Module-3 : Transmission Lecture-5 (4/5/00)
Marc Moonen
Dept. E.E./ESAT, K.U.Leuven
marc.moonen@esat.kuleuven.ac.be
www.esat.kuleuven.ac.be/sista/~moonen/
Prelude
Comments on lectures being too fast/technical
* I assume comments are representative for (+/-)whole group * Audience = always right, so some action needed….
To my own defense :-)
* Want to give an impression/summary of what today’s
transmission techniques are like (`box full of mathematics & signal processing’, see Lecture-1).
Ex: GSM has channel identification (Lecture-6), Viterbi (Lecture-4),...
* Try & tell the story about the maths, i.o. math. derivation.
* Compare with textbooks, consult with colleagues working in
Prelude
Good news
* New start (I): Will summarize Lectures (1-2-)3-4.
-only 6 formulas-
* New start (II) : Starting point for Lectures 5-6 is 1 (simple) input-output model/formula (for Tx+channel+Rx).
* Lectures 3-4-5-6 = basic dig.comms principles, from then on focus on specific systems, DMT (e.g. ADSL), CDMA (e.g. 3G mobile), ...
Bad news :
* Some formulas left (transmission without formulas = fraud) * Need your effort !
* Be specific about the further (math) problems you may have.
Lecture-5 : Equalization
Problem Statement :
• Optimal receiver structure consists of
* Whitened Matched Filter (WMF) front-end
(= matched filter + symbol-rate sampler + `pre-cursor equalizer’ filter)
* Maximum Likelihood Sequence Estimator (MLSE), (instead of simple memory-less decision device)
• Problem: Complexity of Viterbi Algorithm (MLSE)
• Solution: Use equalization filter + memory-less
decision device (instead of MLSE)...
Lecture-5: Equalization - Overview
• Summary of Lectures (1-2-)3-4
Transmission of 1 symbol : Matched Filter (MF) front-end
Transmission of a symbol sequence :
Whitened Matched Filter (WMF) front-end & MLSE (Viterbi)
• Zero-forcing Equalization
Linear filtersDecision feedback equalizers
• MMSE Equalization
• Fractionally Spaced Equalizers
Summary of Lectures (1-2-)3-4
Channel Model:
Continuous-time channel
=Linear filter channel + additive white Gaussian noise (AWGN)
(symbols) a
kaˆ
kn(t) +
AWGN
transmitter receiver (to be defined) h(t)
channel
...
? ?
Summary of Lectures (1-2-)3-4
Transmitter:
* Constellations (linear modulation): n bits -> 1 symbol (PAM/QAM/PSK/..)
receiver (to be defined) ...
s
k E
a .
r(t)
aˆk
transmit pulse
s(t)
n(t)
p(t) +
AWGN transmitter
h(t)
channel ?
E
sa
kp t kT
st
s ( ) . . ( )
a
kSummary of Lectures (1-2-)3-4
Transmitter:
-> piecewise constant p(t) (`sample & hold’) gives s(t) with infinite bandwidth, so not the greatest choice for p(t)..
-> p(t) usually chosen as a (perfect) low-pass filter (e.g. RRC)
s k
E a .
transmit pulse
s(t) p(t)
transmitter discrete-time
symbol sequence
continuous-time transmit signal t
p(t)
t
Example
Summary of Lectures (1-2-)3-4
Receiver:
In Lecture-3, a receiver structure was postulated (front-end filter + symbol-rate sampler + memory-less decision
device). For transmission of 1 symbol, it was found that the front-end filter should be `matched’ to the received pulse.
ˆa0
front-end filter
1/Ts
receiver n(t)
+ AWGN
Es
a .0
transmit pulse
p(t)
transmitter
h(t)
channel
u0
Summary of Lectures (1-2-)3-4
Receiver: In Lecture-4, optimal receiver design was based on a minimum distance criterion :
• Transmitted signal is
• Received signal
• p’(t)=p(t)*h(t)=transmitted pulse, filtered by channel
k
s k
s a
a
a
r t E a p t kT dt
K
2 ,...,ˆ
,ˆ
ˆ
| ( ) . ˆ . ' ( ) |
min
0 1
k
s k
s
a p t kT
E t
s ( ) . . ( )
) ( )
( ' . .
)
( t E a p t kT n t
r
k
s k
s
Summary of Lectures (1-2-)3-4
Receiver: In Lecture-4, it was found that for transmission of 1 symbol, the receiver structure of Lecture 3 is indeed optimal !
2 0 0
ˆ 0
( . ). ˆ
min
a0u E
sg a
ˆa0
p’(-t)*
front-end filter
1/Ts
receiver n(t)
+ AWGN
Es
a .0
transmit pulse
p(t)
transmitter
h(t)
channel
sample at t=0 p’(t)=p(t)*h(t)
u0
Summary of Lectures (1-2-)3-4
• Receiver: For transmission of a symbol sequence, the optimal receiver structure is...
aˆk
p’(-t)*
front-end filter
1/Ts n(t)
+ AWGN
s
k E
a .
transmit pulse
p(t) h(t)
sample at t=k.Ts
uk
K k
k k l
l k K
k
K l
k s
a
a
E a g a a u
K 1
*
1 1
* ,...,ˆ
ˆ
. ˆ . . ˆ 2 ˆ .
min
0Summary of Lectures (1-2-)3-4
Receiver:
• This receiver structure is remarkable, for it is
based on symbol-rate sampling (=usually below Nyquist-rate sampling), which appears to be
allowable if preceded by a matched-filter front-end.
• Criterion for decision device is too complicated.
Need for a simpler criterion/procedure...
Summary of Lectures (1-2-)3-4
Receiver: 1st simplification by insertion of an additional (magic) filter (after sampler).
* Filter = `pre-cursor equalizer’ (see below)
* Complete front-end = `Whitened matched filter’
aˆk
p’(-t)*
front-end filter
1/Ts n(t)
+ AWGN
s
k E
a .
transmit pulse
p(t) h(t)
uk
1/L*(1/z*) yk
2
1 1
,...,ˆ
ˆ
ˆ .
min
0
K m
K k
k m k m
a
a
y a h
K
Summary of Lectures (1-2-)3-4
Receiver: The additional filter is `magic’ in that it turns the complete transmitter-receiver chain into a simple input-
output model:
k k
z H k
k k
k k
k k
w a
z h z
h z
h h
y
w a
h a
h a
h a
h y
. ...) .
. .
(
...
. ..
..
.
) (
3 3
2 2
1 1
0
3 3
2 2
1 1
0
aˆk
p’(-t)*
front-end filter
1/Ts
receiver n(t)
+ AWGN
s
k E
a .
transmit pulse
p(t)
transmitter
h(t)
channel
uk
1/L*(1/z*) yk
Summary of Lectures (1-2-)3-4
Receiver: The additional filter is `magic’ in that it turns the complete transmitter-receiver chain into a simple input-
output model:
= additive white Gaussian noise
means interference from future
(`pre-cursor) symbols has been cancelled, hence only interference from past (`post-cursor’) symbols remains
k k
k k
k
k
h a h a h a h a w
y
0.
1.
1
2.
2
3.
3 ...
w
k0
2
...
1
h
h
Summary of Lectures (1-2-)3-4
Receiver: Based on the input-output model
one can compute the transmitted symbol sequence as
A recursive procedure for this = Viterbi Algorithm Problem = complexity proportional to M^N !
(N=channel-length=number of non-zero taps in H(z) )
k k
k k
k
k
h a h a h a h a w
y
0.
1..
1
2..
2
3.
3 ...
2
1 1
,...,ˆ
ˆ
ˆ .
min
0
K m
K k
k m k
m a
a
y a h
K
Problem statement (revisited)
• Cheap alternative for MLSE/Viterbi ?
• Solution: equalization filter + memory-less decision device (`slicer’)
Linear filters
Non-linear filters (decision feedback)
• Complexity : linear in number filter taps
• Performance : with channel coding, approaches
MLSE performance
Preliminaries (I)
• Our starting point will be the input-output model for transmitter + channel + receiver whitened matched filter front-end
k k
k k
k
k
h a h a h a h a w
y
0.
1.
1
2.
2
3.
3 ...
h
1
h
3h
0
h
2a
ka
k1a
k2a
k3y
kw
kPreliminaries (II)
• PS: z-transform is `shorthand notation’ for discrete-time signals…
…and for input/output behavior of discrete-time systems
) ( )
( ).
( )
(
hence
...
. .
.
.
1 1 2 2 3 30
z W z
A z
H z
Y
w a
h a
h a
h a
h
y
k k k k k k
....
. .
. .
) (
....
. .
. .
) (
2 2
1 1
0 0
0
2 2
1 1
0 0
0
z h z
h z
h z
h z
H
z a z
a z
a z
a z
A
i
i i
i
i i
) (z
A H(z) Y (z )
Preliminaries (III)
• PS: if a different receiver front-end is used (e.g. MF instead of WMF, or …), a similar model holds
for which equalizers can be designed in a similar fashion...
k k
k k
k k
k
h a h a h a h a h a w
y ~ . ... ~
~ .
~ .
~ .
~ .
...
2 2
1 1
0
1 1
2 2
Preliminaries (IV)
PS: properties/advantages of the WMF front end
• additive noise = white (colored in general model)
• H(z) does not have anti-causal taps
pps: anti-causal taps originate, e.g., from transmit filter design (RRC, etc.). practical implementation based on causal filters + delays...
• H(z) `minimum-phase’ :
=`stable’ zeroes, hence (causal) inverse exists &
stable
= energy of the impulse response maximally concentrated
w
k0
2
...
1
h
h
)
1
( z
H
Preliminaries (V)
• `Equalization’: compensate for channel distortion.
Resulting signal fed into memory-less decision device.
• In this Lecture :
- channel distortion model assumed to be known - no constraints on the complexity of the
equalization filter (number of filter taps)
• Assumptions relaxed in Lecture 6
NOISE
ISI
3 3
2 2
1 1
0
.
k.
k.
k.
k...
kk
h a h a h a h a w
y
Zero-forcing & MMSE Equalizers
2 classes :
Zero-forcing (ZF) equalizers
eliminate inter-symbol-interference (ISI) at the slicer input
Minimum mean-square error (MMSE) equalizers
tradeoff between minimizing ISI and minimizing noise at
NOISE
ISI
3 3
2 2
1 1
0
.
k.
k.
k.
k...
kk
h a h a h a h a w
y
Zero-forcing Equalizers
Zero-forcing Linear Equalizer (LE) : - equalization filter is inverse of H(z) - decision device (`slicer’)
• Problem : noise enhancement ( C(z).W(z) large) )
( )
( z H
1z C
H(z)
) (z W
) (z Y
C(z) )
(z
A A ˆ z ( )
Zero-forcing Equalizers
Zero-forcing Linear Equalizer (LE) :
- ps: under the constraint of zero-ISI at the slicer
input, the LE with whitened matched filter front-end is optimal in that it minimizes the noise at the slicer input
- pps: if a different front-end is used, H(z) may have
unstable zeros (non-minimum-phase), hence may
be `difficult’ to invert.
Zero-forcing Equalizers
Zero-forcing Non-linear Equalizer
Decision Feedback Equalization (DFE) :
- derivation based on `alternative’ inverse of H(z) :
(ps: this is possible if H(z) has , which is another property of the WMF model)
- now move slicer inside the feedback loop :
) (z Y
1-H(z) H(z)
) (z W )
(z
A A ˆ z ( )
0
1
h
Zero-forcing Equalizers
moving slicer inside the feedback loop has…
- beneficial effect on noise: noise is removed that would otherwise circulate back through the loop - beneficial effect on stability of the feedback loop:
output of the slicer is always bounded, hence feedback loop always stable
) (z Y
D(z) H(z)
) (z W )
(z A
) ˆ z ( A
) ( 1
)
( z H z
D
Zero-forcing Equalizers
Decision Feedback equalization (DFE) : - general DFE structure
C(z): `pre-cursor’ equalizer
(eliminates ISI from future symbols)
D(z): `post-cursor’ equalizer
(eliminates ISI from past symbols)
) (z Y )
(z A
H(z)
) (z W
C(z) A ˆ z ( )
D(z)
Zero-forcing Equalizers
Decision Feedback equalization (DFE) : - Problem : Error propagation
Decision errors at the output of the slicer cause a corrupted estimate of the postcursor ISI.
Hence a single error causes a reduction of the noise margin for a number of future decisions.
Results in increased bit-error rate.
) (z Y H(z)
) (z W )
(z A
C(z) A ˆ z ( )
D(z)
Zero-forcing Equalizers
`Figure of merit’
• receiver with higher `figure of merit’ has lower error probability
• is `matched filter bound’ (transmission of 1 symbol)
• DFE-performance lower than MLSE-performance, as DFE relies on only the first channel impulse response sample (eliminating all other ‘s), while MLSE uses energy of all taps . DFE benefits from minimum-phase property (cfr.
supra, p.20)
MF MLSE
DFE
LE
MFh
0h
ih
iMMSE Equalizers
• Zero-forcing equalizers: minimize noise at slicer input under zero-ISI constraint
• Generalize the criterion of optimality to allow for residual ISI at the slicer & reduce noise variance at the slicer
=Minimum mean-square error equalizers
MMSE Equalizers
MMSE Linear Equalizer (LE) :
- combined minimization of ISI and noise leads to
2
*
*
*
*
*
*
*
*
1 ) ( ).
(
1 ) ( )
( 1 )
( ).
( ).
(
1 ) ( ).
( )
(
n W
A
A
H z z H
H z z
z S H
z H z S
H z z S z
C
H(z)
) (z W
) (z Y
C(z) )
(z
A A ˆ z ( )
MMSE Equalizers
- signal power spectrum (normalized) - noise power spectrum (white)
- for zero noise power -> zero-forcing
- (in the nominator) is a discrete-time matched filter, often `difficult’ to realize in practice
(stable poles in H(z) introduce anticausal MF)
2
*
*
*
*
*
*
*
*
1 ) ( ).
(
1 ) ( )
( 1 )
( ).
( ).
(
1 ) ( ).
( )
(
W W
A
A
H z z H
H z z
z S H
z H z S
H z z S z
C
1 )
(z S
A
2)
(
WW
z
S
) ( )
( z H
1z C
1 ) (
**
H z
MMSE Equalizers
MMSE Decision Feedback Equalizer :
• MMSE-LE has correlated `slicer errors’
(=difference between slicer in- and output)
• MSE may be further reduced by incorporating a `whitening’
filter (prediction filter) E(z) for the slicer errors
• E(z)=1 -> linear equalizer
• Theory & formulas : see textbooks
) (z Y
H(z)
) (z W )
(z A
C(z)E(z) A ˆ z ( )
1-E(z)
Fractionally Spaced Equalizers
Motivation:
• All equalizers (up till now) based on (whitened) matched filter front-end, i.e. with symbol-rate sampling, preceded by an (analog) front-end filter matched to the received pulse p’(t)=p(t)*h(t).
• Symbol-rate sampling = below Nyquist-rate sampling
(aliasing!). Hence matched filter is crucial for performance !
• MF front-end requires analog filter, adapted to channel h(t), hence difficult to realize...
• A fortiori: what if channel h(t) is unknown ?
• Synchronization problem : correct sampling phase is
Fractionally Spaced Equalizers
• Fractionally spaced equalizers are based on Nyquist-rate sampling, usually 2 x symbol-rate sampling (if excess bandwidth < 100%).
• Nyquist-rate sampling also provides sufficient statistics, hence provides appropriate front-end for optimal receivers.
• Sampler preceded by fixed (i.e. channel independent) analog anti-aliasing (e.g. ideal low-pass) front-end filter.
• `Matched filter’ is moved to digital domain (after sampler).
• Avoids synchronization problem associated with MF
front-end.
Fractionally Spaced Equalizers
• Input-output model for fractionally spaced equalization : `symbol rate’ samples :
`intermediate’ samples :
• may be viewed as 1-input/2-outputs system
k k
k k
k
h a h a h a w
y ~ . ... ~
~ .
~ .
...
0
1 1
2 2
2 / 1 2
2 / 5 1
2 / 3 2
/ 1 2
/
1
~ . ... ~
~ .
~ .
...
k
k
k
kk
h a h a h a w
y
Fractionally Spaced Equalizers
• Discrete-time matched filter + Equalizer (LE) :
• Fractionally spaced equalizer (LE) :
) ˆ z ( 1/2Ts A
2
MF(z)
C(z)
equalizer
) (t
r C(z) A ˆ z ( )
1/2Ts
2
Fractionally spaced equalizer )
(t r
F(f)
F(f)
Fractionally Spaced Equalizers
• Fractionally spaced equalizer (DFE):
• Theory & formulas : see textbooks & Lecture 6
C(z) A ˆ z ( )
D(z) 1/2Ts
2 )
(t r
F(f)
Conclusions
• Cheaper alternatives to MLSE, based on equalization filters + memoryless decision device (slicer)
• Symbol-rate equalizers : -LE versus DFE
-zero-forcing versus MMSE
-optimal with matched filter front-end, but several assumptions underlying this structure are often violated in practice
• Fractionally spaced equalizers (see also Lecture-6)
Assignment 3.1
• Symbol-rate zero-forcing linear equalizer has i.e. a finite impulse response (`all-zeroes’) filter is turned into an infinite impulse response filter
• Investigate this statement for the case of fractionally spaced equalization, for a simple channel model
and discover that there exist finite-impulse response inverses in this
case. This represents a significant advantage in practice. Investigate the
) ( )
( z H
1z
C
2 2
1 1
0
. .
)
( z h h z
h z
H
) .
. /(
1 )
( z h
0 h
1z
1 h
2z
2C
2 2
/ 5 1
2 / 3 2
/ 1 2
/ 1
2 2
1 1
0
. .
.
. .
.
k k
k k
k k
k k