Equalization

Prof. David Johns

University of Toronto

([email protected])
(www.eecg.toronto.edu/~johns)

slide 1 of
University of 70
Toronto © D.A. Johns, 1997

Adaptive Filter Introduction

• Adaptive filters are used in:

• Noise cancellation
• Echo cancellation
• Sinusoidal enhancement (or rejection)
• Beamforming
• Equalization

• Adaptive equalization for data

communications proposed by R.W. Lucky at
Bell Labs in 1965.
• LMS algorithm developed by Widrow and Hoff
in 60s for neural network adaptation

slide 2 of
University of 70
Toronto © D.A. Johns, 1997
 Adaptive Filter Introduction
• A typical adaptive system consists of the
following two-input, two output system
(n)

+
y(n) -
u(n) H(z) e(n)

y(n)

adaptive
algorithm

• u(n) and y(n) are the

filter’s input and
output
• (n) and e(n) are the
slide 3 of
University of
reference
Toronto
and error 70
© D.A. Johns, 1997

signals

Adaptive Filter Goal

• Find a set of filter coefficients to minimize the
power of the error signal, e(n) .
• Normally assume the time-constant of the
adaptive algorithm is much slower than those of
the filter, H(z) .
• If it were instantaneous, it could always set y(n)
equal to (n) and the error would be zero (this is
useless)

• Think of adaptive algorithm as an optimizer

which finds the best set of fixed filter
coefficients that
minimizes the power of the error signal.
University of slide 4 of
70
Toronto © D.A. Johns, 1997
 Noise (and Echo) Cancellation
signal
+
+
H1(z)  noise
noise (n)
H1(z)
+ e(n) = signal
H2(z) H(z) -
u(n) y(n) = H1(z)  noise

H(z) = H1(z)  H2(z)

• Useful in cockpit noise cancelling, fetal

heart monitoring, acoustic noise
cancelling, echo
cancelling, etc.
slide 5 of
University of 70
Toronto © D.A. Johns, 1997

Sinusoidal Enhancement (or Rejection)

sinusoid
+ (n)
noise
y(n) +
 H(z) - noise
u(n) sinusoid
fixed delay
e(n)
• The sinusoid’s frequency and
amplitude are unknown.
• If H(z) is adjusted such that its phase plus the
delay equals 360 degrees at the sinusoid’s
frequency, the sinusoid is cancelled while the
noise is passed.
• The “noise” might be a broadband signal
which should be recovered.
slide 6 of
University of 70
Toronto © D.A. Johns, 1997
 Adaptation Algorithm
• Optimization might be performed by:
• perturb some coefficient in H(z) and check whether the
power of the error signal increased or decreased.
• If it decreased, go on to the next coefficient.
• If it increased, switch the sign of the coefficient change and
go on to the next coefficient.
• Repeat this procedure until the error signal is minimized.

• This approach is a steepest-descent algorithm

but is slow and not very accurate.

• The LMS (Least-Mean-Square) algorithm is

also a steepest-descent algorithm but is more
accurate and simpler to realize

slide 7 of
University of 70
Toronto © D.A. Johns, 1997

Steepest-Descent Algorithm

• Minimize the power of the error signal,

Ee 2 (n)

• General steepest-descent for filter coefficient

pi(n) :
2
pi(n + 1) = pi(n) –   Ee (n)
-----------------------
pi 

• Here   0 and controls the adaptation

rate

slide 8 of
University of 70
Toronto © D.A. Johns, 1997
 Steepest Descent Algorithm
• In the one-dimensional case

Ee 2 (n)

Ee2(n) 
0
pi

pi
*
pi pi(2) pi(1) pi(0)

slide 9 of
University of 70
Toronto © D.A. Johns, 1997

Steepest-Descent Algorithm
• In the two-dimensional case
p2

*
p
2

p1
Ee 2 (n) *
(out of p1
page)
• Steepest-descent path follows
perpendicular to tangents of the contour
lines.
slide 10 of
University of 70
Toronto © D.A. Johns, 1997
 LMS Algorithm
• Replace expected error squared with
instantaneous error squared. Let adaptation
time smooth out result.
2
pi(n + 1) = pi(n) –  e (n)
-------------

- pi

p i(n + 1) = pi (n) – 2e(n) ---e---

 pi 
( n )
-- --- -
• and since e(n) = (n) – y(n) , we have

pi(n + 1) = pi(n) + 2e(n)i(n) where i = y(n)  pi

• e(n) and i(n) are uncorrelated after

convergence.
slide 11 of
University of 70
Toronto © D.A. Johns, 1997

Variants of the LMS Algorithm

• To reduce implementation complexity,
variants are taking the sign of e(n) and/or i(n) .

• LMS — pi(n + 1) = pi(n) + 2e(n)  i(n)

• Sign-data LMS — pi(n + 1) = pi(n) + 2e(n)  sgn i(n)

• Sign-error LMS — pi(n + 1) = pi(n) + 2 sgn e(n) 

i(n)
• Sign-sign LMS — i (n + 1) = pi(n) + 2sgne(n)  sgn i(n)

• However, the sign-data and sign-sign

algorithms
have gradient misadjustment — may not
converge!
• These LMS algorithms have different dc
offset implications
University of in analog realizations. slide 12 of
70
Toronto © D.A. Johns, 1997
 Obtaining Gradient Signals
m pi n
hum(n) hny(n)

u(n) H(z) y(n)

 i(n) = --- y ( n )
(n)  h
pi--- -- ---ny- = h um
(n) 
• H(z) is a LTIu(n)
system where the signal-flow-
graph arm corresponding to coefficient pi is
shown explicitly.
• hum(n) is the impulse response of from u to m
• The gradient signal with respect to element pi
is the convolution of u(n) with hum(n) convolved
slide 13 of
University of
withToronto
hny(n). 70
© D.A. Johns, 1997

Gradient Example
G1

vlp(t)

u(t)
-1 G2 1

vbp(t)
y(t)
G1 G3

1
vlp(t)
-1 G2 1

G3 ---
G1
y---
 y ( t ) lp  y ( t ) bp
-- --- -- -- =
-G --- --- -- -- - =
G
2 3
(--
–v (t) –v (t)
t--)-

slide 14 of
University of 70
Toronto © D.A. Johns, 1997
 Adaptive Linear Combiner
p1(n)
x1(n)

(n)
p2(n)
x2(n)

N +
- e(n)
state
u(n) generator y(n)

pN(n)
xN(n)
y(n) = p (n)x (n)
i i

 y ( n i)
---
p --- -- --- -
i
= x (n)
often, a tapped delay Y
H(z) = U(z)
- ---
line
( z )
-- --- -

slide 15 of
University of 70
Toronto © D.A. Johns, 1997

Adaptive Linear Combiner

• The gradient signals are simply the state
signals (1)

pi(n + 1) = pi(n) + 2e(n)xi(n)

• Only the zeros of the filter are being

adjusted.
• There is no need to check that for filter
stability
(though the adaptive algorithm could go
unstable if 
is too large).
• The performance surface is guaranteed
unimodal (i.e. there is only one minimum so no
need to worry about being stuck in a local
minimum).
slide 16 of
University of
• The performance surface becomes ill-
Toronto
70
© D.A. Johns, 1997

conditioned as the state-signals become

correlated (or have large
 Performance Surface
• Correlation of two states is determined by
multiplying the two signals together and
averaging the output.
• Uncorrelated (and equal power) states
result in a “hyper-paraboloid” performance
surface — good adaptation rate.
• Highly-correlated states imply an ill-
conditioned
performance surface — more residual mean-
2
square error and longer adaptation time.

p
* p1
Ee 2 (n) *
(out of p1
page)

slide 17 of
University of 70
Toronto © D.A. Johns, 1997

Adaptation Rate
• Quantify performance surface — state-
correlation matrix

Ex 1 x1  Ex 1 x2  Ex 1 x3 

R Ex 2 x1  Ex 2 x2  Ex 2 x3 

Ex 3 x1  Ex 3 x2  Ex 3 x3 
• Eigenvalues, i , of R are all positive real —
indicate curvature along the principle axes.
• For adaptation stability, 0    1
- - --- -- but
adaptation rate
max

is determined by least steepest curvature, min .

• Eigenvalue spread indicates performance
surface conditioning.
slide 18 of
University of 70
Toronto © D.A. Johns, 1997
 Adaptation Rate

• Adaptation rate might be 100 to 1000 times

slower than time-constants in programmable
filter.
• Typically use same  for all coefficient
parameters since orientation of performance
surface not usually known.
• A large value of  results in a larger
coefficient “bounce”.
• A small value of  results in slow
adaptation
• Often “gear-shift”  — use a large value at
start-up then switch to a smaller value during
steady-state. slide 19 of
University of 70
• Might need to detect if one should “gear- ©
Toronto D.A. Johns, 1997

shift” again.

Adaptive IIR Filtering

• The poles (and often the zeros) are adjusted
—
useful in applications with long impulse
responses.
• Stability check needed for the adaptive filter
itself to ensure the poles do not go outside the
unit circle for too long a time (or perhaps at
all).
• In general, a multi-modal performance
surface occurs. Can get stuck in local
minimum.
• However, if the order of the adaptive filter is
greater than the order of the system being
matched (and all poles and zeros are being
adapted) —ofthe
University slide 20 of
70
performance
Toronto surface is unimodal. © D.A. Johns, 1997

• To obtain the gradient signals for poles, extra

filters are generally required.
 Adaptive IIR Filtering
• Direct-form structure needs only one
additional filter to obtain all the gradient
signals.
• However, choice of structure for
programmable filter is VERY important —
sensitive structures tend to
have ill-conditioned performance surfaces.
• Equation error structure has unimodal
performance surface but has a bias.
• SHARF (simplified hyperstable adaptive
recursive filter) — the error signal is filtered
to guarantee
adaptation — needs to meet a strictly-
positive-real condition
• There are of
University few commercial use of adaptive slide 21 of
70
Toronto © D.A. Johns, 1997

IIR filters

Digital Adaptive Filters

• FIR tapped delay line is the most
common
p1(n)
x 1(n)
u(n)
(n)
–1
p2(n)
z
x2(n)
+
- e(n)
–1
z y(n)

pN(n)
–1
z xN(n)
y(n) = p (n)x (n)
i i

 y ( n i)
--- --- -- --- -
p i
= x (n)

slide 22 of
University of 70
Toronto © D.A. Johns, 1997
 FIR Adaptive Filters

• All poles at z = 0 and zeros only adapted.

• Special case of an adaptive linear combiner
• Unimodal performance surface
• States are uncorrelated and equal power if
input signal is white — hyper-paraboloid
• If not sure about correlation matrix, can
guarantee adaptation stability by choosing
1
0 -
# of tapsinput signal power

• Usually need an AGC so signal power is

known.

slide 23 of
University of 70
Toronto © D.A. Johns, 1997

FIR Adaptive Filter

• Coefficient word length typically 2 + 0.5log2(# of
bits longer than “bit-equivalent” dynamic
taps)
range
• Example: 6-bit input with 8-tap FIR might
have 10-bit coefficient word lengths.
• Example: 12-bit input with 128-tap FIR might
have 18-bit coefficient word lengths for 72 dB
output SNR.
• Requires multiplies in filter and adaptation
algorithm (unless an LMS variant used or slow
adaptation rate)
— twice the complexity of FIR fixed filter.

slide 24 of
University of 70
Toronto © D.A. Johns, 1997
 Equalization — Training Sequence
u(n) y(n) output data
 Htc(z) H(z) 1
1
known FFE
input data
e(n)
regenerated
delayed FFE = Feed Forward
input data (n)
Equalizer

• The reference signal, (n) is equal to a

delayed version of the transmitted data
• The training pattern should be chosen so as
to ease adaptation — pseudorandom is
common.
• Above is a feedforward equalizer (FFE) since
y(n)is not directly created using derived output
dataUniversity of slide 25 of
70
Toronto © D.A. Johns, 1997

FFE Example
• Suppose channel, Htc(z), has impulse
response 0.3, 1.0, -0.2, 0.1, 0.0, 0.0

tim
e

• If FFE is a 3-tap FIR filter with

y(n) = p1 u(n) + p2 u(n – 1) + p3 u(n – 2) (2)

• Want to force y(1) = 0 , y(2) = 1 , y(3) = 0

• Not possible to force all other y(n) =
0

slide 26 of
University of 70
Toronto © D.A. Johns, 1997
 FFE Example
y(1) = 0 = 1.0p1 + 0.3p2 + 0.0p3

y(2) = 1 = –0.2p1 + 1.0p2 + 0.3p3

y(3) = 0 = 0.1p1 + –0.2p 2 + 1.0p3
(3)

• Solving results in p1 = –0.266 , p2 = 0.886 , p3 = 0.204

• Now the impulse response through both
channel and equalizer is: 0.0, -0.08, 0.0,
1.0, 0.0, 0.05, 0.02, ...
1

tim
e

slide 27 of
University of 70
Toronto © D.A. Johns, 1997

FFE Example
• Although ISI reduced around peak,
introduction of slight ISI at other points
(better overall)
• Above is a “zero-forcing” equalizer — usually
boosts noise too much
• An LMS adaptive equalizer minimizes the
mean
squared error signal (i.e. find low ISI and low
noise)
• In other words, do not boost noise at
expense of leaving some residual
ISI

slide 28 of
University of 70
Toronto © D.A. Johns, 1997
 Equalization — Decision-Directed

 u(n) y(n) output data

1 Htc(z) H(z) 1
input data FFE (n)

e(n)

• After training, the channel might change

during data transmission so adaptation should
be continued.
• The reference signal is equal to the recovered
output data.
• As much as 10% of decisions might be in
error but correct adaptation will occur

slide 29 of
University of 70
Toronto © D.A. Johns, 1997

Equalization — Decision-Feedback
 y(n) output data
1 Htc(z) 1
input data (n)
yDFE(n) H2(z)

e(n) DFE

• Decision-feedback equalizers make use of (n)

in directly creating y(n) .
• They enhance noise less as the derived input
data is used to cancel ISI
• The error signal can be obtained from
either a training sequence or decision-
directed.
slide 30 of
University of 70
Toronto © D.A. Johns, 1997
 DFE Example
• Assume signals 0 and 1 (rather than -1
and +1) (makes examples easier to
explain)
• Suppose channel, Htc(z), has impulse
response 0.0, 1.0, -0.2, 0.1, 0.0, 0.0

tim
e
• If DFE is a 2-tap FIR filter with
yDFE(n) = 0.2(n – 1) + –0.1(n – 2) (4)

• Input to slicer is now 0.0, 1.0, 0.0, 0.0

0.0 0.0
slide 31 of
University of 70
Toronto © D.A. Johns, 1997

FFE and DFE Combined

 u(n) y(n) output data

1 Htc(z) H1(z) 1
input data (n)
e(n) FFE
yDFE(n) H2(z)

e(n) DFE

• Assuming correct operation, output data =

input data
• e(n) same for both FFE and DFE
• e(n) can be either training or decision directed

slide 32 of
University of 70
Toronto © D.A. Johns, 1997
 FFE and DFE Combined
Model as:
nnoise(n)
FFE
x(n) y(n) output data
 Htc(z) H1(z) 1
input
1 data (n)
H2(z) yDFE(n)

DFE

Y
- -- =
(5)
N
H1
Y
-X -- = H t c H 1 +
(6)

H2
• When Htc small, make H2 = 1 (rather
than H1   )

slide 33 of
University of 70
Toronto © D.A. Johns, 1997

DFE and FFE Combined

1

tim
e

precursor ISI postcursor ISI

• FFE can deal with precursor ISI and postcursor

ISI
• DFE can only deal with postcursor ISI
• However, FFE enhances noise while DFE does
not
When both adapt
• FFE trys to add little boost by pushing
precursor into
University of
postcursor ISI (allpass) slide 34 of
70
Toronto © D.A. Johns, 1997
 Equalization — Decision-Feedback

• The multipliers in the decision feedback

equalizer
can be simple since received data is small
number of levels (i.e. +1, 0, -1) — can use
more taps if needed.
• An error in the decision will propagate in
the ISI cancellation — error propagation
• More difficult if Viterbi detection used since
output not known until about 16 sample
periods later (need early estimates).
• Performance surface might be multi-modal
with local minimum if changing DFE affects
output data
slide 35 of
University of 70
Toronto © D.A. Johns, 1997

Fractionally-Spaced FFE
• Feed forward filter is often a FFE sampled at
2 or 3 times symbol-rate — fractionally-
spaced
(i.e. sampled at T  2 or at T  3 )
• Advantages:
— Allows the matched filter to be realized
digitally and also adapt for channel
variations (not possible in symbol-rate
sampling)
— Also allows for simpler timing
recovery
schemes (FFE can take care of phase
recovery)
• Disadvantage
slide 36 of
University of
Costly to
Toronto implement — full and higher 70
© D.A. Johns, 1997

speed multiplies, also higher speed A/D

needed.
 dc Recovery (Baseline Wander)
• Wired channels often ac coupled
• Reduces dynamic range of front-end circuitry
and
also requires some correction if not accounted
for in transmission line-code
+2
+1 +1
-1 -1

• Front end may have to be able to accomodate

twice the input range!
• DFE can restore baseline wander - lower
frequency pole implies longer DFE
• Can use line codes with no dc content

slide 37 of
University of 70
Toronto © D.A. Johns, 1997

Baseline Wander Correction #1

DFE Based
• Treat baseline wander as postcursor
interference
•--z--May
– 1 require 1a –1long
-- ---- ---- = 1 – - -z – -1-z–2DFE
– -1-
–3 IMPULSE INPUT
z –
z – 0.5 2 4 8
0 1 -0.5 -0.25 -0.125 -0.06 ... 0 1 0 0 0 0 ...

0 1 0 0 0 0 ...
0 1 0 0 0 0 ...

DFE
0 0 0.5 0.25 0.125 0.06 ...
1 1 1
- -z–1 + - -z–2 + - -z–3 +

2
4
slide 38 of
University of
Toronto 8 70
© D.A. Johns, 1997
 Baseline Wander Correction #1

DFE Based

z – 1 1 1 1
- - ---- ---- ---- = 1 – - -z–1 – - -z–2 – - - STEP INPUT
z–3 – 
z – 0.5 2 4 8
0 1 0.5 0.25 0.125 0.06 ... 0 1 1 1 1 1 ...

0 1 1 1 1 1 ...
0 1 1 1 1 1 ...

DFE
0 0 0.5 0.75 0.875 0.938 ...
1 1 1
- -z–1 + - -z–2 + - -z–3 +

2
4
8 slide 39 of
University of 70
Toronto © D.A. Johns, 1997

Baseline Wander Correction #2

Analog dc restore
STEP INPUT

0 1 1 1 1 1 ...

0 1 1 1 1 1 ...
0 1 1 1 1 1 ...

• Equivalent to an analog DFE

• Needs to match RC time
constants

slide 40 of
University of 70
Toronto © D.A. Johns, 1997
 Baseline Wander Correction #3
Error Feedback

y(n) output data

1
0 1 1 1 1 1 ... (n)
1
z – 1- e(n)
integrator

• Integrator time-constant should be faster

than ac coupling time-constant
• Effectively forces error to zero with
feedback
• May be difficult to stablilize if too much
in loop (i.e. AGC, A/D, FFE, etc)
slide 41 of
University of 70
Toronto © D.A. Johns, 1997

Analog Equalization

slide 42 of
University of 70
Toronto © D.A. Johns, 1997
 Analog Filters
Switched-capacitor filters
+ Accurate transfer-functions
+ High linearity, good noise performance
- Limited in speed
- Requires anti-aliasing filters
Continuous-time filters
-Moderate transfer-function accuracy
(requires tuning circuitry)
- Moderate linearity
+ High-speed
+ Good noise performance

slide 43 of
University of 70
Toronto © D.A. Johns, 1997

Adaptive Linear Combiner

p1(t)
x1(t)

(t)
p2(t)
x2(t)

N +
- e(t)
state
u(t) generator y(t)

pN(t)
xN(t)
y(t) = p (t)x (t)
i i

 y ( t i)
--- --- -- -- -
p i
= x (t)
Y
H(s) = U(s)
- ---
( s )
-- --- -
slide 44 of
University of 70
Toronto © D.A. Johns, 1997
 Adaptive Linear Combiner
• The gradient signals are simply the state
signals
• If coeff are updated in discrete-time
(7)
pi(n + 1) = pi(n) + 2e(n)xi(n)

• If coeff are updated in cont-time

 (8)

pi(t) =  2e(t)xi(t)dt
• Only the zeros of the filter are being adjusted.
0

• There is no need to check that for filter

stability
(though the adaptive algorithm could go
unstable if 
is too large).
slide 45 of
University of 70
Toronto © D.A. Johns, 1997

Adaptive Linear Combiner

• The performance surface is guaranteed
unimodal (i.e. there is only one minimum so no
need to worry about being stuck in a local
minimum).
• The performance surface becomes ill-
conditioned as the state-signals become
correlated (or have large
power variations).
Analog Adaptive Linear Combiner
• Better to use input summing rather
than output summing to maintain high
speed operation
• Requires extra gradient filter to obtain
gradients
slide 46 of
University of 70
Toronto © D.A. Johns, 1997
 Analog Adaptive Filters
Analog Equalization Advantages
• Can eliminate A/D converter
• Reduce A/D specs if partial equalization done
first
• If continuous-time, no anti-aliasing filter
needed
• Typically consumes less power and silicon for
high- frequency low-resolution applications.
Disadvantages
• Long design time (difficult to “shrink” to new
process)
• More difficult testing
• DC offsets can result in large MSE (discussed
later).
University of slide 47 of
70
Toronto © D.A. Johns, 1997

Analog Adaptive Filter Structures

• Tapped delay lines are difficult to
implement in analog.
To obtain uncorrelated states:
• Can use Laguerre structure — cascade of
allpass first-order filters — poles all fixed at
one location on real axis

• For arbitrary pole locations, can use

orthonormal filter structure to obtain
uncorrelated filter states
[Johns, CAS, 1989].

slide 48 of
University of 70
Toronto © D.A. Johns, 1997
 Orthonormal Ladder Structure

x2(t) – 
x4(t)
2
1 3

1/s 1/s 1/s –4

1/s

2
x 1(t) –1 x3(t) –3
 in
u(t)

• For white noise input, all states are

uncorrelated and have equal power.

slide 49 of
University of 70
Toronto © D.A. Johns, 1997

Analog’s Big Advantage

• In digital filters, programmable filter has
about same complexity as a fixed filter (if not
power of 2 coeff).
• In analog, arbitrary fixed coeff come for
free (use element sizing) but programming
adds complexity.
• In continuous-time filters, frequency
adjustment is
required to account for process variations —
relatively simple to implement.
• If channel has only frequency variation — use
arbitrary fixed coefficient analog filter and adjust
a single control line for frequency adjustment.
• Also possible with switched-C filter by
adjusting clock
University of frequency. slide 50 of
70
Toronto © D.A. Johns, 1997
 Analog Adaptive Filters
• Usually digital control desired — can switch
in caps and/or transconductance values
• Overlap of digital control is better than
missed values

pi pi
better worse
(hysteresis (potential large coeff
effect)
jitter)

digital coefficient control

digital coefficient control

• In switched-C filters, some type of

multiplying DAC needed.
• Best fully-programmable filter approach is
not clear
slide 51 of
University of 70
Toronto © D.A. Johns, 1997

Analog Adaptive Filters — DC Offsets

• DC offsets result in partial correlation of
data and error signals (opposite to opposite
DC offset)
x (k) 
i

m xi   wi(k)

e(k) 
mi
me

• At high-speeds, offsets might even be

larger than signals (say, 100 mV signals and
200mV offsets)
• DC offset effects worse for ill-
conditioned performance surfaces
slide 52 of
University of 70
Toronto © D.A. Johns, 1997
 Analog Adaptive Filters — DC Offsets
• Sufficient to zero offsets in either error or
state-
signals (easier with error since only one error
signal)
• For integrator offset, need a high-gain on error
signal
• error
Use + median-offset cancellation — slice error
comparato
r
signal and set the median of offset-free
offset
output to zero
error

• In most signals,
D/A its
up/ mean equals its median
down
counter

• Experimentally verified (low-frequency)

analog
adaptive with DC offsets more than twice the
size of the signal.
slide 53 of
University of 70
Toronto © D.A. Johns, 1997

DC Offset Effects for LMS Variants

0 2

e
Test Case LMS SD-LMS SE-LMS SS-LMS
Residual Mean Squared Error
input power  1
2 2
no effect
  1  ln
2
2
e e x  no -
x 1
effect SD-LMS
e2  0 e2  0 2   2 4 2   2 2 0
no offsets e e
for   for   x x
0 2 0 -20
LMS
 weakly depends on  2 strongly depends on 
e e

1 multiplier/tap 1 slicer/tap 1 trivial 1 slicer/tap

algorithm multiplier/tap -
1 integrator/tap 1 trivial 1 XOR gate/tap
circuit multiplier/tap 1 integrator/tap 1 counter/tap 3
complexity 1 integrator/tap 1 slicer/filter 1 DAC/tap 0
1 slicer/filter

convergence no gradient gradients no gradient gradients - SS-LMS

misalignment misaligned misalignment misaligned 4
0 SE-LMS

-50 - - -3 -
1 15 14 -2
1
-6
0 0 0 10 0
 Step
10 Size

slide 54 of
University of 70
Toronto © D.A. Johns, 1997
 Coax Cable Equalizer
• Analog adaptive filter used to equalize up to
300m
• Cascade of two 3’rd order filters with a single
tuning control
w1 s  s + p1  

i w2 s  s + p2  ou
n
t
w3 s  s + p3 

highpass filters
• Variable  is tuned to account for cable
length
slide 55 of
University of 70
Toronto © D.A. Johns, 1997

Coax Cable Equalizer

parasitic
Eq
poles
Resp

freq

• Equalizer optimized for 300m

• Works well with shorter lengths by tuning 
• Tuning control found by looking at slope of
equalized waveform
• Max boost was 40 dB
• System included dc recovery circuitry
• Bipolar circuit used — operated up to 300Mb/s

slide 56 of
University of 70
Toronto © D.A. Johns, 1997
 Analog Adaptive Equalization Simulation
noise
{1,0}
1-D
1 0  channel 
r(t) equalizer y(t) PR4
detector
a(k)
1 aˆ (k) 2 3
e(k)
yi(t) _
PR4 S1 +
generator 
4 S2
S1 - for training
S2 - for tracking

• Channel modelled by a 6’th-order Bessel filter with 3

different responses — 3MHz, 3.5MHz and 7MHz
• 20Mb/s data
• PR4 generator — 200 tap FIR filter used to find set of fixed
poles of equalizer
• Equalizer — 6’th-order filter with fixed poles and 5 zeros
adjusted (one left at infinity for high-freq roll-off)

slide 57 of
University of 70
Toronto © D.A. Johns, 1997

Analog Adaptive Equalization Simulation

• Analog blocks simulated with a 200MHz
clock and bilinear transform.
• Switch S1 closed (S2 open) and all poles
and 5 zeros adapted to find a good set of
fixed poles.
2

1.5
1
1

0.5

-1 1 0
[v]
y

-0.5

-1

-1.5

-1 -2
0 20 40 60 80 100 140 160
120 Time [ns]

• Poles and zeros depicted in digital

domain for equalizer filter.
• Residual MSE was -31dB

slide 58 of
University of 70
Toronto © D.A. Johns, 1997
 Equalizer Simulation — Decision Directed
• Switch S2 closed (S1 open), all poles fixed
and 5 zeros adapted using
• e(k) = 1 – y(t) if y(t)  0.5
• e(k) = 0 – y(t) if –0.5  y(t)  0.5
• e(k) = – 1 – y(t) if y(t)  –0.5

• all sampled at the decision time — assumes

clock recovery perfect
• Potential problem — AGC failure might cause
y(t) to always remain below 0.5 and then
adaptation will
force all coefficients to zero (i.e. y(t) = 0 ).
• Zeros initially mistuned to significant eye
closure

slide 59 of
University of 70
Toronto © D.A. Johns, 1997

Equalizer Simulation — Decision Directed

• 3.5MHz Bessel
1.5 1.5

1 1

0.5 0.5
Amplitude

0 0
[V]
y

[V]

-0.5 -0.5

-1 -1

-1.5
-1.5 0 100 150 200
0 20 40 60 80 100 120 140 160
Time [ns] Discrete Time
50
ideal PR4 and equalized pulse
[k]
initial
2
mistuned 20 outputs
18
1.5
PR4
16
1 overall
14
0.5
12

0 10
[v]

[V/V]
Gain
y

-0.5 8

6
-1 channel
4
-1.5
2
equalizer
-2 0
0 20 40 60 80 100 120 160 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
140 N o rma l i z e d R a di a n Fre qu e nc y [rad]
Time [ns]
after adaptation (2e6 after adaptation (2e6
iterations) iterations)

slide 60 of
University of 70
Toronto © D.A. Johns, 1997
 Equalizer Simulation — Decision Directed
• Channel changed to 7MHz Bessel
• Keep same fixed poles (i.e. non-optimum
pole placement) and adapt 5 zeros.
1.5 20

18 PR4
1
16

14
0.5
12

0 10

[V/V]
[V]

Gain
y

8
-0.5
6
channel
4
-1
2 equalizer overall
-1.5 0
0 20 40 60 80 100 120 140 0 0.1 0.2 0.3 0.4 0.5 0.6
160 0.7
Time [ns] N o rma l i z e d R a di a n Fre qu e nc y [rad]

after adaptation (2e6 iterations) after adaptation (2e6

iterations)

• Residual MSE = -29dB

• Note that no equalizer boost needed at
high-freq.
University of slide 61 of
70
Toronto © D.A. Johns, 1997

Equalizer Simulation — Decision Directed

• Channel changed to 3MHz Bessel
• Keep same fixed poles and adapt 5 zeros.
2 20

18
PR4
1.5
16
overall
1
14

0.5 12
[V]

10
[V/V]
Gain
y

0
8

-0.5 6 channel
4
-1 equalizer
2

-1.5 0
0 20 40 60 80 100 120 140 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
160 N o rma l i z e d R a di a n Fre qu e nc y
Time [ns] [rad]

after adaptation (3e6 iterations) after adaptation (3e6

• Residual MSE = -25dB
iterations)

• Note that large equalizer boost needed at

high-freq.
• Probably needs better equalization here
(perhaps move all poles together and let
zeros adapt)of
University slide 62 of
70
Toronto © D.A. Johns, 1997
 BiCMOS Analog Adaptive Filter Example
• Demonstrates a method for tuning the pole-
frequency and Q-factor of a 100MHz filter —
adaptive analog
• Application is a pulse-shaping filter for
data transmission.
• One of the fastest reported integrated
adaptive filters
— it is a Gm-C filter in 0.8um BiCMOS
process
• Makes use of MOS input stage and
translinear- multiplier for tuning
• Large tuning range (approx. 10:1)
• All analog components integrated (digital
left off) slide 63 of
University of 70
Toronto © D.A. Johns, 1997

BiCMOS Transconductor
M1 M2
M3 M4
v v
--- I+i I–i
d2 –---
2
-
M6
I+i
d
- Two styles implemented:
v vo I – i “2-quadrant” tuning (F-
--- M5 M7 M8
o2 –--2-- Cell), “4-quadrant”
-
tuning by cross-coupling
Q1 top
Q3
VC2 input stage (Q-Cell)
VC1 Q2 Q4
+ +
V BIAS  Gm_
_

slide 64 of
University of 70
Toronto © D.A. Johns, 1997
 Biquad Filter

2C 2C
+ + + + + + + +
U Gm2_1 X2 Gm2 2 G m 1_ X1
_ _ _ _ _
Gm 2
_i 2C 2C

+ +
Gm
_ _
b

• fo and Q not independent due to finite

output conductance
• Only use 4 quadrant transconductor
where needed

slide 65 of
University of 70
Toronto © D.A. Johns, 1997

Experimental Results Summary

Transconductor (T.) size 0.14mm x 0.05mm
T. power dissipation 10mW @ 5V
Biquad size 0.36mm x 0.164mm
Biquad worst case 20dB
CMRR
Biquad fo tuning range 10MHz-230MHz @ 5V, 9MHz-135MHz @ 3V
Biquad Q tuning range 1-Infinity
Bq. inpt. ref. noise dens. 0.21 Vrms  H
Biquad PSRR+ 28dB
Biquad PSRR- 21dB
Output 3rd Order
Filter Setting
Intercept Point SFDR
100MHz, Q = 2, Gain = 10.6dB 23dBm 35dB
20MHz, Q = 2, Gain = 30dB 20dBm 26dB
100MHz, Q = 15, Gain = 29.3dB 18dBm 26dB
227MHz, Q = 35, Gain = 31.7dB 10dBm 20dB

slide 66 of
University of 70
Toronto © D.A. Johns, 1997
 Adaptive Pulse Shaping Algorithm
30
Ideal input pulse
(not to scale)
20
lowpas
s
10
output

Voltage
[mV]
bandpas
-10 s output

-20

  =
-30 2.5ns
20 25 30 35 40
• Fo control: sample output pulse shape at nominal zero-
4

crossing and decide if early or late (cutoff frequency too fast or

too slow respectively)
• Q control: sample bandpass output at lowpass nominal zero-
crossing and decide if peak is too high or too small (Q
too large or too small)

slide 67 of
University of 70
Toronto © D.A. Johns, 1997

Experimental Setup
pulse-shaping filter chip
LP
U
CLK
Gm-C +
Data Biquad - logic u/d
counter
DAC
U
Generator Filter
+ logic
u/d
counter DAC
fo Q - off-chip tuning algorithm
BP

Vref CLK

• Off-chip used an external 12 bit DAC.

• Input was 100Mb/s NRZI data 2Vpp
differential.
• Comparator clock was data clock (100MHz)
time delayed by 2.5ns
slide 68 of
University of 70
Toronto © D.A. Johns, 1997
 Pulse Shaper Responses
20 20

0 0

-20 -20

0 5 10 15 20 25 0 5 10 15 20 25

20 20

0 0

-20 -20

0 5 10 15 20 2 0 5 10 15 20 2

Initial — high-freq. high-Q Initial — high-freq. low-Q

20 20

0 0

-20 -20

0 5 10 15 20 25 0 5 10 15 20 25

20 20

0 0

-20 -20

5 10 15 20 0 5 10 15 20 2
0 2
Initial — low-freq. high-Q Initial — low-freq. low-Q

slide 69 of
University of 70
Toronto © D.A. Johns, 1997

Summary
• Adaptive filters are relatively common
• LMS is the most widely used algorithm
• Adaptive linear combiners are almost always
used.
• Use combiners that do not have poor
performance surfaces.
• Most common digital combiner is tapped FIR

Digital Adaptive:
• more robust and well suited for programmable filtering

Analog Adaptive:
• best suited for high-speed, low dynamic range.
• less power
• very good at realizing arbitrary coeff with frequency only
change.
• Be aware of DC offset effects
slide 70 of
University of 70
Toronto © D.A. Johns, 1997