Linear Binary Codes - Aalborg Universitetkom.aau.dk/project/sipcom/SIPCom06/sites/sipcom8/courses/SIPCom… · SIPCom8-1: Information Theory and Coding Linear Binary Codes Ingmar

SIPCom8-1: Information Theory and Coding

Linear Binary Codes

Ingmar Land

Ingmar Land, SIPCom8-1: Information Theory and Coding (2006 Spring) – p.1

Overview

Basic Concepts of Channel Coding

Block Codes I: Codes and Encoding

Communication Channels

Block Codes II: Decoding

Convolutional Codes


Basic Concepts of Channel Coding

System Model

Examples


The Concept of Channel Coding

Source

Channel

Encoder

DecoderSink

PSfrag replacements

u x

yu

Source generates data

Destination accepts estimated data

Channel introduces noise and thus errors

Encoder adds redundancy

Decoder exploits redundancy to detect or correct errors


The Concept of Channel Coding

Objective of channel coding

Reliable transmissionof digital data

over noisy channels

Tools

Introduction of redundancyfor error detectionor error correction


Examples for Channel Coding

Mobile communications (GSM, UMTS, WLAN, Bluetooth)Channel: mobile radio channel

Satellite communications (pictures from Mars)Channel: radio channel

Cable modems (DSL)Channel: wireline, POTS

Compact Disc, DVD (music, pictures, data)Channel: storage medium

Memory Elements (data)Channel: storage medium

In all digital communication systems, channel coding is appliedto protect the transmitted data against transmission errors.


System Model for Channel Coding I

The Source

Binary Symmetric Source (BSS)

u ∈ F2 := {0, 1}pU (u) = 1/2 for u = 0, 1

Binary Info(rmation) Word of length k

u = [u0, u1, . . . , uk−1] ∈ Fk2

BSS ⇒ pU (u) = 1/2k


System Model for Channel Coding II

The Encoder

Binary Info(rmation) Word of length k

u = [u0, u1, . . . , uk−1] ∈ Fk2

Binary Code Word of length nx = [x0, x1, . . . , xn−1] ∈ F

n2

Linear Binary Code of length nC := {set of codewords x}Linear Binary Encoder- one-to-one map u 7→ x

- map is linear with respect to F2

(addition and multiplication modulo 2)- code rate R := k/n (R < 1)


Examples for Binary Linear Codes

Single Parity-Check Code (k = 2, n = 3)

Code word x = [x0, x1, x2]

with code constraint x0 ⊕ x1 ⊕ x2 = 0

Code C := {000, 110, 101, 011}Possible encoder: u = [u0, u1] 7→ x = [u0, u1, u0 ⊕ u1]

Repetition Code of length (k = 1, n = 3)

Code word x = [x0, x1, x2]

with code constraint x0 = x1 = x2

Code C := {000, 111}Encoder: u = [u0] 7→ x = [u0, u0, u0]


System Model for Channel Coding III

The Channel

Binary-Input Memoryless Channelbinary input alphabet x ∈ F2

real-valued output alphabet y ∈ R

transition probabilities pY |X(y|x)

memoryless independent transmissions

Probabilistic mapping

from input x to output y


Examples for BISMCs

Binary Symmetric Channel(BSC)

0 0

11

PSfrag replacements

X Yε

ε

1 − ε

1 − ε

Crossover probability ε

Binary Erasure Channel(BEC)

0

1 1

0

PSfrag replacements

X Y∆δ

δ

1 − δ

1 − δ

Erasure probability δ


Examples for BISMCs

Binary-Input AWGN Channel (BI-AWGNC)

BPSK Map

PSfrag replacements

XX ′

Y

N

0 7→ +11 7→ −1

X ∈ {0, 1}, X ′ ∈ {−1, +1}, N ∈ R, Y ∈ R

Gaussian distributed noise N with noise variance σ2n

pN (n) =1

√

2πσ2n

exp(− n2

2σ2n

)

Conditional pdf: pY |X′(y|x′) = pN (y − x′)


System Model for Channel Coding IV

The Decoder

Received Word of length ny = [y0, y1, . . . , yn−1] ∈ R

n

Decoder- error correction or error detection- estimation of the transmitted info word or code word

Estimated Info Word of length k

u = [u0, u1, . . . , uk−1] ∈ Fk2


Example for Decoding

Repetition Code over BSC

Code: C := {000, 111}Transmitted code word x = [111]

BSC with crossover probability ε = 0.1

Received word y = [011]

Error detectiony /∈ C ⇒ error detected

Error correction- if x = [111] transmitted, then one error- if x = [000] transmitted, then two errors- one error is less likely than two errors

⇒ estimated code word x = [111](maximum-likelihood decoding)


System Model for Channel Coding V

Source

Channel

Encoder

DecoderSink

PSfrag replacementsu x

yu

Word Error Probability

Pw := Pr(u 6= u) = Pr(x 6= x)

Bit Error Probability

Pb :=1

k

k−1∑

i=0

Pr(ui 6= ui)


Problems in Channel Coding

Construction of good codes

Construction of low-complexity decoders

Existence of codes with certain parameters

(Singleton bound, Hamming bound, Gilbert bound,Varshamov bound)

Highest code rate for a given channel such thattransmission is error-free when codelength tends to infinity

(Channel Coding Theorem)


Block Codes I: Codes and Encoding

Properties of Codes and Encoders

Hamming Distances and Hamming Weights

Generator Matrix and Parity-Check Matrix

Examples of Codes


Codes

A not so accurate Definition (Galois Field)A Galois field GF (q) is a finite set of q elements with twooperators (often called addition and multiplication) such that alloperations are similar to ordinary multiplication and addition forreal numbers. (For the exact definition, see any channel coding textbook.)

Example 1GF (5) = {0, 1, 2, 3, 4}a ⊕ b := a + b mod 5 for a, b ∈ GF (5)a � b := a · b mod 5 for a, b ∈ GF (5)

Example 2: Binary FieldF2 = GF (2) = {0, 1}a ⊕ b := a + b mod 2 for a, b ∈ F2

a � b := a · b mod 2 for a, b ∈ F2


Codes

Definition (Hamming distance)The Hamming distance between two vectors a and b (of thesame length) is defined as the number of positions in which theydiffer, and it is denoted by dH(a, b).

Examplea = [0111], b = [1100], dH(a, b) = 3


Codes

Definition (Linear Binary Block Code)A binary linear (n, k, dmin) block code C is a subset of F

n with2k vectors that have the following properties:

Minimum Distance The minimum Hamming distancebetween pairs of distinct vectors in C is dmin, i.e.,

dmin := mina,b∈Ca6=b

dH(a, b)

Linearity Any sum of two vectors in C is again a vectorin C, i.e.,

a ⊕ b ∈ C for all a, b ∈ C.

The ratio of k and n is called the code rate R = k/n.


Codes

CodewordThe elements x of a code are called codewords, and they arewritten as

x = [x0, x1, . . . , xn−1].

The elements xi of a codeword are called code symbols.

InfowordEach codeword x may be associated with a binary vector oflength k. These vectors u are called infowords (informationwords), and they are written as

u = [u0, u1, . . . , uk−1].

The elements ui of an infoword are called info symbols(information symbols).


Codes

ExampleThis code is a linear binary block code.

C = { [000000], [100111], [010001], [110110],

[001110], [101001], [011111], [111000] }

Code parameters: codeword length n = 6

info word length k = 3

minimum distance dmin = 2

Code parameters in short notation (n, k, dmin) : (6, 3, 2)


Encoding

EncoderAn encoder is a one-to-one map from infowords ontocodewords. The encoder inverse is the inverse map fromcodewords to infowords.

encoder : u 7→ x = enc(u)

encoder inverse : x 7→ u = enc−1(x)

Linear Binary EncoderA linear binary encoder is an encoder for a linear binary codesuch that for all infowords u1 and u2

enc(u1 ⊕ u2) = enc(u1) ⊕ enc(u2).


Encoding

Remark 1Also the device implementing the encoding may be called anencoder.

Remark 2An encoder is a linear map over F2.A code is a linear vector subspace over F2.

(Compare to ordinary linear algebra over real numbers.)


Encoding

ExampleThis encoder for the (6, 3, 3) code is linear.

u 7→ x = enc(u)

000 7→ 000000001 7→ 011011010 7→ 001110011 7→ 010101100 7→ 100111101 7→ 111100110 7→ 101001111 7→ 110010


Encoding

Systematic EncoderAn encoder is called a systematic encoder if all info symbols arecode symbols.

ExampleConsider a (6, 3, 3) code with infowords and codewords denotedby u = [u0, u1, u2] and x = [x0, x1, x2, x3, x4, x5]. The encoder

[u0, u1, u2] 7→ [ u0, u1, u2︸︷︷︸

systematicpart

, x3, x4, x5︸︷︷︸

paritypart

]

is a systematic encoder.


Distances and Weights

Definition (Hamming distance)The Hamming distance dH(a, b) between two vectors a and b (ofthe same length) is defined as the number of positions in whichthey differ.

Definition (Hamming weight)The Hamming weight wH(a) of a vector a is defined as thenumber of non-zero positions.

Examplea = [00111], b = [00011]; wH(a) = 3, wH(b) = 2, dH(a, b) = 1.

Notice: dH(a, b) = wH(a b) = wH(a ⊕ b).



Consider a linear binary code C.

The set of Hamming weights of the codewords in C isdenoted by

wH(C) = {wH(b) : b ∈ C}.

The set of Hamming distances between a codeword a ∈ Cand the other codewords in C is denoted by

dH(a, C) = {dH(a, b) : b ∈ C}.



TheoremFor a linear binary code C, we have

wH(C) = dH(a, C)

for all a ∈ C.

Proof(a) dH(a, b) = wH(a ⊕ b)(b) C = a ⊕ C for all a ∈ CThus for all a ∈ C,

{dH(a, b) : b ∈ C} = {wH(a ⊕ b) : b ∈ C} = {wH(b) : b ∈ C}.



The Hamming distances between codewords are closelyrelated to the error correction/detection capabilities of thecode.

The set of Hamming distances is identical to the set ofHamming weights.

IdeaA code should not only be described by the parameters(n, k, dmin), but also by the distribution of the codewordweights.



Definition (Weight Distribution)Consider a linear code of length n.

The weight distribution of this code is the vector

A = [A0, A1, . . . , An]

with Aw denoting the number of codewords with Hammingweight w, w = 0, 1, . . . , n.

The weight enumerating function (WEF) of this code is thepolynomial

A(H) = A0 + A1H + A2H2 + · · · + AnHn,

where Aw are the elements of the weight distribution, and H is adummy variable.



ExampleConsider the (6, 3, 3) code

C = { [000000], [100111], [010101], [110010],

[001110], [101001], [011011], [111100] }

Weight distribution

A = [1, 0, 0, 4, 3, 0, 0]

Weight enumerating function

A(H) = 1 + 4H3 + 3H4



Consider a binary linear (n, k, dmin) code C.

This code C is a k-dimensional vector subspace of then-dimensional vector space F

n2 (due to the linearity).

Every codeword x can be written as a linear combination ofk basis vectors g0, g1, . . . , gk−1 ∈ C:

x = u0g0 ⊕ u1g1 ⊕ . . . ⊕ uk−1gk−1

= [u0, u1, . . . , uk−1]︸︷︷︸

u

g0

g1

· · ·gk−1

︸︷︷︸

G

= uG



Definition (Generator Matrix)Consider a binary linear (n, k, dmin) code C.A matrix G ∈ F

k×n2 is called a generator matrix of C if the set of

“generated” words is equal to the code, i.e., if

C = {x = uG : u ∈ Fk2}.

Remarks

The rows of G are codewords.

The rank of G is equal to k.

The generator matrix defines an encoder:

x = enc(u) := uG.




C = { [000000], [100111], [010101], [110010],

[001110], [101001], [011011], [111100] }

A generator matrix of this code is

G =

1 0 0 1 1 1

0 1 0 1 0 1

0 0 1 1 1 0



Definition (Parity-Check Matrix)Consider a binary linear (n, k, dmin) code C.A matrix H ∈ F

m×n2 is called a parity-check matrix of C if

C = {x ∈ Fn2 : xHT = 0}.

RemarksThe rows of H are orthogonal to the codewords.

The rank of H is equal to (n − k), and thus m ≥ n − k.Often, one selects H such that m = n − k.

The parity-check matrix defines a code:

x ∈ C ⇔ xHT = 0




C = { [000000], [100111], [010101], [110010],

[001110], [101001], [011011], [111100] }

A parity-check matrix of this code is

H =

1 1 1 1 0 0

1 0 1 0 1 0

1 1 0 0 0 1



Interpretation of Parity-Check MatrixThe equation xHT = 0 represents a system of parity-checkequations.

Example

xH = [x0, x1, x2, x3, x4, x5]

1 1 1 1 0 0

1 0 1 0 1 0

1 1 0 0 0 1

T

= 0

can equivalently be written as

x0 ⊕ x1 ⊕ x2 ⊕ x3 = 0

x0 ⊕ x2 ⊕ x4 = 0

x0 ⊕ x1 ⊕ x5 = 0



Definition (Systematic Generator Matrix)Consider a binary linear (n, k, dmin) code C.A systematic generator matrix Gsyst of C represents a systematicencoder and has the structure

Gsyst = [Ik|P ],

where Ik ∈ Fk2 is the identity matrix and P ∈ F

n−k2 .

TheoremConsider a binary linear (n, k, dmin) code C. If G = [Ik|P ] is agenerator matrix of C, then H = [P T|In−k] is a parity-checkmatrix C, and vice versa.




C = { [000000], [100111], [010101], [110010],

[001110], [101001], [011011], [111100] }

A generator matrix and a parity-check matrix of this code aregiven by

G =

1 0 0 1 1 1

0 1 0 1 0 1

0 0 1 1 1 0

and H =

1 1 1 1 0 0

1 0 1 0 1 0

1 1 0 0 0 1



Definition (Dual Code)Consider a binary linear (n, k) code C with generator matrixG ∈ F

k×n2 and parity-check matrix H ∈ F

(n−k)×n2 .

The binary linear (n, n − k) code C⊥ with generator matrixG⊥ = H and parity-check matrix H⊥ = G is called the dualcode of C.

RemarkThe codewords of the original code and those of the dual codeare orthogonal:

abT = 0 for all a ∈ C and b ∈ C⊥.


Examples of Codes

Binary Repetition CodesLinear binary (n, 1, n) codes:

C = {x ∈ Fn2 : x0 = x1 = · · · = xn−1}

Binary Single Parity-Check CodesLinear binary (n, n − 1, 2) codes:

C = {x ∈ Fn2 : x0 ⊕ x1 ⊕ · · · ⊕ xn−1 = 0}

RemarkRepetition codes and single parity-check codes are dual codes.


Examples of Codes

Binary Hamming Codes

Linear binary codes with dmin = 3 and maximal rate.(Dual codes of the binary Simplex codes.)

Defined by the parity check matrix H and an integer r ∈ N:the columns of H are all non-zero binary vectors of length r.

Resulting code parameters:

codeword length n = 2r − 1,infoword length k = 2r − 1 − r,minimum distance dmin = 3;

Thus: (2r − 1, 2r − 1 − r, 3) code.


Examples of Codes

Example: (7, 4, 3) Hamming code

Follows from r = 3; parity-check matrix:

H =

0 0 0 1 1 1 1

0 1 1 0 0 1 1

1 0 1 0 1 0 1


Examples of Codes

Binary Simplex Codes

Linear binary codes with all non-zero codewords having thesame weight.(Dual codes of the binary Hamming codes.)

Defined by the generator matrix H and an integer r ∈ N:the columns of G are all non-zero binary vectors of length r.

Resulting code parameters:

codeword length n = 2r − 1,infoword length k = r,minimum distance dmin = 2r−1.

Thus: (2r − 1, r, 2r−1) code.


Examples of Codes

Example: (7, 3, 4) Simplex code

Follows from r = 3; generator matrix:

G =

0 0 0 1 1 1 1

0 1 1 0 0 1 1

1 0 1 0 1 0 1


Examples of Codes

Repetition codes

Single parity-check codes

Hamming codes

Simplex codes

Golay codes

Reed-Muller codes

BCH codes

Reed-Solomon codes

Low-density parity-check codes

Concatenated codes (turbo codes)


Summary

Linear Binary (n, k, dmin) Code

Linear (Systematic) Encoder

Weight distribution

(Systematic) Generator matrix

Parity-check matrix

Dual code


Communication Channels

Binary Symmetric Channel

Binary Erasure Channel

Binary Symmetric Erasure Channel

Binary-Input AWGN Channel


Binary Symmetric Channel (BSC)PSfrag replacements

00

1 1

X Yε

ε

1 − ε

1 − ε

Channel input symbols X ∈ {0, 1}Channel output symbols Y ∈ {0, 1}Transition probabilities:

pY |X(y|x) =

{

1 − ε for y = x

ε for y 6= x

Channel parameter: crossover probability ε


Binary Erasure Channel (BEC)PSfrag replacements

00

11

∆X Yδ

δ

1 − δ

1 − δ

Channel input symbols X ∈ {0, 1}Channel output symbols Y ∈ {0, ∆, 1} (∆ = “erasure”)

Transition probabilities:

pY |X(y|x) =

1 − δ for y = x

δ for y = ∆

0 for y 6= x and x, y ∈ {0, 1}

Channel parameter: erasure probability δ


Binary Symmetric Erasure Channel (BSEC)PSfrag replacements

00

11

X Y∆ε

ε

δ

δ

1 − ε − δ

1 − ε − δ

Channel input symbols X ∈ {0, 1}Channel output symbols Y ∈ {0, ∆, 1} (∆ = “erasure”)

Transition probabilities:

pY |X(y|x) =

1 − ε − δ for y = x

δ for y = ∆

ε for y 6= x and x, y ∈ {0, 1}

Channel parameters: erasure probability δcrossover probability ε


Binary-Input AWGN Channel (BI-AWGNC)PSfrag replacements

BPSK Map

XX ′

Y

N ∼ N (0, N0/2)

0 7→ +√

Es1 7→ −√

Es

Code symbols X ∈ {0, 1}Modulation symbols X ′ ∈ {−√

Es, +√

Es}with symbol energy Es

White Gaussian noise (WGN) N ∈ R

with noise variance σ2N = N0/2 and pdf pN (n)

Channel output symbols Y ∈ R

Signal-to-noise ratio (SNR) per code symbol: Es/N0

Transition probabilities: pY |X(y|x) = pY |X′(y|x′) = pN (y−x′)



Equivalent normalized representation

PSfrag replacements

BPSK Map

XX ′

Y

N ∼ N (0, N0/2Es)

0 7→ +11 7→ −1

Code symbols X ∈ {0, 1}Modulation symbols X ′ ∈ {−1, +1}White Gaussian noise (WGN) N ∈ R

with noise variance σ2N = N0/2Es

Channel output symbols Y ∈ R

Signal-to-noise ratio (SNR) per code symbol: Es/N0



Something about energies and SNRs

Assume an (n, k, dmin) code with code rate R = k/n.

Energy per code symbol: Es

Energy per info symbol: Eb = 1REs

SNR per code symbol:Es

N0

SNR per info symbol:Eb

N0=

1

R

Es

N0

Logarithmic scale (often used in error-rate plots):[

Es

N0

]

dB= 10 log10

(Es

N0

)

[dB],

[Eb

N0

]

dB= 10 log10

( Eb

N0

)

[dB]



Error Probability

Assume the detection rule

x′ =

{

+1 for y > 0,−1 for y < 0

, x =

{

0 for x′ = +1,1 for x′ = −1

If y = 0, x′ is randomly chosen out of {−1, +1}.

The error probability can be computed as

Pr(X = 0|X = 1) = Pr(X ′ = +1|X ′ = −1) =

= Pr(N ≥ 1) =

∫ ∞

1pN (n)dn = Q(1/σ2

N ) = Q(√

2Es/N0)

with Q(a) := 1/√

2π∫ ∞a exp(−a2/2)da



Conversion of a BI-AWGNC into a BSC

Assume the previous detection rule applied to anBI-AWGNC with SNR Es/N0

The channel between X and X is then a BSC withcrossover probability

ε = Q(√

2Es/N0)


Block Codes II: Decoding

The Tasks

Decoding Principles

Guaranteed Performance

Performance Bounds


The Tasks of Decoding

Objective of error correction

Given a received word,estimate the most likely (or at least a likely)

transmitted infoword or codeword

Objective of error detection

Given a received word,detect transmission errors

Problem

Decoding complexity


Decoding Principles

Optimum Word-wise Decoding⇒ minimization of word-error probability

Maximum-a-posteriori (MAP) word-estimation

Maximum-likelihood (ML) word-estimation

Optimum Symbol-wise Decoding⇒ minimization of symbol-error probability

Maximum-a-posteriori (MAP) symbol-estimation

Maximum-likelihood (ML) symbol-estimation

Remarks:Word-wise estimation is also called sequence estimation.Symbol-wise estimation is also called symbol-by-symbol estimation


Decoding Principles

Maximum A-Posteriori (MAP) Word Estimation

Estimation of the MAP infoword

uMAP = argmaxu∈F

k2

pU |Y (u|y)

Equivalent two-step operation:

Estimation of the MAP codeword andsubsequent determination of the MAP infoword

xMAP = argmaxx∈C

pX|Y (x|y), uMAP = enc-1(xMAP)


Decoding Principles

Maximum-Likelihood (ML) Word Estimation

Estimation of the ML infoword

uML = argmaxu∈F

k2

pY |U (y|u)

Equivalent two-step operation:

Estimation of the ML codeword andsubsequent determination of the ML infoword

xML = argmaxx∈C

pY |X(y|x), uML = enc-1(xML)


Decoding Principles

Remarks

ML word estimation and MAP word estimation areequivalent if the infowords (codewords) are uniformlydistributed, i.e., if

pU (u) = 2−k

The rules for symbol estimation are similar to the rules forword estimation. (For details, see textbooks.)


Decoding Principles

ML Decoding for the BSCA binary linear (n, k, dmin) code C is used for transmission over abinary symmetric channel (BSC) with crossoverprobability ε < 1/2, i.e.,

pY |X(y|x) =

{

1 − ε for y = x,ε for y 6= x.

Likelihood of a codeword x

pY |X(y|x) =

n−1∏

i=0

pY |X(yi|xi) = εdH(y,x)(1 − ε)n−dH(y,x)


Decoding Principles

Log-likelihood of a codeword x

log pY |X(y|x) = log[

εdH(y,x)(1 − ε)n−dH(y,x)]

= dH(y, x) · logε

1 − ε︸︷︷︸

< 0

+n · log(1 − ε)

Maximum-likelihood word estimation

pY |X(y|x) → max

⇔ log pY |X(y|x) → max

⇔ dH(y, x) → min


Decoding Principles

ML Word-Estimation for the BSC

The Hamming distance dH(y, x) is a sufficient statistic forthe received word y.

ML estimation (in two steps)

x = argminx∈C

dH(y, x), u = enc−1(x)

Decoding for the BSC is also called hard-decision decoding.


Decoding Principles

ML Decoding for the BI-AWGNC

Remember: The code symbols xi ∈ F2 are mapped to codesymbols x′

i ∈ {−1, +1} according to the BPSK mapping

x′i = BPSK(x) =

{

+1 for xi = 0,−1 for xi = 1.

For convenience, we define the BPSK-mapped codewordsand the BPSK-mapped code

x′ = BPSK(x) ∈ {−1, +1}n C′ = BPSK(C).

Notice the one-to-one relation between u, x, and x′.


Decoding Principles

ML Word Estimation for the BI-AWGNC

The squared Euclidean distance is a sufficient statistic forthe received word y:

d2E(y, x′) = ||y − x′||2

ML estimation (in three steps)

x′ = argminx′∈C′

d2E(y, x′), x = BPSK−1(x′), u = enc−1(x).

Decoding for the BI-AWGNC is also called soft-decisiondecoding.



System Model

Binary linear (n, k, dmin) code CBinary symmetric channel (BSC) with crossoverprobability ε < 1

2

ML decoder, i.e., a decoder that applies the rule

x = argminx∈C

dH(y, x)

Questions

How many errors t can guaranteed to be corrected.

How many errors r can guaranteed to be detected.



Most Likely ScenarioConsider the transmitted codeword x and a codeword a that hasthe minimum distance from x, i.e., dH(x, a) = dmin.

Examples

PSfrag replacements x a

dmin = 3


dmin = 4



Number of errors that can be corrected for sure:

t =⌊dmin − 1

2

⌋

Number of errors that can be detected for sure:

r = dmin − 1

Examples


dmin = 3


dmin = 4


Decoding Principles

Bounded Minimum Distance (BMD) Decoding

Decoding rule:

If there is a codeword a ∈ C such that

dH(y, a) ≤ t =⌊dmin − 1

2

⌋

,

output the estimated codeword x = a.

Otherwise, indicate a decoding failure.

RemarkReceived words are decoded only if they are within spheresaround codewords with radius t (Hamming distance), calleddecoding spheres.


Performance Bounds

Coding Scheme

Binary linear (n, k, dmin) code C with WEF A(H)

ML decoder, i.e., a decoder that applies the rule

x = argmaxx∈C

pY |X(y|x)

Question

Bounds on the word-error probability

Pw = Pr(X 6= X)

for a given channel model based on A(H)?


Performance Bounds

Approach

The code is linear, and thus

Pr(X 6= X) = Pr(X 6= X|X = 0) =

= Pr(X 6= 0|X = 0) = Pr(X ∈ C\{0}|X = 0)

Lower bound: for any codeword a ∈ C with wH(a) = dmin,

Pr(X ∈ C\{0}|X = 0) ≥ Pr(X = a|X = 0)

Upper bound: by the union-bound argument,

Pr(X ∈ C\{0}|X = 0) ≤∑

a∈C\{0}

Pr(X = a|X = 0)


Performance Bounds for the BSC

Pairwise word-error probability(Bhattacharyya bound for two codewords)

Pr(X = a|X = b) ≤(√

4ε(1 − ε))dH(a,b)

Lower bound

Pw ≥(√

4ε(1 − ε))dmin

Upper bound

Pw ≤n∑

d=dmin

Ad ·(√

4ε(1 − ε))d


Performance Bounds the BI-AWGNC

Pairwise word-error probability

Pr(X = a|X = b) = Q(√

2dH(a, b)REb/N0)

Lower boundPw ≥ Q(

√

2dminREb/N0)

Upper bound

Pw ≤n∑

d=dmin

Ad · Q(√

2dREb/N0)


Performance Bounds

For improving channel quality, the gap between the lowerbound and the upper bound becomes very small. (ForAdmin

= 1, it vanishes.)

Improving channel quality meansfor the BSC: ε → 1

for the BI-AWGNC: Eb /N0 → ∞


Asymptotic Coding-Gain for the BI-AWGNC

Error probability for the uncoded system(n = k = 1, dmin = 1, R = 1)

Pw = Q(√

2Eb/N0)

Error probability for an (n, k, dmin) code of rate R = k/nin the case of high SNR (corresponds to lower bound)

Pw ≈ Q(√

2dminREb/N0)

The asymptotic coding gain Gasy is the gain in SNR(reduction of SNR) such that the same error probability isachieved as for the uncoded system:

Gasy = 10 log10(dminR)


Summary

Guaranteed PerformanceNumber of errors that can guaranteed to be correctedNumber of errors that can guaranteed to be detected

Maximum-likelihood decodingBSC (“hard-decision decoding”): Hamming distanceBI-AWGNC (“soft-decision decoding”): squaredEuclidean distance

Performance boundsbased on the weight enumerating functionasymptotic coding gain for the BI-AWGNC


Convolutional Codes

General Properties

EncodingConvolutional EncoderState Transition DiagramTrellis Diagram

DecodingML Sequence DecodingViterbi Algorithm


Convolutional Codes

General Properties

Continuous encoding of info symbols to code symbols

Generated code symbols depend on current info symboland previously encoded info symbols

Convolutional encoder is a finite state machine (withmemory)

Certain decoders allow continuous decoding of receivedsequence

⇒ Convolutional codes enable continuous transmission,whereas block codes allow only blocked transmission.


Convolutional Encoder

Defined by generator sequences or generator polynomials

g(0) = [101] corr. to g(D) = 1 + D2

g(1) = [111] corr. to g(D) = 1 + D + D2

or by the shorthand octal notation (5, 7)8

Shift register representation of the encoderPSfrag replacements

u

x(0)

x(1)

D D

Memory length m = number of delay elements

Code rate R = (# info-symbols)/(# code-symbols) = 1/2


Convolutional EncoderPSfrag replacements

ut

x(0)t

x(1)t

s(0)t s

(1)t

Objects of encoding step at time index t ∈ {0, 1, 2, . . . }

info symbol ut ∈ F2

encoder state st = [s(0)t , s

(1)t ] ∈ F

22 (Fm

2 )

code-symbol block xt = [x(0)t , x

(1)t ] ∈ F

22 (F1/R

2 )

→ Generalizations are straight-forward (see textbooks).


State Transition Diagram

All possible state transitions of the convolutional encodermay be depicted in a state transition diagram.

Notation: states s(0)t s

(1)t , labels ut/x

(0)t x

(1)t

PSfrag replacements

00

01

10

11

0/00

0/01

0/10

0/11

1/00

1/01

1/10

1/11


Trellis Diagram

All sequences of state transitions and thus all codewordscan be depicted in a trellis diagram.

ut = 0: solid line, ut = 1: dashed line

Free distance dfree:minimal codeword weight of a detour from the all-zero path

PSfrag replacements00

01

10

11

00

00

00

00

00

00

00

00

00

00

00

00

01

01

01

01

01

01

01

01

01

01

01

01

10

10

10

10

10

10

10

10

10

10

10

10

11

11

11

11

11

11

11

11

11

11

11

11


Block Codes from Convolutional Codes

Encoding of an info sequence of length K with a convolutionalencoder of rate R to a code sequence of length N to obtain an(N, K, dmin) block code of rate RBC

Encoding Strategies

Trellis-truncation constructionK trellis sections, s0 = 0, sK = arbitrary=⇒ N = 1/R · K, dmin ≤ dfree

Trellis-termination constructionK + m trellis sections, s0 = sK+m = 0

=⇒ N = 1/R · (K + m), dmin = dfree

Tail-biting constructionK trellis sections, s0 = sK+m

=⇒ N = 1/R · K, dmin ≤ dfree (better than truncation)


Decoding of Convolutional Codes

Goal: ML word estimation

x = argmaxx∈C

pY |X(y|x)

Possible evaluation:

Compute pY |X(y|x) for all x ∈ C and determine thecodeword x with the largest likelihood

Problem: computational complexity

Solution: Viterbi algorithm


Branches, Paths, and Metrics

Branch in the trellis: one state transition

s[t,t+1] = [st, st+1]

Block of code symbols associated to a branch and block ofobservations corresponding to these code symbols:

x(s[t,t+1]) = xt = [x(0)t , x

(1)t ] yt = [y

(0)t , y

(1)t ]

Path “through” the trellis: sequence of state transitions

s[t1,t2+1] = [st1 , st1+1, . . . , st2 , st2+1]

Partial code word associated to a path:

x(s[t1,t2+1]) = [xt1 , xt1+1, . . . , xt2−1, xt2 ]


Branches, Paths, and Metrics

Branch metric: distance measure between code-symbolblock and observations block,e.g. Hamming distance for the BSC:

µ(s[t,t+1]) = dH(yt, xt)

Path metric: distance measure between code-symbolsequence and sequence of observations:

µ(s[t1,t2+1]) = dH

(

[yt1 , . . . , yt2 ], [xt1 , . . . , xt2 ])

,

The metric should be an additive metric to allow for arecursive computation of the path metric:

µ(s[0,t+1])︸︷︷︸

path metric

= µ(s[0,t])︸︷︷︸

path metric

+ µ(st,t+1)︸︷︷︸

branch metricIngmar Land, SIPCom8-1: Information Theory and Coding (2006 Spring) – p.89

Decoding of Convolutional Codes

Consider a convolutional code with T trellis sections.

ML decoding rule

Search for the path with the smallest metricand determine the associated codeword:

s0,T = argmins0,T∈ trellis

µ(s0,T )

x = x(s0,T )

Viterbi algorithm (VA)trellis-based search algorithmrecursive evaluation of the decoding rule in the trellismost efficient way to implement ML decoding


Viterbi Algorithm

1. Compute branch-metrics. (May be done when necessary.)

2. Step through trellis sections, t = 0, 1, . . . , T − 1.For each trellis state,(a) add previous path metric and current branch metric;(b) compare resulting new path metrics;(c) select the survivor (path with the smallest metric).

3. Trace back to find the ML path.

The resulting ML path corresponds to the ML codeword andthus to the ML infoword.

Remark: The VA may be applied in any situation where (i) thesearch space can be represented in a trellis and (ii) an additivemetric can be defined.


Viterbi Algorithm: Example

Consider a terminated (5, 7)8 convolutional code with infowordlength K = 3 and thus T = 5 trellis sections.

PSfrag replacements

00

01

10

11

00

00

00

00

00

00

00

00

00

00

01

01

01

01

01

01

01

01

01

01

10

10

10

10

10

10

10

10

10

10

11

11

11

11

11

11

11

11

11

11

Remove parts contradicting s0 = sK+m = 0.


Viterbi Algorithm: Example

Consider a terminated (5, 7)8 convolutional code with infowordlength K = 3 and thus T = 5 trellis sections.

PSfrag replacements

00

01

10

11

0000

00

000000

01

01

01

01

10

10

1010

1111

11

11

1111


Summary

Convolutional encoder

State transition diagram

Trellis diagram

Path metrics and branch metrics

Viterbi algorithm


Linear Binary Codes - Aalborg Universitetkom.aau.dk/project/sipcom/SIPCom06/sites/sipcom8/courses/SIPCom… · SIPCom8-1: Information Theory and Coding Linear Binary Codes Ingmar

Documents