ICAI – Máster MIT

Sistemas de comunicación I

Chapter 6:
Equalization

Javier Matanza
W. Warzanskyj
Luis Cucala
1
Introduction
 The need for equalization
• We’ve previously seen that in order to minimize bandwidth while constraining
ISI the overall transmitter + channel + receiver frequency response must follow
a raised cosine shape
• The common procedure is to adjust the transmitter frequency response to a
raised cosine shape and assume a channel flat frequency response
• But if the channel response is not flat it must be flattened at the receiver. Such
flattening is called equalization

Equalization – Intuitive idea

Tx Channel: 𝐻( (𝑓) Rx Equalizer: 𝐻!" 𝑓 = #$$!(&) Output

f f f f f
this is what is this is the channel and this is what is then we place an this is the final output
transmitted frequency response received equalizer after the Rx

3
 Classic transmission support for equalization
• For the receiver to estimate the distortion introduced by the channel, the
transmitter sends some known signals, called reference signals.
• There are two main types of reference signals; preambles and pilots:

Preambles Pilots

• When the preambles consist of predefined sequences, they are sometimes

called training sequences (this is the GSM case)
4
 Types of equalizers
Depending on their structure, equalizers can be classified in :
• Non-linear Equalizers:
o For severe ISI
o They include information feedback loops
o They will not be treated in this course
• Linear Equalizers
o They are the most common equalizers
o Two types of linear equalizers will be addressed in the lesson:

• Zero-forcing Equalizer
• MMSE Equalizer
o There are many others (e.g., Rake, for CDMA signals), but the MMSE equalizer, and
its variants, is the most widely used, because of its simplicity

6
2

Zero-forcing equalizer
 Zero-forcing equalizer 𝑠[𝑛] Channel 𝑦[𝑛]

• ISI is generated by previous symbols’ non-zero values at the symbol sampling

time ,
𝑦 𝑡 = ) 𝑎) · 𝑠 𝑡 − 𝑛𝑇 + 𝑤(𝑡)
)*+,

𝑦[𝑘] = 𝑦 𝑡 = 𝑘𝑇 = 𝑎- 5 𝑠[0] + ) 𝑎) · 𝑠 𝑛 − 𝑘 + 𝑤[𝑘]

).- noise
Desired output ISI: Intersymbol interference
• The zero-forcing equalizer is implemented as a K-coefficients FIR filter that for
each symbol forces a zero in its output at the sampling times different to the
symbol sampling instant :
1+# 1+# 1+#
675 34)
𝑔/ 𝑡 = ) 𝜔) · 𝛿 𝑡 − 𝑛𝑇 𝐺/ 𝑓 = ) 𝜔) · 𝑒 +234&)5 𝐺/ 𝑘 = ) 𝜔) · 𝑒 +2·- 9
)*0 )*0 )*0

• The filter sequence is defined as 𝜔 𝑛 = {𝜔! , 𝜔" , … 𝜔# }. The filter length K is

chosen sufficiently large so that all ISI effects are considered

*N is the DFT length

8
 𝑠[𝑛] 𝑦[𝑛] 𝑠[𝑛]
̂
Zero-forcing equalizer Channel Equalizer

• When the input is a single symbol s(𝑡) the equalizer output can be expressed as:
1+# 1+#

𝑠̂ 𝑡 = 𝑦 𝑡 ∗ 𝑔/ 𝑡 = 𝑦 𝑡 ∗ ( 𝜔) · 𝛿 𝑡 − 𝑛𝑇 = ( 𝜔) · 𝑦 𝑡 − 𝑛𝑇
)*0 )*0

• In order to avoid ISI, the equalizer is designed so that its sampled output is:
1+#
1 𝑖𝑓 𝑚 = 𝑛
𝑠̂ 𝑚𝑇 = ( 𝜔) · 𝑠 𝑚𝑇 − 𝑛𝑇 = 0
0 𝑖𝑓 𝑚 ≠ 𝑛
)*0

• Zero-forcing equalizer limitation:

• Forcing-zero crossings is equivalent in the frequency domain to multiplying the
channel transfer function by its inverse, so that the resulting transfer function
becomes effectively flat
• But if the channel attenuation is high at some frequency, inversion implies amplifying
the noise excessively, which should be avoided to keep the SNR at that frequency
within system design levels
equalizer signal + noise
signal noise
channel
𝑓 𝑓 𝑓 𝑓 𝑓
9
 Zero-forcing equalization in OFDM
1. Assuming channel coherence time is longer than the frame duration, a known
pilot symbol is sent at the beginning of every frame: the pilot’s 𝐾 subcarriers’
complex amplitudes vector 𝑋 ! = 𝑋" , 𝑋# , … , 𝑋$ is known
pilot symbol 2. After transmission and the DFT, the pilot vector is converted
𝑓
into 𝑌 ! = 𝑌" , 𝑌# , … , 𝑌$
&:
3. 𝐾 equalization coefficients are calculated as 𝜔% =
':

4. The next symbols in the frame are equalized as follows:

frame duration 𝑡 𝑋% )* = 𝜔% . 𝑋%

DFT 𝜔!
Time Freq
0 0
From channel ·
·
1 X[1] X
A/D S/P · · QAM demodulators
·
· · · & mux
N-1 K X[K] X

subcarriers’ complex amplitudes 𝑋 " = 𝑋# , 𝑋$ , … , 𝑋%

Note: a frame is an organized set of symbols 10
3

MMSE forcing equalizer

 MMSE Minimum Mean Squared Error
• A different way to undo the channel effects is to minimize the difference
between the desired signal and the equalized one. This equalizer is better than
the zero forcing one in the sense that it does not allow excessive noise
amplification
𝑦[𝑛]
𝑠[𝑛] Channel Equalizer 𝑠[𝑛]
̂

• A Minimum Mean Squared Error (MMSE) equalizer seeks to minimize the error
signal
Error ≜ 𝑒 𝑛 = 𝑠 𝑛 − 𝑠[𝑛]
̂
𝑠 𝑛 : a sequence known at the receiver (preamble, training sequence or pilot)
𝑠[𝑛]
̂ : the sequence coming out of the equalizer when the input to the channel is 𝑠 𝑛

• More precisely, MMSE minimizes the Mean Squared Error (MSE):

MSE ≜ 𝐸 𝑒 𝑛 " = 𝐸 𝑠 𝑛 − 𝑠̂ 𝑛 "

12
 The equalized signal
• The equalized signal is
1+#
> 5 · 𝑦[𝑛]
𝑠̂ 𝑛 = 𝑤[𝑛] ∗ 𝑦[𝑛] = ) 𝑤 𝑘 · 𝑦 𝑛 − 𝑘 = 𝑤 ?
-*0

> 5 = [𝑤0, 𝑤#, … , 𝑤1+#] (equalizer coefficients)

𝑤
𝑦𝑛
𝑦 𝑛−1
𝑦[𝑛]
? = ? 5 = 𝑦 𝑛 ,𝑦 𝑛 − 1 ,…,𝑦 𝑛 − 𝐾 + 1
𝑦[𝑛]
………
𝑦[𝑛 − 𝐾 + 1

• Understanding the expression:

• An equalized sample, 𝑠[𝑛],
̂ depends on K received samples, 𝑦 𝑛 , 𝑦 𝑛 − 1 , … , 𝑦[𝑛 − 𝑘 +
1], weighted by K equalizer coefficients, or taps, 𝑤0 , 𝑤# , … , 𝑤1+#
• The current sample out of the equalizer is the current input weighted by the first
coefficient, plus the previous input weighted by the second tap, and so on till the last pair
𝑠̂ 1 = 𝑤0 · 𝑦 1
𝑠̂ 2 = 𝑤0 · 𝑦 2 + 𝑤# · 𝑦 1
𝑠̂ 𝐾 = 𝑤0 · 𝑦 𝐾 + 𝑤# · 𝑦 𝐾 − 1 + . . . + 𝑤1+# · 𝑦[1]
13
 Defining the correlation matrixes
• By definition:
𝑅F; ≜ 𝐸 𝑦[𝑛]
? ? 5
· 𝑦[𝑛] Auto-correlation matrix of signal 𝑦 𝑛 , a KxK symmetric matrix
𝑅?<; ≜ 𝐸 𝑠 𝑛 · 𝑦[𝑛
? Cross-correlation matrix between signals 𝑦 𝑛 and s[𝑛], Kx1
matrix
• Example when two samples are taken into account:
𝑦[𝑛] 𝑦 𝑛 .𝑦 𝑛 𝑦 𝑛 .𝑦 𝑛 − 1 𝑅' 0 𝑅' 1
𝑅2' = 𝐸 𝑦[𝑛]
% % ( =𝐸
· 𝑦[𝑛] · 𝑦[𝑛] 𝑦[𝑛 − 1 = 𝐸{ }=
𝑦 𝑛−1 𝑦 𝑛 − 1 .𝑦 𝑛 𝑦 𝑛 − 1 .𝑦 𝑛 − 1 𝑅' 1 𝑅' 0
(
𝑅%&' = 𝐸 𝑠 𝑛 · 𝑦[𝑛]
% ( =𝐸 𝑠 𝑛 · 𝑦 𝑛 𝑦 𝑛−1 = 𝑅&' 0 𝑅&' 1

• Example when three samples are taken into account:

𝑦𝑛 𝑅' 0 𝑅' 1 𝑅' 2
𝑅2' = 𝐸 𝑦[𝑛]
% % ( =𝐸
· 𝑦[𝑛] 𝑦[𝑛 − 1] · 𝑦 𝑛 𝑦[𝑛 − 1 𝑦 𝑛 − 2 ] = 𝑅' 1 𝑅' 0 𝑅' 1
𝑦 𝑛−2 𝑅' 2 𝑅' 1 𝑅' 0
(
𝑅%&' = 𝐸 𝑠 𝑛 · 𝑦[𝑛]
% ( = 𝐸 𝑠 𝑛 · 𝑦 𝑛 𝑦 𝑛 − 1 𝑦[𝑛 − 2 = 𝑅&' 0 𝑅&' 1 𝑅&' 2

• 𝑅.+ is a symmetric matrix, and all the elements of its main diagonal equal

14
 Example of correlation matrix calculations:
copy and execute the code % Example MMSE. We choose a 3 tap equalizer
%
close all; clear; format compact;
%
fs = 8; Ts = 1/fs;
f2 = 0.5; f3 = 1.8;
Expectation value, 𝐸 N = 64;
t = 0:Ts:(N-1)*Ts;
s = cos(2*pi*f2*t) + cos(2*pi*f3*t);
If the signal is a sequence whose y = 0.5*cos(2*pi*f2*t) + 1.5*cos(2*pi*f3*t);
%
statistics are not known, 𝐸 is % Building of vectors yn and sn
calculated as average value, e.g., yn = fliplr(y); sn = fliplr(s);
%
*+$
)
1 % Matrix Ry: 3x3
𝐸 𝑠𝑛 → 9 𝑠,) Ry0 = mean(yn.*yn);
𝑁 yn_1 = yn(2:N);
#
yn_2 = yn(3:N);
Ry1 = mean(yn(1:N-1).*yn_1);
Ry2 = mean(yn(1:N-2).*yn_2);
1.2848 0.3088 −0.9925 %
𝑅4$ = 0.3088 1.2848 0.3088 %
Ry = [Ry0 Ry1 Ry2; Ry1 Ry0 Ry1; Ry2 Ry1 Ry0]

−0.9925 0.3088 1.2848 % Matrix Rsy: 3x1

Rsy0 = mean(sn.*yn);
Rsy1 = mean(sn(1:N-1).*yn_1);
1.0433 Rsy2 = mean(sn(1:N-2).*yn_2);
𝑅?%$ = 0.3717 %
Rsy = [Rsy0 Rsy1 Rsy2]'
−0.5507

15
 The MSE

𝑠[𝑛]
̂ Demodulate +
𝑦[𝑛] Equalizer
Detect
𝑒[𝑛]
- 𝑠[𝑛]

MSE ≜ 𝐸 𝑒 𝑛 3 = 𝐸 𝑠 𝑛 − 𝑠̂ 𝑛 3 =𝐸 𝑠 𝑛 3 + 𝐸 𝑠̂ 𝑛 3 − 2 · 𝐸 𝑠 𝑛 · 𝑠̂ 𝑛

𝐸 𝑠̂ 𝑛 3 > 5 · 𝑦[𝑛]
=𝐸 𝑤 ? 3 > 5 · 𝑦[𝑛]
=𝐸 𝑤 ? > 5 · 𝑦[𝑛]
𝑤 ? 5 > 5 · 𝑦? 𝑛
=𝐸 𝑤 ? 5 . 𝑤)
(𝑦[𝑛] > =

> 5 · 𝐸 𝑦[𝑛]
=𝑤 ? ? 5 ·𝑤
· 𝑦[𝑛] > 5 · 𝑅F; · 𝑤
> =𝑤 >
5
> 5 · 𝑦[𝑛]
𝐸 𝑠 𝑛 · 𝑠̂ 𝑛 = 𝐸 𝑠 𝑛 · 𝑤 ? ? 5·𝑤
= 𝐸 𝑠 𝑛 · 𝑦[𝑛] > = 𝐸 𝑠 𝑛 · 𝑦[𝑛] > = 𝑅?<; · 𝑤
? 5 ·𝑤 >

5
MSE ≜ 𝐸 𝑒 𝑛 3 =𝐸 𝑠 𝑛 3 > 5 · 𝑅F; · 𝑤
+𝑤 > − 2 · 𝑅?<; . 𝑤
>

A & · 𝑦[𝑛]
Note: 𝑤 ? A & · 𝑦? 𝑛 = 𝑤
is a 1x1 matrix, then 𝑤 A & · 𝑦? 𝑛 &
? & ·𝑤
= 𝑦[𝑛] A 16
 MSE representation
• A reduced expression of the MSE is:
5
> 5 · 𝑅F; · 𝑤
> = 𝜎<3 + 𝑤
MSE ≜ 𝐽 𝑤 > − 2 · 𝑅?<; · 𝑤
>

• 𝐽 𝑤
> is a function that expresses the mean power of the error as a
function of the K equalizer’s taps, it’s a multivariable function.

• Graphical representation for K=2

• The function must have a minimum, 500

otherwise there must exist a point so 400
that 𝐽(𝑤0 , 𝑤# ) is negative, and it is

J(w0,w1)
300
impossible by definition. 200
• The equalizer’s objective is to find 100
the pair of coefficients (𝑤0 , 𝑤# ) that 0
corresponds to the minimum 20
20
0 10
0
-10
w1 -20 -20 w0

17
Appendix

MMSE adaptative equalizer

 MMSE Adaptative Filtering
• The channel usually does not stay constant over time.
• Adaptive filtering changes the equalizer coefficients dynamically seeking the
minimization of the error signal.
• A signal pilot signal is sent periodically to be used as a known reference 𝐬[𝐧].

𝑠[𝑛]
̂ (received pilot)
𝑦[𝑛] Equalizer Demod + Detect
𝑒[𝑛]
- 𝑠[𝑛] (expected pilot, known by the receiver)

• However, computing the Weiner’s coefficient for every incoming sample is

computationally complex or, more likely, matrix inversion is unstable, and
degrades the signal to noise ratio
• A simpler algorithm is needed to perform this procedure adaptively.

20
Steepest Descent Equalization Algorithm
(Algoritmo del gradiente)

• Avoids the computation of an inverse matrix for every incoming sample.

• Basic idea: Compute the gradient at every point and move towards the
opposite direction using small steps.
• The weights vector is updated in steps using the relationship:
𝑤 𝑘 + 1 = 𝑤 𝑘 − 𝜇·∆𝐽 𝑤[𝑘 + 1]

Moving towards the opposite direction of the growth

• 𝜇 is a parameter that controls the speed of the iterative process

o ↑ 𝜇 ⇒ Fast convergence. May cause instability.
o ↓ 𝜇 ⇒ Slow convergence. Higher stability.

21
Least Mean Square (LMS) algorithm
• A problem with the Steepest Descent algorithm is that the gradient might
not be analytically available.
• The LMS algorithm uses an estimation of the gradient given by
∆𝐽 𝑤 ≈ −2𝑒 𝑛 · 𝑦[𝑛]

• Coefficients are updated using:

𝑤 𝑘 + 1 = 𝑤 𝑘 + 2𝜇 · 𝑒 𝑛 · 𝑦[𝑛]
where 𝑦[𝑛] is a vector containing the K latest received samples
• The algorithm for updating 𝑤 becomes
1. Update 𝑦 n with a new received sample
2. Compute equalized signal: 𝑠̂ 𝑛 = 𝑤 𝑘 5 · 𝑦[𝑛]
3. Compute error signal: 𝑒 𝑛 = 𝑠 𝑛 − 𝑠̂ 𝑛
4. Update weights:𝑤 𝑘 + 1 = 𝑤 𝑘 + 2𝜇 · 𝑒 𝑛 · 𝑦[𝑛]

22
 Least mean Square Algorithm

20
Reference signal 20
15 Received Signal
15
Equalized Signal
10 10

5
5
0

w1
0
-5
-5 -10

-10 -15

-20
-15 -20 -10 0 10 20
0 50 100 150 w0
Iteration

23