
=
¯
c
x
(SNR
MMSE−LE
−1)
SNR
2
MMSE−LE
.
(3.169)
The SNR for the unbiased MMSE-LE then becomes (taking the ratio of (3.167) to σ
2
e
)
SNR
MMSE−LE,U
=
(SNRMMSE−LE−1)
2
SNR
2
MMSE−LE
¯
c
x
¯
c
x
(SNRMMSE−LE−1)
SNR
2
MMSE−LE
= SNR
MMSE−LE
−1 , (3.170)
which corroborates the earlier result on the relation of the optimum biased and unbiased SNR’s for any
particular receiver structure (at least for the LE structure). The unbiased SNR is
SNR
MMSE−LE,U
=
¯
c
x
σ
2
MMSE−LE
−1 , (3.171)
which is the performance level that this text always uses because it corresponds to the best error
probability for an SBS detector, as was discussed earlier in Section 3.1. Figure 3.27 illustrates the
concept with the effect of scaling on a 4QAM signal shown explicitly. Again, MMSE receivers (of which
the MMSE-LE is one) reduce noise power at the expense of introducing a bias, the scaling up removes
188
Figure 3.27: Illustration of bias and its removal.
the bias, and ensures the detector achieves the best P
e
it can. Since the ZFE is also an unbiased receiver,
(3.158) must hold. The first inequality in (3.158) tends to equality if the ratio
¯
E
x
σ
2
→∞or if the channel
is free of ISI (Q(D) = 1). The second inequality tends to equality if
¯
E
x
σ
2
→0 or if the channel is free of
ISI (Q(D) = 1).
The unbiased MMSE-LE loss with respect to the MFB is
γ
MMSE−LE
=
SNR
MFB
SNR
MMSE−LE,U
=
_
|p|
2
σ
2
MMSE−LE
σ
2
_
_
¯
c
x
¯
c
x
− σ
2
MMSE−LE
_
. (3.172)
The
p
2
σ
2
MMSE−LE
σ
2
in (3.172) is the increase in noise variance of the MMSE-LE, while the term
¯
E
x
¯
E
x
−σ
2
MMSE−LE
term represents the loss in signal power at the equalizer output that accrues to lower noise enhancement.
The MMSE-LE also requires no additional complexity to implement and should always be used in
place of the ZFE when the receiver uses symbol-by-symbol detection on the equalizer output.
The error is not necessarily Gaussian in distribution. Nevertheless, engineers commonly make this
assumption in practice, with a good degree of accuracy, despite the potential non-Gaussian residual ISI
component in σ
2
MMSE−LE
. This text also follows this practice. Thus,
P
e
≈ N
e
Q
_
_
κSNR
MMSE−LE,U
_
, (3.173)
where κ depends on the relation of
¯
c
x
to d
min
for the particular constellation of interest, for instance
κ = 3/(M −1) for Square QAM. The reader may recall that the symbol Q is used in two separate ways
in these notes, for the Q-function, and for the transform of the matched-filter-pulse-response cascade.
The actual meaning should always be obvious in context.)
3.5.3 Examples Revisited
This section returns to the earlier ZFE examples to compute the improvement of the MMSE-LE on these
same channels.
EXAMPLE 3.5.1 (PAM - MMSE-LE) The pulse response of a channel used with bi-
nary PAM is again given by
P(ω) =
_ √
T
_
1 +.9e
ωT
_
[ω[ ≤
π
T
0 [ω[ >
π
T
. (3.174)
189
Figure 3.28: Comparison of equalizer frequency-domain responses for MMSE-LE and ZFE on baseband
example.
Let us suppose SNR
MFB
= 10dB (=
¯
c
x
|p|
2
/
N0
2
) and that c
x
= 1.
The equalizer is
W(D) =
1
|p|
_
.9
1.81
D
−1
+ 1.1 +
.9
1.81
D
_ . (3.175)
Figure 3.28 shows the frequency response of the equalizer for both the ZFE and MMSE-LE.
Clearly the MMSE-LE has a lower magnitude in its response.
The σ
2
MMSE−LE
is computed as
σ
2
MMSE−LE
=
T

_ π
T

π
T
N0
2
1.81 + 1.8 cos(ωT) + 1.81/10
dω (3.176)
=
^
0
2
1

1.991
2
− 1.8
2
(3.177)
=
^
0
2
(1.175) , (3.178)
which is considerably smaller than σ
2
ZFE
. The SNR for the MMSE-LE is
SNR
MMSE−LE,U
=
1 − 1.175(.181)
1.175(.181)
= 3.7 (5.7dB) . (3.179)
The loss with respect to the MFB is 10dB - 5.7dB = 4.3dB. This is 5.5dB better than the
ZFE (9.8dB - 4.3dB = 5.5dB), but still not good for this channel.
Figure 3.29 compares the frequency-domain responses of the equalized channel. Figure 3.30
compares the time-domain responses of the equalized channel. The MMSE-LE clearly does
not have an ISI-free response, but mean-square error/distortion is minimized and the MMSE-
LE has better performance than the ZFE.
The relatively small energy near the Nyquist Frequency in this example is the reason for
the poor performance of both linear equalizers (ZFE and MMSE-LE) in this example. To
190
Figure 3.29: Comparison of equalized frequency-domain responses for MMSE-LE and ZFE on baseband
example.
Figure 3.30: Comparison of equalized time-domain responses for MMSE-LE and ZFE on baseband
example.
191
Figure 3.31: Comparison of MMSE-LE and ZFE for complex channel example.
improve performance further, decision-assisted equalization is examined in Section 3.6 and
in Chapter 5 (or codes combined with equalization, as discussed in Chapters 10 and 11).
Another yet better method appears in Chapter 4.
The complex example also follows in a straightforward manner.
EXAMPLE 3.5.2 (QAM - MMSE-LE) Recall that the equivalent baseband pulse re-
sponse samples were given by
p
k
=
1

T
_

1
2
,
_
1 +

4
_
, −

2
_
. (3.180)
Again, SNR
MFB
= 10dB (=
¯
c
x
|p|
2
/
N0
2
),
¯
c
x
= 1, and thus
σ
2
MMSE−LE
= w
0
^
0
2
|p|
−1
. (3.181)
The equalizer is, noting that
N0
2
=
¯
E
x
p
2
SNRMFB
= 1.5625/10 = .15625,
W(D) =
1
(Q(D) + 1/SNR
MFB
) |p|
(3.182)
or
=

1.5625
−.25D
−2
+ .625(−1 + )D
−1
+ 1.5625(1 +.1) − .625(1 +)D
1
+ .25D
2
. (3.183)
Figure 3.31 compares the MMSE-LE and ZFE equalizer responses for this same channel. The
MMSE-LE high response values are considerably more limited than the ZFE values. The
roots of Q(D) + 1/SNR
MFB
, or of the denominator in (3.183), are
D = 2.217

−1.632
= −.1356 − 2.213 (3.184)
D = 2.217

.0612
= 2.213 +.1356 (3.185)
D = .451

−1.632
= −.0276 − .4502 (3.186)
D = .451

.0612
= .4502 +.0276 , (3.187)
192
and (3.183) becomes
W(D) =
_√
1.5625D
2
/(.25)
¸
(D − 2.22

−1.632
)(D − 2.22

.0612
)(D − .451

−1.632
)(D −.451

.0612
)
(3.188)
or
W(D) =
A
D −2.22

−1.632
+
B
D −2.22

.0612
+ ... , (3.189)
where the second expression (3.189) has ignored any terms in the partial fraction expansion
that will not contribute to w
0
. Then,
A =

1.5625(−4)(2.22

−1.632
)
2
(2.22

−1.632
− 2.22

.0612
)(2.22

−1.632
−.451

−1.632
)(2.22

−1.632
− .451

.0612
)
=

1.5625(−.805 + 1.20) = 1.0079 +1.5025 (3.190)
B =

1.5625(−4)(2.22

.0612
)
2
(2.22

.0612
− 2.22

−1.632
)(2.22

.0612
−.451

−1.632
)(2.22

.0612
−.451

.0612
)
=

1.5625(1.20 − .805) = −1.5025 −1.0079 . (3.191)
Then, from (3.189),
w
0
= A(−.451

1.632
) +B(−.451

−.0612
) =

1.5625(1.125) . (3.192)
Then
σ
2
MMSE−LE
= 1.125
^
0
2
= 1.125(.15625) = .1758 , (3.193)
SNR
MMSE−LE,U
=
[1 −1.125(.15625)]
1.125(.15625)
= 4.69 ( 6.7 dB ) , (3.194)
which is 3.3 dB below the matched filter bound SNR of 10dB, or equivalently,
γ
MMSE−LE
= 10 log
10
(10/4.69) = 10 log
10
(2.13) = 3.3 dB , (3.195)
but .6 dB better than the ZFE. It is also possible to do considerably better on this channel,
but structures not yet introduced will be required. Nevertheless, the MMSE-LE is one of the
most commonly used equalization structures in practice.
Figure 3.31 compares the equalized channel responses for both the MMSE-LE and ZFE.
While the MMSE-LE performs better, there is a significant deviation from the ideal flat
response. This deviation is good and gains .6 dB improvement. Also, because of biasing,
the MMSE-LE output is everywhere slightly lower than the ZFE output (which is unbiased).
This bias can be removed by scaling by 5.69/4.69.
3.5.4 Fractionally Spaced Equalization
To this point in the study of the discrete time equalizer, a matched filter ϕ

p
(−t) precedes the sampler,
as was shown in Figure 3.7. While this may be simple from an analytical viewpoint, there are several
practical problems with the use of the matched filter. First, ϕ

p
(−t) is a continuous time filter and
may be much more difficult to design accurately than an equivalent digital filter. Secondly, the precise
sampling frequency and especially its phase, must be known so that the signal can be sampled when it is
at maximum strength. Third, the channel pulse response may not be accurately known at the receiver,
and an adaptive equalizer (see Chapter 7), is instead used. It may be difficult to design an adaptive,
analog, matched filter.
For these reasons, sophisticated data transmission systems often replace the matched-filter/sampler/equalizer
system of Figure 3.7 with the Fractionally Spaced Equalizer (FSE) of Figure 3.32. Basically, the sampler
and the matched filter have been interchanged with respect to Figure 3.7. Also the sampling rate has
been increased by some rational number l (l > 1). The new sampling rate is chosen sufficiently high so
193
Figure 3.32: Fractionally spaced equalization.
as to be greater than twice the highest frequency of x
p
(t). Then, the matched filtering operation and the
equalization filtering are performed at rate l/T, and the cascade of the two filters is realized as a single
filter in practice. An anti-alias or noise-rejection filter always precedes the sampling device (usually an
analog-to-digital converter). This text assumes this filter is an ideal lowpass filter with gain

T transfer
over the frequency range −l
π
T
≤ ω ≤ l
π
T
. The variance of the noise samples at the filter output will then
be
N0
2
l per dimension.
In practical systems, the four most commonly found values for l are
4
3
, 2, 3, and 4. The major
drawback of the FSE is that to span the same interval in time (when implemented as an FIR filter, as is
typical in practice), it requires a factor of l more coefficients, leading to an increase in memory by a factor
of l. The FSE outputs may also appear to be computed l times more often in real time. However, only
_
1
l
_
th
of the output samples need be computed (that is, those at the symbol rate, as that is when we need
them to make a decision), so computation is approximately l times that of symbol-spaced equalization,
corresponding to l times as many coefficients to span the same time interval, or equivalently, to l times
as many input samples per symbol.
The FSE will digitally realize the cascade of the matched filter and the equalizer, eliminating the need
for an analog matched filter. (The entire filter for the FSE can also be easily implemented adaptively, see
Chapter 7.) The FSE can also exhibit a significant improvement in sensitivity to sampling-phase errors.
To investigate this improvement briefly, in the original symbol-spaced equalizers (ZFE or MMSE-LE)
may have the sampling-phase on the sampler in error by some small offset t
0
. Then, the sampler will
sample the matched filter output at times kT + t
0
, instead of times kT. Then,
y(kT +t
0
) =

m
x
m
|p| q(kT −mT + t
0
) +n
k
, (3.196)
which corresponds to q(t) →q(t +t
0
) or Q(ω) →Q(ω)e
−ωt0
. For the system with sampling offset,
Q(e
−ωT
) →
1
T

n
Q(ω −
2πn
T
)e
−(ω−
2πn
T
)t0
, (3.197)
which is no longer nonnegative real across the entire frequency band. In fact, it is now possible (for
certain nonzero timing offsets t
0
, and at frequencies just below the Nyquist Frequency) that the two
aliased frequency characteristics Q(ω)e
−ωt0
and Q(ω −

T
)e
−(ω−

T
)t0
add to approximately zero, thus
producing a notch within the critical frequency range (-
π
T
,
π
T
). Then, the best performance of the ensuing
symbol-spaced ZFE and/or MMSE-LE can be significantly reduced, because of noise enhancement. The
resulting noise enhancement can be a major problem in practice, and the loss in performance can be
several dB for reasonable timing offsets. When the anti-alias filter output is sampled at greater than
twice the highest frequency in x
p
(t), then information about the entire signal waveform is retained.
Equivalently, the FSE can synthesize, via its transfer characteristic, a phase adjustment (effectively
interpolating to the correct phase) so as to correct the timing offset, t
0
, in the sampling device. The
symbol-spaced equalizer cannot interpolate to the correct phase, because no interpolation is correctly
performed at the symbol rate. Equivalently, information has been lost about the signal by sampling at
a speed that is too low in the symbol-spaced equalizer without matched filter. This possible notch is an
194
example of information loss at some frequency; this loss cannot be recovered in a symbol-spaced equalizer
without a matched filter. In effect, the FSE equalizes before it aliases (aliasing does occur at the output
of the equalizer where it decimates by l for symbol-by-symbol detection), whereas the symbol-spaced
equalizer aliases before it equalizes; the former alternative is often the one of choice in practical system
implementation, if the extra memory and computation can be accommodated. Effectively, with an FSE,
the sampling device need only be locked to the symbol rate, but can otherwise provide any sampling
phase. The phase is tacitly corrected to the optimum phase inside the linear filter implementing the
FSE.
The sensitivity to sampling phase is channel dependent: In particular, there is usually significant
channel energy near the Nyquist Frequency in applications that exhibit a significant improvement of
the FSE with respect to the symbol-spaced equalizer. In channels with little energy near the Nyquist
Frequency, the FSE is avoided, because it provides little performance gain, and is significantly more
complex to implement (more parameters and higher sampling rate).
Infinite-Length FSE Settings To derive the settings for the FSE, this section assumes that l, the
oversampling factor, is an integer. The sampled output of the anti-alias filter can be decomposed into l
sampled-at-rate-1/T interleaved sequences with D-transforms Y
0
(D), Y
2
(D), ..., Y
l−1
(D), where Y
l
(D)
corresponds to the sample sequence y[kT − iT/l]. Then,
Y
i
(D) = P
i
(D) X(D) + N
i
(D) (3.198)
where P
i
(D) is the transform of the symbol-rate-spaced-i
th
-phase-of-p(t) sequence p[kT − (i − 1)T/l],
and similarly N
i
(D) is the transform of a symbol-rate sampled white noise sequence with autocorrelation
function R
nn
(D) = l
N0
2
per dimension, and these noise sequences are also independent of one another.
A column vector transform is
Y (D) =
_
¸
_
Y
0
(D)
.
.
.
Y
l−1
(D)
_
¸
_ = P(D)X(D) +N(D) . (3.199)
Also
P(D)

=
_
¸
_
P
0
(D)
.
.
.
P
l−1
(D)
_
¸
_ and N(D) =
_
¸
_
N
0
(D)
.
.
.
N
l−1
(D)
_
¸
_ . (3.200)
By considering the FSE output at sampling rate 1/T, the interleaved coefficients of the FSE can also be
written in a row vector W(D) = [W
1
(D) , ..., W
l
(D)]. Thus the FSE output is
Z(D) = W(D)Y (D) . (3.201)
Again, the orthogonality condition says that E(D) = X(D) − Z(D) should be orthogonal to Y (D),
which can be written in vector form as
E
_
E(D)Y

(D
−∗
)
¸
= R
xY
(D) − W(D)R
Y Y
(D) = 0 , (3.202)
where
R
xY
(D)

= E
_
X(D)Y

(D
−∗
)
¸
=
¯
c
x
P

(D
−∗
) (3.203)
R
Y Y
(D)

= E
_
Y (D)Y

(D
−∗
)
¸
=
¯
c
x
P(D)P

(D
−∗
) +l
^
0
2
I . (3.204)
MMSE-FSE filter setting is then
W(D) = R
xY
(D)R
−1
Y Y
(D) = P

(D
−∗
)
_
P(D)P

(D
−∗
) +l/SNR
¸
−1
. (3.205)
The corresponding error sequence has autocorrelation function
¯
R
ee
(D) =
¯
c
x
−R
xY
(D)R
−1
Y Y
(D)R
Y x
(D) =
l
N0
2
P

(D
−∗
)P(D) +l/SNR
. (3.206)
195
Figure 3.33: Passband equalization, direct synthesis of analytic-equivalent equalizer.
The MMSE is then computed as
MMSE
MMSE−FSE
=
T

_ π
T

π
T
l
N0
2

|P(e
−ωT
)|
2
+ l/SNR
= MMSE
MMSE−LE
, (3.207)
where |P(e
−ωT
)|
2
=

l
i=1
[P
i
(e
−ωT
)[
2
. The SNR’s, biased and unbiased, are also then exactly the
same as given for the MMSE-LE, as long as the sampling rate exceeds twice the highest frequency of
P(f).
The reader is cautioned against letting
N0
2
→ 0 to get the ZF-FSE. This is because the matrix
R
Y Y
(D) will often be singular when this occurs. To avoid problems with singularity, an appropriate
pseudoinverse, which zeros the FSE characteristic when P(ω) = 0 is recommended.
Passband Equalization
This chapter so far assumed for the complex baseband equalizer that the passband signal has been de-
modulated according to the methods described in Chapter 2 before filtering and sampling. An alternative
method, sometimes used in practice when ω
c
is not too high (or intermediate-frequency up/down con-
version is used), is to commute the demodulation and filtering functions. Sometimes this type of system
is also called “direct conversion,” meaning that no carrier demodulation occurs prior to sampling. This
commuting can be done with either symbol-spaced or fractionally spaced equalizers. The main reason
for the interchange of filtering and demodulation is the recovery of the carrier frequency in practice.
Postponing the carrier demodulation to the equalizer output can lead to significant improvement in the
tolerance of the system to any error in estimating the carrier frequency, as Chapter 6 investigates. In the
present (perfect carrier phase lock) development, passband equalization and baseband equalization are
exactly equivalent and the settings for the passband equalizer are identical to those of the corresponding
baseband equalizer, other than a translation in frequency by ω
c
radians/sec.
Passband equalization is best suited to CAP implementations (See Section 4 of Chapter 2) where the
complex channel is the analytic equivalent. The complex equalizer then acts on the analytic equivalent
channel and signals to eliminate ISI in a MMSE sense. When used with CAP, the final rotation shown
in Figure 3.33 is not necessary – such a rotation is only necessary with a QAM transmit signal.
The passband equalizer is illustrated in Figure 3.33. The phase splitting is the forming of the analytic
equivalent with the use of the Hilbert Transform, as discussed in Chapter 2. The matched filtering and
equalization are then performed on the passband signal with digital demodulation to baseband deferred
to the filter output. The filter w
k
can be a ZFE, a MMSE-LE, or any other desired setting (see Chapter
4 and Sections 3.5 and 3.6). In practice, the equalizer can again be realized as a fractionally spaced
equalizer by interchanging the positions of the matched filter and sampler, increasing the sampling rate,
and absorbing the matched filtering operation into the filter w
k
, which would then be identical to the
(modulated) baseband-equivalent FSE with output sample rate decimated appropriately to produce
outputs only at the symbol sampling instants.
Often in practice, the cross-coupled nature of the complex equalizer W(D) is avoided by sampling the
FSE at a rate that exceeds twice the highest frequency of the passband signal y(t) prior to demodulation.
(If intermediate frequency (IF) demodulation is used, then twice the highest frequency after the IF.) In
196
this case the phase splitter (contains Hilbert transform in parallel with a unit gain) is also absorbed into
the equalizer, and the same sampled sequence is applied independently to two equalizers, one of which
estimates the real part, and one the imaginary part, of the analytic representation of the data sequence
x
k
e
ωckT
. The two filters can sometimes (depending on the sampling rate) be more cost effective to
implement than the single complex filter, especially when one considers that the Hilbert transform is
incorporated into the equalizer. This approach is often taken with adaptive implementations of the
FSE, as the Hilbert transform is then implemented more exactly adaptively than is possible with fixed
design.
10
This is often called Nyquist Inband Equalization. A particularly common variant of this
type of equalization is to sample at rate 2/T both of the outputs of a continuous-time phase splitter,
one stream of samples staggered by T/4 with respect to the other. The corresponding T/2 complex
equalizer can then be implemented adaptively with four independent adaptive equalizers (rather than
the two filters, real and imaginary part, that nominally characterize complex convolution). The adaptive
filters will then correct for any imperfections in the phase-splitting process, whereas the two filters with
fixed conjugate symmetry could not.
10
However, this structure also slows convergence of many adaptive implementations.
197
Figure 3.34: Decision feedback equalization.
3.6 Decision Feedback Equalization
Decision Feedback Equalization makes use of previous decisions in attempting to estimate the current
symbol (with an SBS detector). Any “trailing” intersymbol interference caused by previous symbols
is reconstructed and then subtracted. The DFE is inherently a nonlinear receiver. However, it can be
analyzed using linear techniques, if one assumes all previous decisions are correct. There is both a MMSE
and a zero-forcing version of decision feedback, and as was true with linear equalization in Sections 3.4
and 3.5, the zero-forcing solution will be a special case of the least-squares solution with the SNR→∞.
Thus, this section derives the MMSE solution, which subsumes the zero-forcing case when SNR →∞.
The Decision Feedback Equalizer (DFE) is shown in Figure 3.34. The configuration contains a linear
feedforward equalizer, W(D), (the settings for this linear equalizer are not necessarily the same as those
for the ZFE or MMSE-LE), augmented by a linear, causal, feedback filter, 1 −B(D), where b
0
= 1. The
feedback filter accepts as input the decision from the previous symbol period; thus, the name “decision
feedback.” The output of the feedforward filter is denoted Z(D), and the input to the decision element
Z

(D). The feedforward filter will try to shape the channel output signal so that it is a causal signal.
The feedback section will then subtract (without noise enhancement) any trailing ISI. Any bias removal
in Figure 3.34 is absorbed into the SBS.
This section assumes that previous decisions are correct. In practice, this may not be true, and
can be a significant weakness of decision-feedback that cannot be overlooked. Nevertheless, the analysis
becomes intractable if it includes errors in the decision feedback section. To date, the most efficient way to
specify the effect of feedback errors has often been via measurement. Section 3.7 provides an exact error-
propagation analysis for finite-length DFE’s that can (unfortunately) require enormous computation on
a digital computer. Section 3.8.1 shows how to eliminate error propagation with precoders.
3.6.1 Minimum-Mean-Square-Error Decision Feedback Equalizer (MMSE-
DFE)
The Minimum-Mean-Square Error Decision Feedback Equalizer (MMSE-DFE) jointly optimizes
the settings of both the feedforward filter w
k
and the feedback filter δ
k
− b
k
to minimize the MSE:
Definition 3.6.1 (Mean Square Error (for DFE)) The MMSE-DFE error signal is
e
k
= x
k
− z

k
. (3.208)
The MMSE for the MMSE-DFE is
σ
2
MMSE−DFE

= min
wk,bk
E
_
[x
k
−z

k
[
2
¸
. (3.209)
The error sequence can be written as
E(D) = X(D) − W(D) Y (D) −[1 −B(D)]X(D) = B(D) X(D) − W(D) Y (D) . (3.210)
198
For any fixed B(D), E[E(D)Y

(D
−∗
)] = 0 to minimize MSE, which leads us to the relation
B(D)
¯
R
xy
(D) − W(D)
¯
R
yy
(D) = 0 . (3.211)
Thus,
W(D) =
B(D)
|p|
_
Q(D) +
1
SNRMFB
_ = B(D) W
MMSE-LE
(D) , (3.212)
for any B(D) with b
0
= 1. (Also, W(D) = W
MMSE-LE
(D) B(D), a consequence of the linearity of
the MMSE estimate, so that E(D) = B(D) E
MMSE-LE
(D))
The autocorrelation function for the error sequence with arbitrary monic B(D) is
¯
R
ee
(D) = B(D)
¯
c
x
B

(D
−∗
) −2'
_
B(D)
¯
R
xy
(D)W

(D
−∗
)
_
+W(D)
¯
R
yy
(D)W

(D
−∗
)(3.213)
= B(D)
¯
c
x
B

(D
−∗
) −W(D)
¯
R
yy
(D)W

(D
−∗
) (3.214)
= B(D)
¯
c
x
B

(D
−∗
) −B(D)
¯
R
xy
(D)
|p|
_
Q(D) +
1
SNRMFB
_B

(D
−∗
) (3.215)
= B(D)R
MMSE−LE
(D)B

(D
−∗
) (3.216)
where R
MMSE−LE
(D) =
N
0
2
p
2
1/(Q(D) + 1/SNR
MFB
) is the autocorrelation function for the error
sequence of a MMSE linear equalizer. The solution for B(D) is then the forward prediction filter
associated with this error sequence as discussed in the Appendix. The linear prediction approach is
developed more in what is called the “linear prediction DFE,” in an exercise where the MMSE-DFE is
the concatenation of a MMSE-LE and a linear-predictor that whitens the error sequence.
In more detail on B(D), the (scaled) inverse autocorrelation has spectral factorization:
Q(D) +
1
SNR
MFB
= γ
0
G(D) G

(D
−∗
), (3.217)
where γ
0
is a positive real number and G(D) is a canonical filter response. A filter response G(D) is
called canonical if it is causal (g
k
= 0 for k < 0), monic (g
0
= 1), and minimum-phase (all its poles
are outside the unit circle, and all its zeroes are on or outside the unit circle). If G(D) is canonical,
then G

(D
−∗
) is anti-canonical i.e., anti-causal, monic, and “maximum-phase” (all poles inside the
unit circle, and all zeros in or on the unit circle). Using this factorization,
¯
R
ee
(D) =
B(D) B

(D
−∗
)
Q(D) + 1/SNR
MFB

N0
2
|p|
2
(3.218)
=
B(D)
G(D)

B

(D
−∗
)
G

(D
−∗
)

N0
2
γ
0
|p|
2
(3.219)
r
ee,0

N0
2
γ
0
|p|
2
, (3.220)
with equality if and only if B(D) = G(D). Thus, the MMSE will then be σ
2
MMSE−DFE
=
N
0
2
γ0p
2
. The
feedforward filter then becomes
W(D) =
G(D)
|p| γ
0
G(D) G

(D
−∗
)
=
1
|p| γ
0
G

(D
−∗
)
. (3.221)
The last step in (3.220) follows from the observations that
r
ee,0
= |
B
G
|
2
N0
2
γ
0
|p|
2
, (3.222)
the fractional polynomial inside the squared norm is necessary monic and causal, and therefore the
squared norm has a minimum value of 1. B(D) and W(D) specify the MMSE-DFE:
199
Lemma 3.6.1 (MMSE-DFE) The MMSE-DFE has feedforward section
W(D) =
1
|p| γ
0
G

(D
−∗
)
(3.223)
(realized with delay, as it is strictly noncausal) and feedback section
B(D) = G(D) (3.224)
where G(D) is the unique canonical factor of the following equation:
Q(D) +
1
SNR
MFB
= γ
0
G(D) G

(D
−∗
) . (3.225)
This text also calls the joint matched-filter/sampler/W(D) combination in the forward path of
the DFE the “Mean-Square Whitened Matched Filter (MS-WMF)”. These settings
for the MMSE-DFE minimize the MSE as was shown above.
3.6.2 Performance Analysis of the MMSE-DFE
Again, the autocorrelation function for the error sequence is
¯
R
ee
(D) =
N0
2
γ
0
|p|
2
. (3.226)
This last result states that the error sequence for the MMSE-DFE is “white” when minimized (since
¯
R
ee
(D) is a constant) and has MMSE or average energy (per real dimension)
N
0
2
p
2
γ
−1
0
. Also,
T

_ π
T

π
T
ln
_
Q(e
−ωT
) +
1
SNR
MFB
_
dω = ln(γ
0
) +
T

_ π
T

π
T
ln
_
G(e
−ωT
)G

(e
−ωT
)
_
dω(3.227)
= ln(γ
0
) . (3.228)
This last result leads to a famous expression for σ
2
MMSE−DFE
, which was first derived by Salz in 1973,
σ
2
MMSE−DFE
=
N0
2
|p|
2
e

T

_
π
T

π
T
ln
_
Q(e
−ωT
) +
1
SNR
MFB
_

. (3.229)
The SNR for the MMSE-DFE can now be easily computed as
SNR
MMSE−DFE
=
¯
c
x
σ
2
MMSE−DFE
= γ
0
SNR
MFB
(3.230)
= SNR
MFB
e
T

_
π
T

π
T
ln
_
Q(e
−ωT
) +
1
SNR
MFB
_

. (3.231)
From the k = 0 term in the defining spectral factorization, γ
0
can also be written
γ
0
=
1 + 1/SNR
MFB
|g|
2
=
1 + 1/SNR
MFB
1 +


i=1
[g
i
[
2
. (3.232)
From this expression, if G(D) = 1 (no ISI), then SNR
MMSE−DFE
= SNR
MFB
+ 1, so that the SNR
would be higher than the matched filter bound. The reason for this apparent anomaly is the artificial
signal power introduced by biased decision regions. This bias is noted by writing
Z

(D) = X(D) − E(D) (3.233)
= X(D) − G(D) X(D) +
1
|p| γ
0
G

(D
−∗
)
Y (D) (3.234)
= X(D) − G(D)X(D) +
Q(D)
γ
0
G

(D
−∗
)
X(D) + N(D)
1
|p|γ
0
G

(D
−∗
)
(3.235)
= X(D) −
1/SNR
MFB
γ
0
G

(D
−∗
)
X(D) +N(D)
1
|p| γ
0
G

(D
−∗
)
. (3.236)
200
The current sample, x
k
, corresponds to the time zero sample of 1 −
1/SNRMFB
γ0·G

(D
−∗
)
, and is equal to (1 −
1
SNRMMSE−DFE
)x
k
. Thus z

k
contains a signal component of x
k
that is reduced in magnitude. Thus,
again using the result from Section 3.2.1, the SNR corresponding to the lowest probability of error
corresponds to the same MMSE-DFE receiver with output scaled to remove the bias. This leads to the
more informative SNR
SNR
MMSE−DFE,U
= SNR
MMSE−DFE
−1 = γ
0
SNR
MFB
− 1 . (3.237)
If G(D) = 1, then SNR
MMSE−DFE,U
= SNR
MFB
. Also,
γ
MMSE−DFE
=
SNR
MFB
SNR
MMSE−DFE,U
=
SNR
MFB
γ
0
SNR
MFB
−1
=
1
γ
0
−1/SNR
MFB
. (3.238)
Rather than scale the decision input, the receiver can scale (up) the feedforward output by
1
1−
1
SNR
MMSE−DFE
.
This will remove the bias, but also increase MSE by the square of the same factor. The SNR will then
be SNR
MMSE−DFE,U
. This result is verified by writing the MS-WMF output
Z(D) = [X(D) |p| Q(D) + N(D)]
1
|p| γ
0
G

(D
−∗
)
(3.239)
where N(D), again, has autocorrelation
¯
R
nn
(D) =
N0
2
Q(D). Z(D) expands to
Z(D) = [X(D) |p| Q(D) +N(D)]
1
|p| γ
0
G

(D
−∗
)
(3.240)
= X(D)
γ
0
G(D) G

(D
−∗
) − 1/SNR
MFB
γ
0
G

(D
−∗
)
+N(D)
1
|p| γ
0
G

(D
−∗
)
(3.241)
= X(D) G(D) −
X(D)
SNR
MFB
γ
0
G

(D
−∗
)
+ N

(D) (3.242)
= X(D)
_
G(D) −
1
SNR
MMSE−DFE
_
+
1
SNR
MMSE−DFE
_
1 −
1
G

(D
−∗
)
_
X(D) + N

(D) (3.243)
= X(D)
_
G(D) −
1
SNR
MMSE−DFE
_
+E

(D) , (3.244)
where N

(D) is the filtered noise at the MS-WMF output, and has autocorrelation function
¯
R
n

n
(D) =
N0
2
Q(D)
|p|
2
γ
2
0
G(D) G

(D
−∗
)
=
¯
c
x
SNR
MMSE−DFE
_
1 −
1/SNR
MFB
Q(D) + 1/SNR
MFB
_
, (3.245)
and
e

k
= −e
k
+
1
SNR
MMSE−DFE
x
k
. (3.246)
The error e

k
is not a white sequence in general. Since x
k
and e

k
are uncorrelated, σ
2
e
= σ
2
MMSE−DFE
=
σ
2
e
+
¯
E
x
SNR
2
MMSE−DFE
. If one defines an “unbiased” monic, causal polynomial
G
U
(D) =
SNR
MMSE−DFE
SNR
MMSE−DFE,U
_
G(D) −
1
SNR
MMSE−DFE
_
, (3.247)
then the MMSE-DFE output is
Z(D) =
SNR
MMSE−DFE,U
SNR
MMSE−DFE
X(D)G
U
(D) +E

(D) . (3.248)
Z
U
(D) removes the scaling
Z
U
(D) = Z(D)
SNR
MMSE−DFE
SNR
MMSE−DFE,U
= X(D)G
U
(D) +E
U
(D) , (3.249)
201
Figure 3.35: Decision feedback equalization with unbiased feedback filter.
where E
U
(D) =
SNRMMSE−DFE
SNRMMSE−DFE,U
E

(D) and has power
σ
2
eU
=
_
σ
2
MMSE−DFE

¯
c
x
SNR
2
MMSE−DFE
_
_
SNR
MMSE−DFE
SNR
MMSE−DFE,U
_
2
(3.250)
=
SNR
2
MMSE−DFE
σ
2
MMSE−DFE

¯
c
x
SNR
2
MMSE−DFE
SNR
2
MMSE−DFE
SNR
2
MMSE−DFE,U
(3.251)
=
¯
c
x
[(SNR
MMSE−DFE
/
¯
c
x
)SNR
MMSE−DFE
σ
2
MMSE−DFE
−1]
SNR
2
MMSE−DFE,U
(3.252)
=
¯
c
x
[SNR
MMSE−DFE
−1]
SNR
2
MMSE−DFE,U
(3.253)
=
¯
c
x
SNR
MMSE−DFE,U
. (3.254)
SNR for this scaled and unbiased MMSE-DFE is thus verified to be
¯
E
x
¯
E
x
/SNRMMSE−DFE,U
= SNR
MMSE−DFE,U
.
The feedback section of this new unbiased MMSE-DFE becomes the polynomial 1−G
U
(D), if the scaling
is performed before the summing junction, but is 1 − G(D) if scaling is performed after the summing
junction. The alternative unbiased MMSE-DFE is shown in Figure 3.35.
All the results on fractional spacing for the LE still apply to the feedforward section of the DFE
– thus this text does not reconsider them for the DFE, other than to state that the characteristic of
the feedforward section is different than the LE characteristic in all but trivial channels. Again, the
realization of any characteristic can be done with a matched filter and symbol-spaced sampling or with
an anti-alias filter and fractionally spaced sampling, as in Section 3.5.4.
3.6.3 Zero-Forcing DFE
The ZF-DFE feedforward and feedback filters are found by setting SNR
MFB
→ ∞ in the expressions
for the MMSE-DFE. The spectral factorization (assuming ln(Q(e
ωT
) is integrable over (−
π
T
,
π
T
), see
Appendix A)
Q(D) = η
0
P
c
(D) P

c
(D
−∗
) (3.255)
then determines the ZF-DFE filters as
B(D) = P
c
(D) , (3.256)
W(D) =
1
η
0
|p| P

c
(D
−∗
)
. (3.257)
P
c
is sometimes called a canonical pulse response for the channel. Since q
0
= 1 = η
0
|P
c
|
2
, then
η
0
=
1
1 +


i=1
[p
c,i
[
2
. (3.258)
202
Equation (3.258) shows that there is a signal power loss of the ratio of the squared first tap magnitude
in the canonical pulse response to the squared norm of all the taps. This loss ratio is minimized for a
minimum phase polynomial, and P
c
(D) is the minimum-phase equivalent of the channel pulse response.
The noise at the output of the feedforward filter has (white) autocorrelation function
N0
2
/(|p|
2
η
0
), so
that
σ
2
ZFDFE
=
N0
2
|p|
2
1
η
0
=
N0
2
|p|
2
e

T

_
π
T

π
T
ln
_
Q(e
ωT
)
_

. (3.259)
The corresponding SNR is then
SNR
ZF−DFE
= η
0
SNR
MFB
=
1
1 +


i=1
p
2
c,i
SNR
MFB
. (3.260)
3.6.4 Examples Revisited
EXAMPLE 3.6.1 (PAM - DFE) The pulse response of a channel used with binary PAM
is again given by
P(ω) =
_ √
T
_
1 +.9e
ωT
_
[ω[ ≤
π
T
0 [ω[ >
π
T
. (3.261)
Again, the SNR
MFB
is 10dB and c
x
= 1.
Then
˜
Q(D)

= Q(D) + 1/SNR
MFB
is computed as
˜
Q(e
−ωT
) =
1.81 + 1.8 cos(ωT) + 1.81/10
1.81
(3.262)
˜
Q(D) =
1
1.81
(1 +.9D)(1 + .9D
−1
) +.1 (3.263)
=
1
1.81
_
.9D + 1.991 +.9D
−1
_
. (3.264)
The roots of
˜
Q(D) are −.633 and −1.58. The function
˜
Q(D) has canonical factorization
˜
Q(D) = .785(1 + .633D)(1 + .633D
−1
) . (3.265)
Then, γ
0
= .785. The feedforward section is
W(D) =
1

1.81(.785)(1 +.633D
−1
)
=
.9469
1 +.633D
−1
, (3.266)
and the feedforward filter transfer function appears in Figure 3.36. The feedback section is
B(D) = G(D) = 1 + .633D , (3.267)
with magnitude in Figure 3.37. The MS-WMF filter transfer function appears in Figure 3.38,
illustrating the near all-pass character of this filter.
The MMSE is
σ
2
MMSE−DFE
=
^
0
2
1
|p|
2
γ
0
= .
.181
1.81 .785
= .1274 (3.268)
and
SNR
MMSE−DFE,U
=
1 −.1274
.1274
= 6.85 (8.4dB) . (3.269)
Thus, the MMSE-DFE is only 1.6dB below the MFB on this channel. However, the ZF-
DFE for this same channel would produce η
0
= 1/|p
c
|
2
=.5525. In this case σ
2
ZFDFE
=
.181/(.5525)1.81 = .181 and the loss with respect to the MFB would be η
0
= 2.6 dB, 1 dB
lower than the MMSE-DFE for this channel.
It will in fact be possible to do yet better on this channel, using sequence detection, as
discussed in Chapter 9, and/or by codes (Chapters 10 and 11). However, what originally
appeared as a a stunning loss of 9.8 dB with the ZFE has now been reduced to a much
smaller 1.6 dB.
203
Figure 3.36: Feedforward filter transfer function for real-channel example.
Figure 3.37: Feedback filter transfer function for real-channel example.
204
Figure 3.38: MS-WMF transfer function for real-channel example.
The QAM example is also revisited here:
EXAMPLE 3.6.2 (QAM - DFE) Recall that the equivalent baseband pulse response
samples were given by
p
k
=
1

T
_

1
2
_
1 +

4
_


2
_
. (3.270)
The SNR
MFB
is again 10 dB. Then,
˜
Q =
−.25D
−2
+.625(−1 +)D
−1
+ 1.5625(1 + .1) −.625(1 +)D + .25D
2
1.5625
. (3.271)
or
˜
Q = (1 −.451

1.632
D)(1 −.451

−.0612
D)(1 − .451

−1.632
D
−1
)(1 − .451

.0612
D
−1
)

−
1.5625 4 (.451

−1.632
)(.451

.0612
)
(3.272)
fromwhich one extracts γ
0
= .7866 and G(D) = 1−.4226(1+)D+.2034D
2
. The feedforward
and feedback sections can be computed in a straightforward manner, as
B(D) = G(D) = 1 − .4226(1 +)D +.2034D
2
(3.273)
and
W(D) =
1.017
1 −.4226(1 − )D
−1
− .2034D
−2
. (3.274)
The MSE is
σ
2
MMSE−DFE
= (.15625)
1
.7866(1.5625)
= .1271 . (3.275)
and the corresponding SNR is
SNR
MMSE−DFE,U
=
1
.1271
− 1 = 8.4dB . (3.276)
205
The loss is (also coincidentally) 1.6 dB with respect to the MFB and is 1.7 dB better than
the MMSE-LE on this channel.
For the ZF-DFE,
Q(D) = (1 − .5

π/2
D)(1 − .5D)(1 − .5

−π/2
D
−1
)(1 −.5D
−1
) (3.277)

−
1.5625 4 (.5

−π/2
)(.5)
(3.278)
and thus η
0
= .6400 and P
c
(D) = 1 − .5(1 + )D + .25D
2
. The feedforward and feedback
sections can be computed in a straightforward manner, as
B(D) = P
c
(D) = 1 −.5(1 +)D + .25D
2
(3.279)
and
W(D) =
1.25
1 −.5(1 −)D
−1
−.25D
−2
. (3.280)
The output noise variance is
σ
2
ZFDFE
= (.15625)
1
.6400(1.5625)
= .1563 . (3.281)
and the corresponding SNR is
SNR
ZFDFE
=
1
.1563
= 6.4 = 8.0dB . (3.282)
The loss is 2.0 dB with respect to the MFB and is .4 dB worse than the MMSE-DFE.
206
Figure 3.39: Polyphase subchannel representation of the channel pulse response.
3.7 Finite Length Equalizers
The previous developments of discrete-time equalization structures presumed that the equalizer could
exist over an infinite interval in time. Equalization filters, w
k
, are almost exclusively realized as finite-
impulse-response (FIR) filters in practice. Usually these structures have better numerical properties
than IIR structures. Even more importantly, adaptive equalizers (see Section 3.8 and Chapter 13) are
often implemented with FIR structures for w
k
(and also for b
k
, in the case of the adaptive DFE). This
section studies the design of and the performance analysis of FIR equalizers.
Both the LE and the DFE cases are examined for both zero-forcing and least-squares criteria. This
section describes design for the MMSE situation and then lets SNR→∞ to get the zero-forcing special
case. Because of the finite length, the FIR zero-forcing equalizer cannot completely eliminate ISI in
general.
3.7.1 FIR MMSE-LE
Returning to the FSE in Figure 3.32, the matched filtering operation will be performed digitally (at
sampling rate l/T) and the FIR MMSE-LE will then incorporate both matched filtering and equalization.
Perfect anti-alias filtering with gain

T precedes the sampler and the combined pulse-response/anti-alias
filter is ˜ p(t). The assumption that l is an integer simplifies the specification of matrices.
One way to view the oversampled channel output is as a set of l parallel T-spaced subchannels whose
pulse responses are offset by T/l from each other as in Figure 3.39. Each subchannel produces one of
the l phases per symbol period of the output set of samples at sampling rate l/T. Mathematically, it is
convenient to represent such a system with vectors as shown below.
The channel output y(t) is
y(t) =

m
x
m
˜ p(t −mT) + n(t) , (3.283)
which, if sampled at time instants t = kT −
iT
l
, i = 0, ..., l −1, becomes
y(kT −
iT
l
) =

m=−∞
x
m
˜ p(kT −
iT
l
− mT) +n(kT −
iT
l
) . (3.284)
The (per-dimensional) variance of the noise samples is
N0
2
l because the gain of the anti-alias filter,

T,
is absorbed into p(t). The l phases per symbol period of the oversampled y(t) are contained in
y
k
=
_
¸
¸
¸
_
y(kT)
y(kT −
T
l
)
.
.
.
y(kT −
l−1
l
T)
_
¸
¸
¸
_
. (3.285)
207
The vector y
k
can be written as
y
k
=

m=−∞
x
m
p
k−m
+ n
k
=

m=−∞
x
k−m
p
m
+ n
k
, (3.286)
where
p
k
=
_
¸
¸
¸
_
˜ p(kT)
˜ p(kT −
T
l
)
.
.
.
˜ p(kT −
l−1
l
T)
_
¸
¸
¸
_
and n
k
=
_
¸
¸
¸
_
n(kT)
n(kT −
T
l
)
.
.
.
n(kT −
l−1
l
T)
_
¸
¸
¸
_
(3.287)
The response ˜ p(t) is assumed to extend only over a finite time interval. In practice, this assumption
requires any nonzero component of p(t) outside of this time interval to be negligible. This time interval
is 0 ≤ t ≤ νT. Thus, p
k
= 0 for k < 0 , and for k > ν. The sum in (3.284) becomes
y
k
= [ p
0
p
1
, ..., p
ν
]
_
¸
¸
¸
¸
¸
¸
_
x
k
x
k−1
.
.
.
.
.
.
x
k−ν
_
¸
¸
¸
¸
¸
¸
_
+n
k
. (3.288)
Each row in (3.288) corresponds to a sample at the output of one of the filters in Figure 3.39. More
generally, for N
f
successive l-tuples of samples of y(t),
Y
k

=
_
¸
¸
¸
_
y
k
y
k−1
.
.
.
y
k−Nf+1
_
¸
¸
¸
_
(3.289)
=
_
¸
¸
¸
_
p
0
p
1
... p
ν
0 0 ... 0
0 p
0
p
1
... ... p
ν
... 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 ... 0 0 p
0
p
1
... p
ν
_
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
_
x
k
x
k−1
.
.
.
.
.
.
x
k−Nf−ν+1
_
¸
¸
¸
¸
¸
¸
_
+
_
¸
¸
¸
_
n
k
n
k−1
.
.
.
n
k−Nf +1
_
¸
¸
¸
_
. (3.290)
This text uses P to denote the large (N
f
l) (N
f
+ ν) matrix in (3.290), while X
k
denotes the data
vector, and N denotes the noise vector. Then, the oversampled vector representation of the channel is
Y
k
= PX
k
+N
k
. (3.291)
When l = n/m (a rational fraction), then (3.291) still holds with P an [N
f

n
m
] (N
f
+ν) matrix that
does not follow the form in (3.290), with each row possibly having a set of coefficients unrelated to the
other rows.
11
An equalizer is applied to the sampled channel output vector Y
k
by taking the inner product of an
N
f
l-dimensional row vector of equalizer coefficients, w, and Y
k
, so Z(D) = W(D)Y (D) can be written
z
k
= wY
k
. (3.292)
11
In this case, (3.283) is used to compute each row at the appropriate sampling instants. N
f
·
n
m
should also be an
integer so that P is constant. Otherwise, P becomes a “time-varying” matrix P
k
.
208
For causality, the designer picks a channel-equalizer system delay ∆ T symbol periods,
∆ ≈
ν + N
f
2
, (3.293)
with the exact value being of little consequence unless the equalizer length N
f
is very short. The delay
is important in finite-length design because a non-causal filter cannot be implemented, and the delay
allows time for the transmit data symbol to reach the receiver. With infinite-length filters, the need for
such a delay does not enter the mathematics because the infinite-length filters are not realizable in any
case, so that infinite-length analysis simply provides best bounds on performance. ∆ is approximately
the sum of the channel and equalizer delays in symbol periods. The equalizer output error is then
e
k
= x
k−∆
−z
k
, (3.294)
and the corresponding MSE is
σ
2
MMSE−LE
= E
_
[e
k
[
2
_
= E¦e
k
e

k
¦ = E
_
(x
k−∆
−z
k
) (x
k−∆
− z
k
)

_
. (3.295)
Using the Orthogonality Principle of Appendix A, the MSE in (3.295) is minimized when
E¦e
k
Y

k
¦ = 0 . (3.296)
In other words, the equalizer error signal will be minimized (in the mean square sense) when this error
is uncorrelated with any channel output sample in the delay line of the FIR MMSE-LE. Thus
E¦x
k−∆
Y

k
¦ − wE¦Y
k
Y

k
¦ = 0 . (3.297)
The two statistical correlation quantities in (3.297) have a very special significance (both y(t) and x
k
are stationary):
Definition 3.7.1 (Autocorrelation and Cross-Correlation Matrices) The FIR MMSE
autocorrelation matrix is defined as
R
Y Y

= E¦Y
k
Y

k
¦ /N , (3.298)
while the FIR MMSE-LE cross-correlation vector is defined by
R
Y x

= E
_
Y
k
x

k−∆
_
/N , (3.299)
where N = 1 for real and N = 2 for complex. Note R
xY
= R

Y x
, and R
Y Y
is not a
function of ∆.
With the above definitions in (3.297), the designer obtains the MMSE-LE:
Definition 3.7.2 (FIR MMSE-LE) The FIR MMSE-LE for sampling rate l/T, delay ∆,
and of length N
f
symbol periods has coefficients
w = R

Y x
R
−1
Y Y
= R
xY
R
−1
Y Y
, (3.300)
or equivalently,
w

= R
−1
Y Y
R
Y x
. (3.301)
In general, it may be of interest to derive more specific expressions for R
Y Y
and R
xY
.
R
xY
= E¦x
k−∆
Y

k
¦ = E¦x
k−∆
X

k
¦P

+E¦x
k−∆
N

k
¦ (3.302)
=
_
0 ... 0
¯
c
x
0 ... 0
¸
P

+ 0 (3.303)
=
¯
c
x
[ 0 ... 0 p

ν
... p

0
0 ... 0 ] (3.304)
209
R
Y Y
= E¦Y
k
Y

k
¦ = PE¦X
k
X

k
¦P

+E¦N
k
N

k
¦ (3.305)
=
¯
c
x
PP

+l
^
0
2
R
nn
, (3.306)
where the (l
N0
2
-normalized) noise autocorrelation matrix R
nn
is equal to I when the noise is white.
A convenient expression is
[0...0 p

ν
...p

0
0...0] = 1


P

, (3.307)
so that 1

is an N
f
+ν-vector of 0’s and a 1 in the (∆+1)
th
position. Then the FIR MMSE-LE becomes
(in terms of P, SNR and ∆)
w = 1


P

_
PP

+
1
SNR
R
nn
_
−1
= 1


_
P

R
−1
nn
P +
l
SNR
I
_
−1
P

R
−1
nn
. (3.308)
The matrix inversion lemma of Appendix B is used in deriving the above equation.
The position of the nonzero elements in (3.304) depend on the choice of ∆. The MMSE, σ
2
MMSE−LE
,
can be found by evaluating (3.295) with (3.300):
σ
2
MMSE−LE
= E
_
x
k−∆
x

k−∆
−x
k−∆
Y

k
w

−wY
k
x

k−∆
+ wY
k
Y

k
w

_
(3.309)
=
¯
c
x
− R
xY
w

−wR
Y x
+wR
Y Y
w

(3.310)
=
¯
c
x
− wR
Y Y
w

(3.311)
=
¯
c
x
− R
xY
R
−1
Y Y
R
Y x
(3.312)
=
¯
c
x
− wR
Y x
, (3.313)
With algebra, (3.313) is the same as
σ
2
MMSE−LE
=
l
N0
2
|p|
2
1


_
P

R
−1
nn
P
|p|
2
+
l
SNR
MFB
I
_−1
1

=
l
N0
2
|p|
2
1


¯
Q
−1
1

, (3.314)
so that the best value of ∆ (the position of the 1 in the vectors above) corresponds to the smallest
diagonal element of the inverted (“Q-tilde”) matrix in (3.314) – this means that only one matrix need be
inverted to obtain the correct ∆ value as well as compute the equalizer settings. A homework problem
develops some interesting relationships for w in terms of P and SNR. Thus, the SNR for the FIR
MMSE-LE is
SNR
MMSE−LE
=
¯
c
x
σ
2
MMSE−LE
, (3.315)
and the corresponding unbiased SNR is
SNR
MMSE−LE,U
= SNR
MMSE−LE
− 1 . (3.316)
The loss with respect to the MFB is, again,
γ
MMSE−LE
=
SNR
MFB
SNR
MMSE−LE,U
. (3.317)
3.7.2 FIR ZFE
One obtains the FIR ZFE by letting the SNR→∞ in the FIR MMSE, which alters R
Y Y
to
R
Y Y
= PP

¯
c
x
(3.318)
and then w remains as
w = R
xY
R
−1
Y Y
. (3.319)
210
Because the FIR equalizer may not be sufficiently long to cancel all ISI, the FIR ZFE may still have
nonzero residual ISI. This ISI power is given by
σ
2
MMSE−ZFE
=
¯
c
x
−wR
Y x
. (3.320)
However, (3.320) still ignores the enhanced noise at the FIR ZFE output.
12
The power of this noise is
easily found to be
FIR ZFE noise power =
^
0
2
l |w|
2
R
nn
, (3.321)
making the SNR at the ZFE output
SNR
ZFE
=
¯
c
x
¯
c
x
− wR
Y x
+
N0
2
l|w|
2
R
nn
. (3.322)
The ZFE is still unbiased in the finite-length case:
13
E[wY
k
/x
k−∆
] = R
xY
R
−1
Y Y
¦PE[X
k
/x
k−∆
] + E[N
k
]¦ (3.323)
= [0 0 ...1 0]E
_
P

(PP

)
−1
PX
k
/x
k−∆
_
(3.324)
= [0 0 ...1 0]P

(PP

)
−1
PE
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
_
x
k
x
k−1
.
.
.
x
k−Nf
don’t care
_
¸
¸
¸
¸
¸
_
/x
k−∆
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
(3.325)
= x
k−∆
. (3.326)
Equation (3.326) is true if ∆ is a practical value and the finite-length ZFE has enough taps. The loss is
the ratio of the SNR to SNR
ZFE
,
γ
ZFE
=
SNR
MFB
SNR
ZFE
. (3.327)
3.7.3 example
For the earlier PAM example, one notes that sampling with l = 1 is sufficient to represent all signals.
First, choose N
f
= 3 and note that ν=1. Then
P =
_
_
.9 1 0 0
0 .9 1 0
0 0 .9 1
_
_
. (3.328)
With a choice of ∆ = 2, then
R
Y x
=
¯
c
x
P
_
¸
¸
_
0
0
1
0
_
¸
¸
_
(3.329)
=
_
_
0
1
.9
_
_
, (3.330)
12
The notation w
2
R
nn
means wRnnw

.
13
The matrix P

(PP

)
−1
P is a “projection matrix” and PP

is full rank; therefore the entry of x
k−∆
passes directly
(or is zeored, in which case ∆ needs to be changed).
211
Figure 3.40: FIR equalizer performance for 1 + .9D
−1
versus number of equalizer taps.
and
R
Y Y
=
¯
c
x
_
PP

+
1
SNR
I
_
(3.331)
=
_
_
1.991 .9 0
.9 1.991 .9
0 .9 1.991
_
_
, (3.332)
The FIR MMSE is
w

= R
−1
Y Y
R
Y x
=
_
_
.676 −.384 .174
−.384 .849 −.384
.174 −.384 .676
_
_
_
_
0
1
.9
_
_
=
_
_
−.23
.51
.22
_
_
(3.333)
Then
σ
2
MMSE−LE
=
_
_
1 − [−.23 .51 .22]
_
_
0
1
.9
_
_
_
_
= .294 (3.334)
The SNR is computed to be
SNR
MMSE−LE,U
=
1
.294
−1 = 2.4 (3.8 dB) . (3.335)
which is 6.2dB below MFB performance and 1.9dB worse than the infinite length MMSE-LE for this
channel.
With sufficiently large number of taps for this channel, the infinite length performance level can be
attained. This performance is plotted as versus the number of equalizer taps in Figure 3.40. Clearly, 15
taps are sufficient for infinite-length performance.
EXAMPLE 3.7.1 For the earlier complex QAM example, sampling with l = 1 is sufficient
212
to represent all signals. First, choose N
f
= 4 and note that ν = 2. Then
P =
_
¸
¸
_
−.5 1 + /4 −/2 0 0 0
0 −.5 1 +/4 −/2 0 0
0 0 −.5 1 + /4 −/2 0
0 0 0 −.5 1 +/4 −/2
_
¸
¸
_
. (3.336)
The matrix (P

P +.15625 I)
−1
has the same smallest element 1.3578 for both ∆ = 2, 3, so
choosing ∆ = 2 will not unnecessarily increase system delay. With a choice of ∆ = 2, then
R
Y x
=
¯
c
x
P
_
¸
¸
¸
¸
¸
¸
_
0
0
1
0
0
0
_
¸
¸
¸
¸
¸
¸
_
(3.337)
=
_
¸
¸
_
−.5
1 + /4
−.5
0
_
¸
¸
_
, (3.338)
and
R
Y Y
=
¯
c
x
_
PP

+
1
SNR
I
_
(3.339)
=
_
¸
¸
_
1.7188 −0.6250 − 0.6250 /4 0
−0.6250 + 0.6250 1.7188 −0.6250 − 0.6250 /4
−/4 −0.6250 + 0.6250 1.7188 −0.6250 − 0.6250
0 −/4 −0.6250 + 0.6250 1.7188
_
¸
¸
_
.
The FIR MMSE is
w

= R
−1
Y Y
R
Y x
=
_
¸
¸
_
0.2570 −0.0422
0.7313 −0.0948
−0.1182 −0.2982
−0.1376 + 0.0409
.
_
¸
¸
_
(3.340)
Then
σ
2
MMSE−LE
= (1 −wR
Y x
) (3.341)
= .2121 . (3.342)
The SNR is computed to be
SNR
MMSE−LE,U
=
1
.2121
− 1 = 3.714 (5.7 dB) , (3.343)
which is 4.3 dB below MFB performance and 2.1 dB worse than the infinite-length MMSE-LE
for this channel. With sufficiently large number of taps for this channel, the infinite-length
performance level can be attained.
3.7.4 FIR MMSE-DFE
The FIR DFE case is similar to the feed-forward equalizers just discussed, except that we now augment
the fractionally spaced feed-forward section with a symbol-spaced feedback section. The MSE for the
DFE case, as long as w and b are sufficiently long, is
MSE = E
_
[x
k−∆
− wY
k
+bx
k−∆−1
[
2
_
, (3.344)
213
where b is the vector of coefficients for the feedback FIR filter
b

= [ b
1
b
2
... b
Nb
] , (3.345)
and x
k−∆−1
is the vector of data symbols in the feedback path. It is mathematically convenient to
define an augmented response vector for the DFE as
˜ w

=
_
w
.
.
. −b
_
, (3.346)
and a corresponding augmented DFE input vector as
˜
Y
k

=
_
Y
k
x
k−∆−1
_
. (3.347)
Then, (3.344) becomes
MSE = E
_
[x
k−∆
− ˜ w
˜
Y
k
[
2
_
. (3.348)
The solution, paralleling (3.299) - (3.301), is determined using the following two definitions:
Definition 3.7.3 (FIR MMSE-DFE Autocorrelation and Cross-Correlation Matrices)
The FIR MMSE-DFE autocorrelation matrix is defined as
R
˜
Y
˜
Y

= E
_
˜
Y
k
˜
Y

k
_
/N (3.349)
=
_
R
Y Y
E[Y
k
x

k−∆−1
]
E[x
k−∆−1
Y

k
]
¯
c
x
I
Nb
_
(3.350)
=
_
¯
c
x
_
PP

+
l·R
nn
SNR
_
¯
c
x
PJ

]
¯
c
x
J


P

¯
c
x
I
Nb
_
, (3.351)
where J

is an (N
f
+ν) N
b
matrix of 0’s and 1’s, which has the upper ∆ + 1 rows zeroed
and an identity matrix of dimension min(N
b
, N
f
+ν −∆− 1) with zeros to the right (when
N
f
+ν −∆−1 < N
b
), zeros below (when N
f
+ν −∆−1 > N
b
), or no zeros to the right or
below exactly fitting in the bottom of J

(when N
f
+ν −∆− 1 = N
b
).
The corresponding FIR MMSE-DFE cross-correlation vector is
R
˜
Y x

= E
_
˜
Y
k
x

k−∆
_
/N (3.352)
=
_
R
Y x
0
_
(3.353)
=
_
¯
c
x
P1

0
_
, (3.354)
where, again, N = 1 for real signals and N = 2 for complex signals.
The FIR MMSE-DFE for sampling rate l/T, delay ∆, and of length N
f
and N
b
has coefficients
˜ w = R
x
˜
Y
R
−1
˜
Y
˜
Y
. (3.355)
Equation (3.355) can be rewritten in detail as
_
w
.
.
. − b
_

¯
c
x

_
PP

+
l
SNR
R
nn
PJ

J


P

I
Nb
_
=
_
¯
c
x
1


P

.
.
.0
_
, (3.356)
which reduces to the pair of equations
w
_
PP

+
l
SNR
R
nn
_
−bJ


P

= 1


P

(3.357)
wPJ

− b = 0 . (3.358)
214
Then
b = wPJ

, (3.359)
and thus
w
_
PP

−PJ

J


P

+
l
SNR
R
nn
_
= 1


P

(3.360)
or
w = 1


P

_
PP

− PJ

J


P

+
l
SNR
R
nn
_
−1
. (3.361)
Then
b = 1


P

_
PP

−PJ

J


P

+
l
SNR
R
nn
_
−1
PJ

. (3.362)
The MMSE is then
σ
2
MMSE−DFE
=
¯
c
x
− ˜ wR
˜
Y x
(3.363)
=
¯
c
x
− wR
Y x
(3.364)
=
¯
c
x
_
1 −1


P

_
PP

− PJ

J


P

+
l
SNR
R
nn
_
−1
P1

_
, (3.365)
which is a function to be minimized over ∆. Thus, the SNR for the unbiased FIR MMSE-DFE is
SNR
MMSE−DFE,U
=
¯
c
x
σ
2
MMSE−DFE
− 1 =
˜ wR
˜
Y x
¯
c
x
− ˜ wR
˜
Y x
, (3.366)
and the loss is again
γ
MMSE−DFE
=
SNR
MFB
SNR
MMSE−DFE,U
. (3.367)
EXAMPLE 3.7.2 (MMSE-DFEs, PAM and QAM) For the earlier example with l =
1, N
f
= 2, and N
b
= 1 and ν = 1, we will also choose ∆ = 1. Then
P =
_
.9 1 0
0 .9 1
_
, (3.368)
J

=
_
_
0
0
1
_
_
, (3.369)
1

=
_
_
0
1
0
_
_
, (3.370)
and
R
Y Y
=
_
1.991 .9
.9 1.991
_
. (3.371)
Then
w = [0 1 0]
_
_
.9 0
1 .9
0 1
_
_
__
1.991 .9
.9 1.991
_

_
0
1
_
[0 1]
_
−1
(3.372)
= [1 .9]
_
1.991 .9
.9 0.991
_
−1
(3.373)
= [.1556 .7668] (3.374)
b = wPJ

(3.375)
= [.1556 .7668]
_
.9 1 0
0 .9 1
_
_
_
0
0
1
_
_
(3.376)
= .7668 . (3.377)
215
Figure 3.41: FIR MMSE-DFE Example.
Then
σ
2
MMSE−DFE
=
_
1 − [.16 .76]
_
1
.9
__
= .157 . (3.378)
The SNR is computed to be
SNR
MMSE−DFE,U
=
1 − .157
.157
= 5.4 (7.3dB) , (3.379)
which is 2.7 dB below MFB performance, but 2.8 dB better than the infinite-lengthMMSE-
LE for this channel! The loss with respect to the infinite-length MMSE-DFE is 1.1 dB. A
picture of the FIR MMSE-DFE is shown in Figure 3.41. With sufficiently large number of
taps for this channel, the infinite-length performance level can be attained. This performance
is plotted versus the number of feed-forward taps (only one feedback tap is necessary for
infinite-length performance) in Figure 3.42. In this case, 7 feed-forward and 1 feedback
taps are necessary for infinite-length performance. Thus, the finite-length DFE not only
outperforms the finite or infinite-length LE, it uses less taps (less complexity) also. For the
QAM example l = 1, N
f
= 2, N
b
= 2, and ν = 2, we also choose ∆ = 1. Actually this
channel will need more taps to do well with the DFE structure, but we can still choose these
values and proceed. Then
P =
_
−.5 1 + /4 −/2 0
0 −.5 1 +/4 −/2
_
, (3.380)
J

=
_
¸
¸
_
0 0
0 0
1 0
0 1
_
¸
¸
_
, (3.381)
1

=
_
¸
¸
_
0
1
0
0
_
¸
¸
_
, (3.382)
and
R
Y Y
=
_
1.7188 −0.6250 − 0.6250
−0.6250 −0.6250 1.7188
_
. (3.383)
216
Then
w = [0 1 0 0]
_
¸
¸
_
−.5 0
1 −/4 −.5
/2 1 − /4
0 /2
_
¸
¸
_
(3.384)
__
1.7188 −0.6250 − 0.6250
−0.6250 − 0.6250 1.7188
_

_
1/4 −1/8 − /2
−1/8 +/2 1.3125
__
−1
= [1 − /2 −.5]
_
1.2277 1.5103 + .3776
1.5103 − .3776 4.4366
_
(3.385)
= [.4720 −.1180 −.6136] (3.386)
b = wPJ

(3.387)
= [.4720 −.1180 .6136]
_
−.5 1 + /4 −/2 0
0 −.5 1 +/4 −/2
_
_
¸
¸
_
0 0
0 0
1 0
0 1
_
¸
¸
_
(3.388)
= [−.6726 − .3894 .3608] . (3.389)
Then
σ
2
MMSE−DFE
= .1917 . (3.390)
The SNR is computed to be
SNR
MMSE−DFE,U
=
1 −.1917
.1917
= 6.23dB , (3.391)
which is 3.77 dB below MFB performance, and not very good on this channel! The reader
may evaluate various filter lengths and delays to find a best use of 3, 4, 5, and 6 parameters
on this channel.
FIR ZF-DFE One obtains the FIR ZF-DFE by letting the SNR→∞ in the FIR MMSE-DFE, which
alters R
˜
Y
˜
Y
to
R
˜
Y
˜
Y
=
_
¯
c
x
˜
P
˜
P

¯
c
x
PJ

¯
c
x
J


P

¯
c
x
I
Nb
_
(3.392)
and then ˜ w remains as
˜ w = R
x
˜
Y
R
−1
˜
Y
˜
Y
. (3.393)
Because the FIR equalizer may not be sufficiently long to cancel all ISI, the FIR ZF-DFE may still have
nonzero residual ISI. This ISI power is given by
σ
2
MMSE−DFE
=
¯
c
x
− ˜ wR
˜
Y x
. (3.394)
However, (3.394) still ignores the enhanced noise at the ˜ w
k
filter output. The power of this noise is
easily found to be
FIR ZFDFE noise variance =
^
0
2
l |w|
2
, (3.395)
making the SNR at the FIR ZF-DFE output
SNR
ZF−DFE
=
¯
c
x
¯
c
x
− ˜ wR
˜
Y x
+
¯
E
x
SNR
l|w|
2
. (3.396)
The filter is w in (3.395) and (3.396), not ˜ w because only the feed-forward section filters noise. The loss
is:
γ
ZF-DFE
=
SNR
MFB
SNR
ZF−DFE
. (3.397)
217
Figure 3.42: FIR MMSE-DFE performance for 1 +.9D
−1
versus number of feed-forward equalizer taps.
3.7.5 An Alternative Approach to the DFE
While the above approach directly computes the settings for the FIR DFE, it yields less insight into
the internal structure of the DFE than did the infinite-length structures investigated in Section 3.6,
particularly the spectral (canonical) factorization into causal and anti-causal ISI components is not
explicit. This subsection provides such an alternative caused by the finite-length filters. This is because
finite-length filters inherently correspond to non-stationary processing.
14
The vector of inputs at time k to the feed-forward filter is again Y
k
, and the corresponding vector
of filter coefficients is w, so the feed-forward filter output is z
k
= wY
k
. Continuing in the same fashion,
the DFE error signal is then described by
e
k
= bX
k
(∆) − wY
k
, (3.398)
where x

k:k−N
b
=
_
x

k
... x

k−Nb
¸
, and b is now slightly altered to be monic and causal
b

= [1 b
1
b
2
... b
Nb
] . (3.399)
The mean-square of this error is to be minimized over b and w. Signals will again be presumed to
be complex, but all developments here simplify to the one-dimensional real case directly, although it is
important to remember to divide any complex signal’s variance by 2 to get the energy per real dimension.
The SNR equals
¯
c
x
/
N0
2
= c
x
/^
0
in either case.
Optimizing the feed-forward and feedback filters For any fixed b in (3.398), the cross-correlation
between the error and the vector of channel outputs Y
k
should be zero to minimize MSE,
E[e
k
Y

k
] = 0 . (3.400)
So,
wR
Y Y
= bR
XY
(∆) , (3.401)
14
A more perfect analogy between finite-lengthand infinite-lengthDFE’s occurs in Chapter 5, where the best finite-length
DFE’s are actually periodic over a packet period corresponding to the length of the feed-forward filter.
218
and
R
XY
(∆) = E
_
X
k
(∆)
_
x

k
x

k−1
... x

k−Nf−ν+1
_
P

_
(3.402)
=
¯
c
x
˜
J


P

, (3.403)
where the matrix
˜
J


has the first ∆ columns all zeros, then an (up to) N
b
N
b
identity matrix at
the top of the up to N
b
columns, and zeros in any row entries below that identity, and possibly zeroed
columns following the identity if N
f
+ν −1 > N
b
. The following matlab commands produce
˜
J


, using
for instance N
f
= 8, N
b
= 5, ∆ = 3, and ν = 5,
>> size=min(Delta,Nb);
>> Jdelta=[ zeros(size,size) eye(size) zeros(size, max(Nf+nu-2*size,0))
zeros(max(Nb-Delta,0),Nf+nu)]
Jdelta =
0 0 0 1 0 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0 0 0 0 0
0 0 0 0 0 1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0
or for N
f
= 8, N
b
= 2, ∆ = 3, and ν = 5,
>> >> size=min(Delta,Nb)
size = 2
>> Jdelta=[ zeros(size,size) eye(size) zeros(size, max(Nf+nu-2*size,0)),
zeros(max(Nb-Delta,0),Nf+nu)]
Jdelta =
0 0 1 0 0 0 0 0 0 0 0 0 0
0 0 0 1 0 0 0 0 0 0 0 0 0
The MSE for this fixed value of b then becomes
σ
2

(∆) = b

_
¯
c
x
I − R
XY
(∆)R
−1
Y Y
R
Y X
(∆)
_
b (3.404)
= b

R

X/Y
(∆)b (3.405)
= b

_
¯
c
x
I
Nb

˜
J


P

_
PP

+
l
SNR
R
nn
_
−1P
˜
J

(∆)
_
b (3.406)
= l
^
0
2
_
b

˜
J


_
P

R
−1
nn
P +
l
SNR
I
_
−1P
˜
J

(∆)
˜
J

b
_
, (3.407)
and R

X/Y
(∆)

=
¯
c
x
I − R
XY
(∆)R
−1
Y Y
R
Y X
(∆), the autocorrelation matrix corresponding to the
MMSE vector in estimating X(∆) from Y
k
. This expression is the equivalent of (3.216). That is
R

X/Y
(∆) is the autocorrelation matrix for the error sequence of length N −b that is associated with a
“matrix” MMSE-LE. The solution then requires factorization of the inner matrix into canonical factors,
which is executed with Cholesky factorization for finite matrices.
By defining
˜
Q(∆) =
_
˜
J


_
P

P +
l
SNR
I
_
−1
˜
J

_
−1
, (3.408)
this matrix is equivalent to (3.217), except for the “annoying”
˜
J

matrices that become identities as N
f
and N
b
become infinite, then leaving
˜
Q as essentially the autocorrelation of the channel output. The
219
˜
J

factors, however, cannot be ignored in the finite-length case. Canonical factorization of
˜
Q proceeds
according to
σ
2

(∆) = bG
−1

S

G
−∗

b

, (3.409)
which is minimized when
b = g(∆) , (3.410)
the top row of the upper-triangular matrix G

. The MMSE is thus obtained by computing Cholesky
factorizations of
˜
Q for all reasonable values of ∆ and then setting b = g(∆). Then
σ
2

(∆) = S
o
(∆) . (3.411)
From previous developments, as the lengths of filters go to infinity, any value of ∆ works and also
S
0
→γ
0
N0
2
/
¯
c
x
to ensure the infinite-length MMSE-DFE solution of Section 3.6.
The feed-forward filter then becomes
w = bR
XY
(∆)R
−1
Y Y
(3.412)
= g(∆)
˜
J

P

_
PP

+
l
SNR
I
_
−1
(3.413)
= g(∆)
˜
J

_
P

P +
l
SNR
I
_
−1
. ¸¸ .
feedforward filter
P

.¸¸.
matched filter
, (3.414)
which can be interpreted as a matched filter followed by a feed-forward filter that becomes 1/G

(D
−∗
)
as its length goes to infinity. However, the result that the feed-forward filter is an anti-causal factor of
the canonical factorization does not follow for finite length. Chapter 5 will find a situation that is an
exact match and for which the feed-forward filter is an inverse of a canonical factor, but this requires the
DFE filters to become periodic in a period equal to the number of taps of the feed-forward filter (plus
an excess bandwidth factor).
The SNR is as always
SNR
MMSE−DFE,U
=
¯
c
x
S
0
(∆)
− 1 . (3.415)
Bias can be removed by scaling the decision-element input by the ratio of SNR
MMSE−DFE
/SNR
MMSE−DFE,U
,
thus increasing its variance by (SNR
MMSE−DFE
/SNR
MMSE−DFE,U
)
2
.
Finite-length Noise-Predictive DFE
Problem 3.8 introduces the “noise-predictive” form of the DFE, which is repeated here in Figure 3.43.
In this figure, the error sequence is feedback instead of decisions. The filter essentially tries to predict
the noise in the feedforward filter output and then cancel this noise, whence the name “noise-predictive”
DFE. Correct solution to the infinite-length MMSE filter problem 3.8 will produce that the filter B(D)
remains equal to the same G(D) found for the infinite-length MMSE-DFE. The filter U(D) becomes the
MMSE-LE so that Z(D) has no ISI, but a strongly correlated (and enhanced) noise. The preditor then
reduces the noise to a white error sequence. If W(D) =
1
γ0·p·G
−∗
(D
−∗
)
of the normal MMSE-DFE, then
it can be shown also (see Problem 3.8) that
U(D) =
W(D)
G(D)
. (3.416)
The MMSE, all SNR’s, and biasing/unbiasing remain the same.
An analogous situation occurs for the finite-length case, and it will be convenient notationally to say
that the number of taps in the feedforward filter u is (N
f
−ν)l. Clearly if ν is fixed as always, then any
number of taps (greater than ν can be investigated without loss of generality with this early notational
abuse. Clearly by abusing N
f
(which is not the number of taps), ANY positive number of taps in u can
220
Figure 3.43: Noise-Predictive DFE
be constructed without loss of generality for any value of ν ≥ 0. In this case, the error signal can be
written as
e
k
= b
k
X
k
(∆) −bZ (3.417)
where
Z
k
=
_
¸
_
z
k
.
.
.
z
k−ν
_
¸
_ . (3.418)
Then,
Z
k
= UY
k
(3.419)
where
U =
_
¸
¸
¸
_
u 0 ... 0
0 u ... 0
.
.
.
.
.
.
.
.
.
.
.
.
0 ... 0 u
_
¸
¸
¸
_
(3.420)
and
Y
k
=
_
¸
_
y
k
.
.
.
y
k−Nf +1
_
¸
_ . (3.421)
By defining the (N
f
)l-tap filter
w

= bU , (3.422)
then the error becomes
e
k
= b
k
X
k
(∆) −wY , (3.423)
which is the same error as in the alternate viewpoint of the finite-length DFE earlier in this section.
Thus, solving for the b and w of the conventional finite-length MMSE-DFE can be followed by the step
of solving Equation (3.422) for U when b and w are known. However, it follows directly then from
Equation (3.401) that
u = R
XY
(∆)R
−1
Y Y
, (3.424)
so following the infinite-case form, u is the finite-length MMSE-LE with (N
f
− ν)l taps. The only
difference from the infinite-length case is the change in length.
221
3.7.6 The Stanford DFE Program
A matlab program has been created, used, and refined by the students of EE379A over a period of many
years. The program has the following call command:
function (SNR, wt) = dfe ( l,p,nff, nbb, delay, Ex, noise);
where the inputs and outputs are listed as
• l = oversampling factor
• p = pulse response, oversampled at l (size)
• nff = number of feed-forward taps
• nbb = number of feedback taps
• delay ≈ delay of system Nff + length of p - nbb ; if delay = -1, then choose the best
delay
• Ex = average energy of signals
• noise = autocorrelation vector (size l*nff) (NOTE: noise is assumed to be stationary).
For white noise, this vector is simply [σ
2
0...0].
• SNR = equalizer SNR, unbiased and in dB
• wt = equalizer coefficients
The student may use this program to a substantial advantage in avoiding tedious matrix calculations.
The program has come to be used throughout the industry to compute/project equalizer performance
(setting nbb=0 also provides a linear equalizer). The reader is cautioned against the use of a number
of rules of thumb (like “DFE SNR is the SNR(f) at the middle of the band”) used by various who
call themselves experts and often over-estimate the DFE performance using such formulas. Difficult
transmission channels may require large numbers of taps and considerable experimentation to find the
best settings.
function [dfseSNR,w_t,opt_delay]=dfsecolorsnr(l,p,nff,nbb,delay,Ex,noise);
% --------------------------------------------------------------
% [dfseSNR,w_t] = dfecolor(l,p,nff,nbb,delay,Ex,noise);
% l = oversampling factor
% p = pulse response, oversampled at l (size)
% nff = number of feed-forward taps
% nbb = number of feedback taps
% delay = delay of system <= nff+length of p - 2 - nbb
% if delay = -1, then choose best delay
% Ex = average energy of signals
% noise = noise autocorrelation vector (size l*nff)
% NOTE: noise is assumed to be stationary
%
% outputs:
% dfseSNR = equalizer SNR, unbiased in dB
% w_t = equalizer coefficients [w -b]
% opt_delay = optimal delay found if delay =-1 option used.
% otherwise, returns delay value passed to function
% created 4/96;
% ---------------------------------------------------------------
size = length(p);
nu = ceil(size/l)-1;
p = [p zeros(1,(nu+1)*l-size)];
222
% error check
if nff<=0
error(’number of feed-forward taps > 0’);
end
if delay > (nff+nu-1-nbb)
error(’delay must be <= (nff+(length of p)-2-nbb)’);
elseif delay < -1
error(’delay must be >= -1’);
elseif delay == -1
delay = 0:1:nff+nu-1-nbb;
end
%form ptmp = [p_0 p_1 ... p_nu] where p_i=[p(i*l) p(i*l-1)... p((i-1)*l+1)
ptmp(1:l,1) = [p(1); zeros(l-1,1)];
for i=1:nu
ptmp(1:l,i+1) = conj((p(i*l+1:-1:(i-1)*l+2))’);
end
%form matrix P, vector channel matrix
P = zeros(nff*l+nbb,nff+nu);
for i=1:nff,
P(((i-1)*l+1):(i*l),i:(i+nu)) = ptmp;
end
%precompute Rn matrix - constant for all delays
Rn = zeros(nff*l+nbb);
Rn(1:nff*l,1:nff*l) = toeplitz(noise);
dfseSNR = -100;
P_init = P;
%loop over all possible delays
for d = delay,
P = P_init;
P(nff*l+1:nff*l+nbb,d+2:d+1+nbb) = eye(nbb);
%P
temp= zeros(1,nff+nu);
temp(d+1)=1;
%construct matrices
Ry = Ex*P*P’ + Rn;
Rxy = Ex*temp*P’;
new_w_t = Rxy*inv(Ry);
sigma_dfse = Ex - real(new_w_t*Rxy’);
new_dfseSNR = 10*log10(Ex/sigma_dfse - 1);
%save setting of this delay if best performance so far
if new_dfseSNR >= dfseSNR
w_t = new_w_t;
dfseSNR = new_dfseSNR;
opt_delay = d;
end
end
223
3.7.7 Error Propagation in the DFE
To this point, this chapter has ignored the effect of decision errors in the feedback section of the DFE. At
low error rates, say 10
−5
or below, this is reasonable. There is, however, an accurate way to compute the
effect of decision feedback errors, although enormous amounts of computational power may be necessary
(much more than that required to simulate the DFE and measure error rate increase). At high error
rates of 10
−3
and above, error propagation can lead to several dB of loss. In coded systems (see Chapters
7 and 8), the inner channel (DFE) may have an error rate that is unacceptably high that is later reduced
in a decoder for the applied code. Thus, it is common for low decision-error rates to occur in a coded
DFE system. The Precoders of Section 3.8 can be used to eliminate the error propagation problem, but
this requires that the channel be known in the transmitter – which is not always possible, especially in
transmission systems with significant channel variation, i.e, digital mobile telephony. An analysis for
the infinite-length equalizer has not yet been invented, thus the analysis here applies only to FIR DFE’s
with finite N
b
.
By again denoting the equalized pulse response as v
k
(after removal of any bias), and assuming that
any error signal, e
k
, in the DFE output can is uncorrelated with the symbol of interest, x
k
, this error
signal can be decomposed into 4 constituent signals
1. precursor ISI - e
pr,k
=

−1
i=−∞
v
i
x
k−i
2. postcursor ISI - e
po,k
=


i=Nb+1
v
i
x
k−i
3. filtered noise - e
n,k
=


i=−∞
w
i
n
k−i
4. feedback errors - e
f,k
=

Nb
i=1
v
i
(x
k−i
− ˆ x
k−i
)
The sum of the first 3 signals constitute the MMSE with power σ
2
MMSE−DFE
, which can be computed
according to previous results. The last signal is often written
e
f,k
=
Nb

i=1
v
i

k−i
(3.425)
where
k

= x
k
−ˆ x
k
. This last distortion component has a discrete distribution with (2M−1)
Nb
possible
points.
The error event vector,
k
, is

k

= [
k−Nb+1

k−Nb+2
...
k
] (3.426)
Given
k−1
, there are only 2M − 1 possible values for
k
. Equivalently, the evolution of an error event,
can be described by a finite-state machine with (2M − 1)
Nb
states, each corresponding to one of the
(2M − 1)
Nb
possible length-N
b
error events. Such a finite state machine is shown in Figure 3.44. The
probability that
k
takes on a specific value, or that the state transition diagramgoes to the corresponding
state at time k, is (denoting the corresponding new entry in
k
as
k
)
P

k
/k−1
= Q
_
d
min
− [e
f,k
(
k−1
)[

MMSE-DFE
_
. (3.427)
There are 2M−1 such values for each of the (2M−1)
Nb
states. Υ denotes a square (2M−1)
Nb
(2M−
1)
Nb
matrix of transition probabilities where the i, j
th
element is the probability of entering state i, given
the DFE is in state j. The column sums are thus all unity. For a matrix of non-negative entries, there is
a famous “Peron-Frobenious” lemma from linear algebra that states that there is a unique eigenvector
ρ of all nonnegative entries that satisfies the equation
ρ = Υρ . (3.428)
The solution ρ is called the stationary state distribution or Markov distribution for the state
transition diagram. The i
th
entry, ρ
i
, is the steady-state probability of being in state i. Of course,
224
Figure 3.44: Finite State Machine for N
b
= 2 and M = 2 in evaluating DFE Error Propagation

(2M−1)
N
b
i=1
ρ
i
= 1. By denoting the set of states for which
k
,= 0 as c, one determines the probability
of error as
P
e
=

i∈E
ρ
i
. (3.429)
Thus, the DFE will have an accurate estimate of the error probability in the presence of error propagation.
A larger state-transition diagram, corresponding to the explicit consideration of residual ISI as a discrete
probability mass function, would yield a yet more accurate estimate of the error probability for the
equalized channel - however, the relatively large magnitude of the error propagation samples usually
makes their explicit consideration more important than the (usually) much smaller residual ISI.
The number of states can be very large for reasonable values of M and N
b
, so that the calculation of
the stationary distribution ρ could exceed the computation required for a direct measurement of SNR
with a DFE simulation. There are a number of methods that can be used to reduce the number of states
in the finite-state machine, most of which will reduce the accuracy of the probability of error estimate.
In the usual case, the constellation for x
k
is symmetric with respect to the origin, and there is
essentially no difference between
k
and −
k
, so that the analysis may merge the two corresponding
states and only consider one of the error vectors. This can be done for almost half
15
the states in
the state transition diagram, leading to a new state transition diagram with M
Nb
states. Further
analysis then proceeds as above, finding the stationary distribution and adding over states in c. There
is essentially no difference in this P
e
estimate with respect to the one estimated using all (2M − 1)
Nb
states; however, the number of states can remain unacceptably large.
At a loss in P
e
accuracy, we may ignore error magnitudes and signs, and compute error statistics for
binary error event vectors, which we now denote ε = [ε
1
, ..., ε
Nb
], of the type (for N
b
= 3)
[ 0 0 0 ] , [ 0 1 0 ] , [ 1 0 0 ] , [ 1 0 1 ] . (3.430)
This reduces the number of states to 2
Nb
, but unfortunately the state transition probabilities no longer
depend only on the previous state. Thus, we must try to find an upper bound on these probabilities that
depends only on the previous states. In so doing, the sum of the stationary probabilities corresponding
15
There is no reduction for zero entries.
225
to states in c will also upper bound the probability of error. c corresponds to those states with a nonzero
entry in the first position (the “odd” states, ε
1
= 1). To get an upperbound on each of the transition
probabilities we write
P
ε1,k=1/εk−1
=

{i | ε1,k(i)is allowed transition from εk−1
P
ε1,k=1/εk−1,ε1,k(i)
P
ε1,k(i)
¦ (3.431)

i | ε1,k(i)is allowed transition from εk−1
max
ε1,k(i)
P
ε1,k=1/εk−1,ε1,k(i)
P
ε1,k(i)
(3.432)
= max
ε1,k(i)
P
ε1,k=1/εk−1,ε1,k(i)
. (3.433)
Explicit computation of the maximum probability in (3.433) occurs by noting that this error probability
corresponds to a worst-case signal offset of
δ
max
(ε) = (M − 1)d
Nb

i=1
[v
i

1,k−i
, (3.434)
which the reader will note is analogous to peak distortion (the distortion is understood to be the worst
of the two QAM dimensions, which are assumed to be rotated so that d
min
lies along either or both
of the dimensions ). As long as this quantity is less than the minimum distance between constellation
points, the corresponding error probability is then upper bounded as
P
ek/εk−1
≤ Q
_
_
d
min
2
− δ
max
(ε)
_
σ
2
MMSE−DFE
_
_
. (3.435)
Now, with the desired state-dependent (only) transition probabilities, the upper bound for P
e
with error
propagation is
P
e

ε
Q
_
_
d
min
2
− δ
max
(ε)
_
σ
2
MMSE−DFE
_
_
P
ε
. (3.436)
Even in this case, the number of states 2
Nb
can be too large.
A further reduction to N
b
+ 1 states is possible, by grouping the 2
Nb
states into groups that are
classified only by the number of leading in ε
k−1
; thus, state i = 1 corresponds to any state of the form
[0 ε], while state i = 2 corresponds to [0 0 ε ...], etc. The upperbound on probability of error for
transitions into any state, i, then uses a δ
max
(i) given by
δ
max
(i) = (M −1)d
Nb

i=i+1
[v
i
[ , (3.437)
and δ
max
(N
b
) = 0.
Finally, a trivial bound that corresponds to noting that after an error is made, we have M
Nb
− 1
possible following error event vectors that can correspond to error propagation (only the all zeros error
event vector corresponds to no additional errors within the time-span of the feedback path). The
probability of occurrence of these error event vectors is each no greater than the initial error probability,
so they can all be considered as nearest neighbors. Thus adding these to the original probability of the
first error,
P
e
(errorprop) ≤ M
Nb
P
e
(first) . (3.438)
It should be obvious that this bound gives useful results only if N
b
is small (that is a probability of error
bound of .5 for i.i.d. input data may be lower than this bound even for reasonable values of M and N
b
).
That is suppose, the first probability of error is 10
−5
, and M = 8 and N
b
= 8, then this last (easily
computed) bound gives P
e
≤ 100 !
226
Figure 3.45: Look-Ahead mitigation of error propagation in DFEs.
3.7.8 Look-Ahead
Figure 3.45 illustrates a “look-ahead” mechanism for reducing error propagation. Instead of using the
decision, M
ν
possible decision vectors are retained. The vector is of dimension ν and can be viewed as
an address A
k−∆−1
to the memory. The possible output for each of the M
ν
is computed and subtracted
from z
U,k−∆
. M
ν
decisions of the symbol-by-symbol detector can then be computed and compared
in terms of the distance from a potential symbol value, namely smallest [E
U,k−∆
[. The smallest such
distance is used to select the decision for ˆ x
k−∆
. This method is called “look-ahead” decoding basically
because all possible previous decisions’ ISI are precomputed and stored, in some sense looking ahead in
the calculation. If M
ν
calculations (or memory locations) is too complex, then the largest ν

< ν taps can
be used (or typically the most recent ν

and the rest subtracted in typical DFE fashion for whatever the
decisions previous to the ν

tap interval in a linear filter. Look-ahead leads to the Maximum Likelihood
Sequence detection (MLSD) methods of Chapter 9 that are no longer symbol-by-symbol based. Look-
ahead methods can never exceed SNR
MMSE−DFE,U
in terms of performance, but can come very close
since error propagation can be very small. (MLSD methods can exceed SNR
MMSE−DFE,U
.)
227
Figure 3.46: The Tomlinson Precoder.
3.8 Precoding
This section discusses solutions to the error-propagation problem of DFE’s. The first is precoding, which
essentially moves the feedback section of the DFE to the transmitter with a minimal (but nonzero)
transmit-power-increase penalty, but with no reduction in DFE SNR. The second approach or partial
response channels (which have trivial precoders) has no transmit power penalty, but may have an SNR
loss in the DFE because the feedback section is fixed to a desirable preset value for B(D) rather than
optimized value. This preset value (usually with all integer coefficients) leads to simplification of the
ZF-DFE structure.
3.8.1 The Tomlinson Precoder
Error propagation in the DFE can be a major concern in practical application of this receiver structure,
especially if constellation-expanding codes, or convolutional codes (see Chapter 10), are used in concate-
nation with the DFE (because the error rate on the inner DFE is lower (worse) prior to the decoder).
Error propagation is the result of an incorrect decision in the feedback section of the DFE that produces
additional errors that would not have occurred if that first decision had been correct.
The Tomlinson Precoder (TPC), more recently known as a Tomlinson-Harashima Precoder, is
a device used to eliminate error propagation and is shown in Figure 3.46. Figure 3.46(a) illustrates
the case for real one-dimensional signals, while Figure 3.46(b) illustrates a generalization for complex
signals. In the second complex case, the two real sums and two one-dimensional modulo operators can
228
Figure 3.47: Illustration of modulo arithmetic operator.
be generalized to a two-dimensional modulo where arithmetic is modulo a two-dimensional region that
tesselates two-dimensional space (for example, a hexagon, or a square).
The Tomlinson precoder appears in the transmitter as a preprocessor to the modulator. The Tom-
linson Precoder maps the data symbol x
k
into another data symbol x

k
, which is in turn applied to
the modulator (not shown in Figure 3.46). The basic idea is to move the DFE feedback section to the
transmitter where decision errors are impossible. However, straightforward moving of the filter 1/B(D)
to the transmitter could result in significant transmit-power increase. To prevent most of this power
increase, modulo arithmetic is employed to bound the value of x

k
:
Definition 3.8.1 (Modulo Operator) The modulo operator Γ
M
(x) is a nonlinear func-
tion, defined on an M-ary (PAM or QAM square) input constellation with uniform spacing
d, such that
Γ
M
(x) = x − Md¸
x +
Md
2
Md
| (3.439)
where ¸y| means the largest integer that is less than or equal to y. Γ
M
(x) need not be an
integer. This text denotes modulo M addition and subtraction by ⊕
M
and ¸
M
respectively.
That is
x ⊕
M
y

= Γ
M
[x +y] (3.440)
and
x ¸
M
y

= Γ
M
[x −y] . (3.441)
For complex QAM, each dimension is treated modulo

M independently.
Figure 3.47 illustrates modulo arithmetic for M = 4 PAM signals with d = 2. The following lemma
notes that the modulo operation distributes over addition:
Lemma 3.8.1 (Distribution of Γ
M
(x) over addition) The modulo operator can be dis-
tributed over a sum in the following manner:
Γ
M
[x + y] = Γ
M
(x) ⊕
M
Γ
M
(y) (3.442)
Γ
M
[x − y] = Γ
M
(x) ¸
M
Γ
M
(y) . (3.443)
229
The proof is trivial.
The Tomlinson Precoder generates an internal signal
˜ x
k
= x
k

i=1
b
i
x

k−i
(3.444)
where
x

k
= Γ
M
[˜ x
k
] = Γ
M
_
x
k

i=1
b
i
x

k−i
_
. (3.445)
The scaled-by-SNR
MMSE−DFE
/SNR
MMSE−DFE,U
output of the MS-WMF in the receiver is an optimal
unbiased MMSE approximation to X(D) G
U
(D). That is
E
_
SNR
MMSE−DFE
SNR
MMSE−DFE,U
z
k
/[x
k
x
k−1
, ..., ]
_
=

i=0
g
U,i
x
k−i
. (3.446)
Thus, B(D) = G
U
(D). From Equation (3.249) with x

(D) as the new input, the scaled feedforward
filter output with the Tomlinson precoder is
z
U,k
=
_
x

k
+

i=1
g
U,i
x

k−i
_
+ e
U,k
. (3.447)
Γ
M
[z
U,k
] is determined as
Γ
M
[z
U,k
] = Γ
M

M
(x
k

i=1
g
U,i
x

k−i
) +

i=1
g
U,i
x

k−i
+ e
U,k
] (3.448)
= Γ
M
[x
k

i=1
g
U,i
x

k−i
+

i=1
g
U,i
x

k−i
+e
U,k
] (3.449)
= Γ
M
[x
k
+e
U,k
] (3.450)
= x
k

M
Γ
M
[e
U,k
] (3.451)
= x
k
+ e

U,k
. (3.452)
Since the probability that the error e
U,k
being larger than Md/2 in magnitude is small in a well-designed
system, one can assume that the error sequence is of the same distribution and correlation properties after
the modulo operation. Thus, the Tomlinson Precoder has allowed reproduction of the input sequence
at the (scaled) MS-WMF output, without ISI. The original Tomlinson work was done for the ZF-DFE,
which is a special case of the theory here, with G
U
(D) = P
c
(D). The receiver corresponding to Tomlinson
precoding is shown in Figure 3.48.
The noise power at the feed-forward output is thus almost exactly the same as that of the cor-
responding MMSE-DFE and with no error progagation because there is no longer any need for the
feedback section of the DFE. As stated here without proof, there is only a small price to pay in increased
transmitter power when the TPC is used.
Theorem 3.8.1 (Tomlinson Precoder Output) The Tomlinson Precoder output, when
the input is an i.i.d. sequence, is also approximately i.i.d., and futhermore the output sequence
is approximately uniform in distribution over the interval [−Md/2, Md/2).
There is no explicit proof of this theorem for finite M, although it can be proved exactly as M → ∞.
This proof notes that the unbiased and biased receivers are identical as M →∞, because the SNR must
also be infinite. Then, the modulo element is not really necessary, and the sequence x

k
can be shown to
be equivalent to a prediction-error or “innovations” sequence, which is known in the estimation literature
to be i.i.d. The i.i.d. part of the theorem appears to be valid for almost any M. The distribution and
autocorrelation properties of the TPC, in closed form, remain an unsolved problem at present.
230
Figure 3.48: Receiver for Tomlinson precoded MMSE-DFE implementation.
Using the uniform distribution assumption, the increase in transmit power for the TPC is from the
nominal value of
(M
2
−1)d
2
12
to the value for a continuous uniform random variable over the output interval
[−Md/2, Md/2), which is
M
2
d
2
12
, leading to an input power increase of
M
2
M
2
−1
(3.453)
for PAM and correspondingly
M
M − 1
(3.454)
for QAM. When the input is not a square constellation, the Tomlinson Precoder Power loss is usually
larger, but never more than a few dB. The number of nearest neighbors also increases to
¯
N
e
= 2 for all
constellations. These losses can be eliminated (and actually a gain is possible) using techniques called
Trellis Precoding and/or shell mapping (See Section 10.6).
Laroia or Flexible Precoder
Figure 3.49 shows the Laroia precoder, which is a variation on Tomlinson precoding introduced by
Rajiv Laroia, mainly to reduce the transmit-power loss of the Tomlinson precoder. The Laroia precoder
largely preserves the shape of the transmitted constellation. The equivalent circuit is also shown in
Figure 3.49 where the input is considered to be the difference between the actual input symbol value x
k
and the “decision” output λ
k
. The decision device finds the closest point in the infinite extension of the
constellation
16
. The extension of the constellation is the set of points that continue to be spaced by d
min
from the points on the edges of the original constellation and along the same dimensions as the original
constellation. m
k
is therefore a small error signal that is uniform in distribution over (−d/2, d/2), thus
having variance d
2
/12 << c
x
.
Because of the combination of channel and feedforward equalizer filters, the feedforward filter output
is
Z
U
(D) = [X(D) + M(D)] G
U
(D) +E
U
(D) = X(D) − λ(D) + E
U
(D) . (3.455)
Processing Z
U
(D) by a decision operation leaves X(D)−λ(D), which essentially is the decision. However,
to recover the input sequence X(D), the receiver forms
Z

(D) =
X(D) − λ(D)
G
U
(D)
= X(D) +M(D) . (3.456)
Since M(D) is uniform and has magnitude always less than d/2, then M(D) is removed by a second
truncation (which is not really a decision, but operates using essentially the same logic as the first
decision).
16
The maximum range of such an infinite extension is actually the sum

ν
i=1
| or g
U,i
| · |
¯
Ree(D) or xmax|.
231
Figure 3.49: Flexible Precoding.
232
Figure 3.50: Flexible Precoding Example.
Example and error propagation in flexible precoding
EXAMPLE 3.8.1 (1 +.9D
−1
system with flexible precoding) The flexible precoder and
corresponding receiver for the 1 + .9D
−1
example appear in Figure 3.50. The bias removal
factor has been absorbed into all filters (7.85/6.85 multiples .633 to get .725 in feedback
section, and multiplies .9469 to get 1.085 in feedforward section). The decision device is
binary in this case since binary antipodal transmission is used. Note the IIR filter in the
receiver. If an error is made in the first SBS in the receiver, then the magnitude of the error
is 2 = [ +1 −(−1)[ = [ −1−(+1)[. Such an error is like an impulse of magnitude 2 added to
the input of the IIR filter, which has impulse response (−.725)
k
u
k
, producing a contribution
to the correct sequence of 2 (−.725)
k
u
k
. This produces an additional 1 error half the time
and an additional 2 errors 1/4 the time. Longer strings of errors will not occur. Thus, the
bit and symbol error rate in this case increase by a factor of 1.5. (less than .1 dB loss).
A minor point is that the first decision in the receiver has an increased nearest neighbor
coefficient of N
e
= 1.5.
Generally speaking, when an error is made in the receiver for the flexible precoder,
(x
k
−λ
k
) − decision(x
k
− λ
k
) = δ
k
(3.457)
where could be complex for QAM. Then the designer needs to compute for each type of such error, its
probability of occurence and then investigate the string (with (1/b)
k
being the impulse response of the
post-first-decision filter in the receiver)
[ (1/b)
k
[ << d
min
/2 . (3.458)
For some k, this relation will hold and that is the maximum burst length possible for the particular type
of error. Given the filter b is monic and minimum phase, this string should not be too long as long as the
233
roots are not too close to the unit circle. The single error may cause additional errors, but none of these
additional errors in turn cause yet more errors (unlike a DFE where second and subsequent errors can
lead to some small probability of infinite-length bursts), thus the probaility of an infinitely long burst
is zero for the flexible precoder situation (or indeed any burst longer than the k that solves the above
equation.
3.8.2 Partial Response Channel Models
Partial-response methods are a special case of precoding design where the ISI is forced to some known
well-defined pattern. Receiver detectors are then designed for such partial-response channels directly,
exploiting the known nature of the ISI rather than attempting to eliminate this ISI.
Classic uses of partial response abound in data transmission, some of which found very early use.
For instance, the earliest use of telephone wires for data transmission inevitably found large intersymbol
interference because of a transformer that was used to isolate DC currents at one end of a phone line
from those at the other (see Prob 3.35). The transformer did not pass low frequencies (i.e., DC), thus
inevitably leading to non-Nyquist
17
pulse shapes even if the phone line otherwise introduced no signficant
intersymbol interference. Equalization may be too complex or otherwise undesirable as a solution, so
a receiver can use a detector design that instead presumes the presence of a known fixed ISI. Another
example is magnetic recording (see Problem 3.36) where only flux changes on a recording surface can be
sensed by a read-head, and thus D.C. will not pass through the “read channel,” again inevitably leading
to ISI. Straightforward equalization is often too expensive at the very high speeds of magnetic-disk
recording systems.
Partial-Response (PR) channels have non-Nyquist pulse responses – that is, PR channels allow in-
tersymbol interference – over a few (and finite number of) sampling periods. A unit-valued sample of
the response occurs at time zero and the remaining non-zero response samples at subsequent sampling
times, thus the name “partial response.” The study of PR channels is facilitated by mathematically
presuming the tacit existence of a whitened matched filter, as will be described shortly. Then, a number
of common PR channels can be easily addressed.
Figure 3.51(a) illustrates the whitened-matched-filter. The minimum-phase equivalent of Figure
3.51(b) exists if Q(D) = η
0
H(D) H

(D
−∗
) is factorizable.
18
This chapter focuses on the discrete-time
channel and presumes the WMF’s presence without explicitly showing or considering it. The signal
output of the discrete-time channel is
y
k
=

m=−∞
h
m
x
k−m
+n
k
, (3.459)
or
Y (D) = H(D) X(D) + N(D) , (3.460)
where n
k
is sampled Gaussian noise with autocorrelation ¯ r
nn,k
= |p|
−2
η
−1
0

N0
2
δ
k
or
¯
R
nn
(D) =
|p|
−2
η
−1
0

N0
2
when the whitened-matched filter is used, and just denoted σ
2
pr
otherwise. In both cases,
this noise is exactly AWGN with mean-square sample value σ
2
pr
, which is conveniently abbreviated σ
2
for the duration of this chapter.
Definition 3.8.2 (Partial-Response Channel) A partial-response (PR) channel is any
discrete-time channel with input/output relation described in (3.459) or (3.460) that also sat-
isfies the following properties:
1. h
k
is finite-length and causal; h
k
= 0 ∀ k < 0 or k > ν, where ν < ∞ and is often
called the constraint length of the partial-response channel,
2. h
k
is monic; h
0
= 1,
3. h
k
is minimum phase; H(D) has all ν roots on or outside the unit circle,
17
That is, they do not satisfy Nyquist’s condition for nonzero ISI - See Chapter 3, Section 3.
18
H(D) = Pc(D) in Section 3.6.3 on ZF-DFE.
234
Figure 3.51: Minimum Phase Equivalent Channel.
;’
4. h
k
has all integer coefficients; h
k
∈ Z ∀ k ,= 0 (where Z denotes the set of integers).
More generally, a discrete finite-length channel satisfying all properties above except the last
restriction to all integer coefficients is known as a controlled intersymbol-interference
channel.
Controlled ISI and PR channels are a special subcase of all ISI channels, for which h
k
= 0 ∀ k > ν. ν is
the constraint length of the controlled ISI channel. Thus, effectively, the channel can be modeled as
an FIR filter, and h
k
is the minimum-phase equivalent of the sampled pulse response of that FIR filter.
The constraint length defines the span in time νT of the non-zero samples of this FIR channel model.
The controlled ISI polynomial (D-transform)
19
for the channel simplifies to
H(D)

=
ν

m=0
h
m
D
m
, (3.461)
where H(D) is always monic, causal, minimum-phase, and an all-zero (FIR) polynomial. If the receiver
processes the channel output of an ISI-channel with the same whitened-matched filter that occurs in
the ZF-DFE of Section 3.6, and if P
c
(D) (the resulting discrete-time minimum-phase equivalent channel
polynomial when it exists) is of finite degree ν, then the channel is a controlled intersymbol interference
channel with H(D) = P
c
(D) and σ
2
=
N0
2
|p|
−2
η
−1
0
. Any controlled intersymbol-interference channel
is in the form that the Tomlinson Precoder of Section 3.8.1 could be used to implement symbol-by-
symbol detection on the channel output. As noted in Section 3.8.1, a (usually small) transmit symbol
energy increase occurs when the Tomlinson Precoder is used. Section 3.8.4 shows that this loss can be
avoided for the special class of polynomials H(D) that are partial-response channels.
19
The D-Transform of FIR channels (ν < ∞) is often called the “channel polynomial,” rather than its “D-Transform”
in partial-response theory. This text uses these two terms interchangeably.
235
Figure 3.52: Simple DFE for some partial-response channels.
Equalizers for the Partial-response channel.
This small subsection serves to simplify and review equalization structure on the partial-response channel.
The combination of integer coefficients and minimum-phase constraints of partial-response channels
allows a very simple implementation of the ZF-DFE as in Figure 3.52. Since the PR channel is already
discrete time, no sampling is necessary and the minimum-phase characteristic of H(D) causes the WMF
of the ZF-DFE to simplify to a combined transfer function of 1 (that is, there is no feedforward sec-
tion because the channel is already minimum phase). The ZF-DFE is then just the feedback section
shown, which easily consists of the feedback coefficients −h
m
for the delayed decisions corresponding to
ˆ x
k−m
, m = 1, ..., ν. The loss with respect to the matched filter bound is trivially 1/|H|
2
, which is easily
computed as 3 dB for H(D) = 1 ±D
ν
(any finite ν) with simple ZF-DFE operation z
k
= y
k
− ˆ x
k−ν
and
6 dB for H(D) = 1 + D −D
2
−D
3
with simple ZF-DFE operation z
k
= y
k
− ˆ x
k−1
+ ˆ x
k−2
+ ˆ x
k−3
.
The ZF-LE does not exist if the channel has a zero on the unit circle, which all partial-response
channels do. A MMSE-LE could be expected to have signficant noise enhancement while existing.
The MMSE-DFE will perform slightly better than the ZF-DFE, but is not so easy to compute as the
simple DFE. Tomlinson or Flexible precoding could be applied to eliminate error propagation for a small
increase in transmit power (
M
2
M
2
−1
).
3.8.3 Classes of Partial Response
A particularly important and widely used class of partial response channels are those with H(D) given
by
H(D) = (1 +D)
l
(1 −D)
n
, (3.462)
where l and n are nonnegative integers.
For illustration, Let l = 1 and n = 0 in (3.462), then
H(D) = 1 + D , (3.463)
236
Figure 3.53: Transfer Characteristic for Duobinary signaling.
which is sometimes called a “duobinary” channel (introduced by Lender in 1960). The Fourier transform
of the duobinary channel is
H(e
−ωT
) = H(D)[
D=e
−ωT = 1 + e
−ωT
= 2e
−ωT/2
cos
_
ωT
2
_
. (3.464)
The transfer function in (3.464) has a notch at the Nyquist Frequency and is generally “lowpass” in
shape, as is shown in Figure 3.53. A discrete-time ZFE operating on this channel would produce infinite
noise enhancement. If SNR
MFB
= 16 dB for this channel, then
Q(e
−ωT
) +
1
SNR
MFB
= (1 + 1/40) + cos ωT . (3.465)
The MMSE-LE will have performance (observe that
N0
2
= .05)
σ
2
MMSE−LE
=
^
0
2
T

_ π
T

π
T
1
|p|
2
(1.025 + cos ωT)
dω =
^
0
2
1/2

1.025
2
−1
2
= 2.22
^
0
2
, (3.466)
The SNR
MMSE−LE,U
is easily computed to be 8 (9dB), so the equalizer loss is 7 dB in this case. For
the MMSE-DFE, Q(D) +
1
SNRMFB
=
1.025
1.64
(1 + .8D)(1 + .8D
−1
), so that γ
0
= 1.025/1.64 = .625, and
thus γ
MMSE-DFE
= 2.2dB. For the ZF-DFE, η
0
=
1
2
, and thus γ
ZF-DFE
= 3dB.
It is possible to achieve MFB performance on this channel with complexity far less than any equalizer
studied earlier in this chapter, as will be shown in Chapter 9. It is also possible to use a precoder with
no transmit power increase to eliminate the error-propagation-prone feedback section of the ZF-DFE.
There are several specific channels that are used in practice for partial-response detection:
EXAMPLE 3.8.2 (Duobinary 1 + D) The duobinary channel (as we have already seen)
has
H(D) = 1 + D . (3.467)
237
The frequency response was already plotted in Figure 3.53. This response goes to zero at the
Nyquist Frequency, thus modeling a lowpass-like channel. For a binary input of x
k
= ±1,
the channel output (with zero noise) takes on values ±2 with probability 1/4 each and 0 with
probability 1/2. In general, for M-level inputs (±1 ±3 ±5 ... ±(M −1)), there are 2M−1
possible output levels, −2M + 2, ..., 0, ..., 2M − 2. These output values are all possible
sums of pairs of input symbols.
EXAMPLE 3.8.3 (DC Notch 1 −D) The DC Notch channel has
H(D) = 1 − D , (3.468)
so that l = 0 and n = 1 in (3.462). The frequency response is
H(e
−ωT
) = 1 − e
−ωT
= 2e
−ω
T
2
sin
ωT
2
. (3.469)
The response goes to zero at the DC (ω = 0), thus modeling a highpass-like channel. For
a binary input of x
k
= ±1, the channel output (with zero noise) takes on values ±2 with
probability 1/4 each and 0 with probability 1/2. In general for M-level inputs (±1 ± 3 ±
5 ... ± (M − 1)), there are 2M − 1 possible output levels, −2M + 2, ..., 0, ..., 2M − 2.
When the 1 − D shaping is imposed in the modulator itself, rather than by a channel, the
corresponding modulation is known as AMI (Alternate Mark Inversion) if a differential
encoder is also used as shown later in this section. AMI modulation prevents “charge” (DC)
from accumulating and is sometimes also called “bipolar coding,” although the use of the
latter term is often confusing because bipolar transmission may have other meanings for some
communications engineers. AMI coding, and closely related methods are used in multiplexed
T1 (1.544 Mbps DS1 or “ANSI T1.403”) and E1 (2.048 Mbps or “ITU-T G.703”) speed
digital data transmission on twisted pairs or coaxial links. These signals were once prevalent
in telephone-company non-fiber central-office transmission of data between switch elements.
EXAMPLE 3.8.4 (Modified Duobinary 1 − D
2
) The modified duobinary channel has
H(D) = 1 − D
2
= (1 + D)(1 −D) , (3.470)
so l = n = 1 in (3.462). Modified Duobinary is sometimes also called “Partial Response
Class IV” or PR4 or PRIV in the literature. The frequency response is
H(e
−ωT
) = 1 −e
−ω2T
= 2e
−ωT
sin(ωT) . (3.471)
The response goes to zero at the DC (ω = 0) and at the Nyquist frequency (ω = π/T),
thus modeling a bandpass-like channel. For a binary input of x
k
= ±1, the channel output
(with zero noise) takes on values ±2 with probability 1/4 each and 0 with probability 1/2.
In general, for M-level inputs (±1 ± 3 ± 5 ... ± (M − 1)), there are 2M − 1 possible
output levels. Modified duobinary is equivalent to two interleaved 1 − D channels, each
independently acting on the inputs corresponding to even (odd) time samples, respectively.
Many commercial disk drives use PR4.
EXAMPLE 3.8.5 (Extended Partial Response 4 and 6 (1 +D)
l
(1 −D)
n
) The EPR4
channel has l = 2 and n = 1 or
H(D) = (1 + D)
2
(1 − D) = 1 +D − D
2
−D
3
. (3.472)
This channel is called EPR4 because it has 4 non-zero samples (Thapar). The frequency
response is
H(e
−ωT
) = (1 + e
−ωT
)
2
(1 −e
−ωT
) . (3.473)
238
Figure 3.54: Transfer Characteristic for EPR4 and EPR6
The EPR6 channel has l = 4 and n = 1 (6 nonzero samples)
H(D) = (1 +D)
4
(1 −D) = 1 + 3D + 2D
2
−2D
3
− 3D
4
−D
5
. (3.474)
The frequency response is
H(e
−ωT
) = (1 + e
−ωT
)
4
(1 −e
−ωT
) . (3.475)
These 2 channels, along with PR4, are often used to model disk storage channels in magnetic
disk or tape recording. The response goes to zero at DC (ω = 0) and at the Nyquist frequency
in both EPR4 and EPR6, thus modeling bandpass-like channels. The magnitude of these two
frequency characteristics are shown in Figure 3.54. These are increasingly used in commercial
disk storage read detectors.
The higher the l, the more “lowpass” in nature that the EPR channel becomes, and the more appropriate
as bit density increases on any given disk.
For partial-response channels, the use of Tomlinson Precoding permits symbol-by-symbol detection,
but also incurs an M
2
/(M
2
− 1) signal energy loss for PAM (and M/(M − 1) for QAM). A simpler
method for PR channels, that also has no transmit energy penalty, is known simply as a “precoder.”
3.8.4 Simple Precoding
The simplest form of precoding for the duobinary, DC-notch, and modified duobinary partial-response
channels is the so-called “differential encoder.” The message at time k, m
k
, is assumed to take on
values m = 0, 1, ..., M − 1. The differential encoder for the case of M = 2 is most simply described as
the device that observes the input bit stream, and changes its output if the input is 1 and repeats the
last output if the input is 0. Thus, the differential encoder input, m
k
, represents the difference (or sum)
between adjacent differential encoder output ( ¯ m
k
and ¯ m
k−1
) messages
20
:
20
This operation is also very useful even on channels without ISI, as an unknown inversion in the channel (for instance,
an odd number of amplifiers) will cause all bits to be in error if (differential) precoding is not used.
239
Figure 3.55: Precoded Partial Response Channel.
Definition 3.8.3 (Differential Encoder) Differential encoders for PAM or QAM mod-
ulation obey one of the two following relationships:
¯ m
k
= m
k
¸ ¯ m
k−1
(3.476)
¯ m
k
= m
k
⊕ ¯ m
k−1
, (3.477)
where ¸ represents subtraction modulo M (and ⊕ represents addition modulo M). For SQ
QAM, the modulo addition and subtraction are performed independently on each of the two
dimensions, with

M replacing M. (A differential phase encoder is also often used for QAM
and is discussed later in this section.)
A differential encoder is shown on the left side of Figure 3.55. As an example if M = 4 the corresponding
inputs and outputs are given in the following table:
m
k
- 3 1 0 2 1 0 3
¯ m
k
0 3 2 2 0 1 3 0
k -1 0 1 2 3 4 5 6
With either PAM or QAM constellations, the dimensions of the encoder output ¯ m
k
are converted
into channel input symbols (±
d
2
, ±
3d
2
...) according to
x
k
= [2 ¯ m
k
− (M − 1)]
d
2
. (3.478)
Figure 3.55 illustrates the duobinary channel H(D) = 1 +D, augmented by the precoding operation
of differential encoding as defined in (3.476). The noiseless minimum-phase-equivalent channel output
˜ y
k
is
˜ y
k
= x
k
+x
k−1
= 2( ¯ m
k
+ ¯ m
k−1
)
d
2
− 2(M −1)
d
2
(3.479)
so
21
˜ y
k
d
+ (M −1) = ¯ m
k
+ ¯ m
k−1
(3.480)
21
(•)
M
means the quantity is computed in M-level arithmetic, for instance, (5)
4
= 1. Also note that Γ
M
(x) = (x)
M
,
and therefore ⊕ is different from the ⊕
M
of Section 3.5. The functions (x)
M
and ⊕ have strictly integer inputs and have
possible outputs 0, ..., M −1 only.
240
__
˜ y
k
d
+ (M −1)
__
M
= ¯ m
k
⊕ ¯ m
k−1
(3.481)
__
˜ y
k
d
+ (M −1)
__
M
= m
k
(3.482)
where the last relation (3.482) follows from the definition in (3.476). All operations are integer mod-M.
Equation (3.482) shows that a decision on ˜ y
k
about m
k
can be made without concern for preceding or
succeeding ˜ y
k
, because of the action of the precoder. The decision boundaries are simply the obvious
regions symmetrically placed around each point for a memoryless ML detector of inputs and the decoding
rules for M = 2 and M = 4 are shown in Figure 3.55. In practice, the decoder observes y
k
, not ˜ y
k
, so
that y
k
must first be quantized to the closest noise-free level of y
k
. That is, the decoder first quantizes
˜ y
k
to one of the values of (−2M + 2)(d/2), ..., (2M − 2)(d/2). Then, the minimum distance between
outputs is d
min
= d, so that
d
min
2σpr
=
d
2σpr
, which is 3 dB below the

MFB =

2
σpr
(d/2) for the 1 + D
channel because |p|
2
= 2. (Again, σ
2
pr
= |p|
−2
η
−1
0

N0
2
with a WMF and otherwise is just σ
2
pr
.) This
loss is identical to the loss in the ZF-DFE for this channel. One may consider the feedback section of the
ZF-DFE as having been pushed through the linear channel back to the transmitter, where it becomes
the precoder. With some algebra, one can show that the TPC, while also effective, would produce a
4-level output with 1.3 dB higher average transmit symbol energy for binary inputs. The output levels
are assumed equally likely in determining the decision boundaries (half way between the levels) even
though these levels are not equally likely.
The precoded partial-response system eliminates error propagation, and thus has lower P
e
than the
ZF-DFE. This elimination of error propagation can be understood by investigating the nearest neighbor
coefficient for the precoded situation in general. For the 1 + D channel, the (noiseless) channel output
levels are −(2M −2) − (2M −4)... 0 ... (2M − 4) (2M − 2) with probabilities of occurrence
1
M
2
2
M
2
...
M
M
2
...
2
M
2
1
M
2
, assuming a uniform channel-input distribution. Only the two outer-most levels have one
nearest neighbor, all the rest have 2 nearest neighbors. Thus,
¯
N
e
=
2
M
2
(1) +
M
2
−2
M
2
(2) = 2
_
1 −
1
M
2
_
. (3.483)
For the ZF-DFE, the input to the decision device ˜ z
k
as
˜ z
k
= x
k
+x
k−1
+ n
k
− ˆ x
k−1
, (3.484)
which can be rewritten
˜ z
k
= x
k
+ (x
k−1
− ˆ x
k−1
) +n
k
. (3.485)
Equation 3.485 becomes ˜ z
k
= x
k
+ n
k
if the previous decision was correct on the previous symbol.
However, if the previous decision was incorrect, say +1 was decided (binary case) instead of the correct
−1, then
˜ z
k
= x
k
− 2 +n
k
, (3.486)
which will lead to a next-symbol error immediately following the first almost surely if x
k
= 1 (and no
error almost surely if x
k
= −1). The possibility of ˜ z
k
= x
k
+2 +n
k
is just as likely to occur and follows
an identical analysis with signs reversed. For either case, the probability of a second error propagating
is 1/2. The other half of the time, only 1 error occurs. Half the times that 2 errors occur, a third error
also occurs, and so on, effectively increasing the error coefficient from N
e
= 1 to
N
e
= 2 = 1(
1
2
) + 2(
1
4
) + 3(
1
8
) + 4(
1
16
) + ... =

k=1
k (.5)
k
=
.5
(1 − .5)
2
. (3.487)
(In general, the formula


k=1
k r
k
=
r
(1−r)
2
, r < 1 may be useful in error propagation analysis.) The
error propagation can be worse in multilevel PAM transmission, when the probability of a second error
is (M −1)/M, leading to
N
e
(error prop)
N
e
(no error prop)
= 1
1
M
+ 2
M − 1
M

1
M
+ 3
_
M − 1
M
_
2

1
M
+... (3.488)
241
=

k=1
k
_
M − 1
M
_
k−1

1
M
(3.489)
= M . (3.490)
Precoding eliminates this type of error propagation, although
¯
N
e
increases by a factor
22
of (1 + 1/M)
with respect to the case where no error propagation occurred.
For the ZF-DFE system,
¯
P
e
= 2(M − 1)Q(
d
2σpr
), while for the precoded partial-response system,
¯
P
e
= 2
_
1 −
1
M
2
_
Q(
d
2σpr
). For M ≥ 2, the precoded system always has the same or fewer nearest
neighbors, and the advantage becomes particularly pronounced for large M. Using a rule-of-thumb
that a factor of 2 increase in nearest neighbors is equivalent to an SNR loss of .2 dB (which holds at
reasonable error rates in the 10
−5
to 10
−6
range), the advantage of precoding is almost .2 dB for M = 4.
For M = 8, the advantage is about .6 dB, and for M = 64, almost 1.2 dB. Precoding can be simpler
to implement than a ZF-DFE because the integer partial-response channel coefficients translate readily
into easily realized finite-field operations in the precoder, while they represent full-precision add (and
shift) operations in feedback section of the ZF-DFE.
Precoding the DC Notch or Modified Duobinary Channels
The ¯ m
k
= m
k
¸ ¯ m
k−1
differential encoder works for the 1 + D channel. For the 1 − D channel, the
equivalent precoder is
¯ m
k
= m
k
⊕ ¯ m
k−1
, (3.491)
which is sometimes also called NRZI (non-return-to-zero inverted) precoding, especially by storage-
channel engineers. In the 1 −D case, the channel output is
˜ y
k
= x
k
−x
k−1
= 2( ¯ m
k
− ¯ m
k−1
)
d
2
(3.492)
˜ y
k
d
= ¯ m
k
− ¯ m
k−1
(3.493)
_
˜ y
k
d
_
M
= ¯ m
k
¸ ¯ m
k−1
(3.494)
_
˜ y
k
d
_
M
= m
k
. (3.495)
The minimum distance and number of nearest neighbors are otherwise identical to the 1 + D case just
studied, as is the improvement over the ZF-DFE. The 1 −D
2
case is identical to the 1 −D case, on two
interleaved 1 −D channels at half the rate. The overall precoder for this situation is
¯ m
k
= m
k
⊕ ¯ m
k−2
, (3.496)
and the decision rule is
_
˜ y
k
d
_
M
= m
k
. (3.497)
The combination of precoding with the 1 −D channel is often called “alternate mark inversion (AMI)”
because each successive transmitted “1” bit value causes a nonzero channel output amplitude of polarity
oppositie to the last nonzero channel output amplitude, while a “0” bit always produces a 0 level at the
channel output.
Precoding EPR4
An example of precoding for the extended Partial Response class is EPR4, which has l = 2 and n = 1
in (3.472), or EPR4. Then,
˜ y
k
= x
k
+ x
k−1
−x
k−2
− x
k−3
(3.498)
22
The ratio of 2(1 − 1/M
2
) with precoding to 2(1 − 1/M) for M-ary PAM with no error propagation effects included
242
˜ y
k
= d( ¯ m
k
+ ¯ m
k−1
− ¯ m
k−2
− ¯ m
k−3
) (3.499)
˜ y
k
d
= ¯ m
k
+ ¯ m
k−1
− ¯ m
k−2
− ¯ m
k−3
(3.500)
_
˜ y
k
d
_
M
= ¯ m
k
⊕ ¯ m
k−1
¸ ¯ m
k−2
¸ ¯ m
k−3
(3.501)
_
˜ y
k
d
_
M
= m
k
(3.502)
where the precoder, from (3.502) and (3.501), is
¯ m
k
= m
k
¸ ¯ m
k−1
⊕ ¯ m
k−2
⊕ ¯ m
k−3
. (3.503)
The minimum distance at the channel output is still d in this case, so P
e
≤ N
e
Q(d/2σ
pr
), but the
MFB = (
d
σpr
)
2
, which is 6dB higher. The same 6dB loss that would occur with an error-propagation-free
ZF-DFE on this channel.
For the 1 + D − D
2
− D
3
channel, the (noiseless) channel output levels are −(4M − 4) − (4M −
6)... 0 ... (4M−6) (4M−4). Only the two outer-most levels have one nearest neighbor, all the rest have
2 nearest neighbors. Thus,
¯
N
e
=
2
M
4
(1) +
M
4
−2
M
4
(2) = 2
_
1 −
1
M
4
_
. (3.504)
Thus the probability of error for the precoded system is
¯
P
e
= 2
_
1 −
1
M
4
_
Q(
1
σpr
). The number of nearest
neighbors for the ZF-DFE, due to error propagation, is difficult to compute, but clearly will be worse.
3.8.5 General Precoding
The general partial response precoder can be extrapolated from previous results:
Definition 3.8.4 (The Partial-Response Precoder) The partial-response precoder
for a channel with partial-response polynomial H(D) is defined by
¯ m
k
= m
k
ν

i=1
(−h
i
) ¯ m
k−i
. (3.505)
The notation

means a mod-M summation, and the multiplication can be performed without
ambiguity because the h
i
and ¯ m
k−i
are always integers.
The corresponding memoryless decision at the channel output is
ˆ m
k
=
_
ˆ y
k
d
+
ν

i=0
h
i
_
M − 1
2
_
_
M
. (3.506)
The reader should be aware that while the relationship in (3.505) is general for partial-response channels,
the relationship can often simplify in specific instances, for instance the precoder for “EPR5,” H(D) =
(1 + D)
3
(1 − D) simplifies to ¯ m
k
= m
k
⊕ ¯ m
k−4
when M = 2.
In a slight abuse of notation, engineers often simplify the representation of the precoder by simply
writing it as
T(D) =
1
H(D)
(3.507)
where T(D) is a polynomial in D that is used to describe the “modulo-M” filtering in (3.505). Of course,
this notation is symbolic. Furthermore, one should recognize that the D means unit delay in a finite
field, and is therefore a delay operator only – one cannot compute the Fourier transform by inserting
D = e
−ωT
into T(D); Nevertheless, engineers commonly to refer to the NRZI precoder as a 1/(1 ⊕D)
precoder.
243
Lemma 3.8.2 (Memoryless Decisions for the Partial-Response Precoder) The partial-
response precoder, often abbreviated by T(D) = 1/H(D), enables symbol-by-symbol decoding
on the partial-response channel H(D). An upper bound on the performance of such symbol-
by-symbol decoding is
¯
P
e
≤ 2
_
1 −
1
M
ν+1
_
Q
_
d

pr
_
, (3.508)
where d is the minimum distance of the constellation that is output to the channel.
Proof: The proof follows by simply inserting (3.505) into the expression for ˜ y
k
and simpli-
fying to cancel all terms with h
i
, i > 0. The nearest-neighbor coefficient and d/2σ
pr
follow
trivially from inspection of the output. Adjacent levels can be no closer than d on a partial-
response channel, and if all such adjacent levels are assumed to occur for an upper bound on
probability of error, then the
¯
P
e
bound in (3.508) holds. QED.
3.8.6 Quadrature PR
Differential encoding of the type specified earlier is not often used with QAM systems, because QAM
constellations usually exhibit 90
o
symmetry. Thus a 90
o
offset in carrier recovery would make the
constellation appear exactly as the original constellation. To eliminate the ambiguity, the bit assignment
for QAM constellations usually uses a precoder that uses the four possibilities of the most-significant bits
that represent each symbol to specify a phase rotation of 0
o
, 90
o
, 180
o
, or 270
o
with respect to the last
symbol transmitted. For instance, the sequence 01, 11, 10, 00 would produce (assuming an initial phase
of 0
o
, the sequence of subsequent phases 90
o
, 0
o
, 180
o
, and 180
o
. By comparing adjacent decisions and
their phase difference, these two bits can be resolved without ambiguity even in the presence of unknown
phase shifts of multiples of 90
o
. This type of encoding is known as differential phase encoding and
the remaining bits are assigned to points in large M QAM constellations so that they are the same for
points that are just 90
o
rotations of one another. (Similar methods could easily be derived using 3 or
more bits for constellations with even greater symmetry, like 8PSK.)
Thus the simple precoder for the 1 + D and 1 − D (or 1 − D
2
) given earlier really only is practical
for PAM systems. The following type of precoder is more practical for these channels:
A Quadrature Partial Response (QPR) situation is specifically illustrated in Figure 3.56 for M = 4
or
¯
b = 1 bit per dimension. The previous differential encoder could be applied individually to both
dimensions of this channel, and it could be decoded without error. All previous analysis is correct,
individually, for each dimension. There is, however, one practical problem with this approach: If the
channel were to somehow rotate the phase by ±90
o
(that is, the carrier recovery system locked on the
wrong phase because of the symmetry in the output), then there would be an ambiguity as to which part
was real and which was imaginary. Figure 3.56 illustrates the ambiguity: the two messages (0, 1) = 1
and (1, 0) = 2 commute if the channel has an unknown phase shift of ±90
o
. No ambiguity exists for
either the message (0, 0) = 0 or the message (1, 1) = 3. To eliminate the ambiguity,the precoder encodes
the 1 and 2 signals into a difference between the last 1 (or 2) that was transmitted. This precoder thus
specifies that an input of 1 (or 2) maps to no change with respect to the last input of 1 (or 2), while an
input of 2 maps to a change with respect to the last input of 1 or 2.
A precoding rule that will eliminate the ambiguity is then
Rule 3.8.1 (Complex Precoding for the 1+D Channel) if m
k
= (m
i,k
, m
q,k
) = (0, 0) or (1, 1)
then
¯ m
i,k
= m
i,k
⊕ ¯ m
i,k−1
(3.509)
¯ m
q,k
= m
q,k
⊕ ¯ m
q,k−1
(3.510)
else if m
k
= (m
i,k
, m
q,k
) = (0, 1) or (1, 0), check the last ¯ m = (0, 1) or (1, 0) transmitted,
call it ¯ m
90
, (that is, was ¯ m
90
= 1 or 2 ?). If ¯ m
90
= 1, the precoder leaves m
k
unchanged
prior to differential encoding according to (3.509) and (3.510). The operations ⊕ and ¸ are
the same in binary arithmetic.
244
Figure 3.56: Quadrature Partial Response Example with 1 + D.
245
If ¯ m
90
= 2, then the precoder changes m
k
from 1 to 2 (or from 2 to 1) prior to encoding
according to (3.509) and (3.510).
The corresponding decoding rule is (keeping a similar state ˆ y
90
= 1, 2 at the decoder)
ˆ m
k
=
_
¸
¸
_
¸
¸
_
(0, 0) ˆ y
k
= [ ±2 ± 2 ]
(1, 1) ˆ y
k
= [ 0 0 ]
(0, 1) (ˆ y
k
= [ 0 ±2 ] or ˆ y
k
= [ ±2 0 ]) and

ˆ y
k

ˆ y
90
= 0
(1, 0) (ˆ y
k
= [ 0 ±2 ] or ˆ y
k
= [ ±2 0 ]) and

ˆ y
k

ˆ y
90
,= 0
. (3.511)
The probability of error and minimum distance are the same as was demonstrated earlier for this type
of precoder, which only resolves the 90
o
ambiguity, but is otherwise equivalent to a differential encoder.
There will however be limited error propagation in that one detection error on the 1 or 2 message points
leads to two decoded symbol errors on the decoder output.
246
Figure 3.57: Basic diversity channel model.
3.9 Diversity Equalization
Diversity in transmission is the use of multiple channels from a single message source to several receivers,
as illustrated in Figure 3.57. Optimally, the principles of Chapter 1 apply directly where the conditional
probability distribution p
y/x
of the channel has a larger dimensionality on the channel output y than on
the input, x, N
y
> N
x
. Often, this diversity leads to a lower probability of error for the same message,
mainly because a greater channel-output minimum distance between possible (noiseless) output data
symbols can be achieved with a larger number of channel output dimensions. However, intersymbol
interference between successive transmissions, along with interference between the diversity dimensions,
again can lead to a potentially complex optimum receiver and detector. Thus, equalization again allows
productive use of suboptimal SBS detectors. The equalizers in the case of diversity become matrix
equivalents of those studied in Sections 3.5 - 3.7.
3.9.1 Multiple Received Signals and the RAKE
Figure 3.57 illustrates the basic diversity channel. Channel outputs caused by the same channel input
have labels y
l
(t), l = 0, ..., L− 1. These channel outputs can be created intentionally by retransmission
of the same data symbols at different times and/or (center) frequencies. Spatial diversity often occurs in
wireless transmission where L spatially separated antennas may all receive the same transmitted signal,
but possibly with different filtering and with noises that are at least partially independent.
With each channel output following the model
y
p,l
(t) =

k
x
k
p
l
(t − kT) +n
l
(t) , (3.512)
247
a corresponding L 1 vector channel description is
y
p
(t) =

k
x
k
p(t −kT) +n
p
(t) , (3.513)
where
y
p
(t)

=
_
¸
¸
¸
_
y
0
(t)
y
1
(t)
.
.
.
y
L−1
(t)
_
¸
¸
¸
_
p(t)

=
_
¸
¸
¸
_
p
0
(t)
p
1
(t)
.
.
.
p
L−1
(t)
_
¸
¸
¸
_
and n
p
(t)

=
_
¸
¸
¸
_
n
0
(t)
n
1
(t)
.
.
.
n
L−1
(t)
_
¸
¸
¸
_
. (3.514)
The generalization of an inner product is readily seen to be
< x(t), y(t) >=

i
_

−∞
x
i
(t)y

i
(t)dt . (3.515)
Without loss of generality, the noise can be considered to be white on each of the L diversity channels,
independent of the other diversity channels, and with equal power spectral densities
N0
2
.
23
A single transmission of x
0
corresponds to a vector signal
y
p
(t) = x
0
p(t) + n(t) = x
0
|p| ϕ
p
(t) +n(t) . (3.516)
This situation generalizes slightly that considered in Chapter 1, where matched-filter demodulators
there combined all time instants through integration, a generalization of the inner product’s usual sum
of products. A matched filter in general simply combines the signal components from all dimensions
that have independent noise. Here, that generalization of combination simply needs to include also the
components corresponding to each of the diversity channels so that all signal contributions are summed
to create maximum signal-to-noise ratio. The relative weighting of the different diversity channels is thus
maintained through L unnormalized parallel matched filters each corresponding to one of the diversity
channels. When several copies are combined across several diversity channels or new dimensions (whether
created in frequency, long delays in time, or space), the combination is known as the RAKE matched
filter of Figure 3.58:
Definition 3.9.1 (RAKE matched filter) A RAKE matched filter is a set of parallel
matched filters each operating on one of the diversity channels in a diversity transmission
system that is followed by a summing device as shown in Figure 3.58. Mathematically, the
operation is denoted by
y
p
(t) =
L−1

l=0
p

l
(−t) ∗ y
l
(t) . (3.517)
The RAKE was originally so named by Green and Price in 1958 because of the analogy of the various
matched filters being the “fingers” of a garden rake and the sum corresponding to the collection of the
fingers at the rake’s pole handle, nomenclature thus often being left to the discretion of the inventor,
however unfortunate for posterity. The RAKE is sometimes also called a diversity combiner, although
the latter term also applies to other lower-performance suboptimal combining methods that do not
maximize overall signal to noise strength through matched filter. One structure, often called maximal
combining, applies a matched filter only to the strongest of the L diversity paths to save complexity.
The equivalent channel for this situation is then the channel corresponding to this maximum-strength
individual path. The original RAKE concept was conceived in connection with a spread-spectrum
transmission method that achieves diversity essentially in frequency (but more precisely in a code-
division dimension to be discussed in the appendix of Chapter 5), but the matched filtering implied is
23
In practice, the noises may be correlatedwith each other on different subchannels and not white with covariance matrix
Rn(t) and power spectral density matrix Rn(f) =
N
0
2
· R
1/2
n
(f)R
∗/2
n
(f). By prefiltering the vector channel output by the
matrix filter R
−1/2
n
(f), the noise will be whitened and the noise equivalent matrix channel becomes P(f) →R
−1/2
n
(f)P(f).
Anaylsis with the equivalent channel can then proceed as if the noise where white, independent on the diversity channels,
and of the same variance
N
0
2
on all.
248
Figure 3.58: The basic RAKE matched filter combiner.
easily generalized. Some of those who later studied diversity combining were not aware of the connection
to the RAKE and thus the multiple names for the same structure, although diversity combining is a
more accurate name for the method.
This text also defines r
l
(t) = p
l
(t) ∗ p

l
(−t) and
r(t) =
L−1

l=0
r
l
(t) (3.518)
for an equivalent RAKE-output equivalent channel and the norm
|p|
2
=
L−1

l=0
|p
l
|
2
. (3.519)
Then, the normalized equivalent channel q(t) is defined through
r(t) = |p|
2
q(t) . (3.520)
The sampled RAKE output has D-transform
Y (D) = X(D) |p|
2
Q(D) + N(D) , (3.521)
which is essentially the same as the early channel models used without diversity except for the additional
scale factor of |p|
2
, which also occurs in the noise autcorrelation, which is
¯
R
nn
(D) =
^
0
2
|p|
2
Q(D) . (3.522)
An SNR
MFB
=
¯
E
x
p
2
N
0
2
and all other detector and receiver principles previously developed in this
text now apply directly.
3.9.2 Infinite-length MMSE Equalization Structures
The sampled RAKE output can be scaled by the factor |p|
−1
to obtain a model identical to that found
earlier in Section 3.1 in Equation (3.26) with Q(D) and |p| as defined in Subsection 3.9.1. Thus, the
249
MMSE-DFE, MMSE-LE, and ZF-LE/DFE all follow exactly as in Sections 3.5 -3.7. The matched-filter
bound SNR and each of the equalization structures tend to work better with diversity because |p|
2
is
typically larger on the equivalent channel created by the RAKE. Indeed the RAKE will work better
than any of the individual channels, or any subset of the diversity channels, with each of the equalizer
structures.
Often while one diversity channel has severe characteristics, like an inband notch or poor transmission
characteristic, a second channel is better. Thus, diversity systems tend to be more robust.
EXAMPLE 3.9.1 (Two ISI channels in parallel) Figure 3.59 illustrates two diversity
channels with the same input and different intersymbol interference. The first upper channel
has a sampled time equivalent of 1 + .9D
−1
with noise variance per sample of .181 (and
thus could be the channel consistently examined throughout this Chapter so
¯
c
x
= 1). This
channel is in effect anti-causal (or in reality, the .9 comes first). A second channel has causal
response 1 + .8D with noise variance .164 per sample and is independent of the noise in
the first channel. The ISI effectively spans 3 symbol periods among the two channels at a
common receiver that will decide whether x
k
= ±1 has been transmitted.
The SNR
MFB
for this channel remains SNR
MFB
=
¯
E
x
·p
2
N
0
2
, but it remains to compute
this quantity correctly. First the noise needs to be whitened. While the two noises are
independent, they do not have the same variance per sample, so a pre-whitening matrix is
_
1 0
0
_
.181
.164
_
(3.523)
and so then the energy quantified by |p|
2
is
|p|
2
= |p
1
|
2
+|˜ p
2
|
2
= 1.81 +
_
.181
.164
_
1.64 = 2(1.81) . (3.524)
Then
SNR
MFB
=
1 2(1.81)
.181
= 13 dB . (3.525)
Because of diversity, this channel has a higher potential performance than the single channel
alone. Clearly having a second look at the input through another channel can’t hurt (even
if there is more ISI now). The ISI is characterized as always by Q(D), which in this case is
Q(D) =
1
2(1.81)
_
(1 + .9D
−1
)(1 + .9D) +
.181
.164
(1 + .8D)(1 +.8D
−1
)
_
(3.526)
= .492D + 1 + .492D
−1
(3.527)
= .7082 (1 +.835D) (1 +.835D
−1
) (3.528)
˜
Q(D) = .492D + (1 + 1/20) + .492D
−1
(3.529)
= .587 (1 + .839D) (1 + .839D
−1
) (3.530)
Thus, the SNR of a MMSE-DFE would be
SNR
MMSE−DFE,U
= .7082(20) − 1 = 13.16 (11.15 dB) . (3.531)
The improvement of diversity with respect to a single channel is about 2.8 dB in this case.
The receiver is a MMSE-DFE essentially designed for the 1+.839D ISI channel after adding
the matched-filter outputs. The loss with respect to the MFB is about 1.8 dB.
250
Figure 3.59: Two channel example.
3.9.3 Finite-length Multidimensional Equalizers
The case of finite-length diversity equalization becomes more complex because the matched filters are
implemented within the (possibly fractionally spaced) equalizers associated with each of the diversity
subchannels. There may be thus many coefficients in such a diversity equalizer.
24
In the case of finite-length equalizers shown in figure 3.60, each of the matched filters of the RAKE
is replaced by a lowpass filter of wider bandwidth (usually l times wider as in Section 3.7), a sampling
device at rate l/T, and a fractionally spaced equalizer prior to the summing device. With vectors p
k
in
Equation (3.288) now becoming l L-tuples,
p
k
= p(kT) , (3.532)
as do the channel output vectors y
k
and the noise vector n
k
, the channel input/output relationship (in
Eq (3.290)) again becomes
Y
k

=
_
¸
¸
¸
_
y
k
y
k−1
.
.
.
y
k−Nf+1
_
¸
¸
¸
_
(3.533)
=
_
¸
¸
¸
_
p
0
p
1
... p
ν
0 0 ... 0
0 p
0
p
1
... ... p
ν
... 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 ... 0 0 p
0
p
1
... p
ν
_
¸
¸
¸
_
_
¸
¸
¸
¸
¸
¸
_
x
k
x
k−1
.
.
.
.
.
.
x
k−Nf−ν+1
_
¸
¸
¸
¸
¸
¸
_
+
_
¸
¸
¸
_
n
k
n
k−1
.
.
.
n
k−Nf +1
_
¸
¸
¸
_
. (3.534)
The rest of Section 3.7 then directly applies with the matrix P changing to include the larger l L-tuples
corresponding to L diversity channels, and the corresponding equalizer W having its 1 L coefficients
corresponding to w
0
... w
Nf
. Each coefficient thus contains l values for each of the L equalizers.
The astute reader will note that the diversity equalizer is the same in principle as a fractionally
spaced equalizer except that the oversampling that creates diversity in the FSE generalizes to simply
any type of additional dimensions per symbol in the diversity equalizer. The dfecolor program can be
used where the oversampling factor is simply l L and the vector of the impulse response is appropriately
organized to have l L phases per entry. Often, l = 1, so there are just L antennas or lines of samples
per symbol period entry in that input vector.
24
Maximal combiners that select only one (the best) of the diversity channels for equalization are popular because they
reduce the equalization complexity by at least a factor of L – and perhaps more when the best subchannel needs less
equalization.
251
Figure 3.60: Fractionally spaced RAKE MMSE-DFE.
3.9.4 DFE RAKE Program
A DFE RAKE program similar to the DFERAKE matlab program has again been written by the
students (and instructor debugging!) over the past several years. It is listed here and is somewhat
self-explanatory if the reader is already using the DFECOLOR program.
%DFE design program for RAKE receiver
%Prepared by: Debarag Banerjee, edited by Olutsin OLATUNBOSUN
% and Yun-Hsuan Sung to add colored and spatially correlated noise
%Significant Corrections by J. Cioffi and above to get correct results --
%March 2005
%
function [dfseSNR,W,b]=dfsecolorsnr(l,p,nff,nbb,delay,Ex,noise);
% ------------------------------
%**** only computes SNR ****
% l = oversampling factor
% L = No. of fingers in RAKE
% p = pulse response matrix, oversampled at l (size), each row corresponding to a diversity path
% nff = number of feedforward taps for each RAKE finger
% nbb = number of feedback taps
% delay = delay of system <= nff+length of p - 2 - nbb
% if delay = -1, then choose best delay
% Ex = average energy of signals
% noise = noise autocorrelation vector (size L x l*nff)
% NOTE: noise is assumed to be stationary, but may be spatially correlated
% outputs:
% dfseSNR = equalizer SNR, unbiased in dB
% ------------------------------
siz = size(p,2);
L=size(p,1);
nu = ceil(siz/l)-1;
p = [p zeros(L,(nu+1)*l-siz)];
% error check
if nff<=0
252
error(’number of feedforward taps must be > 0’);
end
if delay > (nff+nu-1-nbb)
error(’delay must be <= (nff+(length of p)-2-nbb)’);
end
if delay < -1
error(’delay must be >= 0’);
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%if length(noise) ~= L*l*nff
% error(’Length of noise autocorrelation vector must be L*l*nff’);
%end
if size(noise,2) ~= l*nff | size(noise,1) ~= L
error(’Size of noise autocorrelation matrix must be L x l*nff’);
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%form ptmp = [p_0 p_1 ... p_nu] where p_i=[p(i*l) p(i*l-1)... p((i-1)*l+1)
for m=1:L
ptmp((m-1)*l+1:m*l,1) = [p(m,1); zeros((l-1),1)];
end
for k=1:nu
for m=1:L
ptmp((m-1)*l+1:m*l,k+1) = conj((p(m,k*l+1:-1:(k-1)*l+2))’);
end
end
ptmp;
% form matrix P, vector channel matrix
P = zeros(nff*l*L+nbb,nff+nu);
%First construct the P matrix as in MMSE-LE
for k=1:nff,
P(((k-1)*l*L+1):(k*l*L),k:(k+nu)) = ptmp;
end
%Add in part needed for the feedback
P(nff*l*L+1:nff*l*L+nbb,delay+2:delay+1+nbb) = eye(nbb);
temp= zeros(1,nff+nu);
temp(delay+1)=1;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Rn = zeros(nff*l*L+nbb);
for i = 1:L
n_t = toeplitz(noise(i,:));
for j = 1:l:l*nff
for k = 1:l:l*nff
Rn((i-1)*l+(j-1)*L+1:i*l+(j-1)*L, (i-1)*l+(k-1)*L+1:i*l+(k-1)*L) = n_t(j:j+l-1,k:k+l-1);
end
end
253
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Ex*P*P’;
Ry = Ex*P*P’ + Rn;
Rxy = Ex*temp*P’;
IRy=inv(Ry);
w_t = Rxy*IRy;
%Reshape the w_t matrix into the RAKE filter bank and feedback matrices
ww=reshape(w_t(1:nff*l*L),l*L,nff);
for m=1:L
W(m,:)=reshape(ww((m-1)*l+1:m*l,1:nff),1,l*nff);
end
b=-w_t(nff*l*L+1:nff*l*L+nbb);
sigma_dfse = Ex - w_t*Rxy’;
dfseSNR = 10*log10(Ex/sigma_dfse - 1);
For the previous example, some command strings that work are (with whitened-noise equivalent
channel first):
>> p
p = 0.9000 1.0000 0
0 1.0500 0.8400
>> [snr,W,b] = dfeRAKE(1,p,6,1,5,1,[.181 zeros(1,5) ; .181 zeros(1,5)])
snr = 11.1465
W = 0.0213 -0.0439 0.0668 -0.0984 0.1430 0.3546
-0.0027 0.0124 -0.0237 0.0382 0.4137 -0.0000
b = 0.7022
or with colored noise directly inserted
>> p1
p1 = 0.9000 1.0000 0
0 1.0000 0.8000
>> [snr,W,b] = dfeRAKE(1,p1,6,1,5,1,[.181 zeros(1,5) ; .164 zeros(1,5)])
snr = 11.1486
W = 0.0213 -0.0439 0.0667 -0.0984 0.1430 0.3545
-0.0028 0.0130 -0.0249 0.0401 0.4347 0.0000
b = 0.7022
Two outputs are not quite the same because of the finite number of taps.
3.9.5 Multichannel Transmission
Multichannel transmission in Chapters 4 and 5 is the logical extension of diversity when many inputs
may share a transmission channel. In the multichannel case, each of many inputs may affect each of many
outputs, logically extending the concept of intersymbol interference to other types of overlap than simple
ISI. Interference from other transmissions is more generally called crosstalk. Different inputs may for
instance occupy different frequency bands that may or may not overlap, or they may be transmitted
from different antennas in a wireless system and so thus have different channels to some common receiver
or set of receivers. Code division systems (See Chapter 5 Appendix) use different codes for the different
254
sources. A diversity equalizer can be designed for some set of channel outputs for each and every of
the input sequences, leading to multichannel transmission. In effect the set of equalizers attempts
to “diagonalize” the channels so that no input symbol from any source interferes with any other source
at the output of each of the equalizers. From an equalizer perspective, the situation is simply multiple
instances of the diversity equalizer already discussed in this section. However, when the transmit signals
can be optimized also, there can be considerable improvement in the performance of the set of equalizers.
Chapters 4 and 5 (EE379C) develop these concepts.
However, the reader may note that a system that divides the transmission band into several different
frequency bands may indeed benefit from a reduced need for equalization within each band. Ideally,
if each band is sufficiently narrow to be viewed as a “flat” channel, no equalizer is necessary. The
SNR of each of these “sub-channels” relates (via the “gap” approximation) how many bits can be
transferred with QAM on each. By allocating energy intelligently, such a system can be simpler (avoiding
equalization complexity) and actually perform better than equalized wider-band QAM systems. While
this chapter has developed in depth the concept of equalization because there are many wide-band
QAM and PAM systems that are in use and thus benefit from equalization, an intuition that for the
longest time transmission engineers might have been better off never needing the equalizer and simply
transmitting in separate disjoint bands is well-founded. Chapters 4 and 5 support this intuition. Progress
of understanding in the field of transmission should ultimately make equalization methods obsolete. As
in many technical fields, this has become a line of confrontation between those resisting change and
those with vision who see a better way.
255
Exercises - Chapter 3
3.1 Quadrature Modulation.
Consider the quadrature modulator shown below,
Figure 3.61: Quadrature modulator.
a. What conditions must be imposed on H
1
(f) and H
2
(f) if the output signal s(t) is to have no
spectral components between −f
c
and f
c
? Assume that f
c
is larger than the bandwidth of H
1
(f)
and H
2
(f). (2 pts.)
b. Let the input signal be of the form,
x(t) =

n=−∞
a
n
φ(t −nT)
What conditions must be imposed on H
1
(f) and H
2
(f) if the real part of the demodulated signal
is to have no ISI. (3 pts.)
c. Find the impulse responses h
1
(t) and h
2
(t) corresponding to the minimum bandwidth H
1
(f) and
H
2
(f) which simultaneously satisfy (a) and (b). You can express your answer in terms of φ(t). (3
pts.)
3.2 Sampling time and Eye Patterns.
The received signal for a binary transmission system is,
y(t) =

n=−∞
a
n
q(t −nT) +n(t)
where a
n
∈ ¦−1, +1¦ and Q(f) is a triangular, i.e. q(t) = sinc
2
(
t
T
). The received signal is sampled at
t = kT +t
0
, where k is an integer and t
0
is the sampling phase, [t
0
[ <
1
2
T.
a. Neglecting the noise for the moment, find the peak distortion as a function of t
0
. Hint: use
Parseval’s relation. (3 pts.)
b. Consider the following four binary sequences ¦u
n
¦

−∞
, ¦v
n
¦

−∞
, ¦w
n
¦

−∞
, ¦x
n
¦

−∞
:
u
n
= −1 ∀n
v
n
= +1 ∀n
256
w
n
=
_
+1, for n = 0
−1, otherwise
x
n
=
_
−1, for n = 0
+1, otherwise
Using the result of (a), find expressions for the 4 outlines of the binary eye pattern. Sketch the
eye pattern for these four sequences over two symbol periods −T ≤ t ≤ T. (2 pts.)
c. Find the width (horizontal) of the eye pattern at its widest opening. (2 pts.)
d. If the noise variance is σ
2
, find a worst case bound (using peak distortion) on the probability of
error as a function of t
0
. (2 pts.)
3.3 The 379 channel model.
Consider the ISI-model shown in Figure 3.2 with PAM and symbol period T. Let ϕ(t) =
1

T
sinc(
t
T
)
and h(t) = δ(t) −
1
2
δ(t − T).
a. Determine p(t), the pulse response. (1 pt)
b. Find [[p[[ and ϕ
p
(t). (2 pts.)
c. (2 pts.) Find q(t), the function that characterizes how one symbol interferes with other symbols.
Confirm that q(0) = 1 and that q(t) is hermitian (conjugate symmetric). You may find it useful
to note that
sinc
_
t + mT
T
_
∗ sinc
_
t + nT
T
_
= Tsinc
_
t + (m +n)T
T
_
.
d. Use Matlab to plot q(t). For this plot you may assume that T = 10 and plot q(t) for integer values
of t. Assuming that we sample y(t) at the perfect instants (i.e. when t = kT for some integer k),
how many symbols are distorted by a given symbol? Assume that the given symbol was positive,
how will the distorted symbols be affected? Specifically, will they be increased or decreased? (2
pts.)
e. For the rest of the problem consider 8 PAM with d=1. Determine the peak distortion, T
p
. (2 pts.)
f. Determine the MSE distortion, T
mse
. Compare with the peak distortion. (2 pts.)
g. Find an approximation to the probability of error (using the MSE distortion). You may express
your answer in terms of σ.(1 pt.)
3.4 Bias and SNR.
Continuing the setup of the previous problem with 8 PAM and d = 1, let
1
2
^
0
= σ
2
= 0.1.
Assume that a detector first scales the sampled output y
k
by α and chooses the closest x to y
k
α as
illustrated in Figure 3.13. Please express your SNR’s below both as a ratio of powers and in
dB.
a. For which value of α is the receiver unbiased? (1 pt.)
b. For the value of α found in (a), find the receiver signal-to-noise ratio, SNR
R
. (2 pts.)
c. Find the value of α that maximizes SNR
R
and the corresponding SNR
R
. (2 pts.)
d. Show that the receiver found in the previous part is biased. (1 pt.)
e. Find the matched filter bound on signal-to-noise ratio, SNR
MFB
. (1 pt.)
257
f. Discuss the ordering of the three SNR’s you have found in this problem. Which inequalities will
always be true? (1 pt.)
3.5 Raised cosine pulses with Matlab.
For this problem you will need to get all the .m files from /usr/class/ee379a/hw5/ on leland. If you
are logon to leland, you could do a cd and copy the files to your directory. Alternatively, you could do
an anonymous ftp to ftp.stanford.edu, and cd to class/ee379a/hw5 and copy the files.
a. Consider the formula for the raised cosine pulse (Eq. 3.79 of chapter 3):
q(t) = sinc
_
t
T
_

_
cos
_
απt
T
_
1 −
_
2αt
T
_
2
_
There are three values of t that would cause a program, such as Matlab, difficulty because of a
division by zero. Identify these trouble spots and evaluate what q(t) should be for these values. (2
pts)
b. Once you have identified those three trouble spots, a function to generate raised cosine pulses is a
straightforward implementation of Eq. 3.75. We have implemented this for you in mk rcpulse.m.
Executing q = mk rcpulse(a) will generate a raised cosine pulse with α = a. The pulses generated
by mk rcpulse assume T = 10 and are truncated to 751 points. (2 pts)
Use mk rcpulse to generate raised cosine pulses with 0%, 50%, and 100% excess bandwith and
plot them with plt pls lin. The syntax is plt pls lin(q 50, ’title’) where q 50 is the vector of
pulse samples generated by mk rcpulse(0.5) (50% excess bandwidth). Turn in your three plots
and discuss any deviations from the ideal expected frequency domain behavior.
c. Use plt pls dB(q, ’title’) to plot the same three raised cosine pulses as in part (b). Turn in your
plots. Which of these three pulses would you choose if you wanted the one which had the smallest
band of frequencies with energy above -40dB? Explain this unexpected result. (3 pts)
d. The function plt qk (q, ’title’) plots q(k) for sampling at the optimal time and for sampling
that is off by 4 samples (i.e., q(kT) and q(kT + 4) where T = 10). Use plt qk to plot q k for
0%, 50%, and 100% excess bandwidth raised cosine pulses. Discuss how excess bandwidth affects
sensitivity of ISI performance to sampling at the correct instant. (3 pts)
3.6 Noise enhancement: MMSE-LE vs ZFE.
Consider the channel with
|p|
2
= 1 +aa

Q(D) =
a

D
−1
+p
2
+aD
p
2
0 ≤ [a[ < 1.
a. (2 pts) Find the zero forcing and minimum mean square error linear equalizers W
ZFE
(D) and
W
MMSE−LE
(D). Use the variable b = |p|
2
_
1 +
1
SNRMFB
_
in your expression for W
MMSE−LE
(D).
b. (6 pts) By substituting e
−jwT
= D (with T = 1) and taking SNR
MFB
= 10|p|
2
, use Matlab to
plot (lots of samples of) W(e
jw
) for both ZFE and MMSE-LE for a = .5 and a = .9. Discuss the
differences between the plots.
c. (3 pts) Find the roots r
1
, r
2
of the polynomial
aD
2
+bD + a

.
Show that b
2
−4aa

is always a real positive number (for [a[ ,= 1). Hint: Consider the case where
1
SNRMFB
= 0. Let r
2
be the root for which [r
2
[ < [r
1
[. Show that r
1
r

2
= 1.
258
d. (2 pts) Use the previous results to show that for the MMSE-LE
W(D) =
|p|
a
D
(D −r
1
)(D −r
2
)
=
|p|
a(r
1
−r
2
)
_
r
1
D − r
1

r
2
D− r
2
_
. (3.535)
e. (2 pts) Show that for the MMSE-LE, w
0
=
p

b
2
−4aa

. By taking
1
SNRMFB
= 0, show that for the
ZFE w
0
=
p
1−aa

.
f. (4 pts) For
¯
c
x
= 1 and σ
2
= 0.1 find expressions for σ
2
ZFE
, σ
2
MMSE−LE
, γ
ZFE
, and γ
MMSE−LE
.
g. (4 pts) Find γ
ZFE
and γ
MMSE−LE
in terms of the paramter a and calculate for a = 0, 0.5, 1.
Sketch γ
ZFE
and γ
MMSE−LE
for 0 ≤ a < 1.
3.7 DFE is even better.
a. (2 pts) For the channel of problem 3.6, show that the canonical factorization is
Q(D) +
1
SNR
MFB
= γ
0
(1 − r
2
D
−1
)(1 −r

2
D).
What is γ
0
in terms of a and b? Please don’t do this from scratch. You have done much of the
work for this in problem 3.6.
b. (2 pts) Find B(D) and W(D) for the MMSE DFE.
c. (4 pts) Give an expression for γ
MMSE−DFE
. Compute its values for a = 0, .5, 1 for the
¯
c
x
and
σ
2
of problem 3.6. Sketch γ
MMSE−DFE
as in problem 3.6. Compare with your sketches from
Problem 3.6.
3.8 Noise predictive DFE.
Figure 3.62: Noise-Predictive DFE
Consider the equalizer shown above with B(D) restricted to be causal and monic. Assume all decisions
are correct. We use a MMSE criterion to choose U(D) and B(D) to minimize E[[x
k
− z

k
[
2
]. (x
k
is the
channel input.)
a. (5 pts) Show this equalizer is equivalent to a MMSE-DFE and find U(D) and B(D) in terms of
G(D), Q(D), |p|,
¯
c
x
, SNR
MFB
and γ
0
.
259
b. (2 pts) Relate U
NPDFE
(D) to W
MMSE-LE
(D) and to W
MMSE-DFE
(D).
c. (1 pt) If we remove feedback (B(D) = 1), what does this equalizer become?
d. (1 pt) Interpret the name “noise-predictive” DFE by explaining what the feedback section is doing.
e. (2 pts) Is the NPDFE biased? If so, show how to remove the bias.
3.9 Receiver SNR relationships.
a. (3 pts) Recall that:
Q(D) +
1
SNR
MFB
= γ
0
G(D)G

(D
−∗
)
Show that:
1 +
1
SNR
MFB
= γ
0
|g|
2
b. (3 pts) Show that:
|g| ≥ 1
with equality if and only if Q(D) = 1 ( i.e. if the channel is flat) and therefore:
γ
0
≤ 1 +
1
SNR
MFB
c. (3 pts) Let x
0
denote the time-zero value of the sequence X. Show that:
_
_
1
Q(D) +
1
SNR
MFB
_
_
0

1
γ
0
and therefore:
SNR
MMSE-LE,U
≤ SNR
MMSE-DFE,U
(with equality if and only if Q(D) = 1, that is Q(ω) has vestigial symmetry.)
d. (2 pts) Use the results of parts b) and c) to show:
SNR
MMSE-LE,U
≤ SNR
DFE,U
≤ SNR
MFB
3.10 Bias and probability of error.
We have used the fact that the best unbiased receiver has a lower P
e
than the (biased) MMSE
receiver. Here is a simple illustration of this fact.
Figure 3.63: A three point PAM constellation.
Consider the constellation above with additive white noise n
k
having
^
0
2
= σ
2
= 0.1.
Assume that the inputs are independent and identically distributed uniformly over the three possible
values. Assume also that n
k
is zero mean and independent of x
k
.
260
a. (1 pt) Find the mean square error between y
k
= x
k
+n
k
and x
k
. (easy)
b. (1 pt) Find P
e
exactly for the ML detector on the unbiased values y
k
.
c. (2 pts) Find the scale factor α that will minimize the mean square error
E[e
2
k
] = E[(x
k
−αy
k
)
2
].
Prove that your α does in fact provide a minimum by taking the appropriate second derivative.
d. (2 pts) Find E[e
2
k
] and P
e
for the scaled output αy
k
. When you compute P
e
, use the decision
regions for the unbiased ML detector you found in (b). You should find that the biased receiver
of part c has a P
e
which is higher than the (unbiased) ML detector of part b, even though it has
a smaller squared error.
e. (1 pt) Where are the optimal decision boundaries for detecting x
k
from αy
k
(for the MMSE α) in
part c? What is the probability of error for these decision boundaries?
3.11 Bias and the DFE.
Consider the DFE you designed in Problem 3.7. Recall that you (hopefully) found
γ
0
=
b +

b
2
− 4aa

2(1 +aa

)
.
a. (2 pts) Find G
u
in terms of a, b, and r
2
. Assume the same SNR
MFB
as in Prob 3.7.
b. (1 pt) Give a block diagram of the DFE with scaling after the feedback summation.
c. (1 pt) Give a block diagram of the DFE with scaling before the feedback summation.
3.12 Zero forcing DFE.
Consider once again the general channel model of Problems 3.6 and 3.7 with
¯
c
x
= 1 and σ
2
= 0.1.
a. (2 pts) Find η
0
and P
c
(D) so that
Q(D) = η
0
P
c
(D)P

c
(D
−∗
).
b. (2 pts) Find B(D) and W(D) for the ZF-DFE.
c. (1 pt) Find σ
2
ZF−DFE
d. (2 pts) Find the loss with respect to SNR
MFB
for a = 0, .5, 1. Sketch the loss for 0 ≤ a < 1.
3.13 Complex baseband channel.
The pulse response of a channel is band-limited to
π
T
and has
p
0
=
1

T
(1 + 1.1j)
p
1
=
1

T
(0.95 + 0.5j)
p
k
= 0 for k ,= 0, 1
where p
k
= p(kT) and SNR
MFB
= 15 dB.
a. (2 pts) Find |p|
2
and find a so that
Q(D) =
a

D
−1
+ |p|
2
+aD
|p|
2
261
b. (2 pts) Find SNR
MMSE−LE,U
and SNR
MMSE−DFE,U
for this channel. Use the results of Prob-
lems 3.6 and 3.7 wherever appropriate. Note |p|
2
,= 1 + aa

so some care should be exercised in
using the results of Problems 3.6 and 3.7.
c. (2 pts) Use the SNR’s from (b) to compute the NNUB on P
e
for 4-QAM with the MMSE-LE and
the MMSE-DFE.
3.14 Tomlinson Precoding.
We will be working with the 1 + 0.9D channel (i.e. the 1 +aD channel that we have been exploring
at length with a = 0.9). For this problem, we will assume that σ
2
= 0. Furthermore, assume a 4-PAM
constellation with x
k
∈ ¦−3, −1, 1, 3¦.
a. (5 pts) Design a Tomlinson precoder and its associated receiver for this system. Your system will
be fairly simple. How would your design change if noise were present in the system?
b. (5 points) Implement the precoder and its associated receiver using any computer method you
would like (or you can do the following computations by hand if you prefer). Turn in a hard copy
of whatever code you write, and assume no noise.
For the following input sequence compute the output of the precoder, the output of the channel,
and the output of your receiver. Assume that the symbol x

= −3 was sent just prior to the input
sequence below.
¦3, −3, 1, −1, 3, −3, −1¦
c. (3 pts) Now, again with zero noise, remove the modulo operations from your precoder and your
receiver. For this modified system, compute the output of the precoder, the output of the channel,
and the output of your receiver for the same inputs as in the previous part. Turn in hard copies
of whatever code you write.
Did the system still work? What changed? What purpose do the modulo operators serve?
3.15 Flexible Precoding.
This problem considers the 1 + 0.9D channel (i.e. the 1 + aD channel that we have been exploring
at length with a = 0.9). For this problem, we will assume that σ
2
= 0. Furthermore, assume a 4-PAM
constellation with x
k
∈ ¦−3, −1, 1, 3¦.
a. (5 pts) Design a Flexible precoder and its associated receiver for this system. Your system will be
fairly simple. How would your design change if noise were present in the system?
b. (5 points) Implement the precoder and its associated receiver using any computer method you
would like (or you can do the following computations by hand if you prefer). Turn in a hard copy
of whatever code you write, and assume no noise.
For the following input sequence compute the output of the precoder, the output of the channel,
and the output of your receiver. Assume that the symbol x

= −3 was sent just prior to the input
sequence below.
¦3, −3, 1, −1, 3, −3, −1¦
c. (3 pts) Now, again with zero noise, remove the modulo operations from your precoder and your
receiver. For this modified system, compute the output of the precoder, the output of the channel,
and the output of your receiver for the same inputs as in the previous part. Turn in hard copies
of whatever code you write.
Did the system still work? What changed? What purpose do the modulo operators serve?
262
3.16 Finite length equalization and matched filtering.
We design the optimal FIR MMSE-LE without assuming a matched filter. However, it turns out that
we get a “matched filter” anyway. Consider a system whose pulse response is band limited to [w[ <
π
T
that is sampled at the symbol rate T after pass through an anti-alias filter with gain

T.
a. (3 pts)
Show that
w = R
xY
R
−1
Y Y
= (0, . . . , 0, 1, 0, . . . , 0)Φ
p

_
|p|
_
ˆ
Q+l
1
SNR
MFB
I
__
−1
.
where
ˆ
Q =
PP

|p|
2
b. (2 pts) Which terms correspond to the matched filter? Which terms correspond to the infinite
length W
MMSE−LE
?
3.17 Finite Length equalization and MATLAB.
Consider the 1 + 0.25D channel with σ
2
= .1 and
¯
c
x
= 1.
a. (1 pt) Find SNR
MMSE−LE,U
for the infinite length filter.
b. (2 pts) In this problem we will use the MATLAB program mmsele. This program is interactive,
so just type mmsele and answer the questions. When it asks for the pulse response, you can type
[1 0.25] or p if you have defined p = [1 0.25].
Use mmsele to find the best ∆ and the associated SNR
MMSE−LE,U
for a 5 tap linear equalizer
. Compare with your value from part (a). How sensitive is performance to ∆ for this system?
c. (2 pts) Plot [P(e
(jwT)
[ and [P(e
(jwT)
W(e
(jwT)
[ for w ∈ [0,
π
T
]. Discuss the plots briefly.
3.18 Computing finite-length equalizers.
Consider the following system description.
¯
c
x
= 1
^
0
2
=
1
8
φ(t) =
1

T
sinc(
t
T
) h(t) = δ(t) −0.5δ(t −T) l = 1
Feel free to use matlab for any matrix manipulations as you complete the parts below. You may wish to
check your answers with dfecolor.m. However, for this problem, you should go through the calculations
yourself. (You may use matlab for matrix inversion.)
a. (2 pts) We assume perfect anti-alias filtering with gain

T. Find ˜ p(t) = (φ(t) ∗ h(t)) and |˜ p|
2
,
corresponding to the discrete-time channel
y
k
= x
k
−.5 x
k−1
+ n
k
. (3.536)
Also, find |
˜
P(D)|
2
=

k
[ ˜ p
k
[
2
.
b. (1 pt) Compute SNR
MFB
for this channel.
c. (2 pts) Design a 3 tap FIR MMSE-LE for ∆ = 0.
d. (1 pt) Find the σ
2
MMSE−LE
for the equalizer of the previous part.
e. (2 pts) Design a 3 tap FIR ZF-LE for ∆ = 0.
f. (1 pt) Find the associated σ
2
ZF−LE
.
g. (2 pts) Design an MMSE-DFE which has 2 feedforward taps and 1 feedback tap. Again, assume
that ∆ = 0.
263
Figure 3.64: Channel for Problem 3.19.
h. (2 pts) Compute the unbiased SNR’s for the MMSE-LE and MMSE-DFE. Compare these two
SNR’s with each other and the SNR
MFB
.
3.19 Equalizer Design - Final 1996
An AWGN channel has pulse response p(t) =
1

T
_
sinc
_
t
T
_
−sinc
_
t−4T
T

. The receiver anti-alias
filter has gain

T over the entire Nyquist frequency band, −1/2T < f < 1/2T, and zero outside this
band. The filter is followed by a 1/T rate sampler so that the sampled output has D-Transform
y(D) = (1 −D
4
)x(D) + n(D) . (3.537)
x(D) is the D-Transform of the sequence of M-ary PAM input symbols, and n(D) is the D-Transform
of the Gaussian noise sample sequence. The noise autocorrelation function is R
nn
(D) =
N0
2
. Further
equalization of y(D) is in discrete time where (in this case) the matched filter and equalizer discrete
responses (i.e., D-Transforms) can be combined into a single discrete-time response. The target P
e
is
10
−3
.
Let
N0
2
= -100 dBm/Hz (0 dBm = 1 milliwatt = .001 Watt) and let 1/T = 2 MHz.
a. Sketch [P(e
−ωT
)[
2
. (2 pts)
b. In your engineering judgement, what kind of equalizer should be used on this channel?
c. For a ZF-DFE, find the transmit symbol mean-square value (i.e., the transmit energy) necessary
to achieve a data rate of 6 Mbps using PAM and assuming a probability of symbol error equal to
10
−3
. (4 pts)
d. For your transmit energy in part c, how would a MMSE-DFE perform on this channel? (2pts)
3.20 FIR Equalizer Design - Final 1996
A symbol-spaced FIR equalizer (l = 1) is applied to a stationary sequence y
k
at the sampled output
of an anti-alias filter, which produces a discrete-time IIR channel with response given by
y
k
= a y
k−1
+b x
k
+n
k
, (3.538)
where n
k
is white Gaussian noise.
The SNR (ratio of mean square x to mean square n) is
¯
E
x
N
0
2
= 20 dB. [a[ < 1 and both a and b are
real. For all parts of this question, choose the best ∆ where appropriate.
a. Design a 2-tap FIR ZFE. (3 pts)
b. Compute the SNR
zfe
for part a. (2 pts)
c. Compute your answer in part b with that of the infinite-length ZFE. (1 pt)
d. Let a = .9 and b = 1. Find the 2-tap FIR MMSE-LE. (4 pts)
264
e. Find the SNR
mmse−le,u
for part d.
f. Find the SNR
zf−dfe
for an infinite-length ZF-DFE. How does the SNR compare to the matched-
filter bound if we assume there is no information loss incurred in the symbol-spaced anti-alias
filter? Use a = .9 and b = 1. (2 pts)
3.21 ISI quantification - Midterm 1996
For the channel P(ω) =

T(1 + .9e
ωT
) ∀[ω[ < π/T studied repeatedly in this chapter, use binary
PAM with
¯
c
x
= 1 and SNR
MFB
= 10 dB. Remember that q
0
= 1.
a. Find the peak distortion, T
p
. (1 pt)
b. Find the peak-distortion bound on P
e
. (2 pts)
c. Find the mean-square distortion, T
MS
. (1 pt)
d. Approximate P
e
using the T
MS
of part c. (2 pts)
e. ZFE: Compare P
e
in part d with the P
e
for the ZFE. Compute SNR difference in dB between the
SNR based on mean-square distortion implied in parts c and d and SNR
ZFE
. (hint, see example
in this chapter for SNR
ZFE
) (2 pts)
3.22 Precoding - Final 1995
For an ISI channel with
Q(D) +
1
SNR
MFB
= .82
_

2
D
−1
+ 1.25 −

2
D
_
(3.539)
a. Find SNR
MFB
and SNR
MMSE−DFE
. (2 pts)
b. Find G(D) and G
U
(D) for the MMSE-DFE. (1 pt)
c. Design (show/draw) a Tomlinson-Harashimaprecoder, showing fromx
k
through the decision device
in the receiver in your diagram (any M). ( 2 pts)
d. Let M = 4 for your precoder in part c. Find P
e
.
e. Design (show/draw) a Flexible precoder, showing from x
k
through the decision device in the
receiver in your diagram (any M). ( 2 pts)
f. Let M = 4 for your precoder in part e. Find P
e
.
3.23 Finite-delay tree search
A channel with multipath fading has one reflecting path with gain (voltage) 90% of the main path.
The relative delay on this path is approximately T seconds, but the carrier sees a phase-shift of −60
o
that is constant on the second path. Assume binary transmission throught this problem.
Use the model
P(ω) =
_ √
T
_
1 +ae
−ωT
_
[ω[ < π/T
0 elsewhere
to approximate this channel.
N0
2
= .0181 and
¯
c
x
= 1.
a. Find a. (1 pt)
b. Find SNR
MFB
. (1 pt)
c. Find W(D) and SNR
U
for the MMSE-LE. (3 pts)
d. Find W(D), B(D), and SNR
U
for the MMSE-DFE. (3 pts)
e. Show a simple method and compute SNR
zf−dfe
.
265
Figure 3.65: Finite-delay tree search.
f. A finite-delay tree search detector at time k decides ˆ x
k−1
by choosing that ˆ x
k−1
value that mini-
mizes
min
ˆ xk,ˆ xk−1
[z
k
− ˆ x
k
−aˆ x
k−1
[
2
+ [z
k−1
− ˆ x
k−1
− aˆ x
k−2
[
2
,
where ˆ x
k−2
= x
k−2
by assumption. How does this compare with the ZF-DFE (better or worse)?
(1 pt)
g. Find an SNR
fdts
for this channel. (bonus)
h. Could you generalize SNR in part g for FIR channels with ν taps? (bonus)
3.24 Peak Distortion -(5 pts)
Peak Distortion can be generalized to channels without reciever matched filtering (that is ϕ
p
(−t) is
a lowpass anti-alias filter). Let us suppose
y(D) = x(D) (1 −.5D) +N(D) (3.540)
after sampling on such a channel, where N(D) is discrete AWGN. P(D) = 1 −.5D.
a. write y
k
in terms of x
k
, x
k−1
and n
k
. (hint, this is easy. 1 pt)
b. We define peak distortion for such a channel as T
p

= [x
max
[

m=0
[p
m
[. Find this new T
p
for this
channel if [x
max
[ = 1. (2 pts)
c. Suppose
N0
2
, the mean-square of the sample noise, is .05 and
¯
c
x
= 1. What is P
e
for symbol-by-
symbol detection with PAM and M = 2. (2 pts)
3.25 More Equalization - Final 1994
Given an ISI channel with p(t) = 1/

T [sinc(t/T) − sinc[(t − T)/T]],
¯
c
x
= 1, and
N0
2
= .05. for
each equalizer, find the filter transforms W(D) (and B(D)), the unbiased detection SNR, and the loss
with respect to the MFB.
a. What is SNR
MFB
?
b. Find the ZFE.
c. Find the MMSE-LE.
d. Find the ZF-DFE.
e. Find the MMSE-DFE.
f. Draw a diagram illustrating the MMSE-DFE implementation with Tomlinson precoding.
266
g. For P
e
< 10
−6
and using square or cross QAM, choose a design and find the largest data rate you
can transmit using one of the equalizers above.
3.26 Raised Cosine pulses
Show through inverse Fourier transforming that the raised cosine and square-root raised cosine time-
domain responses given in Section 3.3 are correct.
3.27 Expert Understanding - Final 1998
Figure 3.66: Channel for Problem 3.27.
A channel baseband transfer function for PAM transmission is shown in Figure 3.66. PAM transmis-
sion with sinc basis function is used on this channel with a transmit power level of 10 mW =c
x
/T. The
one-sided AWGN psd is -100 dBm/Hz.
a. Let the symbol rate be 20 MHz - find SNR
MFB
.(1 pt)
b. For the same symbol rate as part a, what is SNR
MMSE−LE
? (1 pt)
c. For the equalizer in part b, what is the data rate for PAM if P
e
≤ 10
−6
? (1 pt)
d. Let the symbol rate be 40 MHz - find the new SNR
MFB
.(2 pts)
e. Draw a the combined shape of the matched-filter and feedforward filter for a ZF-DFE corresponding
to the new symbol rate of part d. (1 pt)
f. Estimate SNR
MMSE−DFE,U
for the new symbol rate, assuming a MMSE-DFE receiver is used. (
2 pts) (Hint: you may note the relationship of this transfer function to the channel often used as
an example in EE379A).
g. What is the new data rate for this system at the same probability of error as part c - compare
with the data rate of part c. (1 pt)
h. What can you conclude about incurring ISI if the transmitter is allowed to vary its bandwidth?
3.28 Infinite-Length EQ - 10 pts, Final 1998
An ISI channel with PAM transmission has channel correlation function given by
Q(D) =
.19
(1 +.9D)(1 + .9D
−1
)
, (3.541)
with SNR
MFB
=10 dB, c
x
= 1, and |p|
2
=
1
.19
.
267
a. (3 pts) Find W
ZFE
(D), σ
2
ZFE
, and SNR
ZFE
.
b. (1 pt) Find W
MMSE−LE
(D)
c. (3 pts) Find W
ZF−DFE
(D), B
ZF−DFE
(D), and SNR
ZF−DFE
.
d. (3 pts) Find W
MMSE−DFE
(D), B
MMSE−DFE
(D), and SNR
MMSE−DFE,U
.
3.29 Finite-Length EQ - 9 pts, Final 1998
A baseband channel is given by
P(f) =
_ √
T
_
1 − .7e
2πfT
_
[f[ <
.5
T
0 [f[ ≥
.5
T
(3.542)
and finite-length equalization is used with anti-alias perfect LPR with gain

T followed by symbol-rate
sampling. After the sampling, a maximum complexity of 3 taps TOTAL over a feedforward filter and a
feedback filter can be tolerated. c
x
= 2 and
N0
2
= .01 for a symbol rate of 1/T = 100kHz is used with
PAM transmission. Given the complexity constraint, find the highest data rate achievable with PAM
transmission when the corresponding probability of symbol error must be less than 10
−5
.
3.30 Equalizers - Miterm 2000 - 10 pts
PAM transmission on a filtered AWGN channel uses basic function ϕ(t) =
1

T
sinc
_
t
T
_
with T = 1
and undergoes channel impulse response with Fourier transform ([a[ < 1)
H(ω) =
_
1
1+a·e
ω
[ω[ ≤ π
0 [ω[ > π
(3.543)
and SNR =
¯
E
x
σ
2
= 15 dB.
a. Find the Fourier Transform of the pulse response, P(ω)? (1 pt)
b. Find |p|
2
. (2 pts)
c. Find Q(D), the function characterizing ISI. (3 pts)
d. Find the W(D) for the zero-forcing and MMSE linear equalizers on this channel. (3 pts)
e. If a = 0, what data rate is achievable on this channel according to the gap approximation at
P
e
= 10
−6
? (1 pt)
3.31 Diversity Channel - Final 2000 - 6 pts
We would like to evaluate the channel studied throughout this course and in the text with the same
input energy and noise PSD versus almost the same channel except that two receivers independently
receive:
• an undistorted delayed-by-T signal (that is P(D) = D), and
• a signal that is not delayed, but reduced to 90% of its amplitude (that is P(D) = .9).
a. Find Q(D), |p|, and SNR
MFB
for the later diversity channel. (3 pts)
b. Find the performance of the MMSE-LE, MMSE-DFE for the later diversity channel and compare
to that on the original one-receiver channel and with respect to the matched filter bound. (2 pts)
c. Find the probability of bit error on the diversity channel for the case of 1 bit/dimension transmis-
sion. (1 pt)
3.32 Do we understand basic detection - Final 2000 - 9 pts
A filtered AWGN channel with T = 1 has pulse response p(t) = sinc(t) +sinc(t −1) with SNR
MFB
=
14.2 dB.
268
a. Find the probability of error for a ZF-DFE if a binary symbol is transmitted (2 pts).
b. The Tomlinson precoder in this special case of a monic pulse response with all unity coefficients
can be replaced by a simpler precoder whose output is
x

k
=
_
x

k−1
if x
k
= 1
−x

k−1
if x
k
= −1
(3.544)
Find the possible channel outputs and determine an SBS decoder rule. What is the performance
(probability of error) for this decoder? (2 pts)
c. Suppose this channel is used one time for the transmission of one of the 8 4-dimensional messages
that are defined by [+ ± ± ±], which produce 8 possible 5-dimensional outputs. ML detection
is now used for this multidimensional 1-shot channel. What are the new minimum distance and
number of nearest neighbors and how do they relate to those for the system in part a? What is
the new number of bits per output dimension? (3 pts)
d. Does the probability of error change significantly if the input is extended to arbitrarily long block
length N with only the first sample fixed at +, and all other inputs being equally likely + or −?
What happens to the number of bits per output dimension in this case? (2 pts).
3.33 Equalization - Final 2000 - 10 pts
PAM transmission is used on a filtered AWGN channel with
N0
2
= .01, T = 1, and pulse repsonse
p(t) = sinc(t) + 1.8sinc(t −1) + .81sinc(t −1). The desired probability of symbol error is 10
−6
.
a. Find |p|
2
, SNR
MFB
, AND Q(D) for this channel. (2 pts)
b. Find the MMSE-LE and corresponding unbiased SNR for this channel. (2 pts)
c. Find the ZF-DFE detection SNR and loss with respect to SNR
MFB
. (1 pt)
d. Find the MMSE-DFE. Also compute the MMSE-DFE SNR and maximumdata rate in bits/dimension
using the gap approximation. (3 pts)
e. Draw the flexible (Laroia) precoder for the MMSE-DFE and draw the system from transmit symbol
to detector in the receiver using this precoder. Implement this precoder for the largest number of
integer bits that can be transmitted according to your answer in part (d). (2 pts)
3.34 Finite-Length Equalizer Design - Final 2000 - 15 pts
A filtered AWGN channel has impulse response h(t) =
1
T

1
1+
_
10
7
t
3
_
2
and is used in the transmission
system of Figure 3.67. QAM transmission with 1/T = 1 MHz and carrier frequency f
c
= 600kHz are
used on this channel. The AWGN psd is
N0
2
= −86.5 dBm/Hz. The transmit power is
E
x
T
=1 mW. The
oversampling factor for the equalizer design is chosen as l = 2. Square-root raised cosine filtering with
10% excess bandwidth is applied as a transmit filter, and an ideal filter is used for anti-alias filtering at
the receiver. A matlab subroutine at the course web site may be useful in computing responses in the
frequency domain.
a. Find ν so that an FIR pulse response approximates this pulse response so that less than 5% error
in |p|
2
.
b. Calculate SNR
MFB
. Determine a reasonable set of parameters and settings for an FIR MMSE-LE
and the corresponding data rate at P
e
= 10
−7
. Calculate the data rate using both an integer and
a possibly non-integer number of bits/symbol. (4 pts)
c. Repeat part b for an FIR MMSE-DFE design and draw receiver. Calculate the data rate using
both an integer and a possibly non-integer number of bits/symbol. (3 pts)
269
Figure 3.67: Transmission system for Problem 3.34.
d. Can you find a way to improve the data rate on this channel by changing the symbol rate and or
carrier frequency? (2 pts)
3.35 Telephone-Line Transmission with “T1”
Digital transmission on telephone lines necessarily must pass through two “isolation transformers” as
illustrated in Figure 3.68. These transformers prevent large D.C. voltages accidentally placed on the line
from unintentionally harming the telephone line or equipment attached to it, and they provide immunity
to earth currents, noises, and ground loops. These transformers also introduce ISI.
Figure 3.68: Illustration of a telephone-line data transmission for Problem 3.35.
a. The “derivative taking” combined characteristic of the transformers can be approximately modeled
at a sampling rate of 1/T = 1.544 MHz as successive differences between channel input symbols.
For sufficiently short transmission lines, the rest of the line can be modeled as distortionless. What
is a reasonable partial-response model for the channel H(D)? Sketch the channel transfer function.
Is there ISI?
b. How would a zero-forcing decision-feedback equalizer generally perform on this channel with respect
to the case where channel output energy was the same but there are no transformers?
c. What are some of the drawbacks of a ZF-DFE on this channel?
d. Suppose a Tomlinson precoder were used on the channel with M = 2, how much is transmit
energy increased generally? Can you reduce this increase by good choice of initial condition for
the Tomlinson Precoder?
270
e. Show how a binary precoder and corresponding decoder can significantly simplify the implemen-
tation of a detector. What is the loss with respect to optimum MFB performance on this channel
with your precoder and detector?
f. Suppose the channel were not exactly a PR channel as in part a, but were relatively close. Char-
acterize the loss in performance that you would expect to see for your detector.
3.36 Magnetic Recording Channel
Digital magnetic information storage (i.e., disks) and retrieval makes use of the storage of magnetic
fluexes on a magnetic disk. The disk spins under a “read head” and by Maxwell’s laws the read-head
wire senses flus changes in the moving magnetic field, thus generating a read current that is translated
into a voltage through amplifiers succeeding the read head. Change in flux is often encoded to mean a
“1” was stored and no change means a “0” was stored. The read head has finite band limitations also.
a. Pick a partial-resonse channel with ν = 2 that models the read-back channel. Sketch the magnitude
characteristic (versus frequency) for your channel and justify its use.
b. How would a zero-forcing decision-feedback equalizer generally perform on this channel with respect
to the best case where all read-head channel output energy convenyed either a positive or negative
polarity for each pulse?
c. What are some of the drawbacks of a ZF-DFE on this channel?
d. Suppose a Tomlinson precoder were used on the channel with M = 2, how much is transmit
energy increased generally? Can you reduce this increase by good choice of initial condition for
the Tomlinson Precoder?
e. Show how a binary precoder and corresponding decoder can significantly simplify the implemen-
tation of a detector. What is the loss with respect to optimum MFB performance on this channel
with your precoder and detector?
f. Suppose the channel were not exactly a PR channel as in part a, but were relatively close. Char-
acterize the loss in performance that you would expect to see for your detector.
g. Suppose the density (bits per linear inch) of a disk is to be increased so one can store more files
on it. What new partial response might apply with the same read-channel electronics, but with a
correspondingly faster symbol rate?
3.37 Tomlinson preocoding and simple precoding
In Section 3.8, the simple precoder for H(D) = 1 + D is derived.
a. Design the Tomlinson precoder corresponding to a ZF-DFE for this channel with the possible
binary inputs to the precoder being ±1. (4 pts)
b. How many distinct outputs are produced by the Tomlinson Precoder assuming an initial state
(feedback D element contents) of zero for the precoder. (1 pt)
c. Compute the average energy of the Tomlinson precoder output. (1 pt)
d. How many possible outputs are produced by the simple precoder with binary inputs? (1 pt)
e. Compute the average energy of the channel input for the simple precoder with the input constel-
lation of part (a). (1 pt)
3.38 Flexible precoding and simple precoding
In Section 3.8, the simple precoder for H(D) = 1 + D is derived.
a. Design the Flexible precoder corresponding to a ZF-DFE for this channel with the possible binary
inputs to the precoder being ±1. (4 pts)
271
b. How many distinct outputs are produced by the Flexible Precoder assuming an initial state (feed-
back D element contents) of zero for the precoder. (1 pt)
c. Compute the average energy of the Flexible precoder output. (1 pt)
d. How many possible outputs are produced by the simple precoder with binary inputs? (1 pt)
e. Compute the average energy of the channel input for the simple precoder with the input constel-
lation of part (a). (1 pt)
3.39 Partial Response Precoding and the ZF-DFE
Consider an AWGN with H(D) = (1 − D)
2
with noise variance σ
2
and one-dimensional real input
x
k
= ±1.
a. Determine a partial-response (PR) precoder for the channel, as well as the decoding rule for the
noiseless channel output. (2 pts)
b. What are the possible noiseless outputs and their probabilities? From these, determine the P
e
for
the precoded channel. (4 pts)
c. If the partial-response precoder is used with symbol-by-symbol detection, what is the loss with
respect to the MFB? Ignore nearest neighbor terms for this calculation since the MFB concerns
only the argument of the Q-function. (1 pt)
d. If a ZF-DFE is used instead of a precoder for this channel, so that P
c
(D) = 1 −2D +D
2
, what is
η
0
? Determine also the SNR loss with respect to the SNR
MFB
(2 pts)
e. Compare this with the performance of the precoder, ignoring nearest neighbor calculations. (1 pt)
3.40 Error propagation and nearest neighbors
Consider the H(D) = 1 − D
2
channel with AWGN noise variance σ
2
and d = 2 and 4-level PAM
transmission.
a. State the precoding rule and the noiseless decoding rule. (1 pts)
b. Find the possible noiseless outputs and their probabilities. Find also N
e
and P
e
with the use of
precoding. (4 pts)
c. Suppose a ZF-DFE is used on this system and that at time k = 0 an incorrect decision x
0
−ˆ x
0
= 2
occurs. This incorrect decision affects z
k
at time k = 2. Find the N
e
(taking the error at k = 0
into account) for the ZF-DFE. From this, determine the P
e
with the effect of error propagation
included. (4 pts)
d. Compare the P
e
in part (c) with that of the use of the precoder in part (a). (1 pt)
3.41 Forcing Partial Response
Consider a H(D) = 1 + .9D channel with AWGN noise variance σ
2
. We would like to convert this
to a 1 + D channel
a. Design an equalizer that will convert the channel to a 1 + D channel. (2 pts)
b. The received signal is y
k
= x
k
+ .9x
k−1
+ n
k
where n
k
is the AWGN. Find the autocorrelation of
the noise after going through the receiver designed in part (a). Evaluate r
0
, r
±1
, and r
±2
. Is the
noise white? (3 pts)
c. Do you think that the noise terms would be more or less correlated if we were to convert a 1+.1D
channel to a 1 + D channel? You need only discuss briefly. (1 pt)
272
3.42 Final 2001 - Equalizers - 11 pts
PAM transmission on a filtered AWGN channel uses basis function ϕ(t) =
1

T
sinc
_
t
T
_
with T = 1
and undergoes channel impulse response with Fourier transform ([α[ < 1)
H(ω) =
_
1
1+a·e
ω
[ω[ ≤ π
0 [ω[ > π
(3.545)
and SNR =
¯
E
x
σ
2
= 28 dB.
a. Find the Fourier Transform of the pulse response,P(ω) =? (1 pt)
b. Find |p|
2
. (1 pt)
c. Find Q(D), the function characterizing ISI. (2 pts)
d. Find the filters and sketch the block diagram of receiver for the MMSE-DFE on this channel for
a = .9. (3 pts)
e. Estimate the data rate for uncoded PAM transmission and P
e
< 10
−6
that is achievable with your
answer in part d. (2 pts)
f. Draw a diagram of the better precoder’s (Tomlinson or Laroia), xmit and rcvr, implementations
with d = 2 in the transmitted constellation. (2 pts)
3.43 Final 2001 - Finite-Length Equalizer Design - 16 pts
Transmission lines can sometimes have their transfer function described by the function
H(f) = k e
−α·|f|
(3.546)
where α and k are some positive real constants, here assigned the values α = 6π 10
−7
and k = 3π/10.
Our task in this problem is to design a transmission system for such a transmission line. Initially we
begin with 1/T = 800 kHz and carrier frequency f
c
= 600 kHz on this channel. The AWGN psd is
N0
2
= −92 dBm/Hz. The transmit power is c
x
/T = 1 mW. The oversampling factor for the equalizer
design is chosen as l = 2. Square-root raised cosine filtering with 10% excess bandwidth is applied as a
transmit filter and an ideal lowpass filter is used for anti-alias filtering at the receiver.
a. Find ν so that an FIR pulse response approximates this pulse response so that less than .25 dB
error in |p|
2
(that is, the norm square of the difference between the actual p(t) and the truncated
p
k
is less than (10
.025
−1) |p|
2
. (4 pts)
b. Determine a reasonable set of parameters and settings for an FIR MMSE-LE and the corresponding
data rate at P
e
≤ 10
−7
. Please print w filter coefficients. (2 pts)
c. Repeat part b for an FIR MMSE-DFE design and print coefficients. (2 pts)
d. Can you find a way to improve the data rate on this channel by changing the symbol rate and or
carrier frequency? Try to maximize the data rate with QAM transmission by a suitable choice of
carrier frequency and symbol rate, and provide that data rate - you may maintain the assumption
of oversampling by a factor of 2 and 10% excess bandwidth. (6 pts)
e. Suppose a narrowband Gaussian noise with PSD -62 dBm/Hz were injected over a bandwith of
1 kHz - hereustically, would you expect this to change the performance much? What would you
expect to happen in the filters of the DFE? ( 2 pts).
3.44 Final 2001 - Diversity Concept - 7 pts
An additive white Gaussian noise channel supports QAM transmission with a symbol rate of 1/T =
100 kHz (with 0 % excess bandwidth) anywhere in a total baseband-equivalent bandwidth of 10 MHz
transmission. Each 100 kHz wide QAM signal can be nominally received with an SNR of 14.5 dB without
equalization. However, this 10 MHz wide channel has a 100 kHz wide band that is attenuated by 20
dB with respect to the rest of the band, but the location of the frequency of this notch is not known in
advance to the designer. The transmitters are collocated.
273
a. What is the data rate sustainable at probability of bit error
¯
P
b
≤ 10
−7
in the nominal condition?
(1 pt)
b. What is the maximum number of simultaneous QAM users can share this channel at the perfor-
mance level of part a if the notch does not exist? (1 pt)
c. What is the worst-case probability of error for the number of users in part b if the notch does
exist? (1 pt)
d. Suppose now (unlike part b) that the notch does exist, and the designer decides to send each
QAM signal in two distinct frequency bands. What is the worst-case probability of error for a
diversity-equalization system applied to this channel? (1 pt)
e. For the same system as part c, what is the best SNR that a diversity equalizer system can achieve?
(1 pt)
f. For the system of part c, how many users can now share the band all with probability of bit error
less than 10
−7
? Can you think of a way to improve this number closer to the level of the part b?
(2 pts)
3.45 Final 2002 - Diversity Viewpoint - 12 pts
A time-varying wireless channel is used to send long packets of information that are decoded by a
receiver that uses a symbol-by-symbol decision device. The channel has pulse response in the familiar
form:
P(ω) =
_ √
T
_
1 + a
i
e
−ωT
_
[ω[ ≤
π
T
0 [ω[ >
π
T
(3.547)
where a
i
may vary from packet to packet (but not within a packet). There is also AWGN with constant
power spectral density
N0
2
= .1. QPSK (4QAM) is sent on this channel with symbol energy 2.
a. Find the best DFE SNR and corresponding probability of symbol error for the two cases when
a
1
= .9 and a
2
= .5. (2 pts)
Now, suppose the transmitter can resend the same packet of symbols to the receiver, which can
delay the channel output packet from the first transmission until the symbols from the new packet
arrive. However, the same symbol-by-symbol detector is still to be used by the receiver.
b. Explain why this is a diversity situation. (1 pt)
c. Find a new SNR
MFB
for this situation. (1 pt)
d. Find the new function Q(D) for this diversity situation. (2 pts)
e. Determine the performance of the best DFE this time (SNR and P
e
). ( 2 pts)
f. Compare the P
e
of part e with the product of the two P
e
’s found in part a. Comment. (2 pts)
g. Draw the receiver with RAKE illustrated as well as DFE, unbiasing, and decision device. (phase
splitter can be ignored in diagram and all quantities can be assumed complex.) (2 pts).
3.46 Precoder Diversity - Final 2003 - 7 pts
A system with 4PAM transmits over two discrete-time channels shown with the two independent
AWGN shown. (
¯
c
x
= 1)
a. Find the RAKE matched filter(s) for this system ? (1 pt)
b. Find the single feedforward filter for a ZF-DFE? (1 pt)
c. Show the best precoder for the system created in part b. (1 pt)
274
Figure 3.69: Transmission system for Problem 3.47.
d. Find the SNR at the detector for this precoded system? (1 pt)
e. Find the loss in SNR with respect to the of either single channel. (1 pt)
f. Find the P
e
for this precoded diversity system. (1 pt)
g. Find the probability that a packet of 1540 bytes contains one or more errors. (1 pt)
3.47 Equalizer Performance and Means
Recall that arithmetic mean, geometric mean, and harmonic mean are of the form
1
n

n
i=1
x
i
,
(

n
i=1
x
i
)
1
n
,
_
1
n

n
i=1
1
xi
_
−1
, respectively. Furthermore, they satisfy the following inequalities:
1
n
n

i=1
x
i

_
n

i=1
x
i
_1
n

_
1
n
n

i=1
1
x
i
_
−1
,
with equality when x
1
= x
2
= = x
n
a. (4 pts) Express SNR
ZFE
, SNR
ZF−DFE
, SNR
MFB
in terms of SNR
MFB
and frequency response
of the autocorrelation function q
k
. Prove that SNR
ZFE
≤ SNR
ZF−DFE
≤ SNR
MFB
using above
inequalities. When does equality hold? hint: use
_
b
a
x(t)dt = lim
n→∞

n
k=1
x(
b−a
n
k + a)
b−a
n
b. (4 pts) Similarly, prove that SNR
MMSE−LE,U
≤ SNR
MMSE−DFE,U
≤ SNR
MFB
using above
inequalities. When does equality hold?
c. (4 pts) Compare SNR
ZF−DFE
and SNR
MMSE−LE
. Can you say which scheme has better
performance?
3.48 Unequal Channels (Moshe Malkin- instructor, Midterm 2007)
Please assume the gap approximation works in all SNR ranges for this problem.
275
a. A sequence of 16-QAM symbols with in-phase component a
k
and quadrature component b
k
at time
k is transmitted on a passband channel by the modulated signal
x(t) =

2
__

k
a
k
φ(t −kT)
_
cos(ω
c
t) −
_

k
b
k
φ(t − kT)
_
sin(ω
c
t)
_
, (3.548)
where T = 40 ns, f
c
= 2.4 GHz, and the transmit power is 700 µW.
The waveform φ(t) is given by
φ(t) =
1

T
sinc
_
t
T
_
(3.549)
and the channel response is given by H(f) = e
−j5πfT
and so the received signal is given by
y(t) = h(t) ∗ x(t) + n(t) (3.550)
where n(t) is an additive white Gaussian noise random process with two-sided PSD of -100 dBm/Hz.
(i) (1 pts) What is the data rate for this system?
(ii) (2 pts) What are c
x
,
¯
c
x
and d
min
for this constellation?
(iii) (1.5 pts) What is ˜ x
bb
(t)?
(iv) (1.5 pts) What is the ISI characterizing function q(t)? what is q
k
?
(v) (2 pts) What are P
e
and
¯
P
e
for an optimal ML detector?
(vi) (2 pts) Draw a block diagram of the receiver and label major components.
Unfortunately, the implementation of the baseband demodulator is faulty and the additive
Gaussian noise power in the quadrature component is α = 4 times what it should be. That
is, the noise variance of the quadrature component is α
N0
2
= 2N
0
while the noise variance in
the in-phase dimension is
N0
2
. Answer the following questions for this situation.
(vii) (2.5 pts) What is the receiver SNR?
(viii) (3 pts) What are P
e
and
¯
P
e
?
(ix) (1.5 pts) Instead of using QAM modulation as above, two independent 4-PAM constellations
(with half the total energy allocated for each constellation) are now transmitted on the in-
phase and quadrature channels, respectively. What is P
e
averaged over both of the PAM
constellations?
(x) (3 pts) The transmitter realizes that if he allocates different energies for the in-phase and
quadrature channels (BUT still using QAM modulation with an equal number of
bits for the in-phase and quadrature channels and original power constraint of 700
µW) he could improve the
¯
P
e
performance from (aviii). What is the optimal energy
allocation between the in-phase and quadrature channels? Given that we need
¯
P
e
= 10
−6
,
what is
¯
b and the achievable data rate?
(xi) (2.5 pts) Using the results from part (ax) find the minimal increase in transmit power needed
to guarantee
¯
P
e
= 10
−6
for 16-QAM transmission?
The transmitter now decides to reduce losses. Given this increased noise in the quadra-
ture dimension, the transmitter decides to use only the in-phase dimension for transmission
(but subject to the original transmit power constraint of 700 µW).
(xii) (1.5 pts) What data rate can be then supported that gives
¯
P
e
= 10
−6
?
(xiii) (3 pts) Using a fair comparison, compare the QAM transmission system from part (ax) to
the single-dimensional scheme (N = 1) from part (axii). Which scheme is better?
276
(xiv) (3 pts) Can you derive a new transmission scheme that uses both the in-phase and quadrature
channels to get a higher data rate than part (ax)? (subject to the same
¯
P
e
= 10
−6
, and
original symbol rate and power constraint of 700 µW)
3.49 Equalizer Design (9 pts, Moshe Malkin- instructor, Final 2007)
A transmission system uses the basis function φ(t) = sinc(t) with
1
T
= 1 Hz. The Fourier transform
of the channel impulse response is:
H(f) =
_
1
1−0.5j·e
−j2πf
, for[f[ ≤
1
2
0, for[f[ >
1
2
.
(3.551)
and SNR =
¯ εx
σ
2
= 10dB.
a. (1 pt) Find [[p[[
2
and Q(D).
b. (1 pt) Find W
ZFE
(D) and W
MMSE−LE
(D).
c. (1 pt) Find W(D) and B(D) for ZF-DFE.
d. (2 pts) Find W(D) and B(D) for MMSE-DFE (Hint: the roots of the numerator of Q(D) +
1/SNR
MFB
are 22.4555j and 0.0445j).
e. (0.5 pts) What is SNR
ZF−DFE
?
f. (1.5 pts) Design a Tomlinson-Harashima precoder based on the ZF-DFE. Show both the precoder
and the corresponding receiver (For any M).
g. (2 pts) Design a Laroia precoder based on the ZF-DFE. Assume you are using 4-QAM modulation.
Show both the precoder and the corresponding receiver. Make sure to mention which kind of
constellation is used at each SBS detector in your drawing.
3.50 Diversity (7 pts Moshe Malkin- instructor, Final 2007)
A receiver has access to an infinite number of diversity channels where the output of channel i at
time k is
y
k,i
= x
k
+ 0.9 x
k+1
+n
k,i
, i = 0, 1, 2, . . .
where n
k,i
is a Gaussian noise process independent across time and across all the channels, and with
variance σ
2
i
= 1.2
i
0.181. Also,
¯
c
x
= 1.
a. (1 pts) Find SNR
MFB,i
, the SNR
MFB
for channel i.
b. (2 pts) What is the matched filter for each diversity channel?
c. (1.5 pts) What is the resulting Q(D) for this diversity system?
d. (2.5 pts) What is the detector SNR for this system for a ZF-DFE receiver?
3.51 DFE with Finite Length Feedback Filter (9 pts Moshe Malkin- instructor, Final 2007)
A symbol-rate sampled MMSE-DFE structure, where the received signal is filtered by a continuous
time matched filter, uses an infinite length feedforward filter. However, the MMSE-DFE feedback filter
to be of finite length N
b
.
a. (2 pts) Formulate the optimization problem you would have to solve to find the optimal feedfor-
ward and feedback coefficients to minimize the mean squared error (Hint: Don’t approach this as
a finite-length equalization problem as in section 3.7).
Assume that
N0
2
= .181, c
x
= 1, and P(D) = 1 + .9 D
−1
. (and SNR
MFB
= 10 dB).
277
b. (2 pts) Using only 1 feedback tap (N
b
= 1), determine the optimal feedforward and feedback filters
(Hint: This is easy - you have already seen the solution before).
c. (2 pts) Now consider the ZF-DFE version of this problem. State the optimization problem for this
situation (analogous to part a).
Now, assume that we have P(D) =
1
1+0.9D
−1
(and |p|
2
=
1
1−0.9
2
=
1
0.19
).
d. (3 pts) Determine the optimal feedforward and feedback filters for the ZF-DFE formulation using
only 1 feedback tap (N
b
= 1).
3.52 Packet Processing (10 pts Moshe Malkin- instructor, Final 2007)
Suppose that in a packet-based transmission system, a receiver makes a decision from the set of
transmit symbols x
1
, x
2
, x
3
based on the channel outputs y
1
, y
2
, y
3
, where
_
_
y
3
y
2
y
1
_
_
=
_
_
p
11
p
12
p
13
0 p
22
p
23
0 0 p
33
_
_
_
_
x
3
x
2
x
1
_
_
+
_
_
n
3
n
2
n
1
_
_
=
_
_
0.6 1.9 −3.86
0 1.8 3.3
0 0 1.2
_
_
_
_
x
3
x
2
x
1
_
_
+
_
_
n
3
n
2
n
1
_
_
(3.552)
and where the noise autocorrelation matrix is given by
R
n
= E
_
_
_
_
n
3
n
2
n
1
_
_
_
n

3
n

2
n

1
¸
_
_
=
_
_
0.3 0 0
0 0.5 0
0 0 0.1
_
_
. (3.553)
The inverse of the channel matrix is given by
_
_
0.6 1.9 −3.86
0 1.8 3.3
0 0 1.2
_
_
−1
=
_
_
1.67 −1.76 10.19
0 0.56 −1.52
0 0 0.83
_
_
. (3.554)
The transmit symbols are i.i.d with c
§∞
= c
§∈
= c
§
= ∈/.
a. (1.5 pts) What is the 3-tap zero-forcing equalizer for x
1
based on the observation y
1
, y
2
, y
3
? Repeat
for x
2
and x
3
.
b. (1 pts) What is the matrix ZFE for detecting x
1
, x
2
, and x
3
? (Hint: This is easy - all the work
occurred in the previous part)
c. (1.5 pts) What is the detection SNR for x
1
using a ZFE? Answer the same for x
2
and x
3
.
d. (1 pts) An engineer now realizes that the performance could be improved through use previous
decisions when detecting the current symbol. If the receiver starts by detecting x
1
, then x
2
, and
finally x
3
, describe how this modified ZF detector would work?
e. (1.5 pts) Assuming previous decisions are correct, what is now the detection SNR for each symbol?
f. (1 pts) The approach above does not necessarily minimize the MMSE. How would a packet MMSE-
DFE work? How would you describe the feedback structure? (Don’t solve for the optimal matrices
- just describe the signal processing that would take place).
g. (1 pts) Can the THP also be implemented for this packet based system? Describe how this would
be done (again, don’t solve - just describe how this would work).
278
Appendix A
Useful Results in Linear Minimum
Mean-Square Estimation
This appendix derives and/or lists some key results from linear MinimumMean-Square-Error estimation.
A linear MMSE estimate of some random variable x, given observations of some other random
sequences ¦y
k
¦, chooses a set of parameters w
k
(with the index k having one distinct value for each
observation of y
k
that is used) to minimize
E
_
[e[
2
¸
(A.1)
where
e

= x −
N−1

k=0
w
k
y
k
. (A.2)
One can let N → ∞ without difficulty. The term linear MMSE estimate is derived from the linear
combination of the random variables y
k
in (A.2). More generally, it can be shown that the MMSE
estimate is E[x/¦y
k
¦], the conditional expectation of x, given y
0
, ..., y
N−1
. which will be linear for
jointly Gaussian x and ¦y
k
¦. This text is interested only in linear MMSE estimates. In this text’s
developments, y
i
will usually be successive samples of some channel output sequence, and x will be the
current channel-input symbol that the receiver attempts to estimate.
A.1 The Orthogonality Principle
While (A.1) can be minimized directly by differentiation in each case or problem of interest, it is often
far more convenient to use a well-known principle of linear MMSE estimates - that is that the error
signal e must always be uncorrelated with all the observed random variables in order for the MSE to be
minimized.
Theorem A.1.1 (Orthogonality Principle) The MSE is minimized if and only if the
following condition is met
E[e y

k
] = 0 ∀ k = 0, ..., N − 1 (A.3)
Proof: By writing [e[
2
= ['(e)]
2
+ [·(e)]
2
, we may differentiate the MSE with respect to
both the real and imaginary parts of w
k
for each k of interest. The pertinent parts of the real
and imaginary errors are (realizing that all other w
i
, i ,= k, will drop from the corresponding
derivatives)
e
r
= x
r
−w
r,k
y
r,k
+ w
i,k
y
i,k
(A.4)
e
i
= x
i
− w
i,k
y
r,k
−w
r,k
y
i,k
, (A.5)
279
where subscripts of r and i denote real and imaginary part in the obvious manner. Then, to
optimize over w
r,k
and w
i,k
,
∂[e[
2
∂w
r,k
= 2e
r
∂e
r
∂w
r,k
+ 2e
i
∂e
i
∂w
r,k
= −2 (e
r
y
r,k
+e
i
y
i,k
) = 0 (A.6)
∂[e[
2
∂w
i,k
= 2e
r
∂e
r
∂w
i,k
+ 2e
i
∂e
i
∂w
i,k
= 2 (e
r
y
i,k
− e
i
y
r,k
) = 0 . (A.7)
The desired result is found by taking expectations and rewriting the series of results above
in vector form. Since the MSE is positive-semi-definite quadratic in the parameters w
r,k
and
w
i,k
, this setting must be a global minimum. If the MSE is strictly positive-definite, then
this minimum is unique. QED.
A.2 Spectral Factorization
This section uses D-transform notation for sequences:
Definition A.2.1 (D- Transforms) A sequence ¦x
k
¦ has D-TransformX(D) =

k
x
k
D
k
.
In this chapter, all sequences can be complex. The sequence ¦x

−k
¦ has a D-Transform X

(D
−∗
)

=

k
x

k
D
−k
.
Definition A.2.2 (Autocorrelation and Power Spectrum) If ¦x
k
¦ is any stationary
complex sequence, its autocorrelation function is defined as the sequence R
xx
(D) whose
terms are r
xx,j
= E[x
k
x

k−j
]; symbolically
1
R
xx
(D)

= E
_
X(D) X

(D
−∗
)
¸
. (A.8)
By stationarity, r
xx,j
= r

xx,−j
and R
xx
(D) = R

xx
(D
−∗
). The power spectrum of a
stationary sequence is the Fourier transform of its autocorrelation function, which is written
as
R
xx
(e
−ωT
) = R
xx
(D)[
D=e
−ωT , −
π
T
< ω ≤
π
T
, (A.9)
which is real and nonnegative for all ω. Conversely, we say that any complex sequence R(D)
for which R(D) = R

(D
−∗
) is an autocorrelation function, and any function R(e
−ωT
) that
is real and nonnegative over the interval ¦−
π
T
< ω ≤
π
T
¦ is a power spectrum.
The quantity E
_
[x
k
[
2
¸
is c
x
, or
¯
c
x
per dimension, and is determined by either the autocorrelation
function or the power spectrum as follows:
c
x
= E
_
[x
k
[
2
¸
= r
xx,0
=
T

_ π
T

π
T
R(e
−ωT
)dω . (A.10)
The spectral factorization of an autocorrelation function is:
Definition A.2.3 (Factorizability) An autocorrelation function R(D) will be called fac-
torizable if it can be written in the form
R(D) = S
0
F(D) F

(D
−∗
), (A.11)
where S
0
is a positive real number and F(D) is a canonical filter response. A filter response
F(D) is called canonical if it is causal (f
k
= 0 for k < 0), monic (f
0
= 1), and minimum-
phase (all of its poles are outside the unit circle, and all of its zeroes are on or outside the
unit circle). If F(D) is canonical, then F

(D
−∗
) is anticanonical; i.e., anticausal, monic,
and maximum-phase.
1
The expression Rxx(D)

= E
_
X(D)X

(D
−1
)
¸
is used in a symbolic sense, since the terms of X(D)X

(D
−1
) are of
the form

k
x
k
x

k−j
, implying the additional operation lim
N→∞
[1/(2N + 1)]

−N≤k≤N
on the sum in such terms.
280
If F(D) is a canonical response, then |f|
2

=

j
[f
j
[
2
≥ 1, with equality if and only if F(D) = 1,
since F(D) is monic. This simple inequality has many uses, see Exercises for Chapter 3.
The following lemma from discrete-time spectral factorization theory describes when an autocorre-
lation function is factorizable, and also gives the value of S
0
.
Lemma A.2.1 (Spectral Factorization) If R(e
−ωT
) is any power spectrum such that
both R(e
−ωT
) and logR(e
−ωT
) are integrable over −
π
T
< ω ≤
π
T
, and R(D) is the corre-
sponding autocorrelation function, then there is a canonical discrete-time response F(D) that
satisfies the equation
R(D) = S
0
F(D) F

(D
−∗
), (A.12)
where the finite constant S
0
is given by
logS
0
=
1

_ π
T

π
T
log R(e
−ωT
)dω (A.13)
(where the logarithms can have any common base). For S
0
to be finite, R(e
−ωT
) satisfies
the discrete-time Paley-Wiener criterion
1

_ π
T

π
T
[ logR(e
−ωT
)[dω < ∞ . (A.14)
Linear Prediction
The inverse of R(D) is also an autocorrelation function and can be factored when R(D) also satisfies
the PW criterion with finite γ
0
. In this case, as with the MMSE-DFE in Section 3.6, the inverse
autocorrelation factors as
R
−1
xx
(D) = γ
0
G(D) G

(D
−∗
) . (A.15)
If x(D) is a discrete-time sequence with factorizable inverse autocorrelation function R
−1
xx
(D) =
γ
0
G(D)G

(D
−∗
), and A(D) is any causal and monic sequence, then 1−A(D) is a strictly causal sequence
that may be used as a prediction filter, and the prediction error sequence E(D) is given by
E(D) = X(D) −X(D)[1 − A(D)] = X(D)A(D) . (A.16)
The autocorrelation function of the prediction error sequence is
R
ee
(D) = R
xx
(D) A(D) A

(D
−∗
) =
A(D) A

(D
−∗
)
γ
0
G(D) G

(D
−∗
)
, (A.17)
so its average energy satisfies c
e
= S
0
|1/g ∗ a|
2
≥ S
0
(since A(D)/G(D) is monic), with equality if and
only if A(D) is chosen as the whitening filter A(D) = G(D). The process X(D) G(D) is often called
the innovations of the process X(D), which has mean square value S
0
= 1/γ
0
. Thus, S
0
of the direct
spectral factorization is the mean-square value of the innovations process or equivalent of the MMSE
in linear prediction. X(D) can be viewed as being generated by inputting a white innovations process
U(D) = G(D)X(D) with mean square value S
0
into a filter F(D) so that X(D) = F(D)U(D).
The factorization of the inverse and resultant interpretation of the factor G(D) as a linear-prediction
filter helps develop an interest interpretation of the MMSE-DFE in Section 3.6.
A.2.1 Cholesky Factorization
Cholesky factorization is the finite-length equivalent of spectral factorization. In this finite-length case,
there are really two factorizations, both of which converge to the infinite-length spectral factorization.
281
Cholesky Form 1 - Forward Prediction
Cholesky factorization of a positive semidefinite N N matrix R
N
produces a unique upper triangular
monic (ones along the diagonal) matrix F
N
and a unique diagonal positive-semidefinite matrix S
N
such
that
R
N
= F
N
S
N
F

N
. (A.18)
It is convenient in transmission theory to think of the matrix R
N
as an autocorrelation matrix for N
samples of some random vector process x
k
with ordering
X
N
=
_
¸
_
x
N−1
.
.
.
x
0
_
¸
_ . (A.19)
A corresponding order of F
N
and S
N
is then
F
N
=
_
¸
¸
¸
_
f
N−1
f
N−2
.
.
.
f
0
_
¸
¸
¸
_
and S
N
=
_
¸
¸
¸
_
s
N−1
0 ... 0
0 s
N−2
... 0
0
.
.
.
.
.
. 0
0 0 ... s
0
_
¸
¸
¸
_
. (A.20)
Since F
N
is monic, it is convenient to write
f
i
=
_
1
˜
f
i
_
. (A.21)
The determinant of R
N
is easily found as
γ
2
0
= [R
N
[ =
N−1

n=0
s
n
. (A.22)
(or lnγ
2
0
= ln [R
N
[ for readers taking limits as N → ∞). A convenient description of R
N
’s components
is then
R
N
=
_
r
N
r

1
r
1
R
N−1
_
. (A.23)
The submatrix R
N−1
has Cholesky Decomposition
R
N−1
= F
N−1
S
N−1
F
N−1
, (A.24)
which because of the 0 entries in the triangular and diagonal matrices shows the recursion inherent in
Cholesky decomposition (that is the F
N−1
matrix is the lower right (N − 1) (N − 1) submatrix of
F
N
, which is also upper triangular). The inverse of R
N
(pseudoinverse when singular) has a Cholesky
factorization
R
−1
N
= G

N
S
−1
N
G
N
(A.25)
where F
N
= G
−1
N
is also upper triangular and monic with ordering
G
N
=
_
¸
¸
¸
_
g
N−1
g
N−2
.
.
.
g
0
_
¸
¸
¸
_
. (A.26)
Also,
g
i
= [1 ˜ g
i
] . (A.27)
282
linear prediction A straightforward algorithm for computing Cholesky factorization derives from a
linear prediction interpretation. The innovations, U
N
, of the N samples of X
N
are defined by
X
N
= FU
N
(A.28)
where E[U
N
U

N
] = S
N
, and
U
N
=
_
¸
_
u
N−1
.
.
.
u
0
_
¸
_ . (A.29)
Also, U
N
= GX
N
. The cross-correlation between X
N
and U
N
is
¯
R
ux
= S
N
G

N
, (A.30)
which is lower triangular. Thus,
E
_
u
k
x

k−i
¸
= 0 ∀ i ≥ 1 . (A.31)
Since X
N−1
= F
N−1
U
N−1
shows a reversible mapping from U
N−1
to X
N−1
, then (A.31) relates that
the sequence u
k
is a set of MMSE prediction errors for x
k
in terms of x
k−1
... x
0
(i.e., (A.31) is the
orthogonality principle for linear prediction). Thus,
u
N
= x
N
− r

1
R
−1
N−1
X
N−1
(A.32)
since r
1
is the cross-correlation between x
N
and X
N−1
. Equation (A.32) rewrites as
x
N
= u
N
+ r

1
R
−1
N−1
X
N−1
(A.33)
= U
N
+r

1
G

N−1
S
−1
N−1
. ¸¸ .
˜
f
N
G
N−1
X
N−1
. ¸¸ .
UN−1
(A.34)
= f
N
U
N
, (A.35)
so
f
N
=
_
1 r

1
G

N−1
S
−1
N−1
¸
. (A.36)
Then, from (A.32) and (A.34),
g
N
=
_
1 −
˜
f
N
G
N−1
_
. (A.37)
Finally,
s
N
= r
N

˜
f
N
S
N−1
˜
f

N
. (A.38)
Forward Cholesky Algorithm Summary: For nonsingular R
N
:
Set f
0
= F
0
= g
0
= G
0
= 1, S
0
= s
0
= E[x
0
[
2
For n = 1 ... N:
a.
˜
f
n
= r

1
G

n−1
S
−1
n−1
.
b. F
n
=
_
1
˜
f
n
0 F
n−1
_
.
c. G
n
=
_
1 −
˜
f
n
G
n−1
0 G
N−1
_
.
d. s
n
= r
n

˜
f
N
S
N−1
˜
f

N−1
.
283
A singular R
N
means that s
n
= 0 for at least one index n = i, which is equivalent to u
i
= 0, meaning
that x
i
can be exactly predicted from the samples x
i−1
... x
0
or equivalently can be exactly constructed
from u
i−1
... u
0
. In this case, Cholesky factorization is not unique. One factorization is to save F
i
and G
i
and “restart” the algorithm for n > i realizing that x
i
is equivalent to knowing all lower-index samples.
The new F
n>i
and G
n>i
created will have smaller dimension, but when the algorithm is complete, zeros
can be added to all the “right-most” positions of F and G, and maintaining the previous F
i
and G
i
for
the lower rightmost positions.
That is,
F
N
=
_
F
n>i
0
0 F
i
_
and G
N
=
_
G
n>i
0
0 G
i
_
(A.39)
and
S
N
=
_
_
S
n>i
0 0
0 0 0
0 0 S
i−1
_
_
. (A.40)
The forward Cholesky factorization can also be written
R
N
= G
−1
N
S
N
G
−∗
N
, (A.41)
which is convenient for the alternative finite-length MMSE-DFE interpretations in Section 3.7.
Cholesky Form 2 - Backward Prediction
The backward Cholesky factorization is into a lower-diagonal-upper (instead of the upper-diagonal-lower
in forward prediction). This involves extension of R
N
on the lower right corner instead of at the upper-
left corner. Rather than repeat the development of forward prediction, this section notes that the reverse
backward prediction problem is inherent in the algorithm already given, except that this algorithm is
performed on the inverse (which, granted, might be harder to compute than just repeating the steps
above with the appropriate transpositions).
The inverse matrix Q
N
= R
−1
N
= G

N
S
−1
N
G
N
is also an autocorrelation matrix for some process
Y
N
with innovations V
N
where Y
N
= G

N
V
N
such that E[V
N
V

N
] = S
−1
N
. If Q is singular, then its
pseudoinverse is used by definition here. Thus,
V
N
= F

N
Y , (A.42)
illustrating the backward prediction error v
1−N
=
¯
f

N
Y
N−1
where
¯
f

N
is the N
th
row of F

N
or equiva-
lently
¯
f
N
is the last column of F
N
.
This factorization of Q is lower-diagonal-upper and the filters
¯
f
n
are backward prediction filters for
y
1−n
in terms of y
2−n
... y
0
. So to perform backward-prediction Cholesky, first invert the given Q
autocorrelation matrix to get R and then do forward Cholesky on the inverse (R
N
= F
N
S
N
F

N
=
G
−1
N
SG
−∗
N
. Then, the backward Cholesky factorization uses the G of forward Cholesky (for R = Q
−1
)
in
Q = G

N
S
−1
N
G
N
. (A.43)
Infinite-length convergence
For infinite-length stationary sequences, one takes the limit as N → ∞ in either forward or backward
Cholesky factorization. In this case, the matrix R (and therefore Q) must be nonsingular to satsify
the Paley-Weiner Criterion. The equivalence to spectral factorization is evident from the two linear-
prediction interpretations for finite-length and infinite length series of samples from random processes.
For the stationary case, the concepts of forward and backward prediction are the same so that the
forward predictor is just the time reverse of the coefficents in the backward predictor.
Thus, the inverse autocorrelation function factors as
R
−1
(D) = γ
0
G(D)G

(D
−∗
) , (A.44)
284
where G(D) is the forward prediction polynomial (and its time reverse specified by G

(D
−∗
) is the
backward prediction polynomial). The series of matrices ¦R
n
¦
n=1:∞
formed from the coefficients of
R(D) creates a series of linear predictors ¦g
n
¦
n=1:∞
with D-transforms G
n
(D). In the limit as N →∞
for a stationary nonsingular series,
lim
n→∞
G
n
(D) = G(D) . (A.45)
Similarly,
lim
n→∞
G

n
(D) = G

(D
−∗
) . (A.46)
As N → ∞, the prediction error varianc,es S
N−1
, should tend to a constant. Finally, defining the
geometric-average determinants as S
0,n
= [R[
1/N
and γ
0,N
= [R
−1
[
1/N
lim
n→∞
S
0,n
= S
0
= e
_
T

_
π/T
−π/T
ln(R(e
−ωT
))dω
_
(A.47)
lim
n→∞
γ
0,n
(D) = γ
0
= e

_
T

_
π/T
−π/T
ln(R(e
−ωT
))dω
_
. (A.48)
The convergence to these limits implies that the series of filters converges or that the bottom row (last
column) of the Cholesky factors tends to a constant repeated row.
Interestingly, a Cholesky factorization of a singular process exists only for finite lengths. Using the
modifications to Cholesky factorization suggested above with “restarting,” it becomes obvious why such
a process cannot converge to a constant limit and so only nonsingular processes are considered in spectral
factorization. Factorization of a singular process at infinite-length involves separating that process into
a sum of subprocesses, each of which is resampled at a new sampling rate so that the PW criterion
(nonsingular needs PW) is satisfied over each of the subbands associated with these processes. This is
equivalent to the “restarting” of Cholesky at infinite length. For more, see Chapter 5.
285
Appendix B
Equalization for Partial Response
Just as practical channels are often not free of intersymbol interference, such channels are rarely equal
to some desirable partial-response (or even controlled-ISI) polynomial. Thus, an equalizer may be used
to adjust the channel’s shape to that of the desired partial-response polynomial. This equalizer then
enables the use of a partial-response sequence detector or a symbol-by-symbol detector at its output
(with partial-response precoder). The performance of either of these types of detectors is particularly
sensitive to errors in estimating the channel, and so the equalizer can be crucial to achieving the highest
levels of performance in the partial-response communication channel. There are a variety of ways in
which equalization can be used in conjunction with either sequence detection and/or symbol-by-symbol
detection. This section introduces some equalization methods for partial-response signaling.
Chapter 9 studies sequence detection so terms related to it like MLSD (maximum likelihood sequence
detection) and Viterbi detectors/algorithm are pertinent to readers already familiar with Chapter 9
contents and can be ignored by other readers otherwise in this Appendix.
B.1 Controlled ISI with the DFE
Section 3.8.2 showed that a minimum-phase channel polynomial H(D) can be derived from the feedback
section of a DFE. This polynomial is often a good controlled intersymbol interference model of the
channel when H(D) has finite degree ν. When H(D) is of larger degree or infinite degree, the first ν
coefficients of H(D) form a controlled intersymbol interference channel. Thus, sequence detection on
the output of the feedforward filter of a DFE can be designed using the controlled ISI polynomial H(D)
or approximations to it.
B.1.1 ZF-DFE and the Optimum Sequence Detector
Section 3.1 showed that the sampled outputs, y
k
, of the receiver matched filter form a sufficient statistic
for the underlying symbol sequence. Thus a maximum likelihood (or MAP) detector can be designed
that uses the sequence y
k
to estimate the input symbol sequence x
k
without performance loss. The
feedforward filter 1/(η
0
|p|H

(D
−∗
)) of the ZF-DFE is invertible when Q(D) is factorizable. The re-
versibility theorem of Chapter 1 then states that a maximumlikelihood detector that observes the output
of this invertible ZF-DFE-feedforward filter to estimate x
k
also has no performance loss with respect to
optimum. The feedforward filter output has D-transform
Z(D) = X(D)H(D) + N

(D) . (B.1)
The noise sequence n

k
is exactly white Gaussian for the ZF-DFE, so the ZF-DFE produces an equivalent
channel H(D). If H(D) is of finite degree ν, then a sequence detector can be designed based on the
controlled-ISI polynomial H(D) and this detector has minimum probability of error. Figure B.1 shows
such an optimum receiver. When H(D) has larger degree than some desired ν determined by complexity
constraints, then the first ν feedback taps of H(D) determine
H

(D) = 1 + h
1
D
1
+... + h
ν
D
ν
, (B.2)
286
Figure B.1: The optimum maximum-likelihood detector, implemented with the assistance of the WMF,
the feedforward section of the ZF-DFE
Figure B.2: Limiting the number of states with ZF-DFE and MLSD.
a controlled-ISI channel. Figure B.2 illustrates the use of the ZF-DFE and a sequence detector. The
second feedback section contains all the channel coefficients that are not used by the sequence detector.
These coefficients have delay greater than ν. When this second feedback section has zero coefficients, then
the configuration shown in Figure B.2 is an optimum detector. When the additional feedback section
is not zero, then this structure is intermediate in performance between optimum and the ZF-DFE with
symbol-by-symbol detection. The inside feedback section is replaced by a modulo symbol-by-symbol
detector when precoding is used.
Increase of ν causes the minimum distance to increase, or at worst, remain the same (which is proved
simply by examining a trellis with more states and realizing that the smaller degree H

(D) is a sub trellis
for which distance cannot be greater. Thus, the ZF-DFE with sequence detector in Figure B.2 defines
a series of increasingly complex receivers whose performance approach optimum as ν →∞. A property
of a minimum-phase H(D) is that
ν

i=0
[h
i
[
2
= |h

|
2
(B.3)
is maximum for all ν

≥ 0. No other polynomial (that also preserves the AWGN at the feedforward filter
output) can have greater energy. Thus the SNR of the signal entering the Viterbi Detector in Figure
B.2,
¯
E
x
h

2
N
0
2
, also increases (nondecreasing) with ν. This SNR must be less than or equal to SNR
MFB
.
287
Figure B.3: Partial-response linear equalization.
B.1.2 MMSE-DFE and sequence detection
Symbol-by-symbol detection’s objective is to maximize SNR in an unbiased detector, and so SNR max-
imization was applied in Chapter 3 to the DFE to obtain the MMSE-DFE. A bias in symbol-by-symbol
detector was removed to minimize probability of error. The monic, causal, minimum-phase, and unbi-
ased feedback polynomial was denoted G
U
(D) in Section 3.6. A sequence detector can use the same
structures as shown in Figure B.1 and B.2 with H(D) replaced by G
u
(D). For instance, Figure B.4 is
the same as Figure B.2, with an unbiased MMSE-DFE’s MS-WMF replacing the WMF of the ZF-DFE.
A truncated version of G
U
(D) corresponding to H

(D) is denoted G

U
(D). The error sequence asso-
ciated with the unbiased MMSE-DFE is not quite white, nor is it Gaussian. So, a sequence detector
based on squared distance is not quite optimum, but it is nevertheless commonly used because the exact
optimum detector could be much more complex. As ν increases, the probability of error decreases from
the level of the unbiased MMSE-DFE, SNR
MMSE−DFE,U
when ν = 0, to that of the optimum detector
when ν →∞. The matched filter bound, as always, remains unchanged and is not necessarily obtained.
However, minimum distance does increase with ν in the sequence detectors based on a increasing-degree
series of G
U
(D).
B.2 Equalization with Fixed Partial Response B(D)
The derivations of Section 3.6 on the MMSE-DFE included the case where B(D) ,= G

(D), which this
section reuses.
B.2.1 The Partial Response Linear Equalization Case
In the linear equalizer case, the equalization error sequence becomes
E
pr
(D) = B(D)X(D) − W(D)Y (D) . (B.4)
Section 3.6 minimized MSE for any B(D) over the coefficients in W(D). The solution was found by
setting E
_
E
pr
(D)y

(D
−1
)
¸
= 0, to obtain
W(D) = B(D)
¯
R
xy
(D)
¯
R
yy
(D)
=
B(D)
|p| (Q(D) + 1/SNR
MFB
)
, (B.5)
which is just the MMSE-LE cascaded with B(D). Figure B.3, shows the MMSE-PREQ (MMSE -
“Partial Response Equalizer”). The designer need only realize the MMSE-LE of Section 3.4 and follow
it by a filter of the desired partial-response (or controlled-ISI) polynomial B(D).
1
For this choice of
W(D), the error sequence is
E
pr
(D) = B(D)X(D) − B(D)Z(D) = B(D) [E(D)] (B.6)
1
This also follows from the linearity of the MMSE estimator.
288
where E(D) is the error sequence associated with the MMSE-LE. From (B.6),
¯
R
epr,epr
(D) = B(D)
¯
R
ee
(D)B

(D
−1
) =
B(D)
N0
2
B

(D
−1
)
|p|
2
(Q(D) + 1/SNR
MFB
)
. (B.7)
Thus, the MMSE for the PREQ can be computed as
σ
2
MMSE-PREQ
=
T

_ π
T

π
T
[B(e
−ωT
)[
2 N0
2
|p|
2
(Q(e
−ωT
) + 1/SNR
MFB
)
dω . (B.8)
The symbol-by-symbol detector is equivalent to subtracting (B(D) −1)X(D) before detection based on
decision regions determined by x
k
. The SNR
MMSE-PREQ
becomes
SNR
MMSE-PREQ
=
¯
c
x
σ
2
MMSE-PREQ
. (B.9)
This performance can be better or worse than the MMSE-LE, depending on the choice of B(D); the
designer usually selects B(D) so that SNR
MMSE-PREQ
> SNR
MMSE−LE
. This receiver also has a
bias, but it is usually ignored because of the integer coefficients in B(D) – any bias removal could cause
the coefficients to be noninteger.
While the MMSE-PREQ should be used for the case where symbol-by-symbol and precoding are
being used, the error sequence E
pr
= B(D)E(D) is not a white noise sequence (nor is E(D) for the
MMSE-LE), so that a Viterbi Detector designed for AWGN on the channel B(D) would not be the
optimum detector for our MMSE-PREQ (with scaling to remove bias). In this case, the ZF-PREQ,
obtained by setting SNR
MFB
→∞ in the above formulae, would also not have a white error sequence.
Thus a linear equalizer for a partial response channel B(D) that is followed by a Viterbi Detector
designed for AWGN may not be very close to an optimum detection combination, unless the channel
pulse response were already very close to B(D), so that equalization was not initially necessary. While
this is a seemingly simple observation made here, there are a number of systems proposed for use in
disk-storage detection that overlook this basic observation, and do equalize to partial response, “color”
the noise spectrum, and then use a WGN Viterbi Detector. The means by which to correct this situation
is the PR-DFE of the next subsection.
B.2.2 The Partial-Response Decision Feedback Equalizer
If B(D) ,= G(D) and the design of the detector mandates a partial-response channel with polynomial
B(D), then the optimal MMSE-PRDFE is shown in Figure B.4.
Again, using our earlier result that for any feedback section G
U
(D) = B(D) +
˜
B(D). The error
sequence is the same as that for the MMSE-DFE, and is therefore a white sequence. The signal between
the the two feedback sections in Figure B.4 is input to the sequence or symbol-by-symbol detector. This
signal can be processed on a symbol-by-symbol basis if precoding is used (and also scaling is used to
remove the bias - the scaling is again the same scaling as used in the MMSE-DFE), and
˜
B
U
(D) =
G
U
(D) − B
U
(D), where
B
U
(D) =
SNR
MMSE−DFE
SNR
MMSE−DFE,U
_
B(D) −
1
SNR
MMSE−DFE
_
. (B.10)
However, since the MMSE-PRDFE error sequence is white, and because the bias is usually small so that
the error sequence in the unbiased case is also almost white, the designer can reasonably use a Viterbi
Detector designed for B(D) with white noise.
If the bias is not negligible, then a ZF-PRDFE should be used, which is illustrated in Figure B.2,
and the filter settings are obtained by setting SNR
MFB
→∞ in the above formula.
289
Figure B.4: Partial-response DFE with B(D)
290
Appendix C
The Matrix Inversion Lemma:
The matrix inversion lemma:
[A +BCD]
−1
= A
−1
− A
−1
B
_
C
−1
+DA
−1
B
¸
−1
DA
−1
. (C.1)
291

3.9

Diversity Equalization . . . . . . . . . . . . . . . . . . 3.9.1 Multiple Received Signals and the RAKE . . . 3.9.2 Infinite-length MMSE Equalization Structures 3.9.3 Finite-length Multidimensional Equalizers . . . 3.9.4 DFE RAKE Program . . . . . . . . . . . . . . 3.9.5 Multichannel Transmission . . . . . . . . . . . Exercises - Chapter 3 . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

247 247 249 251 252 254 256

A Useful Results in Linear Minimum Mean-Square Estimation 279 A.1 The Orthogonality Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279 A.2 Spectral Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280 A.2.1 Cholesky Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 B Equalization for Partial Response B.1 Controlled ISI with the DFE . . . . . . . . . . . . . . . . B.1.1 ZF-DFE and the Optimum Sequence Detector . . B.1.2 MMSE-DFE and sequence detection . . . . . . . . B.2 Equalization with Fixed Partial Response B(D) . . . . . . B.2.1 The Partial Response Linear Equalization Case . . B.2.2 The Partial-Response Decision Feedback Equalizer C The Matrix Inversion Lemma: 286 286 286 288 288 288 289 291

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

148

Chapter 3

Equalization
The main focus of Chapter 1 was a single use of the channel, signal set, and detector to transmit one of M messages, commonly referred to as “one-shot” transmission. In practice, the transmission system sends a sequence of messages, one after another, in which case the analysis of Chapter 1 applies only if the channel is memoryless: that is, for one-shot analysis to apply, these successive transmissions must not interfere with one another. In practice, successive transmissions do often interfere with one another, especially as they are spaced more closely together to increase the data transmission rate. The interference between successive transmissions is called intersymbol interference (ISI). ISI can severely complicate the implementation of an optimum detector. Figure 3.1 illustrates a receiver for detection of a succession of messages. The matched filter outputs are processed by the receiver, which outputs samples, z k , that estimate the symbol transmitted at time ˆ k, xk . Each receiver output sample is the input to the same one-shot detector that would be used on an AWGN channel without ISI. This symbol-by-symbol (SBS) detector, while optimum for the AWGN channel, will not be a maximum-likelihood estimator for the sequence of messages. Nonetheless, if the receiver is well-designed, the combination of receiver and detector may work nearly as well as an optimum detector with far less complexity. The objective of the receiver will be to improve the performance of this simple SBS detector. More sophisticated and complex sequence detectors are the topics of Chapters 4, 5 and 9. Equalization methods are used by communication engineers to mitigate the effects of the intersymbol interference. An equalizer is essentially the content of Figure 3.1’s receiver box. This chapter studies both intersymbol interference and several equalization methods, which amount to different structures for the receiver box. The methods presented in this chapter are not optimal for detection, but rather are widely used sub-optimal cost-effective methods that reduce the ISI. These equalization methods try to convert a bandlimited channel with ISI into one that appears memoryless, hopefully synthesizing a new AWGN-like channel at the receiver output. The designer can then analyze the resulting memoryless, equalized channel using the methods of Chapter 1. With an appropriate choice of transmit signals, one of the methods in this chapter - the Decision Feedback Equalizer of Section 3.6 can be generalized into a canonical receiver (effectively achieves the highest possible transmission rate even though not an optimum receiver), which is discussed further in Chapter 5 and occurs only when special conditions are met. Section 3.1 models linear intersymbol interference between successive transmissions, thereby both illustrating and measuring the problem. In practice, as shown by a simple example, distortion from overlapping symbols can be unacceptable, suggesting that some corrective action must be taken. Section 3.1 also refines the concept of signal-to-noise ratio, which is the method used in this text to quantify receiver performance. The SNR concept will be used consistently throughout the remainder of this text as a quick and accurate means of quantifying transmission performance, as opposed to probability of error, which can be more difficult to compute, especially for suboptimum designs. As Figure 3.1 shows, the objective of the receiver will be to convert the channel into an equivalent AWGN at each time k, independent of all other times k. An AWGN detector may then be applied to the derived channel, and performance computed readily using the gap approximation or other known formulae of Chapters 1 and

149

Figure 3. but several outputs. a method for eliminating error propagation in decision feedback equalizers and the related concept of partial-response channels and precoding. as discussed in Chapters 4 and 5. 150 . is yet a higher performance form of equalization for any channel with linear ISI. which is simple to understand but often of limited effectiveness. which will be measured by the SNR of the equivalent AWGN with respect to the best value that might be expected otherwise for an optimum detector.1: The band-limited channel with receiver and SBS detector. specifically introducing the Nyquist Criterion for a linear channel (equalized or otherwise) to be free of intersymbol interference. minimum mean square error linear (MMSE-LE) and decision-feedback equalizers (MMSE-DFE) are discussed in Sections 3.6. such as wireless transmission systems with multiple receive antennas called “diversity. There may be loss of optimality in creating such an equivalent AWGN.4 illustrates the basic concept of equalization through the zero-forcing equalizer (ZFE). Section 3. Section 3.” The area of transmit optimization for scalar equalization. 2 with the SNR of the derived AWGN channel. The more widely used and higher performance. Section 3. Section 3.7 discusses the design of finite-length equalizers.3 discusses some desired types of channel responses that exhibit no intersymbol interference.8 discusses precoding. Section 3.5 and 3.9 generalizes the equalization concepts to systems that have one input.

presumably found by Gram-Schmitt decomposition of the set of (noiseless) channel output waveforms (of which it can be shown K dimensions are sufficient only if N = 1.1. This performance degradation increases as T decreases (or the symbol rate increases) in most communication channels. and it may be possible to compute the ` posteriori probability function with less than exponentially growing complexity. Chapter 9 addresses such “sequence detectors” in detail.1 in view of various receiver corrective actions for ISI. the complexity can become too large for practical implementation. Interference between successive transmissions. including voiceband modems.1 The data rate of Chapter 1 for a communication system that sends one of M possible messages every T time units is R= ∆ log2 (M ) b = T T .2) Equation (3. x(t). and the term “baud” usually now only appears in trade journals and advertisements. complex or real).1. The transmitted signal x(t) corresponding to K successive transmissions is K−1 x(t) = k=0 xk (t − kT ) . From Section 1. and even fiber-optic (where dispersion-limited) cables. 151 . digital mobile radio channels. precedes a block detector that determines the K-dimensional vector symbol transmitted. A GramSchmidt decomposition on the set of M K signals would then be performed and an optimum detector designed accordingly.” although abuse of the term baud (by equating it with data rate even when M = 2) has rendered the term archaic among communication engineers. or intersymbol interference (ISI).3) 1 The symbol rate is sometimes also called the “baud rate. This section introduces a model for intersymbol interference. one for each possible transmitted block symbol. The K successive transmissions could be considered an aggregate or “block” symbol. but may not be optimum. (3. That is. As K → ∞. The designer mathematically analyzes ISI by rewriting (3. digital subscriber loop data transmission. (3. The bank of matched filters. digital microwave channels. where T is called the symbol period.3. storage disks.1 Transmission of Successive Messages Most communication systems re-use the channel to transmit several messages in succession.2 contrasts the SBS detector with the block detector of Chapter 1. and 1/T is called the symbol rate. a An alternative (suboptimal) receiver can detect each of the successive K messages independently. followed by a receiver and an SBS detector. the message transmissions are separated by T units in time.2) as K−1 N x(t) = k=0 n=1 xknϕn(t − kT ) . This section then continues and revisits the equivalent AWGN of Figure 3. Figure 3. conveying one of M K possible messages.1) To increase the data rate in a design. The receiver could attempt to implement MAP or ML detection for this new transmission system with M K messages. (3. can degrade the performance of symbol-bysymbol detection. Such an approach has complexity that grows exponentially (in proportion to M K ) with the block message length K. Decreasing T narrows the time between message transmissions and thus increases intersymbol interference on any band-limited channel.2) slightly abuses previous notation in that the subscript k on xk (t − kT ) refers to the index associated with the kth successive transmission. 3. either b can be increased (which requires more signal energy to maintain Pe ) or T can be decreased. The lower system in Figure 3.2 has a single matched filter to the channel. Such detection is called symbol-by-symbol (SBS) detection. The later system has fixed (and lower) complexity per symbol/sample.1 Intersymbol Interference and Receivers for Successive Message Transmission Intersymbol interference is a common practical impairment found in many transmission and storage systems. the optimal detector might need M K matched filters. with output sampled K times. The complexity would become large or infinite as K becomes large or infinite for the block detector.

3. In some cases.6) (with m a positive integer) are orthogonal for all integer-multiple-of-T time translations. This channel is used.5) or from Chapter 2. For instance. Signal sets for data transmission are usually designed to be orthogonal for any translation by an integer multiple of symbol periods. The filtering of the channel alters the basis functions so that at the channel output the filtered basis functions are no longer orthogonal. (3. when sampled at time instants kT .7. the successive transmissions. (3. and transmission is equivalent to a succession of “one-shot” uses of the channel. however.2 Bandlimited Channels The bandlimited linear ISI channel.Figure 3. however. is the same as the filtered AWGN channel discussed in Section 1. are more accurately modeled by a filtered AWGN channel as discussed in Section 1.1. the baseband equivalent 1 ϕ(t) = √ sinc T t T . for successive transmission of data symbols. shown in Figure 3. In this case. translates of the basis functions are orthogonal. the two bandlimited basis functions ϕ1 (t) ϕ2 (t) = = − 2 cos T 2 sin T mπt T mπt T · sinc · sinc t T t T . In this case symbol-by-symbol detection is optimal. The channel thus introduces ISI.3). In (3. ϕn (t − kT ) and ϕm (t − lT ) may be non-orthogonal when k = l.4) (3. where the transmissions xk (t) are decomposed using a common orthonormal basis set {ϕn(t)}. in QAM. Most linear AWGN channels.3.2: Comparison of Block and SBS detectors for successive transmission of K messages. 152 . are free of ISI. and the MAP detector for the entire block of messages is the same as a MAP detector used separately for each of the K independent transmissions.7 and Chapter 2.

Figure 3. 153 .3: The bandlimited channel. and equivalent forms with pulse response.

xp (t). such a theory would unnecessarily complicate the present development. p(t).12) = −∞ p(t)p∗ (t)dt = p(t). Equalization methods apply a processor.9) For the complex QAM case. in Figure 3.8) where pn (t) = ϕn (t) ∗ h(t).11) The signal in (3. Using (3. (3.g. Symbol-by-symbol detection can then be used on the equalized channel output. 154 . and this text introduces the normalized pulse response: ϕp (t) = where p 2 ∞ ∆ p(t) . p . PAM) or two-dimensional (e.7 considers diversity receivers that may have several observations a single or multiple channel output(s). The remainder of this chapter also presumes that the channel-input symbol sequence xk is independent and identically distributed at each point in time. (3. This was the main motivation for the introduction of bandpass analysis in Chapter 2.4. to the channel output to try to convert {pn(t − kT )} to an orthogonal set of functions. Using the baseband-equivalent systems of Chapter 2. In this way. This presumption will be relaxed in later Chapters.12). Equation (3. and h(t) can be complex time functions. QAM or hexagonal) constellation. the “equalizer”.g. When N = 1 (or N = 2 with complex signals).The (noise-free) channel output. while Section 3. this chapter’s (Chapter 3’s) analysis will also apply to quadrature modulated systems modeled as complex (equivalent to two-dimensional real) channels.14) 2 Chapter 4 considers multidimensional signals and intersymbol interference. An optimum (MAP) detector would need to search a signal set of size M K . ϕ(t). (3.13) The subscript p on ϕp (t) indicates that ϕp (t) is a normalized version of p(t). the functions pn(t − kT ) do not necessarily form an orthonormal basis. (3.2 This chapter handles the ISI-equalizer case for N = 1. While a very general theory of ISI could be undertaken for any N . there is only one pulse response p(t).3 is given by K−1 N xp (t) = k=0 n=1 K−1 N xkn · ϕn (t − kT ) ∗ h(t) (3. The pulse response for the transmitter/channel fundamentally quantifies ISI: Definition 3. The pulse response energy p 2 is not necessarily equal to 1. When h(t) = δ(t).1. The one-dimensional noiseless channel output xp (t) is K−1 xp (t) = k=0 K−1 xk · ϕ(t − kT ) ∗ h(t) (3.1 (Pulse Response) The pulse response of a bandlimited channel is defined by p(t) = ϕ(t) ∗ h(t) .11) is real for a one-dimensional system and complex for a baseband equivalent quadrature modulated system. Further discussion of such equalization filters is deferred to Section 3. nor are they even necessarily orthogonal.11) becomes K−1 xp (t) = k=0 xp.7) = k=0 n=1 ∆ xkn · pn (t − kT ) (3. which is often too complex for implementation as K gets large. the developed theory of ISI and equalization will apply equally well to any one-dimensional (e. p(t) (3.k · ϕp (t − kT ) .10) = k=0 xk · p(t − kT ) .

17) (3. according to h(t) = δ(t) + δ(t − T ). and thus has energy Ep = E |xp.Modified Duobinary) A PAM modulated signal using rectangular pulses is 1 ϕ(t) = √ (u(t) − u(t − T )) . (3.k absorbs the channel gain/attenuation p into the definition of the input symbol.14). Figure 3. consider the pulse response p(t) = 1+t4 and two successive transmissions of opposite polarity (−1 followed by +1) through the corresponding channel. Clearly.1 (Intersymbol interference and the pulse response) As an exam1 ple of intersymbol interference. EXAMPLE 3.4 illustrates the two isolated pulses with correct polarity and also the waveform corresponding to the two transmissions separated by 1 unit in time.k |2 = Ex · p 2. While the functions ϕp (t − kT ) are normalized.1.5 illustrates the resultant sum of the two waveforms for spacings of 1 unit in time. and .1 units in time.k = xk · p ∆ .5 units in time. Higher transmission rates would force successive transmissions to be closer and closer together. where xp. ISI has the effect of severely reducing pulse strength.16) 155 . so symbol-by-symbol detection is not necessarily optimal for the signal in (3.4: Illustration of intersymbol interference for p(t) = 1 1+t4 with T = 1.Figure 3. for example. (3. T The channel introduces ISI. Clearly the peaks of the pulses have been displaced in time and also significantly reduced in amplitude.2 (Pulse Response Orthogonality .15) xp. they are not necessarily orthogonal. Figure 3. EXAMPLE 3.1. thereby reducing immunity to noise. .

5: ISI with increasing symbol rate.Figure 3. 156 .

if 0 < p < ∞.6.k so that Exp = Ex · p 2. the transformation outputs are the filter samples y(kT ).7. by the reversibility theorem of Chapter 1. The following theorem illustrates that there is no loss in performance that is incurred via the matched-filter/sampler combination.k (t)}k∈(−∞. The invertible square-root can be realized as a filter in the beginning of a receiver.∞) followed by Γ−1 is p.e. Then.” An AWGN channel with a notch in H(f) at some frequency is thus equivalent to a “flat channel” with H(f) = 1.. The designer can then construe this filter as being pushed back into.21) (3.k (t)}k∈(−∞. The set {ϕp. ∆ (3. even though ϕ(t) and ϕ(t − T ) were originally orthonormal. xk is scaled by p to form xp. where Γ is the invertible transformation. The power spectral density of the noise is N0 · Sn (f).∞) is a linearly independent set of functions. The transformation and its inverse are written {φp. The concept of noise equivalence allows an analysis for AWGN to be valid (using the equivalent pulse response instead of the original pulse response). but with narrow-band Gaussian noise at the same frequency as the notch.k (t)}k∈(−∞. (i. In Figure 3. no information is lost.1. although correlated Gaussian noise can be included by transforming the correlated-Gaussian-noise channel into an equivalent white Gaussian noise channel using the methods in the previous subsection and illustrated in Figure 3.7 are sufficient to represent the continuous-time ISI-model channel output y(t). The additive noise is white Gaussian. Theorem 3.6. 2T (3. a receiver with minimum Pe can be designed that uses only the samples yk ). The infinite set of filters {φ∗ (−t)}k∈(−∞. In this model.1.k (t) = ϕp (t − kT ) .∞) by an invertible transformation Γ (use Gram-Schmidt an infinite number of times). 3.The resulting pulse response is 1 p(t) = √ (u(t) − u(t − 2T )) T and the normalized pulse response is 1 ϕp (t) = √ (u(t) − u(t − 2T )). but that do have “colored noise.k 157 . the channel as shown in the lower part of Figure 3.1 (ISI-Channel Model Sufficiency) The discrete-time signal samples yk = y(kT ) in Figure 3. Then also “colored noise” is equivalent in its effect to ISI.6 models a channel with additive Gaussian noise that is not white. which often occurs in practice. and furthermore the compensating equalizers that are developed later in this chapter can also be very useful on channels that originally have no ISI. When Sn (f) = 0 (noise is never exactly zero at 2 any frequency in practice).22) ({φp.k (t)}k∈(−∞. The channel output yp (t) is passed through a matched filter ϕ∗ (−t) to generate y(t).∞) is related to a set of orthogonal basis functions {φp.∞) ) Γ −1 (3.∞) = = Γ({ϕp.20) where {ϕp. Noise Equivalent Pulse Response Figure 3.k (t)}k∈(−∞.∞) {ϕp. the noise psd has an invertible square root as in Section 1.k (t)}k∈(−∞.3 The ISI-Channel Model A model for linear ISI channels is shown in Figure 3. 1/2 The noise equivalent pulse response then has Fourier Transform P (f)/Sn (f) for an equivalent filteredAWGN channel.8.∞) ) . y(t) is sampled at the p symbol rate and subsequently processed by a discrete time receiver. and thus a part of. Since this filter is invertible.19) (3.k (t)}k∈(−∞.7.18) The pulse-response translates ϕp (t) and ϕp (t − T ) are not orthogonal. Sketch of Proof: Define ϕp.

6: White Noise Equivalent Channel. Figure 3.7: The ISI-Channel model. 158 .Figure 3.

Referring to Figure 3. y(t) = k p · xk q(t − kT ) + np (t) ∗ ϕ∗ (−t) .9. the set of sampled filter outputs in Figure 3.∞) is orthonormal. so the symbol xk passes at time kT to the output with amplitude scaling p . as illustrated in Figure 3.k (t)}k∈(−∞.8: Equivalent diagram of ISI-channel model matched-filter/sampler equivalent to an infinite set of matched filters to ϕ∗ (−t) p.Figure 3.125 · sinc((t − 2T )/T )). QED.8 are sufficient to represent yp (t).24) The deterministic autocorrelation function q(t) is Hermitian (q∗ (−t) = q(t)). Also. Since the set {φp. By (3.25 · sinc((t − T )/T ) − . The function q(t) can also exhibit ISI.2029 1 .∞) is equivalent to a single matched filter whose output is sampled at t = kT to produce y(kT ). p 2 (3. p (3. the sampled matched filter output y(kT ) is a sufficient representation of the ISI-channel output yp (t). equivalently.1159] or.k ϕ∗ (−t).1159 . q(0) = 1. or 159 . then by the theorem of reversibility in Chapter 1.23) where q(t) = ϕp (t) ∗ ϕ∗ (−t) = p ∆ p(t) ∗ p∗ (−t) .2029 − . to the channel p(t) = 1 T · (sinc(t/T ) + . The plotted q(t) corresponds to qk = [−.7.20) this last set k∈(−∞. p . Since Γ−1 is invertible (inverse is Γ).

) For notational ∆ ∆ ∆ ∆ brevity. nk = n(kT ) where n(t) = np (t) ∗ ϕ∗ (−t). let yk = y(kT ). the magnitudes correspond to actual values.2). addition of the magnitudes of the symbols ignores the phase components. (3. and sampling. which may significantly change the ISI term. Using D-transform notation. The ISI represents a new distortion component not previously considered in the analysis of Chapters 1 and 2 for a suboptimum SBS detector.27) 160 . qk = q(kT ). then the noise sample nk of (one-dimensional) variance N0 at the matched-filter output combines 2 with ISI from the sample times mT (m = k) in corrupting p · xk . the ISI is characterized by both its magnitude and phase. The scaled input is the desired informationbearing signal.25) becomes Y (D) = X(D) · p · Q(D) + N (D) ∆ ∞ (3.26) where Y (D) = k=−∞ yk · Dk .5D)(1 − . The ISI and noise are unwanted signals that act to distort the information being transmitted. However. If the receiver uses symbol-by-symbol detection on the sampled output yk .25D) (the notation P (D) is defined in Appendix A. then the peak distortion is: Dp = |x|max · p · m=0 3 In the real case. ∆ |qm | . This SBS detector is the same detector as in Chapters 1 and 2. (3. except used under the (false) assumption that the ISI is just additional AWGN. Such a receiver can be decidedly suboptimum when the ISI is nonzero. which only has meaning for real-valued q(t): 3 Definition 3. There are two common measures of ISI distortion. (The values for qk can be confirmed by convolving p(t) with its time reverse. normalizing.Figure 3. noise and ISI.9: ISI in q(t) 1 P (D) = √T · (1 + .2 (Peak Distortion Criterion) If |xmax | is the maximum value for |xk |. for complex-valued terms. The first is Peak Distortion. Thus p yk = p · xk scaled input (desired) + nk + p · noise m=k xm qk−m . So.1. ISI (3.25) The output yk consists of the scaled input.

99.078·. (3. this worst case might not occur very often in practice. The peak distortion represents a worst-case loss in minimum distance between signal points in the signal constellation for xk . the mean-square distortion (with Ex = 5) is Dms = 5 p 2(. The fact .9.20292 + √ 2 .34) channels for which 2Dp ≥ p dmin .4 While peak distortion represents the worst-case ISI. or equivalently   d p min − Dp 2  .20292 + .31) Implicitly.99 illustrates that Dms ≤ Dp . and Gaussian.) The mean-square distortion criterion assumes (erroneously5 ) that Dms is the variance of an uncorrelated Gaussian noise that is added to nk . Without loss of generality.6376 ≈ 1.28) σ for symbol-by-symbol detection.1159+. and the resulting error probability. see Chapter 6. assume yk > yk .9 with xmax = 3. the distance interpretation in (3. Nevertheless.109 ≈ . (3.11592 +. These outputs are generated by two different sequences {xk } and {xk }.2029+. Without ISI the distance is yk − yk ≥ p · dmin . With this assumption. the probability of occurrence of the worst-case value (worst level occurring in all 14 ISI contributors) is 4−14 = 3. For instance.d.m · qk−m |2   m=k 2 = Ex · p · m=0 |qm |2 .078).i. well below typical channel Pe ’s in data transmission. data transmission engineers more often use a measure of ISI called Mean-Square Distortion (valid for 1 or 2 dimensions): Definition 3.33) where (3.588 = . with an input alphabet size M = 4 and a q(t) that spans 15 symbol periods.30) yk − yk ≥ p dmin − 2|x|max m=0 |qm| . 5 This 161 . so that this approximation of Gaussian ISI becomes accurate.31) assumes 2Dp ≤ p dmin. In the example of Figure 3. Consider two matched filter outputs yk and yk that the receiver attempts to distinguish by suboptimally using symbol-by-symbol detection.767 < 1.√ For q(t) in Figure 3. xk is approximately i. Pe ≤ Ne Q  (3. Rather than separately compute each possible combination’s reduction of minimum distance. In very well-designed data transmission systems. (3.33) is valid when the successive data symbols are independent and identically distributed with zero mean. Pe is approximated by Pe ≈ Ne · Q 4 On p dmin ¯ 2 σ2 + Dms .3 (Mean-Square Distortion) The Mean-Square Distortion is defined by:     ∆ Dms = E | (3. (The proof of this fact is left as an exercise to the reader.1.32) xp.7×10−9. the worst-case ISI occurs when |2Dp − p dmin | is maximum.2029+. assumption is only true when xk is Gaussian. while with ISI the distance can decrease to   (3.29) The summation term inside the brackets in (3.588. its probability of occurrence.11592) ≈ 5(1.1159) ≈ 3· 1. and consider the difference   yk − yk = p (xk − xk ) + m=k ˜ (xm − xm )qk−m  + n (3. Dp = 3· p (. there may be other ISI patterns of nearly just as bad interference that can also occur.29) represents the change in distance between yk and yk caused by ISI.

The ISI causes the spread among the path traces. Clearly increasing M reduces the eye opening.10 and 3. 162 .11 do not include noise).11. For binary transmission on this channel. The Lorentzian pulse response p(t) = 1/(1 + (3t/T )2) is used in both plots.One way to visualize ISI is through the “eye diagram”. when the trigger is synchronized to the symbol rate. there is a significant opening in the eye in the center of the plot in Figure 3. the openings are much smaller. leading to less noise immunity. The eye diagram is similar to what would be observed on an oscilloscope. The eye diagram is produced by overlaying several successive symbol intervals of the modulated and filtered continuous-time waveform (except Figures 3. more ISI results in a narrower eye opening.10. some examples of which are shown in Figures 3.10 and 3. With 4-level PAM transmission.

Figure 3.11: 4-Level eye diagram for a Lorentzian pulse response. 163 .10: Binary eye diagram for a Lorentzian pulse response.Figure 3.

Use of SNR as a performance measure builds upon the simplifications of considering mean-square distortion.35) is the mean-square error MSE=E|ek |2. independent of M at constant Ex . This SNR is often directly related to probability of error and is a function of both the receiver and the decision regions for the SBS detector.1 Receiver Signal-to-Noise Ratio Definition 3.2.12) is SNRR = ∆ ∆ Ex .12 is the ratio of channel output ¯ ¯ sample energy Ex p 2 to the mean-square distortion σ2 + Dms .2. that is both noise and ISI are jointly considered in a single measure.12: Use of possibly suboptimal receiver to approximate/create an equivalent AWGN. This text uses SNR consistently.2.25) have normalized mean-square ¯ value σ2 + Dms . such a receiving device is inserted at the sampler output.2 discusses. E|ek |2 (3.1 (Receiver SNR) The receiver SNR. Usually. and decision regions based on xk (see Figure 3. The purpose of the receiver is to attempt to convert the channel into an equivalent AWGN that is also shown below the dashed line. the receiver is unbiased (otherwise biased) with respect to the decision regions for xk .12 focuses upon the receiver and specifically the device shown generally as R. The concept of a receiver SNR facilitates evaluation of the performance of data transmission systems with various compensation methods (i. although not exactly so as Subsection 3. Subsection 3. The SNR for the matched filter output yk in Figure 3.e. 3. The denominator of (3. SNRM F B . equalizers) for ISI.3 finishes this section with a discussion of the highest possible SNR that a designer could expect for any filtered AWGN channel. An SNR. smaller deviation from the transmitted symbol means better performance.35) where ek = xk − zk is the receiver error. and a generally good measure 164 . This section shall not be specific as to the content of the box shown as R.2. Such an AWGN is not always exactly achieved.Figure 3. The two right-most terms in (3. When channels have ISI.1 can be used then to analyze the performance of the symbol-by-symbol detector that follows the receiver R. the so-called “matched-filter-bound” SNR. but nonetheless any deviation between the receiver output zk and the channel input symbol xk is viewed as additive white Gaussian noise. 3. as in Subsection 3. replacing probability of error as a measure of comparative performance.2 Basics of the Receiver-generated Equivalent AWGN Figure 3. When E [zk |xk ] = xk . SNRR for any receiver R with (pre-decision) output zk . but later sections will allow both linear and slightly nonlinear structures that may often be good choices because their performance can be close to SNRM F B .2. ¯ SNR is easier to compute than Pe .

When E [zk /xk ] = α · xk = xk .Figure 3.) The following theorem relates the SNR’s of the unbiased and biased decision rules for any receiver R: Theorem 3.4 .” Removal of the bias is easily achieved by scaling by 1/α as also in Figure 3. satisfy zk = α · (xk + uk ) (3. so that E [uk /xk ] = 0. the decision regions in the SBS detector are “biased.13 illustrates a receiver that somehow has tried to reduce the combination of ISI and noise. The SNR is easier to compute and this text assumes that the insertion of the appropriately scaled SNR (see Chapter 1 .13. scaling the signal correctly improves the probability of error on the average if the input constellation has zero mean. 165 (3. Even when this insertion into the Q-function is not sufficiently accurate.Sections 1. then removal of bias by scaling by 1/α improves the probability of error of such a detector as in Chapter 1. the maximum unconstrained SNR corresponding to that same receiver with any biased decision rule is SNRR = SNRR. However. The probability of error is difficult to compute exactly because the distribution of the ISI-plus-noise is not known or is difficult to compute.6)) into the argument of the Q-function approximates the probability of error for the suboptimum SBS detector. comparison of SNR’s for different receivers usually relates which receiver is better.13: Receiver SNR concept. zk . (Even when the noise is not Gaussian as is the case with the ISI component.1 (Unconstrained and Unbiased Receivers) Given an unbiased receiver R for a decision rule based on a signal constellation corresponding to xk .38) .2 Receiver Biases Figure 3. 3.1. of performance: higher SNR means lower probability of error. The uncorrelated distortion has no remnant of the current symbol being decided by the SBS detector.36) where α is some positive scale factor that may have been introduced by the receiver and uk is an uncorrelated distortion uk = rm · xk−m + fm · nk−m .37) m=0 m The coefficients for residual intersymbol interference rm and the coefficients of the filtered noise fk will depend on the receiver and generally determine the level of mean-square distortion.2.U + 1 . (3. the receiver may have found that by scaling (reducing) the xk component in zk by α that the SNR improves (small signal loss in exchange for larger uncorrelated distortion reduction). as is the assumption with the SBS detector. If the distortion is assumed to be Gaussian noise.2. Any time-invariant receiver’s output samples.

(3.U or αopt = 1/(1 + (SNRR. and the appearance of an improved SNR is an artifact of the SNR definition. and the SNR for the “MMSE” case will often be easier to compute.27 in Section 3.40) 2 |α|2σu + |1 − α|2Ex Allowing for complex α with phase θ and magnitude |α|.41) SNRR.13. Proof: From Figure 3. R∈R (3.U = ¯ Ex σu ¯2 . the biased SNR is equal to the unbiased SNR plus 1. A natural question is then “Why compute the biased SNR?” The answer is that the biased receiver corresponds directly to minimizing the mean-square distortion. a receiver designed for maximum SNR can be altered slightly through simple output scaling by 1/αopt to a related receiver that has no bias and has SNR thereby reduced to SNRR. which allows noise to be scaled down without taking into account the fact that actual signal power after scaling has also been reduced.39) The maximum SNR for the biased signal zk prior to the scaling occurs when α is chosen to maximize the unconstrained SNR Ex SNRR = .U ] = max[SNRR ] − 1 = SNRRopt − 1 . the SNR after scaling is easily SNRR.2.where SNRR.43) Proof. Figure 3. (3.1.U )−1).U 1 Clearly θ = 0 for a minimum and differentiating with respect to |α| yields −2 + 2|α|(1 + SNR ) = 0 R. a receiver could scale the channel output 2 by α = 10/11.44) 166 . and the corresponding performance-characterizing SNR can always be found by subtracting 1 from the biased SNR.2. which has MSE=E[|ek |2] = 11 11 1 100 1 121 + 121 (. the SNR maximization over alpha is equivalent to minimizing 1 1 − 2|α|cos(θ) + |α|2(1 + ) .U = SNRR − 1. but not improving Pe since the SBS detector works best when there is no bias) from a receiver that minimizes mean-square distortion (or error) to get an unbiased decision. The resultant new error signal is ek = xk (1 − 10 ) − 10 nk . Theorem 3. so R∈R max[SNRR.2 (Unbiased MMSE Receiver Theorem) Let R be any allowed class of receivers R producing outputs zk . suppose an ISI-free AWGN 2 channel has an SNR=10 with Ex = 1 and σu = N0 = . Then.5 illustrates the usual situation of removing a bias (and consequently reducing SNR. The scaling has done nothing to improve the system. (3.1) = 11 . QED.2 below. Conversely. a receiver R and a corresponding SBS detector that have zero bias will not correspond to a maximum SNR – the SNR can be improved by scaling (reducing) the receiver output by αopt. Then the receiver that achieves the maximum SNR with an unbiased decision rule is also Ropt.U = SNRR − 1. (3. The bias from a receiver that maximizes SNR by equivalently minimizing mean-square error can then be removed by simple scaling and the resultant more accurate SNR is thus found for the unbiased receiver by subtracting 1 from the more easily computed biased receiver.U + 1 . To illustrate the relationship of unbiased and biased receiver SNRs.U is the SNR using the unbiased decision rule. and max SNRR. This concept will be very useful in evaluating equalizer performance in later sections of this chapter.U = SNR(Ropt) − 1. for any R ∈ R.2. Clearly.1.42) Thus. From Theorem 3. R∈R (3. and SNR=11. the relation between the signal-to-noise ratios of unbiased and unconstrained decision rules is SNRR. Removing the bias corresponds to using the actual signal power. and let Ropt be the receiver that achieves the maximum signal-to-noise ratio SNR(Ropt ) over all R ∈ R with an unconstrained decision rule. and is formalized in Theorem 3. Substitution of this value into the expression for SNRR finds SNRR = SNRR.

45) MFB denotes the square of the argument to the Q-function that arises in the equivalent “one-shot” analysis of the channel. The normalized average energy of this sample is p 2Ex . QED. In effect the min MFB forces no ISI by disallowing preceding or successive transmitted symbols. The quantity SNRR. SNRM F B = N0 . xp (t) = k xk · p(t − kT ) . the MFB (in dB) is just the “channel-output” SNR.1 (Binary PAM) For binary PAM. The following example illustrates computation of the MFB for several cases of practical interest: EXAMPLE 3. again. (3. σ2 = N0 2 . p 2 . implying performance is also a function of the symbol rate 1/T .U is the SNR that corresponds to best Pe . When Ex equals (d2 /4)/κ. The possibility of correlating the input sequence {xk } to take advantage of the channel correlation will be considered in Chapters 4 and 5.QED. An optimum detector is used for this “one-shot” case.U + 1 is artificially high because of the bias inherent in the general SNR definition.48) The binary-PAM Pe is then bounded by Pe ≥ Q( SNRM F B ) . (3.46) √ where xk = ± Ex . SNRR. Lemma 3. The minimum distance at the matched-filter output is p · dmin = d2 √ p · d = 2 · p · Ex . 2 2 The probability of error. σ2 (3. while the ¯ Ex p 2 corresponding noise sample energy is N0 .2.2. if xk is an i. so Ex = min and κ = 1. √ ¯ satisfies Pe ≥ Ne · Q( MFB).49) 167 . specifies an upper SNR limit on the performance of data transmission systems with ISI. 3. the maximum output sample ¯ of the matched filter is p · x0.d. SNRM F B . No other (for the same input constellation) receiver for continuous transmission could have better performance. also called the “one-shot” bound. For any SBS detector. since the sequence must incur some level of ISI. only the SNR measures are different.3 The Matched-Filter Bound The Matched-Filter Bound (MFB). then MFB equals SNRM F B · κ. 4 MFB = SNRM F B . This SNR is SNRM F B = ¯ Ex p N0 2 2 (3. Proof: Given a channel with pulse response p(t) and isolated input x0. Then. If the transmitter symbols xk are equal to ±1 (Ex = 1).2. then MFB = where. This theorem implies that the optimum unbiased receiver and the optimum biased receiver settings are identical except for any scaling to remove bias. sequence. The performance is tacitly a function of the transmitter basis functions.i. (3. measured after the matched filter and prior to the symbol-by-symbol detector.1 (Matched-Filter Bound SNR) The SNRM F B is the SNR that characterizes the best achievable performance for a given pulse response p(t) and signal constellation (on an AWGN channel) if the channel is used to transmit only one message.47) Thus for a binary PAM channel.

M −1 (3. Thus.57) so κ = 3/(M − 1). for both (3. M −1 for M ≥ 4.50) (3.. . .2. 2 p 2 . ± ( 2 2 √ {xk } = ± d . xk = ± d . ± ( 2 2 respectively) and M −1)d .52).. MFB = M2 3 SNRM F B . If the transmitter symbols xk are equal to ±1. ± 2 2 d = 2 so κ = 3/(M 2 − 1). ±( M − 1). then MFB = The best QPSK Pe is then approximated by ¯ Pe ≈ Q( SNRM F B ) . . −1 3Ex . xk = ± d ±  d . .51) for M ≥ 2. Thus 3 (3. Thus.. which should be expected since the minimum distance is the same at the transmitter.2 (M-ary PAM) For M-ary PAM. σ2 (3.3 (QPSK) For QPSK.. M2 − 1 (3. so κ = 1.4 (M-ary QAM Square) For M-ary QAM. The M’ary-PAM Pe is then bounded by Pe ≥ 2(1 − 1/M ) · Q( 3 · SNRM F B ) .. M2 − 1 (M −1)d 2 and (3.. {xk } = ± d . ±3.58) SNRM F B .. If the transmitter symbols xk are ±1 ± . ± 3d . then MFB = p 2 . . ±(M − 1). then p 2 MFB = 2 . ¯ 3Ex ..53) ¯ Ex . MFB (in dB) equals the channel output SNR.59) σ The best M’ary QAM Pe is then approximated by MFB = √ ¯ Pe ≈ 2(1 − 1/ M) · Q( 3 · SNRM F B ) .2. σ2 (3. ± 3d .52) Equation (3. and d = 2 2 2 MFB = SNRM F B . ±3. ± 3d .2. 2 (recall that d = 2 and denote real and imaginary parts. If√ real and imaginary components of the transmitter symbols xk equal the ±1. and thus also at the channel output.60) 168 . Thus (3..EXAMPLE 3.55) EXAMPLE 3.54) EXAMPLE 3. (3. . M −1 (3.56) √ M −1)d .48)..48) and (3.. for a QPSK (or 4SQ QAM) channel... . (3...52) is the same result as (3.

SNRU ≤ SNRM F B . unless one allows co-design of the input symbols xk in a channel-dependent way (see Chapters 4 and 5). −1 (3. then the receiver is approaching the bound on performance. The loss with respect to matched filter bound will be determined for any receiver by SNRu /SNRM F B ≤ 1.62) For the suboptimum receivers to come in later sections. It is not always possible to design a receiver that attains SNRM F B . b 31 · 4¯ − 32 (3. even with infinite complexity. in effect determining a loss in signal power because successive transmissions interfere with one another – it may well be that the loss in signal power is an acceptable exchange for a higher rate of transmission. As SNRU → SNRM F B . MFB = For the QAM Cross constellations.61) MFB = = 96 SNRM F B . 169 . ¯ 2E x p 2 31 2 48 M − 3 2 σ b 4¯ 3 SNRM F B .In general for square QAM constellations.

One feature of sinc(t/T ) is that it has minimum bandwidth for no ISI. T ) to be added (i.69) are called “Nyquist pulses. where Qeq (ω).65) (3.71) The function q(kT ) = sinc(k) = δk satisfies the ISI-free condition.63) (3. QED. k=−∞ (3.1 (Nyquist’s Criterion) A channel specified by pulse response p(t) (and ∗ (−t) resulting in q(t) = p(t)∗p 2 ) is ISI-free if and only if p Q(e−ωT ) = 1 T ∞ Q(ω + n=−∞ 2πn )=1 .3.” One function that satisfies Nyquist’s Criterion is t q(t) = sinc . (3. (3. “folded in”).68) a well-known relation between the discrete-time and continuous-time representations of any waveform in digital signal processing. It is now straightforward to specify Nyquist’s Criterion: Theorem 3.69) Proof: By definition the channel is ISI-free if and only if qk = 0 for all k = 0 (recall q0 = 1 by definition).67) The function Qeq (ω) is periodic in ω with period 2π . Writing the Fourier Transform of the sequence qk as ∞ Q(e−ωT ) = k=−∞ qk e−ωkT leads to 1 ∆ · Qeq (ω) = Q(e−ωT ) = T ∞ qk e−ωkT .64) (3. T (3. the equivalent frequency response becomes ∞ Qeq (ω) = n=−∞ ∆ Q(ω + 2πn ) .e. No other function has this same minimum bandwidth and also satisfies 170 . q(kT ) = = = = 1 2π 1 2π 1 2π 1 2π ∞ Q(ω)eωkT dω −∞ ∞ (2n+1)π T (2n−1)π T π T (3. Functions that satisfy (3. This section first reviews some fundamental relationships between q(t) and its samples qk = q(kT ) in the frequency domain.3 Nyquist’s Criterion Nyquist’s Criterion specifies the conditions on q(t) = ϕp (t) ∗ ϕ∗ (−t) for an ISI-free channel on which a p symbol-by-symbol detector is optimal. This function is also known as the folded or aliased T spectrum of Q(ω) because the sampling process causes the frequency response outside of the fundamental π π interval (− T .70) T which corresponds to normalized pulse response 1 ϕp (t) = √ sinc T t T .3. T (3. The proof follows directly by substitution of qk = δk into (3.66) Q(ω)eωkT dω 2πn (ω+ 2πn )kT T dω )e T n=−∞ ∞ Q(ω + n=−∞ π T π −T π −T Qeq (ω)eωkT dω .68).

1/2T represents a minimum bandwidth necessary to satisfy the Nyquist Criterion.Figure 3. resulting in infinite peak distortion. continuously differentiable).6 π T or f = 1 2T is called the 3.1 Vestigial Symmetry In addition to the sinc(t/T ) Nyquist pulse. In fact for q(t) = sinc(t/T ).e. any increase in bandwidth should be as small as possible. the response must occupy a larger than minimum bandwidth that is between 1/2T and 1/T . 6 This 171 .14 for −20T ≤ t ≤ 20T . Thus. In terms of positive-frequencies. the ISI term k=0 q(τ + kT ) with a sampling timing error of τ = 0 is not absolutely summable. A q(t) with higher bandwidth can exhibit significantly faster decay as |t| increases. The envelope of the time domain response decays more rapidly if the frequency response is smooth (i.7 introduces residual ISI with amplitude that only decays linearly with time. (Proof is left as an exercise to the reader. Definition 3.3. thus reducing sensitivity to timing phase errors. or percent roll-off. To meet this smoothness condition and also satisfy Nyquist’s Criterion. the Nyquist Criterion. and thus has a special name in data transmission: Definition 3.) The sinc(t/T ) function is plotted in Figure 3.14: The sinc(t/T) function.3. Of course. 1 The frequency 2T (1/2 the symbol rate) is often construed as the maximum frequency of a sampled signal that can be represented by samples at the sampling rate. data-transmission engineers use responses with up to twice the minimum bandwidth.1 (Nyquist Frequency) The frequency ω = Nyquist frequency. The percent excess bandwidth7. Q(ω) = 0 for |ω| > 2π . any sampling-phase error in the sampling process of Figure 3. where the latter is twice the highest frequency of a signal to be sampled and is not the same as the Nyquist Frequency here. and similarly Q(·) is a measure of ISI and not the integral of a unit-variance Gaussian function – uses should be clear to reasonable readers who’ll understand that sometimes symbols are re-used in obviously differenct contexts. 7 The quantity alpha used here is not a bias factor.3.2 (Percent Excess Bandwidth) The percent excess bandwidth α is determined from a strictly bandlimited Q(ω) by finding the highest frequency in Q(ω) for text distinguishes the Nyquist Frequency from the Nyquist Rate in sampling theory. These wider bandwidth responses T will provide more immunity to timing errors in sampling as follows: The sinc(t/T ) function decays in amplitude only linearly with time. For all these pulses. while still meeting other system requirements. is a measure of the extra bandwidth.

recalling that q(t) = ϕp (t) ∗ ϕ∗ (−t) is Hermitian and has the properties of an autocorrelation p π function. +1 contribute to the folded spectrum and the Nyquist Criterion becomes 1 = = Q(e−ωT ) 1 2π Q ω+ T T (3. (3.15 and 3. (3. In this case in equation (3. An example of a vestigially symmetric response with 100% excess bandwidth is q(t) = sinc2 (t/T ).73) + Q (ω) + Q ω − 2π T − π π ≤ω≤ T T .2 Raised Cosine Pulses The most widely used set of functions that satisfy the Nyquist Criterion are the raised-cosine pulse shapes: 172 .77) .76) π For complex signals.74) Further. That is Q(ω) = nonzero 0 π |ω| ≤ (1 + α) T π |ω| > (1 + α) T . (3.Figure 3. (3. then Q(ω) is real and Q(ω) ≥ 0. (3.78) in the complex case) is said to be vestigially symmetric with respect to the Nyquist Frequency. the pulse q(t) is said to have “15% excess bandwidth. 0.3. data transmission systems have 0 ≤ α ≤ 1.” Usually. which is shown in Figures 3.76) (and (3. For the region 0 ≤ ω ≤ T (for real signals).69). .78) Any Q(ω) satisfying (3. if α = .16.74) reduces to 1 = = Q(e−ωT ) 1 2π Q (ω) + Q ω − T T (3. the negative frequency region (− T ≤ ω ≤ 0) should also have 1 = = Q(e−ωT ) 1 2π Q (ω) + Q ω + T T (3. 3.15: The sinc2 (t/T ) Function – time-domain which there is nonzero energy transfer.72) Thus.15 (a typical value).75) . only the terms n = −1.

Figure 3.16: The sinc2 (t/T ) function – frequency-domain Figure 3.17: Raised cosine pulse shapes – time-domain 173 .

(3. .3.5.80) The raised cosine pulse shapes are shown in Figure 3. the transmit filter (and the receiver matched filters) are squareroots of a Nyquist pulse shape. When α = 0.18 (frequencydomain) for α = 0.82) 1 − sin 2α (|ω| − T ) (1 − α) ≤ |ω| ≤ T (1 + α) 2 T   π 0 T (1 + α) ≤ |ω| 174 . The raised-cosine pulse shapes are then often used in square-root form. Such square-root transmit filtering is often used even when ϕp (t) = ϕ(t) and the pulse response is the convolution of h(t) with the square-root filter.81) so that the matched filter and transmit/channel filters are “square-roots” of the Nyquist pulse.3 (Raised-Cosine Pulse Shapes) The raised cosine family of pulse shapes (indexed by 0 ≤ α ≤ 1) is given by q(t) = sinc and have Fourier Transforms  T  T T Q(ω) = 1 − sin 2α (|ω| −  2 0 t T · cos 1− απt T 2αt 2 T .Figure 3. 3 t 3.3 Square-Root Splitting of the Nyquist Pulse The optimum ISI-free transmission system that this section has described has transmit-and-channel filtering ϕp (t) and receiver matched filter ϕ∗ (−t) such that q(t) = ϕp (t) ∗ ϕ∗ (−t) or equivalently p p Φp (ω) = Q1/2(ω) (3. the function decays as t1 for t → ∞. while for α = 0. (3. which decays asymptotically as 1 for t → ∞.18: Raised Cosine Pulse Shapes – frequency-domain Definition 3. and 1. When h(t) = δ(t) or equivalently ϕp (t) = ϕ(t) .79) π T) π |ω| ≤ T (1 − α) π T (1 − α) ≤ |ω| ≤ π T (1 + α) ≤ |ω| π T (1 + α) .17 (time-domain) and Figure 3. (3.3. which is  √ π T |ω| ≤ T (1 − α)   1/2 T T π π π Q(ω) = . the raised cosine reduces to a sinc function.

5(1 − cos(2θ)). to obtain 4α cos [1 + α] πt + T ϕp (t) = √ · π T 1 − 4αt T T ·sin ([1−α] πt ) T 4αt 2 .This expression can be inverted to the time-domain via use of the identity sin2 (θ) = . as in the exercises.83) 175 . (3.

90) (3. The ZFE is so named because ISI is “forced to zero” at all sampling instants kT except k = 0. based on decision regions for the constellation defined by xk .7. The ZFE sets R equal to a linear time-invariant filter with discrete impulse response wk that acts on yk to produce zk . it is also Gaussian. In so doing. The receiver uses symbol-by-symbol detection. 2 2 N0 ¯ Rnn(D) = Q(D) . The equalizer tries to restore this Nyquist Pulse character to the channel. leaving the ZFE filter characteristic: Definition 3. on the output of the ZFE. qk = δk by Nyquist’s Criterion.85) This discrete-time filter processes the discrete-time sequence corresponding to the matched-filter output. yk = p · xk ∗ qk + nk . Ideally.k ¯ = = −∞ ∞ −∞ ∞ −∞ ∞ E nl n∗ l−k /N ∞ ∞ (3.86) (3. As this noise is produced by a linear filter acting on the discrete Gaussian noise process nk . at the output of the ZFE is important in determining performance. (3. The ZFE is often a first non-trivial attempt at a receiver R in Figure 3.88) (3.4 Linear Zero-Forcing Equalization This section examines the Zero-Forcing Equalizer (ZFE).2.84) In the ZFE case. which is an estimate of xk .3.1 and Figure 3. which is the easiest type of equalizer to analyze and understand.4. The ISI model in Figure 3.87) (3. (3.4. for symbol-by-symbol detection to be optimal. the ZFE ignores the noise and shapes the signal yk so that it is free of ISI. 2 176 (3. which is not zero in practice.91) (3.94) . even if ignored in the ZFE design of Equation (3. but has inferior performance to some other equalizers to be introduced in later sections. has Transform Z(D) = W (D) · Y (D) = W (D) · Q(D) · p · X(D) .1 Performance Analysis of the ZFE The variance of the noise. zk .93) E np (t)n∗ (s) ϕ∗ (t − lT )ϕp (s − (l − k)T )dtds p p N0 δ(t − s)ϕ∗ (t − lT )ϕp (s − (l − k)T )dtds p 2 = −∞ = = = or more simply N0 ϕ∗ (t − lT )ϕp (t − (l − k)T )dt 2 −∞ p N0 ∞ ∗ ϕ (u)ϕp (u + kT )du (letting u = t − lT ) 2 −∞ p N0 ∗ N0 q−k = qk .7 is used to describe the zero-forcing version of the discrete-time receiver.2) Y (D) = p · X(D) · Q(D) .87). so N = 2 for the complex QAM case) for the noise nk as rnn. nk is initially viewed as being zero. From the ISI-channel model of Section 3. 3.89) (3.1 (Zero-Forcing Equalizer) The ZFE transfer characteristic is W (D) = 1 Q(D) · p .12 of Section 3. The D-transform of yk is (See Appendix A. The ZFE output. and will be free of ISI if Z(D) = X(D). The designer can compute the discrete autocorrelation function (the bar denotes normalized to one dimension.92) (3.

with equality if and only if Q(D) = 1.102) (3. (3. 2 then σZF E is computed as 2 σZF E = π T π −T π T π −T (3. Finally. Since zk = xk + nZF E.97) and the fact that E [nZF E. The sign of the loss is often ignored since it is always a negative (or 0) number and only its magnitude is of concern.k is then RZF E (e−ωT ).92). has autocorrelation function ∗ rZF E. the ZFE equalization loss. From (3.where the analysis uses the substitution u = t − lT in going from (3.105) always thus provides a non-negative number.100) σZF E γZF E ¯ Computation of dmin over the signal set corresponding to xk leads to a relation between Ex .k .103) (3.104) SNRM F B γZF E ≥ SNRZF E ≥ 1 .4.0 . so that the ZFE is unbiased for a detection rule based on xk . ¯ ¯ (3. σ ≤ 2 2 σZF E p 2 T = σ 2π 2 dω = σ2 · γZF E Q(e−ωT ) (3.2 (ZFE Equalization Loss) The ZFE Equalization Loss. E [zk /xk] = xk .k = rnn. The (per-dimensional) variance of the noise samples at the output of the ZFE is the (per-dimensional) mean-square error between the desired xk and the ZFE output zk . (3. γZF E in decibels is γZF E = 10 · log10 ∆ SNRM F B SNRZF E = 10 log10 p 2 2 · σZF E = 10 log10 p · wZF E.98) where γZF E π T π −T 1 Q(e−ωT ) dω = w0 · p . 1/γZF E .k is then ¯ N0 Q(D) ¯ RZF E (D) = = 2 p 2 · Q2(D) p N0 2 2 · Q(D) . 2 nZF E. dmin. The SNR at the output of the ZFE is ¯ Ex 1 SNRZF E = 2 = SNRM F B · . and the NNUB on error probability for a symbol-by-symbol detector at the ZFE output is (please.e ≈ Ne · Q dmin 2σZF E π T π −T .97) T 2π T ¯ RZF E (e−ωT )dω = 2π T = 2π N0 2 p 2 · Q(e−ωT ) dω = N0 2 p γ 2 ZF E = N0 2 p · w0 . or if the ZFE is explicity used. has (one-dimensional) sample variance N0 at the normalized matched filter output. The complex baseband-equivalent noise.96) ¯ The power spectral density of the noise samples nZF E.k ] = 0. do not confuse the Q function with the channel autocorrelation Q(D)) PZF E.99) The center tap of a linear equalizer is w0 .k ∗ wk ∗ w−k .105) σ2 and is a measure of the loss (in dB) with respect to the MFB for the ZFE.91) to (3. The noise at the output of the ZFE. defines the SNR reduction from the MFB: Definition 3.k .101) Since symbol-by-symbol detection can never have performance that exceeds the MFB. (3. This tap is easily shown to always have the largest magnitude for a ZFE and thus spotted readily when plotting the impulse response of any linear equalizer. and M . (3. and assumes baseband equivalent signals for the complex case. (3.95) The D-Transform of rZF E.0. 177 . (3. Equation (3. then wZF E.

Figure 3.4. and the consequent poor performance. leading to unacceptable performance degradation.81 + 1.110) . Q(e−ωT ) is zero at ω = T .81 = 12 + .81 = 1.92 . Since W (e−ωT ) = 1/( p · Q(e−ωT )). the author has found that the table in Table 3. a σZF E that is unacceptably large.8 cos(ωT )] dω (3. The following example clearly illustrates the noise-enhancement effect: EXAMPLE 3.1 is useful in recalling basic digital signal processing equivalences related to scale factors between continuous time convolution and discrete-time convolution.20. then any noise energy near the Nyquist Frequency is enhanced (increased in power or energy).1 (1 + . Thus.108) π −T T 2π (3. with Φp (ω) as the Fourier transform of {ϕp (t)}. σZF E can be finite. Even when there is no channel notch at any frequency.109) 1. Φp (ω) = T 1.107) (3. but large. (3.2 Noise Enhancement The design of W (D) for the ZFE ignored the effects of noise.9D−1 .4. (3.19 hypothetically illustrates an example of a lowpass channel with a notch at the Nyquist π Frequency.19: Noise Enhancement 3. 2π T The magnitude of the Fourier transform of the pulse response appears in Figure 3. In actual computation of examples.Figure 3.9eωT P (ω) = . that is.106) π 0 |ω| > T Then p 2 is computed as 1 2π 1 2π π T π −T π T p 2 = = = |P (ω)|2dω T [1.9eωT 0 178 |ω| ≤ |ω| > π T π T . in this case so much 2 2 that σZF E → ∞.ZFE) Suppose the pulse response of a (strictly bandlimited) channel used with binary PAM is √ π |ω| ≤ T T 1 + . This oversight can lead to noise enhance2 ment.81 1 + .

Table 3. 179 .1: Conversion factors of T .

81 1. 180 . 8 2 σZF E = = = T 2π T 2π π T π −T π T π −T N0 2 Q(e−ωT ) · p 2 dω (3. Then. The time-domain sample response of the equalizer is shown in Figure 3.115) .9D . 2 Computation of σZF E allows performance evaluation. 1.81D + .20: Magnitude of Fourier transform of (sampled) pulse response.117) (3.9D−1 + 1. (using the bandlimited property of P (ω) in this example) Q(e−ωT ) = = = Q(D) Then = 1 |Φp (ω)|2 T 1 |1 + .81 + 1.112) (3.9eωT |2 1.111) (3.Figure 3.9D2 The magnitude of the Fourier transform of the ZFE response appears in Figure 3.81) · N0 2 dω 1.116) (3.81 + 1.21.8 cos(ωT ) 1.9 + 1.81 + .81 .114) √ 1.118) (1.8 cos(ωT ) π −π N0 1 2 2π 1 du 1.8 cos u 8 From a table of integrals = √ 2 Tan−1 a2 −b2 √ 1 du a+b cos u a2 −b2 tan( u ) 2 a+b .81D W (D) = .81 + 1. (3.81/1.113) (3.22.81 (3.

Figure 3.22: ZFE time-domain response.Figure 3. 181 .21: ZFE magnitude for example.

81 N0 (5.8 π (3.26) . the loss is almost 10 dB from best performance.0625 + . (−1 + ) .−k = or Q(D) = 1 1.4. thus. −  2 . then we get MFB = leaving γZF E = 10 log10(1. there are good solutions. 2 ¯ Ex 5.120) (3.25 + 1 + .81 · 5.5625 5 5   − D−2 + (−1 + )D−1 + 1. 2 1+  4 .82 2 2 2π 1.2 (QAM: −. Chapter 9 demonstrates a means by which to ensure that there is no loss with respect to the matched filter bound for this channel. 4 8 8 4 (3. The pulse response norm (squared) is p Then qk is given by qk = q(kT ) = T ϕp.82 tan u 2 1.119) 0 N0 π 4 · ) √ 2 − 1.82 2 1.812 − 1.26 N0 2 (3.23.81 = 2 .25) − .5625 . 1.122) .812 − 1. with no information rate loss. σ2 σ (3.√ = = = = The ZFE-output SNR is SNRZF E = N0 2 ( ) 2 2π ( √ 2 1. With alteration of the signal constellation.5625 .k ∗ ϕ∗ p.129) 2 = T (.127) This channel has the transfer function of Figure 3. (3. − (1 + ) . (3.5625 − (1 + )D + D2 4 8 8 4 182 (3.8dB (3. but the simple concept of a ZFE is not a good solution. which is far below the error rate achievable even when the MFB is attained. (3.81 + 1. Another example illustrates the generalization of the above procedure when the pulse response is complex (corresponding to a QAM channel): EXAMPLE 3.5625 −  5 5  .121) (3. T (3. Chapter 10 describes a means that ensures an error rate of 10−5 on this channel.126) p 2 1.5D−1 + (1 + . This is very poor performance on this channel.123) This SNRZF E is also the argument of the Q-function for a binary detector at the ZFE output. Thus.81 N0 1 ( )√ 2 − 1. No matter what b is used.25) = 1.124) the discrete-time channel samples are 1 pk = √ T − 1 .128) .82 Tan −1 1. Assuming the two transmitted levels are xk = ±1.26) ≈ 9. noise enhancement can be a serious problem.5D Channel) Given a baseband equivalent channel 1 p(t) = √ T 1 − sinc 2 t+T T + 1+  sinc 4 t T  − sinc 2 t−T T .125) for any value of the noise variance.130) 1 1.

Then. The real and imaginary parts of the equalizer response in the time domain are shown in Figure 3.5D−1 1.5D−1 (1 − .9412 = 1.5D) 1 − .136) . where A and B are coefficients in the partial-fraction expansion (the residuez and residue commands in Matlab can be very useful here): A = 4(−4) = −1.5625 D2 (−4) {(D + 2) (D − 2) (D − .25 Then 2 σZF E T = 2π π T π −T N0 2 p 2Q(e−ωT ) N0 w0 2 p . dω .5625 1 + .131) (3.135) where w0 is the zero (center) coefficient of the ZFE.5625 .5686 − .5) (D + . (3.5D) 1 Q(D) p .5) 183 (3.625(1 + )D + .24.23: Fourier transform magnitude for pulse response for complex channel example. Q(D) factors into Q(D) = and WZF E (D) = or WZF E (D) = (−.Figure 3. This coefficient can be extracted from WZF E (D) = = √ 1.601 (2 + 2)(2 + .134) or 2 σZF E = (3. (3.5625 − .133) + 1.625(−1 + √ )D−1 1.8293 − 2.25D−2 + .5625 + + terms not contributing to w0 D − 2 D + 2 .5)(1. (3.25D2) 1 (1 − .132) The Fourier transform of the ZFE is in Figure 3.5)} √ A B 1.

Figure 3.Figure 3. 184 .24: Fourier transform magnitude of the ZFE for the complex baseband channel example.25: Time-domain equalizer coefficients for complex channel.

9412 + 1.57) = 1.140) The ZFE loss can be shown to be γZF E = 10 · log10(w0 · p ) = 3.5686 − .5(. then P 185 .5) √ WZF E (D) = 1.B Finally = −4(−4) = .96 (3.9 dB .8293 1. then Pe ≈ 2. (−2 − . the designer calculates ¯ Pe ≈ Q SNRM F B − 3.139) (3.4dB. considerably better performance is also possible on this complex channel.5)(−2 − 2)(−1.5685 = 1. (3. (3.9412) − . but not with the ZFE. If ¯e ≈ 2.5625(1.5686)] 1.9412 + 1.5625 [(−. Nevertheless. (3. SNRM F B = 23.0 × 10−6.5625 A(−.4dB for 16QAM. then Pe ≈ 1. If SNRM F B = 17.5) + + terms 1 − .5)(−1.0 × 10−5.141) which is better than the last channel because the frequency spectrum is not as near zero in this complex example as it was earlier on the PAM example.2 × 10−2 . To compute Pe for 4QAM.137) .142) ¯ ¯ If SNRM F B = 10dB.030 .5D) (3.5D (1 − .5) B(−.138) and √ w0 = = √ 1.9 dB .

as the next subsection computes. it is slightly more complicated to describe and analyze than is the ZFE.1 (Mean Square Error (for the LE)) The LE error signal is given by ek = xk − wk ∗ yk = xk − zk .1 Optimization of the Linear Equalizer E(D) = X(D) − W (D)Y (D) (3.145) Using D-transforms. By the orthogonality principle in Appendix A. This is permissable for stationary (and ergodic) discrete-time sequences.5. the filter output is as close as possible. the error sample ek must be uncorrelated with any equalizer input signal ym . The transfer function for the equalizer W (e−ωT ) is also real and positive for all finite signal-to-noise ratios. N = 2 for Quadrature Modulation)9 ¯ Rxy (D) ¯ Ryy (D) = = ¯ E X(D)Y ∗ (D−∗ ) /N = p Q(D)Ex N0 ¯ E Y (D)Y ∗ (D−∗ ) /N = p 2Q2 (D)Ex + Q(D) = Q(D) 2 ¯ Rxy (D) 1 W (D) = ¯ .143) The Minimum Mean Square Error (MMSE) for the linear equalizer is defined by 2 σM M SE−LE = min E |xk − zk |2 wk ∆ . 186 .3.148). The MMSE-LE uses a linear time-invariant filter wk for R.148) The MMSE-LE differs from the ZFE only in the additive positive term in the denominator of (3.5. That is.146). since the terms of X(D)X ∗ (D −∗ ) are of k ∆ the form xk x∗ .146) Then the MMSE-LE becomes (3.26 repeats Figure 3. using (3. E E(D)Y ∗ (D−∗ ) = 0 . or better than. at any time k.19 with addition of the MMSEπ ¯ LE transfer characteristic. The MMSE-LE is a linear filter wk that acts on yk to form an output sequence zk that is the best MMSE estimate of xk . to the data symbol xk . so that we are implying the additional operation limK→∞ [1/(2K + 1)] k−j −K≤k≤K on the sum in such terms. where (N = 1 for PAM. the filter wk minimizes the Mean Square Error (MSE): Definition 3.144) The MSE criteria for filter design does not ignore noise enhancement because the optimization of this filter compromises between eliminating ISI and increasing noise power. Figure 3. (3. 9 The expression Rxx (D) = E X(D)X ∗(D −∗ ) is used in a symbolic sense. while the ZFE becomes infinite at this same frequency.5 Minimum Mean-Square Error Linear Equalization The Minimum Mean-Square Error Linear Equalizer (MMSE-LE) balances a reduction in ISI with noise enhancement. (3. Evaluating (3. and thus makes the MMSE-LE well defined even when the channel (or pulse response) is zero for some frequencies or frequency bands. Also W (D) = W ∗ (D−∗ ). This small positive term prevents the denominator from ever becoming zero. = p (Q(D) + 1/SNRM F B ) Ryy (D) ¯ p 2Q(D)Ex + N0 2 . This MMSE-LE leads to better performance. The MMSE-LE always performs as well as. Succinctly. in the Minimum MSE sense.145).147) (3. (3. yields ¯ ¯ 0 = Rxy (D) − W (D)Ryy (D) . 3. Instead. but the choice of filter impulse response wk is different than the ZFE. The MMSE-LE transfer function has magnitude Ex /σ2 at ω = T . the ZFE and is of the same complexity of implementation. Nevertheless.

Figure 3.26: Example of MMSE-LE versus ZFE.

3.5.2

Performance of the MMSE-LE
¯ Ree(D) E E(D)E ∗ (D−∗ ) /N ¯ ¯ ¯ ¯ Ex − W ∗ (D−∗ )Rxy (D) − W (D)R∗ (D−∗ ) + W (D)Ryy (D)W ∗ (D−∗ ) xy ∗ −∗ ¯ ¯ Ex − W (D)Ryy (D)W (D ) ¯ Q(D) p · Q(D)Ex + ¯ Ex − 2 p 2 (Q(D) + 1/SNRM F B ) ¯ Ex Q(D) ¯ Ex − (Q(D) + 1/SNRM F B )
2 N0 2 N0 2

The MMSE is the time-0 coefficient of the error autocorrelation sequence = = = = = = (3.149) (3.150) (3.151) (3.152) (3.153) (3.154)

p

2 (Q(D)

+ 1/SNRM F B )

The third equality follows from ¯ W (D)Ryy (D)W ∗ (D−∗ ) = = ¯ and that W (D)Ryy (D)W ∗ (D−∗ )
2 σM M SE−LE ∗

¯ ¯ ¯ W (D)Ryy (D)Ryy (D)−1 R∗ (D−∗ ) xy ¯∗ (D−∗ ) , W (D)Rxy

¯ = W (D)Ryy (D)W ∗ (D−∗ ). The MMSE then becomes
π T π −T

T = 2π

π T π −T

T ¯ Ree(e−ωT )dω = 2π

p

N0 2 dω 2 (Q(e−ωT ) + 1/SNR MF B ) N0 2 /

.

(3.155)

By recognizing that W (e−ωT ) is multiplied by the constant
2 σM M SE−LE = w0 N0 2

p in (3.155), then (3.156)

p

.

From comparison of (3.155) and (3.98), that
2 2 σM M SE−LE ≤ σZF E ,

(3.157)

with equality if and only if SNRM F B → ∞. Furthermore, since the ZFE is unbiased, and SNRM M SE−LE = SNRM M SE−LE,U + 1 and SNRM M SE−LE,U are the maximum-SNR corresponding to unconstrained and unbiased linear equalizers, respectively, SNRZF E ≤ SNRM M SE−LE,U = ¯ Ex
2 σM M SE−LE

− 1 ≤ SNRM F B .

(3.158)

187

One confirms that the MMSE-LE is a biased receiver by writing the following expression for the equalizer output Z(D) = = = W (D)Y (D) 1 (Q(D) p X(D) + N (D)) p (Q(D) + 1/SNRM F B ) 1/SNRM F B N (D) X(D) − X(D) + , Q(D) + 1/SNRM F B p (Q(D) + 1/SNRM F B ) (3.159) (3.160) (3.161)

for which the xk -dependent residual ISI term contains a component signal-basis term = −1/SNRM F B w0 · p · xk p σ2 = −1/SNRM F B · M M SE−LE N0
2 2 σM M SE−LE · xk ¯ Ex 1 = − · xk . SNRM M SE−LE

(3.162)
2

xk

(3.163) (3.164) (3.165) (3.166)

= −

1 xk − ek where ek is the error for unbiased detection and R. The optimum SNRMMSE−LE unbiased receiver with decision regions scaled by 1 − SNR 1 (see Section 3.2.1) has the signal MMSE−LE energy given by 2 1 ¯ 1− Ex . (3.167) SNRM M SE−LE

So zk = 1 −

A new error for the scaled decision regions is ek = (1 − 1/SNRM M SE−LE ) xk −zk = ek −

1 x , SNRMMSE−LE k which is also the old error with the xk dependent term removed. Since ek and xk are then independent, then 2 1 2 2 2 ¯ σe = σM M SE−LE = σe + Ex , (3.168) SNRM M SE−LE

leaving
2 2 σe = σM M SE−LE − 2 ¯ ¯ SNR2 M SE−LE σM M SE−LE − Ex Ex (SNRM M SE−LE − 1) M ¯ Ex = = . 2 SNRM M SE−LE SNRM M SE−LE SNR2 M SE−LE M (3.169) 2 The SNR for the unbiased MMSE-LE then becomes (taking the ratio of (3.167) to σe )

1

2

SNRM M SE−LE,U

¯ Ex SNR2 MMSE−LE = = SNRM M SE−LE − 1 , ¯ (SNRMMSE−LE −1) Ex 2 SNRMMSE−LE

(SNRMMSE−LE −1)2

(3.170)

which corroborates the earlier result on the relation of the optimum biased and unbiased SNR’s for any particular receiver structure (at least for the LE structure). The unbiased SNR is SNRM M SE−LE,U = ¯ Ex
2 σM M SE−LE

−1 ,

(3.171)

which is the performance level that this text always uses because it corresponds to the best error probability for an SBS detector, as was discussed earlier in Section 3.1. Figure 3.27 illustrates the concept with the effect of scaling on a 4QAM signal shown explicitly. Again, MMSE receivers (of which the MMSE-LE is one) reduce noise power at the expense of introducing a bias, the scaling up removes 188

Figure 3.27: Illustration of bias and its removal. the bias, and ensures the detector achieves the best Pe it can. Since the ZFE is also an unbiased receiver, ¯ E (3.158) must hold. The first inequality in (3.158) tends to equality if the ratio σx → ∞ or if the channel 2 ¯ Ex is free of ISI (Q(D) = 1). The second inequality tends to equality if 2 → 0 or if the channel is free of
σ

ISI (Q(D) = 1). The unbiased MMSE-LE loss with respect to the MFB is γM M SE−LE = The
p
2 2 σMMSE−LE σ2

SNRM F B = SNRM M SE−LE,U

2 p 2σM M SE−LE σ2

¯ Ex ¯x − σ2 E M M SE−LE

.

(3.172)
¯ Ex 2 ¯ Ex −σMMSE−LE

in (3.172) is the increase in noise variance of the MMSE-LE, while the term

term represents the loss in signal power at the equalizer output that accrues to lower noise enhancement. The MMSE-LE also requires no additional complexity to implement and should always be used in place of the ZFE when the receiver uses symbol-by-symbol detection on the equalizer output. The error is not necessarily Gaussian in distribution. Nevertheless, engineers commonly make this assumption in practice, with a good degree of accuracy, despite the potential non-Gaussian residual ISI 2 component in σM M SE−LE . This text also follows this practice. Thus, Pe ≈ Ne Q κSNRM M SE−LE,U , (3.173)

¯ where κ depends on the relation of Ex to dmin for the particular constellation of interest, for instance κ = 3/(M − 1) for Square QAM. The reader may recall that the symbol Q is used in two separate ways in these notes, for the Q-function, and for the transform of the matched-filter-pulse-response cascade. The actual meaning should always be obvious in context.)

3.5.3

Examples Revisited

This section returns to the earlier ZFE examples to compute the improvement of the MMSE-LE on these same channels. EXAMPLE 3.5.1 (PAM - MMSE-LE) The pulse response of a channel used with binary PAM is again given by √ π |ω| ≤ T T 1 + .9eωT . (3.174) P (ω) = π 0 |ω| > T 189

The relatively small energy near the Nyquist Frequency in this example is the reason for the poor performance of both linear equalizers (ZFE and MMSE-LE) in this example.175) Figure 3.81 + 1.U = 1 − 1. ¯ Let us suppose SNRM F B = 10dB (= Ex p 2/ N0 ) and that Ex = 1.8dB . 2 2 which is considerably smaller than σZF E . The MMSE-LE clearly does not have an ISI-free response. To 190 . This is 5.5.9 .82 N0 (1.178) N0 1 √ 2 1. The SNR for the MMSE-LE is SNRM M SE−LE.81 D−1 + 1. 2 The equalizer is 1 W (D) = .81/10 dω (3.28 shows the frequency response of the equalizer for both the ZFE and MMSE-LE.4.5dB).1 + 1.176) (3.30 compares the time-domain responses of the equalized channel.8 cos(ωT ) + 1.28: Comparison of equalizer frequency-domain responses for MMSE-LE and ZFE on baseband example.7 (5.9912 − 1. 1. Clearly the MMSE-LE has a lower magnitude in its response.29 compares the frequency-domain responses of the equalized channel.5dB better than the ZFE (9.175) .3dB = 5. but still not good for this channel. .179) The loss with respect to the MFB is 10dB .175(.7dB = 4. Figure 3.181) (3. but mean-square error/distortion is minimized and the MMSELE has better performance than the ZFE.Figure 3.81 D (3.177) (3.175(.9 p · 1. Figure 3.181) = 3. 2 The σM M SE−LE is computed as 2 σM M SE−LE π T π −T = = = T 2π N0 2 1.3dB.7dB) .

191 .Figure 3.29: Comparison of equalized frequency-domain responses for MMSE-LE and ZFE on baseband example. Figure 3.30: Comparison of equalized time-domain responses for MMSE-LE and ZFE on baseband example.

625(1 + )D1 + . improve performance further. or of the denominator in (3.6 and in Chapter 5 (or codes combined with equalization.1356 = . The roots of Q(D) + 1/SNRM F B . Ex = 1.2 (QAM .187) = .184) (3.15625.5625(1 + . The MMSE-LE high response values are considerably more limited than the ZFE values.1) − .0276 .5625 .213 = 2.181) The equalizer is. − − .MMSE-LE) Recall that the equivalent baseband pulse response samples were given by 1   1 pk = √ . Another yet better method appears in Chapter 4.632 = −. SNRM F B N0 p 2 −1 .31: Comparison of MMSE-LE and ZFE for complex channel example.25D−2 + .31 compares the MMSE-LE and ZFE equalizer responses for this same channel.213 + . are D D D D = 2. (3. SNRMF B 1 W (D) = (Q(D) + 1/SNRM F B ) p N0 2 2 = (3.183).Figure 3. 1+ 2 4 2 T 2 N0 ¯ ¯ = 10dB (= Ex p / 2 ). as discussed in Chapters 10 and 11).4502 + .217 −1.632 = −. decision-assisted equalization is examined in Section 3. (3.5. noting that ¯ Ex p = 1.1356 − 2.180) Again.0612 (3.183) Figure 3. and thus 2 σM M SE−LE = w0 . 192 .4502 = . The complex example also follows in a straightforward manner.25D2 (3.5625/10 = .451 .185) (3.217 .625(−1 + )D−1 + 1.182) or = √ 1.451 −1.0612 = 2. −.0276 − . EXAMPLE 3.186) (3.

22 ) B = (2.22 √ (3.20 − .0612 ) −1.451 A D − 2.5025 (3.7.189).20) = 1.22 . Third.632 − .189) has ignored any terms in the partial fraction expansion that will not contribute to w0 .6 dB better than the ZFE.15625)] = = 4.125(.22 . It is also possible to do considerably better on this channel.451 . While this may be simple from an analytical viewpoint.and (3.632)(2. (3.125(. For these reasons. p as was shown in Figure 3.3 dB below the matched filter bound SNR of 10dB. Also. Secondly. Figure 3. 2 [1 − 1.632)2 A = (2. ϕ∗ (−t) is a continuous time filter and p may be much more difficult to design accurately than an equivalent digital filter.125(.22 −1.451 Then 1.190) √ .193) (3. (3.805) = −1. w0 = A(−. matched filter.0612 − . It may be difficult to design an adaptive.69 ( 6.22 −1.22 .195) but .5625(−4)(2.22 −1.22 −1.632 −1.7 dB ) .632 .22 + B D − 2.0612 2 1.69. While the MMSE-LE performs better. there is a significant deviation from the ideal flat response. and an adaptive equalizer (see Chapter 7). √ 1.451 −1..188) (D − 2.0612 − .0079 .69/4.15625) (3.32.451 −.451 .7.451 −1.6 dB improvement.632 )(D − .U N0 = 1.0612 − 2. 3. Nevertheless.183) becomes W (D) = or W (D) = 1.4 Fractionally Spaced Equalization To this point in the study of the discrete time equalizer. Then. the precise sampling frequency and especially its phase.5625(−.22 −1. Basically. there are several practical problems with the use of the matched filter.0612) √ = 1.805 + 1.0612) √ = 1.189) where the second expression (3. the channel pulse response may not be accurately known at the receiver.13) = 3. sophisticated data transmission systems often replace the matched-filter/sampler/equalizer system of Figure 3.125 SNRM M SE−LE.15625) = ..5025 − 1.3 dB .0612 (3. from (3.632 ) + B(−.7 with the Fractionally Spaced Equalizer (FSE) of Figure 3. γM M SE−LE = 10 log10(10/4.632 − . is instead used. must be known so that the signal can be sampled when it is at maximum strength.1758 . 1.69) = 10 log10 (2.0612 √ )= 1. the sampler and the matched filter have been interchanged with respect to Figure 3.632)(2. The new sampling rate is chosen sufficiently high so 193 .25) )(D − 2.451 + .191) Then.5. but structures not yet introduced will be required.22 . .22 .194) which is 3. First. .5625D2 /(. This bias can be removed by scaling by 5.31 compares the equalized channel responses for both the MMSE-LE and ZFE.0612)(D − . Also the sampling rate has been increased by some rational number l (l > 1).5625(1. a matched filter ϕ∗ (−t) precedes the sampler.0079 + 1.125) .5625(1.632)(2.632 − 2. because of biasing. the MMSE-LE output is everywhere slightly lower than the ZFE output (which is unbiased).5625(−4)(2.22 −1.0612)(2. This deviation is good and gains . the MMSE-LE is one of the most commonly used equalization structures in practice. analog. or equivalently. (3.192) 2 σM M SE−LE = 1.

) The FSE can also exhibit a significant improvement in sensitivity to sampling-phase errors. corresponding to l times as many coefficients to span the same time interval. This possible notch is an 194 .T . The symbol-spaced equalizer cannot interpolate to the correct phase. In practical systems. the best performance of the ensuing symbol-spaced ZFE and/or MMSE-LE can be significantly reduced. in the sampling device. a phase adjustment (effectively interpolating to the correct phase) so as to correct the timing offset. so computation is approximately l times that of symbol-spaced equalization. instead of times kT . information has been lost about the signal by sampling at a speed that is too low in the symbol-spaced equalizer without matched filter. Then. the four most commonly found values for l are 4 . t0. the matched filtering operation and the equalization filtering are performed at rate l/T . and 4. leading to an increase in memory by a factor of l. To investigate this improvement briefly. The resulting noise enhancement can be a major problem in practice. y(kT + t0 ) = m xm · p · q(kT − mT + t0) + nk . in the original symbol-spaced equalizers (ZFE or MMSE-LE) may have the sampling-phase on the sampler in error by some small offset t0 . as is typical in practice). it is now possible (for certain nonzero timing offsets t0. T ). and the loss in performance can be several dB for reasonable timing offsets. (3. However. Then. it requires a factor of l more coefficients. Equivalently. (The entire filter for the FSE can also be easily implemented adaptively. see Chapter 7.Figure 3. the sampler will sample the matched filter output at times kT + t0 . 2. and the cascade of the two filters is realized as a single filter in practice. eliminating the need for an analog matched filter.197) which is no longer nonnegative real across the entire frequency band. The FSE will digitally realize the cascade of the matched filter and the equalizer. only 1 th of the output samples need be computed (that is. In fact.196) which corresponds to q(t) → q(t + t0 ) or Q(ω) → Q(ω)e−ωt0 . via its transfer characteristic. and at frequencies just below the Nyquist Frequency) that the two 2π aliased frequency characteristics Q(ω)e−ωt0 and Q(ω − 2π )e−(ω− T )t0 add to approximately zero. Q(e−ωT ) → 1 T Q(ω − n 2πn −(ω− 2πn )t0 T . thus T π π producing a notch within the critical frequency range (. An anti-alias or noise-rejection filter always precedes the sampling device√ (usually an analog-to-digital converter). Then. Equivalently. because of noise enhancement. For the system with sampling offset. because no interpolation is correctly performed at the symbol rate. to l times as many input samples per symbol. as to be greater than twice the highest frequency of xp (t).32: Fractionally spaced equalization. the FSE can synthesize. then information about the entire signal waveform is retained. those at the symbol rate. This text assumes this filter is an ideal lowpass filter with gain T transfer π π over the frequency range −l T ≤ ω ≤ l T . The variance of the noise samples at the filter output will then N0 be 2 l per dimension. The major 3 drawback of the FSE is that to span the same interval in time (when implemented as an FIR filter. as that is when we need l them to make a decision). The FSE outputs may also appear to be computed l times more often in real time. )e T (3. Then. When the anti-alias filter output is sampled at greater than twice the highest frequency in xp (t). or equivalently. 3.

(3. but can otherwise provide any sampling phase. . The sampled output of the anti-alias filter can be decomposed into l sampled-at-rate-1/T interleaved sequences with D-transforms Y0(D). whereas the symbol-spaced equalizer aliases before it equalizes. Pl−1 (D) Nl−1 (D)  (3. which can be written in vector form as E E(D)Y ∗ (D−∗ ) = RxY (D) − W (D)RY Y (D) = 0 . The sensitivity to sampling phase is channel dependent: In particular. Yl−1 (D) Also    P0(D) N0 (D) ∆     . the FSE is avoided. the oversampling factor.200) By considering the FSE output at sampling rate 1/T . . . Wl (D)]. (3. the orthogonality condition says that E(D) = X(D) − Z(D) should be orthogonal to Y (D). the FSE equalizes before it aliases (aliasing does occur at the output of the equalizer where it decimates by l for symbol-by-symbol detection). if the extra memory and computation can be accommodated. The phase is tacitly corrected to the optimum phase inside the linear filter implementing the FSE. and is significantly more complex to implement (more parameters and higher sampling rate).198) where Pi(D) is the transform of the symbol-rate-spaced-ith -phase-of-p(t) sequence p[kT − (i − 1)T /l].204) MMSE-FSE filter setting is then W (D) = RxY (D)R−1 (D) = P ∗ (D−∗ ) P (D)P ∗ (D−∗ ) + l/SNR YY The corresponding error sequence has autocorrelation function . Y (D) =  (3. Then. Ree(D) = Ex − RxY (D)R−1 (D)RY x (D) = ∗ −∗ YY P (D )P (D) + l/SNR 195 (3. In effect. . . . Effectively. the sampling device need only be locked to the symbol rate.. 2 −1 (3.199)  = P (D)X(D) + N (D) . Yl−1 (D).201) Again. this loss cannot be recovered in a symbol-spaced equalizer without a matched filter.205) l · N0 2 ¯ ¯ .202) ¯ E X(D)Y ∗ (D−∗ ) = Ex P ∗ (D−∗ ) N0 ¯ E Y (D)Y ∗ (D−∗ ) = Ex P (D)P ∗ (D−∗ ) + l · I . Y2 (D). the former alternative is often the one of choice in practical system implementation. P (D) =   and N (D) =   . .206) . and these noise sequences are also independent of one another. and similarly Ni (D) is the transform of a symbol-rate sampled white noise sequence with autocorrelation function Rnn(D) = l · N0 per dimension. . there is usually significant channel energy near the Nyquist Frequency in applications that exhibit a significant improvement of the FSE with respect to the symbol-spaced equalizer.203) (3. Yi (D) = Pi (D) · X(D) + Ni (D) (3. In channels with little energy near the Nyquist Frequency. where Yl (D) corresponds to the sample sequence y[kT − iT /l]. Infinite-Length FSE Settings To derive the settings for the FSE. because it provides little performance gain. where RxY (D) RY Y (D) = = ∆ ∆ (3... with an FSE... 2 A column vector transform is   Y0 (D)   .example of information loss at some frequency.. the interleaved coefficients of the FSE can also be written in a row vector W (D) = [W1(D) . is an integer. this section assumes that l. Thus the FSE output is Z(D) = W (D)Y (D) . .

5 and 3. and absorbing the matched filtering operation into the filter wk . An alternative method. The main reason for the interchange of filtering and demodulation is the recovery of the carrier frequency in practice. In practice. or any other desired setting (see Chapter 4 and Sections 3. which would then be identical to the (modulated) baseband-equivalent FSE with output sample rate decimated appropriately to produce outputs only at the symbol sampling instants. The phase splitting is the forming of the analytic equivalent with the use of the Hilbert Transform. This is because the matrix 2 RY Y (D) will often be singular when this occurs. The reader is cautioned against letting N0 → 0 to get the ZF-FSE. Passband Equalization This chapter so far assumed for the complex baseband equalizer that the passband signal has been demodulated according to the methods described in Chapter 2 before filtering and sampling. biased and unbiased. are also then exactly the same as given for the MMSE-LE.Figure 3. Postponing the carrier demodulation to the equalizer output can lead to significant improvement in the tolerance of the system to any error in estimating the carrier frequency. This commuting can be done with either symbol-spaced or fractionally spaced equalizers.” meaning that no carrier demodulation occurs prior to sampling. the cross-coupled nature of the complex equalizer W (D) is avoided by sampling the FSE at a rate that exceeds twice the highest frequency of the passband signal y(t) prior to demodulation. The SNR’s. as long as the sampling rate exceeds twice the highest frequency of P (f).33. The filter wk can be a ZFE. sometimes used in practice when ωc is not too high (or intermediate-frequency up/down conversion is used). the final rotation shown in Figure 3. (3. The MMSE is then computed as MMSEM M SE−F SE = l T 2π π T π −T N0 dω 2 −ωT ) 2 + P (e l· l/SNR = MMSEM M SE−LE . increasing the sampling rate. (If intermediate frequency (IF) demodulation is used. The passband equalizer is illustrated in Figure 3. other than a translation in frequency by ωc radians/sec. Passband equalization is best suited to CAP implementations (See Section 4 of Chapter 2) where the complex channel is the analytic equivalent. To avoid problems with singularity.207) where P (e−ωT ) 2 = i=1 |Pi(e−ωT )|2. The matched filtering and equalization are then performed on the passband signal with digital demodulation to baseband deferred to the filter output. an appropriate pseudoinverse. Often in practice. In the present (perfect carrier phase lock) development. as discussed in Chapter 2. Sometimes this type of system is also called “direct conversion.6).33: Passband equalization. The complex equalizer then acts on the analytic equivalent channel and signals to eliminate ISI in a MMSE sense.) In 196 . a MMSE-LE. is to commute the demodulation and filtering functions. then twice the highest frequency after the IF. passband equalization and baseband equalization are exactly equivalent and the settings for the passband equalizer are identical to those of the corresponding baseband equalizer. When used with CAP. the equalizer can again be realized as a fractionally spaced equalizer by interchanging the positions of the matched filter and sampler.33 is not necessary – such a rotation is only necessary with a QAM transmit signal. as Chapter 6 investigates. which zeros the FSE characteristic when P (ω) = 0 is recommended. direct synthesis of analytic-equivalent equalizer.

that nominally characterize complex convolution). 197 . and one the imaginary part.this case the phase splitter (contains Hilbert transform in parallel with a unit gain) is also absorbed into the equalizer. A particularly common variant of this type of equalization is to sample at rate 2/T both of the outputs of a continuous-time phase splitter. This approach is often taken with adaptive implementations of the FSE. 10 However. real and imaginary part. The two filters can sometimes (depending on the sampling rate) be more cost effective to implement than the single complex filter. whereas the two filters with fixed conjugate symmetry could not. one of which estimates the real part. of the analytic representation of the data sequence xk eωc kT . especially when one considers that the Hilbert transform is incorporated into the equalizer. The corresponding T /2 complex equalizer can then be implemented adaptively with four independent adaptive equalizers (rather than the two filters. as the Hilbert transform is then implemented more exactly adaptively than is possible with fixed design.10 This is often called Nyquist Inband Equalization. one stream of samples staggered by T /4 with respect to the other. and the same sampled sequence is applied independently to two equalizers. The adaptive filters will then correct for any imperfections in the phase-splitting process. this structure also slows convergence of many adaptive implementations.

W (D). and can be a significant weakness of decision-feedback that cannot be overlooked.8. it can be analyzed using linear techniques. This section assumes that previous decisions are correct.” The output of the feedforward filter is denoted Z(D). There is both a MMSE and a zero-forcing version of decision feedback.34 is absorbed into the SBS. The DFE is inherently a nonlinear receiver. Thus. and the input to the decision element Z (D). The Decision Feedback Equalizer (DFE) is shown in Figure 3.34. this section derives the MMSE solution. (3. augmented by a linear.209) The error sequence can be written as E(D) = X(D) − W (D) · Y (D) − [1 − B(D)]X(D) = B(D) · X(D) − W (D) · Y (D) . the most efficient way to specify the effect of feedback errors has often been via measurement.6 Decision Feedback Equalization Decision Feedback Equalization makes use of previous decisions in attempting to estimate the current symbol (with an SBS detector).208) .6. if one assumes all previous decisions are correct. causal.34: Decision feedback equalization. (the settings for this linear equalizer are not necessarily the same as those for the ZFE or MMSE-LE). Any “trailing” intersymbol interference caused by previous symbols is reconstructed and then subtracted. Nevertheless. 3.1 shows how to eliminate error propagation with precoders.210) . and as was true with linear equalization in Sections 3.4 and 3. The feedback section will then subtract (without noise enhancement) any trailing ISI. the name “decision feedback. The MMSE for the MMSE-DFE is 2 σM M SE−DF E = min E |xk − zk |2 wk . feedback filter. The feedforward filter will try to shape the channel output signal so that it is a causal signal.1 (Mean Square Error (for DFE)) The MMSE-DFE error signal is ek = xk − zk . where b0 = 1.5. 1 − B(D).bk ∆ (3. 198 (3. The feedback filter accepts as input the decision from the previous symbol period. Section 3. However.Figure 3. this may not be true. 3. thus. To date. the analysis becomes intractable if it includes errors in the decision feedback section. The configuration contains a linear feedforward equalizer.1 Minimum-Mean-Square-Error Decision Feedback Equalizer (MMSEDFE) The Minimum-Mean-Square Error Decision Feedback Equalizer (MMSE-DFE) jointly optimizes the settings of both the feedforward filter wk and the feedback filter δk − bk to minimize the MSE: Definition 3. Section 3. the zero-forcing solution will be a special case of the least-squares solution with the SNR→ ∞. In practice.7 provides an exact errorpropagation analysis for finite-length DFE’s that can (unfortunately) require enormous computation on a digital computer.6. Any bias removal in Figure 3. which subsumes the zero-forcing case when SNR → ∞.

0 ≥ N0 B(D) · B ∗ (D−∗ ) · 22 Q(D) + 1/SNRM F B p N0 B(D) B ∗ (D−∗ ) 2 · ∗ −∗ · G(D) G (D ) γ0 p N0 2 (3.e. so that E(D) = B(D) · EMMSE-LE (D)) The autocorrelation function for the error sequence with arbitrary monic B(D) is ¯ Ree(D) = = = = ¯ ¯ ¯ B(D)Ex B ∗ (D−∗ ) − 2 B(D)Rxy (D)W ∗ (D−∗ ) + W (D)Ryy (D)W ∗ (D−∗ )(3.217) where γ0 is a positive real number and G(D) is a canonical filter response. anti-causal. a consequence of the linearity of the MMSE estimate. and all its zeroes are on or outside the unit circle).215) 1 p Q(D) + SNRMF B B(D)RM M SE−LE (D)B ∗ (D−∗ ) (3.222) the fractional polynomial inside the squared norm is necessary monic and causal. W (D) = p B(D) 1 Q(D) + SNR MF B (3.. and minimum-phase (all its poles are outside the unit circle. and “maximum-phase” (all poles inside the unit circle. (Also. The linear prediction approach is developed more in what is called the “linear prediction DFE. Using this factorization. and therefore the squared norm has a minimum value of 1. and all zeros in or on the unit circle). p · γ0 · G(D) · G∗(D−∗ ) p · γ0 · G∗ (D−∗ ) (3.213) ¯ ¯ B(D)Ex B ∗ (D−∗ ) − W (D)Ryy (D)W ∗ (D−∗ ) (3.212) for any B(D) with b0 = 1. Thus.219) (3. the MMSE will then be σM M SE−DF E = feedforward filter then becomes γ0 p 2 . The W (D) = G(D) 1 = . Thus. B(D) and W (D) specify the MMSE-DFE: 199 .” in an exercise where the MMSE-DFE is the concatenation of a MMSE-LE and a linear-predictor that whitens the error sequence. 2 with equality if and only if B(D) = G(D). The solution for B(D) is then the forward prediction filter associated with this error sequence as discussed in the Appendix.220) follows from the observations that ree. (3. A filter response G(D) is called canonical if it is causal (gk = 0 for k < 0). If G(D) is canonical. which leads us to the relation ¯ ¯ B(D) · Rxy (D) − W (D) · Ryy (D) = 0 .218) (3.221) The last step in (3.For any fixed B(D).214) ¯ Rxy (D) ¯ B(D)Ex B ∗ (D−∗ ) − B(D) B ∗ (D−∗ ) (3.216) N0 where RM M SE−LE (D) = p2 2 1/(Q(D) + 1/SNRM F B ) is the autocorrelation function for the error sequence of a MMSE linear equalizer. (3.220) N0 2 2 γ0 p 2 . the (scaled) inverse autocorrelation has spectral factorization: Q(D) + 1 = γ0 · G(D) · G∗ (D−∗ ). SNRM F B (3. monic (g0 = 1).0 = B G 2 N0 2 γ0 · p 2 . E [E(D)Y ∗ (D−∗ )] = 0 to minimize MSE. ¯ Ree(D) = = ree. In more detail on B(D). then G∗(D−∗ ) is anti-canonical i. monic. W (D) = WMMSE-LE (D) · B(D).211) = B(D) · WMMSE-LE (D) .

228) 2 This last result leads to a famous expression for σM M SE−DF E .6. then SNRM M SE−DF E = SNRM F B + 1.233) (3.223) (realized with delay.235) (3.231) From the k = 0 term in the defining spectral factorization.232) From this expression.Lemma 3.229) The SNR for the MMSE-DFE can now be easily computed as ¯ Ex SNRM M SE−DF E = = γ0 · SNRM F B 2 σM M SE−DF E = SNRM F B e T 2π π −T π T (3.230) 1 ln Q(e−ωT ) + SNR dω MF B .236) . N0 2 2 σM M SE−DF E = p 2 ·e T − 2π π −T π T 1 ln Q(e−ωT ) + SNR dω MF B . so that the SNR would be higher than the matched filter bound.234) (3. the autocorrelation function for the error sequence is ¯ Ree(D) = γ0 · p 2 . This bias is noted by writing Z (D) = X(D) − E(D) 1 = X(D) − G(D) · X(D) + Y (D) p · γ0 · G∗(D−∗ ) Q(D) 1 = X(D) − G(D)X(D) + X(D) + N (D) γ0 G∗ (D−∗ ) p γ0 G∗ (D−∗ ) 1/SNRM F B 1 = X(D) − X(D) + N (D) . p 0 T 2π π T π −T ln Q(e−ωT ) + 1 SNRM F B dω = = ln(γ0 ) + ln(γ0 ) . (3. (3.1 (MMSE-DFE) The MMSE-DFE has feedforward section W (D) = 1 p · γ0 · G∗(D−∗ ) (3. γ0 can also be written γ0 = 1 + 1/SNRM F B 1 + 1/SNRM F B = g 2 1 + ∞ |gi|2 i=1 . The reason for this apparent anomaly is the artificial signal power introduced by biased decision regions.6. T 2π π T π −T ln G(e−ωT )G∗ (e−ωT ) dω(3.2 Performance Analysis of the MMSE-DFE N0 2 Again.227) (3. These settings for the MMSE-DFE minimize the MSE as was shown above. 3. which was first derived by Salz in 1973. as it is strictly noncausal) and feedback section B(D) = G(D) where G(D) is the unique canonical factor of the following equation: Q(D) + 1 = γ0 · G(D) · G∗ (D−∗ ) . γ0 · G∗(D−∗ ) p · γ0 · G∗ (D−∗ ) 200 (3. SNRM F B (3.224) This text also calls the joint matched-filter/sampler/W (D) combination in the forward path of the DFE the “Mean-Square Whitened Matched Filter (MS-WMF)”. Also. if G(D) = 1 (no ISI). (3.225) (3. (3.226) This last result states that the error sequence for the MMSE-DFE is “white” when minimized (since N0 ¯ Ree(D) is a constant) and has MMSE or average energy (per real dimension) 2 2 γ −1 .

causal polynomial SNR2 MMSE−DF E GU (D) = 1 SNRM M SE−DF E G(D) − SNRM M SE−DF E.240) p · γ0 · G∗(D−∗ ) γ0 · G(D) · G∗ (D−∗ ) − 1/SNRM F B 1 X(D) + N (D) (3. SNRM M SE−DF E (3.242) + N (D) SNRM F B · γ0 · G∗(D−∗ ) 1 1 1 X(D) G(D) − + 1 − ∗ −∗ X(D) + N(3.U 201 (3. Also. SNRM M SE−DF E.239) 1 (3.248) . (3.238) 1 1 Rather than scale the decision input. Thus. has autocorrelation Rnn(D) = Z(D) = [X(D) · p · Q(D) + N (D)] = = = = N0 2 Q(D). SNRMMSE−DF E This will remove the bias. If G(D) = 1. SNRM M SE−DF E (3. The SNR will then be SNRM M SE−DF E.246) 2 2 The error ek is not a white sequence in general.The current sample. corresponds to the time zero sample of 1 − 1/SNRMF B .U SNRM M SE−DF E . σe = σM M SE−DF E = ¯ Ex 2 σe + .249) SNRM M SE−DF E. This leads to the more informative SNR SNRM M SE−DF E. but also increase MSE by the square of the same factor. If one defines an “unbiased” monic.U .243) (D) SNRM M SE−DF E SNRM M SE−DF E G (D ) 1 X(D) G(D) − + E (D) . Since xk and ek are uncorrelated.1. 1 p · γ0 · G∗ (D−∗ ) Z(D) expands to (3. xk .U = SNRM M SE−DF E − 1 = γ0 · SNRM F B − 1 . (3.U = SNRM F B .237) (3.U X(D)GU (D) + E (D) . This result is verified by writing the MS-WMF output Z(D) = [X(D) · p · Q(D) + N (D)] ¯ where N (D).245) 1 xk .247) then the MMSE-DFE output is Z(D) = ZU (D) removes the scaling ZU (D) = Z(D) SNRM M SE−DF E = X(D)GU (D) + EU (D) . the SNR corresponding to the lowest probability of error corresponds to the same MMSE-DFE receiver with output scaled to remove the bias. the receiver can scale (up) the feedforward output by . (3. 1− (3.2.244) SNRM M SE−DF E where N (D) is the filtered noise at the MS-WMF output. and is equal to (1 − γ0 ·G∗ (D −∗ ) 1 )xk . γM M SE−DF E = 1 SNRM F B SNRM F B = = SNRM M SE−DF E.241) γ0 · G∗ (D−∗ ) p · γ0 · G∗(D−∗ ) X(D) X(D) · G(D) − (3. and has autocorrelation function ¯ Rn n (D) = and ek = −ek + N0 2 Q(D) 2 γ0 p 2 · · G(D) · G∗ (D−∗ ) = ¯ Ex SNRM M SE−DF E 1− 1/SNRM F B Q(D) + 1/SNRM F B .U γ0 · SNRM F B − 1 γ0 − 1/SNRM F B . again. Thus zk contains a signal component of xk that is reduced in magnitude. then SNRM M SE−DF E. SNRMMSE−DF E again using the result from Section 3.

253) (3. ∗ η0 · p · Pc (D−∗ ) (3. as in Section 3. The feedback section of this new unbiased MMSE-DFE becomes the polynomial 1−GU (D). The spectral factorization (assuming ln(Q(eωT ) is integrable over (− T . 3.U . All the results on fractional spacing for the LE still apply to the feedforward section of the DFE – thus this text does not reconsider them for the DFE. but is 1 − G(D) if scaling is performed after the summing junction.254) 2 ¯ SNR2 M SE−DF E σM M SE−DF E − Ex SNR2 M SE−DF E M M SNR2 M SE−DF E SNR2 M SE−DF E. Again. if the scaling is performed before the summing junction.256) (3.U M ¯ Ex [SNRM M SE−DF E − 1] SNR2 M SE−DF E.35.35: Decision feedback equalization with unbiased feedback filter.U SNR for this scaled and unbiased MMSE-DFE is thus verified to be = SNRM M SE−DF E. The alternative unbiased MMSE-DFE is shown in Figure 3. other than to state that the characteristic of the feedforward section is different than the LE characteristic in all but trivial channels.250) (3. then η0 = 1 1+ ∞ i=1 |pc. the realization of any characteristic can be done with a matched filter and symbol-spaced sampling or with an anti-alias filter and fractionally spaced sampling.U M M ¯x [(SNRM M SE−DF E /Ex )SNRM M SE−DF E σ2 ¯ E M M SE−DF E − 1] SNR2 M SE−DF E.255) then determines the ZF-DFE filters as B(D) = Pc (D) .257) Pc is sometimes called a canonical pulse response for the channel. T ). (3.U 2 (3.252) (3.6.Figure 3. SNR where EU (D) = SNR MMSE−DF E E (D) and has power MMSE−DF E. see Appendix A) ∗ Q(D) = η0 · Pc (D) · Pc (D−∗ ) (3.5.258) 202 .4.U ¯ Ex ¯ Ex /SNRMMSE−DF E.i|2 .3 Zero-Forcing DFE The ZF-DFE feedforward and feedback filters are found by setting SNRM F B → ∞ in the expressions π π for the MMSE-DFE. SNRM M SE−DF E. W (D) = 1 .U M ¯ Ex .U 2 σeU = = = = = 2 σM M SE−DF E − ¯ Ex SNR2 M SE−DF E M SNRM M SE−DF E SNRM M SE−DF E.251) (3. Since q0 = 1 = η0 · Pc 2 .

1 1.4 Examples Revisited EXAMPLE 3.81(.269) . (3.633D)(1 + .267) with magnitude in Figure 3.58.1274 = 6. In this case σZF DF E = . 1. The feedforward section is W (D) = √ 1 .263) (1 + .1274 Thus. SNRM M SE−DF E.81 = .6 dB.9469 .i · SNRM F B . the ZF2 DFE for this same channel would produce η0 = 1/ pc 2 =. = 1 + .U = It will in fact be possible to do yet better on this channel. and Pc(D) is the minimum-phase equivalent of the channel pulse response.6 dB. The function Q(D) has canonical factorization ˜ Q(e−ωT ) = ˜ Q(D) = .6.265) and the feedforward filter transfer function appears in Figure 3.81 1 ˜ Q(D) = (3. ∆ ˜ Then Q(D) = Q(D) + 1/SNRM F B is computed as 1. γ0 = .785.81 1 = (3.181/(. what originally appeared as a a stunning loss of 9. Then.785 (3.9D−1) + .633 and −1. so 2 that π T ωT T N0 N0 − π ln Q(e ) dω 1 2 2 2 2π − T σZF DF E = = e . The feedback section is B(D) = G(D) = 1 + . The noise at the output of the feedforward filter has (white) autocorrelation function N0 /( p 2 · η0 ).4dB) . The MMSE is 2 σM M SE−DF E = N0 1 .9D−1 .DFE) The pulse response of a channel used with binary PAM is again given by √ π |ω| ≤ T T 1 + .258) shows that there is a signal power loss of the ratio of the squared first tap magnitude in the canonical pulse response to the squared norm of all the taps. as discussed in Chapter 9.9eωT P (ω) = .37. = .259) p 2 η0 p 2 The corresponding SNR is then SNRZF −DF E = η0 · SNRM F B = 1 1+ ∞ i=1 p2 c.81 ˜ ˜ The roots of Q(D) are −.81 + 1.1274 2 p 2γ0 1.266) (3. the SNRM F B is 10dB and Ex = 1.261) π 0 |ω| > T Again.81/10 (3. using sequence detection.633D−1) .268) 1 − .9D + 1. This loss ratio is minimized for a minimum phase polynomial.991 + . the MMSE-DFE is only 1.85 (8. The MS-WMF filter transfer function appears in Figure 3. (3. 1 dB lower than the MMSE-DFE for this channel.785(1 + .36.8 cos(ωT ) + 1.181 =. (3.785)(1 + .633D−1 1.5525.5525)1.38. (3. illustrating the near all-pass character of this filter.9D)(1 + . However. However.181 and the loss with respect to the MFB would be η0 = 2.6. and/or by codes (Chapters 10 and 11).8 dB with the ZFE has now been reduced to a much smaller 1.264) .262) 1. 203 and .6dB below the MFB on this channel.Equation (3. (3.260) 3.81 · .1 (PAM .633D .633D−1) (3.

Figure 3. Figure 3.37: Feedback filter transfer function for real-channel example. 204 .36: Feedforward filter transfer function for real-channel example.

−2 −1 2 ˜ −.451 .5625 (3.270) − 2 4 2 T The SNRM F B is again 10 dB.7866 and G(D) = 1−.2034D2 and W (D) = The MSE is 2 σM M SE−DF E = (.625(1 + )D + . .5625) (3.4226(1 + )D + .451 . (3.0612D)(1 − .0612) (1 − .25D + .272) from which one extracts γ0 = .2034D2 .274) (3.1271 (3.2 (QAM . . Then.451 −1.276) 205 .5625(1 + . as B(D) = G(D) = 1 − .38: MS-WMF transfer function for real-channel example.25D .0612 D−1 ) · (3.2034D−2 1 = .4dB .4226(1+)D+.632 −1.4226(1 − )D−1 − .1) − .5625 · 4 · (.632 D−1 )(1 − . 1 − .15625) (3.451 1. The feedforward and feedback sections can be computed in a straightforward manner.273) 1.451 −.1271 .U = 1 − 1 = 8.7866(1.451 − · 1. The QAM example is also revisited here: EXAMPLE 3.DFE) Recall that the equivalent baseband pulse response samples were given by 1   1 pk = √ 1+ − .017 .275) and the corresponding SNR is SNRM M SE−DF E.Figure 3.271) or ˜ Q = D)(1 − . Q= 1.632)(.625(−1 + )D + 1.6.

6 dB with respect to the MFB and is 1.1563 (3.5 π/2 −π/2 D−1 )(1 − .5(1 − )D−1 − .25D2.5) (1 − . 206 . The feedforward and feedback sections can be computed in a straightforward manner. .5D)(1 − . Q(D) = D)(1 − .280) 1 = .281) and the corresponding SNR is SNRZF DF E = 1 = 6.5625 · 4 · (.4 = 8.6400(1.5 − · 1.15625) (3.277) (3.0 dB with respect to the MFB and is .5D−1 ) · (3.5(1 + )D + .The loss is (also coincidentally) 1.5 −π/2 )(.278) and thus η0 = .25D2 and W (D) = The output noise variance is 2 σZF DF E = (.5625) (3.0dB .4 dB worse than the MMSE-DFE. For the ZF-DFE.25 .6400 and Pc (D) = 1 − .1563 . .279) 1. as B(D) = Pc (D) = 1 − .7 dB better than the MMSE-LE on this channel. 1 − .282) The loss is 2.5(1 + )D + .25D−2 (3.

i = 0. The assumption that l is an integer simplifies the specification of matrices. Because of the finite length. This section describes design for the MMSE situation and then lets SNR→ ∞ to get the zero-forcing special case. are almost exclusively realized as finiteimpulse-response (FIR) filters in practice. The l phases per symbol period of the oversampled y(t) are contained in   y(kT )  y(kT − T )  l   yk =  (3. l − 1. T . Equalization filters. ˜ iT l (3. ˜ One way to view the oversampled channel output is as a set of l parallel T -spaced subchannels whose pulse responses are offset by T /l from each other as in Figure 3. 3.39. . it is convenient to represent such a system with vectors as shown below.8 and Chapter 13) are often implemented with FIR structures for wk (and also for bk . l l l m=−∞ √ The (per-dimensional) variance of the noise samples is N0 · l because the gain of the anti-alias filter. Usually these structures have better numerical properties than IIR structures. wk . Even more importantly.7 Finite Length Equalizers The previous developments of discrete-time equalization structures presumed that the equalizer could exist over an infinite interval in time.32. the matched filtering operation will be performed digitally (at sampling rate l/T ) and the FIR MMSE-LE will then incorporate both matched filtering and equalization. Each subchannel produces one of the l phases per symbol period of the output set of samples at sampling rate l/T . y(kT − l−1 T) l 207 . adaptive equalizers (see Section 3.39: Polyphase subchannel representation of the channel pulse response. 2 is absorbed into p(t).7. √ Perfect anti-alias filtering with gain T precedes the sampler and the combined pulse-response/anti-alias filter is p(t). the FIR zero-forcing equalizer cannot completely eliminate ISI in general.. This section studies the design of and the performance analysis of FIR equalizers. . Mathematically.283) which.284) y(kT − iT iT iT xm · p(kT − ˜ )= − mT ) + n(kT − ) .. . if sampled at time instants t = kT − ∞ . Both the LE and the DFE cases are examined for both zero-forcing and least-squares criteria. The channel output y(t) is y(t) = m xm · p(t − mT ) + n(t) . becomes (3. 3.1 FIR MMSE-LE Returning to the FSE in Figure 3. in the case of the adaptive DFE).285)  .Figure 3..   .

.. w. (3. xk−ν Each row in (3. .286) where   pk =    p(kT ) ˜ p(kT − T ) ˜ l .. .287) The response p(t) is assumed to extend only over a finite time interval. pk = 0 for k < 0 .39. (3... 0 0 .. p1 .290) This text uses P to denote the large (Nf · l) × (Nf + ν) matrix in (3.. and Y k . The sum in (3. . xk−Nf −ν+1 .  yk = [ p0 p1 . .290).. and N denotes the noise vector. with each row possibly having a set of coefficients unrelated to the other rows. . .  . nk−Nf +1 sample at the output of one of the filters in Figure 3. this assumption ˜ requires any nonzero component of p(t) outside of this time interval to be negligible. In practice. . . the oversampled vector representation of the channel is Y k = P Xk + N k .. . . . . ..288) corresponds to a generally. this case. .284) becomes   xk  xk−1     . 0      pν .291) still holds with P an [Nf · m ] × (Nf + ν) matrix that does not follow the form in (3. pν ]  .. . 11 In n m (3.11 An equalizer is applied to the sampled channel output vector Y k by taking the inner product of an Nf l-dimensional row vector of equalizer coefficients.288)   . .   . This time interval is 0 ≤ t ≤ νT .. ... 0   0      . n(kT − l−1 l T) (3. . . (3. Nf · integer so that P is constant.      (3. nk nk−1 . then (3. . Otherwise. pν  xk xk−1 . . .283) is used to compute each row at the appropriate sampling instants. .. while X k denotes the data vector.  + nk ..  . p0 0 pν .290).  (3.. . Thus. so Z(D) = W (D)Y (D) can be written zk = wY k .289)          yk−Nf +1 . and for k > ν. p1 . . ...       (3. for Nf successive l-tuples of  yk  yk−1 ∆  Yk =  . P becomes a “time-varying” matrix P k .. . 0 p1 p0 . .  =      +     p0 0 ..The vector yk can be written as ∞ ∞ yk = m=−∞ xm · pk−m + nk = m=−∞ xk−m · pm + nk ..292) should also be an 208 . p(kT − ˜ l−1 l T)       and nk =    n(kT ) n(kT − T ) l .291) n When l = n/m (a rational fraction). Then. More samples of y(t).

294) .. ∆ is approximately the sum of the channel and equalizer delays in symbol periods.293) with the exact value being of little consequence unless the equalizer length Nf is very short.303) (3.2 (FIR MMSE-LE) The FIR MMSE-LE for sampling rate l/T . (3. 0 P ∗ + 0 ¯ Ex [ 0 .296) In other words.297) have a very special significance (both y(t) and xk are stationary): Definition 3. The equalizer output error is then ek = xk−∆ − zk .298) (3...1 (Autocorrelation and Cross-Correlation Matrices) The FIR MMSE autocorrelation matrix is defined as RY Y = E {Y k Y ∗ } /N . it may be of interest to derive more specific expressions for RY Y and RxY .300) 209 ..295) is minimized when E {ek Y ∗ } = 0 . the designer obtains the MMSE-LE: Definition 3.. ∆≈ ν + Nf 2 . Y In general.. and the delay allows time for the transmit data symbol to reach the receiver.7. (3. the designer picks a channel-equalizer system delay ∆ · T symbol periods. w∗ = R−1Y RY x . ∆ ∆ (3. 0 ] ν 0 (3.297) The two statistical correlation quantities in (3. The delay is important in finite-length design because a non-causal filter cannot be implemented. and of length Nf symbol periods has coefficients w = R∗ x R−1Y = RxY R−1Y Y Y Y or equivalently... p∗ 0 . Thus E {xk−∆ Y ∗ } − wE {Y k Y ∗ } = 0 .301) .295) Using the Orthogonality Principle of Appendix A. k k (3. k (3. the need for such a delay does not enter the mathematics because the infinite-length filters are not realizable in any case. 0 Ex 0 . 0 p∗ ..7. and the corresponding MSE is 2 σM M SE−LE = E |ek |2 = E {ek e∗ } = E (xk−∆ − zk ) (xk−∆ − zk ) k ∗ (3. Note RxY = R∗ x .302) (3.299) where N = 1 for real and N = 2 for complex. the equalizer error signal will be minimized (in the mean square sense) when this error is uncorrelated with any channel output sample in the delay line of the FIR MMSE-LE. With the above definitions in (3.297). delay ∆. With infinite-length filters. and RY Y is not a Y function of ∆. RxY = = = E {xk−∆ Y ∗ } = E {xk−∆X ∗ }P ∗ + E {xk−∆N ∗ } k k k ¯ 0 ..304) (3. the MSE in (3.For causality. (3. so that infinite-length analysis simply provides best bounds on performance. k while the FIR MMSE-LE cross-correlation vector is defined by R Y x = E Y k x∗ k−∆ /N .

0 p∗ .316) 3. 2 (3. (3. σM M SE−LE . nn (3. (3.0] = 1∗ P ∗ .U .310) (3.308) The matrix inversion lemma of Appendix B is used in deriving the above equation.319) 210 ..312) (3.304) depend on the choice of ∆. can be found by evaluating (3.306) where the (l · N0 -normalized) noise autocorrelation matrix Rnn is equal to I when the noise is white.. p 2 ∆ (3.U = SNRM M SE−LE − 1 . 2 A convenient expression is [0.311) (3. 2 The position of the nonzero elements in (3.p∗ 0.2 FIR ZFE ¯ RY Y = P P ∗ Ex One obtains the FIR ZFE by letting the SNR→ ∞ in the FIR MMSE. The loss with respect to the MFB is. The MMSE. (3.317) (3. Then the FIR MMSE-LE becomes (in terms of P .314) so that the best value of ∆ (the position of the 1 in the vectors above) corresponds to the smallest diagonal element of the inverted (“Q-tilde”) matrix in (3. Thus.315) σM M SE−LE and the corresponding unbiased SNR is SNRM M SE−LE.. SNR and ∆) w = 1∗ P ∗ P P ∗ + ∆ 1 Rnn SNR −1 = 1∗ P ∗ R−1 P + ∆ nn l ·I SNR −1 P ∗ R−1 . again..RY Y = = E {Y k Y ∗ } = P E {X k X ∗ }P ∗ + E {N k N ∗ } k k k N0 ∗ ¯ Ex P P + l · · Rnn .300): 2 σM M SE−LE = = = = = ∗ ∗ ∗ ∗ ∗ E xk−∆x∗ k−∆ − xk−∆ Y k w − wY k xk−∆ + wY k Y k w ¯ Ex − RxY w∗ − wRY x + wRY Y w∗ ¯ Ex − wRY Y w∗ ¯ Ex − RxY R−1Y RY x Y ¯ Ex − wRY x . (3.305) (3. γM M SE−LE = SNRM F B SNRM M SE−LE.314) – this means that only one matrix need be inverted to obtain the correct ∆ value as well as compute the equalizer settings.313) is the same as 2 σM M SE−LE l · N0 ∗ 2 = 1 p 2 ∆ P ∗ R−1 P l nn + I p 2 SNRM F B −1 1∆ = l · N0 ∗ ¯ −1 2 1 Q 1∆ .295) with (3.7. (3..318) and then w remains as w = RxY R−1Y Y .313) With algebra.. (3. which alters RY Y to (3.309) (3. the SNR for the FIR MMSE-LE is ¯ Ex SNRM M SE−LE = 2 .307) ν 0 ∆ so that 1∆ is an Nf +ν-vector of 0’s and a 1 in the (∆+1)th position. A homework problem develops some interesting relationships for w in terms of P and SNR.

322) The ZFE is still unbiased in the finite-length case:13 E [wY k /xk−∆] = = RxY R−1Y {P E [X k /xk−∆] + E[N k ]} Y [0 0 . (3. (3. This ISI power is given by 2 ¯ σM M SE−ZF E = Ex − wRY x . [0 0 .326) is true if ∆ is a practical value and the finite-length ZFE has enough taps.9 1 With a choice of ∆ = 2.1 0]P ∗ (P P ∗ )−1P E  .Because the FIR equalizer may not be sufficiently long to cancel all ISI. the FIR ZFE may still have nonzero residual ISI. .328) 0 0 . choose Nf = 3 and note that ν=1. (3.         (3.9  RY x (3.330) 12 The 13 The notation w 2 matrix is a “projection matrix” and P P ∗ is full rank. SNRM F B γZFE = .. P ∗ (P P ∗ )−1 P Rnn means wRnn w ∗ . .    xk−Nf    don’t care xk−∆ . The loss is the ratio of the SNR to SNRZF E .1 0]E P ∗ (P P ∗ )−1 P X k /xk−∆  xk     xk−1    .9 1 0 0 P =  0 .320) However. (3.324) =     /xk−∆         (3.327) SNRZF E 3.326) Equation (3. 2 making the SNR at the ZFE output SNRZF E = ¯ Ex ¯ Ex − wRY x + N0 l w 2 2 Rnn ..12 The power of this noise is easily found to be N0 (3. therefore the entry of xk−∆ passes directly (or is zeored.325) = (3. 211 .321) ·l· w 2 FIR ZFE noise power = Rnn ..7.3 example For the earlier PAM example. Then   .9 1 0  . then  0  0  ¯ = Ex P    1  0   0 =  1  . in which case ∆ needs to be changed).320) still ignores the enhanced noise at the FIR ZFE output. one notes that sampling with l = 1 is sufficient to represent all signals.323) (3.329) (3.. First. (3.

331) = The FIR MMSE is w∗ = R−1Y RY x Y Then 2 σM M SE−LE (3. Clearly.333) (3.9 1. the infinite length performance level can be attained.8 dB) . 0 .335) which is 6.991 .9 0  .51  .849 −.174 0 −. .991 . With sufficiently large number of taps for this channel.4 (3.23 =  −.40: FIR equalizer performance for 1 + .9 1.174 −.384 . This performance is plotted as versus the number of equalizer taps in Figure 3. and RY Y = 1 ¯ Ex P P ∗ + I SNR   1.676 −.9 .2dB below MFB performance and 1. sampling with l = 1 is sufficient 212 .384 .384 .23 .9D−1 versus number of equalizer taps.332)     .22]  1  = .9   (3.U = 1 − 1 = 2.40.Figure 3. EXAMPLE 3.9  .294 .676 .991 (3.51 .384   1  =  . 15 taps are sufficient for infinite-length performance.7.22   0 = 1 − [−.334) The SNR is computed to be SNRM M SE−LE.1 For the earlier complex QAM example.9dB worse than the infinite length MMSE-LE for this channel.294 (3.

7 dB) .7188 −0. 3.714 (5.7313 − 0. the infinite-length performance level can be attained.0948 =  −0.343) which is 4.339) SNR   1.338) =   −.0409  .7188 −0.6250  1. choose Nf = 4 and note that ν = 2.337)  0     0  0   −.3 dB below MFB performance and 2. P =  0 0 −.340) = The FIR MMSE is 0. as long as w and b are sufficiently long.344) . The MSE for the DFE case. 0 and RY Y = 1 ¯ Ex P P ∗ + I (3.7188   (3.6250 + 0.  w∗ = R−1Y RY x Y Then 2 σM M SE−LE = (1 − wRY x ) = .2121 (3.341) (3.3578 for both ∆ = 2.to represent all signals.1 dB worse than the infinite-length MMSE-LE for this channel.5 1 + /4 −/2 0 0   . so choosing ∆ = 2 will not unnecessarily increase system delay.5 1 + /4 −/2 (3.5  1 + /4   (3.1376 + 0. (3.2570 − 0.336) The matrix (P ∗ P + .7.U = 1 − 1 = 3.0422  0. First.2982 −0.6250 − 0.6250 − 0.6250 − 0. Then   −.6250 /4 0  −0.6250 1.5 1 + /4 −/2 0  0 0 0 −.15625 · I)−1 has the same smallest element 1.7188 −0. With sufficiently large number of taps for this channel. 3.1182 − 0. (3.6250 1.342) The SNR is computed to be SNRM M SE−LE.2121 .6250  0 −/4 −0.5  . With a choice of ∆ = 2.6250 /4   .4 FIR MMSE-DFE The FIR DFE case is similar to the feed-forward equalizers just discussed. .6250 + 0.  −/4 −0. is M SE = E |xk−∆ − wY k + bxk−∆−1|2 213 . then   0  0     1  ¯ RY x = Ex P   (3.5 1 + /4 −/2 0 0 0  0 −. except that we now augment the fractionally spaced feed-forward section with a symbol-spaced feedback section.6250 + 0.

7. again. is determined using the following two definitions: Definition 3.0 ∆ .351) where J∆ is an (Nf + ν) × Nb matrix of 0’s and 1’s.3 (FIR MMSE-DFE Autocorrelation and Cross-Correlation Matrices) The FIR MMSE-DFE autocorrelation matrix is defined as R˜ ˜ YY = = = ∆ ˜ ˜∗ E Y k Y k /N RY Y E[Y k x∗ k−∆−1 ] ∗ ¯ E[xk−∆−1Y k ] Ex INb l·R nn ¯ ¯ Ex P P ∗ + Ex P J∆ ] SNR ¯ ∗ ¯ Ex J∆ P ∗ Ex INb (3.(3. or no zeros to the right or below exactly fitting in the bottom of J∆ (when Nf + ν − ∆ − 1 = Nb ).346) The solution.where b is the vector of coefficients for the feedback FIR filter b = [ b1 b2 . (3.345) and xk−∆−1 is the vector of data symbols in the feedback path. ∆ ˜ w = w .299) .347) . and of length Nf and Nb has coefficients ˜ w=R ˜ xY R−1 ˜ ˜ YY . The corresponding FIR MMSE-DFE cross-correlation vector is R˜ Yx = = = ∆ ˜ k−∆ /N E Y k x∗ RY x 0 ¯x P 1∆ E 0 (3. It is mathematically convenient to define an augmented response vector for the DFE as . (3.301).355) can be rewritten in detail as . l P P ∗ + SNR Rnn ∗ J∆ P ∗ P J∆ INb . (3.352) (3. ¯ = Ex · 1∗ P ∗ . (3. zeros below (when Nf + ν − ∆ − 1 > Nb ).353) . and a corresponding augmented DFE input vector as ˜ ∆ Yk= Then. The FIR MMSE-DFE for sampling rate l/T .. (3. Nf + ν − ∆ − 1) with zeros to the right (when Nf + ν − ∆ − 1 < Nb ).358) ..355) Equation (3. −b . ¯ w.349) (3. paralleling (3.348) Yk xk−∆−1 . (3. − b · Ex · .350) . . 214 (3.344) becomes ˜ ˜ M SE = E |xk−∆ − wY k |2 .357) (3. which has the upper ∆ + 1 rows zeroed and an identity matrix of dimension min(Nb . ∆ (3. bNb ] . N = 1 for real signals and N = 2 for complex signals.356) which reduces to the pair of equations w PP∗ + l ∗ Rnn − bJ∆ P ∗ = 1∗ P ∗ ∆ SNR wP J∆ − b = 0 . delay ∆. (3.354) where. (3.

1  0 1  .991 −1 − 0 1 −1 [0 1] (3. (3.367) SNRM M SE−DF E.1556 .1556 .9 1.361) Then ∗ b = 1∗ P ∗ P P ∗ − P J∆ J∆ P ∗ + ∆ l Rnn SNR P J∆ .U EXAMPLE 3. the SNR for the unbiased FIR MMSE-DFE is ¯ ˜ wR ˜ Ex Yx SNRM M SE−DF E.7668 . Then P =  J∆ =   1∆ and RY Y = Then w =  .9 1 0 0 .9] 1. .9 . (3.9 1  0 0  .Then b = wP J∆ .991 .9 0.9 1.7. and Nb = 1 and ν = 1.9 .364) l Rnn SNR −1 P 1∆ .363) (3. (3.991 .377) 215 [. and thus ∗ w P P ∗ − P J∆ J∆ P ∗ + (3.368) (3. (3.991  1.369) (3.373) (3.7668] .371) =  .991 .366) ˜ σM M SE−DF E Ex − wR ˜ Yx and the loss is again SNRM F B γM M SE−DF E = . Nf = 2. (3.362) The MMSE is then 2 σM M SE−DF E ¯ ˜ = Ex − wR ˜ Yx ¯x − wR = E Yx ¯ = Ex ∗ 1 − 1 ∗ P ∗ P P ∗ − P J∆ J∆ P ∗ + ∆ (3.375) (3. PAM and QAM) For the earlier example with l = 1.376) (3.9  0 1 [1 .359) = 1∗ P ∗ ∆ −1 l Rnn SNR (3.991 . (3.372) = = b = = = (3.370) 1. Thus.7668] wP J∆ [. (3.9 .374) (3.2 (MMSE-DFEs. we will also choose ∆ = 1.U = 2 −1= ¯ .9 1   0  0  1 .9 0 [0 1 0]  1 .9 1 0 0 . 0 .365) which is a function to be minimized over ∆.360) or ∗ w = 1 ∗ P ∗ P P ∗ − P J∆ J∆ P ∗ + ∆ l Rnn SNR −1 .

7 feed-forward and 1 feedback taps are necessary for infinite-length performance. In this case.157 . This performance is plotted versus the number of feed-forward taps (only one feedback tap is necessary for infinite-length performance) in Figure 3. (3.7188 −0. For the QAM example l = 1. (3.4 (7.5 1 + /4 −/2  0 0 0 0   .6250 −0. we also choose ∆ = 1.Figure 3.6250 − 0. 0  0 .1 dB.U = 1 − .5 1 + /4 −/2 0 0 −.157 = 5.16 . the infinite-length performance level can be attained.41: FIR MMSE-DFE Example. but 2.9 = . the finite-length DFE not only outperforms the finite or infinite-length LE. Then P =  J∆ =     1∆ =    −. .41. With sufficiently large number of taps for this channel.381) (3.6250 − 0. Actually this channel will need more taps to do well with the DFE structure.76] 1 .378) The SNR is computed to be SNRM M SE−DF E.8 dB better than the infinite-length MMSELE for this channel! The loss with respect to the infinite-length MMSE-DFE is 1.379) which is 2.42. A picture of the FIR MMSE-DFE is shown in Figure 3. (3. Nf = 2.7 dB below MFB performance.157 (3.382) and RY Y = 1. 1 0  0 1  0 1   .3dB) . it uses less taps (less complexity) also. Nb = 2.6250 1.380) (3. Then 2 σM M SE−DF E = 1 − [. Thus. and ν = 2.383) 216 . but we can still choose these values and proceed.7188 .

388) (3. The SNR is computed to be SNRM M SE−DF E.6726 − . FIR ZF-DFE One obtains the FIR ZF-DFE by letting the SNR→ ∞ in the FIR MMSE-DFE.7188 − 1/4 −1/8 − /2 −1/8 + /2 1.23dB .7188 −0. Yx (3.1917 .5  · [0 1 0 0]   /2 1 − /4  0 /2 1.1917 (3.394) ˜ However.395) and (3.5103 + .4720 − .3776 4.3608] . 2 making the SNR at the FIR ZF-DFE output SNRZF −DF E = ¯ ˜ Ex − wR ˜ Yx ¯ Ex + ¯ Ex .384) −1 = = b = = = Then 1. not w because only the feed-forward section filters noise.389) (3. and not very good on this channel! The reader may evaluate various filter lengths and delays to find a best use of 3.5 1 + /4 −/2 0 0 −.396) SNR ˜ The filter is w in (3.387) (3.5103 − . and 6 parameters on this channel.394) still ignores the enhanced noise at the wk filter output.386) (3.6250 − 0.1917 = 6. (3.6250 −0. 4. which alters R ˜ ˜ to YY ¯ ¯ ˜ ˜∗ E PJ E PP R ˜ ˜ = ¯ x ∗ ∗ ¯x ∆ (3.3894 .3776 1.77 dB below MFB performance.6250 − 0. (3.395) ·l · w 2 .396).6136] wP J∆ [1 − /2 − .5 0  1 − /4 −.4720 − .1180 .397) SNRZF −DF E 217 .1180 − .5 1 + /4 −/2 (3.3125  (3.2277 1. .391) which is 3. The loss is: SNRM F B γZF-DFE = . (3. the FIR ZF-DFE may still have nonzero residual ISI.392) YY Ex J∆ P Ex · INb ˜ and then w remains as R−1 ˜ .6136] −.390)  0  0   1 0  0 0   0  1 [−.5] [. 2 σM M SE−DF E = . l w 2 (3.6250 1. The power of this noise is easily found to be N0 FIR ZFDFE noise variance = (3.U = 1 − .4366 [. 5.393) ˜ YY Because the FIR equalizer may not be sufficiently long to cancel all ISI.Then w =  −. This ISI power is given by ˜ w=R ˜ xY 2 ¯ ˜ σM M SE−DF E = Ex − wR ˜ .385) (3.

the cross-correlation between the error and the vector of channel outputs Y k should be zero to minimize MSE. and b is now slightly altered to be monic and causal k:k−Nb k k−N b = [1 b1 b2 .400) 218 . (3.7.401) 14 A more perfect analogy between finite-length and infinite-length DFE’s occurs in Chapter 5. the DFE error signal is then described by ek = bX k (∆) − wY k . ¯ The SNR equals Ex / N0 = Ex /N0 in either case. This subsection provides such an alternative caused by the finite-length filters. and the corresponding vector of filter coefficients is w.42: FIR MMSE-DFE performance for 1 + . This is because finite-length filters inherently correspond to non-stationary processing. bNb ] . it yields less insight into the internal structure of the DFE than did the infinite-length structures investigated in Section 3. Continuing in the same fashion..399) The mean-square of this error is to be minimized over b and w.. x∗ b . where x∗ = x∗ .398) (3. particularly the spectral (canonical) factorization into causal and anti-causal ISI components is not explicit. 3..9D−1 versus number of feed-forward equalizer taps. where the best finite-length DFE’s are actually periodic over a packet period corresponding to the length of the feed-forward filter. ∆ (3.14 The vector of inputs at time k to the feed-forward filter is again Y k . (3. so the feed-forward filter output is zk = wY k . k So..6.5 An Alternative Approach to the DFE While the above approach directly computes the settings for the FIR DFE. although it is important to remember to divide any complex signal’s variance by 2 to get the energy per real dimension. Signals will again be presumed to be complex.Figure 3. E [ek Y ∗ ] = 0 .398). but all developments here simplify to the one-dimensional real case directly. 2 Optimizing the feed-forward and feedback filters For any fixed b in (3. wRY Y = bRXY (∆) .

Nf+nu)] Jdelta = 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 or for Nf = 8. the autocorrelation matrix corresponding to the X Y MMSE vector in estimating X(∆) from Y k . ∆ = 3.406) (3. which is executed with Cholesky factorization for finite matrices..size) eye(size) zeros(size. Nb = 2.size) eye(size) zeros(size. >> size=min(Delta. I −1P J∆ (∆)J∆ b nn SNR (3. and zeros in any row entries below that identity. That is R⊥ /Y (∆) is the autocorrelation matrix for the error sequence of length N − b that is associated with a X “matrix” MMSE-LE..and RXY (∆) = = E X k (∆) x∗ x∗ . then leaving Q 219 . and ν = 5.403) ˜∗ where the matrix J∆ has the first ∆ columns all zeros. then an (up to) Nb × Nb identity matrix at the top of the up to Nb columns.217). The and Nb become infinite. except for the “annoying” J∆ matrices that become identities as Nf ˜ as essentially the autocorrelation of the channel output. Nb = 5. ∆ = 3. and ν = 5. By defining ˜ Q(∆) = ˜∗ J∆ P ∗ P + l I SNR −1 −1 ˜ J∆ .402) (3.0).216). The following matlab commands produce J∆ . zeros(max(Nb-Delta. >> >> size=min(Delta.Nb).0).Nb) size = 2 >> Jdelta=[ zeros(size.Nf+nu)] Jdelta = 0 0 0 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 The MSE for this fixed value of b then becomes σ2(∆) = = = = ¯ b∗ Ex I − RXY (∆)R−1Y RY X (∆) b Y b∗ R⊥ /Y (∆)b X ¯ ˜∗ b∗ Ex INb − J∆ P ∗ P P ∗ + l· N0 2 ˜∗ b∗ J∆ l ˜ Rnn −1P J∆ (∆) b SNR l ˜ ˜ P ∗ R−1 P + . The solution then requires factorization of the inner matrix into canonical factors. max(Nf+nu-2*size.407) ∆ ¯ and R⊥ /Y (∆) = Ex I − RXY (∆)R−1Y RY X (∆).405) (3. >> Jdelta=[ zeros(size. using for instance Nf = 8. (3.0)).0)) zeros(max(Nb-Delta. max(Nf+nu-2*size. x∗ f −ν+1 P ∗ k k−1 k−N ¯ ˜∗ Ex J∆ P ∗ .408) ˜ this matrix is equivalent to (3.404) (3. This expression is the equivalent of (3. (3. and possibly zeroed ˜∗ columns following the identity if Nf + ν − 1 > Nb .

In this figure. However. 2 thus increasing its variance by (SNRM M SE−DF E /SNRM M SE−DF E. G(D) (3. but a strongly correlated (and enhanced) noise.416) The MMSE. whence the name “noise-predictive” DFE. The filter U (D) becomes the MMSE-LE so that Z(D) has no ISI. Then σ2 (∆) = So (∆) . then −∗ it can be shown also (see Problem 3. all SNR’s. (3. The SNR is as always ¯ Ex SNRM M SE−DF E. An analogous situation occurs for the finite-length case. If W (D) = γ0 · p ·G1 (D−∗ ) of the normal MMSE-DFE. Finite-length Noise-Predictive DFE Problem 3. ∆ ∆ which is minimized when b = g(∆) .˜ ˜ J∆ factors. but this requires the DFE filters to become periodic in a period equal to the number of taps of the feed-forward filter (plus an excess bandwidth factor). then any number of taps (greater than ν can be investigated without loss of generality with this early notational abuse. (3. Chapter 5 will find a situation that is an exact match and for which the feed-forward filter is an inverse of a canonical factor.U = −1 . cannot be ignored in the finite-length case.413) P∗ matched filter . Canonical factorization of Q proceeds according to (3. The MMSE is thus obtained by computing Cholesky ˜ factorizations of Q for all reasonable values of ∆ and then setting b = g(∆). The filter essentially tries to predict the noise in the feedforward filter output and then cancel this noise. the error sequence is feedback instead of decisions. ANY positive number of taps in u can 220 .415) S0 (∆) Bias can be removed by scaling the decision-element input by the ratio of SNRM M SE−DF E /SNRM M SE−DF E.412) (3. which is repeated here in Figure 3. Clearly by abusing Nf (which is not the number of taps). 2 The feed-forward filter then becomes w = bRXY (∆)R−1 YY ˜ = g(∆)J∆ P ∗ P P ∗ + ˜ = g(∆)J∆ P ∗ P + l I SNR −1 −1 (3. and biasing/unbiasing remain the same. any value of ∆ works and also ¯ S0 → γ0 N0 /Ex to ensure the infinite-length MMSE-DFE solution of Section 3.U .414) l I SNR feedforward filter which can be interpreted as a matched filter followed by a feed-forward filter that becomes 1/G∗(D−∗ ) as its length goes to infinity. (3.8 will produce that the filter B(D) remains equal to the same G(D) found for the infinite-length MMSE-DFE. (3. Clearly if ν is fixed as always. as the lengths of filters go to infinity.8) that U (D) = W (D) . The preditor then reduces the noise to a white error sequence. Correct solution to the infinite-length MMSE filter problem 3.409) σ2 (∆) = bG−1S∆ G−∗ b∗ . and it will be convenient notationally to say that the number of taps in the feedforward filter u is (Nf − ν)l.410) the top row of the upper-triangular matrix G∆ .6.411) From previous developments. however.U ) .43. the result that the feed-forward filter is an anti-causal factor of the canonical factorization does not follow for finite length.8 introduces the “noise-predictive” form of the DFE.

. . zk−ν   Zk =  Then. . it follows directly then from Equation (3..422) for U when b and w are known...401) that u = RXY (∆)R−1Y ... . solving for the b and w of the conventional finite-length MMSE-DFE can be followed by the step of solving Equation (3. u       (3. Thus. the error signal can be written as ek = bk X k (∆) − bZ (3. (3.. (3. However. u is the finite-length MMSE-LE with (Nf − ν)l taps.424) Y so following the infinite-case form. .423) which is the same error as in the alternate viewpoint of the finite-length DFE earlier in this section. 0  Yk= By defining the (Nf )l-tap filter   . .   .. (3.420)   U =  and . . .43: Noise-Predictive DFE be constructed without loss of generality for any value of ν ≥ 0. (3.419) 0 0 . .421) w = bU . yk .417) where  zk . 221 .422) (3..Figure 3. . 0  0 u . The only difference from the infinite-length case is the change in length. . .418) Zk = U Y k where  u 0 . yk−Nf +1 ∆ (3. In this case. . then the error becomes ek = bk X k (∆) − wY .

this vector is simply [σ2 0. • SNR = equalizer SNR. unbiased in dB % w_t = equalizer coefficients [w -b] % opt_delay = optimal delay found if delay =-1 option used.w_t.Ex. delay.p.nff.0]. oversampled at l (size) • nff = number of feed-forward taps • nbb = number of feedback taps • delay ≈ delay of system Nff + length of p . % --------------------------------------------------------------size = length(p).2 .6 The Stanford DFE Program A matlab program has been created. noise). unbiased and in dB • wt = equalizer coefficients The student may use this program to a substantial advantage in avoiding tedious matrix calculations. Difficult transmission channels may require large numbers of taps and considerable experimentation to find the best settings.noise). % otherwise. wt) = dfe ( l.Ex. % l = oversampling factor % p = pulse response. 222 . and refined by the students of EE379A over a period of many years.nbb.delay.(nu+1)*l-size)]. where the inputs and outputs are listed as • l = oversampling factor • p = pulse response. then choose the best delay • Ex = average energy of signals • noise = autocorrelation vector (size l*nff) (NOTE: noise is assumed to be stationary). used.p.nff. p = [p zeros(1.nbb % if delay = -1.nbb.delay. returns delay value passed to function % created 4/96. nu = ceil(size/l)-1. oversampled at l (size) % nff = number of feed-forward taps % nbb = number of feedback taps % delay = delay of system <= nff+length of p .opt_delay]=dfsecolorsnr(l. Ex. nbb..noise).. if delay = -1.w_t] = dfecolor(l. The program has come to be used throughout the industry to compute/project equalizer performance (setting nbb=0 also provides a linear equalizer).3. then choose best delay % Ex = average energy of signals % noise = noise autocorrelation vector (size l*nff) % NOTE: noise is assumed to be stationary % % outputs: % dfseSNR = equalizer SNR.p. For white noise.nff. function [dfseSNR.nbb . The program has the following call command: function (SNR.7. % -------------------------------------------------------------% [dfseSNR. The reader is cautioned against the use of a number of rules of thumb (like “DFE SNR is the SNR(f) at the middle of the band”) used by various who call themselves experts and often over-estimate the DFE performance using such formulas.

%save setting of this delay if best performance so far if new_dfseSNR >= dfseSNR w_t = new_w_t.1).nff+nu). dfseSNR = new_dfseSNR.1) = [p(1).nff+nu).constant for all delays Rn = zeros(nff*l+nbb). %loop over all possible delays for d = delay.d+2:d+1+nbb) = eye(nbb). Rn(1:nff*l. end %form matrix P. Rxy = Ex*temp*P’. zeros(l-1.. p_nu] where p_i=[p(i*l) p(i*l-1). P(nff*l+1:nff*l+nbb.. end end 223 .. %P temp= zeros(1. vector channel matrix P = zeros(nff*l+nbb. for i=1:nu ptmp(1:l.i:(i+nu)) = ptmp.1)]. temp(d+1)=1.real(new_w_t*Rxy’). new_dfseSNR = 10*log10(Ex/sigma_dfse .% error check if nff<=0 error(’number of feed-forward taps > 0’). opt_delay = d. new_w_t = Rxy*inv(Ry). end %form ptmp = [p_0 p_1 . sigma_dfse = Ex . elseif delay == -1 delay = 0:1:nff+nu-1-nbb. dfseSNR = -100. P = P_init. p((i-1)*l+1) ptmp(1:l. for i=1:nff. P(((i-1)*l+1):(i*l)..i+1) = conj((p(i*l+1:-1:(i-1)*l+2))’).1:nff*l) = toeplitz(noise). %construct matrices Ry = Ex*P*P’ + Rn. end %precompute Rn matrix . P_init = P. elseif delay < -1 error(’delay must be >= -1’). end if delay > (nff+nu-1-nbb) error(’delay must be <= (nff+(length of p)-2-nbb)’).

There is. At low error rates. but this requires that the channel be known in the transmitter – which is not always possible. ek . there are only 2M − 1 possible values for k .e.7.3. in the DFE output can is uncorrelated with the symbol of interest. there is a famous “Peron-Frobenious” lemma from linear algebra that states that there is a unique eigenvector ρ of all nonnegative entries that satisfies the equation ρ = Υρ . The Precoders of Section 3. the inner channel (DFE) may have an error rate that is unacceptably high that is later reduced in a decoder for the applied code. however. can be described by a finite-state machine with (2M − 1)Nb states. Equivalently. (3. The ith entry. ρi .425) where k = xk − xk .44. the evolution of an error event. this is reasonable. At high error rates of 10−3 and above. The probability that k takes on a specific value.epo. especially in transmission systems with significant channel variation. although enormous amounts of computational power may be necessary (much more than that required to simulate the DFE and measure error rate increase).ef. this chapter has ignored the effect of decision errors in the feedback section of the DFE. The last signal is often written Nb ef.8 can be used to eliminate the error propagation problem. feedback errors . filtered noise .426) Given k−1. For a matrix of non-negative entries. k .k = 4. (3.en. is the steady-state probability of being in state i.428) The solution ρ is called the stationary state distribution or Markov distribution for the state transition diagram. this error signal can be decomposed into 4 constituent signals 1. say 10−5 or below. postcursor ISI . it is common for low decision-error rates to occur in a coded DFE system.427) There are 2M − 1 such values for each of the (2M − 1)Nb states. is (denoting the corresponding new entry in k as k ) P k / k−1 =Q dmin − |ef. An analysis for the infinite-length equalizer has not yet been invented. The column sums are thus all unity. thus the analysis here applies only to FIR DFE’s with finite Nb .k = −1 i=−∞ vi xk−i ∞ i=Nb +1 vi xk−i ∞ i=−∞ wink−i (xk−i − xk−i) ˆ Nb i=1 vi 2 The sum of the first 3 signals constitute the MMSE with power σM M SE−DF E .k = 3.. Such a finite state machine is shown in Figure 3. each corresponding to one of the (2M − 1)Nb possible length-Nb error events. given the DFE is in state j. is k =[ ∆ k−Nb +1 k−Nb +2 . and assuming that any error signal. i. In coded systems (see Chapters 7 and 8). or that the state transition diagram goes to the corresponding state at time k. xk . j th element is the probability of entering state i.k = i=1 ∆ vi · k−i (3.k = 2.k ( k−1)| 2σMMSE-DFE . Thus. Of course. digital mobile telephony. By again denoting the equalized pulse response as vk (after removal of any bias). which can be computed according to previous results. 224 . an accurate way to compute the effect of decision feedback errors. precursor ISI . The error event vector. error propagation can lead to several dB of loss.7 Error Propagation in the DFE To this point. This last distortion component has a discrete distribution with (2M − 1)Nb possible ˆ points. Υ denotes a square (2M − 1)Nb × (2M − 1)Nb matrix of transition probabilities where the i.. k] (3.epr.

would yield a yet more accurate estimate of the error probability for the equalized channel . The number of states can be very large for reasonable values of M and Nb . finding the stationary distribution and adding over states in E. 225 . leading to a new state transition diagram with M Nb states.however. [101] . of the type (for Nb = 3) [000] . By denoting the set of states for which Pe = i∈E k = 0 as E. . Thus.. which we now denote ε = [ε1 . Thus.Figure 3. (3. the relatively large magnitude of the error propagation samples usually makes their explicit consideration more important than the (usually) much smaller residual ISI. and there is essentially no difference between k and − k . At a loss in Pe accuracy. so that the analysis may merge the two corresponding states and only consider one of the error vectors. so that the calculation of the stationary distribution ρ could exceed the computation required for a direct measurement of SNR with a DFE simulation..430) This reduces the number of states to 2Nb . There is essentially no difference in this Pe estimate with respect to the one estimated using all (2M − 1)Nb states..429) of error as ρi . the sum of the stationary probabilities corresponding 15 There is no reduction for zero entries. we must try to find an upper bound on these probabilities that depends only on the previous states. and compute error statistics for binary error event vectors. This can be done for almost half15 the states in the state transition diagram. one determines the probability (3. the number of states can remain unacceptably large. εNb ]. Further analysis then proceeds as above. A larger state-transition diagram. There are a number of methods that can be used to reduce the number of states in the finite-state machine. [010] .44: Finite State Machine for Nb = 2 and M = 2 in evaluating DFE Error Propagation (2M −1)Nb i=1 ρi = 1. however. [100] . but unfortunately the state transition probabilities no longer depend only on the previous state. the DFE will have an accurate estimate of the error probability in the presence of error propagation. we may ignore error magnitudes and signs. In so doing. most of which will reduce the accuracy of the probability of error estimate. the constellation for xk is symmetric with respect to the origin. corresponding to the explicit consideration of residual ISI as a discrete probability mass function. In the usual case.

with the desired state-dependent (only) transition probabilities. i.k(i) (3.5 for i.k(i) Pε1.k=1/εk−1 . Pek /εk−1 ≤ Q  2 (3.k (i)is max Pε1.]. so they can all be considered as nearest neighbors.431) ≤ i | ε1.ε1. while state i = 2 corresponds to [0 0 ε . etc. by grouping the 2Nb states into groups that are classified only by the number of leading in εk−1. then this last (easily computed) bound gives Pe ≤ 100 ! 226 .437) and δmax (Nb ) = 0.433) occurs by noting that this error probability corresponds to a worst-case signal offset of Nb δmax (ε) = (M − 1)d i=1 |vi |ε1. As long as this quantity is less than the minimum distance between constellation points.d. state i = 1 corresponds to any state of the form [0 ε].k (i) .433) = ε1.434) which the reader will note is analogous to peak distortion (the distortion is understood to be the worst of the two QAM dimensions..ε1. the corresponding error probability is then upper bounded as   d min − δmax (ε)  .i.to states in E will also upper bound the probability of error. The upperbound on probability of error for transitions into any state. the upper bound for Pe with error propagation is   d min − δmax (ε)  Pε .435) 2 σM M SE−DF E Now. Explicit computation of the maximum probability in (3. E corresponds to those states with a nonzero entry in the first position (the “odd” states. Finally..k=1/εk−1 = {i | ε1. a trivial bound that corresponds to noting that after an error is made.k(i) Pε1. Pe(errorprop) ≤ M Nb Pe (first) .436) Q 2 2 σM M SE−DF E ε Even in this case.k (i)is Pε1. input data may be lower than this bound even for reasonable values of M and Nb ). Thus adding these to the original probability of the first error.438) It should be obvious that this bound gives useful results only if Nb is small (that is a probability of error bound of .k=1/εk−1 . A further reduction to Nb + 1 states is possible.432) (3. thus. The probability of occurrence of these error event vectors is each no greater than the initial error probability.ε1.k (i) (3. (3. which are assumed to be rotated so that dmin lies along either or both of the dimensions ). To get an upperbound on each of the transition probabilities we write Pε1.k−i . That is suppose.k (i) allowed transition from εk−1 max Pε1. (3. the number of states 2Nb can be too large. then uses a δmax (i) given by Nb δmax (i) = (M − 1)d i=i+1 |vi| . the first probability of error is 10−5.k(i) } allowed transition from εk−1 ε1. (3. Pe ≤ (3. we have M Nb − 1 possible following error event vectors that can correspond to error propagation (only the all zeros error event vector corresponds to no additional errors within the time-span of the feedback path).k =1/εk−1 . ε1 = 1). and M = 8 and Nb = 8.

Look-ahead leads to the Maximum Likelihood Sequence detection (MLSD) methods of Chapter 9 that are no longer symbol-by-symbol based.k−∆|. This method is called “look-ahead” decoding basically ˆ because all possible previous decisions’ ISI are precomputed and stored.7. The vector is of dimension ν and can be viewed as an address Ak−∆−1 to the memory. The possible output for each of the M ν is computed and subtracted from zU.k−∆. Lookahead methods can never exceed SNRM M SE−DF E.U in terms of performance.U . then the largest ν < ν taps can be used (or typically the most recent ν and the rest subtracted in typical DFE fashion for whatever the decisions previous to the ν tap interval in a linear filter. If M ν calculations (or memory locations) is too complex. namely smallest |EU. (MLSD methods can exceed SNRM M SE−DF E. M ν possible decision vectors are retained. in some sense looking ahead in the calculation.45 illustrates a “look-ahead” mechanism for reducing error propagation.45: Look-Ahead mitigation of error propagation in DFEs.Figure 3. M ν decisions of the symbol-by-symbol detector can then be computed and compared in terms of the distance from a potential symbol value. 3.8 Look-Ahead Figure 3. The smallest such distance is used to select the decision for xk−∆ . but can come very close since error propagation can be very small.) 227 . Instead of using the decision.

is a device used to eliminate error propagation and is shown in Figure 3.46. The Tomlinson Precoder (TPC). 3.46(b) illustrates a generalization for complex signals.1 The Tomlinson Precoder Error propagation in the DFE can be a major concern in practical application of this receiver structure. the two real sums and two one-dimensional modulo operators can 228 . The second approach or partial response channels (which have trivial precoders) has no transmit power penalty. are used in concatenation with the DFE (because the error rate on the inner DFE is lower (worse) prior to the decoder). 3.8. but with no reduction in DFE SNR. Figure 3. or convolutional codes (see Chapter 10).Figure 3. which essentially moves the feedback section of the DFE to the transmitter with a minimal (but nonzero) transmit-power-increase penalty. while Figure 3. This preset value (usually with all integer coefficients) leads to simplification of the ZF-DFE structure.46(a) illustrates the case for real one-dimensional signals. but may have an SNR loss in the DFE because the feedback section is fixed to a desirable preset value for B(D) rather than optimized value. especially if constellation-expanding codes. more recently known as a Tomlinson-Harashima Precoder.8 Precoding This section discusses solutions to the error-propagation problem of DFE’s. In the second complex case. The first is precoding. Error propagation is the result of an incorrect decision in the feedback section of the DFE that produces additional errors that would not have occurred if that first decision had been correct.46: The Tomlinson Precoder.

46). such that x + Md 2 ΓM (x) = x − M d (3. That is ∆ x ⊕M y = ΓM [x + y] (3. straightforward moving of the filter 1/B(D) to the transmitter could result in significant transmit-power increase.8. The following lemma notes that the modulo operation distributes over addition: Lemma 3. The basic idea is to move the DFE feedback section to the transmitter where decision errors are impossible.439) Md where y means the largest integer that is less than or equal to y.47 illustrates modulo arithmetic for M = 4 PAM signals with d = 2. √ For complex QAM. defined on an M -ary (PAM or QAM square) input constellation with uniform spacing d.1 (Distribution of ΓM (x) over addition) The modulo operator can be distributed over a sum in the following manner: ΓM [x + y] = ΓM [x − y] = ΓM (x) ⊕M ΓM (y) ΓM (x) M ΓM (y) . The Tomlinson precoder appears in the transmitter as a preprocessor to the modulator. However.Figure 3.443) 229 . This text denotes modulo M addition and subtraction by ⊕M and M respectively. ∆ (3.441) Figure 3.440) and x M y = ΓM [x − y] .442) (3. ΓM (x) need not be an integer.8. a hexagon.47: Illustration of modulo arithmetic operator. which is in turn applied to the modulator (not shown in Figure 3. be generalized to a two-dimensional modulo where arithmetic is modulo a two-dimensional region that tesselates two-dimensional space (for example. or a square). (3. The Tomlinson Precoder maps the data symbol xk into another data symbol xk . To prevent most of this power increase. modulo arithmetic is employed to bound the value of xk : Definition 3.1 (Modulo Operator) The modulo operator ΓM (x) is a nonlinear function. each dimension is treated modulo M independently.

d.k] = = = = = ΓM [ΓM (xk − i=1 ∞ gU.. ] = SNRM M SE−DF E.447) ∞ ∞ ΓM [zU. The distribution and autocorrelation properties of the TPC. As stated here without proof.i. sequence. There is no explicit proof of this theorem for finite M . (3. is also approximately i.The proof is trivial.k . the scaled feedforward filter output with the Tomlinson precoder is ∞ zU. remain an unsolved problem at present.i.48. Thus. 230 .ixk−i) + i=1 ∞ gU. .. M d/2). because the SNR must also be infinite.452) ΓM [xk − i=1 gU.249) with x (D) as the new input.k being larger than M d/2 in magnitude is small in a well-designed system. which is known in the estimation literature to be i.k] is determined as xk + i=1 gU. Theorem 3.k] xk + eU.444) where xk = ΓM [˜k ] = ΓM xk − x ∞ bi xk−i i=1 .449) (3. which is a special case of the theory here.k = ΓM [zU.. part of the theorem appears to be valid for almost any M .k . although it can be proved exactly as M → ∞. the Tomlinson Precoder has allowed reproduction of the input sequence at the (scaled) MS-WMF output. the modulo element is not really necessary. That is E SNRM M SE−DF E zk /[xk xk−1 . B(D) = GU (D).k ] xk ⊕M ΓM [eU. one can assume that the error sequence is of the same distribution and correlation properties after the modulo operation.450) (3.d.U ∞ gU.8. in closed form.ixk−i + eU. Then.ixk−i + eU.i.d. and the sequence xk can be shown to be equivalent to a prediction-error or “innovations” sequence..451) (3. i=0 (3. with GU (D) = Pc(D).1 (Tomlinson Precoder Output) The Tomlinson Precoder output. This proof notes that the unbiased and biased receivers are identical as M → ∞.448) (3.k] (3. and futhermore the output sequence is approximately uniform in distribution over the interval [−M d/2.ixk−i + eU. The i. The receiver corresponding to Tomlinson precoding is shown in Figure 3. there is only a small price to pay in increased transmitter power when the TPC is used. Since the probability that the error eU. without ISI. when the input is an i. The original Tomlinson work was done for the ZF-DFE.445) The scaled-by-SNRM M SE−DF E /SNRM M SE−DF E. From Equation (3.ixk−i + i=1 gU.d. The noise power at the feed-forward output is thus almost exactly the same as that of the corresponding MMSE-DFE and with no error progagation because there is no longer any need for the feedback section of the DFE.i.k ] ΓM [xk + eU. The Tomlinson Precoder generates an internal signal ∞ xk = xk − ˜ i=1 bi xk−i (3.ixk−i .446) Thus.U output of the MS-WMF in the receiver is an optimal unbiased MMSE approximation to X(D) · GU (D). (3.

The equivalent circuit is also shown in Figure 3. 16 The maximum range of such an infinite extension is actually the sum ν i=1 | or ¯ gU. the feedforward filter output is ZU (D) = [X(D) + M (D)] · GU (D) + EU (D) = X(D) − λ(D) + EU (D) . the Tomlinson Precoder Power loss is usually ¯ larger.i | · |Ree (D) or xmax |. Laroia or Flexible Precoder Figure 3.454) M −1 for QAM.48: Receiver for Tomlinson precoded MMSE-DFE implementation. which is a variation on Tomlinson precoding introduced by Rajiv Laroia. to recover the input sequence X(D). The decision device finds the closest point in the infinite extension of the constellation16. the increase in transmit power for the TPC is from the 2 −1)d2 nominal value of (M 12 to the value for a continuous uniform random variable over the output interval 2 2 d [−M d/2. d/2). then M (D) is removed by a second truncation (which is not really a decision.453) Since M (D) is uniform and has magnitude always less than d/2. When the input is not a square constellation. The Laroia precoder largely preserves the shape of the transmitted constellation. which essentially is the decision. The number of nearest neighbors also increases to Ne = 2 for all constellations.49 shows the Laroia precoder. the receiver forms Z (D) = X(D) − λ(D) = X(D) + M (D) . These losses can be eliminated (and actually a gain is possible) using techniques called Trellis Precoding and/or shell mapping (See Section 10. M d/2).455) Processing ZU (D) by a decision operation leaves X(D)−λ(D). GU (D) (3. Using the uniform distribution assumption. (3. 231 . leading to an input power increase of M2 M2 − 1 for PAM and correspondingly M (3.6).49 where the input is considered to be the difference between the actual input symbol value xk and the “decision” output λk . Because of the combination of channel and feedforward equalizer filters.Figure 3. mainly to reduce the transmit-power loss of the Tomlinson precoder.456) (3. but operates using essentially the same logic as the first decision). However. mk is therefore a small error signal that is uniform in distribution over (−d/2. which is M12 . thus having variance d2 /12 << Ex. The extension of the constellation is the set of points that continue to be spaced by dmin from the points on the edges of the original constellation and along the same dimensions as the original constellation. but never more than a few dB.

Figure 3. 232 .49: Flexible Precoding.

its probability of occurence and then investigate the string (with (1/b)k being the impulse response of the post-first-decision filter in the receiver) | · (1/b)k | << dmin/2 . The decision device is binary in this case since binary antipodal transmission is used. Example and error propagation in flexible precoding EXAMPLE 3. Note the IIR filter in the receiver. this relation will hold and that is the maximum burst length possible for the particular type of error. the bit and symbol error rate in this case increase by a factor of 1. Given the filter b is monic and minimum phase.50. Thus.725)k · uk . (xk − λk ) − decision(xk − λk ) = · δk (3.85/6.725 in feedback section. this string should not be too long as long as the 233 .5.8. A minor point is that the first decision in the receiver has an increased nearest neighbor coefficient of Ne = 1.9D−1 system with flexible precoding) The flexible precoder and corresponding receiver for the 1 + . then the magnitude of the error is 2 = | + 1 − (−1)| = | − 1 − (+1)|. producing a contribution to the correct sequence of 2 · (−. Generally speaking.1 dB loss).633 to get .85 multiples . The bias removal factor has been absorbed into all filters (7. which has impulse response (−. (less than . If an error is made in the first SBS in the receiver.Figure 3. Such an error is like an impulse of magnitude 2 added to the input of the IIR filter. This produces an additional 1 error half the time and an additional 2 errors 1/4 the time.085 in feedforward section). and multiplies .457) where could be complex for QAM.1 (1 + .5. (3.9469 to get 1.9D−1 example appear in Figure 3. when an error is made in the receiver for the flexible precoder. Longer strings of errors will not occur.458) For some k. Then the designer needs to compute for each type of such error.725)k ·uk .50: Flexible Precoding Example.

Then.roots are not too close to the unit circle.” The study of PR channels is facilitated by mathematically presuming the tacit existence of a whitened matched filter. thus inevitably leading to non-Nyquist17 pulse shapes even if the phone line otherwise introduced no signficant intersymbol interference. In both cases. The signal output of the discrete-time channel is ∞ yk = m=−∞ hm · xk−m + nk . The single error may cause additional errors. will not pass through the “read channel. Straightforward equalization is often too expensive at the very high speeds of magnetic-disk recording systems. 2 this noise is exactly AWGN with mean-square sample value σpr .” again inevitably leading to ISI. The minimum-phase equivalent of Figure 3. hk is monic.2 (Partial-Response Channel) A partial-response (PR) channel is any discrete-time channel with input/output relation described in (3.8. and just denoted σpr otherwise..18 This chapter focuses on the discrete-time channel and presumes the WMF’s presence without explicitly showing or considering it.6.51(b) exists if Q(D) = η0 · H(D) · H ∗ (D−∗ ) is factorizable.e.k = p −2 · η0 · N0 · δk or Rnn(D) = ¯ 2 −1 N0 −2 2 p · η0 · 2 when the whitened-matched filter is used. H(D) has all ν roots on or outside the unit circle.2 Partial Response Channel Models Partial-response methods are a special case of precoding design where the ISI is forced to some known well-defined pattern. A unit-valued sample of the response occurs at time zero and the remaining non-zero response samples at subsequent sampling times. hk is minimum phase.459) or (3. DC). Partial-Response (PR) channels have non-Nyquist pulse responses – that is. Another example is magnetic recording (see Problem 3. = Pc (D) in Section 3. 234 . 2. exploiting the known nature of the ISI rather than attempting to eliminate this ISI. PR channels allow intersymbol interference – over a few (and finite number of) sampling periods. Definition 3. (3. some of which found very early use.459) or Y (D) = H(D) · X(D) + N (D) . Figure 3.C. hk = 0 ∀ k < 0 or k > ν. where ν < ∞ and is often called the constraint length of the partial-response channel. and thus D. which is conveniently abbreviated σ2 for the duration of this chapter. 17 That 18 H(D) is.See Chapter 3. so a receiver can use a detector design that instead presumes the presence of a known fixed ISI.460) −1 ¯ where nk is sampled Gaussian noise with autocorrelation rnn. as will be described shortly. the earliest use of telephone wires for data transmission inevitably found large intersymbol interference because of a transformer that was used to isolate DC currents at one end of a phone line from those at the other (see Prob 3.3 on ZF-DFE. h0 = 1. a number of common PR channels can be easily addressed. but none of these additional errors in turn cause yet more errors (unlike a DFE where second and subsequent errors can lead to some small probability of infinite-length bursts). The transformer did not pass low frequencies (i.460) that also satisfies the following properties: 1.36) where only flux changes on a recording surface can be sensed by a read-head. thus the name “partial response.35). they do not satisfy Nyquist’s condition for nonzero ISI .51(a) illustrates the whitened-matched-filter. 3. hk is finite-length and causal. Classic uses of partial response abound in data transmission.8. thus the probaility of an infinitely long burst is zero for the flexible precoder situation (or indeed any burst longer than the k that solves the above equation. For instance. (3. Equalization may be too complex or otherwise undesirable as a solution. Section 3. Receiver detectors are then designed for such partial-response channels directly. 3.

hk ∈ Z ∀ k = 0 (where Z denotes the set of integers). and hk is the minimum-phase equivalent of the sampled pulse response of that FIR filter. Thus. and if Pc(D) (the resulting discrete-time minimum-phase equivalent channel polynomial when it exists) is of finite degree ν. then the channel is a controlled intersymbol interference −1 channel with H(D) = Pc(D) and σ2 = N0 · p −2 · η0 . As noted in Section 3. a (usually small) transmit symbol energy increase occurs when the Tomlinson Precoder is used.Figure 3. The controlled ISI polynomial (D-transform)19 for the channel simplifies to ν H(D) = m=0 ∆ hm D m .1 could be used to implement symbol-bysymbol detection on the channel output. (3.6.8. the channel can be modeled as an FIR filter.4 shows that this loss can be avoided for the special class of polynomials H(D) that are partial-response channels. Any controlled intersymbol-interference channel 2 is in the form that the Tomlinson Precoder of Section 3. ν is the constraint length of the controlled ISI channel. If the receiver processes the channel output of an ISI-channel with the same whitened-matched filter that occurs in the ZF-DFE of Section 3.’ 4. and an all-zero (FIR) polynomial. a discrete finite-length channel satisfying all properties above except the last restriction to all integer coefficients is known as a controlled intersymbol-interference channel.461) where H(D) is always monic. for which hk = 0 ∀ k > ν.51: Minimum Phase Equivalent Channel.” rather than its “D-Transform” in partial-response theory. hk has all integer coefficients. The constraint length defines the span in time νT of the non-zero samples of this FIR channel model. Controlled ISI and PR channels are a special subcase of all ISI channels. More generally. 19 The D-Transform of FIR channels (ν < ∞) is often called the “channel polynomial. . Section 3. 235 .1. causal.8. effectively.8. This text uses these two terms interchangeably. minimum-phase.

but is not so easy to compute as the simple DFE. For illustration.Figure 3.. A MMSE-LE could be expected to have signficant noise enhancement while existing. which is easily ˆ computed as 3 dB for H(D) = 1 ± Dν (any finite ν) with simple ZF-DFE operation zk = yk − xk−ν and ˆ 6 dB for H(D) = 1 + D − D2 − D3 with simple ZF-DFE operation zk = yk − xk−1 + xk−2 + xk−3.8. This small subsection serves to simplify and review equalization structure on the partial-response channel. Tomlinson or Flexible precoding could be applied to eliminate error propagation for a small M2 increase in transmit power ( M 2 −1 ). . The loss with respect to the matched filter bound is trivially 1/ H 2. Since the PR channel is already discrete time.. The combination of integer coefficients and minimum-phase constraints of partial-response channels allows a very simple implementation of the ZF-DFE as in Figure 3.462) where l and n are nonnegative integers.3 Classes of Partial Response A particularly important and widely used class of partial response channels are those with H(D) given by l H(D) = (1 + D) (1 − D)n . Equalizers for the Partial-response channel. which all partial-response channels do. there is no feedforward section because the channel is already minimum phase). (3. no sampling is necessary and the minimum-phase characteristic of H(D) causes the WMF of the ZF-DFE to simplify to a combined transfer function of 1 (that is. then H(D) = 1 + D .52: Simple DFE for some partial-response channels.463) . ˆ ˆ ˆ The ZF-LE does not exist if the channel has a zero on the unit circle.52. which easily consists of the feedback coefficients −hm for the delayed decisions corresponding to xk−m .462). The ZF-DFE is then just the feedback section shown. m = 1. 3. Let l = 1 and n = 0 in (3. 236 (3. The MMSE-DFE will perform slightly better than the ZF-DFE. ν..

8D)(1 + . so the equalizer loss is 7 dB in this case. Q(D) + = 1. For 1 the MMSE-DFE.64 SNRMF B thus γMMSE-DFE = 2. SNRM F B N0 2 (3.U is easily computed to be 8 (9dB). If SNRM F B = 16 dB for this channel.05) (3.2 (Duobinary 1 + D) The duobinary channel (as we have already seen) has H(D) = 1 + D .464) has a notch at the Nyquist Frequency and is generally “lowpass” in shape.22 dω = .2dB. For the ZF-DFE.466) N0 T 2 2π π T π −T 1 N0 N0 1/2 √ = 2. which is sometimes called a “duobinary” channel (introduced by Lender in 1960). A discrete-time ZFE operating on this channel would produce infinite noise enhancement.467) 237 . (3.8D−1 ).025 (1 + .464) The transfer function in (3. There are several specific channels that are used in practice for partial-response detection: EXAMPLE 3. then Q(e−ωT ) + 1 = (1 + 1/40) + cos ωT . It is also possible to use a precoder with no transmit power increase to eliminate the error-propagation-prone feedback section of the ZF-DFE.64 = .8. η0 = 1 .025/1. as will be shown in Chapter 9.025 + cos ωT ) 2 1. p 2(1. and thus γZF-DFE = 3dB.465) The MMSE-LE will have performance (observe that 2 σM M SE−LE = = .Figure 3. The Fourier transform of the duobinary channel is H(e−ωT ) = H(D)|D=e−ωT = 1 + e−ωT = 2e−ωT /2 · cos ωT 2 .53.0252 − 12 2 The SNRM M SE−LE.53: Transfer Characteristic for Duobinary signaling. so that γ0 = 1. and 1. (3.625. 2 It is possible to achieve MFB performance on this channel with complexity far less than any equalizer studied earlier in this chapter. as is shown in Figure 3.

5 (Extended Partial Response 4 and 6 (1 + D)l (1 − D)n ) The EPR4 channel has l = 2 and n = 1 or H(D) = (1 + D)2 (1 − D) = 1 + D − D2 − D3 .The frequency response was already plotted in Figure 3..544 Mbps DS1 or “ANSI T1.3 (DC Notch 1 − D) The DC Notch channel has H(D) = 1 − D .462). so that l = 0 and n = 1 in (3. The frequency response is H(e−ωT ) = 1 − e−ωT = 2e−ω 2 · sin T (3. 2M − 2. These signals were once prevalent in telephone-company non-fiber central-office transmission of data between switch elements. When the 1 − D shaping is imposed in the modulator itself.4 (Modified Duobinary 1 − D2 ) The modified duobinary channel has H(D) = 1 − D2 = (1 + D)(1 − D) . ± (M − 1))..048 Mbps or “ITU-T G. rather than by a channel. In general. for M -level inputs (±1 ± 3 ± 5 . there are 2M − 1 possible output levels.471) The response goes to zero at the DC (ω = 0) and at the Nyquist frequency (ω = π/T ).. each independently acting on the inputs corresponding to even (odd) time samples.8. Many commercial disk drives use PR4. . Modified duobinary is equivalent to two interleaved 1 − D channels. there are 2M − 1 possible output levels. 238 ... In general. For a binary input of xk = ±1.. The frequency response is (3.462)..403”) and E1 (2. respectively.” although the use of the latter term is often confusing because bipolar transmission may have other meanings for some communications engineers. EXAMPLE 3. (3. . the channel output (with zero noise) takes on values ±2 with probability 1/4 each and 0 with probability 1/2.. The frequency response is H(e−ωT ) = 1 − e−ω2T = 2e−ωT · sin(ωT ) . the channel output (with zero noise) takes on values ±2 with probability 1/4 each and 0 with probability 1/2... for M -level inputs (±1 ± 3 ± 5 . the channel output (with zero noise) takes on values ±2 with probability 1/4 each and 0 with probability 1/2. In general for M -level inputs (±1 ± 3 ± 5 .469) The response goes to zero at the DC (ω = 0). EXAMPLE 3. the corresponding modulation is known as AMI (Alternate Mark Inversion) if a differential encoder is also used as shown later in this section..8... there are 2M − 1 possible output levels. ± (M − 1)). −2M + 2.. (3. 0. (3..468) ωT 2 . −2M + 2. thus modeling a highpass-like channel. and closely related methods are used in multiplexed T1 (1. AMI coding.473) H(e−ωT ) = (1 + e−ωT )2 (1 − e−ωT ) . (3.53. 0.. For a binary input of xk = ±1.472) This channel is called EPR4 because it has 4 non-zero samples (Thapar). These output values are all possible sums of pairs of input symbols. . .470) so l = n = 1 in (3. 2M − 2. For a binary input of xk = ±1.8. thus modeling a lowpass-like channel. AMI modulation prevents “charge” (DC) from accumulating and is sometimes also called “bipolar coding..703”) speed digital data transmission on twisted pairs or coaxial links. Modified Duobinary is sometimes also called “Partial Response Class IV” or PR4 or PRIV in the literature. This response goes to zero at the Nyquist Frequency. ± (M − 1)).. thus modeling a bandpass-like channel. EXAMPLE 3.

475) (3.. mk . along with PR4. the use of Tomlinson Precoding permits symbol-by-symbol detection. the more “lowpass” in nature that the EPR channel becomes. represents the difference (or sum) between adjacent differential encoder output (mk and mk−1) messages20 : ¯ ¯ 20 This operation is also very useful even on channels without ISI. DC-notch. and the more appropriate as bit density increases on any given disk. and modified duobinary partial-response channels is the so-called “differential encoder. as an unknown inversion in the channel (for instance. The response goes to zero at DC (ω = 0) and at the Nyquist frequency in both EPR4 and EPR6. is known simply as a “precoder. A simpler method for PR channels.54.. are often used to model disk storage channels in magnetic disk or tape recording.8. These are increasingly used in commercial disk storage read detectors. Thus. The frequency response is H(e−ωT ) = (1 + e−ωT )4 (1 − e−ωT ) . M − 1.” The message at time k. 1. 239 . The higher the l. is assumed to take on values m = 0. and changes its output if the input is 1 and repeats the last output if the input is 0. The magnitude of these two frequency characteristics are shown in Figure 3. The differential encoder for the case of M = 2 is most simply described as the device that observes the input bit stream.. an odd number of amplifiers) will cause all bits to be in error if (differential) precoding is not used.4 Simple Precoding The simplest form of precoding for the duobinary. thus modeling bandpass-like channels.Figure 3.54: Transfer Characteristic for EPR4 and EPR6 The EPR6 channel has l = 4 and n = 1 (6 nonzero samples) H(D) = (1 + D)4 (1 − D) = 1 + 3D + 2D2 − 2D3 − 3D4 − D5 .” 3. For partial-response channels. that also has no transmit energy penalty. mk . (3.474) These 2 channels. but also incurs an M 2/(M 2 − 1) signal energy loss for PAM (and M/(M − 1) for QAM). . the differential encoder input.

. 2 (3.3 1 0 2 1 0 3 mk 0 3 2 2 0 1 3 0 ¯ k -1 0 1 2 3 4 5 6 With either PAM or QAM constellations.479) 2 2 so21 yk ˜ + (M − 1) d = ¯ mk + mk−1 ¯ (3..55. the dimensions of the encoder output mk are converted ¯ into channel input symbols (± d . (5)4 = 1. the modulo addition and subtraction are performed independently on each of the two √ dimensions. with M replacing M .. and therefore ⊕ is different from the ⊕M of Section 3. ¯ where represents subtraction modulo M (and ⊕ represents addition modulo M ).55: Precoded Partial Response Channel. (A differential phase encoder is also often used for QAM and is discussed later in this section. Definition 3. augmented by the precoding operation of differential encoding as defined in (3.480) 21 (•) M means the quantity is computed in M -level arithmetic. As an example if M = 4 the corresponding inputs and outputs are given in the following table: mk . ± 3d .476) (3. for instance.476). Also note that ΓM (x) = (x)M .3 (Differential Encoder) Differential encoders for PAM or QAM modulation obey one of the two following relationships: mk ¯ mk ¯ = = mk mk−1 ¯ (3.) A differential encoder is shown on the left side of Figure 3.8.477) mk ⊕ mk−1 .478) Figure 3. M − 1 only.. The noiseless minimum-phase-equivalent channel output yk is ˜ d d yk = xk + xk−1 = 2(mk + mk−1) − 2(M − 1) ˜ ¯ ¯ (3.5. 240 . For SQ QAM.Figure 3. .55 illustrates the duobinary channel H(D) = 1 + D.) according to 2 2 xk = [2mk − (M − 1)] ¯ d .. The functions (x)M and ⊕ have strictly integer inputs and have possible outputs 0.

488) . For the 1 + D channel. The precoded partial-response system eliminates error propagation. which is 3 dB below the M F B = σpr (d/2) for the 1 + D −1 2 2 channel because p 2 = 2. the probability of a second error propagating is 1/2. This elimination of error propagation can be understood by investigating the nearest neighbor coefficient for the precoded situation in general. r < 1 may be useful in error propagation analysis. only 1 error occurs.. ˜ ˆ (3. The other half of the time.482) = M where the last relation (3. (1 − . (Again. (3. when the probability of a second error is (M − 1)/M .486) which will lead to a next-symbol error immediately following the first almost surely if xk = 1 (and no error almost surely if xk = −1). would produce a 4-level output with 1. (2M − 2)(d/2).484) k · (. while also effective.5)2 (3.55. All operations are integer mod-M . so that 2σpr = 2σpr .3 dB higher average transmit symbol energy for binary inputs. In practice. not yk .5)k = k=1 . M 2 1 . That is.5 .) The error propagation can be worse in multilevel PAM transmission.. The decision boundaries are simply the obvious ˜ regions symmetrically placed around each point for a memoryless ML detector of inputs and the decoding rules for M = 2 and M = 4 are shown in Figure 3. Then. where it becomes the precoder.483) (3. assuming a uniform channel-input distribution. (2M − 4) (2M − 2) with probabilities of occurrence M 2 M 2 ..) This 2 loss is identical to the loss in the ZF-DFE for this channel. a third error also occurs. 0 .482) shows that a decision on yk about mk can be made without concern for preceding or ˜ succeeding yk . if the previous decision was incorrect..485 becomes zk = xk + nk if the previous decision was correct on the previous symbol. then zk = xk − 2 + nk . say +1 was decided (binary case) instead of the correct −1. Half the times that 2 errors occur.482) follows from the definition in (3. M (3. ˜ (3..yk ˜ + (M − 1) d yk ˜ + (M − 1) d = M mk ⊕ mk−1 ¯ ¯ mk (3. leading to Ne (error prop) Ne (no error prop) = 1· 1 M −1 1 +2· · +3· M M M 241 M −1 M 2 · 1 + .481) (3.. = 2 4 8 16 ∞ ∞ .... the input to the decision device zk as ˜ zk = xk + xk−1 + nk − xk−1 . One may consider the feedback section of the ZF-DFE as having been pushed through the linear channel back to the transmitter. all the rest have 2 nearest neighbors. and thus has lower Pe than the ZF-DFE. the (noiseless) channel output 1 2 levels are −(2M − 2) − (2M − 4). Only the two outer-most levels have one M2 nearest neighbor. the decoder observes yk . Thus. For either case. . the minimum distance between ˜ √ √ d d 2 min outputs is dmin = d. because of the action of the precoder. With some algebra. 2 M2 − 2 1 ¯ Ne = 2 (1) + (2) = 2 1 − 2 2 M M M For the ZF-DFE. ˜ ˆ which can be rewritten zk = xk + (xk−1 − xk−1) + nk . The possibility of zk = xk + 2 + nk is just as likely to occur and follows ˜ an identical analysis with signs reversed.476). the decoder first quantizes yk to one of the values of (−2M + 2)(d/2). σpr = p −2 · η0 · N0 with a WMF and otherwise is just σpr . Equation (3.. M 2 M 2 . one can show that the TPC.. ˜ However.485) Equation 3.. so ˜ that yk must first be quantized to the closest noise-free level of yk .. and so on. The output levels are assumed equally likely in determining the decision boundaries (half way between the levels) even though these levels are not equally likely..487) r (In general.. effectively increasing the error coefficient from Ne = 1 to 1 1 1 1 Ne = 2 = 1( ) + 2( ) + 3( ) + 4( ) + . the formula k=1 k · rk = (1−r)2 .

yk ˜ 22 The = xk + xk−1 − xk−2 − xk−3 (3. the advantage of precoding is almost . The 1 − D2 case is identical to the 1 − D case. on two interleaved 1 − D channels at half the rate. ¯ Precoding eliminates this type of error propagation. which has l = 2 and n = 1 in (3. For M = 8.494) (3. In the 1 − D case. almost 1. Precoding EPR4 An example of precoding for the extended Partial Response class is EPR4. the advantage is about . For the 1 − D channel. ¯ ¯ and the decision rule is yk ˜ d (3. although Ne increases by a factor22 of (1 + 1/M ) with respect to the case where no error propagation occurred. For M ≥ 2.497) The combination of precoding with the 1 − D channel is often called “alternate mark inversion (AMI)” because each successive transmitted “1” bit value causes a nonzero channel output amplitude of polarity oppositie to the last nonzero channel output amplitude. Then. and the advantage becomes particularly pronounced for large M . Pe = 2(M − 1)Q( 2σpr ). and for M = 64.498) ratio of 2(1 − 1/M 2 ) with precoding to 2(1 − 1/M ) for M-ary PAM with no error propagation effects included 242 .2 dB for M = 4. Precoding can be simpler to implement than a ZF-DFE because the integer partial-response channel coefficients translate readily into easily realized finite-field operations in the precoder. while they represent full-precision add (and shift) operations in feedback section of the ZF-DFE. ¯ ¯ (3. while for the precoded partial-response system. the precoded system always has the same or fewer nearest neighbors.2 dB (which holds at reasonable error rates in the 10−5 to 10−6 range).2 dB. 1 d ¯ Pe = 2 1 − M 2 Q( 2σpr ). the channel output is yk ˜ yk ˜ d yk ˜ d yk ˜ d M = = = = xk − xk−1 = 2(mk − mk−1) ¯ ¯ ¯ mk − mk−1 ¯ mk ¯ mk .492) (3. d ¯ For the ZF-DFE system.6 dB.493) (3. especially by storagechannel engineers.472). or EPR4. Using a rule-of-thumb that a factor of 2 increase in nearest neighbors is equivalent to an SNR loss of .491) which is sometimes also called NRZI (non-return-to-zero inverted) precoding. the ¯ ¯ equivalent precoder is mk = mk ⊕ mk−1 .∞ = k=1 k· M −1 M k−1 · 1 M (3. while a “0” bit always produces a 0 level at the channel output. Precoding the DC Notch or Modified Duobinary Channels The mk = mk mk−1 differential encoder works for the 1 + D channel.489) (3. The overall precoder for this situation is mk = mk ⊕ mk−2 .496) = mk . M (3. as is the improvement over the ZF-DFE.495) M The minimum distance and number of nearest neighbors are otherwise identical to the 1 + D case just studied. mk−1 ¯ d 2 (3.490) = M .

this notation is symbolic. due to error propagation. 2 M4 − 2 1 ¯ Ne = 4 (1) + (2) = 2 1 − 4 M M4 M . is difficult to compute.499) (3.500) (3. M (3. from (3.504) 1 ¯ Thus the probability of error for the precoded system is Pe = 2 1 − M 4 Q( σ1 ). ¯ The corresponding memoryless decision at the channel output is mk = ˆ yk ˆ + d ν hi i=0 M −1 2 . (3. For the 1 + D − D2 − D3 channel. the (noiseless) channel output levels are −(4M − 4) − (4M − 6)...8.502) M where the precoder.506) The reader should be aware that while the relationship in (3.. ¯ ¯ ¯ (3. Nevertheless.505) The notation means a mod-M summation. engineers often simplify the representation of the precoder by simply writing it as 1 P(D) = (3. 0 .5 General Precoding The general partial response precoder can be extrapolated from previous results: Definition 3.503) The minimum distance at the channel output is still d in this case.505). all the rest have 2 nearest neighbors. Only the two outer-most levels have one nearest neighbor. so Pe ≤ Ne · Q(d/2σpr ).502) and (3.505) is general for partial-response channels. engineers commonly to refer to the NRZI precoder as a 1/(1 ⊕ D) precoder. one should recognize that the D means unit delay in a finite field.501) (3. Furthermore. The same 6dB loss that would occur with an error-propagation-free pr ZF-DFE on this channel.501). Of course. for instance the precoder for “EPR5. ¯ (3. 3. (4M − 6) (4M − 4). ¯ ¯ In a slight abuse of notation.4 (The Partial-Response Precoder) The partial-response precoder for a channel with partial-response polynomial H(D) is defined by ν mk = mk ¯ i=1 (−hi ) · mk−i . and the multiplication can be performed without ambiguity because the hi and mk−i are always integers. and is therefore a delay operator only – one cannot compute the Fourier transform by inserting D = e−ωT into P(D).” H(D) = (1 + D)3 (1 − D) simplifies to mk = mk ⊕ mk−4 when M = 2.8. which is 6dB higher. but the MFB = ( σd )2 . is mk = mk ¯ mk−1 ⊕ mk−2 ⊕ mk−3 .. Thus. but clearly will be worse.507) H(D) where P(D) is a polynomial in D that is used to describe the “modulo-M ” filtering in (3. 243 . The number of nearest pr neighbors for the ZF-DFE.yk ˜ yk ˜ d yk ˜ d yk ˜ d M = = = = d(mk + mk−1 − mk−2 − mk−3) ¯ ¯ ¯ ¯ ¯ ¯ ¯ mk + mk−1 − mk−2 − mk−3 ¯ mk ⊕ mk−1 ¯ ¯ mk mk−2 ¯ mk−3 ¯ (3. the relationship can often simplify in specific instances.

11. these two bits can be resolved without ambiguity even in the presence of unknown phase shifts of multiples of 90o . the sequence 01. (3.510). and 180o. and it could be decoded without error. To eliminate the ambiguity. the bit assignment for QAM constellations usually uses a precoder that uses the four possibilities of the most-significant bits that represent each symbol to specify a phase rotation of 0o .8. 10.508) Pe ≤ 2 1 − ν+1 Q M 2σpr where d is the minimum distance of the constellation that is output to the channel. however. The operations ⊕ and are the same in binary arithmetic.510) else if mk = (mi.k−1 ¯ mq. while an input of 2 maps to a change with respect to the last input of 1 or 2. 0) transmitted. This type of encoding is known as differential phase encoding and the remaining bits are assigned to points in large M QAM constellations so that they are the same for points that are just 90o rotations of one another. ¯ call it m90 . A precoding rule that will eliminate the ambiguity is then Rule 3. mq. the precoder leaves mk unchanged ¯ ¯ ¯ prior to differential encoding according to (3. No ambiguity exists for either the message (0. then the Pe bound in (3. To eliminate the ambiguity.1 (Complex Precoding for the 1+D Channel) if mk = (mi. check the last m = (0.k ⊕ mi. then there would be an ambiguity as to which part was real and which was imaginary. If m90 = 1.the precoder encodes the 1 and 2 signals into a difference between the last 1 (or 2) that was transmitted. By comparing adjacent decisions and their phase difference.56 illustrates the ambiguity: the two messages (0.k ) = (0. 3. 0) = 0 or the message (1. The nearest-neighbor coefficient and d/2σpr follow trivially from inspection of the output.) Thus the simple precoder for the 1 + D and 1 − D (or 1 − D2 ) given earlier really only is practical for PAM systems.508) holds.k ¯ mq.2 (Memoryless Decisions for the Partial-Response Precoder) The partialresponse precoder.8. 180o .6 Quadrature PR Differential encoding of the type specified earlier is not often used with QAM systems.509) (3. 0) or (1. for each dimension. Figure 3. 0) = 2 commute if the channel has an unknown phase shift of ±90o . 1) = 1 and (1.k ⊕ mq. individually. This precoder thus specifies that an input of 1 (or 2) maps to no change with respect to the last input of 1 (or 2). There is. 90o . Thus a 90o offset in carrier recovery would make the constellation appear exactly as the original constellation. because QAM constellations usually exhibit 90o symmetry.509) and (3. 244 . was m90 = 1 or 2 ?).k−1 ¯ (3. The previous differential encoder could be applied individually to both b dimensions of this channel. or 270o with respect to the last symbol transmitted.k ) = (0. the carrier recovery system locked on the wrong phase because of the symmetry in the output). 1) = 3.Lemma 3.8. 0o . 0). An upper bound on the performance of such symbolby-symbol decoding is d 1 ¯ . For instance. (that is.k . All previous analysis is correct. the sequence of subsequent phases 90o . Proof: The proof follows by simply inserting (3. like 8PSK. often abbreviated by P(D) = 1/H(D). 00 would produce (assuming an initial phase of 0o . and if all such adjacent levels are assumed to occur for an upper bound on ¯ probability of error. 1) or (1. 1) or (1.k ¯ = = mi. Adjacent levels can be no closer than d on a partialresponse channel. mq. 1) then mi.505) into the expression for yk and simpli˜ fying to cancel all terms with hi . enables symbol-by-symbol decoding on the partial-response channel H(D).k . (Similar methods could easily be derived using 3 or more bits for constellations with even greater symmetry. The following type of precoder is more practical for these channels: A Quadrature Partial Response (QPR) situation is specifically illustrated in Figure 3. QED. i > 0. one practical problem with this approach: If the channel were to somehow rotate the phase by ±90o (that is.56 for M = 4 or ¯ = 1 bit per dimension. 180o.

56: Quadrature Partial Response Example with 1 + D. 245 .Figure 3.

509) and (3. The corresponding decoding rule is (keeping a similar state y90 = 1. 2 at the decoder) ˆ  ˆ  (0. (3. which only resolves the 90o ambiguity. but is otherwise equivalent to a differential encoder.511) y ˆ ˆ ˆ  (0. then the precoder changes mk from 1 to 2 (or from 2 to 1) prior to encoding ¯ according to (3. 0) yk = [ ±2 ± 2 ]   (1. 0) (ˆk = [ 0 ± 2 ] or yk = [ ±2 0 ]) and yk − y90 = 0 y ˆ ˆ ˆ The probability of error and minimum distance are the same as was demonstrated earlier for this type of precoder.510). There will however be limited error propagation in that one detection error on the 1 or 2 message points leads to two decoded symbol errors on the decoder output.If m90 = 2. 1) (ˆk = [ 0 ± 2 ] or yk = [ ±2 0 ]) and yk − y90 = 0   (1. 1) yk = [ 0 0 ] ˆ mk = ˆ . 246 .

Thus. but possibly with different filtering and with noises that are at least partially independent. mainly because a greater channel-output minimum distance between possible (noiseless) output data symbols can be achieved with a larger number of channel output dimensions. the principles of Chapter 1 apply directly where the conditional probability distribution py/x of the channel has a larger dimensionality on the channel output y than on the input.3. this diversity leads to a lower probability of error for the same message.57: Basic diversity channel model. 3. . Often. (3.57. as illustrated in Figure 3. These channel outputs can be created intentionally by retransmission of the same data symbols at different times and/or (center) frequencies. Spatial diversity often occurs in wireless transmission where L spatially separated antennas may all receive the same transmitted signal. L − 1.. However.7. Channel outputs caused by the same channel input have labels yl (t).5 .. x. 3. Optimally.9 Diversity Equalization Diversity in transmission is the use of multiple channels from a single message source to several receivers. again can lead to a potentially complex optimum receiver and detector. l = 0.57 illustrates the basic diversity channel.9.512) 247 . The equalizers in the case of diversity become matrix equivalents of those studied in Sections 3. Ny > Nx . along with interference between the diversity dimensions. With each channel output following the model yp. intersymbol interference between successive transmissions.l (t) = k xk pl (t − kT ) + nl (t) .1 Multiple Received Signals and the RAKE Figure 3.Figure 3.. equalization again allows productive use of suboptimal SBS detectors.

A matched filter in general simply combines the signal components from all dimensions that have independent noise. independent on the diversity channels. however unfortunate for posterity. The RAKE is sometimes also called a diversity combiner. Mathematically.23 2 A single transmission of x0 corresponds to a vector signal yp (t) = x0 · p(t) + n(t) = x0 · p · ϕp (t) + n(t) . nL−1(t)    . By prefiltering the vector channel output by the 2 (f ). long delays in time. .9. the combination is known as the RAKE matched filter of Figure 3. the noise can be considered to be white on each of the L diversity channels.  (3. applies a matched filter only to the strongest of the L diversity paths to save complexity. a generalization of the inner product’s usual sum of products. . matrix filter Rn Anaylsis with the equivalent channel can then proceed as if the noise where white.513) where  ∆  yp (t) =    y0 (t) y1 (t) . often called maximal combining. pL−1(t)   ∆    and np (t) =    n0 (t) n1 (t) . The equivalent channel for this situation is then the channel corresponding to this maximum-strength individual path.a corresponding L × 1 vector channel description is yp (t) = k xk p(t − kT ) + np (t) .58. One structure.515) Without loss of generality. Here. but the matched filtering implied is practice.58: Definition 3. the noise will be whitened and the noise equivalent matrix channel becomes P (f ) → Rn (f )P (f ). where matched-filter demodulators there combined all time instants through integration. . the operation is denoted by L−1 yp (t) = l=0 p∗ (−t) ∗ yl (t) . When several copies are combined across several diversity channels or new dimensions (whether created in frequency. the noises may be correlated with each other on different subchannels and not white with covariance matrix 1/2 ∗/2 Rn (t) and power spectral density matrix Rn (f ) = N0 · Rn (f )Rn (f ). that generalization of combination simply needs to include also the components corresponding to each of the diversity channels so that all signal contributions are summed to create maximum signal-to-noise ratio.1 (RAKE matched filter) A RAKE matched filter is a set of parallel matched filters each operating on one of the diversity channels in a diversity transmission system that is followed by a summing device as shown in Figure 3. 2 −1/2 −1/2 23 In 248 . l (3.517) The RAKE was originally so named by Green and Price in 1958 because of the analogy of the various matched filters being the “fingers” of a garden rake and the sum corresponding to the collection of the fingers at the rake’s pole handle. y(t) >= i −∞ ∗ xi(t)yi (t)dt . and of the same variance N0 on all. or space). independent of the other diversity channels. yL−1(t)    ∆    p(t) =    p0 (t) p1 (t) . and with equal power spectral densities N0 .514) The generalization of an inner product is readily seen to be ∞ < x(t). The relative weighting of the different diversity channels is thus maintained through L unnormalized parallel matched filters each corresponding to one of the diversity channels. . (3. .516) This situation generalizes slightly that considered in Chapter 1. (3. although the latter term also applies to other lower-performance suboptimal combining methods that do not maximize overall signal to noise strength through matched filter. nomenclature thus often being left to the discretion of the inventor. The original RAKE concept was conceived in connection with a spread-spectrum transmission method that achieves diversity essentially in frequency (but more precisely in a codedivision dimension to be discussed in the appendix of Chapter 5).     (3. .

26) with Q(D) and p as defined in Subsection 3.1.522) and all other detector and receiver principles previously developed in this text now apply directly. 3. (3.9.9. (3.518) for an equivalent RAKE-output equivalent channel and the norm L−1 p 2 = l=0 pl 2 .58: The basic RAKE matched filter combiner.521) which is essentially the same as the early channel models used without diversity except for the additional scale factor of p 2 . (3.1 in Equation (3. although diversity combining is a more accurate name for the method. which also occurs in the noise autcorrelation. Some of those who later studied diversity combining were not aware of the connection to the RAKE and thus the multiple names for the same structure. the 249 . (3. Thus.520) · Q(D) + N (D) . which is N0 ¯ Rnn(D) = · p 2 ¯ p E An SNRM F B = xN0 2 2 2 · Q(D) . This text also defines rl (t) = pl (t) ∗ p∗ (−t) and l L−1 r(t) = l=0 rl (t) (3.2 Infinite-length MMSE Equalization Structures The sampled RAKE output can be scaled by the factor p −1 to obtain a model identical to that found earlier in Section 3. easily generalized.Figure 3. the normalized equivalent channel q(t) is defined through r(t) = p The sampled RAKE output has D-transform Y (D) = X(D) · p 2 2 · q(t) .519) Then.

Often while one diversity channel has severe characteristics.587 · (1 + . The loss with respect to the MFB is about 1. This channel is in effect anti-causal (or in reality.835D) · (1 + .492D + 1 + .9 comes first). ¯ E · p 2 The SNRM F B for this channel remains SNRM F B = xN0 .527) (3. (3.7082(20) − 1 = 13.9D) + (1 + .523) is .525) .164 per sample and is independent of the noise in the first channel.164 1.81) = 13 dB .492D−1 . which in this case is SNRM F B = Q(D) = = ˜ Q(D) = = = 1 . like an inband notch or poor transmission characteristic. Indeed the RAKE will work better than any of the individual channels.839D−1) (3.5 -3.64 = 2(1.8D−1 ) 2(1. MMSE-LE. and ZF-LE/DFE all follow exactly as in Sections 3.181 .835D−1) . The ISI is characterized as always by Q(D). EXAMPLE 3.MMSE-DFE. (3.8 dB in this case. so a pre-whitening matrix is 1 0 and so then the energy quantified by p p Then 2 2 0 .7.839D) · (1 + .16 (11. The ISI effectively spans 3 symbol periods among the two channels at a common receiver that will decide whether xk = ±1 has been transmitted.81) .492D + (1 + 1/20) + . Thus. 250 .U = . but it remains to compute 2 this quantity correctly.524) = p1 2 + p2 ˜ 2 = 1.59 illustrates two diversity channels with the same input and different intersymbol interference. First the noise needs to be whitened.181 Because of diversity. with each of the equalizer structures.839D ISI channel after adding the matched-filter outputs. diversity systems tend to be more robust.1 (Two ISI channels in parallel) Figure 3.9D−1)(1 + .528) (3. A second channel has causal response 1 + .81) .8D)(1 + . this channel has a higher potential performance than the single channel alone. While the two noises are independent.15 dB) .181 .81 + 1 · 2(1. The first upper channel has a sampled time equivalent of 1 + .9D−1 with noise variance per sample of .531) The improvement of diversity with respect to a single channel is about 2. The receiver is a MMSE-DFE essentially designed for the 1+.181 (and ¯ thus could be the channel consistently examined throughout this Chapter so Ex = 1).8 dB.530) Thus. the . they do not have the same variance per sample.492D−1 . the SNR of a MMSE-DFE would be SNRM M SE−DF E.164 . or any subset of the diversity channels.529) (3.9.8D with noise variance .526) (3. (3.7082 · (1 + . Clearly having a second look at the input through another channel can’t hurt (even if there is more ISI now).164 (3.181 (1 + . The matched-filter bound SNR and each of the equalization structures tend to work better with diversity because p 2 is typically larger on the equivalent channel created by the RAKE. a second channel is better.

7 then directly applies with the matrix P changing to include the larger l · L-tuples corresponding to L diversity channels. The astute reader will note that the diversity equalizer is the same in principle as a fractionally spaced equalizer except that the oversampling that creates diversity in the FSE generalizes to simply any type of additional dimensions per symbol in the diversity equalizer. . . so there are just L antennas or lines of samples per symbol period entry in that input vector. .532) as do the channel output vectors yk and the noise vector nk .. . pν . pν   xk xk−1 . . . . nk nk−1 . . (3.. . and a fractionally spaced equalizer prior to the summing device. p1 . each of the matched filters of the RAKE is replaced by a lowpass filter of wider bandwidth (usually l times wider as in Section 3.290)) again becomes   yk  yk−1    ∆ Yk =  (3. yk−Nf +1 p0 0 .Figure 3. Often.533)  . p0 0 pν .59: Two channel example. . l = 1. .7). a sampling device at rate l/T .288) now becoming l · L-tuples... . .60. .. and the corresponding equalizer W having its 1 × L coefficients corresponding to w0 ..9..   .. 0 p1 p0 . 0   0      . Each coefficient thus contains l values for each of the L equalizers. ... the channel input/output relationship (in Eq (3. p1 .. nk−Nf +1  =      +     .3 Finite-length Multidimensional Equalizers The case of finite-length diversity equalization becomes more complex because the matched filters are implemented within the (possibly fractionally spaced) equalizers associated with each of the diversity subchannels. 3. There may be thus many coefficients in such a diversity equalizer. .... . . . The dfecolor program can be used where the oversampling factor is simply l · L and the vector of the impulse response is appropriately organized to have l · L phases per entry. 251 .. .. . wNf . With vectors pk in Equation (3... . xk−Nf −ν+1         (3. 24 Maximal combiners that select only one (the best) of the diversity channels for equalization are popular because they reduce the equalization complexity by at least a factor of L – and perhaps more when the best subchannel needs less equalization. 0 0 .534) The rest of Section 3. 0      .24 In the case of finite-length equalizers shown in figure 3.. . pk = p(kT ) ..

Ex. but may be spatially correlated % outputs: % dfseSNR = equalizer SNR. It is listed here and is somewhat self-explanatory if the reader is already using the DFECOLOR program.delay.2 .Figure 3. unbiased in dB % -----------------------------siz = size(p.nbb % if delay = -1. oversampled at l (size).4 DFE RAKE Program A DFE RAKE program similar to the DFERAKE matlab program has again been written by the students (and instructor debugging!) over the past several years. Cioffi and above to get correct results -%March 2005 % function [dfseSNR. nu = ceil(siz/l)-1. % error check if nff<=0 252 . L=size(p. 3. edited by Olutsin OLATUNBOSUN % and Yun-Hsuan Sung to add colored and spatially correlated noise %Significant Corrections by J.noise). p = [p zeros(L.nff.1).W. % -----------------------------%**** only computes SNR **** % l = oversampling factor % L = No. then choose best delay % Ex = average energy of signals % noise = noise autocorrelation vector (size L x l*nff) % NOTE: noise is assumed to be stationary.9.(nu+1)*l-siz)]. each row corresponding to a diversity path % nff = number of feedforward taps for each RAKE finger % nbb = number of feedback taps % delay = delay of system <= nff+length of p .b]=dfsecolorsnr(l.nbb.2).60: Fractionally spaced RAKE MMSE-DFE. of fingers in RAKE % p = pulse response matrix.p. %DFE design program for RAKE receiver %Prepared by: Debarag Banerjee.

. temp(delay+1)=1. p((i-1)*l+1) for m=1:L ptmp((m-1)*l+1:m*l. end if delay > (nff+nu-1-nbb) error(’delay must be <= (nff+(length of p)-2-nbb)’). end %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %if length(noise) ~= L*l*nff % error(’Length of noise autocorrelation vector must be L*l*nff’). end %Add in part needed for the feedback P(nff*l*L+1:nff*l*L+nbb.delay+2:delay+1+nbb) = eye(nbb). vector channel matrix P = zeros(nff*l*L+nbb.2) ~= l*nff | size(noise. P(((k-1)*l*L+1):(k*l*L). zeros((l-1).k*l+1:-1:(k-1)*l+2))’). p_nu] where p_i=[p(i*l) p(i*l-1).. end %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %form ptmp = [p_0 p_1 . end end ptmp.1) ~= L error(’Size of noise autocorrelation matrix must be L x l*nff’). temp= zeros(1.nff+nu). %end if size(noise.1). %First construct the P matrix as in MMSE-LE for k=1:nff. (i-1)*l+(k-1)*L+1:i*l+(k-1)*L) = n_t(j:j+l-1. for i = 1:L n_t = toeplitz(noise(i.:)).k:k+l-1). % form matrix P. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Rn = zeros(nff*l*L+nbb).k:(k+nu)) = ptmp.k+1) = conj((p(m.nff+nu). for j = 1:l:l*nff for k = 1:l:l*nff Rn((i-1)*l+(j-1)*L+1:i*l+(j-1)*L. end for k=1:nu for m=1:L ptmp((m-1)*l+1:m*l.1)].. end if delay < -1 error(’delay must be >= 0’). end end 253 ..error(’number of feedforward taps must be > 0’).1) = [p(m.

0213 -0.7022 Two outputs are not quite the same because of the finite number of taps.0027 0.:)=reshape(ww((m-1)*l+1:m*l.1486 W = 0.0000 0 0 1.5) .1. IRy=inv(Ry).181 zeros(1. Different inputs may for instance occupy different frequency bands that may or may not overlap.164 zeros(1.0000 0 1.1.8400 >> [snr.5.0500 0.0028 b = 0.181 zeros(1.0382 0.W.0130 0.1465 W = 0. end b=-w_t(nff*l*L+1:nff*l*L+nbb).1.0439 -0.5) .5)]) snr = 11. sigma_dfse = Ex .w_t*Rxy’.0124 b = 0.0000 or with colored noise directly inserted >> p1 p1 = 0.l*L.p1.W.p.6.1).5 Multichannel Transmission Multichannel transmission in Chapters 4 and 5 is the logical extension of diversity when many inputs may share a transmission channel. .0249 -0. %Reshape the w_t matrix into the RAKE filter bank and feedback matrices ww=reshape(w_t(1:nff*l*L).b] = dfeRAKE(1.b] = dfeRAKE(1.9.5. -0.0984 0.0000 3. each of many inputs may affect each of many outputs.7022 0.0000 0.0401 0.5)]) snr = 11. Ry = Ex*P*P’ + Rn.1430 0. In the multichannel case.[.3545 0.1.1. or they may be transmitted from different antennas in a wireless system and so thus have different channels to some common receiver or set of receivers. dfseSNR = 10*log10(Ex/sigma_dfse . Rxy = Ex*temp*P’.4347 0. . for m=1:L W(m.0213 -0. Interference from other transmissions is more generally called crosstalk. Code division systems (See Chapter 5 Appendix) use different codes for the different 254 .[.1:nff).1430 0.9000 1.nff).0984 0.8000 0 >> [snr.4137 0. w_t = Rxy*IRy.end %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Ex*P*P’.9000 1.181 zeros(1.l*nff).0668 -0.0667 -0.0439 0.6.0237 -0. logically extending the concept of intersymbol interference to other types of overlap than simple ISI.3546 -0. some command strings that work are (with whitened-noise equivalent channel first): >> p p = 0. For the previous example.

Ideally. In effect the set of equalizers attempts to “diagonalize” the channels so that no input symbol from any source interferes with any other source at the output of each of the equalizers. the reader may note that a system that divides the transmission band into several different frequency bands may indeed benefit from a reduced need for equalization within each band. such a system can be simpler (avoiding equalization complexity) and actually perform better than equalized wider-band QAM systems. By allocating energy intelligently. Chapters 4 and 5 (EE379C) develop these concepts. From an equalizer perspective. Progress of understanding in the field of transmission should ultimately make equalization methods obsolete. when the transmit signals can be optimized also. this has become a line of confrontation between those resisting change and those with vision who see a better way. 255 . As in many technical fields. A diversity equalizer can be designed for some set of channel outputs for each and every of the input sequences. While this chapter has developed in depth the concept of equalization because there are many wide-band QAM and PAM systems that are in use and thus benefit from equalization. there can be considerable improvement in the performance of the set of equalizers. no equalizer is necessary. The SNR of each of these “sub-channels” relates (via the “gap” approximation) how many bits can be transferred with QAM on each. the situation is simply multiple instances of the diversity equalizer already discussed in this section. Chapters 4 and 5 support this intuition. an intuition that for the longest time transmission engineers might have been better off never needing the equalizer and simply transmitting in separate disjoint bands is well-founded. if each band is sufficiently narrow to be viewed as a “flat” channel. However.sources. leading to multichannel transmission. However.

Exercises - Chapter 3 3.1 Quadrature Modulation. Consider the quadrature modulator shown below,

Figure 3.61: Quadrature modulator. a. What conditions must be imposed on H1(f) and H2(f) if the output signal s(t) is to have no spectral components between −fc and fc ? Assume that fc is larger than the bandwidth of H1(f) and H2 (f). (2 pts.) b. Let the input signal be of the form,

x(t) =
n=−∞

anφ(t − nT )

What conditions must be imposed on H1(f) and H2(f) if the real part of the demodulated signal is to have no ISI. (3 pts.) c. Find the impulse responses h1 (t) and h2 (t) corresponding to the minimum bandwidth H1(f) and H2(f) which simultaneously satisfy (a) and (b). You can express your answer in terms of φ(t). (3 pts.) 3.2 Sampling time and Eye Patterns. The received signal for a binary transmission system is,

y(t) =
n=−∞

an q(t − nT ) + n(t)

t where an ∈ {−1, +1} and Q(f) is a triangular, i.e. q(t) = sinc2 ( T ). The received signal is sampled at t = kT + t0 , where k is an integer and t0 is the sampling phase, |t0 | < 1 T . 2

a. Neglecting the noise for the moment, find the peak distortion as a function of t0 . Hint: use Parseval’s relation. (3 pts.) b. Consider the following four binary sequences {un}∞ , {vn}∞ , {wn}∞ , {xn}∞ : −∞ −∞ −∞ −∞ un vn = = −1 +1 256 ∀n ∀n

wn xn

= =

+1, for n = 0 −1, otherwise −1, for n = 0 +1, otherwise

Using the result of (a), find expressions for the 4 outlines of the binary eye pattern. Sketch the eye pattern for these four sequences over two symbol periods −T ≤ t ≤ T . (2 pts.) c. Find the width (horizontal) of the eye pattern at its widest opening. (2 pts.) d. If the noise variance is σ2 , find a worst case bound (using peak distortion) on the probability of error as a function of t0 . (2 pts.) 3.3 The 379 channel model. Consider the ISI-model shown in Figure 3.2 with PAM and symbol period T. Let ϕ(t) = and h(t) = δ(t) − 1 δ(t − T ). 2 a. Determine p(t), the pulse response. (1 pt) b. Find ||p|| and ϕp (t). (2 pts.) c. (2 pts.) Find q(t), the function that characterizes how one symbol interferes with other symbols. Confirm that q(0) = 1 and that q(t) is hermitian (conjugate symmetric). You may find it useful to note that t + mT t + nT t + (m + n)T sinc ∗ sinc = T sinc . T T T d. Use Matlab to plot q(t). For this plot you may assume that T = 10 and plot q(t) for integer values of t. Assuming that we sample y(t) at the perfect instants (i.e. when t = kT for some integer k), how many symbols are distorted by a given symbol? Assume that the given symbol was positive, how will the distorted symbols be affected? Specifically, will they be increased or decreased? (2 pts.) e. For the rest of the problem consider 8 PAM with d=1. Determine the peak distortion, Dp . (2 pts.) f. Determine the MSE distortion, Dmse . Compare with the peak distortion. (2 pts.) g. Find an approximation to the probability of error (using the MSE distortion). You may express your answer in terms of σ.(1 pt.) 3.4 Bias and SNR. Continuing the setup of the previous problem with 8 PAM and d = 1, let 1 N0 = σ2 = 0.1. 2 Assume that a detector first scales the sampled output yk by α and chooses the closest x to yk · α as illustrated in Figure 3.13. Please express your SNR’s below both as a ratio of powers and in dB. a. For which value of α is the receiver unbiased? (1 pt.) b. For the value of α found in (a), find the receiver signal-to-noise ratio, SN RR . (2 pts.) c. Find the value of α that maximizes SN RR and the corresponding SN RR . (2 pts.) d. Show that the receiver found in the previous part is biased. (1 pt.) e. Find the matched filter bound on signal-to-noise ratio, SN RM F B . (1 pt.)
1 √ T t sinc( T )

257

f. Discuss the ordering of the three SN R’s you have found in this problem. Which inequalities will always be true? (1 pt.) 3.5 Raised cosine pulses with Matlab. For this problem you will need to get all the .m files from /usr/class/ee379a/hw5/ on leland. If you are logon to leland, you could do a cd and copy the files to your directory. Alternatively, you could do an anonymous ftp to ftp.stanford.edu, and cd to class/ee379a/hw5 and copy the files. a. Consider the formula for the raised cosine pulse (Eq. 3.79 of chapter 3): q(t) = sinc t T · cos 1−
απt T 2αt 2 T

There are three values of t that would cause a program, such as Matlab, difficulty because of a division by zero. Identify these trouble spots and evaluate what q(t) should be for these values. (2 pts) b. Once you have identified those three trouble spots, a function to generate raised cosine pulses is a straightforward implementation of Eq. 3.75. We have implemented this for you in mk rcpulse.m. Executing q = mk rcpulse(a) will generate a raised cosine pulse with α = a. The pulses generated by mk rcpulse assume T = 10 and are truncated to 751 points. (2 pts) Use mk rcpulse to generate raised cosine pulses with 0%, 50%, and 100% excess bandwith and plot them with plt pls lin. The syntax is plt pls lin(q 50, ’title’) where q 50 is the vector of pulse samples generated by mk rcpulse(0.5) (50% excess bandwidth). Turn in your three plots and discuss any deviations from the ideal expected frequency domain behavior. c. Use plt pls dB(q, ’title’) to plot the same three raised cosine pulses as in part (b). Turn in your plots. Which of these three pulses would you choose if you wanted the one which had the smallest band of frequencies with energy above -40dB? Explain this unexpected result. (3 pts) d. The function plt qk (q, ’title’) plots q(k) for sampling at the optimal time and for sampling that is off by 4 samples (i.e., q(kT ) and q(kT + 4) where T = 10). Use plt qk to plot q k for 0%, 50%, and 100% excess bandwidth raised cosine pulses. Discuss how excess bandwidth affects sensitivity of ISI performance to sampling at the correct instant. (3 pts) 3.6 Noise enhancement: MMSE-LE vs ZFE. Consider the channel with p 2 = 1 + aa∗ ∗ −1 Q(D) = a D + p p 2 0 ≤ |a| < 1.

2

+aD

a. (2 pts) Find the zero forcing and minimum mean square error linear equalizers WZF E (D) and WM M SE−LE (D). Use the variable b = p
2

1+

1 SN RMF B

in your expression for WM M SE−LE (D).

b. (6 pts) By substituting e−jwT = D (with T = 1) and taking SN RM F B = 10 p 2, use Matlab to plot (lots of samples of) W (ejw ) for both ZFE and MMSE-LE for a = .5 and a = .9. Discuss the differences between the plots. c. (3 pts) Find the roots r1 , r2 of the polynomial aD2 + bD + a∗ . Show that b2 − 4aa∗ is always a real positive number (for |a| = 1). Hint: Consider the case where 1 ∗ SN RMF B = 0. Let r2 be the root for which |r2 | < |r1 |. Show that r1 r2 = 1.

258

1. 3.6. 259 . (2 pts) Find B(D) and W (D) for the MMSE DFE. ¯ c.6. (4 pts) For Ex = 1 and σ2 = 0. We use a MMSE criterion to choose U (D) and B(D) to minimize E[|xk − zk |2]. 1 for the Ex and σ2 of problem 3. Assume all decisions are correct.5.) a. . γZF E .6. show that the canonical factorization is Q(D) + 1 ∗ = γ0 (1 − r2D−1 )(1 − r2 D). (2 pts) Use the previous results to show that for the MMSE-LE W (D) = p D p = a (D − r1 )(D − r2 ) a(r1 − r2 ) √ p . You have done much of the work for this in problem 3. w0 = ZFE w0 = p .8 Noise predictive DFE. Compare with your sketches from Problem 3. (4 pts) Give an expression for γM M SE−DF E . b2 −4aa∗ r2 r1 − D − r1 D − r2 1 SN RMF B . (3. (2 pts) Show that for the MMSE-LE. (4 pts) Find γZF E and γM M SE−LE in terms of the paramter a and calculate for a = 0. g.d.5.1 find expressions for σZF E .6. 1−aa∗ By taking = 0. Compute its values for a = 0. (2 pts) For the channel of problem 3. 0. b. and γM M SE−LE . show that for the 2 2 ¯ f.62: Noise-Predictive DFE Consider the equalizer shown above with B(D) restricted to be causal and monic. SN RM F B What is γ0 in terms of a and b? Please don’t do this from scratch. (xk is the channel input. a.535) e.7 DFE is even better.6. Sketch γZF E and γM M SE−LE for 0 ≤ a < 1. Q(D). p . σM M SE−LE . Ex . SNRMFB and γ0 . (5 pts) Show this equalizer is equivalent to a MMSE-DFE and find U (D) and B(D) in terms of ¯ G(D). Sketch γM M SE−DF E as in problem 3. Figure 3. 3.

(3 pts) Let x0 denote the time-zero value of the sequence X.e.) d. that is Q(ω) has vestigial symmetry. Figure 3. 2 Assume that the inputs are independent and identically distributed uniformly over the three possible values. show how to remove the bias. if the channel is flat) and therefore: γ0 ≤ 1 + 1 SNRMFB 1 = γ0 g SNRMFB 2 1 = γ0 G(D)G∗ (D−∗ ) SNRMFB c. Show that:   1   ≥ 1 1 γ0 Q(D) + SNR MFB 0 and therefore: SNRMMSE-LE. (2 pts) Is the NPDFE biased? If so. 260 . We have used the fact that the best unbiased receiver has a lower Pe than the (biased) MMSE receiver. (2 pts) Relate UNPDFE (D) to WMMSE-LE (D) and to WMMSE-DFE (D). Assume also that nk is zero mean and independent of xk . e. Here is a simple illustration of this fact. 3.U ≤ SNRMFB 3. (1 pt) Interpret the name “noise-predictive” DFE by explaining what the feedback section is doing. Consider the constellation above with additive white noise nk having N0 = σ2 = 0.1.U ≤ SNRDFE.U (with equality if and only if Q(D) = 1. (3 pts) Show that: g ≥1 with equality if and only if Q(D) = 1 ( i. a. (1 pt) If we remove feedback (B(D) = 1).63: A three point PAM constellation.b.U ≤ SNRMMSE-DFE.9 Receiver SNR relationships. (3 pts) Recall that: Q(D) + Show that: 1+ b. c. what does this equalizer become? d. (2 pts) Use the results of parts b) and c) to show: SNRMMSE-LE.10 Bias and probability of error.

(2 pts) Find E[e2 ] and Pe for the scaled output αyk . b. (2 pts) Find η0 and Pc (D) so that ∗ Q(D) = η0Pc (D)Pc (D−∗ ). You should find that the biased receiver of part c has a Pe which is higher than the (unbiased) ML detector of part b.a. Recall that you (hopefully) found √ b + b2 − 4aa∗ γ0 = . . 1 and find a so that Q(D) = a∗ D−1 + p p 2 2 + aD 261 .1j) T 1 √ (0. c. e.7. 2 c. Sketch the loss for 0 ≤ a < 1. a.12 Zero forcing DFE. a. (2 pts) Find the loss with respect to SN RM F B for a = 0. (2 pts) Find Gu in terms of a. 2(1 + aa∗ ) a. (1 pt) Where are the optimal decision boundaries for detecting xk from αyk (for the MMSE α) in part c? What is the probability of error for these decision boundaries? 3.13 Complex baseband channel.1. When you compute Pe . The pulse response of a channel is band-limited to p0 p1 pk where pk = p(kT ) and SNRM F B = 15 dB. 1. (1 pt) Find σZF −DF E d. (1 pt) Give a block diagram of the DFE with scaling before the feedback summation.5j) T 0 for k = 0. (2 pts) Find B(D) and W (D) for the ZF-DFE.6 and 3. Consider the DFE you designed in Problem 3. and r2 . (2 pts) Find p 2 π T and has = = = 1 √ (1 + 1. d. (1 pt) Give a block diagram of the DFE with scaling after the feedback summation.11 Bias and the DFE. (easy) b. Assume the same SN RM F B as in Prob 3. b. 3. (1 pt) Find Pe exactly for the ML detector on the unbiased values yk . b. even though it has a smaller squared error.7. ¯ Consider once again the general channel model of Problems 3. use the decision k regions for the unbiased ML detector you found in (b). c.95 + 0. (1 pt) Find the mean square error between yk = xk + nk and xk . k Prove that your α does in fact provide a minimum by taking the appropriate second derivative.7 with Ex = 1 and σ2 = 0. 3. (2 pts) Find the scale factor α that will minimize the mean square error E[e2 ] = E[(xk − αyk )2].5.

we will assume that σ2 = 0. Your system will be fairly simple.6 and 3. c. {3. 3}. and the output of your receiver.6 and 3. 3}. assume a 4-PAM constellation with xk ∈ {−3. Turn in hard copies of whatever code you write. and the output of your receiver. (5 points) Implement the precoder and its associated receiver using any computer method you would like (or you can do the following computations by hand if you prefer). We will be working with the 1 + 0. This problem considers the 1 + 0. Did the system still work? What changed? What purpose do the modulo operators serve? 262 . remove the modulo operations from your precoder and your receiver.7 wherever appropriate. −3. How would your design change if noise were present in the system? b. 1.15 Flexible Precoding.7. compute the output of the precoder. For this problem. Furthermore. 3.9D channel (i.9). −3.e. assume a 4-PAM constellation with xk ∈ {−3. (2 pts) Use the SN R’s from (b) to compute the NNUB on Pe for 4-QAM with the MMSE-LE and the MMSE-DFE. (2 pts) Find SN RM M SE−LE. Turn in a hard copy of whatever code you write. and the output of your receiver for the same inputs as in the previous part. −1} c. we will assume that σ2 = 0. (5 pts) Design a Flexible precoder and its associated receiver for this system.U and SN RM M SE−DF E. again with zero noise. Did the system still work? What changed? What purpose do the modulo operators serve? 3.9). −1. 3. For this problem.9D channel (i. −3. (3 pts) Now. −1. −1. the output of the channel. Turn in hard copies of whatever code you write. the output of the channel. and assume no noise. Assume that the symbol x = −3 was sent just prior to the input sequence below. {3. For the following input sequence compute the output of the precoder. remove the modulo operations from your precoder and your receiver. 1. the 1 + aD channel that we have been exploring at length with a = 0. Use the results of Problems 3. For the following input sequence compute the output of the precoder.14 Tomlinson Precoding. the output of the channel. the output of the channel. Turn in a hard copy of whatever code you write.b. (5 points) Implement the precoder and its associated receiver using any computer method you would like (or you can do the following computations by hand if you prefer). compute the output of the precoder. (3 pts) Now. Assume that the symbol x = −3 was sent just prior to the input sequence below. Furthermore. For this modified system. a. For this modified system.e. −3. (5 pts) Design a Tomlinson precoder and its associated receiver for this system. again with zero noise. and assume no noise. a. 3. Your system will be fairly simple.U for this channel. 1. and the output of your receiver for the same inputs as in the previous part. −1. How would your design change if noise were present in the system? b. −1} c. Note p 2 = 1 + aa∗ so some care should be exercised in using the results of Problems 3. 1. the 1 + aD channel that we have been exploring at length with a = 0.

263 . You may wish to check your answers with dfecolor. Compare with your value from part (a). assume that ∆ = 0. We design the optimal FIR MMSE-LE without assuming a matched filter. . (3 pts) Show that w = RxY R−1 = (0.17 Finite Length equalization and MATLAB.25].. 2 d.3. 3. However. Consider a system whose pulse response is band limited to |w| < T √ that is sampled at the symbol rate T after pass through an anti-alias filter with gain T . so just type mmsele and answer the questions. ˜ Also. (2 pts) Which terms correspond to the matched filter? Which terms correspond to the infinite length WM M SE−LE ? 3. you should go through the calculations yourself. you can type [1 0. (1 pt) Compute SN RM F B for this channel. (2 pts) Design an MMSE-DFE which has 2 feedforward taps and 1 feedback tap.U for the infinite length filter. This program is interactive.. (1 pt) Find the σM M SE−LE for the equalizer of the previous part.m. Find p(t) = (φ(t) ∗ h(t)) and p 2. for this problem. (1 pt) Find SN RM M SE−LE. ˜ ˜ corresponding to the discrete-time channel yk = xk − . Use mmsele to find the best ∆ and the associated SN RM M SE−LE.5δ(t − T ) l=1 Feel free to use matlab for any matrix manipulations as you complete the parts below.536) = k |˜k |2. 0)Φp∗ YY where p ˆ Q+l 1 I SN RM F B −1 . When it asks for the pulse response. . . (2 pts) We assume perfect anti-alias filtering with gain T . b. Again. T ].16 Finite length equalization and matched filtering. . a.5 · xk−1 + nk . it turns out that π we get a “matched filter” anyway. (2 pts) Design a 3 tap FIR ZF-LE for ∆ = 0.25] or p if you have defined p = [1 0. Discuss the plots briefly. (2 pts) In this problem we will use the MATLAB program mmsele. g. (You may use matlab for matrix inversion. 1. However. . ¯ Ex = 1 N0 1 = 2 8 1 t φ(t) = √ sinc( ) T T h(t) = δ(t) − 0. (1 pt) Find the associated σZF −LE . (2 pts) Design a 3 tap FIR MMSE-LE for ∆ = 0. p b. ¯ Consider the 1 + 0. PP∗ ˆ Q= p 2 b. (2 pts) Plot |P (e(jwT )| and |P (e(jwT )W (e(jwT ) | for w ∈ [0. Consider the following system description. 2 f.) √ a. 0. e.18 Computing finite-length equalizers. find P (D) 2 (3.U for a 5 tap linear equalizer .1 and Ex = 1. . 0.25D channel with σ2 = . a. How sensitive is performance to ∆ for this system? π c. c.

(3. |a| < 1 and both a and b are 0 real. choose the best ∆ where appropriate. (4 pts) d. what kind of equalizer should be used on this channel? c.001 Watt) and let 1/T = 2 MHz. In your engineering judgement. (2 pts) c. The target Pe is 10−3. (2 pts) b.538) where nk is white Gaussian noise.Final 1996 1 t An AWGN channel has pulse response p(t) = √T sinc T − sinc t−4T . 2 a. a.. For a ZF-DFE. Further 2 equalization of y(D) is in discrete time where (in this case) the matched filter and equalizer discrete responses (i. Compute your answer in part b with that of the infinite-length ZFE.e. ¯ E The SNR (ratio of mean square x to mean square n) is Nx = 20 dB.e.19. Let N0 = -100 dBm/Hz (0 dBm = 1 milliwatt = . −1/2T < f < 1/2T . (3 pts) b. and zero outside this band. Find the 2-tap FIR MMSE-LE. Sketch |P (e−ωT )|2. how would a MMSE-DFE perform on this channel? (2pts) 3. The filter is followed by a 1/T rate sampler so that the sampled output has D-Transform y(D) = (1 − D4 )x(D) + n(D) . For your transmit energy in part c.Final 1996 A symbol-spaced FIR equalizer (l = 1) is applied to a stationary sequence yk at the sampled output of an anti-alias filter.9 and b = 1. Compute the SNRzf e for part a. (2 pts) Compute the unbiased SNR’s for the MMSE-LE and MMSE-DFE.64: Channel for Problem 3.Figure 3. (1 pt) d. and n(D) is the D-Transform of the Gaussian noise sample sequence. find the transmit symbol mean-square value (i. the transmit energy) necessary to achieve a data rate of 6 Mbps using PAM and assuming a probability of symbol error equal to 10−3. Design a 2-tap FIR ZFE. (4 pts) 2 264 . Let a = . 3. Compare these two SNR’s with each other and the SN RM F B ..20 FIR Equalizer Design . For all parts of this question. The noise autocorrelation function is Rnn(D) = N0 . D-Transforms) can be combined into a single discrete-time response. (3.19 Equalizer Design . which produces a discrete-time IIR channel with response given by yk = a · yk−1 + b · xk + nk . h.537) x(D) is the D-Transform of the sequence of M -ary PAM input symbols. The receiver anti-alias T √ filter has gain T over the entire Nyquist frequency band.

u for part d. (3 pts) e. Design (show/draw) a Flexible precoder. and SNRU for the MMSE-DFE. (1 pt) c.82 D−1 + 1. ZFE: Compare Pe in part d with the Pe for the ZFE. Find a. The relative delay on this path is approximately T seconds. ( 2 pts) f. Remember that q0 = 1. Find W (D) and SNRU for the MMSE-LE. N0 2 ¯ = . Show a simple method and compute SNRzf −df e . (2 pts) 3. (2 pts) e. How does the SNR compare to the matchedfilter bound if we assume there is no information loss incurred in the symbol-spaced anti-alias filter? Use a = . Approximate Pe using the DM S of part c. ( 2 pts) d. (2 pts) b. use binary ¯ PAM with Ex = 1 and SNRM F B = 10 dB.e. (1 pt) b. Find SNRM F B and SNRM M SE−DF E . Find SNRM F B .25 − D SNRM F B 2 2 (3. Design (show/draw) a Tomlinson-Harashima precoder.9 and b = 1. f. showing from xk through the decision device in the receiver in your diagram (any M ).9eωT ) ∀|ω| < π/T studied repeatedly in this chapter.23 Finite-delay tree search A channel with multipath fading has one reflecting path with gain (voltage) 90% of the main path. 3. Assume binary transmission throught this problem. see example in this chapter for SNRZF E ) (2 pts) 3.21 ISI quantification . DM S . Find the SNRzf −df e for an infinite-length ZF-DFE. B(D). (1 pt) c. a. Find the mean-square distortion.Final 1995 For an ISI channel with Q(D) + 1   = . Find Pe .0181 and Ex = 1. 265 . Let M = 4 for your precoder in part c.22 Precoding . Find the peak-distortion bound on Pe . Find G(D) and GU (D) for the MMSE-DFE. showing from xk through the decision device in the receiver in your diagram (any M ). (1 pt) d. Find the SNRmmse−le. Find Pe . e.539) a. but the carrier sees a phase-shift of −60o that is constant on the second path. Let M = 4 for your precoder in part e. (3 pts) d. Use the model √ |ω| < π/T T 1 + ae−ωT P (ω) = 0 elsewhere to approximate this channel. (1 pt) b. (2 pts) c. (hint. Find the peak distortion. Find W (D).Midterm 1996 √ For the channel P (ω) = T (1 + . a. Compute SNR difference in dB between the SNR based on mean-square distortion implied in parts c and d and SNRZF E . Dp .

where N (D) is discrete AWGN.24 Peak Distortion -(5 pts) Peak Distortion can be generalized to channels without reciever matched filtering (that is ϕp (−t) is a lowpass anti-alias filter). find the filter transforms W (D) (and B(D)).Figure 3. 266 . What is SNRM F B ? b. e.25 More Equalization . Find this new Dp for this ¯ c. c. ˆ x ˆ x xk . f. Find the ZF-DFE. this is easy.5D) + N (D) after sampling on such a channel. (hint. is . the mean-square of the sample noise.Final 1994 √ ¯ Given an ISI channel with p(t) = 1/ T [sinc(t/T ) − sinc[(t − T )/T ]]. Draw a diagram illustrating the MMSE-DFE implementation with Tomlinson precoding. a. P (D) = 1 − .5D. f. Let us suppose y(D) = x(D) · (1 − . (2 pts) ∆ m=0 (3. How does this compare with the ZF-DFE (better or worse)? ˆ (1 pt) g. Find an SNRf dts for this channel.540) |pm |. 1 pt) b. Find the ZFE. for 2 each equalizer.ˆk−1 ˆ x where xk−2 = xk−2 by assumption.65: Finite-delay tree search. Suppose N0 . and the loss with respect to the MFB. What is Pe for symbol-by2 symbol detection with PAM and M = 2.05. Ex = 1. (2 pts) 3.05 and Ex = 1. A finite-delay tree search detector at time k decides xk−1 by choosing that xk−1 value that miniˆ ˆ mizes min |zk − xk − aˆk−1|2 + |zk−1 − xk−1 − aˆk−2|2 . Find the MMSE-DFE. Find the MMSE-LE. xk−1 and nk . d. the unbiased detection SNR. a. We define peak distortion for such a channel as Dp = |xmax| channel if |xmax | = 1. and N0 = . Could you generalize SNR in part g for FIR channels with ν taps? (bonus) 3. write yk in terms of xk . (bonus) h.

541) = 267 .(2 pts) e. Let the symbol rate be 40 MHz . 3. (1 pt) f. assuming a MMSE-DFE receiver is used. g. a. Draw a the combined shape of the matched-filter and feedforward filter for a ZF-DFE corresponding to the new symbol rate of part d.26 Raised Cosine pulses Show through inverse Fourier transforming that the raised cosine and square-root raised cosine timedomain responses given in Section 3. For Pe < 10−6 and using square or cross QAM.9D−1) 1 . For the equalizer in part b. choose a design and find the largest data rate you can transmit using one of the equalizers above.9D)(1 + . ( 2 pts) (Hint: you may note the relationship of this transfer function to the channel often used as an example in EE379A).10 pts.U for the new symbol rate. PAM transmission with sinc basis function is used on this channel with a transmit power level of 10 mW =Ex /T .28 Infinite-Length EQ .66: Channel for Problem 3. (1 + . Ex = 1.27. What can you conclude about incurring ISI if the transmitter is allowed to vary its bandwidth? 3.19 . and p 2 .27 Expert Understanding . what is the data rate for PAM if Pe ≤ 10−6? (1 pt) d.66. Final 1998 An ISI channel with PAM transmission has channel correlation function given by Q(D) = with SNRM F B =10 dB. For the same symbol rate as part a. A channel baseband transfer function for PAM transmission is shown in Figure 3. 3.3 are correct.19 . What is the new data rate for this system at the same probability of error as part c .compare with the data rate of part c.find SNRM F B . Estimate SNRM M SE−DF E. Let the symbol rate be 20 MHz .g. what is SNRM M SE−LE ? (1 pt) c.find the new SNRM F B . (1 pt) h.(1 pt) b. (3.Final 1998 Figure 3. The one-sided AWGN psd is -100 dBm/Hz.

01 for a symbol rate of 1/T = 100kHz is used with 2 PAM transmission.9). If a = 0. 2 1 1+a·eω 1 √ T · sinc t T with T = 1 0 |ω| ≤ π |ω| > π (3. (3 pts) Find WZF −DF E (D). 3. (3 pts) Find WM M SE−DF E (D). and • a signal that is not delayed. BM M SE−DF E (D). (2 pts) c. After the sampling.29 Finite-Length EQ .542) √ and finite-length equalization is used with anti-alias perfect LPR with gain T followed by symbol-rate sampling. Find the W (D) for the zero-forcing and MMSE linear equalizers on this channel.10 pts PAM transmission on a filtered AWGN channel uses basic function ϕ(t) = and undergoes channel impulse response with Fourier transform (|a| < 1) H(ω) = ¯ E and SNR = σx = 15 dB. a. the function characterizing ISI. 268 . b. BZF −DF E (D). Given the complexity constraint. (3 pts) d. Find the Fourier Transform of the pulse response.Final 2000 . but reduced to 90% of its amplitude (that is P (D) = . MMSE-DFE for the later diversity channel and compare to that on the original one-receiver channel and with respect to the matched filter bound.9 pts. P (ω)? (1 pt) b. Find the performance of the MMSE-LE. σZF E . 3.543) a. and SNRZF −DF E . Find p 2.U . Ex = 2 and N0 = .9 pts A filtered AWGN channel with T = 1 has pulse response p(t) = sinc(t) + sinc(t − 1) with SNRM F B = 14.2 a. Final 1998 A baseband channel is given by √ T · 1 − . Find the probability of bit error on the diversity channel for the case of 1 bit/dimension transmission.Final 2000 . Find Q(D). (1 pt) 3. (3 pts) b.32 Do we understand basic detection . (3 pts) Find WZF E (D). and SNRM M SE−DF E.7e2πf T P (f) = 0 |f| < |f| ≥ .6 pts We would like to evaluate the channel studied throughout this course and in the text with the same input energy and noise PSD versus almost the same channel except that two receivers independently receive: • an undistorted delayed-by-T signal (that is P (D) = D). and SNRZF E .5 T .Miterm 2000 . (2 pts) c. Find Q(D). (1 pt) Find WM M SE−LE (D) c. find the highest data rate achievable with PAM transmission when the corresponding probability of symbol error must be less than 10−5. (3 pts) e. d. p .30 Equalizers . and SNRM F B for the later diversity channel. a maximum complexity of 3 taps TOTAL over a feedforward filter and a feedback filter can be tolerated. what data rate is achievable on this channel according to the gap approximation at Pe = 10−6? (1 pt) 3.2 dB.31 Diversity Channel .5 T (3.

Find the probability of error for a ZF-DFE if a binary symbol is transmitted (2 pts). b. QAM transmission with 1/T = 1 MHz and carrier frequency fc = 600kHz are Ex used on this channel. SNRM F B . The desired probability of symbol error is 10−6. (2 pts) b. The Tomlinson precoder in this special case of a monic pulse response with all unity coefficients can be replaced by a simpler precoder whose output is xk = xk−1 −xk−1 if xk = 1 if xk = −1 (3. (2 pts) 3.8sinc(t − 1) + . AND Q(D) for this channel. and pulse repsonse 2 p(t) = sinc(t) + 1. Calculate the data rate using both an integer and a possibly non-integer number of bits/symbol. The transmit power is T =1 mW. Calculate the data rate using both an integer and a possibly non-integer number of bits/symbol. (3 pts) e. (2 pts) c. Implement this precoder for the largest number of integer bits that can be transmitted according to your answer in part (d).a. Determine a reasonable set of parameters and settings for an FIR MMSE-LE and the corresponding data rate at Pe = 10−7. (3 pts) 269 . (4 pts) c. A matlab subroutine at the course web site may be useful in computing responses in the frequency domain. a. Also compute the MMSE-DFE SNR and maximum data rate in bits/dimension using the gap approximation. ML detection is now used for this multidimensional 1-shot channel.33 Equalization . T = 1.34 Finite-Length Equalizer Design . which produce 8 possible 5-dimensional outputs. and an ideal filter is used for anti-alias filtering at the receiver. Suppose this channel is used one time for the transmission of one of the 8 4-dimensional messages that are defined by [+ ± ± ±].10 pts PAM transmission is used on a filtered AWGN channel with N0 = . Does the probability of error change significantly if the input is extended to arbitrarily long block length N with only the first sample fixed at +. and all other inputs being equally likely + or −? What happens to the number of bits per output dimension in this case? (2 pts). Find the ZF-DFE detection SNR and loss with respect to SNRM F B . Find the MMSE-DFE. Find p 2.5 dBm/Hz. (1 pt) d. What is the performance (probability of error) for this decoder? (2 pts) c.67.81sinc(t − 1). 3. Find the MMSE-LE and corresponding unbiased SNR for this channel. a.01. Find ν so that an FIR pulse response approximates this pulse response so that less than 5% error in p 2. Square-root raised cosine filtering with 10% excess bandwidth is applied as a transmit filter. b.Final 2000 . Calculate SNRM F B .544) Find the possible channel outputs and determine an SBS decoder rule. Repeat part b for an FIR MMSE-DFE design and draw receiver. The 2 oversampling factor for the equalizer design is chosen as l = 2. What are the new minimum distance and number of nearest neighbors and how do they relate to those for the system in part a? What is the new number of bits per output dimension? (3 pts) d.Final 2000 .15 pts 1 A filtered AWGN channel has impulse response h(t) = T · 1 1+ 107 t 3 2 and is used in the transmission system of Figure 3. The AWGN psd is N0 = −86. Draw the flexible (Laroia) precoder for the MMSE-DFE and draw the system from transmit symbol to detector in the receiver using this precoder.

voltages accidentally placed on the line from unintentionally harming the telephone line or equipment attached to it. Can you find a way to improve the data rate on this channel by changing the symbol rate and or carrier frequency? (2 pts) 3. What is a reasonable partial-response model for the channel H(D)? Sketch the channel transfer function. Figure 3.68: Illustration of a telephone-line data transmission for Problem 3. and they provide immunity to earth currents.Figure 3.68.C. For sufficiently short transmission lines. These transformers prevent large D. How would a zero-forcing decision-feedback equalizer generally perform on this channel with respect to the case where channel output energy was the same but there are no transformers? c.34. how much is transmit energy increased generally? Can you reduce this increase by good choice of initial condition for the Tomlinson Precoder? 270 .35 Telephone-Line Transmission with “T1” Digital transmission on telephone lines necessarily must pass through two “isolation transformers” as illustrated in Figure 3. The “derivative taking” combined characteristic of the transformers can be approximately modeled at a sampling rate of 1/T = 1. noises.35. Is there ISI? b. the rest of the line can be modeled as distortionless.67: Transmission system for Problem 3. d.544 MHz as successive differences between channel input symbols. Suppose a Tomlinson precoder were used on the channel with M = 2. a. These transformers also introduce ISI. What are some of the drawbacks of a ZF-DFE on this channel? d. and ground loops.

Compute the average energy of the channel input for the simple precoder with the input constellation of part (a). Pick a partial-resonse channel with ν = 2 that models the read-back channel. g. How many possible outputs are produced by the simple precoder with binary inputs? (1 pt) e.36 Magnetic Recording Channel Digital magnetic information storage (i. how much is transmit energy increased generally? Can you reduce this increase by good choice of initial condition for the Tomlinson Precoder? e. 3. How would a zero-forcing decision-feedback equalizer generally perform on this channel with respect to the best case where all read-head channel output energy convenyed either a positive or negative polarity for each pulse? c.38 Flexible precoding and simple precoding In Section 3. (4 pts) b. a. thus generating a read current that is translated into a voltage through amplifiers succeeding the read head. The read head has finite band limitations also. a.e. Design the Tomlinson precoder corresponding to a ZF-DFE for this channel with the possible binary inputs to the precoder being ±1. Characterize the loss in performance that you would expect to see for your detector.e. The disk spins under a “read head” and by Maxwell’s laws the read-head wire senses flus changes in the moving magnetic field. Design the Flexible precoder corresponding to a ZF-DFE for this channel with the possible binary inputs to the precoder being ±1.8. Show how a binary precoder and corresponding decoder can significantly simplify the implementation of a detector. Show how a binary precoder and corresponding decoder can significantly simplify the implementation of a detector. b.8.37 Tomlinson preocoding and simple precoding In Section 3. Suppose a Tomlinson precoder were used on the channel with M = 2. What new partial response might apply with the same read-channel electronics. How many distinct outputs are produced by the Tomlinson Precoder assuming an initial state (feedback D element contents) of zero for the precoder. Characterize the loss in performance that you would expect to see for your detector. What is the loss with respect to optimum MFB performance on this channel with your precoder and detector? f. (4 pts) 271 . the simple precoder for H(D) = 1 + D is derived. Suppose the channel were not exactly a PR channel as in part a. Change in flux is often encoded to mean a “1” was stored and no change means a “0” was stored. but with a correspondingly faster symbol rate? 3. What are some of the drawbacks of a ZF-DFE on this channel? d. but were relatively close. the simple precoder for H(D) = 1 + D is derived. a. Sketch the magnitude characteristic (versus frequency) for your channel and justify its use. What is the loss with respect to optimum MFB performance on this channel with your precoder and detector? f. Compute the average energy of the Tomlinson precoder output. but were relatively close. Suppose the channel were not exactly a PR channel as in part a. (1 pt) c. disks) and retrieval makes use of the storage of magnetic fluexes on a magnetic disk. (1 pt) 3.. (1 pt) d. Suppose the density (bits per linear inch) of a disk is to be increased so one can store more files on it.

(1 pt) 3. determine the Pe with the effect of error propagation included. State the precoding rule and the noiseless decoding rule. (4 pts) c. (1 pt) d.39 Partial Response Precoding and the ZF-DFE Consider an AWGN with H(D) = (1 − D)2 with noise variance σ2 and one-dimensional real input xk = ±1. Compute the average energy of the channel input for the simple precoder with the input constellation of part (a). what is the loss with respect to the MFB? Ignore nearest neighbor terms for this calculation since the MFB concerns only the argument of the Q-function. as well as the decoding rule for the noiseless channel output.41 Forcing Partial Response Consider a H(D) = 1 + . (1 pt) 272 . (2 pts) b. How many distinct outputs are produced by the Flexible Precoder assuming an initial state (feedback D element contents) of zero for the precoder. determine the Pe for the precoded channel. (4 pts) c. r±1. Determine a partial-response (PR) precoder for the channel.b. (4 pts) d. Design an equalizer that will convert the channel to a 1 + D channel. If a ZF-DFE is used instead of a precoder for this channel. We would like to convert this to a 1 + D channel a. This incorrect decision affects zk at time k = 2. ignoring nearest neighbor calculations. Is the noise white? (3 pts) c. The received signal is yk = xk + . (1 pt) d. so that Pc(D) = 1 − 2D + D2 .40 Error propagation and nearest neighbors Consider the H(D) = 1 − D2 channel with AWGN noise variance σ2 and d = 2 and 4-level PAM transmission. (1 pt) 3. Find the possible noiseless outputs and their probabilities.1D channel to a 1 + D channel? You need only discuss briefly. Suppose a ZF-DFE is used on this system and that at time k = 0 an incorrect decision x0 − x0 = 2 ˆ occurs. Do you think that the noise terms would be more or less correlated if we were to convert a 1 + . (1 pt) c. (2 pts) b. Compare the Pe in part (c) with that of the use of the precoder in part (a). Evaluate r0 . Compare this with the performance of the precoder. Find the autocorrelation of the noise after going through the receiver designed in part (a). What are the possible noiseless outputs and their probabilities? From these.9xk−1 + nk where nk is the AWGN. a. and r±2. Find the Ne (taking the error at k = 0 into account) for the ZF-DFE. Find also Ne and Pe with the use of precoding. How many possible outputs are produced by the simple precoder with binary inputs? (1 pt) e. If the partial-response precoder is used with symbol-by-symbol detection. From this. Compute the average energy of the Flexible precoder output. (1 pt) 3. (1 pts) b. a.9D channel with AWGN noise variance σ2 . what is η0? Determine also the SNR loss with respect to the SNRM F B (2 pts) e.

3.42 Final 2001 - Equalizers - 11 pts PAM transmission on a filtered AWGN channel uses basis function ϕ(t) = and undergoes channel impulse response with Fourier transform (|α| < 1) H(ω) =
¯ E and SNR = σx = 28 dB. 2 1 1+a·eω

1 √ T

sinc

t T

with T = 1

0

|ω| ≤ π |ω| > π

(3.545)

a. Find the Fourier Transform of the pulse response,P (ω) =? (1 pt) b. Find p
2

. (1 pt)

c. Find Q(D), the function characterizing ISI. (2 pts) d. Find the filters and sketch the block diagram of receiver for the MMSE-DFE on this channel for a = .9. (3 pts) e. Estimate the data rate for uncoded PAM transmission and Pe < 10−6 that is achievable with your answer in part d. (2 pts) f. Draw a diagram of the better precoder’s (Tomlinson or Laroia), xmit and rcvr, implementations with d = 2 in the transmitted constellation. (2 pts) 3.43 Final 2001 - Finite-Length Equalizer Design - 16 pts Transmission lines can sometimes have their transfer function described by the function H(f) = k · e−α·|f | (3.546)

where α and k are some positive real constants, here assigned the values α = 6π · 10−7 and k = 3π/10. Our task in this problem is to design a transmission system for such a transmission line. Initially we begin with 1/T = 800 kHz and carrier frequency fc = 600 kHz on this channel. The AWGN psd is N0 = −92 dBm/Hz. The transmit power is Ex /T = 1 mW. The oversampling factor for the equalizer 2 design is chosen as l = 2. Square-root raised cosine filtering with 10% excess bandwidth is applied as a transmit filter and an ideal lowpass filter is used for anti-alias filtering at the receiver. a. Find ν so that an FIR pulse response approximates this pulse response so that less than .25 dB error in p 2 (that is, the norm square of the difference between the actual p(t) and the truncated pk is less than (10.025 − 1) · p 2. (4 pts) b. Determine a reasonable set of parameters and settings for an FIR MMSE-LE and the corresponding data rate at Pe ≤ 10−7. Please print w filter coefficients. (2 pts) c. Repeat part b for an FIR MMSE-DFE design and print coefficients. (2 pts) d. Can you find a way to improve the data rate on this channel by changing the symbol rate and or carrier frequency? Try to maximize the data rate with QAM transmission by a suitable choice of carrier frequency and symbol rate, and provide that data rate - you may maintain the assumption of oversampling by a factor of 2 and 10% excess bandwidth. (6 pts) e. Suppose a narrowband Gaussian noise with PSD -62 dBm/Hz were injected over a bandwith of 1 kHz - hereustically, would you expect this to change the performance much? What would you expect to happen in the filters of the DFE? ( 2 pts). 3.44 Final 2001 - Diversity Concept - 7 pts An additive white Gaussian noise channel supports QAM transmission with a symbol rate of 1/T = 100 kHz (with 0 % excess bandwidth) anywhere in a total baseband-equivalent bandwidth of 10 MHz transmission. Each 100 kHz wide QAM signal can be nominally received with an SNR of 14.5 dB without equalization. However, this 10 MHz wide channel has a 100 kHz wide band that is attenuated by 20 dB with respect to the rest of the band, but the location of the frequency of this notch is not known in advance to the designer. The transmitters are collocated. 273

¯ a. What is the data rate sustainable at probability of bit error Pb ≤ 10−7 in the nominal condition? (1 pt) b. What is the maximum number of simultaneous QAM users can share this channel at the performance level of part a if the notch does not exist? (1 pt) c. What is the worst-case probability of error for the number of users in part b if the notch does exist? (1 pt) d. Suppose now (unlike part b) that the notch does exist, and the designer decides to send each QAM signal in two distinct frequency bands. What is the worst-case probability of error for a diversity-equalization system applied to this channel? (1 pt) e. For the same system as part c, what is the best SNR that a diversity equalizer system can achieve? (1 pt) f. For the system of part c, how many users can now share the band all with probability of bit error less than 10−7? Can you think of a way to improve this number closer to the level of the part b? (2 pts) 3.45 Final 2002 - Diversity Viewpoint - 12 pts A time-varying wireless channel is used to send long packets of information that are decoded by a receiver that uses a symbol-by-symbol decision device. The channel has pulse response in the familiar form: √ π |ω| ≤ T T · 1 + ai · e−ωT P (ω) = (3.547) π 0 |ω| > T where ai may vary from packet to packet (but not within a packet). There is also AWGN with constant power spectral density N0 = .1. QPSK (4QAM) is sent on this channel with symbol energy 2. 2 a. Find the best DFE SNR and corresponding probability of symbol error for the two cases when a1 = .9 and a2 = .5. (2 pts)

Now, suppose the transmitter can resend the same packet of symbols to the receiver, which can delay the channel output packet from the first transmission until the symbols from the new packet arrive. However, the same symbol-by-symbol detector is still to be used by the receiver. b. Explain why this is a diversity situation. (1 pt) c. Find a new SNRM F B for this situation. (1 pt) d. Find the new function Q(D) for this diversity situation. (2 pts) e. Determine the performance of the best DFE this time (SNR and Pe ). ( 2 pts) f. Compare the Pe of part e with the product of the two Pe’s found in part a. Comment. (2 pts) g. Draw the receiver with RAKE illustrated as well as DFE, unbiasing, and decision device. (phase splitter can be ignored in diagram and all quantities can be assumed complex.) (2 pts). 3.46 Precoder Diversity - Final 2003 - 7 pts A system with 4PAM transmits over two discrete-time channels shown with the two independent ¯ AWGN shown. (Ex = 1) a. Find the RAKE matched filter(s) for this system ? (1 pt) b. Find the single feedforward filter for a ZF-DFE? (1 pt) c. Show the best precoder for the system created in part b. (1 pt) 274

Figure 3.69: Transmission system for Problem 3.47. d. Find the SNR at the detector for this precoded system? (1 pt) e. Find the loss in SNR with respect to the of either single channel. (1 pt) f. Find the Pe for this precoded diversity system. (1 pt) g. Find the probability that a packet of 1540 bytes contains one or more errors. (1 pt) 3.47 Equalizer Performance and Means Recall that arithmetic mean, geometric mean, and harmonic mean are of the form (
n i=1

xi ) ,

1 n

1 n

n 1 i=1 xi

−1

1 n

n i=1

xi ,

, respectively. Furthermore, they satisfy the following inequalities: 1 n
n n
1 n

xi
i=1


i=1

xi

1 n

n

−1

i=1

1 xi

,

with equality when x1 = x2 = · · · = xn a. (4 pts) Express SN RZF E , SN RZF −DF E , SN RM F B in terms of SN RM F B and frequency response of the autocorrelation function qk . Prove that SN RZF E ≤ SN RZF −DF E ≤ SN RM F B using above b n inequalities. When does equality hold? hint: use a x(t)dt = limn→∞ k=1 x( b−a k + a) · b−a n n b. (4 pts) Similarly, prove that SN RM M SE−LE,U ≤ SN RM M SE−DF E,U ≤ SN RM F B using above inequalities. When does equality hold? c. (4 pts) Compare SN RZF −DF E and SN RM M SE−LE . Can you say which scheme has better performance? 3.48 Unequal Channels (Moshe Malkin- instructor, Midterm 2007) Please assume the gap approximation works in all SNR ranges for this problem. 275

a. (3. Answer the following questions for this situation. The waveform φ(t) is given by 1 φ(t) = √ sinc T t T (3.4 GHz.549) and the channel response is given by H(f) = e−j5πf T and so the received signal is given by y(t) = h(t) ∗ x(t) + n(t) (3. Ex and dmin for this constellation? (iii) (1.5 pts) What is the receiver SNR? ¯ (viii) (3 pts) What are Pe and Pe ? (ix) (1. Which scheme is better? 276 . the noise variance of the quadrature component is α N0 = 2N0 while the noise variance in 2 the in-phase dimension is N0 .550) where n(t) is an additive white Gaussian noise random process with two-sided PSD of -100 dBm/Hz. What is Pe averaged over both of the PAM constellations? (x) (3 pts) The transmitter realizes that if he allocates different energies for the in-phase and quadrature channels (BUT still using QAM modulation with an equal number of bits for the in-phase and quadrature channels and original power constraint of 700 ¯ µW ) he could improve the Pe performance from (aviii). ¯ and the achievable data rate? what is b (xi) (2. A sequence of 16-QAM symbols with in-phase component ak and quadrature component bk at time k is transmitted on a passband channel by the modulated signal x(t) = √ 2 k ak · φ(t − kT ) · cos(ωc t) − k bk · φ(t − kT ) · sin(ωc t) . the implementation of the baseband demodulator is faulty and the additive Gaussian noise power in the quadrature component is α = 4 times what it should be. compare the QAM transmission system from part (ax) to the single-dimensional scheme (N = 1) from part (axii).548) where T = 40 ns.5 pts) What is the ISI characterizing function q(t)? what is qk ? ¯ (v) (2 pts) What are Pe and Pe for an optimal ML detector? (vi) (2 pts) Draw a block diagram of the receiver and label major components.5 pts) What data rate can be then supported that gives Pe = 10−6? (xiii) (3 pts) Using a fair comparison. fc = 2. respectively. What is the optimal energy ¯ allocation between the in-phase and quadrature channels? Given that we need Pe = 10−6. That is.5 pts) Instead of using QAM modulation as above. ¯ (xii) (1. Given this increased noise in the quadrature dimension. and the transmit power is 700 µW.5 pts) Using the results from part (ax) find the minimal increase in transmit power needed ¯ to guarantee Pe = 10−6 for 16-QAM transmission? The transmitter now decides to reduce losses.5 pts) What is xbb(t)? ˜ (iv) (1. 2 (vii) (2. (i) (1 pts) What is the data rate for this system? ¯ (ii) (2 pts) What are Ex . two independent 4-PAM constellations (with half the total energy allocated for each constellation) are now transmitted on the inphase and quadrature channels. Unfortunately. the transmitter decides to use only the in-phase dimension for transmission (but subject to the original transmit power constraint of 700 µW ).

instructor. 1.instructor. . (1 pt) Find WZF E (D) and WM M SE−LE (D). . e. Show both the precoder and the corresponding receiver.2i · 0. . a.49 Equalizer Design (9 pts.7). where the received signal is filtered by a continuous time matched filter. b.5 pts) What is the detector SNR for this system for a ZF-DFE receiver? 3. 2 (3.5 pts) What is the resulting Q(D) for this diversity system? d.551) = 10dB. Final 2007) A symbol-rate sampled MMSE-DFE structure. Ex = 1.(xiv) (3 pts) Can you derive a new transmission scheme that uses both the in-phase and quadrature ¯ channels to get a higher data rate than part (ax)? (subject to the same Pe = 10−6 . the MMSE-DFE feedback filter to be of finite length Nb . 0. Assume you are using 4-QAM modulation.181. a. (1.5j·e−j2πf . (2.9 · xk+1 + nk. b. Ex = 1.181. and with 2 ¯ variance σi = 1. 277 . (and SN RM F B = 10 dB). However.i is a Gaussian noise process independent across time and across all the channels.0445j). (1 pts) Find SN RM F B.4555j and 0. a.i = xk + 0. (1 pt) Find ||p||2 and Q(D). Assume that N0 2 = . Moshe Malkin. Make sure to mention which kind of constellation is used at each SBS detector in your drawing. 3.5 pts) What is SN RZF −DF E ? f.50 Diversity (7 pts Moshe Malkin. Also. the SN RM F B for channel i. Final 2007) 1 A transmission system uses the basis function φ(t) = sinc(t) with T = 1 Hz. and P (D) = 1 + . 2. Show both the precoder and the corresponding receiver (For any M). i = 0. (2 pts) Formulate the optimization problem you would have to solve to find the optimal feedforward and feedback coefficients to minimize the mean squared error (Hint: Don’t approach this as a finite-length equalization problem as in section 3.51 DFE with Finite Length Feedback Filter (9 pts Moshe Malkin.9 · D−1 . (2 pts) Design a Laroia precoder based on the ZF-DFE. (2 pts) Find W (D) and B(D) for MMSE-DFE (Hint: the roots of the numerator of Q(D) + 1/SN RM F B are 22. (1 pt) Find W (D) and B(D) for ZF-DFE.i. The Fourier transform of the channel impulse response is: H(f) = and SN R = εx ¯ σ2 1 1−0.i . uses an infinite length feedforward filter.5 pts) Design a Tomlinson-Harashima precoder based on the ZF-DFE. Final 2007) A receiver has access to an infinite number of diversity channels where the output of channel i at time k is yk. where nk. d. and original symbol rate and power constraint of 700 µW ) 3. (0. c. g. (1. for|f| ≤ 1 2 for|f| > 1 . (2 pts) What is the matched filter for each diversity channel? c.instructor.

(3 pts) Determine the optimal feedforward and feedback filters for the ZF-DFE formulation using only 1 feedback tap (Nb = 1). (1.instructor.552) 0 0 1. 3 2 1 0 0 0. and finally x3 . Now. y2 . what is now the detection SNR for each symbol? f. State the optimization problem for this situation (analogous to part a).you have already seen the solution before).2 0 0 0. determine the optimal feedforward and feedback filters (Hint: This is easy .86 x3 n3  y2  =  0 p22 p23   x2  +  n2  =  0 1. don’t solve . then x2. d. x2.9D −1 (and p 2 = 1 1−0.19  0 1.9 −3. y2.86 1.i.52  .6 1.just describe how this would work). c.just describe the signal processing that would take place).3  =  0 0.9 −3. If the receiver starts by detecting x1 . How would a packet MMSEDFE work? How would you describe the feedback structure? (Don’t solve for the optimal matrices . a.553) (3.d with E§∞ = E§∈ = E§ = ∈ . (2 pts) Using only 1 feedback tap (Nb = 1). 3. g.3 0 0 n3 Rn = E  n2  n∗ n∗ n∗  =  0 0. assume that we have P (D) = 1 1+0. 0 0 1.554) 278 . a receiver makes a decision from the set of transmit symbols x1. where             y3 p11 p12 p13 x3 n3 0.all the work occurred in the previous part) c.5 0  .56 −1.2 y1 0 0 p33 x1 n1 x1 n1 and where the noise autocorrelation matrix is given by      0. (2 pts) Now consider the ZF-DFE version of this problem.5 pts) What is the detection SNR for x1 using a ZFE? Answer the same for x2 and x3 .83 The transmit symbols are i.76 10. b.5 pts) What is the 3-tap zero-forcing equalizer for x1 based on the observation y1 .67 −1. 0.92 = 1 ). Final 2007) Suppose that in a packet-based transmission system. (1 pts) An engineer now realizes that the performance could be improved through use previous decisions when detecting the current symbol.8 3.3   x2  +  n2  (3. describe how this modified ZF detector would work? e.19 d.8 3. (1. and x3? (Hint: This is easy .1 n1 The inverse of the channel matrix is given by  −1   0. y3? Repeat for x2 and x3 . (1 pts) What is the matrix ZFE for detecting x1 . x2. y3 . x3 based on the channel outputs y1 . (1 pts) The approach above does not necessarily minimize the MMSE.b. (1 pts) Can the THP also be implemented for this packet based system? Describe how this would be done (again. (1. (3.52 Packet Processing (10 pts Moshe Malkin.5 pts) Assuming previous decisions are correct.6 1.

2) One can let N → ∞ without difficulty..k + wi.1 The Orthogonality Principle While (A. given observations of some other random sequences {yk }.5) 279 .. the conditional expectation of x. This text is interested only in linear MMSE estimates.k · yr. which will be linear for jointly Gaussian x and {yk }. More generally.that is that the error signal e must always be uncorrelated with all the observed random variables in order for the MSE to be minimized. . (A.k . it is often far more convenient to use a well-known principle of linear MMSE estimates . (A.k xi − wi.k − wr. it can be shown that the MMSE estimate is E [x/{yk }]. i = k.2).k · yi. chooses a set of parameters wk (with the index k having one distinct value for each observation of yk that is used) to minimize E |e|2 (A. In this text’s developments.3) Proof: By writing |e|2 = [ (e)] + [ (e)] . N − 1 2 2 (A. .1.1) can be minimized directly by differentiation in each case or problem of interest.1 (Orthogonality Principle) The MSE is minimized if and only if the following condition is met ∗ E [e · yk ] = 0 ∀ k = 0. Theorem A.Appendix A Useful Results in Linear Minimum Mean-Square Estimation This appendix derives and/or lists some key results from linear Minimum Mean-Square-Error estimation. A. we may differentiate the MSE with respect to both the real and imaginary parts of wk for each k of interest. will drop from the corresponding derivatives) er ei = = xr − wr.. and x will be the current channel-input symbol that the receiver attempts to estimate. yi will usually be successive samples of some channel output sequence. The term linear MMSE estimate is derived from the linear combination of the random variables yk in (A... yN −1 ..k · yi. given y0 .1) where N −1 ∆ e = x− k=0 wk · yk .4) (A.k · yr. A linear MMSE estimate of some random variable x. The pertinent parts of the real and imaginary errors are (realizing that all other wi .

8) ∗ By stationarity.j = rxx. ∆ In this chapter. then F ∗ (D−∗ ) is anticanonical. The sequence {x∗ } has a D-Transform X ∗ (D−∗ ) = −k ∗ −k .2 Spectral Factorization k This section uses D-transform notation for sequences: Definition A. to optimize over wr.k − ei yr. then this minimum is unique. implying the additional operation limN →∞ [1/(2N + 1)] k−j −N ≤k≤N on the sum in such terms. monic.Transforms) A sequence {xk } has D-Transform X(D) = xk D k . − < ω ≤ . k xk D Definition A. Then.11) where S0 is a positive real number and F (D) is a canonical filter response.k ∂wr. all sequences can be complex. and is determined by either the autocorrelation function or the power spectrum as follows: Ex = E |xk|2 = rxx. its autocorrelation function is defined as the sequence Rxx(D) whose terms are rxx.3 (Factorizability) An autocorrelation function R(D) will be called factorizable if it can be written in the form R(D) = S0 · F (D) · F ∗ (D−∗ ). ¯ The quantity E |xk |2 is Ex . Since the MSE is positive-semi-definite quadratic in the parameters wr. which is written as π π Rxx(e−ωT ) = Rxx (D)|D=e−ωT .9) T T which is real and nonnegative for all ω. and maximum-phase.k = = 2er 2er ∂er ∂ei + 2ei = −2 (er yr. (A. If the MSE is strictly positive-definite.k and wi. (A. If F (D) is canonical. i. and minimumphase (all of its poles are outside the unit circle. since the terms of X(D)X ∗(D −1 ) are of k ∆ the form xk x∗ . (A.k (A. 1 The expression Rxx (D) = E X(D)X ∗ (D −1 ) is used in a symbolic sense.k .k and wi.2.k ∂wi. this setting must be a global minimum. ∂|e|2 ∂wr. 280 .k ) = 0 . Conversely.2. or Ex per dimension.−j and Rxx (D) = R∗ (D−∗ ). anticausal.10) The spectral factorization of an autocorrelation function is: Definition A.where subscripts of r and i denote real and imaginary part in the obvious manner. QED.e. A filter response F (D) is called canonical if it is causal (fk = 0 for k < 0). we say that any complex sequence R(D) for which R(D) = R∗ (D−∗ ) is an autocorrelation function.6) (A. and all of its zeroes are on or outside the unit circle). ∂wi. rxx. and any function R(e−ωT ) that π π is real and nonnegative over the interval {− T < ω ≤ T } is a power spectrum.2 (Autocorrelation and Power Spectrum) If {xk } is any stationary complex sequence..0 = T 2π π T π −T R(e−ωT )dω . A.k ) = 0 ∂wr.k ∂er ∂ei + 2ei = 2 (er yi.k . monic (f0 = 1). (A.7) The desired result is found by taking expectations and rewriting the series of results above in vector form.1 (D. symbolically1 k−j Rxx(D) = E X(D) · X ∗ (D−∗ ) ∆ .k ∂|e|2 ∂wi.j = E[xkx∗ ].2.k + ei yi. The power spectrum of a xx stationary sequence is the Fourier transform of its autocorrelation function.

and R(D) is the corresponding autocorrelation function. A.12) where the finite constant S0 is given by log S0 = 1 2π π T π −T ∆ log R(e−ωT )dω (A. and the prediction error sequence E(D) is given by E(D) = X(D) − X(D)[1 − A(D)] = X(D)A(D) . 281 . there are really two factorizations. which has mean square value S0 = 1/γ0. In this case. (A. and also gives the value of S0 . X(D) can be viewed as being generated by inputting a white innovations process U (D) = G(D)X(D) with mean square value S0 into a filter F (D) so that X(D) = F (D)U (D). This simple inequality has many uses. as with the MMSE-DFE in Section 3. The factorization of the inverse and resultant interpretation of the factor G(D) as a linear-prediction filter helps develop an interest interpretation of the MMSE-DFE in Section 3.1 (Spectral Factorization) If R(e−ωT ) is any power spectrum such that π π both R(e−ωT ) and log R(e−ωT ) are integrable over − T < ω ≤ T . R(e−ωT ) satisfies the discrete-time Paley-Wiener criterion 1 2π π T π −T | log R(e−ωT )|dω < ∞ . The following lemma from discrete-time spectral factorization theory describes when an autocorrelation function is factorizable. the inverse autocorrelation factors as R−1 (D) = γ0 · G(D) · G∗ (D−∗ ) .2.17) (A. with equality if and only if A(D) is chosen as the whitening filter A(D) = G(D). The autocorrelation function of the prediction error sequence is Ree(D) = Rxx(D) · A(D) · A∗ (D−∗ ) = A(D) · A∗ (D−∗ ) .6. since F (D) is monic. For S0 to be finite. In this finite-length case. Thus. S0 of the direct spectral factorization is the mean-square value of the innovations process or equivalent of the MMSE in linear prediction.6. then f 2 = j |fj |2 ≥ 1.14) Linear Prediction The inverse of R(D) is also an autocorrelation function and can be factored when R(D) also satisfies the PW criterion with finite γ0 .2.16) so its average energy satisfies Ee = S0 1/g ∗ a 2 ≥ S0 (since A(D)/G(D) is monic). γ0 · G(D) · G∗ (D−∗ ) (A.1 Cholesky Factorization Cholesky factorization is the finite-length equivalent of spectral factorization. then 1−A(D) is a strictly causal sequence that may be used as a prediction filter.If F (D) is a canonical response. (A. The process X(D) · G(D) is often called the innovations of the process X(D). with equality if and only if F (D) = 1. then there is a canonical discrete-time response F (D) that satisfies the equation R(D) = S0 · F (D) · F ∗ (D−∗ ). and A(D) is any causal and monic sequence. (A.15) xx If x(D) is a discrete-time sequence with factorizable inverse autocorrelation function R−1(D) = xx γ0 G(D)G∗ (D−∗ ). both of which converge to the infinite-length spectral factorization.13) (where the logarithms can have any common base). Lemma A. see Exercises for Chapter 3.

(A.   .Cholesky Form 1 . . . . 0 . A convenient description of RN ’s components is then r∗ rN 1 RN = .23) r1 RN −1 0 sN −2 ..21) sn . . .22) The submatrix RN −1 has Cholesky Decomposition RN −1 = F N −1 S N −1 F N −1 . .25) N N where F N = G−1 is also upper triangular and monic with ordering N   g N −1  g N −2    GN =   . (A. 0  s0 (A..  0 0    .    0 . (A. XN =  (A...24) which because of the 0 entries in the triangular and diagonal matrices shows the recursion inherent in Cholesky decomposition (that is the F N −1 matrix is the lower right (N − 1) × (N − 1) submatrix of F N .Forward Prediction Cholesky factorization of a positive semidefinite N × N matrix RN produces a unique upper triangular monic (ones along the diagonal) matrix F N and a unique diagonal positive-semidefinite matrix S N such that RN = F N S N F ∗ . . it is convenient to write ˜ fi = 1 fi The determinant of RN is easily found as N −1 2 γ0 = |RN | = n=0 2 (or ln γ0 = ln |RN | for readers taking limits as N → ∞). (A.19)  .. ˜ gi = [1 gi ] . (A.18) N It is convenient in transmission theory to think of the matrix RN as an autocorrelation matrix for N samples of some random vector process xk with ordering   xN −1   . The inverse of RN (pseudoinverse when singular) has a Cholesky factorization −1 R−1 = G∗ S N GN (A. 0 f0 Since F N is monic. x0 A corresponding order of F N and S N is then    f N −1 sN −1  f N −2   0    FN =   and S N =  . .20) .. g0 Also. . . which is also upper triangular).. (A.27) (A.26) 282 . .

and N UN  uN −1   . =  . The cross-correlation between X N and U N is ¯ Rux = S N G∗ .28) (A. (A. Thus. . 283 . Forward Cholesky Algorithm Summary: For nonsingular RN : Set f 0 = F 0 = g0 = G0 = 1.34) U N−1 (A. F n = c.32) (A.. . U N = GX N . Equation (A. from (A. ˜ 1 −f n Gn−1 0 GN −1 ˜∗ ˜ d. E uk x∗ k−i = 0 ∀ i ≥ 1 . u0  (A.. The innovations. Gn = ˜ 1 fn 0 F n−1 . uN = xN − r∗ R−1 X N −1 1 N −1 since r1 is the cross-correlation between xN and X N −1 .32) rewrites as xN = = uN + r∗ R−1 X N −1 1 N −1 UN + −1 r ∗ G∗ −1 S N −1 GN −1 X N −1 1 N ˜ fN (A.31) Since X N −1 = F N −1 U N −1 shows a reversible mapping from U N −1 to X N −1 . then (A. of the N samples of XN are defined by XN = F UN where E [U N U ∗ ] = S N .35) = so fNUN .33) (A. (A. . x0 (i. 1 n−1 b. .linear prediction A straightforward algorithm for computing Cholesky factorization derives from a linear prediction interpretation.e. U N .36) (A. −1 f N = 1 r ∗ G∗ −1 S N −1 1 N . ˜ gN = 1 − f N GN −1 Finally. N which is lower triangular.31) relates that the sequence uk is a set of MMSE prediction errors for xk in terms of xk−1 . ˜∗ ˜ sN = rN − f N S N −1 f N .32) and (A.38) Then.30) (A. S 0 = s0 = E|x0|2 For n = 1 .34). sn = rn − f N S N −1 f N −1 . f n = r∗ G∗ Sn−1 ... (A. Thus.. N : −1 ˜ a.37) (A.29) Also.31) is the orthogonality principle for linear prediction).

The new F n>i and Gn>i created will have smaller dimension. zeros can be added to all the “right-most” positions of F and G. That is. first invert the given Q autocorrelation matrix to get R and then do forward Cholesky on the inverse (RN = F N S N F ∗ = N G−1SG−∗ . Thus. except that this algorithm is performed on the inverse (which.7. (A. (A. The inverse matrix QN = R−1 = G∗ S −1GN is also an autocorrelation matrix for some process N N N −1 Y N with innovations V N where Y N = G∗ V N such that E [V N V ∗ ] = SN . this section notes that the reverse backward prediction problem is inherent in the algorithm already given.42) (A.Backward Prediction The backward Cholesky factorization is into a lower-diagonal-upper (instead of the upper-diagonal-lower in forward prediction). y0. the inverse autocorrelation function factors as R−1 (D) = γ0 G(D)G∗ (D−∗ ) . but when the algorithm is complete. ¯ This factorization of Q is lower-diagonal-upper and the filters f n are backward prediction filters for y1−n in terms of y2−n . So to perform backward-prediction Cholesky. If Q is singular. (A.. then its N N pseudoinverse is used by definition here. might be harder to compute than just repeating the steps above with the appropriate transpositions)... For the stationary case. the matrix R (and therefore Q) must be nonsingular to satsify the Paley-Weiner Criterion. granted. one takes the limit as N → ∞ in either forward or backward Cholesky factorization. In this case. the backward Cholesky factorization uses the G of forward Cholesky (for R = Q−1 ) N N in −1 Q = G∗ S N GN . (A.. N N which is convenient for the alternative finite-length MMSE-DFE interpretations in Section 3. meaning that xi can be exactly predicted from the samples xi−1 . the concepts of forward and backward prediction are the same so that the forward predictor is just the time reverse of the coefficents in the backward predictor. Cholesky Form 2 .44) 284 . The equivalence to spectral factorization is evident from the two linearprediction interpretations for finite-length and infinite length series of samples from random processes. In this case.. Rather than repeat the development of forward prediction. u0 .41) ¯∗ ¯∗ illustrating the backward prediction error v1−N = f N Y N −1 where f N is the N th row of F ∗ or equivaN ¯ lently f N is the last column of F N . This involves extension of RN on the lower right corner instead of at the upperleft corner.A singular RN means that sn = 0 for at least one index n = i. Then. Thus. One factorization is to save F i and Gi and “restart” the algorithm for n > i realizing that xi is equivalent to knowing all lower-index samples.40) The forward Cholesky factorization can also be written RN = G−1 S N G−∗ .39) 0 Fi 0 Gi and SN S n>i = 0 0  0 0 0 0 0 S i−1   .. F n>i 0 Gn>i 0 FN = and GN = (A. V N = F∗ Y N . which is equivalent to ui = 0. Cholesky factorization is not unique.43) N Infinite-length convergence For infinite-length stationary sequences. x0 or equivalently can be exactly constructed from ui−1 . and maintaining the previous F i and Gi for the lower rightmost positions.

45) n→∞ Similarly.N = |R−1|1/N lim S0.n(D) γ0 = e T 2π π/T −π/T . lim Gn(D) = G(D) . n→∞ lim G∗ (D) = G∗ (D−∗ ) . Factorization of a singular process at infinite-length involves separating that process into a sum of subprocesses.” it becomes obvious why such a process cannot converge to a constant limit and so only nonsingular processes are considered in spectral factorization.48) The convergence to these limits implies that the series of filters converges or that the bottom row (last column) of the Cholesky factors tends to a constant repeated row.n = |R|1/N and γ0. see Chapter 5. (A. In the limit as N → ∞ for a stationary nonsingular series. This is equivalent to the “restarting” of Cholesky at infinite length. Interestingly. defining the geometric-average determinants as S0.46) As N → ∞. For more. Finally. 285 .47) ln(R(e−ωT ))dω lim γ0. Using the modifications to Cholesky factorization suggested above with “restarting. should tend to a constant. (A. a Cholesky factorization of a singular process exists only for finite lengths.where G(D) is the forward prediction polynomial (and its time reverse specified by G∗ (D−∗ ) is the backward prediction polynomial).n = = S0 = e − n→∞ T 2π π/T −π/T ln(R(e−ωT ))dω n→∞ (A.es SN −1 . each of which is resampled at a new sampling rate so that the PW criterion (nonsingular needs PW) is satisfied over each of the subbands associated with these processes. The series of matrices {Rn}n=1:∞ formed from the coefficients of R(D) creates a series of linear predictors {gn}n=1:∞ with D-transforms Gn(D). the prediction error varianc. n (A.

Thus a maximum likelihood (or MAP) detector can be designed that uses the sequence yk to estimate the input symbol sequence xk without performance loss. There are a variety of ways in which equalization can be used in conjunction with either sequence detection and/or symbol-by-symbol detection.Appendix B Equalization for Partial Response Just as practical channels are often not free of intersymbol interference. an equalizer may be used to adjust the channel’s shape to that of the desired partial-response polynomial.2 showed that a minimum-phase channel polynomial H(D) can be derived from the feedback section of a DFE.1 ZF-DFE and the Optimum Sequence Detector Section 3. (B. The feedforward filter 1/(η0 p H ∗(D−∗ )) of the ZF-DFE is invertible when Q(D) is factorizable. sequence detection on the output of the feedforward filter of a DFE can be designed using the controlled ISI polynomial H(D) or approximations to it. The reversibility theorem of Chapter 1 then states that a maximum likelihood detector that observes the output of this invertible ZF-DFE-feedforward filter to estimate xk also has no performance loss with respect to optimum.1) The noise sequence nk is exactly white Gaussian for the ZF-DFE. so the ZF-DFE produces an equivalent channel H(D). Figure B. 286 (B.8. such channels are rarely equal to some desirable partial-response (or even controlled-ISI) polynomial. of the receiver matched filter form a sufficient statistic for the underlying symbol sequence. Thus.1 shows such an optimum receiver. + hν Dν .1. B.1 showed that the sampled outputs. then a sequence detector can be designed based on the controlled-ISI polynomial H(D) and this detector has minimum probability of error. This polynomial is often a good controlled intersymbol interference model of the channel when H(D) has finite degree ν. then the first ν feedback taps of H(D) determine H (D) = 1 + h1D1 + . Thus.1 Controlled ISI with the DFE Section 3. B. Chapter 9 studies sequence detection so terms related to it like MLSD (maximum likelihood sequence detection) and Viterbi detectors/algorithm are pertinent to readers already familiar with Chapter 9 contents and can be ignored by other readers otherwise in this Appendix. This equalizer then enables the use of a partial-response sequence detector or a symbol-by-symbol detector at its output (with partial-response precoder). The feedforward filter output has D-transform Z(D) = X(D)H(D) + N (D) . yk .2) . This section introduces some equalization methods for partial-response signaling. the first ν coefficients of H(D) form a controlled intersymbol interference channel. and so the equalizer can be crucial to achieving the highest levels of performance in the partial-response communication channel. The performance of either of these types of detectors is particularly sensitive to errors in estimating the channel.. If H(D) is of finite degree ν.. When H(D) is of larger degree or infinite degree. When H(D) has larger degree than some desired ν determined by complexity constraints.

3) is maximum for all ν ≥ 0.1: The optimum maximum-likelihood detector. When this second feedback section has zero coefficients.2: Limiting the number of states with ZF-DFE and MLSD. Increase of ν causes the minimum distance to increase. The second feedback section contains all the channel coefficients that are not used by the sequence detector. Thus. When the additional feedback section is not zero. remain the same (which is proved simply by examining a trellis with more states and realizing that the smaller degree H (D) is a sub trellis for which distance cannot be greater. the feedforward section of the ZF-DFE Figure B. Thus the SNR of the signal entering the Viterbi Detector in Figure ¯ h 2 E B.2 defines a series of increasingly complex receivers whose performance approach optimum as ν → ∞. A property of a minimum-phase H(D) is that ν |hi|2 = h i=0 2 (B. No other polynomial (that also preserves the AWGN at the feedforward filter output) can have greater energy. also increases (nondecreasing) with ν. a controlled-ISI channel.2 is an optimum detector. xN0 . The inside feedback section is replaced by a modulo symbol-by-symbol detector when precoding is used. Figure B. 2 287 . implemented with the assistance of the WMF.Figure B. This SNR must be less than or equal to SNRM F B . or at worst. the ZF-DFE with sequence detector in Figure B. These coefficients have delay greater than ν. then this structure is intermediate in performance between optimum and the ZF-DFE with symbol-by-symbol detection.2.2 illustrates the use of the ZF-DFE and a sequence detector. then the configuration shown in Figure B.

288 .3. (B. causal.1. the equalization error sequence becomes Section 3.4 and follow it by a filter of the desired partial-response (or controlled-ISI) polynomial B(D).2 Equalization with Fixed Partial Response B(D) The derivations of Section 3. The monic. to obtain ¯ Rxy (D) B(D) W (D) = B(D) ¯ .4) In the linear equalizer case. A truncated version of GU (D) corresponding to H (D) is denoted GU (D).Figure B.2. The error sequence associated with the unbiased MMSE-DFE is not quite white.2 with H(D) replaced by Gu (D). which this section reuses.6) also follows from the linearity of the MMSE estimator.3: Partial-response linear equalization. with an unbiased MMSE-DFE’s MS-WMF replacing the WMF of the ZF-DFE. B. A sequence detector can use the same structures as shown in Figure B. and unbiased feedback polynomial was denoted GU (D) in Section 3. and so SNR maximization was applied in Chapter 3 to the DFE to obtain the MMSE-DFE. So. SNRM M SE−DF E. The solution was found by setting E Epr (D)y∗ (D−1 ) = 0.U when ν = 0. remains unchanged and is not necessarily obtained.6 minimized MSE for any B(D) over the coefficients in W (D). = p (Q(D) + 1/SNRM F B ) Ryy (D) (B.6 on the MMSE-DFE included the case where B(D) = G (D). However.2. B. The matched filter bound. Figure B.6. to that of the optimum detector when ν → ∞.5) which is just the MMSE-LE cascaded with B(D). minimum distance does increase with ν in the sequence detectors based on a increasing-degree series of GU (D).1 The Partial Response Linear Equalization Case Epr (D) = B(D)X(D) − W (D)Y (D) .2 MMSE-DFE and sequence detection Symbol-by-symbol detection’s objective is to maximize SNR in an unbiased detector. minimum-phase. B.1 and B. the error sequence is Epr (D) = B(D)X(D) − B(D)Z(D) = B(D) [E(D)] 1 This (B. For instance.1 For this choice of W (D). the probability of error decreases from the level of the unbiased MMSE-DFE.4 is the same as Figure B. nor is it Gaussian. Figure B. As ν increases. The designer need only realize the MMSE-LE of Section 3. but it is nevertheless commonly used because the exact optimum detector could be much more complex. A bias in symbol-by-symbol detector was removed to minimize probability of error. shows the MMSE-PREQ (MMSE “Partial Response Equalizer”). a sequence detector based on squared distance is not quite optimum. as always.

the designer usually selects B(D) so that SNRMMSE-PREQ > SNRM M SE−LE . the error sequence Epr = B(D)E(D) is not a white noise sequence (nor is E(D) for the MMSE-LE). While this is a seemingly simple observation made here. then a ZF-PRDFE should be used. This receiver also has a bias. obtained by setting SNRM F B → ∞ in the above formulae. While the MMSE-PREQ should be used for the case where symbol-by-symbol and precoding are being used. The error sequence is the same as that for the MMSE-DFE. 289 . so that a Viterbi Detector designed for AWGN on the channel B(D) would not be the optimum detector for our MMSE-PREQ (with scaling to remove bias). ¯ ¯ Repr .7) |B(e−ωT )|2 N0 2 dω . unless the channel pulse response were already very close to B(D). since the MMSE-PRDFE error sequence is white. ˜ Again. In this case. From (B.2. Thus a linear equalizer for a partial response channel B(D) that is followed by a Viterbi Detector designed for AWGN may not be very close to an optimum detection combination. (B. The means by which to correct this situation is the PR-DFE of the next subsection.8) The symbol-by-symbol detector is equivalent to subtracting (B(D) − 1)X(D) before detection based on decision regions determined by xk . but it is usually ignored because of the integer coefficients in B(D) – any bias removal could cause the coefficients to be noninteger. B.4. This signal can be processed on a symbol-by-symbol basis if precoding is used (and also scaling is used to ˜ remove the bias . and then use a WGN Viterbi Detector. so that equalization was not initially necessary.U SNRM M SE−DF E . The SNRMMSE-PREQ becomes ¯ Ex SNRMMSE-PREQ = 2 σ MMSE-PREQ . depending on the choice of B(D). The signal between the the two feedback sections in Figure B. (B.6). and the filter settings are obtained by setting SNRM F B → ∞ in the above formula.2. using our earlier result that for any feedback section GU (D) = B(D) + B(D). and is therefore a white sequence. the designer can reasonably use a Viterbi Detector designed for B(D) with white noise. and BU (D) = GU (D) − BU (D). then the optimal MMSE-PRDFE is shown in Figure B. “color” the noise spectrum. p 2 (Q(D) + 1/SNRM F B ) (B. and do equalize to partial response. which is illustrated in Figure B.2 The Partial-Response Decision Feedback Equalizer If B(D) = G(D) and the design of the detector mandates a partial-response channel with polynomial B(D).where E(D) is the error sequence associated with the MMSE-LE.4 is input to the sequence or symbol-by-symbol detector. p 2 (Q(e−ωT ) + 1/SNRM F B ) (B. where BU (D) = 1 SNRM M SE−DF E B(D) − SNRM M SE−DF E. and because the bias is usually small so that the error sequence in the unbiased case is also almost white.9) This performance can be better or worse than the MMSE-LE. the ZF-PREQ. there are a number of systems proposed for use in disk-storage detection that overlook this basic observation. the MMSE for the PREQ can be computed as T σMMSE-PREQ = 2π 2 π T π −T B(D) N0 B ∗ (D−1 ) 2 .the scaling is again the same scaling as used in the MMSE-DFE).10) However. If the bias is not negligible.epr (D) = B(D)Ree (D)B ∗ (D−1 ) = Thus. would also not have a white error sequence.

4: Partial-response DFE with B(D) 290 .Figure B.

1) 291 .Appendix C The Matrix Inversion Lemma: The matrix inversion lemma: [A + BCD] −1 = A−1 − A−1 B C −1 + DA−1 B −1 DA−1 . (C.

Sign up to vote on this title
UsefulNot useful