Introduction to information theory and coding

Louis WEHENKEL

Channel coding (data transmission)

1. Intuitive introduction
2. Discrete channel capacity
3. Differential entropy and continuous channel capacity
4. Error detecting and correcting codes
5. Turbo-codes

IT 2000-8, slide 1
1. Intuitive introduction to channel coding
How can we communicate in reliable fashion over noisy channels ?

Examples of noisy communication channels :

1. Phone line (twisted pair : thermal noise, distortions, cross-talk. . . )
2. Satellite (cosmic rays. . . )
3. Hard-disk (read or write errors, imperfect material)
Simplistic model : binary symmetric channel (p = probability of error)
1−p
0 0
p

p
1 1
1−p

IT 2000-8, slide 2
Let us suppose that : p = 0.1 (one error every 10 bits, on the average)
But in order to use the channel (say, a hard-disk) : we want to make sure that during
the whole lifecycle of 100 disks there is no more than one error.
E.g. : life period = 10 years. And let us suppose that we transfer 1GB every day on
the disk :
⇒ Pe < 10−15 (requested)

Two engineering approaches :

1. Physical aproach : better circuits, lower density, better cooling, increase signal
power. . .
2. System approach : compensate for the bad performances by using the disk in an
“intelligent” way

W Xn CHANNEL Yn Ŵ
ENCODER DECODER
P (Y|X ) Estimated
Message
message

IT 2000-8, slide 3
Information theory (and coding) ⇒ system approach (and solutions)
Add redundancy to the input message and exploit this redundancy when decoding the
received message
Information theory :
Which are the possibilities (and limitations) terms of performance tradeoffs ?
⇒ analysis problem
Coding theory :
How to build practical error-compensation systems ?
⇒ design problem
(Cf. analogy with “Systems theory vs Control theory”)

IT 2000-8, slide 4
Some preliminary remarks
Analysis problem : - more or less solved for most channels (but not for networks)
Design problam : - Very difficult in general
- Solved in a satisfactory way for a subset of problems only.
Cost of the system approach : - performance tradeoff, computational complexity,
- loss of real-time characteristics (in some cases).
Cost of the physical approach : - investment, power (energy).
⇒ Tradeoff depends on application contexts

Example : (deep space communication)

Each dB of coding gain (which allows to reduce power, size . . . of antennas) leads to
decrease in cost of $80.000.000 for the Deep Space Network!
⇒ coding can make “infeasible” projects “feasible”

IT 2000-8, slide 5
Error handling systems
Retransmission request

Error Channel
avoidance Coded Error
or Demod- Error Error
and Modu- conceal-
Storage ulation detection correction
redundancy lation ment
Medium
encoding
Abstract channel
(Error concealment : exploits natural redundancy to interpolate “missing values”)
NB:
Error detection with retransmission request is an alternative to error correction, but is
not always possible.
In some protocols, error detection merely leads to dropping packets.
⇒ We will come back later to the discussion of “error detection and retransmission”
versus “forward error correction”.

IT 2000-8, slide 6
Codes for detecting and/or correcting errors on the binary symmetric channel
1. Repetition codes :
Source Code
0 000 Decoder : majority vote.
1 111
Example of transmission : T = 0010110.
s 0 0 1 0 1 1 0
x 000 000 111 000 111 111 000
(b: noise vector)
b 000 001 000 000 101 000 000
y 000 001 111 000 010 111 000

Decoding : T̂ = 0010010
Pe (per source bit) : p3 + 3p2 (1 − p) = 0.028 and code rate : R = 1/3
NB: to reach Pe ≤ 10−15 we need R ≤ 1/60 . . .
Other properties : correction of single errors, detection of double errors.

IT 2000-8, slide 7
2. Linear block codes (Hamming (7, 4))
We would like to maximize the code rate under the reliability constraint Pe ≤ 10−15
Block codes : to a block of K source bits we associate a codeword of length N ≥ K.
Example : Hamming (7, 4)
s x s x s x s x
0000 0000000 0100 0100110 1000 1000101 1100 1100011
0001 0001011 0101 0101101 1001 1001110 1101 1101000
0010 0010111 0110 0110001 1010 1010010 1110 1110100
0011 0011100 0111 0111010 1011 1011001 1111 1111111

This code may be written in compact way by (s and x denote line vectors)
1 0 0 0 1 0 1
 
 0 1 0 0 1 1 0 
x = sG with G = 
 0
 = [I 4 P ]
0 1 0 1 1 1 
0 0 0 1 0 1 1

⇒ linear block code (additions and multiplications modulo 2)

IT 2000-8, slide 8
Definition (linear code) We say that a code is linear if all linear combinations of
codewords are also codewords.
Binary codewords of length n form an n−dimensional linear space. A linear code
consists of a linear subspace of this space.
In our example ⇒ first 4 bits = source word, last 3 bits = parity control bits.
E.g. : 5th bit = parity (sum mod. 2) of the first 3 bits.

Some more definitions

Hamming distance d(x, y) of two vectors : nb of bits different.
Hamming weight of a bit-vector : number of bits equal to 1.
Minimum distance of a code : minimum distance of any two codewords
For a linear code : minimum distance = minimum weight of non-zero codewords.
⇒ Minimum distance of Hamming (7, 4) code : = 3

IT 2000-8, slide 9
Decoding :
Let r = x + b denote the received word (b = error vector)
Maximum likelihood decoding :
Guess that codeword x̂ was sent which maximizes the probability p(x̂|r).
Assuming that all codewords are equiprobable, this is equivalent to maximizing p(r|x̂)

If d(x̂, r) = k then p(r|x̂) = pk (1 − p)n−k (here n = 7)

Hence : p(r|x̂) maximal ⇔ d(x̂, r) minimal (supposing that p < 0.5).
For the Hamming (7, 4) code :
If Hamming weight of b ≤ 1 : correct decoding.
If Hamming weight of b ≤ 2 : correct detection.
Otherwise, minimum distance decoding will make errors.

IT 2000-8, slide 10
Pictorial illustration
Repetition code (3, 1)
011 111

001 101

010 110

000
100
⇒ maximum likelihood decoding ≡ nearest neighbor decoding
But : there is a more efficient way to do it than searching explicitly for the nearest
neighbor (syndrom decoding)
IT 2000-8, slide 11
Let us focus on correcting single errors with the Hamming (7, 4) code
Meaning of the 7 bits of a codeword

4 signal bits 3 parity check bits

s1 s2 s3 s4 p1 p2 p3

If error on any of the 4 first bits (signal bits) → two or three parity check violations.
If single error on one of the parity check bits : only this parity check is violated.
In both cases, we can identify erronous bit.
IT 2000-8, slide 12
Alternative representation : parity circles

p2 s4
p3

s3
s2 s1

p1

Parity within each circle must be even if no error.

If this is not the case, we should flip one single received bit so as to realize this
condition (always possible).
For instance, if parity in the green (upper right) and blue (lower) circles are not even
⇒ flip bit s1

IT 2000-8, slide 13
Syndrom decoding
Syndrom : difference between the received parity bit vector and those which are
recomputed from the received signal bits.
⇒ Syndrom is a vector of three bits ⇒ 23 = 8 possible syndroms
The syndrom contains all the information needed for optimal decoding :
8 possible syndroms → 8 most likely error patterns (can be precomputed).
E.g. : suppose that r = 0101111 :
- signal bits 0101 → codeword 0101101 (parity 101)
- syndrom : 101 + 111 = 010 (bit per bit)
- more probable error pattern : 0000010
- decoded word : 0101101
E.g. : suppose that r = 0101110 :
- signal bits 0101 → codeword 0101101 (parity 101)
- syndrom : 101 + 110 = 011 (bit per bit)
- more probable error pattern : 0001000
- decoded signal bits : 0100 (code 0100110).
IT 2000-8, slide 14
Summary
Like the repetition code, the Hamming code (7, 4) corrects single errors and detects
double errors, but uses longer words (7 bits instead of 3) and has a higher rate.
If p = 0.1 : probability of error per code word : 0.14
→ (signal) bit error rate (BER) of : 0.07
Less good in terms of BER but better in terms of code-rate : R = 4/7.
There seems to be a compromize between BER and Code-rate.
Intuitively : limPe →0 R(Pe ) = 0
(this is what most people still believed not so long ago...)
And then ?
...Shannon came...

IT 2000-8, slide 15
Second Shannon theorem
States that if R < C(p) = 1 − H2 (p) then Pe = 0 may be attained.
Third Shannon theorem (rate distorsion : Pe > 0 tolerated)
Using irreversible compression we can further increase the code rate by a factor
1
1−H2 (Pe ) if we accept to reduce realiability (i.e. Pe ր).

C(p)
Conclusion : we can operate in a region satisfying R ≤ 1−H2 (Pe )

log Pe
0.1
Possible
0.01
Impossible

1e-11
Conclusion : we only need two
R disks and a very good code to
0 0.53 1 reach Pe ≤ 10−15 .

IT 2000-8, slide 16
2nd Shannon theorem (channel coding)
What is a channel ?
Later and/or elsewhere

..... Channel ....

Noisy Version of input

Abstract model :
Y1 , Y2 , . . .
P (Y1 , Y2 , . . . |X1 , X2 , . . .)
X1 , X2 , . . .

in other words, the specification of (all) the conditional probability distributions

P (Y1 , . . . , Ym |X1 , . . . , Xn ),
defined ∀m, n = 1, 2, . . .

IT 2000-8, slide 17
Simplifications
Causal channel : if ∀m ≤ n
P (Y1 , . . . , Ym |X1 , . . . , Xn ) = P (Y1 , . . . , Ym |X1 , . . . , Xm ). (1)

Causal and memoryless channel : if ∀k ≥ 2

P (Yk |X1 , . . . , Xk , Y1 , . . . , Yk−1 ) = P (Yk |Xk ), (2)

Causal, memoryless and stationary channel : if ∀k ≥ 1 we have

P (Yk |Xk ) = P (Y|X ), (3)
⇒ this will be our working model
1 symbol enters at time k → 1 symbol comes out at time k.
If stationary process at the input → stationary process out
If ergodic process at the input → ergodic process at the output
(NB: one can generalize to stationary channels of finite memory...)
IT 2000-8, slide 18
Information capacity of a (stationary memoryless) channel
By definition :
C = max I(X ; Y). (4)
P (X )

Remarks
This quantity relates to one single use of the channel (one symbol)
I(X ; Y) depends both on source and channel properties.
C solely depends on channel properties.
We will see later that this quantity coincides with the notion of operational capacity.
NB: How would you generalize this notion to more general classes of channels ?

IT 2000-8, slide 19
Examples of discrete channels and values of capacity
Channel transition matrix :
P (Y1 |X1 ) · · · P (Y|Y| |X1 )
 

[P (Yj |Xi )] = 
 .. .. .. 
. . . 
P (Y1 |X|X | ) · · · P (Y|Y| |X|X | )

1. Binary channel without noise

 
1 0
Input and output alphabets are binary : [P (Yj |Xi )] = .
0 1

I(X ; Y) = H(X ), maximal when H(X ) is maximal (=1 Shannon).

Achievable rate (without errors) : 1 source symbol/channel use.
Can we do better ?
No, unless we admit Pe > 0.

IT 2000-8, slide 20
2. Noisy channel without overlapping outputs
E.g. : transition matrix
 
p (1 − p) 0 0
.
0 0 q (1 − q)
H(X |Y) = 0 ⇒ I(X ; Y) = H(X ) ⇒ C = 1. (Achievable...)
3. Noisy type-writer
Input alphabet : a, b, c,. . . , z Output alphabet : a, b, c,. . . , z
P (a|a) = 0.5, P (b|a) = 0.5, P (b|b) = 0.5, P (c|b) = 0.5, . . . P (z|z) = 0.5, P (a|z) =
0.5
I(X ; Y) = H(Y) − H(Y|X ) with H(Y|X ) = 1 ⇒ max if outputs are equiprobable
E.g. if inputs are equiprobable : H(Y) = log2 26 ⇒ C = log2 13
Achievable : just use the right subset of input alphabet...
(NB: this is THE idea of channel coding)

IT 2000-8, slide 21
4. Binary symmetric channel
1−p
0 0
p  
1−p p
[P (Yj |Xi )] = .
p 1−p
p
1 1
1−p

Information capacity of this channel :

X
I(X ; Y) = H(Y) − H(Y|X ) = H(Y) − P (X)H(Y|X)
X∈X
X
= H(Y) − P (X)H2 (p) = H(Y) − H2 (p) ≤ 1 − H2 (p),
X∈X

Equal to 0, if p = 0.5 and equal to 1 if p = 0.0. Symmetric : C(p) = C(1 − p).

Achievable ? : less trivial...
IT 2000-8, slide 22
Main properties of the information capacity

1. C ≥ 0.
2. C ≤ min{log |X |, log |Y|}.

Moreover, one can show that I(X ; Y) is continuous and concave with respect to
P (X ).
Thus every local maximum must be a global maximum (on the convex set of input
probability distributions P (X )).
Since the function I(X ; Y) is upper bounded capacity must be finite.
One can use powerful optimisation techniques to compute information capacity for a
large class of channels with the desired accuracy.
In general, the solution can not be obtained analytically.

IT 2000-8, slide 23
Communication system

W Xn CHANNEL Yn Ŵ
ENCODER DECODER
P (Y|X ) Estimated
Message
message

A message W (finite set of possible messages W = {1, 2, . . . , M }) is encoded by

the encoder into a sequence of n channel input symbols, denoted by X n (W ).
At the other end of the channel another (random) sequence of channel output symbols
Y n is received (distributed according to P (Y n |X n (W )).)

The sequence Y n is then decoded by the decoder, who chooses an element Ŵ (Y n ) ∈

W → the receiver makes an error if Ŵ (Y n ) 6= W .
In what follows, we will suppose that the encoder and the decoder operate in a deter-
ministic fashion :
- X n (W ) is the coding rule (or function);
- Ŵ (Y n ) is the decoding rule (or function).

IT 2000-8, slide 24
Memoryless channel specification
Input and output alphabets X and Y, and the P (Yk |Xk ) are given

Channel is used without feedback :

Qn
In this case : P (Y n |X n ) = i=1 P (Yi |Xi ).

Definitions that will follow :

- Channel code (M, n)
- Different kinds of error rates...
- Communication rate.
- Achievable rates.
- Operational capacity.

IT 2000-8, slide 25
(M, n) Code
An (M, n) code for a channel (X , P (Y|X ), Y) is defined by

1. A set of indices {1, . . . , M };

2. A coding function X n (·) : {1, . . . , M } → X n , which gives the codebook
X n (1), . . . , X n (M ).
3. A decoding function
g(·) : Y n → {1, . . . , M }, (5)
which is a deterministic mapping from all possible output strings into an input
index g(Y n ).

⇒ M code words coded using n input symbols.

IT 2000-8, slide 26
Decoding error rates
1. Conditional Probability of error given that index i was sent
X
n n n
λi = P (g(Y ) 6= i|X = X (i)) = P (Y n |X n (i))(1 − δg(Y n ),i )
Y n ∈Y n

2. Maximal error probability for and (M, n) code :

λ(n) = max λi
i∈{1,...,M }

3. The (arithmetic) average probability of error :

M
1 X
Pe(n) = λi
M i=1

⇒ expected error rate if i has a uniform distribution.

IT 2000-8, slide 27
Optimal decoding rule
By definition : the decoding rule which minimises expected error rate.
For a received word Y n → choose i such that P (X n (i)|Y n ) is maximal.
⇒ maximises a posteriori probability (MAP)
⇒ minimises for each Y n error probability
⇒ minimises the expected error rate.
⇒ general principle in decision theory : Bayes rule
We use information Y (random variable which is observed).
We want to guess (decide on) a certain variable D (choose among M possibilities).
Correct decision : D∗ a random variable ⇒ P (D∗ |Y) known.
Cost of the taken decision : 0 if correct, 1 if incorrect.
Optimal decision based on information Y : D̂(Y ) = argD max{P (D|Y )}

IT 2000-8, slide 28
For our channel :
n n P (Y n |X n (i))P (X n (i))
P (X (i)|Y ) = PM
n n n
i=1 P (Y |X (i))P (X (i))

PM
Since i=1 P (Y n |X n (i))P (X n (i)) does not depend on the decision, this is the same
than maximizing P (Y n |X n (i))P (X n (i)).
Discussion
P (Y n |X n (i)) : channel specification.
P (X n (i)) : source specification.
If non-redundant source : P (X n (i)) independent of i ⇒ maximize P (Y n |X n (i)).
(n)
⇒ Maximum likelihood rule : minimize Pe
Quasi optimal, if source is quasi non-redundant.
E.g. if we code long source messages (cf. AEP)

IT 2000-8, slide 29
Communication rate : denoted R
The communication rate (denoted by R) of an (M, n) code is defined by R = lognM
Shannon/channel use ≡ input entropy per channel use if inputs are uniformly dis-
tributed.
Achievable rate (more subtle notion)
R is said to be achievable if ∃ a sequence of (M (n), n), n = 1, 2, . . . codes suchthat

1. M (n) = ⌈2nR ⌉ and 2. limn→∞ λ(n) = 0

⇒ codes of rate ≈ R, eventually become “quasi perfect” and remain so (when using
very long codewords)
Remark
Definition is independent of the source distribution (cf. maximal probability of error).
Operational capacity : Co = is the supremum of all achievable rates.
R = 0 is achievable, but we need to check that it is possible to have Co > 0.

IT 2000-8, slide 30
Second Shannon theorem
Objective : prove that information capacity C is equal to operational capacity Co .

Hypothesis : the pair (X n , Y n ) satisfy the AEP (stationary and ergodic : OK for
stationary finite memory channels and ergodic inputs.)
Information capacity (per channel use) :
1 n n
C = lim max {I(X ; Y )}
n→∞ n P (X n )

(with X n stationary and ergodic)

For memoryless channels the maximum is obtained for independent source symbols
and the above definition yields indeed : C = maxP (X ) I(X ; Y)

Basic idea : for large block lengths, every channel looks like the noisy typewriter :
- it has a subset of inputs that produce essentially disjoint sequences at the output
- the rest of the proof is matter of counting and packing...

IT 2000-8, slide 31
Outline of the proof (Shannon’s random coding ideas)
Let us fix for a moment P (X ), n and M , and construct a codebook by generating
random signals according to P (X ) (M × n drawings) ⇒ if n is large the codewords
as well as the received sequences must be typical (AEP)
Xn Yn

X1n

X2n

X3n

Typical transition Non-typical transition

IT 2000-8, slide 32
Let us count and pack (approximately and intuitively...)

At the input : 2nH(X ) possible typical messages (we choose M of them at random to
construct our codebook).
At the output : for each X n (typical) 2nH(Y|X ) possible typical output sequences.

The total number of possible (typical outputs) is 2nH(Y)

If we want that there is no overlap : we must impose that
2nH(Y)
≥ 2nH(Y|X ) ⇒ M ≤ 2nI(Y;X )
M

Now, let us choose the P (X ) which maximizes I(Y; X ) = C : it should be possible

to find M ≤ 2nC input sequences such that the output regions do not overlap.
Conclusion
Using long codewords (n → ∞) it is possible to exploit redundancy (correlation
between inputs and outputs) to transmit information in a reliable way.

IT 2000-8, slide 33
Second Shannon theorem
Statement in two parts :

Forwards : R < C ⇒ R achievable (λ(n) → 0).

Backwards : R > C ⇒ R not achievable.

The proof is based on random coding and the joint AEP.

One can show that for a long codewords, a large proportion of the random codes
generated by drawing M = 2n(C−ǫ) words are actually very good.
Caveat
In practice random coding is not a feasible solution because of computational limi-
tations, just like typical message data compression is not a feasible approach to data
compression.

IT 2000-8, slide 34