Lecture

Digital Communications II
Prof. Dr.-Ing. Dipl.-Wirt.-Ing. Stephan Pachnicke
Overview

• Organization

• Lecture Contents

• Fundamentals of Linear FIR Equalizers

2 01 - Introduction
Organization
• The lecture will take place in presence (F-SR-III) between 11.00
and 12.30 starting on Monday, 24.10.22.

• Excercises will also take place in presence

(starting on Thursday, 3.11.22) and will be led by Mr. Ali
Derakhshandeh.

• Exam:
1.5h written exam

3 01 - Introduction
Contact Information

Slides of the lectures and exercises can be

downloaded from OLAT (https://lms.uni-kiel.de).
Direct link:
https://lms.uni-kiel.de/auth/RepositoryEntry/1765474312/CourseNode/94280050222606

E-Mail: [email protected]
E-Mail (for questions regarding exercises):
[email protected]

4 01 - Introduction
Lecture Contents (1)
Channel Equalization
• Fundamentals of linear FIR-equalizers
• Adaptive equalization
• Blind equalization using the constant modulus algorithm (CMA)
• Optimum receivers (MAP & ML)
• Maximum likelihood sequence estimation (MLSE)

Code Division Multiple Access (CDMA)

• Working principle
• Pseudo-noise sequences
• Code synchronisation

Multi Carrier Transmission Schemes (OFDM, DMT)

• Working principle
• Digital implementation of the FFT
• Further system aspects

5 01 - Introduction
Lecture Contents (2)
Multiple Input Multiple Output Systems (MIMO)
• Fundamental description and system model
• Channel capacity
• Maximum ratio combining
• Space-time-codes
• Spatial division multiplexing

Introduction to Channel Coding

• Fundamentals
• Block codes
• Convolutional codes
• Trellis coded modulation (TCM)
• Channel capacity and Shannon limit

Application Examples
• Digital audio broadcasting (DAB)
• Digital video broadcasting (DVB)
• Digital subscriber line (DSL)
• Wireless LAN (WIFI)
• Bluetooth
• Optical wireless transmission
• Fiber-optical high speed transmission
6 01 - Introduction
Literature

• Proakis, J.G.: Digital Communications. 5th ed., McGraw-Hill,

2008.

• Glover, I.A.; Grant, P.M.: Digital Communications. 3rd ed.,

Prentice Hall, Harlow, 2009.

• Barry, J.R.; Lee, E.A., Messerschmitt, D.G.: Digital

Communications. 3rd ed., Springer, 2004.

In German only:

• Kammeyer, K.-D.: Nachrichtenübertragung, 6th ed., Vieweg-

Teubner, 2018.

• P. A. Höher, „Grundlagen der digitalen Informationsübertragung:

Von der Theorie zu Mobilfunkanwendungen“, 2nd ed.,
Vieweg+Teubner, 2013.

7 01 - Introduction
 Equalization

Methods to be used in receiver to eliminate unwanted influences of the channel

(mostly ISI).

We considered so far:

1.) ISI-free transmission → 1. Nyquist criterion

 1 t =0

g (t ) = 0 t =kTb , k =±1, ±2,
arbitrary elsewhere

with
g ( t ) = hS ( t ) ∗ hK ( t ) ∗ hE ( t ) = hSK ( t ) ∗ hE ( t )

!
2.) optimum noise performance ( S / N =Max) for AWGN channel → matched filter
hE (t ) ~ hSK ( −t )

| H E ( f )|~| HSK ( f )|

1) and 2) fulfilled e.g. for G(f) = raised-cosine filter and H E f = HSK f = G f bg bg bg

8 01 - Introduction
Recap: Channel Model
nWGN(t)

dk x(t) y(t) y(kTb) d^k

hS(t) hK(t) + hE(t)

clk(t) nu_51_digmod.dsf

pulsformer (AWGN) channel Receive Sampling Decision (slicer)

in transmitter impulse response filter (clock Tb) =quantizer

Further simplified model:

• Noise is disregarded
• Combination to a single impulse response

• assumption: linear system!

• dk is typically an iid binary sequency

9 01 - Introduction
 Recap: Matched Filter
• Optimum design for ISI-free transmission ⇒ 1st Nyquist criterion
• Optimum design of receiver filter for best signal-to-noise ratio SNR ⇒ How to design?
Model with AWGN channel
nu_55_modell_mit_awgn.dsf

nWGN(t)

dk y(t)+n(t) y(νTb)+n(νTb) d^k

hSK(t) + hE(t)
Tb

hSK(t)=hS(t)*hK(t) clk(t)
Pulsformer Receive Sampling Decision (slicer)
in transmitter filter (clock Tb) =quantizer
plus AWGN channl
impulse response Assumptions:
• ISI-free
• non-causal (for simplification, no restriction)
• ideal sampling t=νTb in center of eye, no jitter

z
+∞
• Impulse response with finite energy | h(t )|2 dt < ∞
−∞

g (t ) = hS (t ) ∗ hK (t ) ∗ hE (t ) = hSK (t ) ∗ hE (t ) =
RSg(0) t = t0 = 0
T0 t = kTb , k ≠ 0

10 01 - Introduction
 Optimum Filter Dimensioning

SNR is maximum for equal sign in Schwarz’s inequality. Therefore we derive the

Matched Filter
!
hE (t ) = K ⋅ hSK ( − t ) (5.4-5a)
*
H E ( f ) =⋅
K H SK ( f ) =⋅
K H SK (− f ) (5.4-5b)
hSK(t) hE(t)
nu_54_matched_filter.dsf

t t

Receiver filter hE (t ) acc. to Eq. (5.4-5) is matched to the pulse shape hSK ( t ) at
receiver input. Matched in the sense that the m.f. results in the best possible SNR.

Considering causality and hE(t) of finite length T we have:

hE (t ) =
K ⋅ hSK (T − t )

HE ( f ) = *
K ⋅ H SK ( f ) ⋅ e j 2π f T

11 01 - Introduction
 Practical Systems

No ISI and use of matched filters are theoretical and analytically derived optimisation criteria for
bg
designing hE t . For practical implementations often not applicable without modifications or

adaptations. For example channel is often unknown (e.g. only statistical description → mobile
channel) or changing (e.g. non equidistant cabling lengths in networks)) → adaptive solutions

Frequently used in practice:

• hs (t ) approximately rectangular or raised cosine pulse shape

• bg
Analog receiver filter hE t (front end) with (rough) approximation of matched filter condition

ignoring the channel ( hE (t ) ~ hS (−t ) ). Results in improved noise performance

• Channel influence (mainly ISI) is reduced by an additional equalisation through a separate

equaliser system
• Equaliser is often implemented within a digital signal processor, i.e. after applying ADC
(analog-to-digital converter)
• We thus consider equaliser design based on time discrete signals with sampling clock TA = Ts
12 01 - Introduction
 Signal Representation by
Vectors, Convolutional Matrix
Time discrete signals and systems

h(k ) = {h0 , h1 , ... , hN }

Convolution:
N
(k )
y= ∑ h(κ ) ⋅ x(k −=
κ) x ( k ) * h( k )
κ =0

 x(k ) 
 x(k − 1) 
x( k ) =  
  
 
 x(k − N ) 
 h0 
h 
h= 1
  
 
 hN 

y ( k ) = x( k )T ⋅ h = h T ⋅ x( k ) = x ( k ) * h( k )

13 01 - Introduction
 Example

Example, N=2

N
=y (k ) ∑ h (κ )x ( k − κ )
κ =o

e.g. N = 2
 x(k ) 
T
= h x(k )= ( h0 , h1, h2 )  x(k − 1)= h0 x(k ) + h1x(k − 1) + h2 x(k − 2)
 x(k − 2) 
 

14 01 - Introduction
 Convolutional Matrix
Example: Impulse response length N+1=3, input signal length L+1=4 samples, output
signal observation: N+L+1=6

 y (0)   h(0) 0 0 0 
 y (1)   h(1) h(0) 0 0   x(0) 
  
 y (2)   h(2) h(1) h(0) 0   x(1) 
   ⋅
 y (3)   0 h (2) h (1) h (0)   x(2) 
 y (4)   0  
0 h(2) h(1)   x(3) 
   
 y (5)   0 0 0 h(2) 

y = Hx
or
y = Xh

(due
= to y (k ) x=
( k ) * h( k ) h( k ) * x ( k ) )

15 01 - Introduction
 Fundamentals of Linear FIR
Equalizers
!
here ISI = 0  Zero-Forcing Equaliser
Equaliser = linear digital non-recursive (FIR)-filter after sampling of receiver signal
with clock
= TA Tb / w, w ∈  → suitable for linear channels!

Sampling rate Decision=

kTA FIR filter reduction Slicer
dk x(t) y(t) yd(kTb) d^k
hS(t) hK(t) hE(t) e(k) w

Sequence
Transmitter Channel Receive TA≤Tb Equaliser with bit clock Tb
filter
Tb=wTA nu_571_linearer_entz.dsf

g(t) w=integer

16 01 - Introduction
 Design of Impulse Response

Impulse response of equaliser ( e (=

kTA ) 0 for k > n )

b g l
e kTA = e0 , e1 ,..., en q filter length n + 1
Impulse response (approximately) finite length g(t) (non-recursive channel)
g ( kTA ) = g ( kTb / w ) =  hs ( t ) ∗ hK ( t ) ∗ hE ( t )  kTb (= 0 for k > m)
t=
w

b g l
g kTA = g0 , g1 ,..., gm q Filter length m + 1

g(t) contains channel as most important parameter (hs und hE are assumed to be
known). Therefore g(t) is denoted as „channel impulse response“ (non-recursive
channel)

Task: design of e(k) such that for the overall system a Nyquist-impulse response
results. After sampling rate reduction by w we obtain the sequence:

1 k = k 0
yd ( kTb ) =
= {0, 0,..., 0,1, 0,..., 0} (in bit clock)
0 else

17 01 - Introduction
 Computation of Equalizer
Coefficients
nu_571_zero_forcing.dsf kTA w=2
g(t) e w

g(kTA) yd(kTA)
δ0

t t Tb t
. TA

compute coefficients of equaliser filter from

m
g ( kTA ) ∗=
e ( kTA ) ∑ g (κ TA )e((k −=
κ )TA ) yd ( kTA )
(6.7-1)
κ =0
↑ ↑ ↑
given ? conditions from Nyquist criterion


system of linear equations with a number NB of equations to be solved for
determination of the n+1 equaliser coefficients
18 01 - Introduction
 Synchronous Linear FIR-
Equalization
For w=1 (Sampling with bit clock = synchron. "Echo canceller"). Advantage: lowest sampling rate
However, no exact solution (zero ISI) possible as system of linear equations is overdetermined (more
equations than unknowns) Nyquist criterion can be fulfilled only approximately  Best solution:
minimise mean square error (MMSE)!
Equaliser coefficients: (FIR filter, length n+1 is assumed)

b
e = e0 , e1 , , en gT
Received impulse response:

b
g = g0 , g1 , , gm gT
Set of linear equations (convolutional matrix) from eq. (6.7-1) plus an additional vector ε (error vector).

19 01 - Introduction
 Example
Example = n 4 ; length of g*e: m+n+1=8
m 3,=

 g0   0   ε0 
    ε 
 g1 g0 0  e  0  1 
 g2    0
g1 g0  0  ε2 
   1e    
 g3 g2 g1 g0  ⋅  e  =  1  +  ε3 
 g3 g2 g1 g0   2   0  ε4 
   e3     
 g3 g2 g1     0   ε5 
e
 0 g3 g2   4  0 ε 
     6 
 
 g3   0   ε7 
G ⋅ e = yd + ε
This overdetermined set of equations has to be solved for e so that the mean square
error (=power of error) ε is minimised:
n+m !
T 2
= ε=
FMSE ε ∑ ε=
k Min.
k =0

Solution:

( )
−1
e = GT G GT y d (6.7-2)

Power of residual error:

1 − y dT G ⋅ e
Min{ε T ε } = (6.7-3)
20 01 - Introduction
 Derivation
!
εT ε = (Ge − y d )T (Ge − y d ) = (eT G T − y d T )(Ge − y d ) = Min

therefore:
e e=
= arg min{εT ε}
opt
e

εT=
ε eT G

T
G e − y Td Ge − eT G T y d + 1

A b

With quadratic supplement (completion of the square)

εT ε= (eT A − bT ) A −1 ( Ae − b) − bT A −1b + 1

Only first product term on right hand side of the equation depends on e. It is of
quadratic form and thus non-negative. Minimum by setting equal to zero

Ae −=
b G T Ge − G T =
yd 0

21 01 - Introduction
 Linear FIR-Equalizers with
Oversampling
A unique solution requires at least w=2 (2-fold oversampling) "Fractionally spaced
Equaliser" (2 samples per bit period  increased complexity!)

Equaliser coefficients:

b
e = e0 , e1 , , en gT
Received impulse response (sampling at kTb / 2 ):

b
g = g0 , g1/ 2 , g1 , g3/ 2  , gm / 2 gT
Set of linear equations (Example m=4, n=4)

 g0   0 
   
 g1/ 2 g0 0   yd 1 
 g1 g1/ 2 g0   e0   0 
     
 g3 / 2 g1 g1/ 2 g0   e1   yd 3 
 g g3 / 2 g1 g1/ 2 g0  ⋅  e2  = 1 
 2     
 g2 g3 / 2 g1 g1/ 2   e3   yd 5 
 g2 g3 / 2 g1   e4   0 
   
 0 g2 g3 / 2   yd 7 
 g 2   0 
  

22 01 - Introduction
 Solution
 g0   0 
   
 g1/ 2 g0 0   yd 1 
 g1 g1/ 2 g0   e0   0 
     
 g3 / 2 g1 g1/ 2 g0   e1   yd 3 
 g g3 / 2 g1 g1/ 2 g0  ⋅  e2  = 1 
 2     
 g2 g3 / 2 g1 g1/ 2   e3   yd 5 
 g2 g3 / 2 g1   e4   0 
   
 0 g2 g3 / 2   yd 7 
 g 2   0 
  
Values of yd1, yd 3 , yd 5 , yd 7 arbitrary  only every second row results in a useful equation  delete rows!

 5 unknowns, 5 equations:
 g0 0   e0   0 
     
 g1 g1/ 2 g0   e1   0 
 g2 g3 / 2 g1 g1/ 2 g0  ⋅  e2  = 1
     
 g2 g3 / 2 g1   e3   0 

 0 g 2   e4   0 
G ⋅ e = yd

Solution:

e = G −1 y d (6.7-4)
23 01 - Introduction
 Decision Feedback Equalizers

Decision Feedback (DF) Equaliser → nonlinear equaliser. Compared to linear

equalisers the DFE usually has
• less implementation complexity (shorter filter lengths)
• less noise enhancement at upper frequency range
• exact equalisation with only 1 sample per bit (w=1) is possible

24 01 - Introduction
 Derivation
+∞
Received signal: bg
yt = ∑ d bk g gbt − kTb g
k =−∞

Sampling with bit clock Tb with delay to (corresponds to sampling phase shift ϕ = 2π to Tb ) at time instants
t = to + νTb
+∞
b g
y to + νTb = ∑ d bk g gcto + bν − k gTb h
k =−∞
+∞
= ∑ d bν − l g gbto + lTb g
l =−∞

(with Subst: ν-k=l)

Impulse response g(t) of finite length and causal
g(t)
g1 nu_574_Impulsantwort.dsf

g0
 g=( l ) gl 0≤l ≤m g2
g ( to + lTb ) =
0 else
t
to Tb

m m
b
 y to + νTb = g ∑ d bν − l g gbl g = d bν gg (0) + ∑ d bν − l g gbl g (7.1-5)
l =0 l =1
25 01 - Introduction
 ISI Removal
g(t)
g1 nu_574_Impulsantwort.dsf

g0
g2

 t
m m
to Tb
b g
y to + νTb = ∑ d bν − l g gbl g = d bν gg (0) + ∑ d bν − l g gbl g (7.1-5)
l =0 l =1

Sample no. ν contains d(ν) (bit no. ν with weight g(0)=g0 ) (actual bit) and the m
previous bits (with weights given by the tail of the impulse response g) as the
unwanted ISI component

If we know the previous bits (from previous data decisions) and the impulse
response, g(t) the actual bit d(ν) can be derived ISI free (exactly) from eq. (7.1-5)
m
g ( 0 ) d (ν ) =y (to + ν Tb ) − ∑ g ( l ) d (ν − l )
=1
l
FIR-Filter, filters the
bits d (ν - l ) after decision
26 01 - Introduction
 DFE Block Diagram
m
g ( 0 ) d (ν ) =y (to + ν Tb ) − ∑ g ( l ) d (ν − l )
=1
l
FIR-Filter, filters the
bits d (ν - l ) after decision

Block diagram DF-Equaliser: (Example: m=3, unipolar transmission)

1

y(t)
y(ν) +
+
-
d(ν)g(0)
0
γ
d(ν) {01 {
t=to+νΤb

d(ν−3) d(ν−2) d(ν−1)

Tb Tb Tb

g(3) g(2) g(1)

nu_574_BlockschaltbildDF_entz.dsf

+
• System with feedback (recursive) with FIR-Filter in feedback path, nonlinear due to decision device
(quantiser, slicer)
• Error propagation: one bit error may cause successive errors (recursive system!) (usually not a restriction if
Pe<10 –3)
27 01 - Introduction
 Assessment
g(t)
m
g ( 0 ) d (ν ) =y (to + ν Tb ) − ∑ g ( l ) d (ν − l ) g1 nu_574_Impulsantwort.dsf

=1
l g0
g2
FIR-Filter, filters the
bits d (ν - l ) after decision
t
to Tb

The actual symbol d(ν) is weighted with g(0)  large g(0) results in improved decision
accuracy

Therefore: DF-equaliser is well suited for impulse responses with g(0) > g(1), g(2), ...=
compensation of "post-curser-ISI" (after main lobe of pulse)
not so well suited for impulse responses with "pre-cursor-ISI" (in this case g(0) is
small  small effective eye opening)

28 01 - Introduction
 Improvement by FIR-Filter

Remedy: DF-equaliser with FIR-forward filter for minimising pre-cursor ISI new
overall impulse response g= g ∗ e
nu_574_DF_vorentz.dsf

y(ν) ye(ν) yd(ν)

^
d(ν)
e +
+ -
Forward filter

b
yb(ν)
Feedback filter

29 01 - Introduction
 Tomlinson-Harashima Pre-
Coding
Error propagation is the major problem of Decision Feedback-equalizers. If a symbol is decided
erroneously due to noise or distortions, several bit errors are caused due to the feedback structure.

Error propagation is especially strong for large error rates BER > 10−3 .

In many receiver implementations forward error correction (FEC) is used. The BER before FEC
decoding typically lies in that range (e.g. BER > 10−3 ). That is why the use of FEC and DFE is

problematic.

Solution: Shifting the feedback structure to the transmistter side, where undistorted data can be fed
back and the problem of error propagation does not exist.

30 01 - Introduction
 Recap: DFE

1
{01
y(ν) + d(ν) g(0) {
+ 0 d(ν)
y(t) - γ
t=t o+ νΤb

d(ν−3) d(ν−2) d(ν−1)

Tb Tb Tb

g(3) g(2) g(1)

nu_574_BlockschaltbildDF_entz.dsf

The backwards filters contains the channel coefficients g(k) for k=1…m, but not

g(0). Assume g(0)=1 (normalized) then for the backwards filter results:
Impulse response g (k ) − 1γ 0 (k )

z-transform: G( z) − 1
where g (k )  − G ( z ) describes the channel

31 01 - Introduction
 Derivation of the THP
Precoder
Derivation of the THP-precoder from the DFE structure:

G(z)
-
Channel

G(z)-1 a)

G(z)
-
b)
Channel

G(z)-1

32 01 - Introduction
 G(z)
-
Channel

G(z)-1

G(z)
-
Channel

G(z)-1 c)

33 01 - Introduction
 Block Diagram of the TH
Precoder

G(z) d)
-
Channel

G(z)-1

34 01 - Introduction
 Explanation

a) DFE with backwards filter and decider in the feedback loop

b) The feedback structure realizes the inverse channel. For illustration purposes of the THP principle perfect
equalization and no noise are assumed. In that case the signal is perfectly equalized in front of the decider and
input and output values of the decider are identical. The decider can be elliminated or shifted to the output.
(attention: if noise or residual ISI occur shifting of the decider does not lead to identical system properties!)

c) The linear system blocks can be shifted to the transmitter.

d) If the channel G(z) fullfils the minimum phase condition, it has zeros outside the unitary circle so that poles of

the inverse channel 1/G(z) compensate for them. That leads to instability of the inverse channel and the input
signal into the channel could grow to inifinity. To avoid that, the interval is limited to a maximum of M by
applying a modulo function (mod 2M). M is the resolution per dimension (i.e. real or imaginary part oft he
constellation) of the symbols. The modulo operation needs to be reversed at the receiver side.

35 01 - Introduction
 Remarks concerning THP

• THP can be implemented as a stable structure (s. before).

• A feedback channel is required, to inform about the channel estimation results at
the receiver side.
• The Tx signal is not quantized with high granularity anymore. E.g. for 4-ASK
(without THP) only 4 levels exist and a 4-level (2 bit DAC) is sufficient at the Tx.
Using THP a much finer granularity is required.
• As for DFE equalization of pre-cursor ISI is difficult. That is why also for THP most
often an FIR forward filter is used for elimination of the pre-cursors. This can be
implemented either on the Tx or Rx side.

36 01 - Introduction
 Example
a. b. Example of THP using 16-QAM-transmission

c. d.

37 01 - Introduction
 Application example:
Optical Channel

Application:
Equalization of an optical channel
with dispersion impairments
(fiber) and low-pass filtering
(pulse shaping and Rx filter) in the
transmission channel.

38 01 - Introduction
 PN Diagram

39 01 - Introduction
 Comparison: THP vs. DFE

12.5 GBaud +7% FEC overhead  50 Gbit/s net data rate

40 01 - Introduction
 Results

Equalizers B(z) and F(z) are limited to 15 coefficients. This allows equalization up to
a few 100 km (dispersion and related group-velocity impairments increase with the
transmission reach!). From 800…900km ISI occurs, because 15 coefficients are not
sufficient anymore for a good equalization. Then DFE error propagation occurs.

41 01 - Introduction
 Definitions

42 01 - Introduction
 BER & EVM

Source: R.A. Shafik, et al, „On the Extended

1 Relationships Among EVM, BER and SNR as
𝐸𝐸𝐸𝐸𝐸𝐸𝑟𝑟𝑟𝑟𝑟𝑟 = Performance Metrics”, 2006.
𝑆𝑆𝑆𝑆𝑆𝑆

43 01 - Introduction