Equalization

52

 Delay spread causes intersymbol interference (ISI), which can

cause an irreducible error when the symbol time is on the same
order as the channel delay spread ( .
 Equalization defines any signal processing technique used at
the receiver to alleviate the ISI problem caused by delay
spread.
 Mitigation of ISI is required when the symbol time Ts is on the
order of the channel’s rms delay spread .
 Equalizer design must balance ISI mitigation with noise
enhancement, since both the signal and the noise pass through
the equalizer, which can increase the noise power.
 Equalization…
53

 Nonlinear equalizers suffer less from noise enhancement than linear

equalizers but typically entail higher complexity.
 Equalizers require an estimate of the channel impulse or frequency
response to mitigate the resulting ISI.
 Since the wireless channel varies over time, the equalizer must
 Learn the frequency or impulse response of the channel (training)
 Then update its estimate of the frequency response as the channel changes
(tracking)
 The process of equalizer training and tracking is often referred to as
adaptive equalization, since the equalizer adapts to the changing
channel.
 Equalizer training and tracking can be quite difficult if the channel is
changing rapidly.
 Equalizer Noise Enhancement
54

 Equalizer must balance ISI mitigation so that noise power in the received
signal is not enhanced.

 Consider a simple analog equalizer shown in the figure above.

 Consider a signal s(t) that is passed through a channel with frequency
response H(f).
 At the receiver front end, white Gaussian noise n(t) is added to the signal
and so the signal input to the receiver is
Y(f ) = S(f)H(f) + N(f)
 Equalizer Noise Enhancement…
55

where N(f ) is white noise with power spectral density (PSD) N0/2.
 If the bandwidth of s(t) is B, then the noise power within the signal
bandwidth of interest is N0B.
 In order to completely remove ISI, an analog equalizer is introduced in
the receiver that is defined by

 The received signal Y(f ) after passing through this equalizer becomes

where N’(f) is colored Gaussian noise with power spectral density .

 Equalizer Noise Enhancement…
56

 Now all ISI is removed but it encounters noise enhancement under

certain conditions.
 If H(f) has a spectral null (H(f0) = 0 for some f0) at any
frequency within the bandwidth of s(t), then the power of the
noise N’(f) is infinite.
 Even without a spectral null, if some frequencies in H(f) are
greatly attenuated then the equalizer Heq(f ) = 1/H(f) will
greatly enhance the noise power at those frequencies.
 In this case, even though the ISI effects are removed, the
equalized system will perform poorly because of its greatly
reduced SNR.
 Equalizer Noise Enhancement…
57

 Linear digital equalizers work by approximately inverting the

channel frequency response and thus have the most noise
enhancement.

 Nonlinear equalizers do not invert the channel frequency

response, so they tend to suffer much less from noise
enhancement.
 Equalizer Types
58

 Equalization techniques fall into two broad categories: linear

and nonlinear.
 Linear Techniques:
 The simplest to implement and to understand conceptually.
 Suffer from more noise enhancement than nonlinear equalizers and
hence are not used in most wireless applications.
 Non-Linear Techniques:
 Decision-feedback equalization (DFE) is the most common because it
is fairly simple to implement and usually performs well.
 On channels with low SNR, the DFE suffers from poor performance. The
optimal equalization technique is maximum likelihood sequence
estimation (MLSE).
 Equalizer Types…
59

 Linear and nonlinear equalizers are typically implemented using

a transversal or lattice structure.
 The transversal structure is a filter with N −1 delay elements and
N taps featuring tunable complex weights.
 The lattice filter uses a more complex recursive structure.
 Adaptive equalizers require algorithms for updating the filter
tap coefficients during training and tracking.
 These algorithms generally incorporate trade-offs between
complexity, convergence rate, and numerical stability.
 Equalizer Types…
60

 Equalizers can also be categorized as symbol-by-symbol (SBS)

or sequence estimators (SEs).
 SBS equalizers remove ISI from each symbol and then detect
each symbol individually.
 All linear equalizers as well as the DFE are SBS equalizers.
 Sequence estimators detect sequences of symbols, so the effect
of ISI is part of the estimation process.
 Maximum likelihood sequence estimation is the optimal form of
sequence detection
 Linear Equalizer
61

 We assume a linear equalizer implemented via a 2L +1 = N-

tap transversal filter:
 Linear Equalizer…
62

 For a given equalizer size N, the equalizer design must specify

(i) the tap weights {wi}Li=−L for a given channel frequency
response and (ii) the algorithm for updating these tap weights as
the channel varies.
 Even though the performance metric in wireless systems is outage
probability, it is difficult to optimize the equalizer coefficients
subject to this criteria.
 Hence an indirect optimization that balances ISI mitigation with
the prevention of noise enhancement is employed.
 Zero-Forcing (ZF) Equalizers
63
 Zero-Forcing (ZF) Equalizers…
64

 The samples {yn} input to the equalizer can be represented

based on the combined impulse response f(t) = h(t) g (−t) as

where Ng(z) is the z-transform of the noise samples at the output of

the matched filter G m(1/z ) and

 The zero-forcing equalizer removes all ISI introduced in the

composite response f(t).
 Zero-Forcing (ZF) Equalizers…
65

 To accomplish this, the equalizer impulse response should be

 The power spectrum N(z) of the noise samples at the equalizer

output is given by
 Zero-Forcing (ZF) Equalizers…
66

 If the channel H(z) is sharply attenuated at any frequency within

the signal bandwidth of interest – as is common on frequency-
selective fading channels – the noise power will be significantly
increased.
 This requires an equalizer design that better optimizes between
ISI mitigation and noise enhancement. One such equalizer is
the MMSE equalizer.
 Since linear equalizers are implemented as transversal tap
filters, we can find a set of coefficients {wi} that best
approximates the zero-forcing equalizer.
 Problem
67

Q) Consider a channel with impulse response

Find a two-tap ZF equalizer for this channel.

SOLUTION:
Sampled version of h(t), 

h[n] = 1+ e−Ts/τδ[n −1] + e−2Ts/τδ[n − 2]+· · ·

−1
H(z) = 1+ e−Ts/τz + e−2Ts/τz−2 + e−3Ts/τz−3 +· · ·
 Problem…
68

Heq(z) = 1/H(z) = 1 − e−Ts/τz −1.

The two-tap ZF equalizer therefore has tap weight coefficients w0 = 1

and w1 = e−Ts/τ.
 Minimum Mean-Square Error
(MMSE) Equalizers
69

 The goal of the MMSE equalizer design is to minimize the

average mean square error (MSE) between the transmitted
symbol dk and its estimate at the output of the equalizer.
 i.e. the {wi} are chosen to minimize E[ ]2.
 Since the MMSE equalizer is linear, its output is a linear
combination of the input samples y[k]:

 The noise input to the equalizer has a power spectrum

N0|G m(1/z )|2.
 Minimum Mean-Square Error
(MMSE) Equalizers…
70

 Therefore the equalizer response can be expanded into two

components
 A noise-whitening component 1/G∗m(1/z∗)
 An ISI-removal component
 Minimum Mean-Square Error
(MMSE) Equalizers…
71

 The purpose of the noise-whitening filter is to whiten the noise so

that the noise component output from this filter has a constant
power spectrum.
 Since the noise input to this filter has power spectrum
N0|G m(1/z )|2, the appropriate noise-whitening filter is
1/G m(1/z ).
 The noise power spectrum at the output of the noise-whitening

filter is then N0|G m(1/z )|2/|G m(1/z )|2 = N0.

 We assume that the filter , with input vn, is a linear filter
with N = 2L +1 taps:
 Minimum Mean-Square Error
(MMSE) Equalizers…
72

 Our goal is to design the filter coefficients {wi} so as to minimize

E[dk − ]2.
 Let vT = (v[k+L], v[k+L−1], . . . , v[k−L]) = (vk+L, vk+L−1, . . . , vk−L)
be the row vector of inputs to the filter used to obtain
the filter output and define wT = (w−L, . . . ,wL) as the row
vector of filter coefficients.

 Let J be the mean square error to be minimized.

 Minimum Mean-Square Error
(MMSE) Equalizers…
73

 Define Mv = E[vvH] and vd = E[vHdk].

 The matrix Mv is an N × N Hermitian matrix, and vd is a length-N
row vector.
 Assume E|dk|2 = 1. Then the MSE - J is

 We obtain the optimal tap vector w by setting the gradient wJ

= 0 and solving for w.
 Minimum Mean-Square Error
(MMSE) Equalizers…
74

 Setting this to zero yields wTMv = vd or, equivalently, that the

optimal tap weights are given by

 Substituting in these optimal tap weights, we obtain the minimum

mean-square error as

 For an equalizer of infinite length, vT = (vn+∞, . . . , vn, . . . , vn−∞)

and wT = (w−∞, . . . , w0, . . . ,w∞).
 Minimum Mean-Square Error
(MMSE) Equalizers…
75

 Then wTMv = vd can be written as

 Taking z-transforms and noting that is the z-transform of

the filter coefficients w yields

 Solving for , we obtain

 Minimum Mean-Square Error
(MMSE) Equalizers…
76

 Since the MMSE equalizer consists of the noise-whitening filter

1/G m(1/z ) plus the ISI removal component , it follows
that the full MMSE equalizer
 Adaptive Equalizers: Training
and Tracking
77

 The ZFE and MMSE are designed based on a known value of the
combined impulse response h(t) = g(t) c(t).
 Mostly, the channel c(t) in generally not known when the
receiver is designed. So the equalizer must be tunable so it can
adjust to different values of c(t).
 Also in wireless channels, c(t) will change over time, the system
must periodically estimate the channel c(t) and update the
equalizer coefficients accordingly.
 This process is called equalizer training or adaptive
equalization.
 The equalizer can also use the detected data to adjust the
equalizer coefficients, a process known as equalizer tracking.
 Adaptive Equalizers: Training
and Tracking…
78

 Equalizers that do not use training and learn the channel

response via the detected data only are called blind
equalizers.
 During training, the coefficients of the equalizer are updated at
time k based on a known training sequence [dk−M, . . . , dk] that
has been sent over the channel.
 The length M +1 of the training sequence depends on the
number of equalizer coefficients that must be determined and
the convergence speed of the training algorithm.
 The equalizer must be retrained when the channel
decorrelates – that is, at least every Tc seconds, where Tc is the
channel coherence time.
 Adaptive Equalizers: Training
and Tracking…
79

 If the training algorithm is slow relative to the channel coherence

time then the channel may change before the equalizer can
learn the channel.
 In this case equalization is not an effective counter measure for
ISI, and some other technique (e.g., multicarrier modulation or
CDMA) is needed.
 Let { } denote the bit decisions output from the equalizer given
a transmitted training sequence {dk}.
 Our goal is to update the N equalizer coefficients at time k +1
based on the training sequence we have received up to time k.
 Adaptive Equalizers: Training
and Tracking…
80

 A wide range of algorithms exist to adapt the filter coefficients:

 MMSE(Minimum Mean Square Error): More complex but convergence
is very fast.
 LMS(Least Mean Square): Reduced complexity, but convergence is
slow for small values of . For large values of , the algorithm
becomes unstable.
 RLS(Root Least Square): Complexity and performance lies between
that of MMSE and LMS algorithms.
 The symbol decisions output from the equalizer are typically
passed through a threshold detector to round the decision to the
nearest constellation point.
 Adaptive Equalizers: Training
and Tracking…
81
 The resulting roundoff error can be used to adjust the equalizer coefficients
during data transmission, a process called equalizer tracking.

 Tracking is based two premises:

 That if the roundoff error is nonzero then the equalizer is not perfectly
trained.

 That the roundoff error can be used to adjust the channel estimate inherent in
the equalizer.

 The equalizer output bits and threshold detector output bits are used
to adjust an estimate of the equivalent lowpass composite channel F(z).

 The updated version of F(z) is then taken to equal the composite channel
and used to update the equalizer coefficients accordingly.
 LMS Algorithm
82

 The criterion used is the minimization of the mean square

error(MSE) between the desired and the actual equalizer output.
 LMS Algorithm…
83

 The prediction error is given by

ek = dk - = xk -
ek = xk – wkTyk = xk – ykTwk
 The mean squared error at time instant k is calculated as
 = E[ek*ek] = E[ek2]
 The LMS algorithm aims at minimizing the mean squared error.
|ek|2 = xk2 + wkTykykTwk -2xkykTwk
E[|ek|2] = E[xk2] + wkTE[ykykT]wk -2E[xkykT]wk
 Let p denote the cross correlation vector between the desired
response and the input signal.
 LMS Algorithm…
84

p = E[xkyk] = E[xkyk xkyk-1 xkyk-2 …. xkyk-N ] T

*
 Let R denote the input correlation matrix E[ykyk ] called the input
covariance matrix.
 Mean square error,  = E[xk ] + w Rw -2p w
2 T T

 It can be seen that the MSE is a function of wN.

 Let the cost function J(wN) denote the MSE as a function of wN.

 In order to minimize MSE

J(wN) = -2pN +2RNNwN

RNN =pN
 The above equation is called normal equation and the optimum
weight vector is obtained as = RNN-1pN
 LMS Algorithm…
85

 Inverting the square matrix requires a large number of

arithmetic operations.
 In practice, the minimization of MSE is carried out by use of
stochastic gradient descent algorithm introduced by Widrow.
 This is more commonly known as Least Mean Square (LMS)
algorithm.
 The LMS algorithm requires only 2N + 1 arithmetic operations
per iteration.
 The filter weights are updated by the update equation given
below:
 LMS Algorithm…
86

 The filter weights are updated by the update equation given

below:

where n denotes the sequence of iteration, N denotes the number

of delay stages in the equalizer and  is the step size which
controls the convergence rate and stability of the algorithm
 To prevent the adaptation from being unstable, the value of 

is chosen as 0 <  < where  is the ith Eigen value of

𝐢
RNN.