Algebraic Coding Theory PDF

29• Information Theory
29• Information Theory
Algebraic Coding Theory Abstract | Full Text: PDF (222K)
Cryptography Abstract | Full Text: PDF (116K)
Data Compression Codes, Lossy Abstract | Full Text: PDF (190K)
Estimation Theory Abstract | Full Text: PDF (161K)
Image Codes Abstract | Full Text: PDF (399K)
Information Theory Abstract | Full Text: PDF (131K)
Information Theory of Data Transmission Codes Abstract | Full Text: PDF (265K)
Information Theory of Modulation Codes and Waveforms Abstract | Full Text: PDF (794K)
Information Theory of Multiaccess Communications Abstract | Full Text: PDF (204K)
Information Theory of Radar and Sonar Waveforms Abstract | Full Text: PDF (410K)
Information Theory of Spread-Spectrum Communication Abstract | Full Text: PDF (141K)
Information Theory of Stochastic Processes Abstract | Full Text: PDF (279K)
Maximum Likelihood Imaging Abstract | Full Text: PDF (290K)
Queueing Theory Abstract | Full Text: PDF (456K)
Signal Detection Theory Abstract | Full Text: PDF (152K)
Trellis-Coded Modulation Abstract | Full Text: PDF (143K)
file:///N|/000000/0WILEY%20ENCYCLOPEDIA%20OF%20ELE...CTRONICS%20ENGINEERING/29.Information%20Theory.htm17.06.2008 14:13:03
file:///N|/000000/0WILEY%20ENCYCLOPEDIA%20OF%20ELECTRICAL...RONICS%20ENGINEERING/29.%20Information%20Theory/W4205.htm
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP
Home / Engineering / Electrical and Electronics Engineering
Wiley Encyclopedia of Electrical and ● Recommend to Browse this title

Your Librarian
Electronics Engineering ● Save title to My ● Search this title
Profile
Algebraic Coding Theory ● Email this page Enter words or
Standard Article phrases
● Print this page
Tor Helleseth Torleiv KlØve1
1University of Bergen, Bergen, Norway
Copyright © 1999 by John Wiley & Sons, Inc. All rights ❍ Advanced Product
reserved. Search
DOI: 10.1002/047134608X.W4205 ❍ Search All Content
Article Online Posting Date: December 27, 1999 ❍ Acronym Finder
Abstract | Full Text: HTML PDF (222K)
Abstract
The sections in this article are
Linear Codes
Some Bounds on Codes
Galois Fields
Cyclic Codes
BCH Codes
Automorphisms
The Weight Distribution of a Code
The Binary Golay Code
Decoding
Reed–Solomon Codes
Nonlinear Codes from Codes over Z

4
file:///N|/000000/0WILEY%20ENCYCLOPEDIA%20OF%20...NGINEERING/29.%20Information%20Theory/W4205.htm (1 of 2)17.06.2008 14:13:16

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

Copyright © 1999-2008John Wiley & Sons, Inc. All Rights Reserved.

402 ALGEBRAIC CODING THEORY
ALGEBRAIC CODING THEORY
In computers and digital communication systems, informa-

tion is almost always represented in a binary form as a se-
quence of bits each having the values 0 or 1. This sequence of
bits is transmitted over a channel from a sender to a receiver.
In some applications the channel is a storage medium like a
CD, where the information is written to the medium at a cer-
tain time and retrieved at a later time. Due to physical limita-
tions of the channel, some of the transmitted bits may be cor-
rupted (the channel is noisy) and thus make it difficult for the
receiver to reconstruct the information correctly.
In algebraic coding theory we are mainly concerned with
developing methods for detecting and correcting errors that
typically occur during transmission of information over a
noisy channel. The basic technique to detect and correct er-
rors is by introducing redundancy in the data that are to be
transmitted. This is similar to communicating in a natural
language in daily life. One can understand the information
while listening to a noisy radio or talking on a bad telephone
line due to the redundancy in the language.
For an example, suppose the sender wants to communicate
one of 16 different messages to a receiver. Each message m
can then be represented as a binary quadruple m ⫽ (c0, c1,
c2, c3). If the message (0101) is transmitted and the first posi-
tion is corrupted such that (1101) is received, this leads to an
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright # 1999 John Wiley & Sons, Inc.
ALGEBRAIC CODING THEORY 403
uncorrectable error since this quadruple represents a differ-

ent valid message than the message that was sent across the 1
channel. The receiver will have no way to detect and correct
a corrupted message in general, since any quadruple repre-
sents a valid message. 0 1
Therefore, to combat errors the sender encodes the data by 0
introducing redundancy into the transmitted information. If
M messages are to be transmitted, the sender selects a subset
of M binary n-tuples, where M ⬍ 2n. Each of the M messages 1 1 0
is encoded into one of the selected n-tuples. The set consisting
of the M n-tuples obtained after encoding is called a binary
(n, M) code and the elements are called codewords. The code- Figure 2. Example of the encoding procedure given in Fig. 1. The
words are sent over the channel. message (0011) is encoded into (0011110). Note that there is an even
It is customary for many applications to let M ⫽ 2k, such number of ones within each circle.
that each message can be represented uniquely by a k-tuple
of information bits. To encode each message the sender can
append n ⫺ k parity bits depending on the message bits and
use the resulting n bit codeword to represent the correspond- area contains a bit ci for i ⫽ 0, 1, . . ., 6. Each of the 16
ing message. possible messages, denoted by (c0, c1, c2, c3), is encoded into a
A binary code C is called a linear code if the sum (modulo codeword (c0, c1, c2, c3, c4, c5, c6), in such a way that the sum of
2) of two codewords is again a codeword. This is always the the bits in each circle has an even parity.
case when the parity bits are linear combinations of the infor- In Fig. 2, an example is shown of encoding the message
mation bits. In this case, the code C is a vector space of di- (0011) into the codeword (0011110). Since the sum of two
mension k over the binary field of two elements, containing codewords also obeys the parity checks and thus is a
M ⫽ 2k codewords, and is called an [n, k] code. The main rea- codeword, the code is a linear [7, 4] code.
son for using linear codes is that these codes have more alge- Suppose, for example, that the transmitted codeword is
braic structure and are therefore often easier to analyze and corrupted in the bit c1 such that the received word is
decode in practical applications. (0111110). Then, calculating the parity of each of the three
The simplest example of a linear code is the [n, n ⫺ 1] circles, we see that the parity fails for the upper circle as well
even-weight code (or parity-check code). The encoding consists as for the leftmost circle while the parity of the rightmost
of appending a single parity bit to the n ⫺ 1 information bits circle is correct. Hence, from the received vector we can in-
so that the codeword has an even number of ones. Thus the deed conclude that bit c1 is in error and should be corrected.
code consists of all 2n⫺1 possible n-tuples of even weight, In the same way, any single error can be corrected by this
where the weight of a vector is the total number of ones in its code.
components. This code can detect all errors in an odd number
of positions, since if such an error occurs the received vector
will also have odd weight. The even-weight code, however, can LINEAR CODES
only detect errors. For example, if (000 . . . 0) is sent and the
first bit is corrupted, then (100 . . . 0) is received. Also, if An (n, M) code is simply a set of M vectors of length n with
(110 . . . 0) was sent and the second bit was corrupted, then components from a finite field F2 ⫽ 兵0, 1其, where addition and
(100 . . . 0) is received. Hence, there is no way the receiver multiplication are done modulo 2. For practical applications
can correct this single error or, in fact, any other error. it is desirable that the code is provided with more structure.
An illustration of a code that can correct any single error Therefore, linear codes are often preferred. A linear [n, k]
is shown in Fig. 1. The three circles intersect and divide the code C is a k-dimensional subspace C of F2n, where F2n is
plane into seven finite areas and one infinite area. Each finite the vector space of n-tuples with coefficients from the finite
field F2.
A linear code C is usually described in terms of a generator
matrix or a parity-check matrix. A generator matrix G of C is
a k ⫻ n matrix whose row space is the code C. That is,
c4
C = {xxG|xx ∈ F2k }
c1 c2
c0 A parity-check matrix H is an (n ⫺ k) ⫻ n matrix such that
c5 c6 C = {cc ∈ F2n |cc H tr = 0 }

c3
where Htr denotes the transpose of H.
Figure 1. The message (c0, c1, c2, c3) is encoded into the codeword
(c0, c1, c2, c3, c4, c5, c6), where c4, c5, c6 are chosen such that there is an Example. The codewords in the code in the previous section
even number of ones within each circle. are the vectors (c0, c1, c2, c3, c4, c5, c6) that satisfy the following
system of parity-check equations: Therefore, finding the minimum distance of a linear code is
equivalent to finding the minimum nonzero weight among all
c0 + c1 + c2 + c4 = 0 codewords in the code.
c0 + c1 + c3 + c5 = 0 If w(c) ⫽ i, then cHtr is the sum of i columns of H. Hence,
c0 + c2 + c3 + c6 = 0 an alternative description of the minimum distance of a linear
code is as follows: the smallest d such that there exists d lin-
where all additions are modulo 2. Each of the three parity- early dependent columns in the parity-check matrix. In par-
check equations correspond to one of the three circles. ticular, to obtain a linear code of minimum distance at least
The coefficient matrix of the parity-check equations is the three, it is sufficient to select the columns of a parity-check
parity-check matrix matrix to be distinct and nonzero.
  Sometimes we include d in the notation and refer to an
1 1 1 0 1 0 0 [n, k] code with minimum distance d as an [n, k, d] code. If t
  components are corrupted during transmission of a codeword,
H = 1 1 0 1 0 1 0  (1)
we say that t errors have occurred or that an error e of weight
1 0 1 1 0 0 1
t has occurred [where e ⫽ (e0, e1, . . ., en⫺1) 僆 F2n, where ei ⫽
1 if and only if the ith component was corrupted—that is, if
The code C is therefore given by
c was sent, c ⫹ e was received].
C = {cc = (c0 , c1 , · · ·, c6 )|ccH tr = 0 } The error-correcting capability of a code is defined as

d−1
A generator matrix for the code in the previous example is t=
2
given by
  where x denotes the largest integer ⱕ x.
1 0 0 0 1 1 1
0 1 0 0 0
A code with minimum distance d can correct all errors of
 1 1 
G=  weight t or less. This is due to the fact that if a codeword c is
0 0 1 0 1 0 1 transmitted and an error e of weight e ⱕ t occurs, the received
0 0 0 1 0 1 1 vector r ⫽ c ⫹ e is closer in Hamming distance to the trans-
mitted codeword c than to any other codeword. Therefore, de-
Two codes are equivalent if the codewords in one of the coding any received vector to the closest codeword corrects all
codes can be obtained by a fixed permutation of the positions errors of weight ⱕ t.
in the codewords in the other code. If G (respectively, H) is a The code can also be used for error detection only. The code
generator (respectively, parity-check) matrix of a code, then is able to detect all errors of weight ⬍ d since if a codeword
the matrices obtained by permuting the columns of these ma- is transmitted and the error has weight ⬍ d, then the re-
trices in the same way give the generator matrix (respec- ceived vector is not another codeword.
tively, parity-check) matrix of the permuted code. The code can also be used for a combination of error correc-
The Hamming distance between x ⫽ (x0, x1, . . ., xn⫺1) and tion and error detection. For a given e ⱕ t, the code can cor-
y ⫽ (y0, y1, . . ., yn⫺1) in F2n is the number of positions in which rect all errors of weight ⱕ e and in addition detect all errors
they differ. That is, of weight at most d ⫺ e ⫺ 1. This is due to the fact that no
vector in F2n can be at distance ⱕ e from one codeword and at
d(xx, y ) = |{i|xi = yi , 0 ≤ i ≤ n − 1}|
the same time at a distance ⱕ d ⫺ e ⫺ 1 from another
codeword. Hence, the algorithm in this case is to decode a
The Hamming distance has the properties required to be a
received vector to a codeword at distance ⱕ e if such a
metric:
codeword exists and otherwise detect an error.
If C is an [n, k] code, the extended code Cext is the [n ⫹ 1,
1. d(x, y) ⱖ 0 for all x, y 僆 F2n and equality holds if and
k] code defined by
only if x ⫽ y.

2. d(x, y) ⫽ d(y, x) for all x, y 僆 F2n.
Cext = (cext , c0 , c1 , . . ., cn−1 )
(c0 , c1 , . . ., cn−1 ) ∈ C,
3. d(x, z) ⱖ d(x, y) ⫹ d(y, z) for all x, y, z 僆 F2n.

n−1
For any code C one of the most important parameters is cext = ci
its minimum distance, defined by i=0
d = min{d(xx, y )|xx = y , x , y ∈ C} That is, each codeword in C is extended by one parity bit such
that the Hamming weight of each codeword becomes even. In
The Hamming weight of a vector x in F2n is the number of particular, if C has odd minimum distance d, then the mini-
nonzero components in x ⫽ (x0, x1, . . ., xn⫺1). That is, mum distance of Cext is d ⫹ 1. If H is a parity-check matrix
for C, then a parity-check matrix for Cext is
w(xx ) = |{i|xi = 0, 0 ≤ i ≤ n − 1}| = d(xx, 0 )

Note that since d(x, y) ⫽ d(x ⫺ y, 0) ⫽ w(x ⫺ y) for a linear 1 1
code C, it follows that 0 tr H
d = min{w(zz )|zz ∈ C, z = 0 } where 1 ⫽ (11 . . . 1).

For any linear [n, k] code C, the dual code C⬜ is the [n, n where e ⫽ (d ⫺ 1)/2. This follows from the fact that the M
⫺ k] code defined by spheres
C⊥ = {xx ∈ F2n |(xx, c ) = 0 for all c ∈ C} Sc = {xx|d(xx, c ) ≤ e}
where (x, c) ⫽ 兺i⫽0 xici. We say that x and c are orthogonal

n⫺1
centered at the codewords c 僆 C are disjoint and that each
if (x, c) ⫽ 0. Therefore, C⬜ consists of all n-tuples that are sphere contains
orthogonal to all codewords in C and vice versa—that is,
e

(C⬜)⬜ ⫽ C. It follows that C⬜ has dimension n ⫺ k since it n
consists of all vectors that are solutions of a system of equa- i
i=0
tions with coefficient matrix G of rank k. Hence, the parity-
check matrix of C⬜ is a generator matrix of C, and similarly vectors.
the generator matrix of C⬜ is a parity-check matrix of C. In If the spheres fill the whole space, that is,
particular, GHtr ⫽ O [the k ⫻ (n ⫺ k) matrix of all zeros].

Sc = F2n
Example. Let C be the [n, n ⫺ 1, 2] even-weight code where c ∈C
 
1 0 ··· 0 0 1 then C is called perfect. The binary linear perfect codes are
0 1 ··· 0 0 1 as follows:
 
G=
 ..

.
.. ..
.
.. .. ... 

. . . • The [n, 1, n] repetition codes for all odd n
0 0 ··· 0 1 1 • The [2m ⫺ 1, 2m ⫺ 1 ⫺ m, 3] Hamming codes Hm for all
mⱖ2
and • The [23, 12, 7] Golay code G23
H = (1 1 ··· 1 1 1)
We will return to the Golay code later.
⬜
Then C has H and G as its generator and parity-check matri-
ces, respectively. It follows that C⬜ is the [n, 1, n] repetition GALOIS FIELDS
code consisting of the two codewords (00 ⭈ ⭈ ⭈ 000) and (11
⭈ ⭈ ⭈ 111). There exist finite fields, also known as Galois fields, with pm
elements for any prime p and any positive integer m. A Galois
Example. Let C be the [2m ⫺ 1, 2m ⫺ 1 ⫺ m, 3] code, where field of a given order pm is unique (up to isomorphism) and is
H contains all nonzero m-tuples as its columns. This is known denoted by Fpm.
as the Hamming code. In the case when m ⫽ 3, a parity-check For a prime p, let Fp ⫽ 兵0, 1, . . ., p ⫺ 1其 denote the inte-
matrix is already described in Eq. (1). Since all columns of gers modulo p with the two operations addition and multipli-
the parity-check matrix are distinct and nonzero, the code has cation modulo p.
minimum distance at least 3. The minimum distance is in- To construct a Galois field with pm elements, select a poly-
deed 3 since there exist three columns whose sum is zero, in nomial f(x) with coefficients in Fp that is irreducible over Fp;
fact the sum of any two columns of H equals another column that is, f(x) cannot be written as a product of two polynomials
in H for this particular code. with coefficients from Fp of degree ⱖ 1 (irreducible polynomi-
The dual code C⬜ is the [2m ⫺ 1, m, 2m⫺1] simplex code all als of any degree m over Fp exist).
of whose nonzero codewords have weight 2m⫺1. This follows Let
since the generator matrix has all nonzero vectors as its col-
umns. In particular, taking any linear combination of rows, Fp m = {am−1 xm−1 + am−2 xm−2 + · · · + a0 |a0 , . . ., am−1 ∈ Fp }
the number of columns with odd parity in the corresponding
subset of rows equals 2m⫺1 (and the number with even parity Then Fpm is a finite field when addition and multiplication of
is 2m⫺1 ⫺ 1). the elements (polynomials) are done modulo f(x) and modulo
The extended code of the Hamming code is a [2m, 2m ⫺ 1 p. To simplify the notations let 움 denote a zero of f(x), that
⫺ m, 4] code. Its dual code is a [2m, m ⫹ 1, 2m⫺1] code that is is, f(움) ⫽ 0. If such an 움 exists, it can formally be defined as
known as the first-order Reed-Muller code. the equivalence class of x modulo f(x). For coding theory, p ⫽
2 is by far the most important case, and we assume this from
now on. Note that for any a, b 僆 F2m,
SOME BOUNDS ON CODES
(a + b)2 = a2 + b2
The Hamming bound states that for any (n, M, d) code we
Example. The Galois field F24 can be constructed as follows.
have
Let f(x) ⫽ x4 ⫹ x ⫹ 1 that is an irreducible polynomial over
F2. Then 움4 ⫽ 움 ⫹ 1 and
e
n
M ≤ 2n
i=0
i F2 4 = {a3 α 3 + a2 α 2 + a1 α + a0 |a0 , a1 , a2 , a3 ∈ F2 }
Computing the powers of 움, we obtain To verify this, one simply computes the coefficients and
uses the preceding table of F24 in the computations. For exam-
α = α · α = α(α + 1) = α + α,
5 4 2
ple,
a6 = α · α 5 = α(α 2 + α) = α 3 + α 2 ,
α 7 = α · α 6 = α(α 3 + α 2 ) = α 4 + α 3 = α 3 + α + 1 m5 (x) = (x + α 5 )(x + α 10 ) = x2 + (α 5 + α 10 )x + α 5 · α 10
= x2 + x + 1
and, similarly, all higher powers of 움 can be expressed as a
linear combination of 움3, 움2, 움, and 1. In particular, 움15 ⫽ 1.
We get the following table of the powers of 움. In the table the This also leads to a factorization into irreducible polynomials:
polynomial a3움3 ⫹ a2움2 ⫹ a1움 ⫹ a0 is represented as a3a2a1a0.
4
14
i 움i i 움i i 움i x2 + x = x (x + α j )
j=0
0 0001 5 0110 10 0111
1 0010 6 1100 11 1110 = x(x + 1)(x2 + x + 1)(x4 + x + 1)
2 0100 7 1011 12 1111 (x4 + x3 + x2 + x + 1)(x4 + x3 + 1)
3 1000 8 0101 13 1101
4 0011 9 1010 14 1001 = xm0 (x)m1 (x)m3 (x)m5 (x)m7 (x)
m
Hence, the elements 1, 움, 움2, . . ., 움14 are all the nonzero In fact, it holds in general that x2 ⫹ x is the product of all
elements in F24. Such an element 움 that generates the nonzero irreducible polynomials over F2 of degree that divides m.
elements of F2m is called a primitive element in F2m. An irreduc- Let Ci ⫽ 兵i2j (mod n) 兩 j ⫽ 0, 1, . . .其, which is called the
ible polynomial g(x) with a primitive element as a root is cyclotomic coset of i (mod n). Then the elements of the cyclo-
called a primitive polynomial. Every finite field has a primi- tomic coset Ci (mod 2m ⫺ 1) correspond to the exponents of
tive element, and therefore the multiplicative subgroup of a the zeros of mi(x). That is,
finite field is cyclic.
m
All elements in F2m are roots of the equation x2 ⫹ x ⫽ 0.
mi (x) = (x − α j )
Let 웁 be an element in F2m. It is important to study the poly-
j∈C i
nomial m(x) of smallest degree with coefficients in F2 that has
웁 as a zero. This polynomial is called the minimal polynomial
of 웁 over F2. The cyclotomic cosets (mod n) are important in the next sec-
␬
First, observe that if m(x) ⫽ 兺i⫽0 mixi has coefficients in tion when cyclic codes of length n are discussed.
F2 and 웁 as a zero, then
2
κ
κ κ
CYCLIC CODES
m(β ) =
2
mi β =
2i
mi β =
2 2i
mi β i
= (m(β ))2 = 0
i=0 i=0 i=0
␬⫺1
Many good linear codes that have practical and efficient de-
Hence, m(x) has 웁, 웁2, . . ., 웁2 , as zeros, where ␬ is the coding algorithms have the property that a cyclic shift of a
␬
smallest integer such that 웁2 ⫽ 웁. Conversely, the polynomial codeword is again a codeword. Such codes are called cyclic
with exactly these zeros can be shown to be a binary irreduc- codes.
ible polynomial. We can represent the set of n-tuples over F2n as polynomials
Example. We will find the minimal polynomial of all the ele- of degree ⬍ n in a natural way. The vector c ⫽ (c0, c1, . . .,
ments in F24. Let 움 be a root of x4 ⫹ x ⫹ 1 ⫽ 0; that is, 움4 ⫽ cn⫺1) is represented as the polynomial c(x) ⫽ c0 ⫹ c1x ⫹ c2x2 ⫹
움 ⫹ 1. The minimal polynomials over F2 of 움i for 0 ⱕ i ⱕ 14 ⭈ ⭈ ⭈ ⫹ cn⫺1xn⫺1. A cyclic shift
are denoted mi(x). Observe by the preceding argument that
m2i(x) ⫽ mi(x), where the indices are taken modulo 15. It fol- σ (cc ) = (cn−1 , c0 , c1 , . . ., cn−2 )
lows that
m0 (x) = (x + α 0 ) = x + 1, of c is then represented by the polynomial
m1 (x) = (x + α)(x + α 2 )(x + α 4 )(x + α 8 )

σ (c(x)) = cn−1 + c0 x + c1 x2 + · · · + cn−2 xn−1
= x4 + x + 1,
= x(cn−1 xn−1 + c0 + c1 x + · · · + cn−2 xn−2 ) + cn−1 (xn + 1)
m3 (x) = (x + α )(x + α )(x + α )(x + α )
3 6 12 9
≡ xc(x) (mod xn + 1)
= x + x + x + x + 1,
4 3 2
m5 (x) = (x + α )(x + α )
5 10
= x2 + x + 1, Example. Rearranging the columns in the parity-check ma-
trix of the [7, 4] Hamming code in Eq. (1), an equivalent code
m7 (x) = (x + α 7 )(x + α 14 )(x + α 13 )(x + α 11 )
is obtained with parity-check matrix
= x4 + x3 + 1,
 
m9 (x) = m3 (x), 1 0 0 1 0 1 1
 
m11 (x) = m7 (x), H = 0 1 0 1 1 1 0 (2)
m13 (x) = m7 (x) 0 0 1 0 1 1 1
This code contains 16 codewords, which are represented Since the generator polynomial of a cyclic code divides
next in polynomial form: xn ⫹ 1, it is important to know how to factor xn ⫹ 1 into irre-
ducible polynomials. Let n be odd. Then there is an integer m
1000110 ↔ x5 + x4 + 1 = (x2 + x + 1)g(x) such that 2m ⬅ 1 (mod n) and there is an element 움 僆 F2m of
0100011 ↔ x6 + x5 + x = (x3 + x2 + x)g(x) order n [if 웆 is a primitive element of F2m, then 움 can be taken
1010001 ↔ x6 + x2 + 1 = (x3 + x + 1)g(x) m
to be 움 ⫽ 웆(2 1)/n].
1101000 ↔ x3 + x + 1 = g(x) We have
0110100 ↔ x4 + x2 + x = xg(x)
0011010 ↔ x5 + x3 + x2 = x2 g(x)
n−1
0001101 ↔ x6 + x4 + x3 = x3 g(x) xn + 1 = (x + α i )
0010111 ↔ x6 + x5 + x4 + x2 = (x3 + x2 )g(x) i=0
1001011 ↔ x6 + x5 + x3 + 1 = (x3 + x2 + x + 1)g(x)

1100101 ↔ x6 + x4 + x + 1 = (x3 + 1)g(x) Let mi(x) denote the minimal polynomial of 움i; that is, the
1110010 ↔ x5 + x2 + x + 1 = (x2 + 1)g(x) polynomial of smallest degree with coefficients in F2 and hav-
0111001 ↔ x6 + x3 + x2 + x = (x3 + x)g(x) ing 움i as a zero. The generator polynomial g(x) can be written
1011100 ↔ x4 + x3 + x2 + 1 = (x + 1)g(x) as
0101110 ↔ x5 + x4 + x3 + x = (x2 + x)g(x)
0000000 ↔ 0= 0 g(x) = (x + α i )
1111111 ↔ x6 + x5 + · · · + x + 1 = (x3 + x2 + 1)g(x) i∈I
By inspection it is easy to verify that any cyclic shift of a where I is a subset of 兵0, 1, . . ., n ⫺ 1其, called the defining
codeword is again a codeword. Indeed, the 16 codewords in set of C with respect to 움. Then mi(x) divides g(x) for all i 僆
I. Further, g(x) ⫽ ⌸ j⫽1 mij(x) for some i1, i2, . . ., il.
l
the code are 0, 1 and all cyclic shifts of (1000110) and
(0010111). The unique nonzero polynomial in the code of low- We can therefore describe the cyclic code in alternative
est possible degree is g(x) ⫽ x3 ⫹ x ⫹ 1, and g(x) is called the equivalent ways as
generator polynomial of the cyclic code. The code consists of
all polynomials c(x) that are multiples of g(x). Note that the C = {c(x)|mi (x) divides c(x), for all i ∈ I},
degree of g(x) is n ⫺ k ⫽ 3 and that g(x) divides x7 ⫹ 1 since C = {c(x)|c(α i ) = 0, for all i ∈ I},
x7 ⫹ 1 ⫽ (x ⫹ 1)(x3 ⫹ x ⫹ 1)(x3 ⫹ x2 ⫹ 1).
The code therefore has a simple description in terms of the C = {cc ∈ F2n |ccH tr = 0 }
set of code polynomials as
where
C = {c(x)|c(x) = u(x)(x3 + x + 1), deg(u(x)) < 4}
 
1 α i1 α 2i 1 ··· α (n−1)i 1
This situation holds in general for any cyclic code. 1 α i2 α 2i 2 ··· α (n−1)i 2 
 
For any cyclic [n, k] code C, we have H=
 .. .. .. .. .. 

. . . . . 
C = {c(x)|c(x) = u(x)g(x), deg(u(x)) < k}
1 α il α 2i l ··· α (n−1)i l
for a polynomial g(x) of degree n ⫺ k that divides xn ⫹ 1.
We can show this as follows: Let g(x) be the generator poly- The encoding for cyclic codes is usually done in one of two
nomial of C, which is the nonzero polynomial of smallest de- ways. Let u(x) denote the information polynomial of degree ⬍
gree r in the code C. Then the cyclic shifts g(x), xg(x), ⭈ ⭈ ⭈ , k. The two ways are as follows:
xn⫺r⫺1g(x) are codewords as well as any linear combination
u(x)g(x), where deg(u(x)) ⬍ n ⫺ r. These are the only 2n⫺r code- 1. Encode into u(x)g(x).
words in the code C, since if c(x) is a codeword then 2. Encode into c(x) ⫽ xn⫺ku(x) ⫹ s(x), where s(x) is the poly-
nomial such that
c(x) = u(x)g(x) + s(x), where deg(s(x)) < deg(g(x))
• s(x) ⬅ xn⫺ku(x) (mod g(x)) [thus g(x) divides c(x)]
• deg(s(x)) ⬍ deg(g(x))
By linearity, s(x) is a codeword and therefore s(x) ⫽ 0 since
deg(s(x)) ⬍ deg(g(x)) and g(x) is the nonzero polynomial of
smallest degree in the code. It follows that C is as described The last of these two methods is systematic; that is, the last
previously. Since C has 2n⫺r codewords, it follows that n ⫺ k bits of the codeword are the information bits.
r ⫽ k; that is, deg(g(x)) ⫽ n ⫺ k.
Finally, we show that g(x) divides xn ⫹ 1. Let c(x) ⫽ c0 ⫹
c1x ⫹ ⭈ ⭈ ⭈ ⫹ cn⫺1xn⫺1 be a nonzero codeword shifted such that BCH CODES
cn⫺1 ⫽ 1. Then a cyclic shift of c(x) given by ␴(c(x)) ⫽ cn⫺1 ⫹
c0x ⫹ c1x ⫹ ⭈ ⭈ ⭈ ⫹ cn⫺2xn⫺1 is also a codeword and An important task in coding theory is to design codes with
a guaranteed minimum distance d that correct all errors of
σ (c(x)) = xc(x) + ϑ (xn + 1) Hamming weight (d ⫺ 1)/2. Such codes were designed inde-
pendently by Bose and Ray-Chaudhuri (1960) and by Hoc-
Since both of the codewords c(x) and ␴(c(x)) are divisible by quenghem (1959) and are known as BCH codes. To construct
g(x), it follows that g(x) divides xn ⫹ 1. a BCH code of designed distance d, the generator polynomial
is chosen to have d ⫺ 1 consecutive powers of 움 as zeros Similarly, a binary triple-error correcting BCH code of the
same length is obtained by choosing the generator polynomial
α ,α
b b+1
, · · ·, α b+d−2
g(x) = m1 (x)m3 (x)m5 (x)

That is, the defining set I with respect to 움 contains a set of
d ⫺ 1 consecutive integers (mod n). The parity-check matrix = (x4 + x + 1)(x4 + x3 + x2 + x + 1)(x2 + x + 1)
of the BCH code is = x10 + x8 + x5 + x4 + x2 + x + 1
 
1 αb α 2b ··· α (n−1)b
The main interest in BCH codes is due to the fact that they
1 α b+1 α 2(b+1) ··· α (n−1)(b+1) 
  have a very fast and efficient decoding algorithm. We describe
H=  .. .. .. .. .. 
 this later.
. . . . . 
(n−1)(b+d−2)
1 α b+d−2
α 2(b+d−2)
··· α
AUTOMORPHISMS
To show that this code has minimum distance at least d,
it is sufficient to show that any d ⫺ 1 columns are linear Let C be a binary code of length n. Consider a permutation 앟
independent. Suppose there is a linear dependency between of the set 兵0, 1, . . ., n ⫺ 1其; that is, 앟 is a one-to-one function
the d ⫺ 1 columns corresponding to 움i1b, 움i2b, . . ., 움id⫺1b. In this of the set of coordinate positions onto itself.
case the (d ⫺ 1) ⫻ (d ⫺ 1) submatrix obtained by retaining For a codeword c 僆 C, let
these columns in H has determinant

π (cc ) = (cπ (0) , cπ (1) , . . ., cπ (n−1) )

α i1 b α i2 b ··· α i d −1 b

i (b+1)

α1 α i 2 (b+1) ··· α i d −1 (b+1)
That is, the coordinates are permuted by the permutation 앟.

.. .. .. ..

If

. . . .

i (b+d−2)

α 2 α2i (b+d−2)
· · · α d −1
i (b+d−2)
{π (cc )|cc ∈ C} = C

1 1 ··· 1

α i1 α i2
· · · α i d −1
then 앟 is called an automorphism of the code C.

b(i 1 +i 2 +...+i d −1 )
=α

.. .. .. ..

. . . .
Example. Consider the following (nonlinear code):

α (d−2)i 1 α (d−2)i 2 · · · α (d−2)i d −1
C = {101, 011}
= α b(i 1 +i 2 +...+i d −1 ) (ai k − α i r ) = 0
k<r
The actions of the six possible permutations on three ele-
since the elements 움 , 움 , ⭈ ⭈ ⭈ , 움
i1 i2 id⫺1
are distinct (the last ments are given in the following table. The permutations that
equality follows from the fact that the last determinant is a are automorphisms are marked by a star.
Vandermonde determinant). It follows that the BCH code has
minimum Hamming distance at least d. 앟(0) 앟(1) 앟(2) 앟((101)) 앟((011))
If b ⫽ 1, which is often the case, the code is called a nar- 0 1 2 101 011
row-sense BCH code. If n ⫽ 2m ⫺ 1, the BCH code is called a 0 2 1 110 011
primitive BCH code. A binary single error-correcting primi- 1 0 2 011 101
tive BCH code is generated by g(x) ⫽ m1(x). The zeros of g(x) 1 2 0 011 110
i
are 움2 , i ⫽ 0, 1, . . ., m ⫺ 1. The parity-check matrix is 2 0 1 110 101
m −2 2 1 0 101 110
H = (1 α 1 α2 ... α2 )
This code is equivalent to the Hamming code since 움 is a In general, the set of automorphisms of a code C is a group,
primitive element of F2m. the Automorphism group Aut(C). We note that
To construct a binary double error-correcting primitive
BCH code, we let g(x) have 움, 움2, 움3, 움4 as zeros. Therefore,
n−1
n−1
xi yi = xπ (i) yπ (i)
g(x) ⫽ m1(x)m3(x) is a generator polynomial of this code. The
i=0 i=0
parity-check matrix of a double error-correcting BCH code is
m
and so (x, y) ⫽ 0 if and only if (앟(x), 앟(y)) ⫽ 0. In particular,
1 α1 α2 · · · α 2 −2
H= m this implies that
1 α 3 α 6 · · · α 3(2 −2)
Aut(C) = Aut(C⊥ )
In particular, a binary double-error correcting BCH code of
length n ⫽ 24 ⫺ 1 ⫽ 15 is obtained by selecting That is, C and C⬜ have the same automorphism group.
g(x) = m1 (x)m3 (x)
= (x4 + x + 1)(x4 + x3 + x2 + x + 1) For a cyclic code C of length n, we have by definition ␴(c)
僆 C for all c 僆 C, where ␴(i) ⬅ i ⫺ 1 (mod n). In particular,
= x8 + x7 + x6 + x4 + 1 ␴ 僆 Aut(C). For n odd, the permutation 웃 defined by 웃( j) ⫽
2j (mod n) is also contained in the automorphism group. To The automorphism group is transitive since the code is cy-
show this it is easier to show that 웃 ⫺1 僆 Aut(C). We have clic, but not doubly transitive. For example, there is no auto-
morphism 앟 such that 앟(0) ⫽ 0 and 앟(3) ⫽ 1 since 0 and 1 are
δ −1 (2 j) = j for j = 0, 1, . . ., (n − 1)/2, not equivalent modulo 3. A simple counting argument shows
δ −1 (2 j + 1) = (n + 1)/2 + j for j = 0, 1, . . ., (n − 1)/2 − 1 that Aut(C) has order 1296: First choose 앟(0); this can be done
in 9 ways. There are then 2 ways to choose 앟(3) and 앟(6). Next
Let g(x) be a generator polynomial for C, and let 兺i⫽0 cixi ⫽
n⫺1 choose 앟(1); this can be done in 6 ways. There are again 2
a(x)g(x). Since xn ⬅ 1 (mod xn ⫹ 1), we have ways to choose 앟(4) and 앟(7). Finally, there are 3 ⭈ 2 ways to
choose 앟(2), 앟(5), 앟(8). Hence, the order is 9 ⭈ 2 ⭈ 6 ⭈ 2 ⭈ 3 ⭈ 2 ⫽

n−1
(n−1)/2
(n−1)/2−1 1296.
cδ −1 (i) xi ≡ c j x2 j + c(n+1)/2+ j x2 j+1+n
i=0 j=0 j=0 Example. Consider the extended Hamming code Hext m . The

(n−1)/2
n−1 positions of the codewords correspond to the elements of F2m
= c j x2 j + c j x2 j and are permuted by the affine group
j=0 j=(n+1)/2
= a(x2 )g(x2 ) = (a(x2 )g(x))g(x), (mod xn + 1) AG = {π|π (x) = ax + b, a, b ∈ F2 m , a = 0}
and so 웃⫺1(c) 僆 C; that is, 웃⫺1 僆 Aut(C) and so 웃 僆 Aut(C). This is the automorphism group of Hext
m . It is double transitive.
The automorphism group Aut(C) is transitive if for each
pair (i, j) there exists a 앟 僆 Aut(C) such that 앟(i) ⫽ j. More
general, Aut(C) is t-fold transitive if, for distinct i1, i2, . . ., it
and distinct j1, j2, . . ., jt, there exists a 앟 僆 Aut(C) such that THE WEIGHT DISTRIBUTION OF A CODE
앟(i1) ⫽ j1, 앟(i2) ⫽ j2, . . ., 앟(it) ⫽ jt.
Let C be a binary linear [n, k] code. As we noted before,
Example. Any cyclic [n, k] code has a transitive automor-
phism group since ␴ repeated s times, where s ⬅ i ⫺ j (mod d(xx, y ) = d(xx − y , 0 ) = w(xx − y )
n), maps i to j.
If x, y 僆 C, then x ⫺ y 僆 C by the linearity of C. In particular,
Example. The (nonlinear) code C ⫽ 兵101, 011其 was consid- this means that the set of distances from a fixed codeword to
ered previously. Its automorphism group is not transitive all the other codewords is independent of which codeword we
since there is no automorphism 앟 such that 앟(0) ⫽ 2. fix; that is, the code looks the same from any codeword. In
particular, the set of distances from the codeword 0 is the set
Example. Let C be the [9, 3] code generated by the matrix of Hamming weights of the codewords. For i ⫽ 0, 1, . . ., n,
let Ai denote the number of codewords of weight i. The se-
 
0 0 1 0 0 1 0 0 1 quence
 
0 1 0 0 1 0 0 1 0
A0 , A1 , A2 , . . ., An
1 0 0 1 0 0 1 0 0
is called the weight distribution of the code C. The correspond-
This is a cyclic code and we will determine its automorphism ing polynomial
group. The all zero and the all one vectors in C are trans-
formed into themselves by any permutation. The vectors of AC (z) = A0 + A1 z + A2 z2 + · · · + An zn
weight 3 are the rows of the generator matrix and the vectors
of weight 6 are the complements of these vectors. Hence, we
see that 앟 is an automorphism if and only if it leaves the set is known as the weight enumerator polynomial of C.
of the three rows of the generator matrix invariant, that is, if The polynomials AC(z) and AC⬜(z) are related by the funda-
and only if the following conditions are satisfied: mental MacWilliams identity:

π (0) ≡ π (3) ≡ π (6) (mod 3), 1−z
AC ⊥ (z) = 2−k (1 + z)n AC
1+z
π (1) ≡ π (4) ≡ π (7) (mod 3),
π (2) ≡ π (5) ≡ π (8) (mod 3).
Example. The [2m ⫺ 1, m] simplex code has the weight enu-
m⫺1
merator polynomial 1 ⫹ (2m ⫺ 1)z2 . The dual code is the
Note that the two permutations ␴ and 웃 defined previously [2 ⫺ 1, 2 ⫺ 1 ⫺ m] Hamming code with weight enumerator
m m
satisfy these conditions, as they should. They are listed ex- polynomial
plicitly in the following table
 
2 m−1
i 0 1 2 3 4 5 6 7 8 2 −m
(1 + z) 2 m −1 1 + (2 − 1) 1 − z
m 
1+z
␴(i) 8 0 1 2 3 4 5 6 7
m −1 m−1 m−1 −1
웃(i) 0 2 4 6 8 1 3 5 7 = 2−m (1 + z)2 + (1 − 2−m )(1 − z)2 (1 + z)2
For example, for m ⫽ 4, we get the weight distribution of the The weight distribution of G23 is given by the following
[15, 11] Hamming code: table:
1 + 35z3 + 105z4 + 168z5 + 280z6 + 435z7 + 435z8 + 280z9 i Ai

+ 168z 10
+ 105z 11
+ 35z 12
+z 15
0, 23 1
7, 16 253
8, 15 506
Consider a binary linear code C that is used purely for er- 11, 12 1288
ror detection. Suppose a codeword c is transmitted over a bi-
nary symmetric channel with bit error probability p. The The automorphism group Aut(G23) of the Golay code is the
probability of receiving a vector r at distance i from c is
Mathieu group M23, a simple group of order 10200960 ⫽ 27 ⭈
pi(1 ⫺ p)n⫺i, since i positions are changed (each with probabil-
32 ⭈ 5 ⭈ 7 ⭈ 11 ⭈ 23, which is four-fold transitive.
ity p) and n ⫺ i are unchanged (each with probability 1 ⫺
Much information about G23 can be found in the book by
p). If r is not a codeword, then this will be discovered by the
MacWilliams and Sloane (see Reading List).
receiver. If r ⫽ c, then no errors have occurred. However, if
r is another codeword, then an undetectable error has oc-
curred. Hence, the probability of undetected error is given by DECODING

Pue (C, p) = pd(cc ,cc ) (1 − p)n−d(cc ,cc ) Suppose that a codeword c from the [n, k] code C was sent
c =c
and that an error e occurred during the transmission over the
)
= pw(cc ) (1 − p)n−w(cc noisy channel. Based on the received vector r ⫽ c ⫹ e, the
c =0 receiver has to make an estimate of what was the transmitted

n codeword. Since error patterns of lower weight are more prob-
= Ai pi (1 − p)n−i able than error patterns of higher weight, the problem is to
i=1 estimate an error ê such that the weight of ê is as small as

p possible. He will then decode the received vector r into ĉ ⫽
= (1 − p)n AC − (1 − p)n
1− p r ⫹ ê.
If H is a parity-check matrix for C, then cHtr ⫽ 0 for all
From the MacWilliams identity we also get codewords c. Hence,
Pue (C⊥ , p) = 2−k AC (1 − 2p) − (1 − p)n r H tr = (cc + e )H tr = c H tr + e H tr = e H tr (3)
Example. For the [2m ⫺ 1, 2m ⫺ 1 ⫺ m] Hamming code Hm, The vector

we get
m−1 m −1 s = e H tr
Pue (Hm , p) = 2−m (1 + (2m − 1)(1 − 2p)2 ) − (1 − p)2
is known as the syndrome of the error e; Eq. (3) shows that s
More information on the use of codes for error detection can can be computed from r. We now have the following outline
be found in the book by Kløve and Korzhik (see Reading List). of a decoding strategy:
1. Compute the syndrome s ⫽ rHtr.

THE BINARY GOLAY CODE 2. Estimate an error ê of smallest weight corresponding to
the syndrome s.
The Golay code G23 has received much attention. It is practi- 3. Decode to ĉ ⫽ r ⫹ ê.
cally useful and has a number of interesting properties. The
code can be defined in various ways. One definition is that
G23 is the cyclic code generated by the irreducible polynomial The hard part is, of course, step 2.
For any vector x 僆 F2n, the set 兵x ⫹ c 兩 c 僆 C其 is a coset of
x11 + x9 + x7 + x6 + x5 + x + 1 C. All the elements of the coset have the same syndrome—
namely, xHtr. There are 2n⫺k cosets, one for each syndrome in
which is a factor of x23 ⫹ 1 over F2. Another definition is the F2n⫺k, and the set of cosets is partition of F2n. We can rephrase
following: Let H denote the [7, 4] Hamming code and let H* step 2 as follows: Find a vector e of smallest weight in the
be the code whose codewords are the reversed of the code- coset with syndrome s.
words of H. Let
Example. Let C be the [6, 3, 3] code with parity-check ma-
C = {(u u, v ∈ H ext , x ∈ (H ∗ )ext}
u + x , v + x , u + v + x )|u trix
 
where Hext is the [8, 4] extended Hamming code and (H*)ext is 1 1 0 1 0 0
the [8, 4] extended H*. The code C is a [24, 12, 8] code. Punc-  
H = 1 0 1 0 1 0
turing the last position, we get a [23, 12, 7] code that is
0 1 1 0 0 1
(equivalent to) the Golay code.
A standard array for C is the following array (the eight col- This implies that S13 ⫽ S3 ⫹ 움i움jS1 ⬆ S3. Furthermore,
umns to the right): x1 ⫽ 움⫺i and x2 ⫽ 움⫺j are roots of the equation
000 111000 001011 010101 011110 100110 101101 110011 111000
S31 + S3 2
110 100000 101011 110101 111110 000110 001101 010011 011000 1 + S1 x + x =0 (4)
101 010000 011011 000101 001110 110110 111101 100011 101000 S1
011 001000 000011 011101 010110 101110 100101 111011 110000
100 000100 001111 010001 011010 100010 101001 110111 111100 This gives the following procedure to correct two errors:
010 000010 001001 010111 011100 100100 101111 110001 111010
001 000001 001010 010100 011111 100111 101100 110010 111001
• Compute S1 and S3.
111 100001 101010 110100 111111 000111 001100 010010 011001
• If S1 ⫽ S3 ⫽ 0, then assume that no errors have occurred.
Each row in the array is a listing of a coset of C; the first row • Else, if S3 ⫽ S13 ⬆ 0, then one error has occurred in the
is a listing of the code itself. The vectors in the first column ith position determined by S1 ⫽ 움i.
have minimal weight in their cosets and are known as coset • Else (if S3 ⬆ S13), consider the equation
leaders. The choice of coset leader may not be unique. For
example, in the last coset there are three vectors of minimal 1 + S1 x + (S31 + S3 )/S1 x2 = 0
weight. Any entry in the array is the sum of the codeword at
If the equation has two roots 움⫺i and 움⫺j, then errors
the top of the column and the coset leader (at the left in the
have occurred in positions i and j.
row). Each vector of F 26 is listed exactly once in the array. The
standard array can be used for decoding: Locate r in the array Else (if the equation has no roots in F2m), then more
and decode to the codeword at the top of the corresponding than two errors have occurred.
column (that is, the coset leader is assumed to be the error
pattern). However, this is not a practical method; except for Similar explicit expressions (in terms of the syndrome) for
small n, the standard array of 2n entries is too large to store the coefficients of an equation with the error positions as
(also locating r may be a problem). A step in simplifying the roots can be found for t error-correcting BCH codes when t ⫽
method is to store a table of coset leaders corresponding to 3, t ⫽ 4, etc., but they become increasingly complicated. How-
the 2n⫺k syndromes. In the preceding table this is illustrated ever, there is an efficient algorithm for determining the equa-
by listing the syndromes at the left. Again this is a possible tion, and we describe this is some detail next.
alternative only if n ⫺ k is small. For carefully designed Let 움 be a primitive element in F2m. A parity-check matrix
codes, it is possible to compute e from the syndrome. The sim- for the primitive t error-correcting BCH code is
plest case is single errors: If e is an error pattern of weight 1,  
where the 1 is in the ith position, then the corresponding syn- 1 α α2 ··· α n−1
 
drome is the ith column of H; hence, from H and the syn- 1 α3 α6 ··· α 3(n−1) 
 
drome we can determine i. H = . .. .. .. 
. .. 
. . . . . 
Example. Let H be the m ⫻ (2m ⫺ 1) parity-check matrix 1 α 2t−1 α 2(2t−1) ··· α (2t−1)(n−1)
where the ith column is the binary expansion of the integer i
for i ⫽ 1, 2, . . ., 2m ⫺ 1. The corresponding [2m ⫺ 1, 2m ⫺
where n ⫽ 2m ⫺ 1. Suppose errors have occurred in positions
1 ⫺ m, 3] Hamming code corrects all single errors. Decoding i1, i2, . . ., i␶, where ␶ ⱕ t. Let Xj ⫽ 움ij for j ⫽ 1, 2, . . ., ␶. The
is done as follows: Compute the syndrome s ⫽ (s0, s1, . . .,
error locator polynomial ⌳(x) is defined by
sm⫺1). If s ⬆ 0, then correct position i ⫽ 兺j⫽0 sj 2 j.
m⫺1

τ
τ
Example. Let (x) = (1 + X j x) = λ l xl
j=1 l=0
1 α α2 ··· α n−1
H= The roots of ⌳(x) ⫽ 0 are X⫺1
j . Therefore, if we can determine
1 α3 α6 ··· α 3(n−1)
⌳(x), then we can determine the locations of the errors. Ex-
panding the expression for ⌳(x), we get
where 움 僆 F2m and n ⫽ 2m ⫺ 1. This is the parity-check matrix
for the double error-correcting BCH code. It is convenient to λ0 = 1,
have a similar representation of the syndromes:
λ1 = X1 + X2 + · · · + Xτ ,
s = (S1 , S3 ) where S1 , S3 ∈ F2 m λ2 = X1 X2 + X1 X3 + X2 X3 + · · · + Xτ −1 Xτ ,
λ3 = X1 X2 X3 + X1 X2 X4 + X2 X3 X4
Depending on the syndrome, there are several cases:
+ · · · + Xτ −2 Xτ −1 Xτ ,
1. If no errors have occurred, then clearly S1 ⫽ S3 ⫽ 0. ..
.
2. If a single error has occurred in the ith position (that
is, the position corresponding to 움i), then S1 ⫽ 움i and λτ = X1 X2 · · · Xτ ,
S3 ⫽ 움3i. In particular, S3 ⫽ S13. λl = 0 for l > τ
3. If two errors have occurred in positions i and j, then
Hence ␭l is the lth elementary symmetric function of X1, X2,
S1 = α i + α j , S3 = α 3i + α 3 j . . ., X␶.
From the syndrome we get S1, S3, . . ., S2t⫺1, where will have the required property. If ⌳(r)(x) ⬆ 1, then there ex-
ists a maximal positive integer ␳ ⬍ r such that 웆(2␳␳)⫹1 ⬆ 0 and
S1 = X1 + X2 + · · · + Xτ , we add a suitable multiple of ⌳(␳):
S2 = X12 + X22 + · · · + Xτ2 ,
(r+1) (x) = (r) (x) + ω2r+1
(r) ρ
(ω2ρ+1 )−1 x2r−2ρ (ρ ) (x)
S3 = X13 + X23 + ··· + Xτ3 ,
.. We note that this implies that
.

S2t = X12t + X22t + · · · + Xτ2t (r+1) (x)S(x) = wl(r) xl + ω2r+1
(r) (ρ )
(ω2ρ+1 )−1 ωl(ρ ) xl+2r−2ρ
l≥0 l≥0
Further,
Hence for odd l we get
S2r = X12r + X22r + · · · + Xτ2r = (X1r + X2r + · · · + Xτr )2 = S2r  (r)

 ωl = 0


for all r. Hence, from the syndrome we can determine the 
 for 1 ≤ l ≤ 2r − 2ρ − 1,

 ω (r) + ω (r) (ω (ρ ) )−1 ω (ρ )
polynomial =0+0=0
ωl(r+1) = l 2r+1 2ρ+1 l−2r+2ρ

 for 2r − 2ρ + 1 ≤ l ≤ 2r − 1,

 (r) (r) (ρ ) −1 (ρ ) (r) (r)
S(x) = 1 + S1 x + S2 x2 + · · · + S2t x2t  ω
 2r+1 + ω (ω ) ω = ω + ω2r+1 =0

 2r+1 2ρ+1 2ρ+1 2r+1
for l = 2r + 1
The Newton equations are a set of relations between the
power sums Sr and the symmetric functions ␭l —namely, We now formulate these ideas as an algorithm (in a Pas-
cal-like syntax). In each step we keep the present ⌳(x) [the

l−1
superscript (r) is dropped] and the modifying polynomial
Sl− j λ j + lλl = 0 for l ≥ 1
[x2r⫺2␳⫺1 or (웆(2␳␳)⫹1)⫺1x2r⫺2␳⫺1⌳(␳)(x)], which we denote by B(x).
j=0
Let Berlekamp–Massey algorithm in the binary case

(x) = S(x)(x) = ω l x (5) Input: t and S(x).
l≥0
⌳(x) :⫽ 1; B(x) :⫽ 1;
Since 웆l ⫽ 兺
l⫺1
j⫽0 Sl⫺j␭j ⫹ ␭l, the Newton equations imply that for r :⫽ 1 to t do
begin
ωl = 0 for all odd l, 1 ≤ l ≤ 2t − 1 (6) 웆 :⫽ coefficient of x2r⫺1 in S(x)⌳(x);
if 웆 ⫽ 0 then B(x) :⫽ x2B(x)
The Berlekamp–Massey algorithm is an algorithm that, else [⌳(x), B(x)] :⫽ [⌳(x) ⫹
given S(x), determines the polynomial ⌳(x) of smallest degree 웆xB(x), x⌳(x)/웆]
such that Eq. (6) is satisfied, where the 웆l are defined by Eq. end;
(5). The idea is, for r ⫽ 0, 1, . . ., t, to determine polynomials
⌳(r) of lowest degree such that The assignment following the else is two assignments to be
done in parallel; the new ⌳(x) and B(x) are computed from the
ωl(r) = 0 for all odd l, 1 ≤ l ≤ 2r − 1 old ones.
The Berlekamp–Massey algorithm determines the polyno-
mial ⌳(x). To find the roots of ⌳(x) ⫽ 0, we try all possible
where elements of F2m. In practical applications, this can be effi-
ciently implemented using shift registers (usually called the
ωl(r) xl = S(x)(r) (x) Chien search).
l≥0
Example. We consider the [15, 7, 5] double-error correcting

For r ⫽ 0, we can clearly let ⌳(0)(x) ⫽ 1. We proceed by
BCH code; that is, m ⫽ 4 and t ⫽ 2. As a primitive element,
induction. Let 0 ⱕ r ⬍ t, and suppose that polynomials ⌳(␳)(x)
we choose 움 such that 움4 ⫽ 움 ⫹ 1. Suppose that we have
have been constructed for 0 ⱕ ␳ ⱕ r. If 웆(r)
2r⫹1 ⫽ 0, then we can
received a vector with syndrome (S1, S3) ⫽ (움4, 움5). Since
choose
S3 ⬆ S13, at least two errors have occurred. Equation (4) be-
comes
(r+1) (x) = (r) (x)
1 + α 4 x + α 10 x2 = 0
If, on the other hand, 웆(r)
2r⫹1 ⬆ 0, then we modify ⌳ (x) by add-
(r)
ing another suitable polynomial. There are two cases to con- which has the zeros 움⫺3 and 움⫺7. We conclude that the re-
sider. First, if ⌳(r)(x) ⫽ 1 [in which case ⌳(r)(x) ⫽ 1 for 0 ⱕ ceived vector has two errors (namely, in positions 3 and 7).
␶ ⱕ r], then
Now consider the Berlekamp–Massey algorithm for the
(r+1) (x) = 1 + ω2r+1
(r)
x2r+1
same example. First we compute S2 ⫽ S12 ⫽ 움8 and S4 ⫽ S22 ⫽
움. Hence Thus, the Reed–Solomon codes satisfy the Singleton bound

with equality n ⫺ k ⫽ d ⫺ 1. That is, they are MDS codes.
S(x) = 1 + α x + α x + α x + αx
4 8 2 5 3 4
The weight distribution of the Reed–Solomon code is (for
The values of r, 웆, ⌳(x), and B(x) after each iteration of the i ⱖ d)
for-loop in the Berlekamp–Massey algorithm are shown in
the following table: n i−d
j i
Ai = (−1) (2m(i−d− j+1) − 1)
i j=0 j
r 웆 ⌳(x) B(x)
1 1 The encoding of Reed–Solomon codes is similar to the en-
1 움4 1 ⫹ 움4x 움11x coding of binary cyclic codes. The decoding is similar to the
2 움14 1 ⫹ 움4x ⫹ 움10x2 움x ⫹ 움5x2 decoding of binary BCH codes with one added complication.
Using a generalization of the Berlekamp–Massey algorithm,
Hence, ⌳(x) ⫽ 1 ⫹ 움4x ⫹ 움10x2 (as before). we determine the polynomials ⌳(x) and ⍀(x). From ⌳(x) we
Now consider the same code with syndrome of received can determine the locations of the errors. In addition, we have
vector (S1, S3) ⫽ (움, 움9). Since S3 ⬆ S13, at least two errors have to determine the value of the errors (in the binary case the
occurred. We get values are always 1). The value of the error at location Xj can
easily be determined using ⍀(x) and ⌳(x); we omit further de-
(x) = 1 + αx + x2 tails.
However, the equation 1 ⫹ 움x ⫹ x2 ⫽ 0 does not have any
roots in F24. Hence, at least three errors have occurred, and NONLINEAR CODES FROM CODES OVER Z4
the code is not able to correct them.
In the previous sections we have mainly considered binary
REED–SOLOMON CODES linear codes; that is, codes where the sum of two codewords is
again a codeword. The main reason has been that the linear-
In the previous sections we have considered binary codes ity greatly simplified construction and decoding of the codes.
where the components of the codewords belong to the finite A binary nonlinear (n, M, d) code C is simply a set of M
field F2 ⫽ 兵0, 1其. In a similar way we can consider codes with binary n-tuples with pairwise distance at least d, but without
components from any finite field Fq. any further imposed structure. In general, to find the mini-
The Singleton bound states that for any [n, k, d] code with mum distance of a nonlinear code one has to compute the dis-
components from Fq, we have tance between all pairs of codewords. This is, of course, more
complicated than for linear codes, where it suffices to find the
d ≤ n−k+1
minimum weight among all the nonzero codewords. The lack
A code for which d ⫽ n ⫺ k ⫹ 1 is called maximum distance of structure in a nonlinear code also makes it quite difficult
separable (MDS). The only binary MDS codes are the trivial to decode in an efficient manner.
[n, 1, n] repetition codes and [n, n ⫺ 1, 2] even-weight codes. There are, however, some advantages to nonlinear codes.
However, there are important nonbinary MDS codes (in par- For given values of length n and minimum distance d, it is
ticular, the Reed–Solomon codes, which we now will de- sometimes possible to construct nonlinear codes with more
scribe). codewords than is possible for linear codes. For example, for
Reed–Solomon codes are t error-correcting cyclic codes n ⫽ 16 and d ⫽ 6 the best linear code has dimension k ⫽ 7
with symbols from a finite field Fq, even though they can be (i.e., it contains 128 codewords). The code of length 16 ob-
constructed in many different ways. They can be considered tained by extending the double-error-correcting primitive
as the simplest generalization of BCH codes. Since the most BCH code has these parameters.
important case for applications is q ⫽ 2m, we consider this In 1967, Nordstrom and Robinson found a nonlinear code
case here. Each symbol is then an element in F2m and can be with parameters n ⫽ 16 and d ⫽ 6 containing M ⫽ 256 code-
considered as an m-bit symbol. words, which has twice as many codewords as the best linear
The construction of a cyclic Reed–Solomon code is as fol- code for the same values of n and d.
lows: Let 움 be a primitive element of F2m. Since 움i 僆 F2m for In 1968, Preparata generalized this construction to an in-
all i, the minimal polynomial of 움i over F2m is just x ⫹ 움i. The finite family of codes having parameters
generator polynomial of a (primitive) t error-correcting Reed–
Solomon code of length 2m ⫺ 1 has 2t consequtive powers of 움 (2m+1, 22
m+1 −2m−2
, 6), m odd, m ≥ 3
as zeros:

2t−1 A few years later, in 1972, Kerdock gave another generaliza-
g(x) = (x + α b+i ) tion of the Nordstrom–Robinson code and constructed an-
i=0 other infinite class of codes with parameters
= g0 + g1 x + · · · + g2t−1 x2t−1 + x2t
(2m+1, 22m+2, 2m − 2 (m−1)/2 ), m odd, m ≥ 3
The code has the following parameters:
The Preparata code contains twice as many codewords as
Block length: n ⫽ 2m ⫺ 1 the extended double-error-correcting BCH code and is optimal
Number of parity-check symbols: n ⫺ k ⫽ 2t in the sense of having the largest possible size for the given
Minimum distance: d ⫽ 2t ⫹ 1 length and minimum distance. The Kerdock code has twice as
many codewords as the best known linear code. In the case powers of 웁 in terms of 1, 웁, and 웁2, as follows:
m ⫽ 3 the Preparata code and the Kerdock codes both coin-
cide with the Nordstrom–Robinson code. β 3 = 2β 2 + 3β + 1
The Preparata and Kerdock codes are distance invariant.
This means that the distance distribution from a given β 4 = 3β 2 + 3β + 2
codeword to all the other codewords is independent of the β 5 = β 2 + 3β + 3
given codeword. In particular, since they contain the all-zero
β 6 = β 2 + 2β + 1
codeword, their weight distribution equals their distance dis-
tribution. β7 = 1
In general, there is no natural way to define the dual code
of a nonlinear code, and thus the MacWilliams identities have Consider the code C over Z4 with generator matrix given by
no meaning for nonlinear codes. However, one can define the
weight enumerator polynomial A(z) of a nonlinear code in the
1 1 1 1 1 1 1 1
same way as for linear codes and compute its formal dual G=
B(z) from the MacWilliams identities: 0 1 β β2 β3 β4 β5 β6
 
1 1 1 1 1 1 1 1
1 1−z 0 1
B(z) = (1 + z)n A  0 0 1 2 3 1

M 1+z = 
0 0 1 0 3 3 3 2
The polynomial B(z) obtained in this way has no simple inter- 0 0 0 1 2 3 1 1
pretation. In particular, it may have coefficients that are non-
integers or even negative. For example, if C ⫽ 兵(110), (101), where the column corresponding to 웁i is replaced by the coef-
(111)其, then A(z) ⫽ 2z2 ⫹ z3 and B(z) ⫽ (3 ⫺ 5z ⫹ z2 ⫹ z3)/3. ficients in its expression in terms of 1, 웁, and 웁2. Then the
An observation that puzzled the coding theory community Nordstrom–Robinson code is the Gray map of C .
for a long time was that the weight enumerator of the Prepar- The dual code C ⬜ of a code C over Z4 is defined similarly
ata code A(z) and the weight enumerator of the Kerdock code as for binary linear codes, except that the inner product of the
B(z) satisfied the MacWilliams identities, and in this sense vectors x ⫽ (x1, x2, . . ., xn) and y ⫽ (y1, y2, . . ., yn) with
these nonlinear codes behaved like dual linear codes. components in Z4 is defined by
Hammons, Kumar, Calderbank, Sloane, and Solé (IEEE
Trans. Information Theory 40: 301–319, 1994) gave a signifi-
n
cantly simpler description of the family of Kerdock codes. (xx, y ) = xi yi (mod 4)
They constructed a linear code over Z4 ⫽ 兵0, 1, 2, 3其, which is i=1
an analog of the binary first-order Reed–Muller code. This
code is combined with a mapping called the Gray map that The dual code C ⬜ of C is then
maps the elements of Z4 into binary pairs. The Gray map ␾ is
defined by C ⊥ = {xx ∈ Zn4 (xx, c ) = 0 for all c ∈ Zn4 }
φ(0) = 00, φ(1) = 01, φ(2) = 11, φ(3) = 10

For a linear code C over Z4, there is a MacWilliams rela-
The Lee weight of an element in Z4 is defined by tion that determines the Lee weight distribution of the dual
code C ⬜ from the Lee weight distribution of C . Therefore, one
wL (0) = 0, wL (1) = 1, wL (2) = 2, wL (3) = 1 can compute the relation between the Hamming weight dis-
tributions of the nonlinear codes C ⫽ ␾(C ) and C⬜ ⫽ ␾(C ⬜),
Extending ␾ in a natural way to a map ␾: Z4n 씮 Z2n 2 , one
and it turns out that the MacWilliams identities hold.
observes that ␾ is a distance preserving map from Z4n (under Hence, to find nonlinear binary codes related by the Mac-
the Lee metric) to Z2n
2 , (under the Hamming metric).
Williams identities, one can start with a pair of Z4-linear dual
A linear code over Z4 is a subset of Z4n such that any linear codes and apply the Gray map. For any odd integer m ⱖ 3,
combination of two codewords is again a codeword. From a the Gray map of the code K m over Z4 with generator matrix
linear code C of length n over Z4, one obtains a binary code
C ⫽ ␾(C ) of length 2n by replacing each component in a 1 1 1 1 ··· 1
codeword in C by its image under the Gray map. This code is G= m −2
0 1 β β2 ··· β2
usually nonlinear.
The minimum Hamming distance of C equals the mini-
mum Lee distance of C and is equal to the minimum Lee is the binary nonlinear (2m⫹1, 22m⫹2, 2m ⫺ 2(m⫺1)/2) Kerdock code.
weight of C since C in linear over Z4. The Gray map of K m⬜ has the same weight distribution as the
m⫹1
(2m⫹1, 22 ⫺2m⫺2, 6) Preparata code. It is, however, not identical
Example. To obtain the Nordstrom–Robinson code, we will to the Preparata code and is therefore denoted the ‘‘Prepar-
construct a code over Z4 of length 8 and then apply the Gray ata’’ code. Hence the Kerdock code and the ‘‘Preparata’’ code
map. are the Z4-analogy of the first-order Reed–Muller code and
the extended Hamming code, respectively.
Let f(x) ⫽ x3 ⫹ 2x2 ⫹ x ⫹ 3 僆 Z4[x]. Let 웁 be a zero of Hammons, Kumar, Calderbank, Sloane, and Solé also
f(x); that is, 웁3 ⫹ 2웁2 ⫹ 웁 ⫹ 3 ⫽ 0. Then we can express all showed that the binary code defined by C ⫽ ␾(C ), where C is
ALGORITHM THEORY 415
the quaternary code with parity-check matrix given by ALGORITHMS FOR BACKTRACKING. See BACK-
  TRACKING.
1 1 1 1 ··· 1 ALGORITHMS FOR RECURSION. See RECURSION.
 m −2 
H = 0 1 β β2 ··· β2  ALGORITHMS, GENETIC. See GENETIC ALGORITHMS.
3(2 m −2) ALGORITHMS, MULTICAST. See MULTICAST ALGO-
0 2 2β 3 2β 6 ··· 2β
RITHMS.
m⫹1
is a binary nonlinear (2m⫹1, 22 ⫺3m⫺2, 8) code whenever m ⱖ 3 ALGORITHMS, ONLINE. See ONLINE OPERATION.
is odd. This code has the same weight distribution as the Goe-
thals code, which is a nonlinear code that has four times as
many codewords as the comparable linear extended triple-er-
ror-correcting primitive BCH code. The code C⬜ ⫽ ␾(C ⬜) is
identical to a binary nonlinear code that was constructed in a
much more complicated way by Delsarte and Goethals more
than 20 years ago.
To analyze codes obtained from codes over Z4 in this man-
ner, one is led to study Galois rings instead of Galois fields.
Similar to a Galois field, a Galois ring can be defined as
Zpe[x]/( f(x)), where f(x) is a monic polynomial of degree m that
is irreducible modulo p. The richness in structure of the Ga-
lois rings has led to several recently discovered good nonlin-
ear codes that have an efficient and fast decoding algorithm.
Reading List
R. Blahut, The Theory and Practice of Error Control Codes. Reading,
MA: Addison-Wesley, 1983.
R. Hill, A First Course in Coding Theory. Oxford: Clarendon Press,
1986.
T. Kløve and V. I. Korzhik, Error-Detecting Codes, Boston, MA:
Kluwer Academic, 1995.
R. Lidl and H. Niederreiter, Finite Fields, vol. 20 of Encyclopedia of
Mathematics and Its Applications. Reading, MA: Addison-Wesley,
1983.
S. Lin and D. J. Costello, Jr., Error Control Coding, Fundamentals
and Applications. Englewood Cliffs, NJ: Prentice-Hall, 1983.
F. J. MacWilliams and N. J. A. Sloane, The Theory of Error-Correcting
Codes, Amsterdam: North-Holland, 1977.
W. W. Peterson and E. J. Weldon, Jr., Error-Correcting Codes. Cam-
bridge, MA: MIT Press, 1972.
M. Purser, Introduction to Error-Correcting Codes, Boston, MA: Ar-
tech House, 1995.
J. H. van Lint, Introduction to Coding Theory, New York: Springer-
Verlag, 1982.
H. van Tilborg, Error-Correcting Codes—A First Course, Lund: Stu-
dentlitteratur, 1993.
S. A. Vanstone and P. C. van Oorschot, An Introduction to Error-
Correcting Codes with Applications, Boston, MA: Kluwer Aca-
demic, 1989.
TOR HELLESETH TORLEIV KLØVE

University of Bergen
ALGEBRA, LINEAR. See LINEAR ALGEBRA.

ALGEBRA, PROCESS. See PROCESS ALGEBRA.
ALGORITHMIC DIFFERENTIATION. See AUTOMATIC
DIFFERENTIATION.
ALGORITHMS. See DIVIDE AND CONQUER METHODS.
ALGORITHMS AND DATA STRUCTURES. See DATA
STRUCTURES AND ALGORITHMS.
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Cryptography ● Email this page Enter words or
● Print this page
Yvo G. Desmedt
1University of Wisconsin—Milwaukee, Milwaukee, WI
2University of London, UK ❍ Advanced Product
Copyright © 1999 by John Wiley & Sons, Inc. All rights Search
reserved. ❍ Search All Content
DOI: 10.1002/047134608X.W4202 ❍ Acronym Finder
Article Online Posting Date: December 27, 1999
Abstract
Fundamentals
Tools
Algorithms Based on Number Theory and Algebra
Conclusion
Reading List

file:///N|/000000/0WILEY%20ENCYCLOPEDIA%20OF%20ELE...20ENGINEERING/29.%20Information%20Theory/W4202.htm17.06.2008 14:13:33
CRYPTOGRAPHY 425
CRYPTOGRAPHY
Cryptography is the science and study of the security aspects

of communications and data in the presence of a malicious
adversary. Cryptanalysis is the study of methods used to
break cryptosystems. Cryptographic schemes and protocols
are being and have been developed to protect data. Until
1974, only privacy issues were studied, and the main users
were diplomats and the military (1). Systems are also being
deployed to guarantee integrity of data, as well as different
aspects of authenticity and to identify individuals or comput-
ers (called entity authenticity). Emerging topics of study in-
clude anonymity and traceability, authorized wiretapping
(called law enforcement), copyright, digital contracts, freedom
of speech, revocation of rights, timestamping, witnessing, etc.
Related disciplines are computer security, network security,
physical security (including tempest), spread spectrum, and
steganography.
Fast computers and advances in telecommunications have
made high-speed, global, widespread computer networks pos-
sible, in particular the Internet, which is an open network. It
has increased the access to databases, such as the open World
Wide Web. To decrease communication cost and to be user-
friendly, private databases containing medical records, pro-
prietary information, tax information, etc., are often accessi-
ble via the Internet by using a low-security password scheme.
The privacy of data is obviously vulnerable during commu-
nication, and data in transit can be modified, in particular in
426 CRYPTOGRAPHY
open networks. Because of the lack of secure computers, such cryptanalyst is allowed to try to inject fraudulent messages
concerns extend to stored data. Data communicated and/or and attempt to alter the data. Therefore one calls the cryptan-
accessible over such networks include bank and other finan- alyst an active eavesdropper. To protect the data, one ap-
cial transactions, love letters, medical records, proprietary in- pends a message authentication code, abbreviated as MAC. If
formation, etc., whose privacy must be protected. The authen- there is no concern for privacy, the message itself is sent in
ticity of (the data in) contracts, databases, electronic the clear. Only the legitimate sender should be allowed to
commerce, etc. must be protected against modifications by an generate a MAC. Therefore the sender needs to know a secret
outsider or by one of the parties involved in the transaction. key k. If the key were not secret, anybody could impersonate
Modern cryptography provides the means to address these the sender. So, the authenticator generation algorithm has
issues. the message and the sender’s secret key as input. To check
the authenticity of a message, the receiver runs a verification
FUNDAMENTALS algorithm. If the algorithm’s outputs ‘‘false,’’ then the mes-
sage is definitely not authentic and must be rejected and dis-
To protect data, one needs to know what type of attacks the carded. If the output is ‘‘satisfactory,’’ very likely the message
untrusted party (enemy) can use. These depend on the secu- is authentic and is accepted. One cannot give a 100% guaran-
rity needs. The two main goals of modern cryptography are tee that the message is authentic because the active eaves-
privacy and authenticity. The issue of protecting privacy is dropper could be very lucky, but one can approach the 100%
discussed now. margin as closely as desired. If the receiver wants to verify
the authenticity of messages originating from different send-
Privacy ers, the verification algorithm must use a parameter k⬘, speci-
The threat undermining privacy is eavesdropping. The un- fying the sender, as extra input. For historical reasons this
trusted party, called the eavesdropper, will have access to the parameter has been called a key, which is discussed in more
transmitted or stored data, for example, by tapping the line or detail later.
capturing (even rather minimal) electromagnetic interference In all types of attacks the active eavesdropper is allowed
from a screen. To protect the data, called the plaintext or to see one (or more) authenticated message(s). In chosen-text
cleartext, it is transformed into ciphertext. This transforma- attacks, the cryptanalyst can choose a text which the sender
tion is called encryption. To achieve security, it should be dif- will authenticate and/or send messages with a (fictitious)
ficult for the eavesdropper to cryptanalyze, that is, to recover MAC(s). In the latter case, it is assumed that the active
the plaintext from the ciphertext. However, to guarantee use- eavesdropper can find out whether the message was accepted
fulness, the legitimate receiver should be able to recover the or rejected.
plaintext. Such an operation is called decryption and uses a
key k. To guarantee that only the legitimate receiver is able Public Key Systems
to decrypt, obviously this key must remain secret. If the
sender wants to send data to different receivers, the en- One can wonder whether k⬘ must remain secret, which is dis-
cryption algorithm must use a parameter k⬘, specifying the cussed now. If it is easy to compute k from k⬘, it is obvious
receiver, as extra input. For historical reasons this parameter that k⬘ must also remain secret. Then the key must be unique
has been called a (encryption) key, which is discussed in more to a sender–receiver pair. This introduces a key management
detail later. problem, since this key has to be transmitted in a secure way.
The person who attempts a cryptanalysis, called a cryptan- In this case, the cryptosystem is called a conventional or sym-
alyst, may in some circumstances know a previously en- metric cryptosystem and k, k⬘ usually coincide.
crypted plaintext when trying to break the current ciphertext. On the other hand, if it is hard to compute k from k⬘ and
Such an attack is called a known-plaintext attack, distin- hard to compute a k, which allows partial cryptanalysis, then
guishing it from the more basic ciphertext-only attack in the key k⬘ can be made public. This concept was invented by
which only the ciphertext is available to the cryptanalyst. Diffie and Hellman (2) and independently by Merkle (3). Such
Even more powerful attacks, especially in the commercial a system is called a public key (or sometimes an asymmetric
world, are feasible, such as a chosen-plaintext attack, in cryptosystem). This means that for privacy protection each
which the cryptanalyst chooses one (or more) plaintext(s). A receiver R publishes a personal kR, and for authentication, the
company achieves this by sending a ciphertext to a local sender S makes kS public. In the latter case the obtained au-
branch of a competing company that will most likely send the thenticator is called a digital signature because anyone who
corresponding plaintext to its headquarters and encrypt it knows the correct public key kS can verify the correctness.
with a key the first party wants to break (1). In a variant of Note that the sender can claim that the secret key was
this type of attack the cryptanalyst sends a chosen ciphertext stolen or that kS was published without consent. That would
to the receiver. The plaintext is likely to be garbled and allow a denial of ever having sent a message (4). Such situa-
thrown in the bin. If the garbage collectors collaborate with tions must be dealt with by an authorized organization. If
the cryptanalyst, the latter has started a chosen-ciphertext high security is desired, the MAC of the message must be
attack. In the strongest subtype of chosen-text attacks the deposited with a notary public. Another solution is digital
text chosen may depend on (previous or) other texts, and time stamping (5) based on cryptography (the signer needs to
therefore it is called adaptive. alert an authority that his public key must have been stolen
or lost).
Authenticity
If the public key is not authentic, the one who created the
A document is authentic if it originated from the claimed fake public key can decrypt messages intended for the legiti-
source and if its content has not been modified. So, now the mate receiver or can sign claiming to be the sender (6). So
CRYPTOGRAPHY 427
then the security is lost. In practice, this problem is solved as disciplines (mainly algebra, combinatorics, number theory,
follows. A known trusted entity(ies), for example, an author- and probability theory) and our state-of-the-art knowledge
ity, certifies that the key KS corresponds to S, and therefore of computer science (in particular, the study of (efficient)
signs (S, KS). This signature is called a certificate. algorithms, algorithmic number theory, and computational
complexity). Software engineering is used to design software
Security Levels implementations. Electrical engineering plays a role in hard-
ware implementations, and information theory is also used, in
There are different levels of security in modern cryptography,
particular, to construct unconditionally secure cryptosystems.
depending on whether information theory, physics (in particu-
Some of the main tools are explained briefly now.
lar quantum physics), computational complexity theory, or
heuristics, has been used. To be more precise, when the com-
puter power of the opponent is allowed to be unbounded and The One-Time Pad
one can mathematically prove that a formal definition of secu- The one-time pad (9), also called the Vernam scheme, was
rity is satisfied, then one is speaking about unconditional se- originally designed to achieve privacy. Shannon (10), who in-
curity. Information theory and probability theory is used to vented information theory to study cryptography, proved the
achieve this level of security. Evidently the formal definition unconditional security of the scheme when used for privacy.
of security must sufficiently model real-world security The scheme has become a cornerstone of cryptography and is
need(s). used as a principle in a wide range of seemingly unrelated
In quantum cryptography one assumes the correctness of contexts.
the laws of quantum physics (7). Shannon defined an encryption system as perfect when, for
A system or protocol is proven secure, relative to an as- a cryptanalyst not knowing the secret key, the message m is
sumption, when one can mathematically prove the following independent of the ciphertext c.
statement. The latter being, if the assumption is true, then a In the original scheme the plaintext is represented in bi-
formal security definition is satisfied for that system or proto- nary. Before encrypting the binary message, the sender and
col. Such an assumption is typically an unproven claim in receiver have obtained a secret key, a binary string chosen
computational complexity theory, such as the presumed hard- uniformly at random. When mi is the ith plaintext bit, ki the
ness of factoring large integers, or to compute discrete loga- ith key bit and ci the ith ciphertext bit, in the Vernam
rithm in finite groups. In this model the users and the oppo- scheme ci ⫽ mi 丣 ki, where 丣 is the exclusive-or, also known
nent have only a computer power bounded by a polynomial in as exor. To decrypt, the receiver computes mi ⫽ ci 丣 k⫺1 i ,
function of the length of a security parameter and one states where in the case of the exclusive-or k⫺1 ⫽ k. The key is used
that a system is secure if it requires superpolynomial (that is, only once. This implies that if the sender needs to encrypt a
growing faster to infinity than any polynomial) time to break new message, then a new key is chosen, which explains the
it. One should note that this model is limited. Indeed, when terminology: one-time pad. In modern applications, the exor
using a cryptosystem, one needs to choose a security parame- is often replaced by a group operation.
ter which fixes the length.
In practice, a system is secure if the enemy needs the com-
puter time of all computers on earth working in parallel, and Secret Sharing
the user needs, varying from application to application, 1 na- A different interpretation of the one-time pad has recently
nosecond up to a few minutes. However, modern theoretical been given (11–13). Suppose that one would like to make a
computer science cannot guarantee that a certain number of backup of a secret m with bits mi. If it is put into only one
basic operations are needed to break a cryptosystem. So, new safe, a thief who breaks open the safe will find it. So, it is put
algorithms may be developed that break cryptosystems faster in two safes so that a thief who breaks open one safe is unable
than the previously best algorithms. Moreover, new technol- to recover the secret.
ogy makes computers faster each day. The impact of new al- The solution to this problem is to choose a uniformly ran-
gorithms and new hardware is clear from the following exam- dom string of bits ki (as many as there are bits in the mes-
ple. In 1977, it was estimated that factoring a 129 digit sage). One stores the bits ki in the first safe and the bits ci ⫽
integer (product of two primes) would take 40 quadrillion mi 丣 ki in the second. Given the content of both safes, one can
(that is 4 ⫻ 1016) years, whereas it was actually factored in easily recover the secret.
1993–1994 using the idle time of approximately 1600 comput- In the previous discussion, it is assumed that two safes
ers on the Internet for 234 days (8). would not be broken into, but only one at the most. If one
A cryptosystem or protocol is as secure as another if one fears that the thief may succeed in opening more, one could
can mathematically prove that a new attack on the first proceed as follows. Choose uniformly random (t ⫺ 1) ele-
scheme implies a new attack against the other and vice versa.
ments s1, s2, . . ., st⫺1 in a finite group S(⫹) and (assuming
Finally, the weakest form of security is called heuristic. A
m 僆 S) construct st ⫽ m ⫺ (s1 ⫹ s2 ⫹ ⭈ ⭈ ⭈ ⫹ st⫺1). Put si (1 ⱕ
system is heuristically secure if no (significant) attack has
i ⱕ t) in safe i. An example of such a group is GF(2n)(⫹) where
been found. Many modern but practical cryptosystems have
n is the length of the message m. When t ⫽ 2, this corre-
such a level of security.
sponds to the one-time pad. One calls si (1 ⱕ i ⱕ t) a share of
the secret m, and the one who knows the share is called a
TOOLS shareholder or participant. Then it is easy to prove that the
eavesdropper who opens (t ⫺ 1) safes learns nothing about
Many tools are used to achieve the desired security proper- the secret. Only by opening all the safes is one able to recover
ties. These are based on discrete mathematics from several the secret m.
428 CRYPTOGRAPHY
A major disadvantage of this scheme is that it is unreli- Hash Function

able. Indeed if one share is destroyed, for example, by an
A hash function h is a function with n bits of input and m
earthquake, the secret m is lost. A t-out-of-l secret sharing
bits of output, where m ⬍ n. A cryptographic hash function
scheme is the solution. In such a scheme, one has l shares,
needs to satisfy the following properties:
but only t are required to recover the secret, whereas (t ⫺ 1)
are useless. An example of such a secret sharing scheme is
discussed later on. 1. It is a one-way function.
The concept of secret sharing was generalized, allowing 2. Given x, it is hard to find an x⬘ ⬆ x such that h(x) ⫽
one to specify in more detail who can recompute the secret h(x⬘).
and who cannot (14). Although previous secret sharing 3. It is hard to find an x and an x⬘ ⬆ x such that h(x) ⫽
schemes protect reliability and privacy, they do not protect h(x⬘).
correctness and authenticity. Indeed, a shareholder could re-
veal an incorrect share, which (very likely) implies the recon- Note that the second property does not necessarily imply the
struction of an incorrect secret. When one can demonstrate third.
the correctness of the shares, it is called verifiable secret Several modes of block ciphers allow one to make crypto-
sharing. graphic hash functions. A cryptographic hash function is an
important tool for achieving practical authentication schemes.
One-Way Functions When signing a message digitally, first one pads it, and then
one uses a cryptographic hash function before using the secret
Cryptography based on computational complexity relies on key to sign.
one-way functions. A function(s) f is one-way if it is easy to Universal hash functions are another type of hash func-
compute f, and, given an image y, it is hard to find an x such tion. These are used in unconditionally secure settings.
that y ⫽ f(x). When referring to a hash function in applied cryptography,
The state-of-the-art of computational complexity does not one means a cryptographic hash function.
allow one to prove that one-way functions exist. For some
functions f no efficient algorithm has been developed so far to
invert f, and in modern cryptography it is often assumed that Pseudonoise Generators and Stream Ciphers
such functions are one-way. A problem with the one-time pad is that the key can be used
One-way functions have many applications in modern only once. The key must be transported by a secure path. In
cryptography. For example, it has been proven that a neces- the military and diplomatic environment, this is often done
sary and sufficient condition for digital signatures is a one- by a trusted courier (using secret sharing, trust in the courier
way function(15,16). can be reduced). However, these requirements are unrealis-
tic commercially.
Block Ciphers The goal of a pseudonoise (or pseudorandom) generator is
to output a binary string whose probability distribution is
A blockcipher is a cryptosystem in which the plaintext and (computationally) indistinguishable from a uniformly random
ciphertext are divided into strings of equal length, called binary string. The pseudonoise generator starts from a seed,
blocks, and each block is encrypted one at a time with the which is a relatively short binary string chosen uniformly
same key. random.
To obtain acceptable security, a block cipher requires a When one replaces the one-time key in the Vernam scheme
good mode (17). Indeed, patterns of characters are very com- by the output of a pseudorandom generator, this is called a
mon. For example, subsequent spaces are often used in text stream cipher. Then the sender and receiver use the seed as
processors. Common sequences of characters are also not un- the secret key. It has been proven that if the pseudonoise is
usual. For example, ‘‘ from the ’’ corresponds to 10 characters, (computationally) indistinguishable from uniform, the privacy
which is 80 bits. In the Electronic Code Book (ECB) mode, the protection obtained is proven secure. This means that if an
plaintext is simply divided into blocks that are then en- unproven computational complexity hypothesis is satisfied, no
crypted. Frequency analysis of these blocks allows one to find modern computer can find information about the plaintext
such very common blocks. This method allows one to find a from the ciphertext. It has also been demonstrated that a one-
good fraction of the plaintext and often the complete plaintext way function is needed to build a pseudorandom generator.
if the plaintext that has been encrypted is sufficiently long. Moreover, given any one-way function, one can build a pseu-
Good modes have been developed based on feedback and dorandom generator. Unfortunately, the latter result is too
feedforward. theoretical to be used for building efficient pseudorandom
Many block ciphers have been designed. Some of the most generators.
popular ones are the US Data Encryption Standard (DES), Linear-feedback shift-register sequences are commonly
the Japanese NTT (Nippon Telegraph and Telephone Corpo- used in software testing. However, these are too predictable
ration), Fast Encipherment ALgorithm (FEAL), the ‘‘Interna- to be useful in cryptography and do not satisfy the previous
tional Data Encryption Algorithm’’ (IDEA) designed by Lai definition. Indeed, using linear algebra and having observed
(Switzerland), RC2, and RC5. DES (18), an ANSI (American a sufficient number of outputs, one can compute the seed and
National Standards Institute) and NIST (National Institute predict the next outputs.
of Standards and Technology, US) standard for roughly 20 Many practical pseudorandom generators have been pre-
years, is being replaced by the Advanced Encryption Stan- sented. Some of these have been based on nonlinear combina-
dard (AES), currently under development. tions of linear-feedback shift-registers others on recurrent lin-
CRYPTOGRAPHY 429
ear congruences. Many of these systems have been broken. ple, demonstrate that a public key was chosen following the
Using the output feedback (OFB) mode (17) of a block cipher specifications. A straightforward, but unacceptable solution,
one can also obtain pseudonoise generators. An example of a would be to reveal the secret key used.
pseudonoise generator based on number theory is discussed The solution to this problem is to use interaction (19). In
later on. many of these interactive protocols, the prover commits to
something. The verifier asks a question [if the question is cho-
Key Distribution sen randomly then the protocol is called an Arthur–Merlin
game (20)]. Then the prover replies and may be asked to open
Public key systems, when combined with certificates, solve the commitment. This may be repeated.
the key distribution problem. In many applications, however, To be a (interactive) proof, it is necessary that the verifier
replaying old but valid signatures should be impossible. In- will accept if the statement is true and the prover and verifier
deed, for example, one should not allow a recorded and re- follow the described protocol. This property is called complete-
played remote authenticated login to be accepted in the ness. It is also necessary that the verifier will reject the proof
future. A solution to this problem is to require a fresh session if the statement is false, even if the prover behaves differently
key, used only for a particular session. Another reason to use than specified and the dishonest prover A⬘ has infinite com-
session keys is that public key systems are slow, and so puter power. This requirement is known as soundness. In a
sender and receiver need to agree on a common secret key. variant of interactive proofs, called arguments, the last condi-
When conventional cryptography is used, the problem of tion has been relaxed.
key management is primary. Freshness remains important. An important subset of interactive proofs are the zero-
The problem is how two parties who may have never commu- knowledge ones. Then the view of a possibly dishonest verifier
nicated with each other can agree on a common secret key. can be simulated, so the verifier does not learn any informa-
Many protocols have been presented. Designing secure tion that can be used off-line. Zero-knowledge interactive
ones is very tricky. Different security levels exist. A key dis- proofs have been used toward secure identification (entity au-
tribution protocol based on number theory is discussed fur- thentication) protocols. An example of such a protocol is dis-
ther on. cussed later.
Note that several mechanisms for turning interactive zero-
Zero-Knowledge knowledge proofs into noninteractive ones have been studied
In many practical protocols one must continue using a key both from a theoretical and practical viewpoint.
without endangering its security. Zero-knowledge (19) has
been invented to prevent a secret(s) which has been used in a Cryptanalysis
protocol by party (parties) A to leak to other parties B. Cryptanalysis uses its own tools. The classical tools include
If B is untrusted, one gives the dark side of B the name statistics and discrete mathematics.
B⬘. More scientifically, machines B adhere to their specified Even if a cryptographic scheme is secure (that is, has not
protocol. To specify parties that will interact with A, but be- been broken), an inappropriate use of it may create a security
have differently, we need to speak about B⬘. breach. A mode or protocol may allow a cryptanalyst to find
When untrusted parties (or a party), let us say specified by the plaintext, impersonate the sender, etc. Such problems are
B⬘, are involved in a protocol, they see data being communi- called ‘‘protocol failures.’’ An incorrect software implementa-
cated to them and they also know the randomness they have tion often enables a hacker to make an attack, and a poor
used in this protocol. This data pulled together is called the hardware implementation may imply, for example, that the
view of B⬘. To this view corresponds a probability distribution plaintext or the key leaks due to electromagnetic radiation
(a random variable), because of the randomness used in the or interference.
protocol. When both parties A and B have x as common input, The most popular modern cryptanalytic tool against asym-
this random variable is called ViewA,B⬘(x). If x is indetermi- metric cryptosystems, based on the geometry of numbers, is
nate, we have a family of such random variables, denoted the Lenstra–Lenstra–Lovasz (LLL) lattice reduction algo-
兵ViewA,B⬘(x)其. One says that the protocol is zero-knowledge rithm (21). It has, for example, been used to break several
(does not leak anything about the secret of A) if one can simu- knapsack public key systems and many protocols (22). When
late the view of B. This means that there is a computer (poly- analyzing the security of block ciphers, the differential (23)
nomial-time machine) without access to the secret that can and linear cryptanalytic (24) methods are very important.
generate strings with a distribution that is indistinguishable Specially developed algorithms to factor and compute discrete
from 兵ViewA,B⬘(x)其. One form of indistinguishability is called log have been developed, for example, the quadratic sieve
perfect, meaning that the two distributions are identical. method (25).
There is also statistical and computational indistinguish-
ability.
So, zero-knowledge says that whatever party B⬘ learned ALGORITHMS BASED ON NUMBER THEORY AND ALGEBRA
could be simulated off-line. So party B did not receive any
information it can use after the protocol terminated. This is Although many of these algorithms are rather slow, they are
an important tool when designing proven secure protocols. becoming very popular. Attempts to break them have allowed
scientists to find better lower bounds on the size of keys for
which no algorithm exists and unlikely will be invented in the
Commitment and Interactive Proofs
near future to break these cryptosystems. However, if a true
In many cryptographic settings, a prover A needs to prove to quantum computer can be built, the security of many of these
a verifier B that something has been done correctly, for exam- schemes is in jeopardy.
430 CRYPTOGRAPHY
When writing a 僆R S, one means that a is chosen uni- ElGamal Encryption

formly random in the set S.
The ElGamal scheme (28) is a public key scheme. Let g and
We assume that the reader is familiar with basic knowl-
q be as in the Diffie–Hellman scheme. If g and q differ from
edge of number theory and algebra.
user to user, then these should be extra parts of the public
key.
RSA To make a public key, one chooses a 僆R Zq, computes y :⫽
ga in this group, and makes y public. To encrypt m 僆具g典,
RSA is a very popular public key algorithm invented by
knowing the public key yA, one chooses k 僆R Zq, computes
Rivest, Shamir, and Adleman (26).
(c1, c2) :⫽ (gk, m ⭈ yAk ) in the group, and sends c ⫽ (c1, c2). To
To generate a public key, one chooses two random and dif-
decrypt, the legitimate receiver (using the secret key a) com-
ferent primes p and q which are large enough (512 bits at
putes m⬘ :⫽ c2 ⭈ (c1a)⫺1 in this group.
least). One computes their product n :⫽ p ⭈ q. Then one
The security of this scheme is related to the Diffie–
chooses e 僆R Z*␾(n), where ␾(n) ⫽ (p ⫺ 1)(q ⫺ 1), computes Hellman problem.
d :⫽ e⫺1 mod ␾(n) and publishes (e, n) as a public key. The
number d is the secret key. The numbers p, q, and ␾(n) must
ElGamal Signatures
also remain secret or be destroyed.
To encrypt a message m 僆 Zn, one finds the authentic pub- The public and secret key are similar as in the ElGamal en-
lic key (e, n) of the receiver. The ciphertext is c :⫽ me mod n. cryption scheme. The group used is Z*p , where p is a prime.
To decrypt the ciphertext, the legitimate receiver computes Let M be the message and m the hashed and processed
m⬘ :⫽ cd mod n using the secret key d. The Euler–Fermat version of M. To sign, the sender chooses k 僆R Z*p⫺1, computes
theorem (and the Chinese Remainder theorem) guarantees r :⫽ gk mod p, computes s :⫽ (m ⫺ ar)k⫺1 mod(p ⫺ 1), and
that m⬘ ⫽ m. sends (M, r, s). To verify the signature, the receiver computes
To sign with RSA, one processes the message M, hashes it m from M and accepts the signature if gm ⫽ rs ⭈ yr mod p; oth-
with h to obtain m, computes s :⫽ md mod n, and sends (M, erwise rejects.
s), assuming that h has been agreed upon in advance. The Several variants of this scheme have been proposed, for
receiver, who knows the correct public key (e, n) of the sender, example, the US Digital Signature Standard (29).
can verify the digital signature. Given (M⬘, s⬘), one computes
m⬘ from M⬘, using the same preprocessing and hash function Pseudonoise Generator
as in the signing operation, and accepts the digital signature Several pseudorandom generators have been presented, but
if m⬘ ⫽ (s⬘)e mod n. If this fails, the receiver rejects the we discuss only one. In the Blum–Blum–Shub (30) generator,
message. a large enough integer n ⫽ pq is public, where p and q have
Many popular implementations use e ⫽ 3, which is not rec- secretly been chosen. One starts from a seed s 僆 Z*n and sets
ommended at all for encryption. Other special choices for e x :⫽ s, and the first output bit b0 of the pseudorandom genera-
are popular, but extreme care with such choices is called for. tor is the parity bit of s. To compute the next output bit, com-
Indeed many signature and encryption schemes have suffered pute x :⫽ x2 mod n and output the parity bit. More bits can
severe protocol failures. be produced in a similar manner.
More efficient pseudorandom generators have been pre-
Diffie–Hellman Key Distribution sented (31).
Let 具g典 be a finite cyclic group of large enough order generated Shamir’s Secret Sharing Scheme
by g. We assume that q, a multiple of the order of the ord(g)
(not necessarily a prime), is public. Let t be the threshold, m be the secret, and l the number
The first party, let us say A, chooses a 僆R Zq, computes of shareholders.
x :⫽ ga in this group, and sends x to the party with which it In this scheme (12), one chooses a1, a2, . . ., at⫺1 僆R GF(q),
wants to exchange a key, say B. Then B chooses a 僆R Zq, and lets f(0) ⫽ a0 ⫽ m, where f(x) ⫽ a0 ⫹ a1 ⭈ x ⫹ a2 ⭈ x2 ⫹
computes y :⫽ gb in this group, and sends y to A. Now both ⭈ ⭈ ⭈ ⫹ at⫺1 ⭈ xt⫺1 is a polynomial over GF(q) and q ⱖ l ⫹ 1. The
parties can compute a common key. Indeed, A computes z1 :⫽ share si ⫽ f(xi) where xi ⬆ 0 and the xi are distinct. This corre-
ya in this group, and B computes z2 :⫽ xb in this group. Now sponds to a Reed–Solomon code in which the message con-
z2 ⫽ z1, as is easy to verify. tains the secret and (t ⫺ 1) uniformly chosen elements. Given
It is very important to observe that this scheme does not t shares it is easy to compute f(0), the secret, using Lagrange
provide authenticity. A solution to this very important prob- interpolation. One can easily prove that given (t ⫺ 1) (or less
lem has been described in Ref. 27. shares), one has perfect secrecy, that is, any (t ⫺ 1) shares
are independent of the secret m.
The cryptanalyst needs to compute z ⫽ glogg(x)ⴱlogg(y) in 具g典.
This is believed to be difficult and is called the Diffie–
Hellman search problem. GQ
An example of a group which is considered suitable is a Fiat and Shamir (32) suggested using zero-knowledge proofs
subgroup of Z*p , the Abelian group for the multiplication of to achieve identification. We discuss a variant of their scheme
elements modulo a prime p. Today it is necessary to have at invented by Guillou and Quisquater (33).
least a 1024 bit value for p, and q should have a prime factor Let n ⫽ pq, where p and q are distinct primes and v is a
of at least 160 bits. Other groups being used include elliptic positive integer. To each prover one associates a number I,
curve groups. relatively prime to n which has a vth root. The prover, usually
CRYPTOGRAPHY 431
called Alice, will prove that I has a vth root and will prove appeared in journals are scattered. Unfortunately, some pres-
that she knows a vth root s such that svI ⬅ 1 mod n. If she tigious journals have accepted several articles of poor quality.
can prove this, then a receiver will conclude that the person
in front must be Alice. One has to be careful with such a con-
clusion (34). The zero-knowledge interactive proof is as BIBLIOGRAPHY
follows.
1. D. Kahn, The Codebreakers, New York: Macmillan, 1967.
The verifier first checks whether I is relatively prime to
n. The prover chooses r 僆R Z*n , computes z :⫽ rv mod n, and 2. W. Diffie and M. E. Hellman, New directions in cryptography,
IEEE Trans. Inf. Theory, IT-22: 644–654, 1976.
sends z to the verifier. The verifier chooses q 僆R Zv and sends
it to the prover. If q 僆 Zv, the prover halts. Else, the prover 3. R. C. Merkle, Secure communications over insecure channels,
Commun. ACM, 21: 294–299, 1978.
computes y :⫽ rsq mod n and sends y to the verifier. The veri-
fier checks that y 僆 Z*n and that z ⫽ yvIq mod n. If one of these 4. J. Saltzer, On digital signatures, ACM Oper. Syst. Rev., 12 (2):
12–14, 1978.
tests fails, the protocol is halted.
This protocol must be repeated to guarantee soundness. 5. S. Haber and W. S. Stornetta, How to time-stamp a digital docu-
ment, J. Cryptol., 3 (2): 99–111, 1991.
Avoiding such repetitions is a practical concern, addressed in
Ref. 35. If the protocol did not halt prematurely, the verifier 6. G. J. Popek and C. S. Kline, Encryption and secure computer
accepts the prover’s proof. networks, ACM Comput. Surv., 11 (4): 335–356, 1979.
7. C. H. Bennett and G. Brassard, An update on quantum cryptog-
raphy, Lect. Notes Comput. Sci., 196: 475–480, 1985.
CONCLUSION 8. D. Atkins et al., The magic words are squeamish ossifrage, Lect.
Notes Comput. Sci., 917: 263–277, 1995.
More encryption schemes and many more signature schemes 9. G. S. Vernam, Cipher printing telegraph systems for secret wire
exist than we were able to survey. The tools we discussed are and radio telegraphic communications, J. Amer. Inst. Electr. Eng.,
used in a broad range of applications, such as electronic funds 45: 109–115, 1926.
transfer (36), electronic commerce, threshold cryptography 10. C. E. Shannon, Communication theory of secrecy systems, Bell
(37,38) (which allows companies to have public keys and re- Syst. Tech. J., 28: 656–715, 1949.
duce the potential of abuse by insiders), private e-mail. Cryp- 11. G. R. Blakley, Safeguarding cryptographic keys, AFIPS Conf.
tography has evolved from a marginally important area in Proc., 48: 313–317, 1979.
electrical engineering and computer science to a crucial com- 12. A. Shamir, How to share a secret, Commun. ACM, 22: 612–613,
ponent. 1979.
13. G. R. Blakley, One-time pads are key safeguarding schemes, not
cryptosystems, Proc. IEEE Symp. Security Privacy, CA, 1980,
READING LIST
pp. 108–113.
14. M. Ito, A. Saito, and T. Nishizeki, Secret sharing schemes realiz-
Several books on practical cryptography have appeared in the
ing general access structures, Proc. IEEE Global Telecommun.
last few years. The book by Menezes et al. (39) can be consid-
Conf. (GLOBECOM ’87), 1987, pp. 99–102.
ered the best technical survey on the topic of applied cryptog-
15. M. Naor and M. Yung, Universal one-way hash functions and
raphy printed so far. A more academic book, although not so
their cryptographic applications, Proc. 21st Annu. ACM Symp.
exclusive, is Stinson’s (40). Several interesting chapters, in Theory Comput. (STOC), 1989, pp. 33–43.
particular the one on cryptanalysis (22) have appeared in the
16. J. Rompel, One-way functions are necessary and sufficient for se-
book edited by Simmons (41). Several chapters on cryptogra- cure signatures, Proc. 22nd Annu. ACM Symp. Theory Comput.
phy will appear in Ref. 42. (STOC), 1990, pp. 387–394.
Unfortunately, no good book on theoretical cryptography
17. National Bureau of Standards, DES Modes of Operation, FIPS
has appeared so far. Books which have appeared in this area Publ. No. 81 (Fed. Inf. Process. Stand.), Washington, DC: US De-
are only readable by experts in the area, or their authors have partment of Commerce, 1980.
only written about their own contributions. 18. National Bureau of Standards, Data Encryption Standard, FIPS
Although outdated, the tutorial by Brassard (43) balances Publ. No. 46 (Fed. Inf. Process. Stand.), Washington, DC: US De-
theory and practical aspects and is still worth reading. The partment of Commerce, 1977.
book by Kahn (1) overviews historical cryptosystems. Al- 19. S. Goldwasser, S. Micali, and C. Rackoff, The knowledge complex-
though the new edition discusses modern cryptography, there ity of interactive proof systems, SIAM J. Comput., 18 (1): 186–
are too few pages on the topic to justify buying the new edi- 208, 1989.
tion if one has the old one. Other more general books are 20. L. Babai, Trading group theory for randomness, Proc. 17th Annu.
available (44,45). ACM Symp. Theory Comput. (STOC), 1985, pp. 421–429.
The main conferences on the topic of cryptography are Eu- 21. A. K. Lenstra, Jr., H. W. Lenstra, and L. Lovasz, Factoring poly-
rocrypt and Crypto. Nowadays, there are many specialized or nomials with rational coefficients, Math. Ann., 261: 515–534,
more local conferences, such as Asiacrypt (which absorbed 1982.
Auscrypt), the Workshop on Fast Software Encryption, the 22. E. F. Brickell and A. M. Odlyzko, Cryptanalysis: A survey of re-
Workshop on Cryptographic Protocols, the ACM Conference on cent results, in G. J. Simmons (ed.), Contemporary Cryptology,
Computer and Communications Security, and the IMA Confer- New York: IEEE Press, 1992, pp. 501–540.
ence on Cryptography and Coding in Britain. The proceedings 23. E. Biham and A. Shamir, Differential cryptanalysis of DES-like
of many of these conferences are published by Springer Ver- cryptosystems, J. Cryptol., 4 (1): 3–72, 1991.
lag. Many conferences have sessions on cryptography, such 24. M. Matsui, Linear cryptanalysis method for DES cipher, Lect.
as IEEE-ISIT, IEEE-FOCS, ACM-STOC. Articles that have Notes Comput. Sci., 765: 386–397, 1994.
432 CULTURAL IMPACTS OF TECHNOLOGY
25. C. Pomerance, The quadratic sieve factoring algorithm, Lect.

Notes Comput. Sci., 209: 169–182, 1985.
26. R. L. Rivest, A. Shamir, and L. Adleman, A method for obtaining
digital signatures and public key cryptosystems, Commun. ACM,
21: 294–299, 1978.
27. P. C. van Oorschot, W. Diffie, and M. J. Wiener, Authentication
and authenticated key exchanges, Des., Codes Cryptogr., 2: 107–
125, 1992.
28. T. ElGamal, A public key cryptosystem and a signature scheme
based on discrete logarithms, IEEE Trans. Inf. Theory, 31: 469–
472, 1985.
29. National Institute of Standards and Technology, Digital Signa-
ture Standard, FIPS Publ. No. 186 (Fed. Inf. Process. Stand.),
Springfield, VA: U.S. Department of Commerce, 1994.
30. L. Blum, M. Blum, and M. Shub, A simple unpredictable pseudo-
random number generator, SIAM J. Comput., 15 (2): 364–383,
1986.
31. A. W. Schrift and A. Shamir, The discrete log is very discreet,
Proc. 22nd Annu. ACM Symp. Theory Comput. (STOC), 1990,
pp. 405–415.
32. A. Fiat and A. Shamir, How to prove yourself: Practical solutions
to identification and signature problems, Lect. Notes Comput. Sci.,
263: 186–194, 1987.
33. L. C. Guillou and J.-J. Quisquater, A practical zero-knowledge
protocol fitted to security microprocessor minimizing both trans-
mission and memory, Lect. Notes Comput. Sci., 330: 123–128,
1988.
34. S. Bengio et al., Secure implementations of identification sys-
tems, J. Cryptol., 4 (3): 175–183, 1991.
35. M. V. D. Burmester, An almost-constant round interactive zero-
knowledge proof, Inf. Process. Lett., 42 (2): 81–87, 1992.
36. S. Brands, Electronic money, in M. Atallah (ed.), Handbook of
Algorithms and Theory of Computation, Boca Raton, FL: CRC
Press, in press, 1998.
37. Y. G. Desmedt, Threshold cryptography, Eur. Trans. Telecom-
mun., 5 (4): 449–457, 1994.
38. Y. Desmedt, Some recent research aspects of threshold cryptogra-
phy, Lect. Notes Comput. Sci., 1396: 158–173, 1997.
39. A. Menezes, P. van Oorschot, and S. Vanstone, Applied Cryptog-
raphy, Boca Raton, FL: CRC Press, 1996.
40. D. R. Stinson, Cryptography: Theory and Practice, Boca Raton,
FL: CRC Press, 1995.
41. G. J. Simmons (ed.), Contemporary Cryptology, New York: IEEE
Press, 1992.
42. M. Atallah (ed.), Handbook of Algorithms and Theory of Computa-
tion, Boca Raton, FL: CRC Press, in press, 1998.
43. G. Brassard, Modern Cryptology, Lect. Notes Comput. Sci.,
Springer-Verlag, New York, 1988, p. 325.
44. B. Schneier, Applied Cryptography. Protocols, Algorithms, and
Source Code in C, 2nd ed. New York: Wiley, 1996.
45. C. P. Schnorr, Efficient signature generation for smart cards, J.
Cryptol., 4 (3): 239–252, 1991.
YVO G. DESMEDT
University of
Wisconsin—Milwaukee
University of London
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Data Compression Codes, Lossy ● Email this page Enter words or
● Print this page
Ken Chu1
1Apple Computers Inc., Cupertino, CA
reserved. Search
Abstract
Lossy Versus Lossless
Why Lossy?
Periodic Sampling
Aliasing
Quantization
Vector Quantization
Transform Coding
Discrete Cosine Transform
Subband Coding
Predictive Coding
Rate Distortion Theory

Acknowledgments


686 DATA COMPRESSION CODES, LOSSY
DATA COMPRESSION CODES, LOSSY
In this article we introduce lossy data compression. We con-

sider the overall process of converting from analog data to
digital so that the data are processed in digital form. Our goal
is to achieve the most compression while retaining the high-
est possible fidelity. First we consider the requirements of sig-
nal sampling and quantization. Then we introduce several ef-
fective and popular lossy data compression techniques. At the
end of this article we describe the theoretical limits of lossy
data compression performance.
Lossy compression is a process of transforming data into a
more compact form in order to reconstruct a close approxima-
tion to the original data. Let us start with a description using
a classical information coding system model. A common and
general data compression system is illustrated in Fig. 1.
As shown in Fig. 1, the information source data, S, is first
transformed by the compression process to compressed signal,
which usually is a more compact representation of the source
data. The compact form of data offers tremendous advantages
in both communication and storage applications. For exam-
ple, in communication applications, the compressed signal is
transmitted to a receiver through a communication channel
with lower communication bandwidth. In storage applica-
tions, the compressed signal takes up less space. The stored
data can be retrieved whenever they are needed. After re-
ceived (or retrieved) signal is received (retrieved), it is pro-
Compressed
S Compression signal
Source process
data
Transmission
or storage
^
S Decompression
process Received
Reconstructed
data (or retrieved)
signal
Figure 1. General data compression system.
DATA COMPRESSION CODES, LOSSY 687
cessed by the decompression process, which reconstructs the

xa(t) Sampling xs[n] Quantization xq[n] Coding 10011
.
.
.
original data with the greatest possible fidelity. In lossy com-
stage stage stage
pression systems, the original signal, S, cannot be perfectly
retrieved from the reconstructed signal, Ŝ, which is only a
close approximation. Analog Discrete-time Discrete-time Binary
data continuous-valued discrete-valued digital data
LOSSY VERSUS LOSSLESS Figure 2. Analog-to-digital converter.
In some applications, such as in compressing computer binary

executables, database records, and spreadsheet or word pro- a moderated implementation complexity. Because of this,
cessor files, the loss of even a single bit of data can be cata- the conferencing speech signals can be transmitted to the
strophic. For such applications, we use lossless data compres- destination through a lower bandwidth network at a rea-
sion techniques so that an exact duplicate of the input data sonable cost. For music centric entertainment applications
is generated after the compress/decompress cycle. In other that require near CD-quality audio, the amount of informa-
words, the reconstructed signal, Ŝ, is identical to the original tion loss that can be tolerated is significantly lower. However,
signal, S, it is still not necessary to restrict compression to lossless tech-
niques. The European MUSICAM and ISO MPEG digital
S = Ŝ audio standards both incorporate lossy compression yet pro-
duce high-fidelity audio. Similarly a perfect reconstruction of
Lossless data compression is also known as noiseless data
the original sequence is not necessary for most of the visual
compression. Naturally, it is always desirable to recreate per-
applications as long as the distortion does not result in
fectly the original signal after the transmission or storage
annoying artifacts.
process. Unfortunately, this requirement is difficult, costly,
Most signals in our environment, such as speech, audio,
and sometimes infeasible for some applications. For example,
video, radio, and sonar emissions, are analog signals. We
for audio or visual applications, the original source data are
have just discussed how lossy compression techniques are es-
analog data. The digital audio or video data we deal with are
pecially useful for compressing digital representations of ana-
already an approximation of the original analog signal. After
log data. Now let us discuss how to effectively convert an ana-
the compress/decompress cycle, there is no way to reconstruct
log signal to digital data.
an exact duplicate of the original continuous analog signal.
Theoretically converting an analog signal to the desired
The best we can do is to minimize the loss of fidelity during
digital form is a three-stage process, as illustrated in Fig. 2.
the compress/decompress process. In reality we do not need
In the first stage, the analog data (continuous-time and con-
the requirement of S⫽Ŝ for audio and video compression
tinuous-valued) are converted to discrete-time and continu-
other than for some medical or military applications. The In-
ous-valued data by taking samples of the continuous-time sig-
ternational Standards Organization (ISO) has published the
nal at regular instants, t⫽nT1,
JPEG (Joint Photographic Experts Group) standard for still
image compression (1) and the MPEG (Moving Pictures Ex-
xs [n] = xa (nT1 ) for n = 0, ±1, ±2, . . .
pert Group) standard for moving picture audio and video com-
pression (2, 3). Both JPEG and MPEG standards concern
where T1 is the sampling interval. In the quantization stage,
lossy compression, even though JPEG also has a lossless
the discrete-time continuous-valued signals are further con-
mode. The International Telecommunication Union (ITU) has
verted to discrete-time discrete-valued signals by represent-
published the H-series video compression standards, such as
ing the value of each sample with one of a finite set of possible
H.261 (4) and H.263 (5), and the G-series speech compression
values. The difference between the unquantized sample xs[n]
standards, such as G.723 (6) and G.728 (7). Both the H-series
and the quantizer output xq[n] is called the quantization error.
and G-series standards are also for lossy compression.
In reality quantization is a form of lossy data compression.
Finally, in the coding stage, the quantized value, xq[n], is
WHY LOSSY? coded to a binary sequence, which is transmitted through the
communication channel to the receiver. From a compression
Lossy compression techniques involve some loss of source in- point of view, we need an analog-to-digital conversion system
formation, so data cannot be reconstructed in the original that generates the shortest possible binary sequence while
form after they are compressed by lossy compression tech- still maintaining required fidelity. Let us discuss the signal
niques. However, we can generally get a much higher com- sampling stage first.
pression ratio and possibly a lower implementation com-
plexity.
For many applications, a better compression ratio and a PERIODIC SAMPLING
lower implementation complexity are more desirable than
the ability to reconstruct perfectly the original data. For The typical method of converting a continuous-time signal to
example, in audio-conferencing applications, it is not neces- its discrete-time representation is through periodic sampling,
sary to reconstruct perfectly the original speech samples at with a sequence of samples,xs[n], obtained from the continu-
the receiving end. In general, telephone quality speech is ous-time signal xa(t) according to the following relationship
expected at the receiver. By accepting a lower speech qual-
ity, we can achieve a much higher compression ratio with xs [n] = xa (nT1 ) for all integers n
xa(t) xa(t) where T1 is the period of s(t). The properties of impulse func-
xs[4] xs[2] tions imply that the idealized sampled waveform is easily ex-
pressed as
xs (t) = xa (t)s(t)
∞
–2T –T 0 T 2T 3T 4T t –2T 0 2T 3T 4T t = xa (t) δ(t − nT1 )
n=−∞ (1)
(a) (b)
∞
= xa (nT1 )δ(t − nT1 )
Figure 3. Continuous-time signal xa(t) sampled to discrete-time sig-
n=−∞
nals at the sampling period of (a) T, and (b) 2T.
To summarize, the idealized sampled data signal is defined

where n is an integer, T1 is the sampling period, and its recip- as a product of the original signal and a samping function and
rocal n1⫽1/T1 is the sampling frequency, in samples per sec- is composed of a series of equally spaced impulses weighted
ond. To visualize this process, consider embedding the sam- by the values of the original continuous-time signal at the
ples in an idealized impulse train to form an idealized sampling instants, as depicted in Fig. 4.
앝
continuous time sampled waveform xs(t) ⫽ 兺n⫽⫺앝xs[n]웃(t⫺nT1), Now let us make a Fourier analysis of xs(t). The Fourier
where each impulse or Dirac 웃 function can be thought of as transform pair (8) is defined as
an infinitesimally narrow pulse of unit area at time t ⫽ nT1
which is depicted as an arrow with height 1 corresponding to +∞
the area of the impulse. Then xs(t) can be drawn as a sequence x(t) = X ( f )e j2π f t d f (2)
−∞
of arrows of height xs[n] at time t ⫽ nT1, as shown with the
+∞
original signal xa(t) in Fig. 3 for sampling periods of T and 2T.
X(f) = x(t)e− j2π f t dt (3)
The sampling process usually is not an invertible process. −∞
In other words, given a discrete-time sequence, xs[n], it is not
always possible to reconstruct the original continuous-time
where X( f) is the Fourier transform of x(t), or symbolically,
input of the sampler, xa(t). It is very clear that the sampling
X( f) ⫽ T (x(t)), and x(t) is the inverse Fourier transform of
process is not a one-to-one mapping function. There are many
X( f), x(t) ⫽ T ⫺1(X( f)). A standard result of generalized Fourier
continuous-time signals that may produce the same discrete-
analysis is that
time sequence output unless they have same bandwidth and
sampled at Nyquist rate.
1 +∞
s(t) = e j2nπ f 1 t (4)
T1 n=−∞
ALIASING
In order to get better understanding of the periodic sampler, After substitution of Eq. (4) into Eq. (1), the sampled data,
let us look at it from frequency domain. First, consider the ␹s(t, yield
idealized sampling function, a periodic unit impulse train sig-
nal, s(t):
xs (t) = xa (t)s(t)

+∞
1 ∞
(5)
s(t) = δ(t − nT1 ) = xa (t)e j2nπ f 1 t
n=−∞ T1 n=−∞
+•
s(t)= δ (t – nT1) Unit impulse train
–•
Output of periodic sampler

xs(t) = xa(t)s(t)
+•
. . ., –2T1, –T1, 0, T1, 2T1, . . . t = xa(t) δ (t – nT1)
+• –•
= xa(nT1) δ (t – nT1)
–•
Continuous-time
xa(t) input signal xa(t)
Figure 4. Periodic sampled continuous-time signal xa(t). . . ., –2T1, –T1, 0, T1, 2T1, . . . t . . ., –2T1, –T1, 0, T1, 2T1, . . . t
| xa(f )| | xa(f )|
– fh fh f – fh f
f1 – fh fh f1
| xs(f )|
Figure 6. Spectrum of the sampled data sequence xs(t) for the case
of f h ⬎ f 1 ⫺ f h.
– fh fh f1 – fh, f1, f1 + fh 2f1 – fh, 2f1, 2f1 + fh 3f1 – fh, 3 f1,3 f1 + fh f Nyquist Sampling Theorem. If xa(t) is a bandlimited continu-
ous-time signal with X( f) ⫽ 0 for 兩f兩 ⬎ f h, then xa(t) can be
Figure 5. Spectrum of the sampled data sequence xs(t). uniquely reconstructed from the periodically sampled se-
quence xa(nT), ⫺앝 ⬍ n ⬍ 앝, if 1/T ⬎ 2f h.
On the other hand, if the signal is not bandlimited, theo-
retically there is no avoiding the aliasing problem. All real-
Now, taking the Fourier transform of xs(t) in Eq. (5), the re- life continuous-time signals, such as audio, speech, or video
sult is emissions, are approximately bandlimited. A common prac-
tice is to get a close approximation of the original signals by

1
+∞ +∞ filtering the continuous-time input signal with a low-pass fil-
Xs ( f ) = xa (t)e j2nπ f 1 t e − j2π f t dt ter before the sampling stage. This low-pass filter ensures
−∞ T1 n=−∞
that the filtered continuous-time signal meets the bandlim-
+∞ +∞
1 ited criterion. With this presampling filter and a proper sam-
= xa (t)e− j2π ( f −n f 1 )t dt (6)
T1 n=−∞ −∞ pling rate, we can ensure that the spectral components of in-
terest are within the bounds for which the signal can be
1 +∞
recovered, as illustrated in Fig. 7.
= Xa ( f − n f 1 )
T1 n=−∞
We see from Eq. (6) that the spectrum of a sampled-data sig- QUANTIZATION
nal consists of the periodically repeated copies of the original
signal spectrum. Each copy is shifted by integer multiples of In the quantization stage discrete-time continuous-valued sig-
the sampling frequency. The magnitudes are multiplied by nals are converted to discrete-time discrete-valued signals. In
1T1. the quantization process, amplitudes of the samples are quan-
Let us assume that the original continuous-time signal tized by dividing the entire amplitude range into a finite set
xa(t) is bandlimited to 0 ⱕ兩f兩 ⱕ f h, then the spectrum of the of amplitude ranges. Each amplitude range has a representa-
sampled data sequence xs[n]takes the form illustrated in Fig. tive amplitude value. The representative amplitude value for
5. In the case where f h ⬎ f 1 ⫺ f h, or f 1 ⬍ 2f h, there is an the range is assigned to all samples falling into the given
overlap between two adjacent copies of the spectrum as illus- range. Quantization is the most important step to removing
trated in Fig. 6. Now the overlapped portion of the spectrum the irrelevant information during lossy compression process.
is different from the original spectrum, and therefore it be- Therefore the performance of the quantizer plays a major role
comes impossible to recover the original spectrum. As a result of overall performance of a lossy compression system.
the reconstructed output is distorted from the original contin- There are many different types of quantizers. The simplest
uous-time input signal. This type of the distortion is usually and most popular one is the uniform quantizer, in which the
referred to as aliasing. quantization levels and ranges are distributed uniformly. In
general, a signal with amplitude x is specified by index k if x
To avoid aliasing a bandlimited continuous-time input, it
falls into the interval
is necessary to sample the input at the sampling frequency
f 1 ⱖ 2f h. This is stated in the famous Nyquist sampling theo-
rem (10). Ik : {x : xk ≤ x < xk+1 }, k = 1, 2, 3, . . ., L (7)
xa(t) Low-pass Sampling xs[n] Quantization xq[n] Coding 10011

.
.
.
filter stage stage stage
Analog Discrete-time Discrete-time Binary Figure 7. Sampling a continuous-time signal

data continuous-valued discrete-valued digital data that is not bandlimited.
lk lk a lower-quality output, and the bandwidth requirement is

lower accordingly. This quantizer which changes adaptively
is called an adaptive quantizer.
VECTOR QUANTIZATION
x x
We have just introduced different ways of quantizing the out-
put of a source. In all cases we discussed, the quantizer inputs
were scalar values. In other words, the quantizer takes a sin-
gle output sample of the source at a time and converts it to a
(a) (b)
quantized output. This type of quantizer is called a scalar
Figure 8. Examples of (a) a nonuniform quantizer, (b) an 8-level uni- quantizer.
form quantizer. Consider a case where we want to encode a consecutive
sequence of samples from a stationary source. It is well-
known from Shannon information theory that encoding a
In this process, the continuous valued signal with amplitude block of samples is more efficient than encoding each individ-
x is mapped into an L-ary index k. In most cases the L-ary ual sample separately. In other words, during the quantiza-
index, k, is coded into binary numbers at the coding stage and tion stage we wish to generate a representative index for a
transmitted to the receiver. Often, at the coding stage, effi- block of samples instead of for each separate sample. The ba-
cient entropy coding is incorporated to generate variable sic concept is to generalize the idea from quantizing one sam-
length codewords in order to reach the entropy rate of quan- ple at a time to quantizing a set of samples at a time. The set
tized signals. Figure 8(a) and 8(b) gives examples of a nonuni- of the samples is called a vector, and this type of quantization
form quantizer and an 8-level (L ⫽ 8) uniform quantizer. process is called vector quantization.
At the receiver, the index k is translated into an ampli- Vector quantization is one of the most popular lossy data
tude Ik that represents all the amplitudes of signals fall into compression techniques. It is widely used in image, audio,
the interval Ik, namely and speech compression applications. The most popular vec-
tor quantization is fixed-length vector quantization. In the
x̂k = lk if x ∈ Ik (8)
quantization process, consecutive input samples are grouped
into fixed-length vectors first. As an example, we can group L
wherê xk is the output of the decoder. The amplitude lk is samples of input speech as one L-dimensional vector, which
called the representation level, and the amplitude xk is called forms the input vector to the vector quantizer. For a typical
the decision level. The difference between the input signal and vector quantizer, both the encoder and the decoder share a
the decoded signal,̂ xk ⫺ x, is called the quantization error, or common codebook, C ⫽ 兵ci; i ⫽ 1, . . ., N其, which can be prede-
quantization noise. Figure 9 gives an example of a quantized fined, fixed, or changed adaptively. Each entry of the code-
waveform and the corresponding quantization noise. book, ci, is called a code-vector, which is carefully selected as
Quantization steps and ranges can be changed adaptively one of N representatives of the input vectors. Each code vec-
during the compression process. As an example, for video con- tor, ci, is also assigned an index, i. During the quantization
ferencing application, the compressed audio and video bit stage the input vector, x, is compared against each code-vec-
streams are transmitted through a network to the destina- tor, ci, in the codebook. The ‘‘closest’’ code-vector, ck, is then
tion. Under the condition that the network is out of band- selected as the representative code-vector for the input vector,
width, one cannot possibly transmit all the compressed data and the corresponding index, k, is transmitted to the receiver.
to the decoder in a timely manner. One easy solution is to In other words, ck is selected as the representative code-vector
increase the quantization step, such that quantizer generates if
d(xx, c k ) ≤ d(xx, c i ) for all c i ∈ C (9)

Amplitude
where x ⫽ (x1, x2, . . . , xL) is the L-ary input vector and C ⫽
兵ci; i ⫽ 1, . . . , N其 is the shared codebook, with ith code-
vector, ci. The idea of vector quantization is identical to that
Input signal
of scalar quantization, except the distortion is measured on
Quantizer output an L-dimensional vector basis. In Fig. 10 we show an example
of a two-dimensional vector space quantized by a vector quan-
tizer with L ⫽ 2, and N ⫽ 16. The code-vector ck represents
the input vector if it falls into the shaded vector space where
Eq. (9) is satisfied. Since the receiver shares the same code-
Time
book with the encoder, and with received index, k, the decoder
can easily retrieve the same representative code-vector, ck.
How do we measure the closeness, d(x, y), or distortion,
between two L-ary vectors, x and y, during the vector quanti-
Quantization noise = Input signal zation process? The answer is dependent on the application.
- Quantizer output
A distortion measure usually quantifies how well a vector
Figure 9. Quantization and quantization noise. quantizer can perform. It is also critical to the implementa-
where x is the original data block and T⫺1 is the inverse trans-
form of T. In the transform domain we refer to the compo-
ck nents of y as the transform coefficients. Suppose that the
transform T has the characteristic that most of the transform
coefficients are very small. Then the insignificant transform
coefficients need not to be transmitted to decoder and can be
eliminated during the quantization stage. As a result very
good compression can be achieved with the transform coding
approach. Figure 11 shows a typical lossy transform coding
data compression system.
In Fig. 11 the input data block, x, passes through the for-
ward transform, T, with transform coefficients, y, as its out-
put. T has the characteristics that most of its output, y, are
Figure 10. Two-dimensional vector space quantized by a vector small and insignificant and that there is little statistical cor-
quantizer. relation among the transform coefficients, which usually re-
sults in efficient compression by simple algorithms. The
transform coefficients, y, are quantized by the quantizer, Q.
tion of the vector quantizer, since measuring the distortion Small and insignificant coefficients have a zero quantized
between two L-dimensional vectors is one of the most compu- value; therefore only few nonzero coefficients need to be coded
tationally intensive parts of the vector quantization algo- and transmitted to the decoder. For the best compression ra-
rithm. There are several ways of measuring the distortion. tio, efficient entropy coding can be applied to the quantized
The most widely used distortion measure is the mean square coefficients at the coding stage. After receiving the signal
error (MSE), which is defined as from the network, the decoder decodes and inverse quantizes
the received signal and reconstructs the transform coeffi-
1 L
cients, ŷ. The reconstructed transform coefficients passes
d(xx, y ) = (x − yi )2
L i=1 i through the inverse transform, T⫺1, which generates the re-
constructed signal, x̂.
Another popular distortion measure is the mean absolute dif- In general, transform coding takes advantage of the linear
ference (MAD), or mean absolute error (MAE), and it is defined dependency of adjacent input samples. The linear transform
as actually converts the input samples to the transform domain
for efficient quantization. In the quantization stage the trans-
1 L
form coefficients can be quantized with a scalar quantizer or
d(xx, y ) = |x − yi |
L i=1 i a vector quantizer. However, bit allocation among transform
coefficients is crucial to the performance of the transform cod-
There are various ways of generating the vector quantization ing. A proper bit allocation at the quantization stage can
codebook. Each method generates the codebook with different achieve the output with a good fidelity as well as a good com-
characteristics. The LBG algorithm (11) or the generalized pression ratio.
Lloyd algorithm, computes a codebook with minimum average There are quite a few transform coding techniques. Each
distortion for a given training set and a given codebook size. has its characteristics and applications. The discrete Fourier
Tree-structured VQ (vector quantitization) imposes a tree transform (DFT) is popular and is commonly used for spectral
structure on the codebook such that the search time is re- analysis and filtering (18). Fast implementation of the DFT,
duced (12,13,14). Entropy-constrained vector quantization also known as fast Fourier transform (FFT), reduces the
(ECVQ) minimizes the distortion for a given average transform operation to n(n log2 n) for an n-point transform
codeword length rather than a given codebook size (15). Fi- (19). The Karhunen–Loeve transform (KLT) is an optimal
nite-state vector quantization (FSVQ) can be modeled as a fi- transform in the sense that its coefficients contain a larger
nite-state machine where each state represents a separate VQ fraction of the total energy compared to any other transform
codebook (16). Mean/residual VQ (M/RVQ) predicts the origi- (20). There is no fast implementation of the KLT, however,
nal image based on a limited data set, and then forms a resid-
ual by taking the difference between the prediction and the
original image (17). Then the data used for prediction are Transmitted
x T y Q signal
coded with a scalar quantizer, and the residual is coded with Forward Encoder
a vector quantizer. Quantizer
transform
TRANSFORM CODING Transmission

or
We just considered the vector quantization, which effectively storage
quantizes a block of data called a vector. Suppose that we have
a reversible orthogonal transform, T, that transforms a
block of data to a transform domain with the transform pair as x^ T y^ Q–1
Forward Inverse Decoder
transform quantizer Received
y = T (xx ) signal
x = T −1 (yy ) Figure 11. Basic transform coding system block diagram.
and its basis functions are target dependent. Because of this

Bandpass x1(t) x1[n] Encoder y1[t]
the KLT is not widely used. The Walsh–Hadamard transform Q
filter 1 1
(WHT) offers a modest decorrelating capability, but it has a
M
very simple implementation (21). It is quite popular, espe-
cially for hardware implementation. Bandpass x2(t) x2[n] Encoder y1[t]
Transform coding plays a very important role in the recent Q U
Input filter 2 2 y[n]
lossy compression history. In the next section we will intro-
duce the discrete cosine transform (DCT), which is the most x(f )
X
popular transform for transform coding techniques.
Bandpass xm(t) xm[n] Encoder ym[n]
Q
filter M M
DISCRETE COSINE TRANSFORM
Figure 12. Block diagram of a typical subband coder.
The most important transform for transform coding is the dis-
crete cosine transform (DCT) (22). The one-dimensional DCT
F of a signal f is defined as follows (23,24): tion and bit allocation are applied to the transform coeffi-
r2
N−1
(2 j + 1)kπ
cients in the transform domain. One of the drawbacks of
transform coding is that it has high computational complex-
F(k) = c(k) f ( j) cos ,
N j=0
2N ity. Now we introduce another compression technique—
subband coding, which usually has lower complexity than
k − 0, 1, 2, 3, . . ., N − 1
transform coding.
Just like transform coding, subband coding uses a fre-
where c(0) ⫽ 1/ 兹2 and c(k) ⫽ 1 for k ⬆ 0. The inverse DCT
quency domain approach. The block diagram of a typical sub-
(IDCT) is given by
r2
N−1
(2n + 1)kπ

band encoder is illustrated in Fig. 12. The input signal, x(t),
is first filtered by a bank of M bandpass filters. Each band-
f (n) = c(k)F(k) cos , pass filter produces a signal, xk(t), with limited ranges of spa-
N k=0
2N tial frequencies. Each filtered signal is followed by a quan-
n = 0, 1, 2, 3, . . ., N − 1 tizer and a bandpass encoder, which encodes the signal, xk(t),
with different encoding techniques according to the properties
A two-dimensional DCT for an image is formed by first taking of the subband. It may be encoded with different bit rates,
the one-dimensional DCT of all rows of an image, and then quantization steps, entropy codings, or error distributions.
taking the one-dimension DCT of all columns of the re- The coding techniques we introduced in the previous sections,
sulting image. such as the vector quantization and entropy coding, are often
The DCT has fast implementations with a computational used at the encoder. Finally the multiplexer combines all the
complexity of O(n log n) for an n-point transform. It has subband coder output, yk[n], together and sends it through the
higher compression efficiency, since it avoids the generation communication channel to the decoder.
of spurious spectral components. The DCT is the most widely A subband decoder has the inverse stages of its encoder,
used transform in transform coding for many reasons. It has as shown in Fig. 13. When a signal, ŷ[n], is received from the
superior energy compaction characteristics for most corre- communication channel, it goes through demultiplexing, de-
lated source (25), especially for Markov sources with high cor- coding, and bandpass filtering prior to subband addition.
relation coefficient ␳, Subband coding has many advantages over other compres-
sion techniques. By controlling the bit allocations, quantiza-
E[xn xn+1] tion levels, and entropy coding separately for each subband,
ρ=
E[x2n ] we can fully control the quality of the reconstructed signal.
For this reason we can fully utilize the bandwidth of the com-
where E denotes expectation. Since many sources can be mod- munication channel. With an appropriate subband coding
eled as Markov sources with a high correlation coefficient technique, we can achieve a good reconstruction signal qual-
value, the superior energy compaction capability has made ity, along with good compression. To take an example, for
the DCT the most popular transform coding technique in the
field of data compression. The DCT also tends to reduce the
statistical correlation among coefficients. These properties
y^ 1[n] ^
Decoder x1[n]
Bandpass
make DCT-based lossy compression schemes very efficient. In D filter 1+
1
addition the DCT can be implemented with reasonably low Interpolator 1
complexity. Because of this the DCT transform coding tech- E
nique is widely used for both image and audio compression y^ 2[n] ^ Bandpass
^
y[n] Decoder x2[n] Output
applications. The JPEG (1) and MPEG (2,3) published by ISO, M filter 2 +
2
and H.261 (4) and H.263 (5) published by ITU, are based on Interpolator 2 ^
x(t)
DCT transform coding compression techniques. U
^ ^ Bandpass
X ym[n] Decoder xm[n]
SUBBAND CODING filter M +
M
Interpolator M
In the last section we introduced transform coding, which con-
verts the input samples to the transform domain. Quantiza- Figure 13. Subband decoder.
High-pass ley (27), Croisier, Easteban, and Galand (28), Johnson (29),
band and Smith and Barnwell (30).
High-pass The idea of QMF is to allow the aliasing caused by overlap-
filter
High-low ping filters in the encoder (analysis filter) canceled exactly by
High-pass
filter band the filter banks in the decoder (synthesis filter). The filters
Input are designed such that the overall amplitude and phase dis-
Low-pass
xs[n] filter tortion is minimized. Then overall subband coding system
Low-pass Low-high with QMF filter bank is almost aliasing-free.
filter band
High-pass
filter
Low-low PREDICTIVE CODING
band
Low-pass In this section we introduce another interesting compression
filter technique—predictive coding. In the predictive coding sys-
tems, we assume a strong correlation between adjacent input
Figure 14. Four-band filter bank for uniform subband coding.
data, which can be scalar, vector, or even block samples.
There are many types of predictive coding systems. The most
audio and speech applications low-frequency components are popular one is the linear predictive coding system based on
usually critical to the reconstructed sound quality. The sub- the following linear relationship:
band coding technique enables the encoder to allocate more
bits to lower subbands, and to quantize them with finer quan-
k−1
x̂[k] = αi x[i] (10)
tization steps. As a result the reconstructed data retains
i=0
higher fidelity and higher signal-to-noise ratio (SNR).
A critical part of subband coding implementation is the fil-
where the x[i] are the input data, the 움i are the prediction
ter bank. Each filter in the filter bank isolates certain fre-
coefficients, and x̂[k] is the predicted value of x[k]. The differ-
quency components from the original signal. Traditionally the
most popular bandpass filter used in subband coding con- ence between the predicted value and the actual value, e[k],
sisted of cascades of low-pass filters (LPFs) and high-pass fil- is called the prediction error:
ters (HPFs). A four-band filter bank for uniform subband cod-
ing is shown in Fig. 14. The filtering is usually accomplished e[k] = x[k] − x̂[k] (11)
digitally, so the original input is the sampled signal. The cir-
cled arrows denote down sampled by 2, since only half the It is found that the prediction error usually has a much lower
samples from each filter are needed. The total number of sam- variance than the original signal, and is significantly less cor-
ples remains the same. An alternative to a uniform subband related. It has a stable histogram that can be approximated
decomposition is to decompose only the low-pass outputs, as by a Laplacian distribution (31). With linear predictive cod-
in Fig. 15. Here the subbands are not uniform in size. A de- ing, one can achieve a much higher SNR at a given bit rate.
composition of this type is an example of a critically sampled Equivalently, with linear predictive coding, one can reduce
pyramid decomposition or wavelet decomposition (26). Two- the bit rate for a given SNR. There are three basic compo-
dimensional wavelet codes are becoming increasingly popular nents in the predictive coding encoder. They are predictor,
for image coding applications and include some of the best quantizer, and coder, as illustrated in Fig. 16.
performing candidates for JPEG-2000. As shown in Fig. 16, the predicted signal, x̂[k], is sub-
Ideally the filter bank in the encoder would consist of a tracted from the input data, x[k]. The result of the subtraction
low-pass and a high-pass filter set with nonoverlapping, but is the prediction error, e[k], according to Eq. (11). The predic-
contiguous, unit gain frequency responses. In reality the ideal tion error is quantized, coded, and sent through communica-
filter is not realizable. Therefore, in order to convert the full tion channel to the decoder. In the mean time the predicted
spectrum, it is necessary to use filters with overlapping fre- signal is added back to quantized prediction error, eq[k], to
quency response. As described earlier, the overlapping fre- create reconstructed signal, x̃. Notice that the predictor
quency response will cause aliasing. The problem is resolved makes the prediction according to Eq. (10), with previously
by using exact reconstruction filters such as the quadrature reconstructed signal, x̃’s.
mirror filters (QMF), as was suggested by Princey and Brad-
xs[n] Output to
HPF x[k] e[k] eq[k] channel
Quantizer Coder
HPF + –
^
x[k] ~ +
LPF HPF x
Predictor
+
LPF
LPF
Figure 16. Block diagram of a predictive coder.
Figure 15. Filter bank for nonuniform subband coding.

+ Reconstructed signal The rate of the quantizer R(q) has two useful definitions. If a
Decoder fixed number of bits is sent to describe each quantizer level,
+
then
Received signal
from channel
Predictor R(q) = log2 M
Figure 17. Predictive coding decoder. where M is the number of possible quantizer outputs. On the
other hand, if we are allowed to use a varying number of bits,
Just like the encoder, the predictive coding decoder has a then Shannon’s lossless coding theorem says that
predictor, as shown in Fig. 17, which also operates in the
same way as the one in the encoder. After receiving the pre- R(q) = H(q(x))
diction error from the encoder, the decoder decodes the re-
ceived signal first. Then the predicted signal is added back to The entropy of the discrete quantizer output is the number of
create the reconstructed signal. Even though linear prediction bits required on the average to recover q(x). Variable length
coding is the most popular predictive coding system, there are codes can provide a better trade-off of rate and distribution,
many variations. If the predictor coefficients remain fixed, since more bits can be used on more complicated data and
then it is called global prediction. If the prediction coefficients fewer bits on low-complexity data such as silence or back-
change on each frame basis, then it is called local prediction. ground. Whichever definition is used, we can define the opti-
If they change adaptively, then it is called adaptive prediction. mal performance at a given bit rate by
The main criterion of a good linear predictive coding is to
have a set of prediction coefficients that minimize the mean- (r) = min D(q)
q : R(q)≤r
square prediction error.
Linear predictive coding is widely used in both audio and
By the operational distortion-rate function, or by the dual
video compression applications. The most popular linear pre-
function,
dictive codings are the differential pulse code modulation
(DPCM) and the adaptive differential pulse code modulation R(d) = min R(q)
(ADPCM). q : D(q)≤d
RATE DISTORTION THEORY That is, a quantizer is optimal if it minimizes the distortion
for a given rate, and vice versa. In a similar fashion we could
In the previous sections we have briefly introduced several define the optimal performance ⌬k(r) or Rk(d) using vector
lossy data compression techniques. Each of them has some quantizers of dimension k as providing the optimal rate-dis-
advantages for a specific environment. In order to achieve the tortion trade-off. Last we could ask for the optimal perfor-
best performance, one often combines several techniques. For mance, say ⌬앝(r) or R앝(d), when one is allowed to use quantiz-
example, in the MPEG-2 video compression, the encoder in- ers of arbitrary length and complexity:
cludes a predictive coder (motion estimation), a transform
∞ (r) = mink (r)
coder (DCT), an adaptive quantizer, and an entropy coder k
(run-length and Huffman coding). In this section we consider R∞ (d) = min Rk (d)
how well a lossy data compression can perform. In other k
words, we explore the theoretical performance trade-offs be-

tween fidelity and bit rate. where the ⌬k and Rk are normalized to distortion per sample
The limitation for lossless data compression is straightfor- (pixel) and bits per sample (pixel). Why study such optimiza-
ward. By definition, the reconstructed data for lossless data tions? Because they give an unbeatable performance bound to
compression must be identical to the original sequence. all real codes and they provide a benchmark for comparison.
Therefore lossless data compression algorithms need to pre- If a real code is within 0.25 dB of ⌬앝(r), it may not be worth
serve all the information in the source data. From the lossless any effort to further improve the code.
source coding theorem of Shannon information theory, we Unfortunately, ⌬앝 and R앝 are not computable from these
know that the bit rate can be made arbitrarily close to the definitions, the required optimization is too complicated.
entropy rate of the source that generated the data. So the Shannon rate-distortion theory (32) shows that in some cases
entropy rate, defined as the entropy per source symbol, is the ⌬앝 and R앝 can be found. Shannon defined the (Shannon) rate-
lower bound of size of the compressed data. distortion function by replacing actual quantizers by random
For lossy compression, distortion is allowed. Suppose that mappings. For example, a first-order rate-distortion function
a single output X of a source is described by a probability is defined by
density source function f x(x) and that X is quantized by a
quantizer q into an approximate reproduction x̂ ⫽ q(x). Sup- R(d) = min I(X , Y )
pose also that we have a measure of distortion d(x, x̂) ⱖ 0
such as a square error 兩x ⫺ x̂兩2 that measures how bad x̂ is as where the minimum is over all conditional probability density
a reproduction of x. Then the quality of the quantizer q can functions f Y兩X(y兩x) such that
be quantized by the average distortion
Ed(X , Y ) = fY |X (y|x) f X (x)d(x, y) dxdy
D(q) = Ed(x, q(x)) = f x (x)d(x, q(x)) dx
≤d
The dual function, the Shannon distortion-rate function D(r) 3. B. B. Haskell, A. Puri, and A. N. Netravali, Digital Video: An
is defined by minimizing the average distortion subject to a Introduction to MPEG-2, London: Chapman & Hall, 1997.
constraint on the mutual information. Shannon showed that 4. Recommendation H.261: Video Codec for Audiovisual Services at
for a memoryless source that p ⫻ 64 kbits/s. ITU-T (CCITT), March 1993.
5. Draft Recommendation H.263: Video Coding for Low Bitrate Com-
R∞ (d) = R(d) munication, ITU-T (CCITT), December 1995.
6. Draft Recommendation G.723: Dual Rate Speech Coder for Multi-
That is, R(d) provides an unbeatable performance bound over media Communication Transmitting at 5.3 and 6.3 Kbits/s, ITU-
all possible code, and the bound can be approximately T (CCITT), October 1995.
achieved using vector quantization of sufficiently large di- 7. Draft Recommendation G.728: Coding of Speech at 16 Kbit/s Us-
mension. ing Low-Delay Code Excited Linear Prediction (LD-CELP), ITU-T
For example, if the source is a memoryless Gaussian (CCITT), September 1992.
source with zero mean and variance ␴2, then 8. R. N. Bracewell, The Fourier Transform and Its Applications, 2nd
ed., New York: McGraw-Hill, 1978, pp.6–21.
1 σ2 9. R. N. Bracewell, The Fourier Transform and Its Applications, 2nd
R(d) = log , 0 ≤ d ≤ σ2
2 d ed., New York: McGraw-Hill, 1978, pp. 204–215.
10. H. Nyquest, Certain topics in telegraph transmission theory,
or equivalently, Trans. AIEE, 47: 617–644, 1928.
11. Y. Linde, A. Buzo, and R. M. Gray, An algorithm for vector quan-
D(r) = σ 2 e−2R
tizer design, IEEE Trans. Commun., 28: 84–95, 1980.
which provides an optimal trade-off with which real systems 12. R. M. Gray, Vector quantization, IEEE Acoust. Speech Signal Pro-
cess., 1 (2): 4–29, 1984.
can be compared. Shannon and others extended this approach
to sources with memory and a variety of coding structures. 13. J. Makhoul, S. Roucos, and H. Gish, Vector quantization in
The Shannon bounds are always useful as lower bounds, speech coding, Proc. IEEE, 73: 1551–1588, 1985.
but they are often over conservative because they reflect only 14. A. Gersho and R. M. Gray, Vector Quantization and Signal Com-
in the limit of very large dimensions and hence very compli- pression, Norwell, MA: Kluwer, 1992.
cated codes. An alternative approach to the theory of lossy 15. P. A. Chou, T. Lookabaugh, and R. M. Gray, Entropy-constrained
compression fixes the dimension of the quantizers but as- vector quantization, IEEE Trans. Acoust., Speech Signal Process.,
sumes that the rate is large and hence that the distortion is 37: 31–42, 1989.
small. The theory was developed by Bennett (33) in 1948 and, 16. J. Foster, R. M. Gray, and M. O. Dunham, Finite-state vector
as with Shannon rate-distortion theory, has been widely ex- quantization for waveform coding, IEEE Trans. Inf. Theory, 31:
348–359, 1985.
tended since. It is the source of the ‘‘6 dB per bit’’ rule of
thumb for performance improvement of uniform quantizers 17. R. L. Baker and R. M. Gray, Differential vector quantization of
with bit rate, as well as of the common practice (which is achromatic imagery, Proc. Int. Picture Coding Symp., 1983, pp.
105–106.
often misused) of modeling quantization error as white noise.
For example, the Bennett approximation for the optimal 18. W. K. Pratt, Digital Image Processing, New York: Wiley-Intersci-
ence, 1978.
distortion using fixed rate scalar quantization on a Gaussian
source is (34) 19. E. O. Brigham, The Fast Fourier Transform, Englewood Cliffs,
NJ: Prentice-Hall, 1974.
1 √ 20. P. A. Wintz, Transform Picture Coding, Proc. IEEE, 60: 809–
δ(r) ∼
= 6π 3σ 2 2−2R 820, 1972.
12
21. W. K. Pratt, Digital Image Processing, New York: Wiley-Intersci-
which is strictly greater than the Shannon distortion-rate ence, 1978.
function, although the dependence of R is the same. Both the 22. W. H. Chen and W. K. Pratt, Scene adaptive coder, IEEE Trans.
Shannon and Bennett theories have been extremely useful in Commun., 32: 224–232, 1984.
the design and evaluation of lossy compression systems. 23. N. Ahmed, T. Natarajan, and K. R. Rao, Discrete cosine trans-
form, IEEE Trans. Comput., C-23: 90–93, 1974.
ACKNOWLEDGMENTS 24. N. Ahmed and K. R. Rao, Orthogonal Transforms for Digital Sig-
nal Processing, New York: Springer-Verlag, 1975.
The author wishes to thank Professor Robert M. Gray of Stan- 25. N. S. Jayant and P. Noll, Digital Coding of Waveforms, Englewood
ford University for providing valuable information and en- Cliffs, NJ: Prentice-Hall, 1984.
lightening suggestions. The author also wishes to thank Allan 26. M. Vetterli and J. Kovacevic, Wavelets and Subband Coding, Up-
Chu, Chi Chu, and Dr. James Normile for reviewing his per Saddle River, NJ: Prentice-Hall PTR, 1995.
manuscript. 27. J. Princey and A. Bradley, Analysis/synthesis filter bank design
based on time domain aliasing cancellation, IEEE Trans. Acoust.
Speech Signal Process., 3: 1153–1161, 1986.
BIBLIOGRAPHY 28. A. Croisier, D. Esteban, and C. Galand, Perfect channel splitting
by use of interpolation/decimation techniques, Proc. Int. Conf. Inf.
1. W. B. Pennebaker and J. L. Mitchell, JPEG: Still Image Data Sci. Syst., Piscataway, NJ: IEEE Press, 1976.
Compression Standard, New York: Van Nostrand Reinhold, 1993. 29. J. D. Johnson, A filter family designed for use in quadrature mir-
2. J. L. Mitchell, et al., MPEG Video Compression Standard, Lon- ror filter banks, Proc. IEEE Int. Conf. Acoust., Speech Signal Pro-
don: Chapman & Hall, 1997. cess., Piscataway, NJ: IEEE Press, 1980, pp. 291–294.
696 DATA-FLOW AND MULTITHREADED ARCHITECTURES
30. M. J. T. Smith and T. P. Barnwell III, A procedure for designing

exact reconstruction filter banks for tree structured subband cod-
ers, Proc. IEEE Int. Conf. Acoust., Speech Signal Process., Piscata-
way, NJ: IEEE Press, 1984.
31. A. Habibi, Comparison of nth-order DPCM encoder with linear
transformation and block quantization techniques, IEEE Trans.
Commun. Technol., 19: 948–956, 1971.
32. C. E. Shannon, Coding theorems for a discrete source with a fi-
delity criterion, IRE Int. Convention Rec., pt. 4, 7: 1959, 142–163.
33. A. Gersho, Asymptotically optimal block quantization, IEEE
Trans. Inf. Theory, 25: 373–380, 1979.
34. A. Gersho, Principles of Quantization, IEEE Trans. Circuits Syst.,
25: 427–436, 1978.
Reading List
1. R. M. Gray and D. L. Neuhoff, Quantization, IEEE Trans. Inf.
Theory, 1998.
2. M. Rabbani and P. W. Jones, Digital Image Compression Tech-
niques, Tutorial Texts in Optical Engineering, vol. 7, Belling-
ham, WA: SPIE Optical Eng. Press, 1991.
3. J. L. Mitchell, et al., MPEG Video Compression Standard, London:
Chapman & Hall, 1997.
4. B. G. Haskell, A. Puri, and A. N. Netravali, Digital Video: An Intro-
duction to MPEG-2, London: Chapman & Hall, 1997.
5. W. B. Pennebaker and J. L. Mitchell, JPEG: Still Image Data Com-
pression Standard, New York: Van Nostrand Reinhold, 1993.
6. T. M. Cover and J. A. Thomas, Elements of Information Theory,
New York: Wiley, 1991.
KEN CHU
Apple Computers Inc.
DATA CONVERTERS. See ANALOG-TO-DIGITAL CON-

VERSION.
DATA DETECTION. See DEMODULATORS.
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Estimation Theory ● Email this page Enter words or
● Print this page
Anatoli Juditsky1
1INRIA Rhone-Alpes, Montbonnot Saint Martin, France
reserved. Search
Abstract
Basic Concepts
Asymptotic Behavior of Estimators
Methods of Producing Estimators
Nonparametric Estimation
Model Selection

ESTIMATION THEORY 161
considered below. The practical usefulness of the concepts

used is not comprehensively discussed. One can refer to the
treatises (3) and (9) for thorough motivations of these con-
cepts from the application point of view.
What follows considers only the statistical framework, that
is, it is supposed that the noisy environment, where observa-
tions are taken, is of a stochastic (random) nature. Situations
when this assumption does not hold are addressed by mini-
max estimation methods.
Depending on how much prior information about the sys-
tem to be identified is available, one may distinguish between
two cases:
1. The system can be specified up to an unknown parame-

ter of finite dimension. Then the problem is called the
parametric estimation problem. For instance, such a
problem arises when the parameters of a linear system
of bounded dimension are to be estimated.
2. However, rather often, one has to infer relationships be-
tween input and output data of a system, when very
little prior knowledge is available. In engineering prac-
tice, this problem is known as black-box modeling. Lin-
ear system of infinite dimension and general nonlinear
systems, when the input/output relation cannot be de-
fined in terms of a fixed number of parameters, provide
examples. In estimation theory, these problems are re-
ferred to as those of nonparametric estimation.
Consider now some simple examples of mathematical state-

ments of estimation problems.
Example 1. Let X1, . . ., Xn be independent random vari-

ables (or observations) with a common unknown distribution
P on the real line. One can consider several estimates (i.e.,
functions of the observations (Xi), i ⫽ 1, . . ., n) of the mean
␪ ⫽ 兰 xdP :
X
1. The empirical mean
n
1
X̃ = Xi (1)
n i=1
2. The empirical median m ⫽ median (X1, . . ., Xn), which

is constructed as follows: Let Z1, . . ., Zn be an increas-
ing rearrangement of X1, . . ., Xn. Then m ⫽ Z[(n⫹1)/2] for
n odd and m ⫽ (Zn/2 ⫹ Zn/2⫹1)/2 for n even (here [x]
ESTIMATION THEORY stands for the integer part of x).
3. g ⫽ (max1ⱕiⱕn(Xi) ⫹ min1ⱕiⱕn(Xi))/2
It is often the case in control and communication systems that
the mathematical model describing a particular phenomenon Example 2. The (linear regression model). The variables yi,
is completely specified, except some unknown quantities. Xik, i ⫽1, . . ., n, k ⫽ 1, . . ., d are observed, where
These quantities must be estimated. Identification, adaptive
control, learning systems, and the like, provide examples. Ex- yi = θ1 Xi1 + · · · + θd Xid + ei
act answers are often difficult, expensive, or merely impossi-
ble to obtain. However, approximate answers that are likely The ei are random disturbances and ␪1, . . ., ␪d should be esti-
to be close to the exact answers may be fairly easily obtain- mated. Let us denote Xi ⫽ (Xi1, . . ., Xid)T, ␪ ⫽ (␪1, . . ., ␪d)T.
X
able. Estimation theory provides a general guide for obtaining The estimate
such answers; above all, it makes mathematically precise n
such phrases as ‘‘likely to be close,’’ ‘‘this estimator is optimal θ̂n = arg min ( yi − θ T Xi )2
(better than others),’’ and so forth. θ
i=1
Though estimation theory originated from certain practical
problems, only the mathematical aspects of the subject are of ␪ is referred to as the least squares estimate.
162 ESTIMATION THEORY
Example 3. Let f(x) be an unknown signal, observed at the ␪ˆ 0 ⬅ ␪0, for some fixed ␪0 (independent of observations). Evi-
points, Xi ⫽ i/n, i ⫽1, . . ., n with additive noise: dently, the estimator ␪ˆ * possessing the property
yi = f (Xi ) + ei , i = 1, . . ., n (2)
Eq(θ̂ ∗ (X ) − θ ) ≤ Eq(θ̂ (X ) − θ ), for any θ ∈
This problem is referred to as nonparametric regression. Sup-

pose that f is square-integrable and periodic on [0,1]. Then for any estimate ␪ˆ may be considered as optimal. The trouble
one can develop f into Fourier series is that such estimators do not exist (indeed, any ‘‘reasonable’’
f (x) =
X∞
ck φk (x)
estimator cannot stand the comparison with the ‘‘fixed’’ esti-
mator ␪ˆ 0 at ␪0). Generally, in this method of comparing the
quality of estimators, many estimators prove to be incompara-
k=0 ble. Estimators can be compared by their behavior at ‘‘worst’’
points: an estimator ␪ˆ * of ␪ is called minimax estimator rela-
where, for instance, ␾0(x) ⫽ 0, ␾2l⫺1(x) ⫽ 兹2sin(2앟lx), and tive to the quadratic loss function q( ⭈ ) if
␾2l(x) ⫽ 兹2cos(2앟lx) for l ⫽ 1, 2, . . .. Then one can compute
the empirical coefficients
ĉk =
1 X
n
yi φk (Xi ) (3)
sup Eq(θ̂ ∗ (X ) − θ ) = inf sup Eq(θ̂ (X ) − θ )
θ ∈ θ̂ θ ∈
n where the lower bound is taken over all estimators ␪ˆ of ␪.

k=1
In the Bayesian formulation of the problem the unknown
and construct an estimate f̂n of f by substituting the estimates parameter is considered to represent values of the random
of the coefficients in the Fourier sum of the length M: variables with prior distribution Q on ⌰. In this case, the best
fˆn (x) =
XM
ĉk φk (x) (4)
estimator ␪ˆ * relative to the quadratic loss is defined by the
relation
k=1 Z
Eq(θ̂ ∗ (X ) − θ ) = Eq(θ̂ ∗ (X ) − θ )Q(dθ )

Examples 1 and 2 above are simple parametric estimation Z
problems. Example 3 is a nonparametric problem. Typically, = inf Eq(θ̂ (X ) − θ )Q(dθ )
θ̂
one chooses the order M of the Fourier approximation as a
function of total number of observations n. This way, the
problem of function estimation can be seen as that of para- and the lower bound is taken over all estimators ␪ˆ .
metric estimation, though the number of parameters to be es- As a rule, it is assumed that in parametric estimation
timated is not bounded beforehand and can be large. problems the elements of the parametric family 兵P ␪: ␪ 僆 ⌰其
The basic ideas of estimation theory will now be illus- possess the density p(x, ␪). If the density is sufficiently
trated, using parametric estimation examples. Later, it shall smooth function of ␪ and the Fisher information matrix
be seen how they can be applied in the nonparametric esti-
Z
dp

dp
T dx
mation.
I(θ ) = (x, θ ) (x, θ )
dθ dθ p(x, θ )
BASIC CONCEPTS
exists. In this case, the estimation problem is said to be regu-
Note the abstract statement of the estimation problem. It is lar, and the accuracy of estimation can be bounded from be-
assumed an observation of X is a random element, whose un- low by the Cramér-Rao inequality: if ␪ 僆 R, then for any esti-
known distribution belongs to a given family of distributions mator ␪ˆ ,
P. The family can always be parametrized and written in the
2
form 兵P ␪: ␪ 僆 ⌰其. Here the form of dependence on the parame- db
1+ (θ )
ter and the set ⌰ are assumed to be known. The problem of dθ
E|θ̂ − θ|2 ≥ + b2 (θ ) (5)
estimation of an unknown parameter ␪ or of the value g(␪) of I(θ )
a function g at the point ␪ consists of constructing a function
␪ˆ (X) from the observations, which gives a sufficiently good ap- where b(␪) ⫽ E␪ˆ ⫺ ␪ is the bias of the estimate ␪ˆ . An analo-
proximation of ␪ (or of g(␪)).
gous inequality holds in the case of multidimensional parame-
A commonly accepted approach to comparing estimators,
ter ␪. Note that if the estimate ␪ is unbiased, that is, E␪ˆ ⫽ ␪,
resulting from A. Wald’s contributions, is as follows: consider
then
a quadratic loss function q(␪ˆ (X) ⫺ ␪) (or, more generally, a
nonnegative function w(␪ˆ (X), ␪)), and given two estimators
␪ˆ 1(X) and ␪ˆ 1(X), the estimator for which the expected loss E|θ̂ − θ|2 ≥ I −1 (θ )
(risk) Eq(␪ˆ i(X) ⫺ ␪), i ⫽ 1, 2 is the smallest is called the better,
with respect to the quadratic loss function q (or to w). Moreover, the latter inequality typically holds asymptotically,
Obviously, such a method of comparison is not without its even for biased estimators when I(␪) ⫽ I does not depend on
defects. For instance, the estimator that is good for one value ␪. It can be easily verified that for independent observations
of the parameter ␪ may be completely useless for other values. X1, . . ., Xn with common regular distribution P␪, if I(␪) is the
The simplest example of this kind is given by the ‘‘estimator’’ Fisher information on one observation, then the Fisher infor-
mation on the whole sample In(␪) ⫽nI(␪), and the Cramér– on. As a consequence, for a long time there have been at-
Rao inequality takes the form tempts to develop a general procedure of constructing esti-
mates which are not necessarily optimal for a given finite
2
db amount of data, but which approach optimality asymptoti-
1+ (θ )
dθ cally (when the sample size increases or the signal-to-noise
E|θ̂ − θ|2 ≥ + b2 (θ )
nI(θ ) ratio goes to zero).
For the sake of being explicit, a problem such as in Exam-
where ␪ˆ ⫽ ␪ˆ (X1, . . ., Xn). ple 2 is examined, in which ⌰ 僆 Rd. It is to be expected that,
Return to Example 1. Let Xi be normal random variables when n 씮 앝, ‘‘good’’ estimators will get infinitely close to the
with distribution density parameter being estimated. Let P␪ denote the distribution of
observations y1, X1, . . . for a fixed parameter ␪. A sequence
1 (x − θ )2 of estimators ␪ˆ n is called a consistent sequence of estimators of
√ exp −
2π 2σ 2 ␪, if ␪ˆ n 씮 ␪ in the probability P␪ for all ␪ 僆 ⌰. Note that the
estimators, proposed for Examples 1 and 2 above, are con-
If ␴2 is known, then the estimator X̄ is an unbiased estimator sistent.
of ␪, and E(X̄ ⫺ ␪)2 ⫽ ␴2 /n. On the other hand, the Fisher Note that the notion of the minimax estimator can be re-
information of the normal density I(␪) = ␴⫺2. Thus X̄ is in this fined when the asymptotic framework is concerned. An esti-
situation the best unbiased estimator of ␪. mator ␪ˆ n, for which the quantity
If, in the same example, the distribution P possesses the
Laplace density lim sup sup Eq(θ̂n − θ )
n→∞ θ ∈
1
|x − θ|

exp − is minimized is referred to as the asymptotically minimax es-
2a a
timator in ⌰, relative to the quadratic loss q. At first glance,
this approach seems to be excessively ‘‘cautious’’: if the num-
then the Fisher information on one observation I(␪) ⫽ a⫺1. In
ber n of observations n is large, a statistician can usually lo-
this case E(X̄ ⫺ ␪)2 ⫽ 2a/n. However, the median estimator
calize the value of parameter ␪ with sufficient reliability in a
m, as n grows to infinity, satisfies nE(m ⫺ ␪)2 씮 a. Therefore,
small interval around some ␪0. In such a situation, it would
one can suggest that m is an asymptotically better estimator
seem unnecessary to limit oneself to the estimators that ‘‘be-
of ␪, in this case.
have nicely’’ for values ␪ that are far away from ␪0. Thus one
The error ␪ˆ n ⫺ ␪ of the least-squares estimator ␪ˆ in Exam-
may consider locally asymptotic minimax estimators at a
ple 2, given the observations y1, X1, . . ., yn, Xn, has the covari-
given point ␪0, that is, estimators that become arbitrarily
ance matrix
close to the asymptotically minimax estimators in a small
X
n
!−1 neighborhood of ␪0. However, it is fortunate that, in all inter-
E(θ̂n − θ )(θ̂ − θ ) = σ
T 2
Xi XiT esting cases, asymptotically minimax estimators in ⌰ are also
i=1 asymptotically minimax in any nonempty open subset of ⌰.
Detailed study of asymptotic properties of statistical estima-
This estimator is the best unbiased estimator of ␪ if the tors is a subject of asymptotic theory of estimation. Refer to
disturbances ei obey normal distribution with zero mean and (7) and (10) for exact statements and thorough treatment of
variance ␴2. correspondent problems.
Note that, if the Fisher information I(␪) is infinite, the esti-
mation with the better rate than 1/n is possible. For instance, METHODS OF PRODUCING ESTIMATORS
if in Example 1 the distribution P is uniform over [␪ ⫺1/2,
␪ ⫹ 1/2], then the estimate g satisfies Let p(X, ␪) stand for the density of the observation measure
1 P ␪. The most widely used maximum-likelihood method rec-
E( g − θ )2 = ommends that the estimator ␪ˆ (X) be defined as the maximum
2(n + 1)(n + 2)
point of the random function p(X, ␪). Then ␪ˆ (X) is called the
maximum-likelihood estimator. When the parameter set ⌰ ⱕ
ASYMPTOTIC BEHAVIOR OF ESTIMATORS Rd, the maximum-likelihood estimators are to be found
among the roots of the likelihood equation
Accepting the stochastic model in estimation problems makes
d
it possible to use the power of limit theorems (the law of large ln p(X , θ ) = 0
numbers, the central limit theorem, etc.) of probability theory, dθ
in order to study the properties of the estimation methods.
if these roots are inner points of ⌰ and p(X, ␪) is continuously
However, these results holds asymptotically, that is, when
differentiable. In Example 1, X̄ in (1) is the maximum-likeli-
certain parameters of the problem tend to limiting values
hood estimator if the distribution P is Gaussian. In Example
(e.g., when the sample size increases indefinitely, the inten-
2, if the disturbances ei have Laplace density, the maximum-
sity of the noise approaches zero, etc.). On the other hand, the
X
likelihood estimator mn satisfies
solution of nonasymptotic problems, although an important
task in its own right, cannot be a subject of a sufficiently gen- n
eral mathematical theory: the correspondent solution depends mn = arg min | yi − mT Xi |
m
heavily on the specific noise distribution, sample size, and so i=1
Another approach consists to suppose that the parameter random component in the estimation error. A general ap-
␪ obeys a prior distribution Q on ⌰. Then one can take a proach to the problem is the following: one first chooses an
Bayesian estimator ␪ˆ relative to Q, although the initial formu- approximation method, that is, substitutes the function in
lation is not Bayesian. For example, if ⌰ ⫽ Rd, it is possible question by its approximation. For instance, in Example 3,
to estimate ␪ by means the approximation with a Fourier sum is chosen (it is often
R referred to as the projection approximation, since the func-
Rd
θ p(X , θ ) dθ tion f is approximated by its projection on a final-dimensional
p(X , θ ) dθ subspace, generated by M first functions in the Fourier basis).
Then one estimates the parameters involved in this approxi-
This is a Bayesian estimator relative to the uniform prior dis- mation. This way the problem of function estimation is re-
tribution. duced to that of parametric estimation, though the number of
The basic merit of maximum-likelihood and Bayesian esti- parameters to be estimated is not fixed beforehand and can
mators is that, given certain general conditions, they are con- be large. To limit the number of parameters some smoothness
sistent, asymptotically efficient, and asymptotically normally or regularity assumptions have to be stated concerning f.
distributed. The latter means that is ␪ˆ is an estimator, then Generally speaking, smoothness conditions require that the
the normalized error I(␪)1/2(␪ˆ ⫺ ␪) converges in distribution to unknown function f belongs to a restricted class, such that,
a Gaussian random variable with zero mean, and the identity given an approximation technique, any function from the
covariance matrix. class can be ‘‘well’’ approximated, using a limited number of
The advantages of the maximum-likelihood estimators jus- parameters. The choice of the approximation method is cru-
tify the amount of computation involved in the search for the cial for the quality of estimation and heavily depends on the
maximum of the likelihood function p(X, ␪). However, this can prior information available about the unknown function f [re-
be a hard task. In some situations, the least-squares method fer to Ref. (8) for a more extensive discussion]. Now see how
can be used instead. In Example 1, it recommends that the the basic ideas of estimation theory can be applied to the non-
minimum point of the function
X
parametric estimation problems.
n
(Xi − θ )2 Performance Measures for Nonparametric Estimators
i=1
The following specific issues are important:
be used as the estimator. In this case, X̄ in Eq. (1) is the least-
squares estimate. In Example 2, the least squares estimator 1. What plays the role of Cramér-Rao bound and Fisher
␪ˆ n coincides with the maximum-likelihood solution if the Information Matrix in this case? Recall that the
noises ei are normally distributed. Cramér-Rao bound [Eq. (5)] reveals the best perfor-
Often, the exact form of density p(X, ␪) of observations is mance one can expect in identifying the unknown pa-
unknown. However, the information that p(X, ␪) belongs to rameter ␪ from sample data arising from some parame-
some convex class P is available. The robust approach estima- terized distribution P␪, ␪ 僆 ⌰, where ⌰ is the domain
tion recommends to find the density p*(X, ␪), which maxi- over which the unknown parameter ␪ ranges. In the
mizes the risk of the least-squares estimate on P, and then to nonparametric (as well as in the parametric) case, lower
take bounds for the best achievable performance are pro-
vided by minimax risk functons. These lower bounds will
θ̂ ∗ (X ) = arg min p∗ (X , θ ) be introduced and associated notions of optimality will
θ be discussed.
as the estimator. The ␪ˆ * is referred to as the robust estimate. 2. For parametric estimation problems, a quadratic loss
Suppose, for instance, that in Example 1 the distribution P function is typical to work with. In functional estima-
satisfies 兰(x ⫺ ␪)P (dx) ⱕ ␴2. Then the empirical mean X̄ is tion, however, the choice is much wider. One can be in-
the robust estimate. If p(x ⫺ ␪) is the density of P , and it is terested in the behavior of the estimate at one particu-
known that p( ⭈ ) is unimodal and for some a ⬎ 0 p(0) ⱖ lar point x0, or in the global behavior of the estimate.
(2a)⫺1, then the median m is the robust estimator of ␪ [for Different distance measures should be used in these two
more details, refer to (5)]. different cases.
In order to compare different nonparametric estimators, it

NONPARAMETRIC ESTIMATION is necessary to introduce suitable figures of merit. It seems
first reasonable to build on the mean square deviation (or
Consider the problem of nonparametric estimation. To be con- mean absolute deviation) of some semi-norm of the error, it is
crete, consider Eq. (2) in Example 3 above. There are two fac- denoted by 储 f̂N ⫺ f 储. A semi-norm is a norm, except it does
tors that limit the accuracy with which the signal f can be not satisfy the condition: 储 f 储 ⫽ 0 implies f ⫽ 0. The following
recovered. First, only a finite number of observation points semi-norms are commonly used: 储 f 储 ⫽ (兰 f 2(x)dx)1/2, (L2-
n
(Xi)i⫽1 are available. This suggests that f(x), at other points x norm), 储 f 储 ⫽ supx 兩 f(x)兩 (uniform norm, C- or L앝-norm),
than those which are observed, must be obtained from the 储 f 储 ⫽ 兩 f(x0) 兩 (absolute value at a fixed point x0). Then consider
observed points by interpolation or extrapolation. Second, as the risk function
in the case of parametric estimation, at the points of observa-
tion, Xi, i ⫽ 1, . . ., n, f(Xi) is observed with an additive noise
Ra N ( fˆN , f ) = E[a−1 ˆ
N f N − f ]
2
(6)
ei ⫽ yi ⫺ f(Xi). Clearly, the observation noises ei introduce a
where aN is a normalizing positive sequence. Letting aN de- Consider Example 3. The following result can be acknowl-
crease as fast as possible so that the risk still remains edged; refer to (7): Consider the Sobolev class Ws(L) on [0, 1],
bounded yields a notion of a convergence rate. Let F be a set which is the family of periodic functions f(x), x 僆 [0, 1], such
of functions that contains the ‘‘true’’ regression function f, that
then the maximal risk raN( f̂N) of estimator f̂N on F is defined
as follows: X∞
(1 + j 2s )|c j |2 ≤ L2 (8)
ra N ( fˆN ) = sup Ra N ( fˆN , f )
j=0
f ∈F
(here cj are the Fourier coefficients of f). If
If the maximal risk is used as a figure of merit, the optimal ;R 1/2
estimator f̂*N is the one for which the maximal risk is mini- g = | g(x)|2 dx , or g = | g(x0 )|
mized, that is, such that
then n⫺s/2s⫹d is a lower rate of convergence for the class Ws(L)
ra N ( fˆN∗ ) = min sup Ra N ( fˆN , f ) in the semi-norm 储 ⭈ 储.
fˆN f ∈F On the other hand, one can construct an estimate f̂n [refer
to (2)], such that uniformly, over f 僆 Ws(L),
f̂*N is called the minimax estimator and the value
E fˆn − f 22 ≤ O(L, σ )n−2s/(2s+1) (9)
min sup Ra N ( fˆN , f )
fˆN f ∈F
Note that the condition [Eq. (8)] on f means that the function
can be ‘‘well’’ approximated by a finite Fourier sum. Indeed,
is called the minimax risk on F . Notice that this concept is due to the Parseval equality, Eq. (8) implies that if
X
consistent with the minimax concept used in the parametric
case. M
The construction of minimax nonparametric regression es- f (x) = c j φ j (x)
timators for different sets F is a difficult problem. However, j=1
letting aN decrease as fast as possible so that the minimax
risk still remains bounded yields a notion of a best achievable then 储 f ⫺ f 储22 ⫽ O(M⫺2s). The upper bound, Eq. (9), appears
convergence rate, similar to that of parametric estimation. rather naturally if one considers the following argument: If
More precisely, one may state the following definition: one approximates the coefficients cj by their empirical esti-
mates ĉj in Eq. (3), the quadratic error in each j is O(n⫺1).
1. The positive sequence aN is a lower rate of convergence Thus, if the sum, Eq. (4) of M terms of the Fourier series is
for the set F in the semi-norm 储 ⭈ 储 if used to approximate f, the ‘‘total’’ stochastic error is order of
M/n. The balance between the approximation (the bias) and
lim inf ra N ( fˆN∗ ) = lim inf inf sup E[a−1 fˆN − f ]2 ≥ C0 the stochastic error gives the best choice M ⫽ O(n1/(2s⫹1)) and
N
N→∞ N→∞ fˆN f ∈F the quadratic error O(n⫺2s/(2s⫹1)). This simple argument can be
(7) used to analyze other nonparametric estimates.
for some positive C0.

2. The positive sequence aN is called minimax rate of con- MODEL SELECTION
vergence for the set F in semi-norm 储 ⭈ 储, if it is a lower
rate of convergence, and if, in addition, there exists an So far the concern has been with estimation problems when
estimator f̂*N achieving this rate, that is, such that the model structure has been fixed. In the case of paramet-
ric estimation, this corresponds to the fixed (a priori known)
lim sup ra N ( fˆN∗ ) < ∞ model order; in functional estimation this corresponds to
N→∞ the known functional class F , which defines the exact ap-
proximation order. However, rather often, this knowledge is
The inequality [Eq. (7)] is a kind of negative statement that not accessible beforehand. This implies that one should be
says that no estimator of function f can converge to f faster able to provide methods to retrieve this information from
than aN. Thus, a coarser, but easier approach consists in as- the data, in order to make estimation algorithms ‘‘implem-
sessing the estimators by their convergence rates. In this set- entable.’’ One should distinguish between two statements of
ting, by definition, optimal estimators reach the lower bound the model (order) selection problem: the first one arises
as defined in Eq. (7) (recall that the minimax rate is not typically in the parametric setting, when one suppose that
unique: it is defined to within a constant). the exact structure of the model is known up to unknown
It holds that the larger the class of functions, the slower dimension of the parameter vector; the second one is essen-
the convergence rate. Generally, it can be shown that no tially nonparametric, when it is assumed that the true
‘‘good’’ estimator can be constructed on too rich functional model is of infinite dimension, and the order of a finite-
class which is ‘‘too rich’’ [refer to (4)]. Note, however, that con- dimensional approximation is to be chosen to minimize a
vergence can sometimes be proved without any smoothness prediction error (refer to the choice of the approximation
assumption, though the convergence can be arbitrary slow, order M in Eq. (4) of Example 3). These two approaches
depending on the unknown function f to be estimated. are illustrated in a simple example.
Example 4. Consider the following problem: where
λ(n) λ(n)
1. Let ␪ ⫽ (␪0, . . . ␪d⫺1)T be coefficients of a digital filter of lim inf >1 and →0
n log log n n
unknown order d, that is,
yi =
X
d−1
θi xi−k+1 + ei
gives a consistent estimate of the true dimension d in the
problem 1 of Example 4.
Another approach is proposed in (12) and (14). It consists
k=0
to minimize, with respect to d, the total length of the incoding
of the sequence yi, Xi (MML—minimum message length, or
We assume that xi are random variables. The problem
MDL—minimum description length). This code length should
is to retrieve ␪ from the noisy observations (yi, xi), i ⫽
also take into account the incoding of ␪ˆ d,n. This leads to the
1, . . ., n. If one denotes Xi ⫽ (xi, . . ., xi⫺d⫹1)T, then the
criterion (the first-order approximation)
estimation problem can be reformulated as that of the
linear regression in Example 2. If the exact order d was dn = arg min BIC(d, n)
known, then the least-squares estimate ␪ˆ n could be used d≤n
to recover ␪ from the data. If d is unknown, it should be

estimated from the data. where
2. A different problem arises when the true filter is of in- 2σe2 d log(n)

finite order. However, all the components of the vector BIC(d, n) = S2d,n +
n
␪ of infinite dimension cannot be estimated. In this case
one can approximate the parameter ␪ of infinite dimen- As was shown in (13), the Bayesian approach (MAP—
sion by an estimate ␪ˆ d,n which has only finite number d maximum a posteriori probability) leads to the minimization
of nonvanishing entries: of BIC(d, n), independently of the distribution of the parame-
ter d.
θ̂n = (θ̂n(1) . . ., θ̂n(d) , 0, 0 . . .)T
Then the ‘‘estimate order’’ d can be seen as a nuisance BIBLIOGRAPHY

parameter to be chosen, in order to minimize, for in-
stance, the mean prediction error E[(␪ˆ d,n ⫺ ␪)TXn]2. 1. H. Akaike, Statistical predictor identification, Ann. Inst. Math.
Statist., 22: 203–217, 1970.
Suppose that ei are independent and Gaussian random
2. N. Cencov, Statistical decision rules and optimal inference, Amer.
⫽ n⫺1兺i⫽1(yi ⫺ ␪ˆ Td,nXi)2. If d is unknown,
2 n
variables. Denote Sd,n Math. Soc. Transl., 53: 1982.
2
one cannot minimize Sd,n with respect to d directly: the result
3. H. Cramer, Mathematical Methods of Statistics, Princeton, NJ:
of such a brute-force procedure would give an estimate Princeton University Press, 1946.
␪ˆ d,n(x), which perfectly fits the noisy data (this is known as
4. L. Devroye and L. Györfi, Nonparametric Density Estimation L1
‘‘overfitting’’ in the neural network literature). The reason is View, New York: Wiley, 1985.
that Sd,n 2
is a biased estimate of E(yi ⫺ ␪ˆ Td,nXi)2. The solution
5. P. Huber, Robust Statistics, New York: Wiley, 1981.
rather consists to modify S2(d, n) to obtain an unbiased esti-
6. E. J. Hannan, Estimation of the order of an ARMA process, Ann.
mate of the prediction error. This can be achieved by intro-
Stat., 8: 339–364, 1980.
ducing a penalty which is proportional to the model order d:
7. I. A. Ibragimov and R. Z. Khas’minskii, Statistical Estimation:
2σ 2 d
Asymptotic Theory, New York: Springer, 1981.
AIC(d, n) = S2d,n + e 8. A. Juditsky et al., Nonlinear black-box modelling in system iden-
n tification: Mathematical foundations, Automatica, 31 (12): 1725–
1750, 1995.
which is an unbiased (up to the terms of the higher order) 9. M. G. Kendall and A. Stuart, The Advanced Theory of Statistics,
estimate of the error up to terms that do not depend on d, Griffin, 1979.
AIC(d, n). One can look for dn such that 10. L. LeCam, Asymptotic Methods in Statistical Decision Theory,
Springer Series in Statistics, Vol. 26, New York: Springer-Ver-
lag, 1986.
dn = arg min AIC(d, n)
d<n 11. C. Mallows, Some comments on Cp, Technometrics, 15: 661–
675, 1973.
This technique leads to the celebrated Mallows–Akaike crite- 12. J. Rissanen, Stochastic Complexity in Statistical Inquiry, Series
rion (1, 11): in Computer Science, Vol. 15, World Scientific, 1989.
Unfortunately, dn is not a consistent estimate of d. Thus it 13. G. Schwartz, Estimating the dimension of a model, Ann. Stat., 6
does not give a solution to the first problem of Example 4 (2): 461–464, 1978.
above. On the other hand, it is shown in (6) that minimization 14. C. S. Wallace and P. R. Freeman, Estimation and inference by
over d of the criterion compact coding, J. Royal Stat. Soc., Ser. B, 49 (3): 240–265,
1987.
2σe2 λ(n)d

HIC(d, n) = S2d,n + ANATOLI JUDITSKY
n INRIA Rhone-Alpes
ETHERNET 167
ESTIMATION THEORY. See CORRELATION THEORY; KAL-

MAN FILTERS.
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Image Codes ● Email this page Enter words or
● Print this page
Yung-Kai Lai1 and C.-C. Jay Kuo1
1University of Southern California, Los Angeles, CA
reserved. Search
Abstract
General Lossy Image/Video Compression Framework
Compression Standards
Other Compression Techniques

J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering
Copyright c 1999 John Wiley & Sons, Inc.
IMAGE CODES
Although usually acquired from analog optical devices, images are often sampled into the digital form, because
they are more easily stored, transmitted, and processed digitally. The major difficulty with digital image storage
and transmission, however, is the size of bits required to record the image data. For a 512 × 512 gray-scale
image with eight-bit resolution, 256 kbytes of storage space are required. With color images or digital video
data, the amount of data is enormously greater. For example, the bit rate is 8.70 Mbytes/s for a color video
sequence with 352 × 288 pixels per picture, eight bits per color channel, and 30 pictures/s. For a 30 s video clip
at such a bit rate, the total data takes up 261 Mbytes of storage space, or 21.12 h of transmission time with a
28800-baud modem. Therefore it is desirable to use data compression techniques to reduce the amount of data
for digital images and video.
There are some important differences between digital image/video compressions and other digital data
compression. First, for most other data compression applications, it is desirable to have the data themselves
unaltered. In digital image compression, on the other hand, some information loss is allowed as long as
the visual appearance of the image or video is not severely impaired. In some cases, though, lossless image
compression is required. For example, it may be preferred that medical images be losslessly compressed,
because small deviations from the original image may affect the doctor’s diagnosis. The second difference is
that natural images contain much redundant information that is very useful for compression. The background
of a natural image, for instance, contains a lot of pixels with similar luminance or color components. These
background pixels are represented more efficiently by using various image compression techniques.
Generally speaking, digital image/video compression techniques are classified into two categories: lossless
compression and lossy compression. Lossless image/video compression uses many lossless compression tech-
niques mentioned in Data compression, lossy. Lossy image/video compression is more important in image/video
compression because the compression ratio is more flexibly adjusted without having to preserve every detail
in the image/video. This section primarily focuses on this category, and so do many international standards to
be introduced later in this section.
General Lossy Image/Video Compression Framework
The most important issue in image/video compression is reducing the redundancy in the image/video. Most
of state-of-the-art lossy image/video compression techniques use transform coding for this purpose. A general
image/video compression framework using transform coding includes four major parts: color space conversion,
transform, quantization, and entropy coding, as shown in Fig. 1.
Color Coordinates and Chrominance Subsampling. Images are often displayed by the cathode
ray tube (CRT) using red (R), green (G), and blue (B) phosphor emissions. In compression, however, the
RGB color coordinate is not the most efficient for representing the color components of images or video. It is
known that the luminance (the intensity of the light, the gray-scale projection of the image) is more important
than the chrominance (colors hue and saturation) components in human visual perception. Therefore it is
1
2 IMAGE CODES
Fig. 1. Block diagram of a general image transform coder. The decoder performs the inverse steps to reconstruct the
image.
preferable to transform the color components from the RGB color coordinate to some luminance-chrominance
representation so that we put more emphasis on the luminance and discard more unimportant information
from the chrominance components without affecting much of the visual perception of compressed images/video.
Three often used luminance-chrominance color coordinate systems are YUV, YIQ, and YCbCr color spaces.
The YUV color coordinate was developed by National Television Systems Committee (NTSC) and now
used in Phase Alternation Line (PAL) and Sequentiel Couleur Avec Mémoire (SECAM) color television systems.
NTSC later developed the YIQ coordinate by rotating the U and V components in YUV space to further reduce
the color component bandwidths. The luminance Y and the color components U, V and I, Q in their respective
coordinates can be transformed via
and
The YCbCr color coordinate was developed as the standard color coordinate system for digital video, as
described in ITU-R Recommendation 601 (1). It is an offset and scaled version of the YUV coordinate to limit
the dynamic range of luminance and chrominance components within the dynamic ranges of the original RGB
components:
Because the human visual system is less sensitive to chrominance, the two chrominance components Cb and
Cr are usually subsampled to reduce the data volume before compression. Several subsampling formats are
commonly used. Two of them, the 4:2:2 format and 4:2:0 format, are used in image/video coding standards, such
IMAGE CODES 3
Fig. 2. Different subsampling formats differ in both the ratio of luminance and chrominance samples and their rela-
tive positions. Two YCbCr subsampling formats are shown in this figure: (a) 4:2:2, (b) 4:2:0 format. Luminance (Y) and
chrominance (Cb, Cr) samples are represented by circles and squares, respectively.
as JPEG (Joint Photographic Experts Group) and MPEG (Motion Picture Experts Group). These two formats
are shown in Fig. 2.
Transform. Transform coding is the most popular coding scheme in scholarly research and industry
standards. The purpose of the transformation is to map the digital image from the spatial domain to some
transform domain so that its total energy is packed in a small portion of transform coefficients, whereas most
other transform coefficients are very close to zero. We can coarsely quantize these unimportant coefficients or
simply throw them away in later steps to achieve the goal of compression.
There are several other desirable properties for the transforms used in transform coding. First, a unique
inverse transform is required because we have to recover the image from its transform domain. Second, the
transform should conserve the total energy in the image. Unitary transforms satisfy these requirements.
But not all unitary transforms are suitable for image compression. The energy compaction property and the
computational complexity of the transforms are always as important in practical implementation. The optimal
transform for energy compaction is known as the Karhunen–Loève transform (KLT), but the computational
complexity is too high to be practical. Most of the compression standards use the discrete cosine transform
(DCT). It has a good energy compaction property, and fast algorithms for its forward and inverse transforms
are available. Wavelet transform is another promising transform for transform coding and is described in a
later section.
Quantization. Most of the transforms convert integral pixel values into floating-point transform coef-
ficients. Encoding these coefficients as floating-point numbers is not economic for lossy compression. Quan-
tization is the process of converting continuous numbers to discrete-value samples. Most transform coding
techniques use scalar quantization. The principles of quantization are described in Data compression codes—
Lossy and are not discussed in detail here. The output of the scalar quantizer is the index of the reconstruction
level. Because quantization is a many-to-one mapping, this is the primary step that causes loss of information
in the whole transform coding process.
Entropy Coding. The final stage of the transform coder is to losslessly compress the quantization
indexes using an entropy coder for further reduction of compressed bit-stream volume. Two often used entropy
coders are the Huffman coder and the arithmetic coder. The details of entropy coders are described in Data
compression, lossy.
Image Error Measures. To evaluate the performance of image compression techniques, proper image
error measures, which evaluate the difference between the original and compressed images, are necessary. A
commonly used image error measure is the mean square error (MSE), defined as
where E{·} is the expectation operator, X and X represent the original and compressed images, respectively,
and i, j are the image coordinates of the pixel. The peak signal-to-noise ratio (PSNR) is more frequently used
4 IMAGE CODES
Fig. 3. Block diagram of a general video transform coder using motion-compensated predictive coding. A motion predictor
is used to find the motion vector and the estimation error is also transform coded.
in image compression:
where P is the peak input amplitude. For example, for an eight-bit gray-scale image, P = 255. A lower MSE or
higher PSNR value means that the compressed image has higher fidelity to the original image. Both MSE and
PSNR are conventionally used for gray-scale image error evaluation. There are no consensual error measures
for color image compression yet.
Motion-Compensated Predictive Coding. The temporal domain is involved in digital video compres-
sion. A digital video is a series of images, called pictures or frames, to be played sequentially. A straightforward
way of compressing digital video by image compression techniques is to treat each frame as independent im-
ages and compress them separately. However, digital video has redundancies in the temporal domain that
are exploited for further compression. Unless there are scene changes, videos usually have many of the same
objects in adjacent frames, though the spatial locations on each frame may differ because of object motion. It
is a waste to code the same objects on different frames repeatedly. We can encode the object on the first frame
and only the direction and distance of object motion in subsequent frames. At the decoder end, after the first
frame is decoded, the object on subsequent frames is reconstructed by pasting the object at different spatial
locations according to the object motion information. The objection motion direction and distance information
are called the motion vector (MV), the process to estimate the motion vector between adjacent frames is called
motion estimation (ME), and the scheme to perform ME and paste the object with respect to the motion vector
is called motion compensation (MC).
The same object appearing on adjacent frames, however, may appear differently because of light reflection,
object rotation, cameras panning or zooming, and so on. Furthermore, new objects may appear which cannot
be well estimated with other objects on the previous frame. Therefore motion compensation is only a prediction
from previous frames. The difference between the prediction and the actual pixel values has to be computed
and encoded. The prediction error, however, would be quite small as long as the ME algorithm finds the
minimal-error match from previous frames. The error histogram usually has its peak around zero with small
probabilities at large values. The prediction error can be quantized and encoded very efficiently. The block
diagram of a general video coder using motion-compensated predictive coding is shown in Fig. 3.
IMAGE CODES 5
Compression Standards
JPEG Standard. The Joint Photographic Experts Group (JPEG) is a working group formed in 1982
under the International Organization for Standardization (ISO). This group joined the International Organi-
zation for Standardization Consultative Committee (CCITT) Special Rapporteur Group (SRG) on New Forms
of Image Communication to establish an international standard for digital color image compression. After eval-
uating numerous proposals, they completed the draft technical specification in 1990, and the draft became an
international standard in 1992. Some further extensions were developed in 1994. The resulting standard (2),
also called JPEG, is now used worldwide for still, continuous-tone, monochrome, and color image compression.
The original JPEG quality requirement is to have indistinguishable images when compressed at 1.50 bits
to 2.00 bits per pixel (bpp), excellent image quality at 0.75 bpp to 1.50 bpp, good to very good quality at 0.50
bpp to 0.75 bpp, and moderate to good quality at 0.25 bpp to 0.50 bpp. There are four modes of JPEG operation.
They are the sequential baseline DCT-based mode, the progressive DCT-based mode, the sequential lossless
mode, and the hierarchical mode. These four modes provide different compression techniques for applications
with different requirements. The baseline mode, however, is the mode most often used. The three other modes
are rarely used so that many JPEG decode software programs do not even support them.
Baseline Mode. The block diagram of the sequential baseline DCT-based mode JPEG coder and decoder
is similar to that shown in Fig. 1. The color image is first converted into the YCbCr coordinates, and then
the three components are compressed separately. The core transform used is the discrete cosine transform
(DCT), which transforms spatial-domain pixel values into frequency-domain coefficients. To represent the DCT
coefficients with 11-bit precision for eight-bit input image (and 15-bit precision for 12-bit input), the three color
components in the YCbCr space are level shifted by subtracting 128 (or 2048 for 12-bit input) before performing
the DCT. For computational efficiency, the whole input image is partitioned into square blocks of 8 × 8 pixels
each. Then the two-dimensional 8 × 8 DCT is performed on each block separately:
where s and S are the 2-D spatial-domain pixel values and the 2-D DCT coefficients, respectively. The sub-
scripts yx and vu are the spatial-domain and frequency-domain coordinates, respectively. S00 is called the DC
coefficient, and the rest of the 63 coefficients are called AC coefficients. The 8 × 8 inverse discrete cosine
transform (IDCT) used at the decoder end is given by
and 128 (or 2048 for 12-bit input) is added back to restore the original pixel value levels. Numerous fast DCT
and IDCT algorithms are available but are not discussed in detail here.
After the DCT operation, the 64 DCT coefficients are quantized by using a lookup table, called the
quantization matrix. The default quantization matrices for luminance and chrominance are different because
the Human Visual System (HVS) has different luminance and color perception characteristics. These two
default quantization matrices Qvu are given in Table 1, where the lowest frequency components Q00 ’s are in
the upper left corners. The encoder is also free to define its own quantization matrices, but they have to be
6 IMAGE CODES
included in the compressed data for the decoder to reconstruct the DCT coefficients. The quantization indexes
Sqvu are obtained by dividing the floating-point DCT coefficients by the quantization matrix and rounding the
quotient to the nearest integer:
The reconstructed DCT coefficients Rvu are obtained at the decoder side by multiplying the quantization indexes
by the quantization matrix:
Because the DC coefficients represent the mean values of the partitioned 8 × 8 blocks, these coefficients
among adjacent blocks are usually quite close to each other in natural images. Thus they are encoded with
differential coding in the raster-scan order to take advantage of this property. With the DCT energy compaction
property, most of the energy of each 8 × 8 block is concentrated in low frequencies. Therefore the 63 AC
coefficients are encoded in a zigzag order so that the significant coefficients are likely to be encoded first, and
in most cases, there are consecutive zero AC coefficients near the end of the block so that they are encoded very
efficiently. The differential coding of DC coefficients and the zigzag coding order of AC coefficients is shown in
Fig. 4.
Two DC and two AC Huffman tables are used for entropy coding the DCT coefficients. The DC Huffman
table for eight-bit resolution is shown in Table 2. The differential (DIFF) values range from −2047 to 2047, are
classified into 12 categories, denoted as SSSS, and are coded by variable-length codewords. The AC coefficients
range from −1023 to 1023 and are classified into 10 nonzero SSSS categories. Because runs of zeros are likely
at the high frequencies along the zigzag scan, the lengths of zero runs are encoded with 15 four-bit categories,
denoted as RRRR. The combination of RRRRSSSS is encoded using the AC Huffman table with 162 possible
IMAGE CODES 7
Fig. 4. Differential coding of DC coefficients (left) and zigzag scan order of AC coefficients (right). The differential coding of
DC coefficients takes advantage of the cross-block correlation of DC values, whereas the zigzag scan order takes advantage
of the energy compaction property so that it is very likely to have consecutive zeros toward the end of the block.
codes, which are not listed here. A particular RRRRSSSS code is used for zero runs with lengths exceeding 15,
and another particular code is used to denote the end of block (EOB) when all remaining quantized coefficients
in the block are zero.
Progressive Mode. In the baseline mode, the blocks in an image are encoded and decoded in the raster-
scan order, that is, from left to right and from top to bottom. The decoder has to receive all the details for one
block, decode it, and then proceed to the next block. If, for some reason, the bit stream is cut midway during
8 IMAGE CODES
the transmission, there is no way the decoder-end user would know the content of the rest of the image. The
progressive DCT-based mode uses multiple scans through the image. The DC coefficients of all blocks from the
whole image are transmitted first, then the first several AC coefficients of the whole images are transmitted,
and so on. In this way, even if the bit stream is cut midway during transmission, it is possible that the whole
image with coarser resolution is already available for the decoder-end user to perceive the image content.
Progressive image transmission is particularly preferable for image browsing over transmission channels with
limited bandwidth. The user can decode the rough image to see if this image carries the required information.
If it does, the user can continue the transmission to add more and more details to the image. Otherwise, the
user can stop the transmission.
In practice, all DCT coefficients are computed as in the baseline mode and stored in a buffer. The encoder
is free to choose the scan number and the coefficients to be transmitted in each scan. For example, the encoder
may choose to send the DC coefficients S00 in the first scan, S01 and S10 in the second scan, S20 , S11 , S02 in the
third run, S03 , S12 , S21 , S30 in the fourth run, and the rest of AC coefficients in the fifth. This choice, called
spectral selection, is up to the encoder and can be specified explicitly in the scan header of the compressed bit
stream.
In addition to intercoefficient spectral selection, intracoefficient successive approximation is also used for
progressive transmission. In short, the successive approximation scheme quantizes the coefficient with lower
precision so that a shorter bit stream is transmitted. In every subsequent stage, one more truncated bit is
added back to improve the precision of the coefficients by one bit until full precision is reached.
Lossless Mode. Although primarily focused on lossy image compression, JPEG also provides a lossless
compression mode. Rather than using the float-point DCT process that introduces error with integral quanti-
zation, the lossless mode uses 2-D predictive coding. This predictive coding method uses the upper, left, and
upper left neighbor pixels to predict the present pixel value. One of the seven prediction types is chosen and
specified in the scan header. The pixels are encoded according to the predictor selected. The lossless mode
allows input precision from 2 to 16 bits/sample. The difference between the prediction value and the input is
computed modulo 216 and encoded using the Huffman table in Table 2, except that extra entries are added at
the end of the table to code the SSSS value from 0 to 16. Arithmetic coding of the modulo difference is also
allowed but not discussed in detail here.
Hierarchical Mode. The last mode, the hierarchical mode, is used to generate subimages of smaller
size and coarser resolution. First the original image is successively downsampled by a factor of 2 horizontally,
vertically, or both. The subimages are smaller versions of the original image with lower resolution. The smallest
subimage is transmitted. Then it is upsampled and interpolated bilinearly to form the prediction of the next
higher resolution image. The prediction error is encoded and transmitted. The process is repeated until the
original image size and resolution are achieved. At the decoder end, a similar process is used to reconstruct the
original image by upsampling and adding the prediction error to form multiresolution images. The encoding
method can be one of the other three modes: sequential, progressive, or lossless. The hierarchical mode is
preferable for platforms with a lower resolution display device or with limited computational power insufficient
for reconstructing full-sized images.
JPEG 2000 Standard. New image compression techniques have emerged in recent years since the
development of JPEG standard. In addition, JPEG either does not support or does not perform well for some
recent applications, such as side-channel information and very low bit-rate coding. All of these encourage the
creation of second-generation image compression standards. JPEG 2000, aiming to become an International
Standard (IS) in year 2000, is the ongoing project for this purpose (3).
The goal of JPEG 2000 is to create a unified system for compressing different types of images (bilevel,
gray-scale, or color) with various characteristics and conditions. The purpose of this standard is to complement
but not replace the current JPEG standard. Therefore it will focus mainly on the applications for which JPEG
fails to provide acceptable quality or performance. The new features of JPEG 2000 will likely include
IMAGE CODES 9
• High-performance, low bit-rate compression: JPEG performs poorly at a low bitrate, where apparent block-
ing artifacts appear. JPEG 2000 intends to improve the rate-distortion performance at low bit rates, for
example, below 0.25 bpp for gray-scale images, while keeping the excellent performance at higher bit rates.
This is the most important function of JPEG 2000.
• Various-type image compression: JPEG focuses mainly on lossy gray-scale and color image compression. The
standard for bilevel (such as text) image compression is currently the Joint Bilevel Image Experts Group
(JBIG) standard. In addition, although providing a lossless mode, JPEG does not perform particularly well
in that aspect. JPEG 2000 aims to provide efficient lossless and lossy compression of images with a wide
dynamic range, such as bilevel, gray-scale, and color images, all within a unified architecture.
• Robustness to errors: JPEG performs poorly if there is bit error in the bit stream during transmission
and storage, for example, when compressed data is transmitted through a noisy channel. JPEG 2000 will
incorporate error-resilience or error-correction capabilities into the standard so that the compression bit
stream is robust to unexpected errors.
• Content-based description and MPEG-4 interface: one of the most challenging topics in image compression
is extracting the semantic content of images and its objects. It benefits applications, such as image retrieval
and object-based compression. JPEG 2000 intends to shed light on this problem and hopefully to find some
solution for it. In addition, the new video compression method MPEG-4 will use a descriptive language to
describe the objects and provide methods to code them. JPEG 2000 is expected to provide an interface for
object description and compression for objects.
Other features, including (but not limited to) fixed bit-rate compression, image security, such as image
watermarking and encryption, side channel (such as alpha channel and transparency plane) information,
random access and processing on arbitrary regions, are also expected to be incorporated into this standard.
Though the transform, even the whole framework of JPEG 2000 is likely to be quite different from that used
in the existing JPEG, it is desirable that JPEG 2000 should be backward-compatible for JPEG.
MPEG-1 and MPEG-2. The Moving Picture Experts Group (MPEG), another working group under
ISO, was formed in 1988. It developed the original video coding standard, which was also commonly called
MPEG later, for video and associated audio compression. Most of the MPEG parts became an international
standard in 1992 (4). Different from JPEG, the second-generation project of MPEG started right after MPEG
was completed. To distinguish among generations of MPEGs, the first generation of MPEG is often called
MPEG-1. The three essential parts of MPEGs are : video, audio, and systems. We focus only on the video part
of this coding standard.
The objective of MPEG-1 is to provide approximately VHS quality of compressed video with a medium
bandwidth for 1 to 1.8 Mbps (Mbits/s). It is used for strictly noninterlaced (progressively scanned) video and is
optimized for CD-ROM, video CD, and CD-interactive (CD-i) applications. The dimension limits for parameter-
constrained video are 768 (h) × 576 (v) × 30 (frames/s, fps).
An MPEG-1 video stream includes several hierarchical layers. The whole video sequence is partitioned
into at least one group of pictures (GOP), intended to allow random access into the sequence. Each GOP consists
of a certain number of pictures. The picture, also called a frame, is the primary coding unit of a video sequence.
Each picture consists of three rectangular matrices representing luminance (Y) and two chrominance (Cb, Cr)
values. In MPEG-1, the YCbCr matrices are sampled by the 4:2:0 format, that is, the Cb and Cr matrices
are subsampled by two horizontally and vertically, therefore their sizes are one-quarter of the Y matrix. Each
picture consists of one or more slices. The major purpose of slices is to isolate the error in the transmitted bit
stream. In the case of a noisy channel, the decoder can skip the erroneous slice and start decoding with the
next slice. In an error-free environment, the encoder usually assigns the whole picture to one slice. Each slice
is composed of one or more macroblocks. The macroblock is the basic coding unit in MPEG. The size of each
macroblock is 16 × 16. In the 4:2:0 format, it consists of six 8 × 8 blocks, four of which are luminance (Y)
10 IMAGE CODES
Fig. 5. Video stream data hierarchical layers in MPEG. The four blocks in the macroblocks represent the luminance (Y)
blocks in the macroblock. The two 8 × 8 chrominance (Cb and Cr) blocks are downsampled from the 16 × 16 macroblock
and are not shown in this figure.
blocks, and the other two are downsampled chrominance (Cb and Cr) blocks. This video stream data hierarchy
is shown in Fig. 5.
MPEG-1 uses a DCT-based transform coding scheme to reduce the spatial redundancy, and the motion-
compensation technique to reduce the temporal redundancy. MPEG defines four types of pictures: intra (I),
predicted (P), bidirectional (B), and DC (D). The D pictures are used only for the fast-forward mode, in which
only the DC value of each block is encoded for fast decoding. This type of picture cannot be used in conjunction
with the other three types of pictures. It is seldom used thus is not discussed in detail here. A GOP must contain
at least one I picture, and may be followed by any number of I, P, B pictures. The I picture is intracoded, which
means that it is coded using a still image-coding technique without any temporal dependency on other pictures.
The P picture is predicted from a previous I or P picture with its motion information. The B picture is inserted
between two I or P pictures (or one of each) and is bidirectionally interpolated from both pictures. With this
dependency of I, P, and B pictures, B pictures must be encoded after I or P pictures even though they are
displayed before them. Figure 6 shows the picture dependency and the difference between video stream order
and display order. Note that this figure only serves as an example of the way the three types of pictures are
organized. The actual number of I, P, B pictures in a GOP can be specified arbitrarily by the encoder.
The I pictures are encoded using a JPEG-like image coding scheme. Each 8 × 8 block is level-shifted and
transformed by using the 8 × 8 DCT and then quantized by using a default or user-defined intraquantization
matrix. The default intraquantization matrix is shown in Table 3(a). The intraquantization matrix can be
multiplied by a quantizer scale factor from 1 to 31 from macroblock to macroblock to achieve different bit rates,
but the quantization step of DC coefficient is always set to eight. The DC coefficients are differential-coded,
whereas the AC coefficients are zigzag scanned and encoded by using run-length coding and Huffman coding,
which is similar (but not identical) to JPEG.
The motion-compensation technique is used in coding P pictures. For each 16 × 16 macroblock in the P
picture, the most similar macroblock is found in the preceding I or P picture as the prediction or estimation
IMAGE CODES 11
Fig. 6. Interpicture dependency (left) and picture order (right) in MPEG. While the encoder transmits the P frames before
the B frames in order to provide interpolation information, it is necessary for the decoder to put the first decoded P frames
in a buffer until the subsequent B frames are decoded and to rearrange them for display.
of the target macroblock. A motion vector is used to record the spatial displacement between the original and
estimated macroblocks in their respective pictures.
The MPEG-1 standard does not define the similarity between two macroblocks and the method of search-
ing for the most similar macroblock in the reference picture. The encoder is free to develop its own similarity
criterion and motion vector searching algorithm. Two distortion definitions are commonly used to provide the
similarity measure:
• Mean squared error (MSE):

12 IMAGE CODES
• Mean absolute error (MAE):
where S1 and S2 are the target and reference pictures, u1 , v1 are the upper left coordinates in S1 and S2 ,
and (k, l) is the MV.
When comparing two MSEs or MAEs, the division of 16 × 16 is a common factor and thus can be dropped.
The smaller the MSE or MAE, the more similar the two macroblocks. The MAE has lower computational
complexity and therefore is used more often in MPEG coders. The purpose of the motion vector searching
algorithm is to find the MV with the smallest MSE or MAE, and choose it as the best MV representing the
motion of the macroblock.
MPEG-1 allows the motion vector to take a large range of displacements in the picture from the reference
macroblock. The computational cost to search the whole range, however, is too high to be practical. An efficient
encoder usually limits its search to a reasonable range, say, in a 32 × 32 neighborhood region. This 32 × 32 region
is conventionally called a [−16,15] searching window because the horizontal and vertical displacements for the
MV are confined to the [−16, 15] range. The simplest searching algorithm is the full search, that is, to sweeping
this searching window pixel-by-pixel and finding the macroblock with the least error. The computational cost
is greatly reduced with the logarithmic searching algorithm. In the first step of the logarithmic searching
algorithm, eight MVs with large displacement from the starting pixel are selected. The eight errors with
respect to these eight MVs are computed and compared to find the smallest one. In the second step, the
starting point is taken as the coordinate associated with the smallest error, and the searching displacement
is halved. A similar process is repeated until the smallest displacement (one pixel) is met. The process of the
full search and the logarithmic search are shown in Fig. 7. MPEG-1 allows half-pixel MV precision if it gives
better matching results. The pixel values are bilinearly interpolated to achieve this half-pixel precision. The
searching algorithms have to be modified accordingly and the searching range is four times as large with half-
pixel precision. Generally speaking, MV search is the most computationally expensive part of the whole MPEG
coding process. Many new and efficient searching algorithms have been developed and adopted by commercial
encoders.
Depending on the magnitude of motion vector and the prediction error, various coding types can be
selected for P picture coding based on the following decisions. How to make these selection decisions is left to
the encoder and not explicitly defined in MPEG-1.
• Intra/nonintra: if the intracoding of the macroblock takes less bits than coding the motion vector and the
prediction error, we may simply use intracoding as used in I pictures, else nonintracoding of the MV’s and
prediction errors is used.
• MC/no MC: if the motion vector is very close to zero, we may avoid using MC to save the bits for encoding
MVs.
• New/old quantizer scale: if the currently used quantizer scale is not adequate for coding, for example,
unable to satisfy the current bit-rate constraint, it may be changed to a new value.
• Coded/not coded: in the nonintracoding case, if the coefficients of a block are all zero, the whole block is
not coded. In this way a significant number of bits is saved. Because a macroblock includes six blocks, if at
least one block in a macroblock has to be coded, a coded block pattern has to be included in the bit stream
to inform the decoder which blocks in the macroblock are actually encoded.
IMAGE CODES 13
Fig. 7. Two MV searching algorithms: full search (left) and logarithmic search (right), where 1, 2, and 3 indicate the step
numbers. In this illustrative example, it takes only 21 MV computations to obtain the best motion vector. On the other
hand, all the 81 possible MVs have to be computed in the full search algorithm.
These four decisions generate eight different macroblock types and each is coded differently. A noteworthy
macroblock type is the skipped block type, which is not intracoded, having all zero coefficients after quanti-
zation, and no MV and quantizer scale change is needed. In other words, the macroblock is identical to the
macroblock in the previous I or P picture at exactly the same location. No variable-length code (VLC) is needed
for this type of macroblocks
In most cases, both the MVs and the prediction errors have to be coded. The MVs of adjacent macroblocks
are likely to have similar values because adjacent macroblocks are likely to move coherently in the same
direction. Therefore the MVs are encoded with differential coding and further entropy-coded by a variable-
length Huffman code. The 16 × 16 prediction error is encoded by using the transform coding technique similar
to that used in encoding I pictures but with a nonintraquantization matrix. The default nonintraquantization
matrix is shown in Table 3(b). Another difference from intrablock coding is that the DC coefficients are encoded
together with all AC coefficients rather than using a separate differential coder.
The B pictures are obtained by bidirectionally interpolating I or P pictures. At first, one of three MC modes
(forward, backward, interpolated) is selected. If the forward mode is selected, a macroblock from the previous I
or P picture and the forward MV is used as the prediction. If the backward mode is selected, a macroblock from
the future I or P picture and the backward MV is used. If the interpolated mode is selected, one macroblock
from the previous and one from the future pictures are bilinearly interpolated to yield the prediction, and both
the forward and backward MVs are transmitted.
Similar to P picture coding, three decisions (except for the MC/no MC decision, because all macroblocks
have to be motion compensated) have to be made. There are a total of 11 macroblock types in coding B pictures.
Different from P picture coding, a skipped block in a B picture uses the same motion vector and same macroblock
type as its previous block.
The DCT, quantization, and variable-length codes for B pictures are the same as those of P pictures.
MPEG-2, the second generation of MPEG, focuses primarily on high-quality compressed video for broad-
casting. The typical bit rate is 4 Mbps to 15 Mbps, compared to the 1.5 Mbps for MPEG-1. The major applications
include digital video disk (DVD), digital video broadcasting (DVB), and TV systems, such as NTSC and PAL.
There was a plan to develop MPEG-3 for high definition television (HDTV) applications, but it was later merged
14 IMAGE CODES
with MPEG-2. Three parts (systems, video, and audio) of MPEG-2 became IS in 1994, and the other parts were
adopted from 1995 to 1997.
Instead of simply providing video compression as in MPEG-1, MPEG-2 focuses on providing more func-
tionality to various applications. MPEG-2 supports both interlaced and noninterlaced video and has a wider
range of picture sizes, called levels. MPEG-2 also includes several profiles to provide different functionality.
The combination of levels and profiles for different applications is shown in Table 4. MPEG-2 also provides four
scalable modes in Main+ and High profiles for decoders with different capabilities. The spatial scalable mode
provides two spatial resolution video layers. The SNR scalable mode provides two video layers of the same
resolution but different quality. The temporal scalable mode has one lower layer coded at the basic temporal
rate, and the second enhancement layer uses the lower layer as its prediction. The data partitioning mode uses
progressive transmission which is similar to JPEG’s progressive mode.
IMAGE CODES 15
The MPEG-2 coding scheme (5) is similar to that of MPEG-1 with some modifications and additions for
the extra functionality not handled in MPEG-1. The MPEG-2 video syntax (for the main profile) is a superset
of that of MPEG-1.
MPEG-4 and MPEG-7. MPEG-4 is an ongoing project of the MPEG family. MPEG-4 (6) focuses mainly
on very low bit-rate (64 kbps or less) applications. The major applications are mobile or other telecommunication
video applications (5 kbps to 64 kbps) and TV/film applications (up to 2 Mbps). It is expected to be finalized by
November 1998.
The three major features of MPEG-4 are content-based interactivity, compression, and universal access.
MPEG-4 plans to provide the functionality of these features to bridge among the TV/film audiovisual data,
wireless telecommunication, and computer interactivity.
Topics involved in content-based interactivity include the following:
• Content-based multimedia data access tools: the objects in the coding sequence are segmented and called
audio-visual objects (AVO). The video is separated into video object planes (VOP). Each VOP may have
different spatial and temporal resolutions, may have sub-VOPs, and can be associated with different
degrees of accessibility, and may be either separated or overlapping.
• Content-based manipulation and bit-stream editing: manipulation and/or editing are allowed to be per-
formed on a VOP, such as spatial position change, spatial scaling, moving speed change, VOP insertion and
deletion, etc.
• Synthetic and natural hybrid coding (SNHC): MPEG-4 is intended to compress both natural and synthetic
(cartoon, graphics, etc.) video. Issues include the processing of synthetic data in geometry and texture,
real-time synchronization and control, integration of mixed media types, and temporal modeling.
• Improved temporal random access.
In compression, several important issues are addressed. The coding efficiency is improved to reduce the
bit rate under 64 kbps for mobile applications and 2 Mbps for high-quality TV/film applications. In addition
to the objective error measures, such as PSNR, the subjective quality should also be higher compared with
existing standards. The ability to encode multiple concurrent data streams, such as the multiple views of a
stereo video scene, is also provided. The most important breakthrough in compression, however, should be the
object-based coding support. The encoder should be able to encode VOPs with arbitrary shapes, transmit the
shape and transparency information of each VOP, support I, P, and B frames of VOPs, and encode the input VOP
sequences at fixed and variable frame rates. The coding scheme of MPEG-4 is still block-based DCT coding.
To support content-based coding, a square bounding box for an object is first found from the segmentation
result. Motion compensation is performed on the macroblocks inside the bounding box. If the macroblock is
inside the object, conventional block matching, as in MPEG-1, is used. If the macroblock is complete outside
the object, no matching is performed. If the macroblock is partly outside and partly inside the object (that is,
the macroblock is on the boundary of the object), some reference pixels have to be padded onto the nonobject
area in the macroblock for block matching. Another method is to approximate the boundary macroblocks with
polygons and perform polygon matching rather than square block matching.
Two issues are stressed in the universal access feature. One is the robustness in error-prone environments.
Because one of MPEG-4’s major applications is telecommunication, the channel error over wired and wireless
networks has to be considered. The bit stream should be robust for severe error conditions, such as long bursty
errors. The other issue is content-based scalability. Scalability in content (spatial, temporal, etc.), quality, and
complexity should be allowed.
Because MPEG-4 is intended to provide compression for various types of applications with different
bit rates, quality, source material, algorithms, and so on, a toolbox approach is adopted for the purpose of
integration. In the toolbox approach, a proper profile is chosen to satisfy the requirements of the application.
The coder selects a compression algorithm according to the profile and picks suitable tools to realize this
16 IMAGE CODES
algorithm. The MPEG-4 system description language (MSDL) allows the transmission in the bit stream of
the object structure, rules for the decoder, and the tools not available at the decoder. MPEG-4 also has a close
relationship with the virtual reality modeling language (VRML) to transmit the description of a 2-D or 3-D
scene.
MPEG-7 (7), the newest standard, is currently under development. It will focus mainly on multimedia
database management, such as image query, indexing, and retrieval. It is expected to be closely tied with
MPEG-4 content-based coding techniques to serve database management purposes.
H.263 Standard. In the 1980s, the International Telephone and Telegraph Consultative Committee
(CCITT) started its research on low bit-rate videophone and videoconferencing, intended to be transmitted
by communication channels with very low bandwidth, such as the telephone lines. The resulting ITU-T H-
series recommendations include two H.32X recommendations and their subsystems. The first H.32X system
Recommendation H.320, “Narrow-band visual telephone systems and terminal equipment,” was finalized in
1990. It targets the bit rate p × 64 kbps (p = 1–30). The more recent Recommendation H.324, “Multimedia
terminal for low bit-rate visual telephone services over the PSTN,” was finalized in 1995. It focuses on bit rates
below 64 kbps. The comparison of these two Recommendations is shown in Table 5. We focus here only on the
video coding standards H.263 (8).
The acceptable picture formats for H.263 are listed in Table 6. The 4:2:0 YCbCr format is used to represent
color frames. Similar to the MPEG family, hierarchical layer representation is also used in H.263. Four layers
are used in H.263: the picture layer, the group of blocks (GOB) layer, the macroblock layer, and the block layer.
Each picture is divided into GOBs. The height of a GOB is defined as 16 pixels for SQCIF, QCIF, and CIF formats,
32 pixels for the 4CIF format, and 64 pixels for the 16CIF format. Each GOB is divided into macroblocks of
size 16 × 16. Each macroblock includes four 8 × 8 luminance (Y) blocks and two 8 × 8 chrominance (Cb and
Cr) blocks. The building blocks of the encoder include motion-compensation, transform-coding, quantization,
and variable-length codes. There are two prediction modes. The intermode uses the information in a previous
frame for MC. The intramode uses only information present in the picture itself. Similar to the picture types in
the MPEG family, there are also I, P, and B pictures in the H.263 standard. However, because there is no GOP
layer in H.263, the sequence is allowed to extend to an arbitrary number of frames without recurring patterns.
However, because the prediction error propagates with each P picture coding, H.263 requires the insertion of at
least one I picture in every 132 pictures. H.263 supports half-pixel precision MVs with the searching window
of [−16,15.5].
IMAGE CODES 17
Except for the basic mode described above, there are four negotiable modes in H.263, which make the
major differences between H.263 and H.261 (and the MPEG family): the syntax-based arithmetic coding mode,
the unrestricted motion-vector mode, the advanced-prediction mode, and the PB-frame mode.
In the syntax-based arithmetic coding mode, an arithmetic code is used instead of VLC. Arithmetic codes
usually provide better entropy-coding performance. The average gain for interframes is about 3% to 4%. For
intrablocks and frames, the gain averages about 10%.
The unrestricted MV mode allows the MVs to point outside the picture. This mode is very useful when
an object moves along or beyond the edge of the picture, especially for the smaller picture formats. The pixels
on the edge row (or column) are replicated to cover the area outside the picture for block matching. This mode
includes an extension of the motion-vector range from [−16,15.5] to [−31.5,31.5] so that larger motion vectors
can be used.
The advanced prediction mode is used in conjunction with the unrestricted MV mode to achieve better
motion prediction. This mode turns on the four MV option and the overlapped-block motion-compensation
(OBMC) option. With the four MV option, four 8 × 8 vectors instead of one 16 × 16 vector are used for some
macroblocks in the picture. The MV for the two chrominance blocks is obtained by averaging the four MVs then
further dividing the average by two. The OBMC option is used for the luminance component of P pictures. For
each 8 × 8 luminance prediction block, the MV of the current luminance block and two remote MVs are chosen.
One remote MV is selected from the two MVs of the blocks to the left and right of the current luminance block
and the other from the two MVs of the blocks above and below the current luminance block. Each pixel value in
the prediction block is a weighted sum of the three predicted values obtained from the three MVs. The remote
MV selection depends on the pixel position in the current block. If the pixel is on the left half of the block, the
left-block MV is chosen, otherwise the right-block MV is chosen. The same is true for the top/bottom selection.
If one of the surrounding blocks was not coded or was coded in intramode, the corresponding remote MV is set
to zero. If the current block is at the border of the picture and therefore a surrounding block is not present, the
corresponding remote MV is replaced by the current MV. The advantage of OBMC is that the blocking artifact
is greatly reduced since every pixel is predicted by three overlapped blocks.
In the PB-frame mode, two pictures are coded as one unit, called a PB-frame. One P picture is predicted
from the last decoded P picture. It is followed by one B-picture predicted from both the last decoded P picture
and the P picture currently being decoded. Because most of the information transmitted is for the P picture in
the PB-frame, the frame rate can be doubled with this mode without increasing the bit rate much for relatively
simple sequences. For sequences with a lot of motion, however, PB-frames do not work as well as the B pictures
in MPEG.
With H.263, it is possible to achieve the same quality as H.261 with 30–50% of the bit usage due to the
half-pixel prediction and negotiable options in H.263. There are also less overhead and improved VLC tables
in H.263. H.263 also outperforms MPEG-1/MPEG-2 for low resolution and low bit rates due to the use of the
18 IMAGE CODES
four negotiable options. Another reason is that H.263 is less flexible than MPEG thus much less overhead is
needed.
More enhancements were made for H.263 and summarized in the new H.263+, a near-term version of
H.263. H.263+ supports more picture formats and provides more coding options. The advanced intracoding
mode allows predicting an intrablock using neighboring intrablocks. The deblocking filter mode further reduces
the blocking artifact by postfiltering the 8 × 8 block edges with a low-pass filter. The slice-structured mode
provides MPEG-like slice structures which act as resynchronization points for bit error and packet loss recovery.
The PB-frame mode is also improved to enhance predictive performance. It also adds temporal, SNR, and spatial
scalability to H.263. Other enhancements are also made.
H.263L, another refinement of H.263, will be completed in a longer time frame than H.263+. A large
change from H.263 and H.263+ is expected, and H.263L might be aligned with the MPEG-4 development and
carry similar functionalities.
Other Compression Techniques
Wavelet Compression. Wavelet transform recently attracted a lot of attention in the transform coding
field. It provides better performance than DCT-based coding techniques, both in high and low bit rates. Both
JPEG 2000 and MPEG-4 are likely to adopt wavelet transforms in their codec.
The wavelet theory states that a signal can be represented by a series of translations and dilations of
a basis function that meets certain mathematical requirements. Instead of using the global transformation
as in DCT, the wavelet transform uses finite-impulse-response (FIR) filters to capture the space-frequency
localization characteristics of the signal. This is usually accomplished by using the filter bank approach. The
signal is passed through a quadrature mirror filter (QMF) bank consisting of a low- and high-pass filter pair
denoted, respectively, by h(k) and g(k) with g(k) = (−1)k h(1 − k). The forward transform is written as
and
whereas the inverse transform takes the form
The h(n − 2k) and g(n − 2k) are implemented by filtering followed by downsampling operations, when per-
forming the forward transform, or filtering preceded by upsampling operations when performing the inverse
transform. The low- or high-passed signals are called the subband signals. The data amount in each of these is
half of that of the original signal because of the downsampling process. The 2-D wavelet transform is performed
by cascading a horizontal filtering operation and a vertical filtering operation. Thus each subband after the 2-D
transform has one-quarter the number of coefficients. The wavelet transform and the subband representation
are shown in Fig. 8(a).
IMAGE CODES 19
Fig. 8. Wavelet transform: (a) Filter bank implementation. The signal is filtered by high-pass and low-pass filters and
downsampled by two. Each of the resulting low-pass and high-pass signals has half of the number of samples. (b) Pyramidal
wavelet transforms. Each LL subband is further decomposed to form a regular pyramidal structure. (c) Wavelet packet
transform. The subbands are further decomposed according to some criterions, for example, the energy distribution. They
do not necessarily form a regular structure, therefore additional structural information has to be coded.
An advantage of wavelet transform in image processing is its flexibility to further decompose the image
in the subbands of interest. With the desirable energy compaction property in mind, we can further decompose
the subbands with higher energies to refine the bit-allocation strategy in these bands. This approach is called
the wavelet packet transform (WPT). For most images, however, the low-pass subband usually has the highest
energy. Therefore the successive decomposition of the LL band gives fairly good performance. This approach is
called the pyramid wavelet transform. These two approaches are shown in Fig. 8(b) and (c).
Wavelet image compression started quite early with performance comparable to DCT compression. The
advent of a series of modern wavelet coders, however, boosted the performance while providing a nice embedded
bit-stream property. In an embedded bit stream, the transform coefficients are quantized successively so that the
most significant coefficients are quantized and transmitted first. More details of the image can be successively
added to refine the image if the bit rate allows. In this way, the bit rate is precisely controlled down to the bit
level while keeping good performance.
The performance breakthough of modern wavelet coders results from exploiting the correlation between
parent and child subbands. Shapiro’s embedded zerotree wavelet (EZW) coder (9) partitioned the subbands
into parent-child groups with same horizontal-vertical wavelet decomposition. If one coefficient in the parent
subband is less than some threshold, then the coefficients in the corresponding child subbands are most likely
also smaller than this threshold. Therefore only coding the zerotree root is enough. After the quantization and
grouping procedure, the wavelet coefficients are represented by four symbols, positive, negative, isolated zero,
and the zerotree root. They are coded with an arithmetic coder. Both the subjective quality and PSNR are
greatly improved.
20 IMAGE CODES
Several embedded wavelet coders followed the EZW and made more improvements in both the perfor-
mance and coding complexity. The coefficient representation and prediction scheme were refined by the layer
zero coding (LZC) technique (10). In LZC, the coefficients were simply represented by zero and one according
to their significance, rather than the four-symbol representation of the EZW. The prediction of wavelet coeffi-
cients is implemented in the context choice of the adaptive arithmetic coder. The parent-child relationship and
the zerotree structure were further exploited by set partitioning in the hierarchical tree (SPHIT) algorithm
(11), which identified more special classes of tree structures of bits in the significant trees. The multithreshold
wavelet coding (MTWC) technique (12) uses multiple quantization thresholds for each subband for better bit
allocation and rearranges the transmission order of wavelet coefficients to achieve better performance. The
latter two have low computational complexity and can be implemented for real-time image compression.
Vector Quantization. Different from scalar quantization, vector quantization (VQ) uses a quantization
index (codeword) to represent a vector to be quantized. Therefore VQ reduces the redundancy if the vectors
are closely correlated. VQ is applied to image coding in two ways. One is to use VQ as the replacement of the
scalar quantizer in transform coding schemes, and the other is to treat clusters of image pixels as the vectors
and perform the VQ.
The first method is useful if the transform coefficients are correlated in some way. For DCT coding, the
correlation among DCT coefficients in the same block or across adjacent blocks is not very strong so that VQ
cannot improve the performance too much. For wavelet coding, the coefficients are more closely correlated
among nearby coefficients in the same subband or among parent-child subbands. Thus using VQ to quantize
the wavelet coefficients can improve the performance to a larger extent.
In the second method, the correlation among adjacent pixels in the spatial domain is exploited. To perform
VQ, a fixed block size is chosen. The pixel values from the block are chosen as the vector. All of the vectors from
several “typical” images are collected as the training vectors, and a training algorithm is chosen. These vectors
are trained to form a codebook with a number of representative vectors, called codewords. When compressing
an image, every block from the image is extracted as a vector and the nearest codeword from the codebook is
found. The index of the codeword is transmitted and the corresponding codeword is used as the reconstruction
vector.
The disadvantage of VQ is that the training time required may be too long to form a codebook. Considering
the training cost, the proper block size may be about 4 × 4. In addition, the performance depends on which
images are used in the training set and which image is to be compressed. If a universal codebook is used, the
performance is not optimal. The performance of traditional VQ is usually a lot worse than the transform coders.
The advantage of the VQ coder is that once the training is completed, the encoding speed is quite fast if the size
of the codebook is not too large. The decoding speed is extremely fast because no real computation is needed.
There are several variants of VQ that improve the performance and reduce the computational complexity (see
Data compression, lossy).
Fractal Compression. Fractal compression schemes (13) exploit the spatial redundancy by utilizing
the self-similarity in the same image. Given a target region in an image, there could be another region similar
to this region with different rotation, scale, and contrast. If we could find this approximation, we could encode
the transformation (rotation, scale, contrast, and the displacement) from the target region. This could be very
efficient because only a small amount of information needs to be encoded. The coding and decoding of the
image is based on the partitioned iterated function system (PIFS). The classical way of separating the image
into regions is to partition it into fixed-size blocks and find a similar block with some transformations. The
transformation information is coded and transmitted. At the decoder end, an initial image (it could even be a
random image!) is chosen, and the transformation is applied iteratively to each corresponding block. According
to its mathematical theory, the image will converge to the original image after iterations.
The advantage of fractal compression is that it is good for low bit-rate compression because most of the
information is included in the image itself and only a small amount of transformation information is needed
for encoding a large block. It also alleviates the blocking artifacts in other block-based coding schemes. Another
IMAGE CODES 21
advantage is that it can be used to enhance the resolution of images even beyond the original resolutions of the
images because the iterative process can be extended to subpixel levels. The disadvantage is that the encoding
may be too time-consuming in finding a similar block. Rather than pixel-wise matching to find a matched block,
fractal compression has to perform all possible transformations to a block in the searching window to find the
best match. The decoding time, on the other hand, is relatively fast without too much computation involved.
An advantage of the iterative process is that we can cut off the decoding process at an arbitrary number of
iterations. But, of course, the result may not be good if too few iterations are performed.
BIBLIOGRAPHY
1. ITU-R Recommendation BT.601, Encoding parameters of digital television for studios, 1982.
2. ISO/IEC JTC1 10918-1, Information technology—digital compression and coding of continuous-tone still images: re-
quirements and guidelines, 1994.
3. ISO/IEC JTC1/SC29/WG1 N390R, New work item: JPEG 2000 image coding system, 1997.
4. ISO/IEC-11172, Information technology—Coding of moving pictures and associated audio for digital storage media at
up to about 1.5 Mbps, 1992.
5. ISO/IEC-13818—ITU-T Rec. H.262: Information technology: Generic coding of moving pictures and associated audio,
1996.
6. ISO/IEC JTC1/SC29/WG11 N1730, Overview of the MPEG-4 standard, 1997.
7. ISO/IEC JTC1/SC29/WG11 N1920, MPEG-7: Context and objectives, 1997.
8. ITU-T Recommendation H.263, Video coding for low bit-rate communication, 1995.
9. J. Shapiro Embedded image coding using zerotrees of wavelet coefficients, IEEE Trans. Signal Process., 41: 3445–3462,
1993.
10. D. Taubman A. Zakhor Multirate 3-D subband coding of video, IEEE Trans. Image Process., 3: 572–588, 1994.
11. A. Said W. A. Pearlman A new, fast, and efficient image codec based on set partitioning in hierarchical trees, IEEE
Trans. Circuits Syst. Video Technol., 6: 243–250, 1996.
12. H.-J. Wang C.-C. J. Kuo A multi-threshold wavelet coder (MTWC) for high fidelity image, IEEE Signal Process. Soc.,
1997 Int. Conf. Image Process., 1997.
13. Y. Fisher, ed. Fractal Image Compression: Theory and Applications, New York: Springer-Verlag, 1995.
YUNG-KAI LAI
C.-C. JAY KUO
University of Southern California
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Information Theory ● Email this page Enter words or
● Print this page
Pramod K. Varshney1
1Syracuse University, Syracuse, NY
reserved. Search
Abstract
Communication System Model
Entropy
Source Coding
Mutual Information
Relative Entropy
Channel Capacity
Channel Coding Theorem
Differential Entropy
Capacity of Gaussian Channels
Rate Distortion Theory
Acknowledgment



INFORMATION THEORY 139
received signals. He only provided an existence proof stating

that such procedures exist but did not specify an approach to
design the best encoders and decoders. Also, he did not dis-
cuss the implementation complexity. These results have pro-
vided the impetus for researchers to try to design encoding
and decoding procedures that approach the fundamental lim-
its given by information theory.
While information theory was primarily developed as a
mathematical model for communications, it has had an im-
pact on a wide variety of fields that include physics, chemis-
try, biology, psychology, linguistics, statistics, economics, and
computer science. For example, languages provide a means
for communication between human beings, and application of
information theory to linguistics arises naturally. Examples
of application of information theory to computer science in-
clude the design of efficient decision trees and introduction of
redundancy in computer systems to attain fault-tolerant com-
puting.
COMMUNICATION SYSTEM MODEL
The main components of a digital communication system are

shown in Fig. 1. The source is assumed to be a digital source
in that a symbol from a finite alphabet is generated in dis-
crete time. An analog source can be converted to a digital
source by sampling and quantization. Data from the source
are processed by the source encoder, which represents the
source data in an efficient manner. The objective of the source
encoding operation is to represent the source output in a com-
pact form with as high fidelity as possible (i.e., with as little
information loss as possible). The sequence of source code-
words generated by the source encoder is fed to the channel
INFORMATION THEORY encoder, which yields the sequence of channel codewords. The
channel encoder adds redundancy to provide error control
capabilities. The goal is to exploit the redundancy in the most
The primary goal of a communication system is to convey in-
effective manner by achieving a high degree of error control
formation-bearing messages from an information source to a
capability for a specified amount of redundancy. In some en-
destination over a communication channel. All real channels
coding schemes, the input data stream is divided into blocks
are subject to noise and other channel impairments that limit
of fixed length, and then some additional symbols are added
communication system performance. The receiver attempts to to each block to yield channel codewords. These codes are
reproduce transmitted messages from the received distorted known as block codes. In the class of codes known as tree
signals as accurately as possible. codes, the encoding process exhibits memory in that a block
In 1948, Shannon proposed a mathematical theory for the of input data stream is encoded based on the past blocks also.
communication process. This theory, known as information In either case, the output of the channel encoder is a string
theory, deals with the fundamental limits on the representa- of symbols to be transmitted. The modulator converts source
tion and transmission of information. Information theory was codeword symbols to analog waveforms suitable for transmis-
a remarkable breakthrough in that it provided a quantitative sion over the channel. The received waveforms are distorted
measure for the rather vague and qualitative notion of the due to noise and other interference processes present over the
amount of information contained in a message. Shannon sug- channel. The demodulator converts the received waveform
gested that the amount of information conveyed by the occur- into symbols and then furnishes received words to the chan-
rence of an event is related to the uncertainty associated with nel decoder. Due to channel noise, the received word may be
it and was defined to be inversely related to the probability in error. The channel decoder exploits the redundancy intro-
of occurrence of that event. Information theory also provides duced at the channel encoder to detect and/ or correct errors
fundamental limits on the transmission of information and on in the received word. This corrected word is the best estimate
the representation of information. These fundamental limits of the source codeword, which is delivered to the destination
are employed as benchmarks and are used to evaluate the after performing the inverse of the source encoding operation.
performance of practical systems by determining how closely Information theory is based on a probabilistic model of this
these systems approach the fundamental limits. communication system.
In his celebrated work, Shannon laid the foundation for
the design and analysis of modern communication systems. ENTROPY
He proved that nearly error-free information transmission
over a noisy communication link is possible by encoding sig- Let the discrete random variable S represent the output of
nals prior to transmission over the link and by decoding the a source generating a symbol every signaling interval in a
140 INFORMATION THEORY
Source Channel
Source Modulator
encoder encoder
Noise Channel
Source Channel
Destination Demodulator
decoder decoder
Figure 1. Block diagram of a communication system.
statistically independent manner. This discrete memoryless It is plotted in Fig. 2 as a function of p0. Note that H(S) is
source (DMS) is assumed to generate symbols from a fixed zero when p0 ⫽ 0 or 1. This corresponds to no uncertainty.
finite alphabet 兵s1, . . ., sK其 with probabilities P(S ⫽ sk) ⫽ pk, When p0 ⫽ , H(S) ⫽ 1. This corresponds to maximum uncer-
k ⫽ 1, . . ., K. The amount of information gained after ob- tainty since symbols 0 and 1 are equally likely.
serving the symbol sk is defined by the logarithmic function
I(sk ) = log(1/pk ) SOURCE CODING
One of the important problems in communications is an effi-

It is inversely related to the probability of a symbol occur-
cient representation of symbols generated by a DMS. Each
rence. The base of the logarithm is usually taken to be 2 and
symbol sk is assigned a binary codeword of length ᐉk. For an
the unit is called a bit. In this article, the base of all loga-
efficient representation, it is desirable to minimize the aver-
rithms is assumed to be 2. Some properties of I(sk) are as
age codeword length L, where
follows:

K
1. If the outcome of an event is certain, no information L= pk k
gain occurs; that is, k=1
I(sk ) = 0 if pk = 1 Shannon’s first theorem, also known as the source coding the-
orem, provides a fundamental limit on L in terms of the en-
tropy of the source.
2. Information gain from the occurrence of an event is
nonnegative; that is,
Source Coding Theorem: Given a DMS with entropy
I(sk ) ≥ 0 for 0 ≤ pk ≤ 1 H(S), the average codeword length L for any source encoding
scheme is bounded as
3. Occurrence of less probable events results in more infor- L ≥ H(S)
mation gain; that is,
I(sk ) > I(s ) if pk < p Thus, entropy of a DMS provides a fundamental limit on the
average number of bits per source symbol necessary to repre-
The average information per source symbol for a DMS is
obtained by determining the average of I(s1), . . ., I(sK).
1

K
H(S) = pk log(1/pk ) 0.9
k=1
0.8
This quantity is known as the entropy of the DMS. It charac- 0.7

terizes the uncertainty associated with the source and is a 0.6
H(p0)
function of source symbol probabilities. The entropy is

0.5
bounded as 0 ⱕ H(S) ⱕ log2 K. The lower bound is attained
when one of the symbols occurs with probability one and the 0.4
rest with probability zero. The upper bound is realized when 0.3
all the symbols are equally likely.
0.2
Example: Consider a binary DMS whose output symbols are 0.1

zero and one with associated probabilities of occurrence given 0
by p0 and p1, respectively. The entropy is given by 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
p0
H(S) = −p0 log p0 − p1 log p1 Figure 2. Binary entropy function.
Table 1. Illustration of Huffman Coding Algorithm

Probabilities at Different Stages
Source Symbols 1 2 3 4 5 Codewords
1
s0 0.3 0.3 0.45 0.55 1.0 11
1 0
s1 0.25 0.25 0.3 0.45 10
0
1
s2 0.25 0.25 0.25 01
1 0
s3 0.1 0.2 001
0
s4 0.1 000
sent the DMS. Based on this lower bound on L, we can ex- In this case,
press the coding efficiency of a source encoder as
H(S) = − 0.3 log 0.3 − 0.25 log 0.25
H(S) − 0.25 log 0.25 − 0.1 log 0.1 − 0.1 log 0.1
η=
L = 2.1855 bits/symbol
A source encoder that is able to attain the lower bound has

and L ⫽ 2.2 bits/symbol. Thus, L ⬎ H(S) and ␩ ⫽ 0.9934.
an efficiency of one.
An important requirement for source codes is that they be
uniquely decodable so that perfect reconstruction is possible
from the encoded binary sequence. One class of uniquely de- MUTUAL INFORMATION
codable codes is the class of prefix-free codes. In these codes,
no codeword is a prefix of any other codeword. Huffman code Let X and Y be two discrete random variables that take val-
is an example of such a source code in which L approaches ues from 兵x1, . . ., xJ其 and 兵y1, . . ., yK其, respectively. The con-
H(S). This code is optimum in that no other uniquely decoda- ditional entropy H(X兩Y) is defined as
ble code has a smaller L for a given DMS. The basic procedure
for Huffman coding can be summarized as follows:
K
J
H(X |Y ) = p(x j , yk ) log[1/p(x j |yk )]
k=1 j=1
1. Arrange the source symbols in decreasing order of prob-
abilities.
This quantity represents the amount of uncertainty re-
2. Assign a 0 and a 1 to the two source symbols with low- maining about X after observing Y. Since H(X) represents the
est probability. original uncertainty regarding X, information gained regard-
3. Combine the two source symbols into a new symbol ing X by observing Y is obtained by the difference of H(X)
with probability equal to the sum of two original proba- and H(X兩Y). This quantity is defined as the mutual informa-
bilities. Place this new symbol in the list according to tion I(X; Y).
its probability.
4. Repeat this procedure until there are only two source I(X;Y ) = H(X ) − H(X |Y )
symbols in the list. Assign a 0 and a 1 to these two
symbols. Some important properties of I(X; Y) are as follows:
5. Find the codeword for each source symbol by working
backwards to obtain the binary string assigned to each
1. The mutual information is symmetric with respect to X
source symbol. and Y; that is,
Example: Consider a DMS with an alphabet consisting of I(X;Y ) = I(Y; X )

five symbols with source probabilities, as shown in Table 1.
Different steps of the Huffman encoding procedure and the
2. The mutual information is nonnegative; that is,
resulting codewords are also shown. Codewords have been ob-
tained by working backward on the paths leading to individ-
ual source symbol. I(X; Y ) ≥ 0
3. I(X; Y) is also given as 1

0.9
I(X; Y ) = H(Y ) − H(Y |X )
0.8
0.7
Channel capacity
RELATIVE ENTROPY
0.6
The relative entropy or discrimination is a measure of the 0.5
distance between two probability distributions. Let p( ⭈ ) and
0.4
q( ⭈ ) be two probability mass functions. Then relative entropy
or Kullback Leibler distance between the two is defined as 0.3
0.2

K
p(xk )
D(pq) = p(xk ) log 0.1
k=1
q(x k)
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
The relative entropy is always nonnegative and is zero only if Transition probability, p
p and q are identical.
The mutual information I(X; Y) can be interpreted as the Figure 4. Capacity of a binary symmetric channel.
relative entropy between the joint distribution p(xj, yk) and
the product distribution p(xj) p(yk). That is,
Channel capacity is a function only of the channel transition
I(X ; Y ) = D(p(x j , yk )p(x j )p( yk )) probabilities and its units are bits per channel use.
Example: The capacity of a BSC as a function of the error

probability p is given by
CHANNEL CAPACITY
C = 1 − H(p)
Consider a discrete channel with input X and output Y, where
X and Y are discrete random variables taking values from and is shown in Fig. 4. When p ⫽ 0 or p ⫽ 1, the channel
(x1, . . ., xJ) and (y1, . . ., yK), respectively. This channel is capacity is maximum and is equal to 1 bit. Note that p ⫽ 1
known as a discrete memoryless channel (DMC) if the output also corresponds to a deterministic channel in that a zero is
symbol at any time depends only on the corresponding input always received as a one and vice versa. When p ⫽ , the
symbol and not on any prior ones. This channel can be com- channel is very noisy and the capacity is zero.
pletely characterized in terms of channel transition probabili-
ties, p(yk兩xj); j ⫽ 1, . . ., J; k ⫽ 1, . . ., K.
CHANNEL CODING THEOREM
Example: An important example of a DMC is the binary
To combat the effects of noise during transmission, the incom-
symmetric channel (BSC) shown in Fig. 3. In this case, both
ing data sequence from the source is encoded into a channel
the input and the output take values from 兵0, 1其 and the two
input sequence by introducing redundancy. At the receiver,
types of errors (receiving a zero when a one is sent, and re-
the received sequence is decoded to reconstruct the data se-
ceiving a one when a zero is sent) are equal.
quence. Shannon’s second theorem, also known as the chan-
nel coding theorem or the noisy coding theorem, provides the
For a DMC, mutual information I(X; Y) is the amount of
fundamental limits on the rate at which reliable information
input source uncertainty reduced after observing the output.
transmission can take place over a DMC.
The channel capacity of a DMC is defined as the maximum
mutual information for any signaling interval, where the Channel Coding Theorem
maximization is performed over all possible input probability
distributions. That is, (i) Let a DMS with entropy H(S) produce a symbol every
Ts seconds. Let a DMC have capacity C and be used
C = max I(X;Y ) once every Tc seconds. Then, if
{ p(x j )}
H(S) C
≤
Ts Tc
1−p
0 0 there exists a coding scheme with which source output
can be transmitted over the channel and be recon-
p structed at the receiver with an arbitrarily small proba-
bility of error. Here, error refers to the event that a
p transmitted symbol is reconstructed incorrectly.
(ii) Conversely, if
1 1
1−p H(S) C
>
Figure 3. Binary symmetric channel. Ts Tc
it is not possible to transmit data with an arbitrarily where B is the bandwidth of the channel, P is the average
small probability of error. transmitted signal power, and the noise power spectral den-
sity is equal to N0 /2.
It must be emphasized that the foregoing result only states
the existence of ‘‘good’’ codes but does not provide methods The capacity provides a fundamental limit on the rate at
to construct such codes. Development of efficient codes has which information can be transmitted with arbitrarily small
remained an active area of research and is discussed else- probability of error. Conversely, information cannot be trans-
where in this volume. In error-control coding, redundant sym- mitted at a rate higher than C bits/s with arbitrarily small
bols are added to the transmitted information at the trans- probability of error irrespective of the coding scheme em-
mitter to provide error detection and error correction ployed.
capabilities at the receiver. Addition of redundancy implies
increased data rate and thus an increased transmission band-
width. RATE DISTORTION THEORY
DIFFERENTIAL ENTROPY Previously, the problem of source coding that required perfect
reconstruction of a DMS was considered. It was seen that the
Thus far, only discrete random variables were considered. entropy provided the minimum rate at which perfect recon-
Now we define information theoretic quantities for continuous struction is possible. A question arises as to what happens
random variables. Consider a continuous random variable X when the allowed rate is less than the lower bound. Also,
with probability density function f(x). Analogous to the en- what if the source is continuous, because a finite representa-
tropy of a discrete random variable, the differential entropy tion of such a source can never be perfect? These questions
of a continuous random variable X is defined as give rise to rate distortion theory. A distortion measure needs
∞ to be defined to quantify the distance between the random
variable and its representation. For a given source distribu-
h(x) = f (x) log[1/ f (x)] dx
−∞ tion and distortion measure, the fundamental problem in rate
distortion theory is to determine the minimum achievable ex-
Example: For a Gaussian random variable with probability pected distortion at a given rate. An equivalent problem is to
density function, find the minimum rate required to attain a given distortion.
This theory is applicable to both continuous and discrete ran-
1 x2 dom variables.
f (x) = √ exp − 2
σ 2π 2σ Consider a source with alphabet X that produces a se-
quence of independent identically distributed random vari-
the differential entropy can be computed to be ables X1, X2, . . .. Let X̂1, X̂2, . . . be the corresponding repro-
ductions with reproduction alphabet denoted as Xˆ . The
h(x) =
1
log 2πeσ 2 bits single-letter distortion measure d(x, x̂) is a mapping d: X xXˆ
2 씮 R⫹ from the source alphabet-reproduction alphabet pair
into the set of nonnegative real numbers. It quantifies the
In an analogous manner, mutual information for two con- distortion when x is represented by x̂. Two commonly used
tinuous random variables X and Y can be defined as distortion measures are as follows:
∞ ∞ Hamming distortion measure:
f (x, y)
I(X; Y ) = f (x, y) log dx dy
−∞ −∞ f (x) f ( y)
0 if x = x̂
d(x, x̂) =
1 if x = x̂
CAPACITY OF GAUSSIAN CHANNELS
Squared error distortion measure:
Earlier, the fundamental limit on error-free transmission over
a DMC was presented. Here we present the channel capacity
d(x, x̂) = (x − x̂)2
theorem for band-limited and power-limited Gaussian chan-
nels. This theorem is known as Shannon’s third theorem or as
the Shannon–Hartley theorem. It is an extremely important The single-letter distortion measure can be extended to define
result with great practical relevance because it expresses the the distortion measure for n-tuples as follows:
channel capacity in terms of system parameters channel
bandwidth, average signal power, and noise power spectral 1 n
d(xn , x̂n ) = d(xi , x̂i )
density. n i=1
Channel Capacity Theorem: The capacity of a band-lim- This is the average of the per symbol distortion over the ele-
ited additive white Gaussian noise (AWGN) channel is given ments of the n-tuple.
by Now we consider the encoding of the source output se-
P
quence of length n, Xn, and then its decoding to yield X̂n. To
C = B log 1 + bits/s accomplish this we define a (2nR, n) rate distortion code that
N0 B consists of an encoding function and a decoding function, as
given by 3.5
fn : X n
→ {1, 2, . . ., 2nR } 3
gn : {1, 2, . . ., 2 nR
} → X̂ n
2.5
where R is the number of bits available to represent each
2
source symbol. The expected distortion for this rate distortion
R(D)
code is given by
1.5

Dn = p(xn )d(xn , gn ( f n (xn )))
1
xn
where p( ⭈ ) is the probability density function associated with 0.5

the source.
A rate distortion pair (R, D) is said to be achievable if there 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
exists a rate distortion code with rate R such that D
lim Dn ≤ D Figure 6. Rate distortion function for the Gaussian source.
n→∞
The rate distortion function R(D) is the infimum of rates R

such that (R, D) is achievable for a given D. Next, we present Example: Consider a zero-mean Gaussian source with vari-
the fundamental theorem of rate distortion theory. ance ␴2. For the squared error distortion measure, the rate
distortion function is given by
Rate Distortion Theorem: The rate distortion function for 
an independent identically distributed source X with distribu-  1 σ2
R(D) = log 0 ≤ D ≤ σ2
tion p(x) and bounded distortion function d(x, x̂) is given by  2 0 D D > σ2
R(D) = min I(X; X̂ )
p(x̂|x): p(x) p(x̂|x)d(x, x̂)≤D
(x, x̂ ) It is plotted in Fig. 6.
Thus, R(D) is the minimum achievable rate at distortion D. The rate distortion function R(D) is a nonincreasing convex
Conversely, if R is less than R(D), we cannot achieve a distor- function of D. For the binary source, when D ⫽ 0, the mini-
tion less than or equal to D. mum rate required for perfect reconstruction is given by
H(p). As D increases, minimum required rate R decreases.
Example: Consider a binary source that produces an output Similar observations can also be made for the Gaussian
of 1 with probability p. For the Hamming distortion measure, source.
its R(D) is given by

H(p) − H(D) 0 ≤ D ≤ min( p, 1 − p) ACKNOWLEDGMENT
R(D) =
0 D > min( p, 1 − p)
I would like to thank Qian Zhang for his help in the prepara-
tion of this article. This article was written while the author
It is illustrated in Fig. 5 for p ⫽ 0.5. was a Visiting Scientist at the Air Force Research Laboratory
at Rome Research Site, AFRL/IFG, 525 Brooks Road, Rome,
NY 13441-4505.
1
0.9
BIBLIOGRAPHY
0.8
0.7 For a detailed discussion of information theory and its applications,
the reader is referred to the sources listed below. Recent results on
0.6
this topic are published in the IEEE Transactions on Information
R(D)
0.5 Theory.
T. Berger, Rate Distortion Theory, Englewood Cliffs, NJ: Prentice-
0.4
Hall, 1971.
0.3 R. E. Blahut, Principles and Practice of Information Theory, Reading,
0.2 MA: Addison-Wesley, 1987.
R. E. Blahut, Theory and Practice of Error Control Codes, Reading,
0.1
0 T. M. Cover and J. A. Thomas, Elements of Information Theory, New
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
York: Wiley, 1991.
D
R. G. Gallagher, Information Theory and Reliable Communication,
Figure 5. Rate distortion function for the binary source. New York: Wiley, 1968.
INFORMATION THEORY OF DATA TRANSMISSION CODES 145
R. W. Hamming, Coding and Information Theory, Englewood Cliffs,

S. Lin and D. J. Costello, Jr., Error Control Coding: Fundamentals
and Applications, Englewood Cliffs, NJ: Prentice-Hall, 1983.
M. Mansuripur, Introduction to Information Theory, Englewood Cliffs,
R. J. McEliece, The Theory of Information Theory and Coding, Read-
ing, MA: Addison-Wesley, 1977.
C. E. Shannon, A Mathematical Theory of Communication, Bell Syst.
Techn. J., vol. 27, pp. 379–423 (part I), and pp. 623–656 (Part
II), 1949.
PRAMOD K. VARSHNEY
Syracuse University
INFORMATION THEORY. See MAXIMUM LIKELIHOOD DE-

TECTION.
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Information Theory of Data Transmission Codes ● Email this page Enter words or
● Print this page
George Thomas1
1University of Southwestern Louisiana, Lafayette, LA
reserved. Search
Abstract
Data Sources and Channels
Block Coding
Convolutional Codes
Additional Topics
Applications

pression codes. Both strive to minimize the number of bits

transmitted per unit of time, the former without loss of fidel-
ity and the latter with possible, controlled reduction in fidel-
ity. This source encoder is followed by the channel encoder,
which uses data transmission codes to control the detrimental
effects of channel noise. Controlled amounts of redundancy is
introduced into the data stream in a manner that affords er-
ror correction. These data transmission codes are the focus of
this article. Further down in the cascade, we have the modu-
lator which maps output strings from the channel encoder
into waveforms that are appropriate for the channel. (Tradi-
tionally, modulation has evolved as an art disjoint from cod-
ing, but some recent research has indicated the merits of com-
bined coding and modulation. We will touch upon this aspect
toward the end of this article.) Following the channel, the de-
modulator and the decoders have the corresponding inverse
functions which finally render the desired information to the
receiver. In brief, this article concentrates on data transmis-
sion codes.
Binary Symmetric Channels

INFORMATION THEORY OF DATA
TRANSMISSION CODES Each binary digit or (group of digits) at the input of the modu-
lator is transmitted as a waveform signal over the transmis-
The basic task of a communication system is to extract rele- sion channel. Physical transmission channels may distort the
vant information from a source, transport the information signal, the net result of which is to occasionally reproduce at
through a channel and to reproduce it at a receiver. Shannon, the output of the demodulator a binary string that is different
in his ground-breaking A Mathematical Theory of Communica- from what was actually sent. In many practical cases, the er-
tions (1), quantified the notions of the information rate of a ror events in successive binary digit positions are mutually
source and the capacity of a channel. He demonstrated the statistically independent. And in many such binary memo-
highly non-intuitive result that the fundamental restrictive ryless channels the probability of error, ⑀, is the same for a
effect of noise in the channel is not on the quality of the infor- transmitted 0 as well as for a transmitted 1 (Fig. 1). Such a
mation, but only on the rate at which information can be binary symmetric channel (BSC) is an important abstraction
transmitted with perfect quality. Shannon considered coding in data transmission coding.
schemes, which are mappings from source outputs to trans- If a binary n-vector is transmitted sequentially (i.e., bit by
mission sequences. His random-coding arguments established bit) over a binary symmetric channel with bit error probabil-
the existence of excellent codes that held the promise of ity ⑀, the number of errors is a random variable with a Ber-
nearly zero error rates over noisy channels while transmitting noulli distribution:
data at rates close to the channel capacity. Shannon’s exis-
tence proofs did not, however, provide any guidelines toward n i
P(i) = (1 − )n−1 , 0≤i≤n
actually constructing any of these excellent codes. A major i
focus of research in information theory (as Shannon’s theory
came to be known) over the past 50 years following Shannon’s If ⑀ ⬍ , as is the case for most practically useful channels,
seminal work has been on constructive methods for channel P(i) is seen to diminish exponentially in i as (⑀ /1 ⫺ ⑀)i. This
coding. A number of later books (e.g., Refs. 2–9) journalize implies that P(0) ⬎ P(1) ⬎ P(2) ⬎ ⭈ ⭈ ⭈ ⬎ P(n). More specifi-
the development of information theory and coding. In this ar- cally, P(0) is typically very large, P(1) is O(⑀) (i.e., on the order
ticle we present a broad overview of the state of the art in of ⑀), P(2) is O(⑀2), and so forth. Thus, even minimal levels of
such data transmission codes. error correction can bring about a significant performance im-
provement on the BSC.
DATA SOURCES AND CHANNELS
Hamming Distance and Hamming Weight
A very general schematic representation of a communication The BSC can be modeled as a mod-2 additive noise channel
link consists of the following cascade: the source, a source en- characterized by the relation Y ⫽ X 丣 N, where X is the
coder, a channel encoder, a modulator, the channel, the de-
modulator, the channel decoder, the source decoder, and the
receiver. The source encoder typically converts the source in- 1–
formation into an appropriate format taking into account the 0 0
quality or fidelity of information required at the receiver.
Sampling, quantization and analog-to-digital conversion of an
analog source, followed possibly by coding for redundancy re- 1 1
1–
moval and data compression, is an example. The codes used
here are usually referred to as data compaction and data com- Figure 1. The binary symmetric channel with error probability ⑀.
146 INFORMATION THEORY OF DATA TRANSMISSION CODES
transmitted binary digit, N is a noise bit, Y is the correspond- The essence of code design is the selection of a sufficient
ing output bit, and ‘‘丣’’ denotes mod-2 addition. The Ham- number of n-vectors sufficiently spaced apart in binary n-
ming weight of a binary n-vector is defined as the number of space. Decoding can in principle be done by table lookup but
1’s in it, and the Hamming distance [in honor of R. W. Ham- is not feasible in practice as the code size grows. Thus we
ming, a coding theory pioneer (10)] between two binary vec- are motivated to look for easily implemented decoders. Such
tors is defined as the number of bit positions where the ele- practical coding schemes fall generally into two broad catego-
ments of the two vectors are different. It is easy to see that ries: block coding and convolutional coding.
the mod-2 sum of two binary n-vectors has a Hamming weight
equal to the Hamming distance between the two vectors. If a
BLOCK CODING
binary input n-vector Xn to a BSC produces the output n-vec-
tor Yn, then the noise pattern Nn ⫽ Xn 丣 Yn is a binary n-
Linear Block Codes
vector whose Hamming weight is equal to the Hamming dis-
tance between Xn and Yn. (If a 丣 b ⫽ c, then a ⫽ b 丣 c). An (n, k) block code maps every k-bit data sequence into a
Consider the n-space Sn of all binary n-vectors. Out of the corresponding n-bit codeword, k ⬍ n. There are 2k distinct n-
total 2n n-vectors in Sn, if we choose only a few vectors well vector codewords in a linear block code. The code rate R ⫽
separated from each other, we can hope that noise-corrupted k/n is a measure of the data efficiency of the code. A linear
versions of one codeword will not be confused with another block code has the property that for any two codewords Xin
valid codeword. To illustrate, suppose we choose a code with and Xjn, their bitwise mod-2 sum Xin 丣 Xjn is also a codeword.
two binary n-vector codewords X1n and X2n which are mutually Using a geometric perspective, we can view the code as a k-
at a Hamming distance d. In Fig. 2 we have shown the exam- dimensional linear subspace of the n-dimensional vector
ple of S4 with X14 ⫽ (0000) and X24 ⫽ (1111) at Hamming dis- space Sn, spanned by k basis vectors. Using matrix notation,
tance d ⫽ 4. (S4 is a four-dimensional hypercube, with each we can then represent the linear encoding operation as Yn ⫽
node having four neighbors at unit distance along the four XkG, where the k-vector Xk is the data vector, Yn is the corre-
orthogonal axes.) It can be seen that if codeword 0000 has at sponding n-vector codeword, and G is the k ⫻ n binary-valued
most one bit altered by the channel, the resulting 4-tuple generator matrix. The rows of G are a set of basis vectors for
(e.g., 0001) is still closer to 0000 than to 1111 so that a near- the k-space and thus are mutually linearly independent. Lin-
est-codeword decoding rule decodes correctly. But if 0000 en- ear codes have the important feature that the minimum dis-
counters two bit errors (e.g., 0011), the resulting word is at tance of the code is equal to the smallest among the nonzero
equal distance from either codeword; and if there are three Hamming weights of the codewords. (The all-zero n-vector is
bit errors (e.g., 1101), the nearest codeword now is 1111 and necessarily a codeword in each linear n-vector code.) If the
a decoding error results. In general, two codewords at a Ham- codewords are of the specific concatenated form Yn ⫽
ming distance d can be correctly decoded if the number of (XkPn⫺k), where Pn⫺k is a parity vector comprising n ⫺ k parity
bits which are solely functions of Xk (i.e., if the codeword Yn
errors incurred in the BSC is at most (d ⫺ 1)/2, where x
contains the data word Xk explicitly), then the code is termed
is the integer part of the number x. If in fact there are more
systematic. Systematic linear block codes have generator ma-
than two codewords in the code, it should be obvious that the
trices with the special structural form G ⫽ [IkP], where Ik is
pair of codewords with the minimum Hamming distance de-
the k ⫻ k identity matrix and P is a k ⫻ n ⫺ k parity genera-
termine the maximum number of bit errors tolerated. Thus,
tor matrix. Any linear block code can be put into an equiva-
a code with minimum distance dmin can correct all error pat-
lent code that is also systematic. A (7,4) Hamming code (dis-
terns of Hamming weight not exceeding t ⫽ (dmin ⫺ 1)/2.
cussed below) is one example of a linear block code with the
following generator matrix in its systematic form:
 
1111
1 0 0 0 1 1 0
1011 0 1 0 0 0 1 1
 
G= 
0 0 1 0 1 0 1
0 0 0 1 1 1 1
1001 1101
For every linear (n, k) block code, there is a parity check
1110 matrix H which is an (n ⫺ k ⫻ n) binary valued matrix with
1010
0011 the property that GHT ⫽ 0. Given G ⫽ [IkP], the correspond-
0111 ing parity check matrix has the structure H ⫽ [PTIn⫺k]. The
parity check matrix for the (7,4) systematic Hamming code is
1000 1100
as follows:
0001 0101  
1 0 1 1 1 0 0
 
0010 0110 H = 1 1 0 1 0 1 0
0 1 1 1 0 0 1
0000 0100 The condition GHT ⫽ 0 implies that every row in G, and con-
Figure 2. Minimum distance decoding. The two codewords 0000 and sequently every codeword, is orthogonal to every row in H.
1111 are at Hamming distance 4 in the space of binary 4-tuples. Every codeword Xn satisfies the parity check condition XnHT
⫽ 0. For an arbitrary Yn appearing at the output of a BSC, and is clearly an important performance parameter, espe-
the n ⫺ k vector S(Yn) ⫽ YnHT is called the syndrome of Yn. cially when codes are used for error detection only.
The 2n⫺k syndromes have a one-to-one correspondence with For the general class of linear block codes, the encoder im-
a set of 2n⫺k n-vector error patterns that the (n, k) linear code plements the multiplication of the data vector by the genera-
is capable of correcting. If n is small, a table lookup will suf- tor matrix. Decoding consists of computing the syndrome (by
fice to find the error pattern from the syndrome. A standard matrix multiplication) and looking up the corresponding coset
array (11) as shown below helps to mechanize this procedure: leader in the standard array. These lookup procedures be-
come difficult for codes with moderate to large values of block
  length n. This motivates the study of a subclass of linear
0 X1 X2 ··· Xi ··· X2 k −1
 block codes, namely, cyclic codes, with features that facilitate
 N1 ··· ··· ··· ··· ··· ···  
  more easily implemented decoders.
 N2 ··· ··· ··· ··· ··· ··· 
 
 ··· ··· ··· ··· ··· ··· ···  Cyclic Codes
 
 
 Nj ··· ··· ··· Yn ··· ···  A cyclic code is a linear block code with the special property
 
 ··· ··· ··· ··· ··· ··· ···  that every cyclic shift of a codeword is also a codeword. Cyclic
 
 ··· ··· ··· ··· ··· ··· ···  codes were first proposed by Prange (13). Polynomial algebra,
 
where binary vectors are represented by polynomials with bi-
N2 n−k −1 ··· ··· ··· ··· ··· ···
nary coefficients, is a useful framework for characterization
of cyclic codes. A binary n-vector Xn ⫽ (x1, x2, . . ., xn) has the
The top row of the standard array consists of the 2k code- polynomial representation X(D) ⫽ x1Dn⫺1 ⫹ x2Dn⫺2 ⫹ ⭈ ⭈ ⭈ ⫹
words. The first element, N1 in the next row is chosen to be xn, with degree not exceeding n ⫺ 1. For instance, (0101) cor-
an n-vector error pattern that the code is expected to correct. responds to X(D) ⫽ D2 ⫹ 1. Analogous to the generator matrix
It must not be one of the elements in the preceding row(s). of a linear block code, a cyclic code can be characterized in
The succeeding elements of this row are obtained by adding terms of a generator polynomial G(D) such that every
this error pattern to the corresponding codeword in the top codeword has a polynomial representation of the form X(D)
row. Additional rows are formed by repeating this procedure, ⫽ G(D)R(D). Here G(D) is a polynomial of degree n ⫺ k and
each time choosing the first element of the row to be a pattern G(D) is a factor of Dn ⫺ 1. The polynomial R(D) has degree
that has not appeared already in the rows above. Each row of k ⫺ 1 or less, representing the k-bit data vector (r1, . . ., rk)
the resulting 2n⫺k ⫻ 2k standard array is called a coset, and being encoded. A code polynomial is generated by multiplying
the first element in each row a coset leader. For a BSC with the data polynomial R(D) by the generator polynomial G(D).
It can be verified that multiplication of polynomials corre-
error probability ⑀ ⬍ , it is natural to choose the coset leaders
sponds to convolution of the corresponding vectors. This ob-
N to have the least Hamming weight possible. Given the stan-
servation leads to simple implementation of encoders using
dard array, the output of a BSC, Yn, is located in the standard
shift-register based digital circuits.
array. The codeword Xi at the top of the column that Yn be-
Denoting G(D) ⫽ g0 ⫹ g1D ⫹ g2D2 ⫹ ⭈ ⭈ ⭈ ⫹ gn⫺kDn⫺k, the
longs to is declared as the transmitted codeword, with the generator matrix G for the linear block code generated by
error pattern produced by the BSC being the coset leader Nj G(D) has the following form:
for the coset that Yn belongs to. If the BSC produces an error
 
pattern which is not one of the coset leaders, the decoder will g0 g1 g2 . . . 0 0
clearly make a decoding error. In the standard array for the 0 g g1 . . . . . 
 0 
(7,4) Hamming code, the coset leaders can be chosen to be the  
G=. . . ... . . 
set of all 7-bit patterns with Hamming weight 1. Hence this  
. . . ... gn−k 0 
code corrects all single error patterns and none else.
The matrix HT generates an (n, n ⫺ k) code (comprising all 0 0 0 . . . gn−k−1 gn−k
linear combinations of its n ⫺ k linearly independent rows).
As an example, binary cyclic codes of length n ⫽ 7 are gen-
The codes generated by G and HT are referred to as dual codes
erated by the factors of D7 ⫺ 1 ⫽ (D ⫹ 1)(D3 ⫹ D ⫹ 1)(D3 ⫹
of each other. The weight spectrum of a block code of length
D2 ⫹ 1). The first degree polynomial G1(D) ⫽ D ⫹ 1 generates
n is defined as the (n ⫹ 1)-vector (A0, . . ., An), where Ai is the (7, 6) code with a single overall parity bit, while G2(D) ⫽
the number of codewords with Hamming weight i. The Mac- D3 ⫹ D ⫹ 1 results in the (7, 4) Hamming code.
Williams identities (12) link the weight spectra of dual codes. The polynomial H(D) such that G(D)H(D) ⫽ Dn ⫺ 1 is
In particular, if k Ⰷ n ⫺ k, the weight spectrum of the (n, k) known as the parity check polynomial for the code generated
code with 2k codewords may be obtained more easily from the by G(D). [Since H(D) is also a factor of Dn ⫺ 1, it also can
weight spectrum of the dual code with only 2n⫺k codewords, by generate a cyclic code, which is the dual of the code generated
means of the MacWilliams identities. The weight spectrum of by G(D).] The polynomial H(D) ⫽ h0 ⫹ h1D ⫹ h2D2 ⫹ ⭈ ⭈ ⭈ ⫹
a linear block code determines the probability of undetected hkDk specifies the form of the parity check matrix H of the
error when the code is used over a BSC. Whenever the n- code as follows:
vector error pattern generated by the BSC coincides with one  
of the codewords, the error becomes undetectable by the code. hk hk−1 . . . 0
0 hk . . . h0 
This undetected-error probability is  
 
H=0 0 . . . h1 

 
n
. . . ... . . . . . .
PUDE = Ai i (1 − )n−i
i=0 0 0 . . . hk
Un is t ⫽ N ⫺ K symbol errors. Since the code can in particular

D D D
correct t consecutive symbol errors or erasures, it is especially
effective against burst errors. The Reed–Solomon codes are
gn–k gn–k–1 g1 g0
maximum-distance separable; that is, for the admissible
choices of n and k, the Reed–Solomon codewords are spaced
apart at the maximum possible Hamming distance.
+
Perfect Codes
Xn
Figure 3. A generic cyclic encoder. An (n, k) linear code can correct 2n⫺k error patterns. For some
integer t, if the set of error patterns consists of exactly all
error patterns of Hamming weight t or less and no other error
patterns at all, such a code is termed as a perfect t-error cor-
The special structure of G and H for cyclic codes greatly
recting code. This would require, for binary codes, that n, k,
simplifies the implementation of the encoder and the syn-
and t satisfy the following equality:
drome computer. A generic encoder for a cyclic code is shown
in Fig. 3. The k-bit data vector Uk is pipelined through the t

n
shift register for n clock times, thus generating the codeword = 2n−k
Xn at the output. The encoder utilizes n ⫺ k single-bit delay i=0
i
units D, binary multipliers, and binary adders. The circuit
complexity is seen to grow only linearly in block length n. For The only known perfect codes are the Hamming codes, the
decoding, we can develop a circuit for syndrome calculation double-error-correcting ternary Golay code, and the triple-er-
for a received n-vector, structured very similarly to Fig. 3. ror-correcting binary Golay code, described below. Tietvainen
Logic circuits are used to expand the n ⫺ k bit syndrome into (24) proved that no other perfect codes exist.
an n-bit error pattern which is then added to the received
codeword to effect error correction. Hamming Codes
Hamming codes are single error correcting codes. For t ⫽ 1,
BCH Codes
the condition for a perfect code becomes 1 ⫹ n ⫽ 2m where
Bose and Ray-Chaudhuri (14) and independently Hoquen- m ⫽ n ⫺ k. For integers m, Hamming single-error-correct-
ghem (15) discovered a remarkably powerful subclass of cyclic ing codes exist with m parity check bits and block length n ⫽
codes, referred to as BCH codes. The BCH code family is prac- 2m ⫺ 1. The rate of the (2m ⫺ 1, 2m ⫺ m ⫺ 1) code is R ⫽
tically the most powerful class of linear codes, especially for (2m ⫺ m ⫺ 1)/(2m ⫺ 1), which approaches 1 as m increases.
small to moderate block lengths. BCH codes can be designed The generator matrix G that was displayed earlier while de-
for a guaranteed design distance 웃 (which of course cannot scribing linear codes is indeed a generator matrix for a Ham-
exceed the true minimum distance dmin of the resulting code). ming (7, 4) code. It is possible to rearrange the generator ma-
Specifically, given 웃 and hence t ⫽ (웃 ⫺ 1)/2, and for any trix of the code so that the decimal value of the m-bit
integer m, there is a t-error correcting binary BCH code with syndrome word indicates the position of the (single) errored
block length n ⫽ 2m ⫺ 1 for which the number of parity check bit in the codeword. Adding an overall parity bit to a basic
bits is no more than mt. Powerful algorithms exist for decod- Hamming code results in a (2m, 2m ⫺ m ⫺ 1) code capable of
ing BCH codes. The polynomial algebraic approach to BCH detecting double errors in addition to correcting single errors.
decoding was pioneered by Peterson (16) for binary BCH Such codes are particularly effective in data transmission
codes and extended to nonbinary BCH codes by Gorenstein with automatic repeat request (ARQ). If the bit error rate is
and Zierler (17). Major contributions came later from Chien ⑀ ⬍ , then the single error patterns which appear with proba-
(18), Berlekamp (19), and Massey (20). An alternative ap- bility O(⑀) are the most common and are corrected by the
proach to BCH decoding based on finite field Fourier trans- code, thus avoiding the need for retransmission. The double
forms has gained attention recently, from the work of Blahut error patterns have a lower probability O(⑀2) and are flagged
(21). for retransmission requests. Other error patterns occur with
negligibly small probabilities O(⑀3) or less.
Reed–Solomon Codes Hamming codes are cyclic codes. For block lengths n ⫽
2m ⫺ 1, their generator polynomials are the primitive poly-
Reed–Solomon codes (22,23) are a very powerful generaliza-
nomials of degree m over the binary field. (An mth degree
tion of BCH codes. Binary Reed–Solomon codes can be de-
primitive polynomial with binary coefficients has the property
fined, for any integer m, as follows. Serial data is organized
that its m roots can be characterized as the m primitive ele-
into m-bit symbols. Each symbol can take one of n ⫽ 2m val-
ments in a finite field of 2m elements.) In fact, Hamming codes
ues. The Reed–Solomon code block length is N ⫽ 2m ⫺ 1 sym-
are BCH codes as well. However, they are decodable by far
bols, or Nm bits. Out of these, K symbols are data symbols
simpler methods than the general BCH decoding algorithms.
(i.e., k ⫽ mK), and N ⫺ K symbols are parity symbols com-
Most notably, Hamming codes are perfect codes.
puted according to the algebraic description of the code.
Reed–Solomon decoding can recover from up to t ⫽ (N ⫺
Golay Codes
K)/2 symbol errors. If symbol erasures are marked as such
(i.e., if additional side information is available as to whether Two codes discovered by Golay (25) are the only other perfect
a symbol is in error or not, though it is not known what the codes, apart from the Hamming codes mentioned above. For
errors are), then the Reed–Solomon erasure correction limit n ⫽ 23 and t ⫽ 3, the total number of 0-, 1-, 2-, and 3-error
binary patters of length 23 add up to 2048 ⫽ 211. Choosing weight choice of coset leaders (correctable error patterns) in
m ⫽ 11, we can form the (23, 12) triple-error-correcting Golay the standard array. In a burst noise channel, errored bit posi-
code. It is also possible to form an (11, 6) double-error-correct- tions tend to cluster together to form error bursts. Codes de-
ing perfect code over ternary alphabet. The Golay codes also signed to correct minimum weight error patterns are not di-
are BCH codes, and hence cyclic codes. rectly useful in presence of burst errors. Interleaved coding
is a technique that allows random-error-correcting codes to
Extended and Shortened Codes effectively combat burst errors. Interleaving renders the burst
noise patterns of the channel as apparent random errors to
Earlier we indicated that a single-error-correcting Hamming the decoder.
code can be made double-error-detecting as well by adding an Figure 4 depicts an elementary block interleaver to illus-
extra overall parity bit. This results in increasing the block trate the idea. Suppose a burst noise channel is known to
length by one. Such modified codes are known as extended generate error bursts spanning at most six consecutive bits.
codes. Adding an overall parity bit to a code of length n ⫽ Further, suppose that a (7, 4) Hamming single-error-correct-
2m ⫺ 1 results in a codeword whose length is a power of two. ing code is to be used. First we read 24 consecutive data bits
This may be advantageous in byte-oriented data handling, or row-wise into a 6 ⫻4 array, as in Fig. 4. The column size is
in matching a prespecified field length in a data packet. Ex- chosen to be six so as to at least equal the maximum burst
tended Reed–Solomon codes are another practical example. length. The row size is chosen to be exactly equal to the num-
As already seen, the natural block length of Reed–Solomon ber of information digits in the code. Next we extend each row
codes is 2m ⫺ 1 m-bit symbols, and frequently there are rea- by three more positions by appending the three parity bits
sons to have a block length which is a power of two. Adding appropriate for a Hamming (7, 4) codeword. The extended
an overall parity symbol accomplishes this task. Extended transmission array is therefore 6 ⫻ 7. The transmission se-
codes may have smaller minimum distance than their original quence is column-wise—that is, in the sequence 1, 5, 9, 13,
counterparts, but in many instances the minimum distance 17, 21, 2, 6, 10, 14, . . . (see Fig. 4). The array transmits 42
remains unchanged. An example is the (7, 4) Hamming code bits for each 24-bit input. A six-bit error burst may affect the
and its (8, 4) extended version, both of which have minimum successively transmitted bits 10, 14, 18, 22, 3 and 7, as shown
distance 4. in Fig. 4. Notice that each bit in this error burst belongs to a
Shortened codes result from the reverse process where we different Hamming codeword. Thus all such errors are cor-
seek to reduce the block length of a basic code. For example, rected if the error burst does not exceed six bits and there are
the Reed–Solomon code with 8-bit symbols has a natural no other burst error within this span of 42 bits. Suitably sized
block length of 255 symbols. If the encoded data is to be trans- block interleavers are often effective in burst noise environ-
ported in fixed length packets of 240 symbols each, we can set ments.
15 information symbols to zeroes and then delete them before
transmission. Shortening can increase minimum distance in Concatenated Codes
general. The shortened code has fewer information bits and
the same number of parity bits, so that the error correction Consider the following example. Let n ⫽ 8 and N ⫽ 2n ⫺ 1 ⫽
capability normalized with respect to block length increases 255. An (N, K) Reed–Solomon code has N ⫽ 255 code symbols
upon shortening. which can be represented as 8-bit binary words. Transmitted
through a binary symmetric channel with bit error probabil-
Product Codes ity ⑀, each 8-bit symbol can be in error with probability ⌬ ⫽
1 ⫺ (1 ⫺ ⑀)8. The code can recover from up to t0 ⫽ (N ⫺
Elias (26) showed how to combine two block codes into a prod- K)/2 symbol errors. The probability of successful decoding is,
uct code. Cyclic product codes were studied by Burton and therefore,
Weldon (27). Suppose we have an (n1, k1) code and an (n2, k2)
code. Arrange k ⫽ k1k2 data bits in an array of k1 rows and

t0
N
k2 columns. Extend each row of k2 bits into an n2 bit codeword Ps0 = i (1 − )255−i
i=0
i
using the (n2, k2) code. Next, extend each of the resulting n2
columns to n1 bits using (n1, k1) code. The resulting array of We can now concatenate the Reed–Solomon code with a
n ⫽ n1n2 bits is the product encoding for the original k bits. Hamming (8, 4) single-error correcting/double-error detecting
The rate of the product code is the product of the rates of the
constituent codes. If the constituent codes respectively have
minimum distances d1 and d2, the product code has a mini-
mum distance dmin ⫽ d1d2. Product codes are frequently capa- 1 2 3 4 P1 P2 P3
ble of correcting not only all error patterns of weight (d1d2 ⫺
5 6 7 8 P4 P5 P6
1)/2 but also many higher weight patterns. However, the
simplistic approach of row-wise decoding first, followed by col- 9 10 11 12 P7 P8 P9
umn-wise decoding, may not achieve the full error correction
capability of product codes. 13 14 15 16 ... ... ...
Interleaved Coding 17 18 19 20 ... ... ...
The binary symmetric channel models the random error pat- 21 22 23 24 ... ... ...
terns which are bit-to-bit independent. That the bit error
probability ⑀ is less than 0.5 is the basis for the minimum Figure 4. An illustration of interleaved coding.
code, as follows. Each 8-bit symbol of the Reed–Solomon code

is chosen as a Hamming (8, 4) codeword. To do this, we orga- 1
nize raw data into consecutive blocks of four bits and encode
each such into a Hamming (8, 4) codeword. Then each set of K
consecutive 8-bit Hamming codewords is encoded into a 255-
Hamming bound
Rate, R
symbol Reed–Solomon codeword. The probability of a symbol
error now becomes smaller, 웃 ⫽ 1 ⫺ (1 ⫺ ⑀)8 ⫺ 8 僆 (1 ⫺ ⑀)7,
Elias bound
assuming triple and higher errors are negligibly infrequent.
Besides, the double-error detection feature identifies the erro- Gilbert bound
red symbols in the Reed–Solomon code. With this side infor-
mation, the Reed–Solomon code can now recover from a
greater number of symbol errors, t1 ⫽ 255 ⫺ K. The probabil-
ity of successful decoding is now found as 0 0.5 1.0
t1 Normalized minimum distance, δ (R)

N Figure 5. Bounds on the minimum distance of block codes.
Ps1 = δ i (1 − δ)255−i
i=1
i
The power of this concatenated coding approach is evident This can be expressed in the following form of the Hamming
from comparing the above two expressions for the probability upper bound on code rate R:
of successful decoding. The Reed–Solomon code is the outer
t

1 n
code and the Hamming code is the inner code. The inner code R ≤ 1 − log
cleans up the milder error events and reserves the outer code n i=0
i
for the more severe error events. In particular, in a burst
noise environment, a long codeword may have some parts Asymptotically for large n, this reduces to the form
completely obliterated by a noise burst while other parts may δ(R)
be affected by occasional random errors. The inner code typi- R ≤ 1 − H2
cally corrects most of the random errors, and the outer Reed– 2
Solomon code combats the burst noise. In most applications
the outer code is a suitably sized Reed–Solomon code. The where H2(x) ⫽ ⫺x log x ⫺ (1 ⫺ x) log(1 ⫺ x) is the binary
inner code is often a convolutional code (discussed below), entropy function; 웃(R) is the largest value of the normalized
minimum distance of a rate-R code as the block length n goes
though block codes can be used as well, as shown above.
to vary large values (i.e., lim supN씮앝 dmin /n); and the logarithm
The invention of concatenated coding by Forney (28) was a
is to the base 2 so that R is in bits of information per binary
major landmark in coding theory. Later, Justesen (29) used
digit.
the concatenation concept to obtain the first constructive
The Hamming upper bound asserts that for a given 웃(R),
codes with rates that do not vanish asymptotically for large
no code can exceed the rate given by the bound above. The
block lengths.
Gilbert bound, on the other hand, is a constructive bound
which states that it is possible, for a given 웃(R), to construct
Performance Limits of Block Codes codes with rates at least as large as the value R specified by
the bound. The asymptotic Gilbert bound states that
The key performance parameters of a block code are the code
rate and the minimum distance. In this section we highlight R ≥ 1 − H2 (δ(R))
some of the known bounds on these parameters. The Ham-
ming bound, also known as the sphere packing bound, is a The Elias bound is a tighter upper bound on the feasible code
direct consequence of the following geometrical view of the rates compared to the Hamming bound. In its asymptotic
code space. Let an (n, k) code have minimum distance d. form the Elias bound is stated as follows:
There are 2k codewords in this code. Around each codeword
we can visualize a ‘‘sphere’’ comprising all n-vectors that are δ(R) ≤ 2λR (1 − λR )
within Hamming distance (d ⫺ 1)/2 from that codeword.
Each such sphere consists of all the n-tuples that result from where
perturbations of the codeword at the center of the sphere by
Hamming weight at most (d ⫺ 1)/2. Any two such spheres R = 1 − H2 (λR )
around two distinct codewords must be mutually exclusive if
unambiguous minimum-distance decoding is to be feasible. These bounds are shown in Fig. 5. The feasibility region of
Thus the total ‘‘volume’’ of all 2k such mutually exclusive ‘‘good’’ block codes lies between the Gilbert and Elias bounds.
spheres must not exceed the total number of possible n- Hamming bound originally appeared in Ref. 10, and the Gil-
tuples, 2n. Thus, bert bound in Ref. 30. The Elias bound was first developed by
Elias circa 1959 but appeared in print only in 1967 paper by
t
Shannon, Gallager, and Berlekamp (see Ref. 5, p. 3). Proofs
n
2 ≥2
n k
for these bounds are found in many coding theory books (e.g.,
i=0
i Ref. 3). It had been conjectured for some time that the Gilbert
bound was asymptotically tight—that is, that it was an upper output (K ⫽ 3) or the minimum number of delay elements
bound as well as a lower bound and that all long, good codes needed to implement the encoder (K ⫽ 2).
would asymptotically meet the Gilbert bound exactly. This It must be noted that Fig. 6 could have been redrawn with
perception was disproved for nonbinary codes by the work of only two memory elements to store the two previous bits; the
Tsfasman et al. (31). Also McEliece et al. (32) obtained some current input bit could be residing on the input line. The
improvements on the Elias bound. See also Ref. 33 for tabula- memory order of the encoder in Fig. 6 is thus only two, and
tions of the best-known minimum distances of block codes. the encoder output is determined by the state of the encoder
(which is the content of the two memory registers) and by the
new input bit. Whether we use an extra memory register to
CONVOLUTIONAL CODES hold the incoming new bit or not is similar in spirit to the
distinction between the Moore and the Mealy machines in the
Convolutional Encoders theory of finite-state sequential machines (39).
The impulse response of the encoder at the upper output
Convolutional codes were originally proposed by Elias (34).
of the convolutional encoder in Fig. 6 is ‘1 1 1’ and that at the
Probabilistic search algorithms were developed by Fano (35)
lower output line is ‘1 0 1’. The output sequences at these
and Wozencraft and Reiffan (36) as practical decoding algo-
lines are therefore the discrete convolutions of the input
rithms. Massey (37) proposed the use of threshold decoding
stream with these impulse responses. The following infinite-
for convolutional codes as a simpler though less efficient al-
dimensional generator matrix represents the mapping of the
ternative. The Viterbi algorithm (38) was developed later as
infinite input sequence (x0, x1, x2, . . .) into the infinite output
an efficient decoder for short convolutional codes. We will
sequence (y0, y1, y2, . . .) where y2n and y2n⫹1 are the two output
briefly outline Wozencraft’s sequential decoding and the Vit-
bits corresponding to input bit xn:
erbi algorithm, after examining the basic structure of convo-
lutional codes. 
Convolutional coding is based on the notion of passing an 1 1 1 0 1 1 0 0
0
arbitrarily long sequence of input data bits through a linear  0 1 1 1 0 1 1

sequential machine whose output sequence has memory prop- 0 0 0 0 1 1 1 0

erties and consequent redundancies that allow error correc- 0 0 0 0 0 0 1 1

tion. A linear sequential machine produces each output sym- 
x0 x1 x2 . . .]  0 0 0 0 0 0 0 0
bol as a linear function of the current input and a given 
0 0 0 0 0 0 0 0
number of the immediate past inputs, so that the output sym- 
0
bols have ‘‘memory’’ or temporal correlation. Certain symbol  0 0 0 0 0 0 0

patterns are more likely than others, and this allows error . . . ... ... ... ... ... ... ...
correction based on maximum likelihood principles. The out- ... ... ... ... ... ... ... ...
put of the linear sequential machine is the convolution of its 
0 0 0 0 0 0 ...
impulse reponse with the input bit stream, hence the name.
0 0 0 0 0 0 . . .

Block codes and convolutional codes are traditionally viewed 
as the two major classes error correction codes, although we 1 1 0 0 0 0 . . .

will recognize shortly that it is possible to characterize finite 1 0 1 1 0 0 . . .


length convolutional codes in a formalism similar to that used 1 1 1 0 1 1 . . . = [ y0 y1 y2 . . .

to describe block codes. 0 0 1 1 1 0 . . .

. . .
A simple convolutional encoder is shown in Fig. 6. For ev-
0 0 0 0 1 1 
ery input bit, the encoder produces two output bits. The code 
rate is hence . (More generally, a convolutional encoder may ... ... ... ... ... ... . . .
accept k input bits at a time and produce n output bits, imple- ... ... ... ... ... ... ...
menting a rate k/n code.) The output of the encoder in Fig. 6
is a function of the current input and the two previous inputs. Also, in terms of the impulse response polynomials G1(D) ⫽
One input bit is seen to affect three successive pairs of output 1 ⫹ D ⫹ D2 and G2(D) ⫽ 1 ⫹ D2, respectively, for the upper
bits. We say that the constraint length of the code is therefore and lower output lines in Fig. 6, we can relate the input poly-
K ⫽ 6. There are other definitions of the constraint length, nomial X(D) to the respective output polynomials as
as the number of consecutive input bits that affect a given
Yi (D) = X (D)Gi (D), i = 1, 2
Output 1 However, these matrix and polynomial algebraic approaches

+ are not as productive here as they were for the block codes.
More intuitive insight into the nature of convolutional codes
can be furnished in terms of its tree and trellis diagrams.
Input D D
Trees, State Diagrams, and Trellises

+
The most illuminating representation of a convolutional code
Output 2
is in terms of the associated tree diagram. The encoding pro-
Figure 6. A convolutional encoder. cess starts at the root node of a binary tree, as shown in Fig.
00 Data sequences drive the encoder through various se-

00 quences of state transitions. The pattern of all such possible
1
11 state transition trajectories in time is known as a trellis dia-
00
1 10 gram. In Fig. 9 we have the trellis diagram for an encoder
11
2 that starts in state 00 and encodes a 7-bit input sequence
01 whose last two bits are constrained to be zeroes. This con-
00
1 11 straint, useful in Viterbi decoding to be described below, ter-
10
3 minates all paths in state 00. The trellis diagram in Fig. 9
00 thus contains 25 distinct paths of length 7 beginning and end-
11
2 01 ing in state 00.
01
0 4
10 Weight Distribution for Convolutional Codes
1 00 An elegant method for finding the weight distribution of con-
11
1 1 volutional codes is to redraw the state transition diagram
11
10 such as in Fig. 8, in the form shown in Fig. 10 with the all-
3 10 zero state (00 in our example) split into two, a starting node
00 and an ending node. To each directed path between two
2
01
11 states, we assign a ‘‘gain’’ Wi, where W is a dummy variable
2 11 and the exponent i is the Hamming weight of the binary se-
1 = 00 01 quence emitted by the encoder upon making the indicated
3
00
2 = 01 01 state transition. For example, in Fig. 10, the transition from
4 01 1 to 2 causes the bit pair 11 to be emitted, with Hamming
3 = 10
10
4 weight i ⫽ 2, so that the gain is W2. In transitions that emit
4 = 11 10 01 or 10, the gain is W and in the case where 00 is emitted,
the gain is W0 ⫽ 1. We can now use a ‘‘signal flow graph’’
Figure 7. A tree diagram for the convolutional encoder in Fig. 6.
technique due to Mason (40) to obtain a certain ‘‘transfer
function’’ of the encoder. In the signal flow graph method, we
postulate an input signal Sin at the starting state and com-
7 for the encoder in Fig. 6. Each node spawns two branches. pute the output signal Sout at the ending state, using the fol-
Each successive input bit causes the process to move to one lowing relations among signal flow intensities at the various
of the next higher level nodes. If the input bit is a zero, the nodes:
upper branch is taken, otherwise the lower one. The labeling
on each branch shows the bit pair produced at the output for Sout = S2W 2
each branch. Tracing an input sequence through the tree, the S2 = (S3 + S4 )W
concatenation of the branch labels for that path produces the
S4 = (S3 + S4 )W
corresponding codeword.
Careful inspection of the tree diagram in Fig. 7 reveals a S3 = SinW 2
certain repetitive structure depending on the ‘‘state’’ of the
encoder at each tree node. The branching patterns from any The transfer function T(W) ⫽ Sout /Sin can be readily found to
two nodes with identical states are seen to be identical. This be
allows us to represent the encoder behavior most succinctly
in terms of a state transition diagram in Fig. 8. The state of W5 ∞
T (W ) = = 2iW 5+i = W 5 + 2W 6 + 4W 7 + · · ·
the encoder is defined as the contents of the memory elements (1 − 2W ) i=0
at any time. The encoder in Fig. 7 has four states, 1 ⫽ 00,
2 ⫽ 01, 3 ⫽ 10 and 4 ⫽ 11. The solid lines in Fig. 8 indicate Each term in the above series corresponds to a set of paths
state transitions caused by a zero input, and the dotted lines of a given weight. The coefficient 2i gives the number of paths
indicate input one. The labels on the branches are the output of weight 5 ⫹ i. There is exactly one path of weight 5, two
bit pairs, as in the tree diagram in Fig. 7. paths of weight 6, four of weight 7, and so on. There are no
paths of weight less than 5. The path with weight 5 is seen
to be the closest in Hamming distance to the all-zero
codeword. This distance is called the free distance, dfree, of the
10
code. In the present example, dfree ⫽ 5. The free distance of a
11 convolutional code is a key parameter in defining its error
01 01 correction, as will be seen in the next section.
00
01 10
10 Maximum Likelihood (Viterbi) Decoding
11 11 for Convolutional Codes
00
00 Each path in a trellis diagram corresponds to a valid code
sequence in a convolutional code. A received sequence with
Figure 8. The state transition diagram for the convolutional encoder bit errors in it will not necessarily correspond exactly to any
in Fig. 6. one particular trellis path. The Viterbi algorithm (38) is a
00 00 00 00 00 00 00
1 1 1 1 1 1 1 1
11 11 11 11 11
2 2 2 2 2
01 01 01 01 01
10 10 10 10
11 11 11 11 11
00 00 00
3 3 3 3 3
10 10 10 10
4 4 4 4 Figure 9. The trellis diagram for the con-

01 01 01 volutional encoder in Fig. 6.
computationally efficient way for discovering the most likely tal, for any path emanating from state 1 at time t ⫽ 3, the
transmitted sequence for any given received sequence of bits. prefix with the lower cumulative distance is clearly the better
With reference to the trellis diagram in Fig. 9, suppose that choice. Thus at this point we discard the path 1–1–1–1 from
we have received the sequence 11 01 10 01 01 10 11. Starting further consideration and retain the unique survivor path
in state 1 at time t ⫽ 0, the trellis branches out to states 1 or 1–3–2–1 in association with state 1 at the current time. Simi-
2 in time t ⫽ 1, and from there to all four states 1, 2, 3, 4, in larly we explore the two contending paths converging at the
time t ⫽ 2. At this point there is exactly one unique path to other three states at this time (t ⫽ 3) and identify the mini-
each of the four current possible states staring from state 1. mum distance (or maximum likelihood, for the BSC) survivors
In order to reach state 1 at time t ⫽ 2, the trellis path indi- for each of those states.
cates the transmitted prefix sequence 00 00 which is at a The procedure now iterates. At each successive stage, we
Hamming distance three from the actual received prefix 11 identify the survivor paths for each state. If the code sequence
01. The path reaching state 2 in time t ⫽ 2 in the trellis dia- were infinite, we would have four infinitely long parallel path
gram similarly corresponds to the transmitted prefix se- traces through the trellis in our example. In order to choose
quence 11 01 which is seen to be at Hamming distance zero one of the four as the final decoded sequence, we require the
from the corresponding prefix of the received sequence. Simi- encoder to ‘‘flush out’’ the data with a sequence of zeroes, two
larly we can associate Hamming distances 3 and 2 respec- in our example. The last two zeroes in the seven-bit input
tively to the paths reaching states 3 and 4 in time t ⫽ 2 in data to the encoder cause the trellis paths to converge to state
the trellis diagram. 1 or 2 at time t ⫽ 6 and to state 1 at t ⫽ 7. By choosing the
Now we extend the trellis paths to time t ⫽ 3. Each state survivors at these states, we finally have a complete trellis
can be reached at time t ⫽ 3 along two distinct paths. For path starting from state 1 at time t ⫽ 0 and ending in state
instance, in order to reach state 1 in time t ⫽ 3, the encoder 1 at time t ⫽ 7. The output labels of the successive branches
could have made a 1 to 1 transition, adding an incremental along this path gives the decoder’s maximum likelihood esti-
Hamming distance of one to the previous cumulative value of mate of the transmitted bits corresponding to the received se-
three; or it could have made the 2 to 1 transition, adding one quence.
unit of Hamming weight to the previous value of zero. Thus The average bit error rate of the Viterbi decoder, Pb, can
at time t ⫽ 3, there are two distinct paths merging at state be shown to be bounded by an exponential function of the free
1: the state sequence 1–1–1–1 with a cumulative Hamming distance of the code, as below:
distance of four from the given received sequence, and the
√ √
sequence 1–3–2–1 with a cumulative Hamming distance of Pb ≈ Nd [2 (1 − )]d free ≈ Nd [2 ]d free
one. Since the Hamming weights of the paths are incremen- free free
This applies to codes that accept one input bit at a time, as

W in Fig. 6. Ndfree is the total number of nonzero information bits
on all trellis paths of weight dfree, and it can in general be
found via an extension of the signal flow transfer function
4
method outlined above. The parameter ⑀ is the BSC error
W W
1 3 2 1
probability and is assumed to be very small in the above ap-
proximation.
2 W2
W W The Viterbi algorithm needs to keep track of only one sur-
vivor path per state. The number of states, however, is an
Figure 10. The signal flow graph for the convolutional encoder in exponential function of the memory order. For short convolu-
Fig. 6. tional codes of modestly sized state space, the Viterbi algo-
rithm is an excellent choice for decoder implementation. A mum distance decoding. Burst error channels are another im-
memory order of 7 or 8 is typically the maximum feasible. portant class of transmission channels encountered in
This, in turn, limits the free distance and hence the bit error practice, both in wireline and wireless links. Errored bit posi-
probability. For long convolutional codes, the survivor path tions tend to cluster together in such channels, making direct
information storage required per state becomes large. In prac- application of much of the foregoing error correction codes fu-
tice, we may choose to retain only some most recent segment tile in such cases. We have already mentioned interleaved
of the history of each survivor path. The resulting ‘‘truncated’’ coding as a practical method for breaking the error clusters
Viterbi algorithm is no longer the theoretically ideal maxi- into random patterns and then using random-error correcting
mum likelihood decoder, though its performance is usually codes. Also we noted that Reed–Solomon codes have an in-
close to the ideal decoder. All these considerations restrict the trinsic burst error correction capability. In addition, there
application of the Viterbi algorithm to short convolutional have been error correction codes specifically developed for
codes with small constraint lengths. Within these limitations, burst noise channels. For a detailed treatment of this subject,
however, the Viterbi algorithm affords excellent performance. see, for example, Ref. 8, Chap. 9.
Sequential Decoding for Convolutional Codes Intersymbol Interference Channels and Precoding
Tree diagrams lead to one viable strategy for decoding convo- Binary data are transmitted by mapping the 0’s and 1’s into
lutional codes. Given a received sequence of bits (possibly con- baseband or radio-frequency (RF) pulses. In a bandwidth-lim-
taining errors), the decoder attempts to map it to one path ited channel, the channel response waveform corresponding
along the tree, proceeding node by node and keeping track of to one input pulse tends to overlap those of succeeding pulses,
the cumulative Hamming distance of the path from the re- if the input pulse rate is high. This intersymbol interference
ceived sequence. Along a wrong path, the cumulative Ham- (ISI) can be controlled by appropriately shaping the input
ming distance exceeds a preset threshold after a few nodes, spectrum by precoding the input pulse waveform. By suitably
whereupon the decoder backtracks to the previous node and constraining the 0/1 transition patterns, it becomes possible
explores another path. The time to decode any given sequence to receive the input bit stream despite the overlap of the pulse
in this scheme is a random variable, but its expected value response waveforms. This technique has been important in
remains bounded for code rates below a number Rcomp ⬍ C, high-speed modem designs for the wireline channel. Because
where Rcomp is the computational cutoff rate and C is the chan- of the recent interest in digital subscriber lines, there has
nel capacity. This technique, known as sequential decoding, been much activity in this area. We cite Ref. 41 as an example
is an appropriate technique for decoding very long convolu- of recent work and for pointers to earlier work in this impor-
tional codes. tant area.
A sequential decoder executes a random number of compu-
tations to decode a received sequence—unlike the Viterbi de-
Synchronization Codes
coder, which executes a fixed number of computations per
code sequence. This can be a strength or a weakness, de- Coding techniques described so far implicitly assume synchro-
pending on the average noise intensity. If the noise level is nization; that is, the decoder knows the exact times when one
high, the sequential decoder typically has to explore many codeword ends and the next begins, in a stream of binary
false paths before it discovers the correct path. But the Vit- data. In real life this of course cannot be assumed. Codes that
erbi algorithm produces an output after a fixed number of can self-synchronize are therefore important. Key results in
computations, possibly faster than the sequential decoder. On this direction is summarized in standard coding theory
the other hand, if the noise level is low, the Viterbi algorithm sources such as Ref. 4. However, the practical use of these
still needs to execute all of its fixed set of computations synchronization codes does appear to be limited, compared to
whereas the sequential decoder will typically land on the more advanced timing and synchronization techniques used
right tree path after only a few trials. Also, sequential decod- in modern digital networks.
ing is preferred in applications where long codes are needed
to drive the postdecoding error probability to extremely low Soft Decision Decoding
values. In such cases, complexity considerations eliminate
Viterbi algorithm as a viable choice. The actual inputs and outputs of the physical channel are
An efficient approach to implementing sequential decoding analog waveforms. The demodulator processes the noisy
is the stack algorithm. The key idea here is that the pre- waveform output of the physical channel and furnishes a
viously explored paths and their likelihood metrics can be noisy estimate of the currently transmitted bit or symbol. A
stored in a stack ordered according to the likelihood metric hard decision is made at the demodulator output when a
value, with the most likely path at the top. The topmost path threshold device maps the noisy analog data into a 0 or a 1
is then extended to the set of branches extending from that (in the binary case). Instead, we can retain the analog value
node, metrics are recomputed, and the stack is updated with (or a finely quantized version of it) and then make an overall
the new information. decision about the identity of an entire codeword from these
soft decision data. Clearly, the soft decision data retain more
information, and hence the overall decision made on an entire
ADDITIONAL TOPICS
codeword can be expected to be more reliable than the concat-
enation of bit-by-bit hard decisions. Analysis and practical im-
Burst Noise Channels
plementations have borne out this expectation, and soft deci-
In the foregoing discussion, we had mostly assumed the bi- sion decoding enables achievement of the same bit error rate
nary symmetric channel model which was the basis for mini- with a lower signaling power requirement than that for hard
decision decoding. Many recent text books on digital commu- standards use Reed–Solomon coding for delivery of com-
nications (e.g., Ref. 42) contain details of this approach. pressed, high-rate digital video (48). Almost all of the recent
digital wireless technologies, such as GSM, IS-54 TDMA, IS-
Combined Coding and Modulation 95 CDMA, cellular digital packet data (CDPD), and others
(49), have found it advantageous to make use of error correc-
As mentioned earlier, coding and modulation have tradition- tion coding to mitigate the excessive noisiness of the wire-
ally developed in mutual isolation. Ungerboeck (43) proposed less channel.
the idea that redundancy for error correction coding may be In summary, over the past 50 years following the inception
embedded into the design of modulation signal constellations, of information theory (1), not only has the art of data trans-
and combined decoding decisions may be based on the Euclid- mission codes matured into a variety of applications technolo-
ean distance between encoded signal points rather than on gies, but also we are remarkably close to the ultimate theoret-
Hamming distance. The approach has been found to be capa- ical limits of coding performance predicted in Ref. 1.
ble of significant improvements in the performance of coded
communications. For more details on this topic, see Ref. 44 or
one of the more recent textbooks in digital communications BIBLIOGRAPHY
such as Ref. 42.
1. C. E. Shannon, A mathematical theory of communications, Bell
Turbo Codes Syst. Tech. J., 27: 379–423, 623–656, 1948.
2. R. G. Gallager, Information Theory and Reliable Communication,
The discovery of ‘‘turbo codes’’ (45) is perhaps the most spec- New York: Wiley, 1968.
tacular event in coding research in recent times. Turbo codes
3. E. R. Berlekamp, Algebraic Coding Theory, New York: McGraw-
have made it possible to approach the ultimate Shannon lim- Hill, 1989.
its for communication much more closely than was previously
4. W. W. Peterson and E. J. Weldon, Jr., Error Correcting Codes,
possible. Turbo codes are essentially a battery of parallel con- 2nd ed., Cambridge, MA: MIT Press, 1972.
catenated encoders. The outputs of these component encoders
5. E. R. Berlekamp, Key Papers in the Development of Coding Theory,
are merged by interleaving, and they are punctured as needed
New York: IEEE Press, 1974.
to get the desired code rate. Relatively simple, iterative soft
6. F. J. MacWilliams and N. J. A. Sloane, The Theory of Error Cor-
decision decoding methods provide surprisingly superior per-
recting Codes, Amsterdam, The Netherlands: North-Holland,
formance. 1977.
7. R. E. Blahut, Theory and Practice of Error Control Codes, Reading,
APPLICATIONS MA: Addison-Wesley, 1983.
8. S. Lin and D. J. Costello, Jr., Error Control Coding: Fundamen-
Data transmission coding is intrinsically an applications-ori- tals and Applications, Englewood Cliffs, NJ: Prentice-Hall, 1983.
ented discipline, its deep mathematical foundations notwith- 9. A. M. Michelson and A. H. Levesque, Error Control Techniques
standing. Early work by Hamming (10) and others were for Digital Communications, New York: Wiley, 1985.
quickly applied to computer data storage and transmission. 10. R. W. Hamming, Error detecting and error correcting codes, Bell
The minimum-distance decoding approach is ideally suited ro Syst. Tech. J., 29: 147–160, 1950.
random, independent-error environments such as found in 11. D. E. Slepian, A class of binary signaling alphabets, Bell Syst.
space communications, where coding applications registered Tech. J., 35: 203–234, 1956.
some early success (e.g., see Chapter 3 of Ref. 23). The mark- 12. F. J. MacWilliams, A theorem on the distribution of weights in a
edly clustered or bursty nature of error patterns in terrestrial systematic code, Bell Syst. Tech. J., 42: 79–94, 1963.
wireline and radio channels initially limited coding applica- 13. E. Prange, Cyclic error-correcting codes in two symbols, AFCRC-
tions in this arena to only error detection and retransmission TN-57-103, Air Force Cambridge Research Center, Cambridge,
(8, Chapter 15), and not forward error correction. Cyclic re- MA, 1957.
dundancy check (CRC) codes are cyclic codes used for error 14. R. C. Bose and D. K. Ray-Chaudhuri, On a class of error correct-
detection only, and they are ubiquitous in modern data trans- ing binary group codes, Inf. Control, 3: 68–79, 1960.
mission protocols. Also, more recently, the needs of the high- 15. A. Hocquenghem, Codes correcteurs d’erreurs, Chiffres, 2: 147–
speed modems created more opportunities for error correction 156, 1959, in French.
applications. Many specific codes have lately been adopted 16. W. W. Peterson, Encoding and decoding procedures for the Bose–
into international standards for data transmission. Chapters Chaudhuri codes, IRE Trans. Inf. Theory, 6: 459–470, 1960.
16 and 17 in Ref. 8 furnish an excellent summary of applica- 17. D. C. Gorenstein and N. Zierler, A class of error-correcting codes
tions of block and convolutional codes, respectively. Applica- in pm symbols, J. Soc. Ind. Appl. Math (SIAM), 9: 207–214, 1961.
tions of Reed–Solomon codes to various situations are well 18. R. T. Chien, Cyclic decoding procedure for the BCH codes, IEEE
documented in Ref. 23, including its use in compact disc play- Trans. Inf. Theory, 10: 357–363, 1964.
ers, deep space communications, and frequency hop spread 19. E. R. Berlekamp, On decoding binary BCH codes, IEEE Trans.
spectrum packet communications. Inf. Theory, 11: 577–580, 1965.
The advent of the broadband integrated service digital net- 20. J. L. Massey, Shift register synthesis and BCH decoding, IEEE
work (B-ISDN) has created ample instances of coding applica- Trans. Inf. Theory, 15: 122–127, 1969.
tions (46). The asynchronous transfer mode (ATM) adaptation 21. R. E. Blahut, Transform techniques for error control codes, IBM
layer 1 (AAL-1) has a standardized option for the use of J. Res. Develop., 23: 299–315, 1979.
Reed–Solomon coding for recovery from ATM cell losses (47, 22. I. S. Reed and G. Solomon, Polynomial codes over certain finite
p. 75). The recently adopted high-definition television (HDTV) fields, J. Soc. Ind. Appl. Math. (SIAM), 8: 300–304, 1960.
156 INFORMATION THEORY OF MODULATION CODES AND WAVEFORMS
23. S. B. Wicker and V. K. Bhargava, Reed–Solomon Codes and Their

Applications, Piscataway, NJ: IEEE Press, 1994.
24. A. Tietvainen, A short proof for the nonexistence of unknown per-
fect codes over GF(q), q ⬎ 2, Ann. Acad. Sci. Fenn. A, 580: 1–6,
1974.
25. M. J. E. Golay, Binary coding, IRE Trans. Inf. Theory, 4: 23–28,
1954.
26. P. Elias, Error-free coding, IRE Trans. Inf. Theory, 4: 29–37,
1954.
27. H. O. Burton and E. J. Weldon, Jr., Cyclic product codes, IRE
Trans. Inf. Theory, 11: 433–440, 1965.
28. G. D. Forney, Jr., Concatenated Codes, Cambridge, MA: MIT
Press, 1966.
29. J. Justesen, A class of constructive asymptotically algebraic
codes, IEEE Trans. Inf. Theory, 18: 652–656, 1972.
30. E. N. Gilbert, A comparison of signaling alphabets, Bell Syst.
Tech. J., 31: 504–522, 1952.
31. M. A. Tsfasman, S. G. Vladut, and T. Zink, Modular curves, Shi-
mura curves and Goppa codes which are better than the Varshar-
mov–Gilbert bound, Math. Nachr., 109: 21–28, 1982.
32. R. J. McEliece et al., New upper bounds on the rate of a code via
the Delsarte–MacWilliams inequalities, IEEE Trans. Inf. Theory,
23: 157–166, 1977.
33. T. Verhoeff, An updated table of minimum-distance bounds for
binary linear codes, IEEE Trans. Inf. Theory, 33: 665–680, 1987.
34. P. Elias, Coding for noisy channels, IRE Conv. Rec., 4: 37–47, 1955.
35. R. M. Fano, A heuristic discussion of probabilistic decoding, IEEE
Trans. Inf. Theory, 9: 64–74, 1963.
36. J. M. Wozencraft and B. Reiffan, Sequential Decoding, Cam-
37. J. L. Massey, Threshold Decoding, Cambridge, MA: MIT Press,
1963.
38. A. J. Viterbi, Error bounds for convolutional codes and an asymp-
totically optimum decoding algorithm, IEEE Trans. Inf. Theory,
13: 260–269, 1967.
39. G. D. Forney, Jr., Convolutional codes I: Algebraic structure,
IEEE Trans. Inf. Theory, 16: 720–738, 1970.
40. S. J. Mason, Feedback theory—Further properties of signal flow
graphs, Proc. IRE, 44: 920–926, 1956.
41. R. F. H. Fischer and J. B. Huber, Comparison of precoding
schemes for digital subscriber lines, IEEE Trans. Commun., 45:
334–343, 1997.
42. S. G. Wilson, Digital Modulation and Coding, Upper Saddle
River, NJ: Prentice-Hall, 1996.
43. G. Ungerboeck, Channel coding with amplitude/phase modula-
tion, IEEE Trans. Inf. Theory, 28: 55–67, 1982.
44. G. Ungerboeck, IEEE Commun. Mag., 25 (2): 12–21, 1987.
45. C. Berrou and A. Glavieux, Near-optimum error correcting coding
and decoding: Turbo codes, IEEE Trans. Commun., 44: 1261–
1271, 1996.
46. E. Ayanoglu, R. D. Gitlin, and N. C. Oguz, Performance improve-
ment in broadband networks using a forward error correction for
lost packet recovery, J. High Speed Netw., 2: 287–304, 1993.
47. C. Partridge, Gigabit Networks, Reading, MA: Addison-Wesley,
1994.
48. T. S. Rzeszewski, Digital Video: Concepts and Applications across
Industries, Piscataway, NJ: IEEE Press, 1995.
49. T. S. Rappaport, Wireless Communications: Principles and Prac-
tice, Upper Saddle River, NJ: Prentice-Hall, 1996.
GEORGE THOMAS
University of Southwestern
Louisiana
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Information Theory of Modulation Codes and Waveforms ● Email this page Enter words or
● Print this page
Costas N. Georghiades1
1Texas A&M University
reserved. Search
DOI: 10.1002/047134608X.W4207.pub2 ❍ Search All Content
Article Online Posting Date: August 17, 2007 ❍ Acronym Finder
Abstract
The fundamental problem of communication is the conveying of information (which may take several different forms) from a
generating source through a communication medium to a desired destination. This conveyance of information, invariably, is
achieved by transmitting signals that contain the desired information in some form and that efficiently carry the information
through the communication medium. We refer to the process of superimposing an information signal onto another for efficient
transmission as modulation.
Introduction
Analog Modulation
Digital Modulation
Keywords: analog modulation; digitally modulated signal; binary modulation; baseband pulse-amplitude modulation; quadrature
amplitude modulation; continuous-phase modulation

INFORMATION THEORY OF MODULATION Currently, a proliferation of products make use of modu-
CODES AND WAVEFORMS lation to transmit information efficiently. Perhaps the most
prevalent and oldest examples are commercial broadcast
stations that use frequency modulation (FM) or ampli-
tude modulation (AM) to transmit audio signals through
INTRODUCTION
the atmosphere. Another example are data modems that
are used to transmit and receive data through telephone
The fundamental problem of communication is the con-
lines. These two examples have obvious similarities but
veying of information (which may take several different
also some very important differences. In the broadcast sta-
forms) from a generating source through a communication
tion example, the information to be communicated (an au-
medium to a desired destination. This conveyance of in-
dio signal) is analog and is used to directly modulate a
formation, invariably, is achieved by transmitting signals
radio-frequency (RF) carrier, which is an example of ana-
that contain the desired information in some form and that
log modulation. On the other hand, the data communicated
efficiently carry the information through the communica-
through a modem come from the serial port of a computer
tion medium. We refer to the process of superimposing an
and are discrete (in fact they are binary; that is, they take
information signal onto another for efficient transmission
two possible values, “0” or “1”), which results in a digitally
as modulation.
modulated signal. Clearly, the difference between analog
Several factors dictate modulating the desired informa-
and digital modulation is not in the nature of the trans-
tion signal into another signal more suitable for transmis-
mitted signals, because the modulation signals are analog
sion. The following factors affect the choice of modulation
in both cases. Rather, the difference is in the nature of the
signals:
set of possible modulation signals, which is discrete (and
in fact finite) for digitally modulated signals and infinitely
1. The need to use signals that efficiently propagate
uncountable for analog modulation.
through the communication medium at hand. For ex-
The simplest possible digital modulation system con-
ample, if the communication medium is the atmo-
sists of two modulation signals. One signal corresponds to
sphere (or free space), one might use a radio fre-
the transmission of a “0” and the other of a “1,” which is
quency (RF) signal at some appropriate frequency,
called binary modulation. Binary digits (bits) are commu-
whereas for underwater communications, one might
nicated using binary modulation by assigning a signal in a
use an acoustical signal.
one-to-one correspondence to each of the two possible logi-
2. Communication media invariably distort stochas- cal values of a bit. This mapping between bits and signals
tically signals transmitted through them, which is done at a rate equal to the bit rate (i.e., the number of
makes information extraction at the receiver nonper- bits/second arriving at the input of the modulator). In re-
fect and most often nonperfect. Thus, a need exists sponse to each transmitted modulation signal, the channel
to design modulation signals that are robust to the produces a received signal at its output, which is a ran-
stochastic (and other) effects of the channel, to mini- domly distorted replica of the transmitted signal. To ex-
mize its deleterious effects on information extraction. tract the information superimposed on the modulation sig-
3. It is highly desirable that communication systems nals, a processor, called a receiver or a detector, processes
convey large amounts of information per unit time. the noisy signal received. The function of the detector is
The price we pay in increasing the in formation rate is to decide which of the two (in this case) possible signals
often an increase in the required transmitted signal was transmitted, and in doing so correctly, it recovers the
bandwidth. We are interested in modulation signals correct value for the transmitted bit. Because of the pres-
that can accommodate large information rates at as ence of stochastic noise in the received signal, the receiver
small a required bandwidth as possible. may make an incorrect decision for some transmitted bits.
4. The power requirements (i.e., average power and The probability of making a decision error in extracting
peak power) of the transmitted signals to achieve a the transmitted bits is known as the bit-error probability
certain level of performance in the presence of noise or the bit-error rate (BER). The performance of communi-
introduced during transmission are of paramount cation systems using digital modulation is invariably mea-
importance, especially in power-limited scenarios, sured by their achieved BER, as a function of the transmit-
such as portable radio and deepspace communicated energy per information bit. Receivers that achieve the
tions. Our preference is for signals that require as smallest possible BER for a given channel and modulation
little power as possible for a desired performance signal set are called optimal.
level. Binary modulation systems are the simplest to imple-
ment and detect, but they are not necessarily the most
The problem of designing modulation signals that pos- efficient in communicating information. Modulators with
sibly optimize some aspect of performance, or satisfy some larger signal sets use a smaller bandwidth to transmit a
constraints imposed by the communication medium or the given information bit rate. For example, one can envision
hardware, is known generally as signal design. Signal de- having a modulation signal set containing four (instead of
sign problems are important and widely prevalent in com- two) signals: s1 (t), s2 (t), s3 (t), s4 (t). With four signals, we
munications. can assign to each a two-bit sequence in a one-to-one cor-
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright © 2007 John Wiley & Sons, Inc.
2 Information Theory of Modulation Codes and Waveforms
respondence, for example, as follows: We assume, which is a good assumption in practice, that
W < < f c . In amplitude modulation (AM), the information
s1 (t) ⇔ 00
signal modulates the amplitude of the carrier according to:
s2 (t) ⇔ 01
s3 (t) ⇔ 10 u(t) = Am(t)cos(2π f c t + φ) (2)
s4 (t) ⇔ 11
where φ is some fixed carrier phase. Insight into the process
In this case, each time a transmitted signal is detected cor-
of modulation is obtained by looking at the Fourier trans-
rectly, the receiver extracts two (correct) bits. The bit rate
form of the modulated signal, given by (see, for example,
has also doubled compared with a binary modulator for the
Reference (1))
same signaling rate (transmitted signals per second). Be-
cause bandwidth is proportional to the signaling rate, we A
U( f ) = [M( f − f c )e jφ + M( f + f c )e− jφ ] (3)
have effectively doubled our transmission efficiency using 2
a modulator with four signals instead of two. Of course, the
Figure 1plots the magnitude of the Fourier transform of
job of the receiver is now harder because it has to make a
the modulated signal for a simple choice (for presentation
four-way decision, instead of just a binary decision, and ev-
purposes) of the Fourier transform of the information sig-
erything else being the same, the probability of making an
nal. It is easy to see that, whereas the information signal
erroneous decision increases. We refer to the above modu-
has bandwidth W, the modulated signal has a bandwidth of
lator as a 4-ary modulator (or a quaternary modulator).
2W. Also, as can be observed from equation 3, no unmodu-
Clearly, the idea can be extended to modulation sig-
lated carrier component exists, which would be manifested
nal sets that contain M = 2k signals, for some integer
as delta functions at the carrier frequency fc . We refer to
k = 1, 2, 3 · · · . In this case, each transmitted signal carries
this scheme as double-sideband, suppressed-carrier (DSB-
k bits. We refer to modulators that use M signals as M-ary
SC), amplitude modulation.
modulators. As in the 4-ary modulator example above, the
Demodulation of DSB-SC amplitude modulated signals
advantage of a larger number of modulation signals is that
can be achieved by multiplying the received signal by a lo-
the number of signals per second that needs to be trans-
cally generated replica of the carrier, which is generated
mitted to accommodate a certain number of bits per second
by a local oscillator (LO). For best performance, the locally
decreases as M increases. Because the number of signals
generated carrier must match as closely as possible the
per second determines to a large extent the bandwidth re-
frequency and phase of the received carrier. It is usually
quired, more signals means a smaller required bandwidth
reasonable to assume the receiver generated carrier fre-
for a given number of transmitted bits per second, which is
quency matches the carrier frequency in the received sig-
a desirable result. The price paid for large signal sets is in
nal well.1 Neglecting noise, for simplicity, and assuming
complexity and, as previously pointed out, in possibly re-
perfect frequency synchronization, the demodulator is de-
duced performance for the same expended average energy
scribed mathematically by
per bit.
Although many analog modulation (communication) z(t) = u(t)cos(2π f c t + φ̂)
systems are still in use, the trend is for systems to become A A (4)
= m(t)cos(φ − φ̂) + m(t)cos(4π f c t + φ + φ̂)
digital. Currently, two prominent examples of analog sys- 2 2
tems becoming digital are cellular phones and digital TV where φ̂ is the phase of the locally generated carrier. Now
broadcasts. Digital modulation techniques are by far the the component in equation 4at twice the carrier frequency
more attractive. is easily filtered out by low-pass filtering to yield
A
ANALOG MODULATION m̂(t) = m(t)cos(φ − φ̂) (5)
2
The most prevalent medium for everyday communication which is a scaled version of the modulation signal. In the
is through RF (sinusoidal) carriers. Three quantities exist presence of noise, to maximize the signal-to-noise ratio
whose knowledge determines exactly the shape of an RF (SNR), it is important that the phase error (φ − φ̂) be small.
signal: (1)its amplitude; (2)its phase; and (3)its frequency, The problem of phase synchronization is an important one
as indicated in equation 1: and is often practically achieved using a phase-locked loop
(PLL) (see, for example, References (2)–(5).) When the lo-
s(t) = A(t) cos[2π f c t + φ(t)] (1) cally generated carrier is perfectly phase and frequency
locked to the phase and frequency of the received signal,
where fc is the frequency of the sinusoidal signal in Hertz.
detection of the information is referred to as coherent. This
Information can be conveyed by modulating the amplitude,
is in contrast to noncoherent detection, when the phase of
the instantaneous frequency, or the phase of the carrier (or
the locally generated carrier does not match that of the re-
combinations of the three quantities).
ceived signal. Clearly, coherent detection achieves the ulti-
mate limit in performance. It can be approached in practice
Amplitude Modulation
by using sophisticated algorithms, at the cost of increased
Let the information signal m(t) be baseband and bandlim- complexity.
ited to some bandwidth W Hz. A baseband signal bandlim- A simpler, noncoherent, detector can be used if the trans-
ited to W Hz has a frequency spectrum centered at the mitted carrier contains an unmodulated component (or a
origin and contains substantially no energy above W HZ. “pilot tone”) resulting in what is referred to as DSB mod-
Information Theory of Modulation Codes and Waveforms 3
Figure 1. The magnitude of the Fourier transform of a DSB-SC amplitude-modulated signal. (Figure is not to scale).
ulation. In conventional AM (such as in broadcasting), the general form of a single-sideband signal is

modulated signal takes the form
u(t) = A[m(t)cos(2π f c t) ± m̂(t)sin(2π f c t)] (6)
u(t) = A[1 + am(t)]cos(2π f c t + φ)
where m̂(t) is the Hilbert transform of m(t) given by
with the constraint that |m(t)| ≤ 1; a, 0 ≤ a ≤ 1, is the mod- 1
ulation index. Figure 2shows an example of a convention- m̂t = m(t) ∗ ⇔ M̂( f ) = M( f )H( f )
πt
ally modulated AM signal. Clearly, the modulated signal
for conventional AM has a strong unmodulated component where H(f) is the Fourier transform of h(t) = 1/πt and is
at the carrier frequency that carries no information but given by
uses power, and thus, a significant power penalty exists −j f >0
in using it. The benefit resulting from the reduced power H( f ) = { j, f <0
efficiency is that simple receivers can now be used to de- 0, f = 0,
tect the signal. The loss in power efficiency can be justi-
fied in broadcasting, where conventional AM is used, be- In equation 6 , the plus or minus sign determines
cause in this case only one high-power transmitter draws whether the upper or the lower sideband is chosen. Fig-
power from the power grid, with millions of (now simpler ure 4shows the spectrum of an upper sideband SSB signal.
and therefore less costly) receivers. For a more complete exposition to SSB, including modu-
lation and demodulation methods, consult References (1)
and (6–9).
Demodulation of AM Signals Another amplitude modulation scheme, widely used in
The most popular detection method for conventional AM TV broadcasting, is vestigial sideband (VSB). The reader is
is envelope detection. This method consists of passing the referred to References (1) and (6–9) for more information.
received modulated signal [usually after RF amplification
and down conversion to some intermediate frequency (IF)] Angle Modulation
through a rectifier followed by a simple low-pass filter (in Angle modulation of a sinusoidal carrier includes phase
the form of a simple, passive, RC circuit). This simple demodulation (PM) and frequency modulation (FM). In phase
tector is shown in Fig. 3. modulation, the information signal modulates the in-
Double-sideband amplitude modulation is wasteful in stantaneous phase of a high-frequency sinusoidal carrier,
bandwidth, requiring a bandwidth that is twice the base- whereas in frequency modulation, the information signal
band signal bandwidth. It can be shown that the two side- directly modulates the instantaneous frequency of the car-
bands are redundant, and that the information signal can rier. As the instantaneous frequency and phase of a sig-
be obtained if only one sideband was transmitted, which nal are simply related (the instantaneous frequency is the
reduces the required bandwidth by a factor of two com- scaled derivative of the instantaneous phase), clearly PM
pared with DSB-AM. At the same time, an improvement and FM are also closely related and have similar proper-
in power efficiency, occurs because transmitting the redun- ties. For angle modulation, the modulated signal is given
dant sideband requires not only extra bandwidth but also by
extra power. When only one sideband is transmitted, the
resulting signal is referred to as single sideband (SSB). The u(t) = Acos[2π f c t + φ(t)]
Figure 2. Illustration of a conventionally amplitude-modulated signal.
Figure 3. A simple demodulator for conventional AM signals.
where stants, given by
d p m(t) PM φ = d p · |m(t)|
t
φ(t) = { and
2πd f m(τ)dτ FM
−∞
f = d f · |m(t)|
The constants dp and df are the phase and frequency de-
In turn, the peak deviation constants define the phase
viation constants, respectively. These constants, along with
and frequency modulation indices according to
the peak amplitude of the information signal, define the
peak phase deviation and peak frequency deviation con- β p = φ
Figure 4. The spectrum of an upper sideband SSB signal.
and DIGITAL MODULATION
f A wide variety of digital modulation methods exists, de-

βf =
W pending on the communication medium and the mode
of communication, both of which impose constraints on
where W is the bandwidth of the information signal signal the nature of transmitted signals. For example, for opti-
m(t). As an example, the peak frequency deviation for FM cal systems that use an optical carrier [generated by a
broadcasts is 75 KHz, and the signal bandwidth is limited light-emitting diode (LED) or a laser], various modulation
to 15 KHz, which yields a modulation index of 5. For illus- schemes are particularly suitable, which may not be suit-
tration, Fig. 5shows typical waveforms for frequency and able for RF communications systems. Similarly, modula-
phase modulation. tion schemes used in magnetic recording systems may not
The spectrum of an angle-modulated signal is much be suitable for other systems. Generally, as indicated in
more difficult to obtain mathematically than in the AM the Introduction, the modulation must be matched to the
case because angle modulation is nonlinear. Moreover, channel under consideration.
strictly speaking, angle-modulated signals have an infi-
nite bandwidth. However, an approximation for the effec- Signal Space
tive bandwidth (i.e., the frequency band containing most of
the signal energy) of angle-modulated signals is given by In designing and describing digital modulation schemes, it
Carson’s rule: is often desirable to consider modulation signals as points
in some appropriate signal space, spanned by a set of
B = 2(β + 1)W orthonormal-basis signals. The dimensionality of the sig-
nal space equals the number of orthonormal-basis signals
where β is the phase- or frequency-modulation index and W that span it.
is the bandwidth of the information signal. The bandwidth A set of signals {φ1 (t), φ2 (t), · · · , φN (t)}, for 0 ≤ t ≤ T is
of the modulated signal increases linearly as the modula- orthonormal if the following condition holds:
tion index increases. FM systems with a small modulation T
index are called narrowband FM, whereas systems with a 1, i= j
φi (t)φ j (t)dt = {
large modulation index are called wideband FM. One pop- 0
0, i = j
ular and practical way to generate wideband FM is to first If s(t) is any signal in the N-dimensional space spanned
generate a narrowband FM signal (which is easily gener- by these signals, then it can be expressed as
ated) and then, through frequency multiplication, to con-
vert it into a wideband FM signal at an appropriate carrier
N
frequency. Wideband FM is used in broadcasting, and nar- s(t) = si φi (t)

rowband FM is used in point-to-point FM radios. i=1
Detection of FM or PM signals takes several different for some set or real numbers s1 , s2 , ·, sN . The N coefficients
forms, including (PLLs) and discriminators, which convert uniquely describing s(t) are obtained using
FM into AM that is then detected as such. For more infor- T
mation on ways to modulate and demodulate angle modu-
sk = s(t)φk (t)dt, k = 1, 2, · · · , N
lated signals, consult References (1,3,5), and (9). 0
Figure 5. Illustration of frequency- and phase-modulated signals.
Figure 6illustrates the concept of signal space for the spe- where
cial case of two dimensions. In the figure, four distinct sig- 2π(i − 1)
nals are represented as points in the signal space. θi =
M
Perhaps the most widely known and used modulation
schemes are those pertaining to RF communication, some and
of which are examined next. T
E= si2 (t)dt
0
Phase-Shift Keying
is the signal energy. Equation (7)is rewritten in a slightly
Under phase-shift keying (PSK), the information bits de- different form as
termine the phase of a carrier, which takes values from a
discrete set in accordance with the information bits. The √ 2 2
si (t) = E[cos(θi ) cos(2π f c t) − sin(θi ) sin(2π f c t)]
general form of M-ary PSK signals (i.e., a PSK signal set √ T T
containing signals) is given by = E[cos(θi )φ1 (t) − sin(θi )φ2 (t)]
where φ1 (t) and φ2 (t) are easily observed to be orthonormal.
2E
si (t) = cos(2π f c t + θi ), i = 1, 2, · · · , M, 0 ≤ t ≤ T Thus, PSK signals are points in a two-dimensional space
T spanned by φ1 (t) and φ2 (t). Figure 7illustrates various PSK
(7) signal constellations, including binary PSK (BPSK) and 4-
Figure 6. Illustration of the concept of signal space. The two signals on top are the basis signals. Signals a(t), b(t), c(t), and d(t) are
represented in signal space as points in the two-dimensional space spanned by the two basis signals.
ary PSK, also known as quadrature PSK (QPSK). The fig- finding the modulation signal that maximizes
ure also illustrates the mapping of information bits to each T
signal in the constellation. The illustrated mapping, known l1 = r(t)si (t)dt
as Gray coding, has the property that adjacent signals are 0
assigned binary sequences that differ in only one bit. This
This signal is the well-known correlation receiver, where
property is desirable in practice, because, when a detection
the most likely signal transmitted is chosen as the one
error is made, it is more likely to be to a signal adjacent to
most correlated with the received signal. The correlation
the transmitted signal. Then Gray coding results in a sin-
receiver involves a multiplication operation, followed by in-
gle bit error for the most likely signal errors.
tegration. Because processing is linear, it is possible to ob-
tain the same result by passing the received signal through
Performance in Additive Gaussian Noise. The simplest a linear filter with an appropriate impulse response and
channel for data transmission is the additive, white, Gaus- sampling it at an appropriate instant. The impulse re-
sian noise (AWGN) channel. For this channel, the transmit- sponse hi (t) of the linear filter is easily derived as
ted signal is corrupted by and additive Gaussian process,
resulting in a received signal given by hi (t) = si (T − t)
r(t) = si (t) + n(t), 0 ≤ t ≤ T (8) This linear filter implementation of the optimum receiver
is called a matched-filter receiver.
where n(t) is zero-mean, white Gaussian noise of spectral For binary PSK, the probability that the optimal re-
density N0 /2. ceiver makes a decision error is given by
For PSK signals, the optimum receiver (detector), also
known as a maximum-likelihood (ML) receiver, decides 1 E
PBPSK (e) = erfc( ) (9)
which of the M possible PSK signals was transmitted by 2 N0
Figure 7. Signal space representation of various PSK constellations. The bit assignments correspond to Gray coding.
where by
x
2 −y2 si (t) = (2i − 1 − M) E p(t), i = 1, 2, · · · , M, 0 ≤ t ≤ T
erfc(x) = 1 − √ e dy
π 0
where p(t) is a unit-energy baseband pulse. Figure 9shows
is the complimentary error-function. In equation 9 , the the signal-space representation of PAM signals assuming
ratio E/N0 is the SNR, which determines performance. The E = 1. In contrast to PSK signals, clearly not every signal
performance of QPSK is also derived easily and is given by has the same energy; in which case, the constellation is
described by its average energy:
E
M
PQPSK (e) = PBPSK (e)[2 − PBPSK (e)] M2 − 1
Eav = (2i − 1 − M)2 = ( )E
M 3
i=1
where PBPSK (e) is as given in equation 9 . An exact expres-
sion for the error probability of larger PSK constellations
also exists and is found, for example, in Chapter 9 of Ref- Performance in Additive Gaussian Noise. Based on the
erence (1). Figure 8shows the error probability of various data r(t) received (as given in equation 8 ), the maximum-
PSK constellations as a function of the SNR per informa- likelihood receiver for PAM signaling chooses as the most
tion bit. likely signal transmitted the signal that maximizes
√
E
Baseband Pulse-Amplitude Modulation li = (2i − 1 − M) · r − (2i − 1 − M)2
2
Pulse-amplitude modulation (PAM) is the digital equiva-
where
lent of AM. The difference is that now only discrete ampli- T
tudes are allowed for transmission. M-ary PAM is a one-
r= r(t) p(t)dt
dimensional signaling scheme described mathematically 0
Figure 8. Symbol error probability for BPSK, QPSK, and 8-PSK as a function of the SNR per bit.
Figure 9. The signal space representation of various PAM constellations.
In signal space, the decision boundaries for this receiver Quadrature Amplitude-Modulation
are midway between constellation points, and a decision is Quadrature amplitude modulation (QAM) is a popular
made accordingly, based on where r falls on the real line. scheme for high-rate, high-bandwidth efficiency systems.
The error probability for M-ary PAM signals is given by QAM is a combination of both amplitude and phase mod-
ulation. Mathematically, M-ary QAM is described by
√
(M − 1) 3 Eav si (t) = E p(t)[Ai cos(2π f c t) + Bi sin(2π f c t)], 0 ≤ t ≤ T,
PPAM (e) = erfc( )
M M 2 − 1 N0
i = 1, 2, · · · , M
The error probability for various PAM constellations is where Ai and Bi take values from the set {±1, ±3, ±5, · · · }
shown in Fig. 10as a function of SNR per bit. and E and p(t) are as defined earlier. The signal space rep-
Figure 10. Symbol error probability for 2-, 4-, and 8-PAM as a function of SNR per bit.
resentation of QAM signals is shown in Fig. 11for various and

values of M, which are powers of 2; that is, M = 2k , k = T
2, 3, · · · . For even values of k, the constellations are square, rs = r(t) p(t)sin(2π f c t)dt
whereas for odd values, the constellations have a cross 0
shape and are thus called cross constellations. For square For square constellations that correspond to independent
constellations, QAM corresponds to the independent am- PAM of each carrier, an exact error probability is derived
plitude modulation of an in-phase carrier (i.e., the cosine easily and is given by
carrier) and a quadrature carrier (i.e., the sine carrier). 2
1 3 Eav
PQAM (e) = 1 − [1 − (1 − √ )erfc( · )]
M 2(M − 1) N0
Performance in Additive Gaussian Noise. The optimum
For cross constellations, tight upper bounds and good ap-
receiver for QAM signals chooses the signal that maxi-
proximations are available. Figure 12plots the symbol er-
mizes
ror probability of various square QAM constellations as a
√ function of SNR per bit.
E 2
li = Ai rc + Bi rs − (Ai + Bi2 )
4 Frequency-Shift Keying
As the name implies, frequency-shift keying (FSK) modu-
where lates the frequency of a carrier to convey information. FSK
is one of the oldest digital modulation techniques and was
T
the modulation of choice for the first, low-rate modems. Its
rc = r(t) p(t)cos(2π f c t)dt
0
main attribute, which makes it of interest in some appli-
Figure 11. Signal space representation of various QAM constellations.
cations, is that it can be detected noncoherently (as well as For binary (orthogonal) signaling, the error probability is
coherently), which reduces the cost of the receiver. Mathe- given simply by
matically, the modulated M-ary FSK signal is described by
1 E
PFSK (e) = erfc ( ), (coherent FSK)
2 2N0
2E
si (t) = cos[2π( f c + f i )t], 0 ≤ t ≤ T, i = 1, 2, · · · , M
T which is 3 dB worse than BPSK. For M-ary signaling, an
exact expression exists in integral form and is found, for ex-
where ample, in Reference (10). Noncoherent detection does not
2i − 1 − M assume phase coherence and does not attempt to phase-
fi = ( )r f lock the locally generated carrier to the received signal.
2
In this case, it is easy to argue that the phase difference
f is the minimum frequency separation between modu- between the LO carrier and the received carrier is com-
lation tones. For orthogonal signaling (i.e., when the cor- pletely randomized. An optimum receiver is also derived
relation between all pairs of distinct signals is zero), the in this case, and it is one that maximizes over the set of
minimum tone spacing is 1/2T. This a condition is often frequency tones
imposed in practice. Orthogonal signaling performs well
as a function of energy per bit, but it is also bandwidth- li = rci2 + rsi2
inefficient, which makes it impractical for high-speed, band
where
limited applications.
T
rci2 = r(t)cos[2π( f c + f i )t]dt
Performance in Additive Gaussian Noise. FSK is detected 0
coherently or incoherently. Coherent detection requires a
and
carrier phase synchronization subsystem at the receiver
T
that generates locally a carrier phase-locked to the received
carrier. The optimum receiver for coherent detection makes rsi2 = r(t)sin[2π( f c + f i )t]dt
0
decisions by maximizing the following (implementation as-
sumes phase-coherence): The exact error-probability performance of this noncoher-
ent receiver is available in analytical form, but it is com-
T
plicated to compute for the general M-ary case (see, for ex-
li = r(t)si (t)dt
0
ample, Reference (10)). For the binary case, the error prob-
Figure 12. Symbol error probability as a function of SNR per bit for 4-, 16-, and 64-QAM.
ability has a simple form given by modulation (CPM) signals. These signals constrain the
phase of the transmitted carrier to be continuous, thereby
1 − 2NE
PFSK (e) = e 0 (noncoherent FSK) reducing the spectral sidelobes of the transmitted signals.
2 Mathematically, the modulation signals for CPM are de-
Figure 13compares the performance of coherent and inco- scribed by the expression
herent binary FSK. At an error probability of about 10−6 ,
noncoherent detection is inferior only slightly more than u(t) = Acos[2π f c t + φ(t; d)]
half a decibel compared with coherent detection. However,
this small loss is well compensated for by the fact that no where
carrier phase synchronization is needed for the former.
n
φ(t; d) = 2π dk hk q(t − kT ), nT ≤ t ≤ (n + 1)T

Continuous-Phase Modulation k=−∞
All modulation schemes described so far are memoryless, The dk are the modulation symbols and hk are the modu-
in the sense that the signal transmitted in a certain symbol lation indices, which may vary from symbol to symbol. For
interval does not depend on any past (or future) symbols. binary modulation, the modulation symbols are either 1
In many cases, for example, when a need exists to shape or −1. Finally, q(t) is the integral of some baseband pulse
the transmitted signal spectrum to match that of the chan- p(t) containing no impulses (thus guaranteeing that q(t) is
nel, it is necessary to constrain the transmitted signals in continuous)
some form. Invariably, the imposed constraints introduce t
memory into the transmitted signals. One important class
q(t) = p(τ)dτ
of modulation signals with memory are continuous-phase −∞
Figure 13. Error probability comparison between coherent and noncoherent FSK.
When p(t) is zero for t ≥ T , we have what is called full- Modulation Codes
response CPM, otherwise, we have partial-response CPM. Another technique for shaping the spectrum of transmitted
In general, partial-response CPM achieves better spectral modulation signals is putting constraints on the sequence
sidelobe reduction than does full-response CPM. A special of bits sent to the modulator. This coding of bits to shape the
case of CPM in which the modulation indices are all equal spectrum of the transmitted modulation signals is called
and p(t) is a rectangular pulse of duration T seconds is modulation coding or line coding. Important examples of
called continuous-phase FSK (CPFSK). If, h = 1/2, we have the use of such codes are in magnetic and optical recording
what is called minimum-shift keying (MSK). A variation channels. Simple examples of modulation codes are found
of MSK, in which the rectangular baseband pulse is first in the baseband transmission of binary data where a pulse
passed through a filter with a Gaussian-shape impulse re- is sent for a binary “1” and its negative for a “0” (called an-
sponse for further reduction in the spectral sidelobes, is tipodal signaling). If the pulse amplitude does not return to
called Gaussian MSK (GMSK). Various simple ways for de- zero in response to consecutive similar bits, then we have
tecting GMSK are available, which combined with its spec- nonreturn-to zero (NRZ) signaling. If the pulse returns to
tral efficiency, has made it a popular modulation scheme. zero, then we have return-to-zero (RZ) signaling. The en-
In particular, it is the modulation scheme originally used coding of bits using NRZ and RZ signaling is illustrated in
for the European digital cellular radio standard, known as Fig. 14.
GSM. For more information on CPM signaling, including It is often desirable to have a transmitted pulse se-
spectral characteristics and performance in noise, refer to quence, in response to random input bits, with no spectral
Reference (10). component at zero frequency (i.e., in dc). This condition
is desirable, for example, when the modulation signals are
sent through a channel with a null at dc. If the bits arriving
Figure 14. Illustration of NRZ, RZ, and Manchester coding.
at the input of the modulator are truly random (each with ber of consecutive zeros between ones and are also called
probability 1/2 of being zero or one) and independent, then (d, k) codes, where d is the minimum number of zeros and
the expected value of the dc component of the transmitted k is the maximum number of zeros between ones. The min-
signal is zero. However, at any given time (even though the imum number of zeros between ones ensures that ISI is
average is zero), a significant dc component may be caused kept small, and the maximum number of zeros between
by the transmission of a long sequence of zeros or ones. Be- ones ensures that the transmitted signal has enough tran-
sides the creation of a dc component, these long sequences sitions in it to aid in timing recovery. RLL codes (and in fact
of zeros or ones also negatively affect the performance of a much larger class of codes) are conveniently described by
the timing recovery system at the receiver, whose function finite-state machines (FSMss). An FSM consists of a set of
is to establish time synchronization (essential before data interconnected states that describe the allowable bit tran-
detection). sitions (paths). The interconnections between all possible
Biphase or Manchester pulses have the property of zero pairs of states are often described by a two-dimensional
dc over each bit interval. These pulses and their encoding state transition matrix, which is known also as the adja-
are illustrated in Fig. 14, along with NRZ and RZ signaling. cency matrix. A one at the i,j position in the matrix means
An important property of a line code that describes the dc that there is a path from state i to state j. A zero means
variations of a baseband signal is the running digital sum that no path exists between the two states. Figure 15shows
(RDS) (11). The RDS is the running sum of the baseband the FSM for the (1,3) (d, k) code. It consists of four states,
amplitude levels. It has been shown that, if the RDS for a and its adjacency matrix is given by
modulation code is bounded, then the code has a null at dc
(12). This process facilitates transmission of the modulated ⎛ ⎞
0 1 0 0
data through channels with a null at dc and avoids a form ⎜1 0 1 0⎟
of intersymbol-interference (ISI) known as baseline wan- A=⎝ ⎠
1 0 0 1
der. A converse result also shows that modulation codes, 1 0 0 0
generated by finite-state machines, which have a spectral
null at dc, have a bounded RDS (13).
Clearly, the constraints imposed on the binary sequences
(in the form of d and k) limit the number of possible se-
Run-Length Limited Codes. Run-length limited (RLL) quences of a given length n, which satisfy the constraint
codes are an important class of modulation codes, which to a subset of the total number of 2n possible sequences. If
are often used in magnetic recording systems. RLL codes the number of sequences of length n satisfying the (d, k)
impose constraints on the minimum and maximum num- constraints is M(n), then the capacity of the code is defined
Figure 15. The finite-state machine for the ((1, 3)) RLL code.
by BIBLIOGRAPHY
1
C(d, k) = n → ∞lim log 2 [M(n)] (10) 1. Proakis J.; Salehi M. Communication Systems Engineering;
n Prentice-Hall: Englewood Cliffs, NJ, 1994.
For a fixed n, the ratio on the right-hand side of equation 2. Gardner F. M. Phaselock Techniques; Wiley: New York, 1966.
10is called the rate of the code (which is the fraction of 3. Viterbi A. J. Principles of Coherent Communications; McGraw-
information bits per transmitted bit). It can be shown that Hill: New York, 1966.
the rate of the code is monotonically nondecreasing in n. 4. Lindsey W. C. Synchronization Systems in Communications;
Thus, the capacity of the code is the largest achievable rate. Prentice-Hall: Englewood Cliffs, NJ, 1972.
Shannon (14,15) has shown that the capacity of a FSM (the 5. Blanchard A. Phase-Locked Loops: Application to Coherent Re-
(d, k) code is just an example) is given by ceiver Design; Wiley: New York, 1976.
C(d, k) = log 2 (λmax ) 6. Stremler F. G. Introduction to Communication Systems; 3rd
ed.; Addison-Wesley: Reading, MA, 1990.
where λmax is the largest real eigenvalue of the adjacency 7. Haykin S. Communication Systems, 3rd ed.; Wiley: New York,
matrix of the FSM. As an example, the eigenvalues of the 1994.
adjacency matrix for the (1,3) code are 1.4656, −1.0000, 8. Roden M. S. Analog and Digital Communication Systems;
−0.2328 + 0.7926i, and −0.2328 − 0.7926i The largest real Prentice-Hall: Englewood Cliffs, NJ, 1991.
eigenvalue is 1.4656, and thus, the capacity of the code is 9. Couch L. W. Modern Communication Systems; Prentice-Hall:
log 2 (1.4656) = 0.5515. For an excellent overview of infor- Englewood Cliffs, NJ, 1995.
mation theory, including Shannon’s result above, consult 10. Proakis J. Digital Communications, 3rd ed.; McGraw-Hill:
Reference (16). New York, 1995.
The fact that an FSM is found that produces sequences 11. Franaszek P. A. Sequence-State Coding for Digital Transmis-
satisfying the necessary constraints does not automatically sion. Bell Syst. Tech. J., 1968, 47, 143.
imply that a code has been constructed. The problem of as- 12. Calderbank A. R.; Mazo J. Spectral Nulls and Coding with
signing information bits to encoded bits still exists. The Large Alphabets. IEEE Commun. Mag.December 1991.
problem of constructing such codes from their FSM repre- 13. Yoshida S.; Yajima Y. On the Relationship Between Encoding
sentation has been studied by Adler et al. (17). An excel- Automaton and the Power Spectrum of its Output Sequence.
lent tutorial paper on the topic can be found in Reference Trans. IECE 1976, E59,p. 97.
(18). Practical examples of applying the results of Refer- 14. Shannon C. E. A Mathematical Theory of Communication. Bell
ence (17) are, for example, in References (19) and (20). An- Syst. Tech. J. 1948, 27,pp 379–423.
other important class of codes that shapes the spectrum 15. Shannon C. E. A Mathematical Theory of Communication. Bell
of the transmitted data and achieves a coding gain in the Syst. Tech. J. 1948, 27,pp. 623–656.
process is the class of matched spectral null (MSN) codes. 16. Cover T. M.; Thomas J. A. Elements of Information Theory;
The interested reader is referred to the paper by Karabed Wiley Interscience: New York, 1991.
and Siegel (21)for more details. 17. Adler R. L.; Coppersmith D.; Hassner M. Algorithms for Slid-
Yet, another, very important class of modulation signals ing Block Codes. IEEE Trans. Inform. Theory 1983, IT-29,pp.
includes those signals that combine coding and modula- 5–22.
tion for improved performance. These combined modula- 18. Marcus B. H.; Siegel P. H.; Wolf J. K. Finite-State Modulation
tion and coding techniques and, in particular, trellis-coded Codes for Data Storage. IEEE J. Select. Areas Commun. 1992,
modulation (TCM) became better known from the break- 10,pp. 5–37.
through paper of Unger-boeck (22). In contrast to previ- 19. Calderbank A. R.; Georghiades C. N. Synchronizable Codes
ous classic coding techniques that separate the coding and for the Optical OPPM Channel. IEEE Trans. Inform. Theory
modulation problems, TCM achieves a coding gain (i.e., im- 1994, 40,pp. 1097–1107.
proved performance) without expanding bandwidth. It is 20. Soljanin E.; Georghiades C. N. Coding for Two-Head Recording
thus very appealing in band limited applications, such as Systems. IEEE Trans. Inform. Theory 1995, 41,pp. 747–755.
telephone modems, where it has been widely employed.
1 This assumption is not as easy to justify when the receiver moves
relative to the transmitter, because of the frequency offset caused by the Doppler effect.
21. Karabed R.; Siegel P. Matched-Spectral Null Codes for Par-

tial Response Channels. IEEE Trans. Inform. Theory 1991,
IT-37,pp. 818–855.
22. Ungerboeck G. Channel Coding with Multilevel/Phase Sig-
nals. IEEE Trans. Inform. Theory 1982, IT-28,pp. 55–67.
COSTAS N. GEORGHIADES
Texas A&M University
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Information Theory of Multiaccess Communications ● Email this page Enter words or
● Print this page
Tien M. Nguyen1, Hung Nguyen2, Boi N. Tran3
1The Aerospace Corporation, El Segundo, CA
2Mountain Technology Inc., Milpitas, CA ❍ Advanced Product
3The Boeing Company, Long Beach, CA Search
Copyright © 1999 by John Wiley & Sons, Inc. All rights ❍ Search All Content
reserved. ❍ Acronym Finder
DOI: 10.1002/047134608X.W4211
Abstract
Fixed-Assignment Multiple Access
Random-Assignment Multiple Access
Performance Comparison of Multiple Access Techniques
Applications of Random-Access Techniques in Cellular Telephony
Keywords: communications resources; multiple access; multiplexing; frequency-division multiple access; time-division multiple
access; code-division multiple access; space-division multiple access; polarization-division multiple access; fixed-assignment
multiple access; random access; controlled random access; guard band; direct-sequence spread spectrum; frequency-hopping
spread spectrum; processing gain; near–far problem; crosscorrelation; random-assignment multiple access; pure ALOHA; slotted
ALOHA; reservation ALOHA; carrier-sense multiple access; data-sense multiple access; 1-persistent carrier-sense multiple access;
nonpersistent carrier-sense multiple access; p-persistent carrier-sense multiple access; polling technique; token passing; cellular
digital packet data; controlled random-assignment multiple access; carrier-sense multiple access with busy-tone signaling; cellular
telephone system; mobile telephone switching office; public switched telephone network; mobile switching service center;
common air interface; roaming service; mobile identification number; electronic serial number; advanced mobile phone system;
extended European total access cellular system IS-54; US digital cellular; global mobile system, IS-95


166 INFORMATION THEORY OF MULTIACCESS COMMUNICATIONS
INFORMATION THEORY OF
MULTIACCESS COMMUNICATIONS
The basic communications resources available to users are
frequency and time. The efficient allocation of these communi-
cations resources lies in the domain of communications multi-
ple access. The term ‘‘multiple access’’ means the remote
sharing of a communications resource (e.g., satellite). The
term multiple access is often confused with the term multi-
plexing. Multiplexing indicates the local sharing of a commu-
nications resource (e.g., a circuit board). Normally, for multi-
plexing, the resource allocation is normally assigned a priori.
This article focuses on the theory of multiple access. High
level description of various multiple access techniques and a
comparison among them will be given.
For multiple access, there are three basic techniques for
distributing the communications resources: frequency-divi-
sion multiple access (FDMA), time-division multiple access
(TDMA), and code-division multiple access (CDMA). For
FDMA, one specifies the subbands of frequency to be allocated
to users. For TDMA, periodically recurring time slots are
identified and then allocated to users. This technique allows
users to access the resource at fixed or random times, de-
pending on the systems. For CDMA, full channel bandwidth
is utilized simultaneously with the time resource. In addition,
two other techniques for multiple access are also available,
namely, space-division multiple access (SDMA) and polariza-
tion-division multiple access (PDMA). SDMA, also referred to
as multibeam frequency reuse multiple access, uses spot
beam antennas to separate radio signals by pointing them in
different directions, which allows for reuse of the same fre-
quency band. PDMA, or dual polarization frequency reuse,
employs orthogonal polarization to separate signals, which
also allows for reuse of the same frequency band.
The three basic multiple access schemes are implemented
with various multiuser access algorithms to form fixed-as-
signment or random-access schemes. In a fixed-assignment
access scheme, a fixed allocation of communication resources,
frequency or time, or both, is made on a predetermined basis
to a single user. The random-access scheme allows the users
to access communications resources randomly. When the ran-
dom-access algorithm exercises some control over the access
method to improve the efficiency of the uncontrolled random
access methods, the result is referred to as the controlled ran-
dom access technique.
This article describes the underlying theory behind the
multiple access techniques and their applications in satellite
and cellular systems. Both fixed- and random-access tech-
niques will be described with their associated applications.
Since the article is intended for readers who are unfamiliar
INFORMATION THEORY OF MULTIACCESS COMMUNICATIONS 167
with this field, only high level descriptions with minimum

technical details are presented.
FIXED-ASSIGNMENT MULTIPLE ACCESS
Guard time
Guard time
Frequency
As mentioned earlier, multiple access techniques are required
for multiple users to efficiently share remote communications Time slot Time slot Time slot ...
resources. There are two major categories of multiple access 1 2 3
methods: fixed assignment and random access. This section
describes the three basic approaches for fixed-assignment
multiple access: FDMA, TDMA, and CDMA. For complete-
ness, brief descriptions of SDMA and PDMA are also pre-
Time
sented.
Figure 2. Illustration of time-division multiple access technique.
Frequency-Division Multiple Access
The frequency-division multiple access (FDMA) technique is
caused by AM–AM and AM–PM distortions (1,5). This
derived based on the frequency-division multiplexing (FDM)
means it is power inefficient.
method. The FDM method involves mixing (or heterodyning)
the signals at the same frequency band with fixed frequencies • FDMA is involved with narrowband technology, which
from local oscillators to different frequency bands and then also involves narrowband filters that may not be realiz-
combining the resulting multiple signals (at different fre- able in very large scale integrated (VLSI) digital circuits.
quency bands) for transmission as a single signal with a This means higher cost for terminals even under volume
wider bandwidth (1). Figure 1 shows the FDM scheme (1, Fig. production conditions.
9.3, page 480). Note that ‘‘guard bands’’ between the fre- • It is inflexible due to limited fixed bit rate per channel.
quency assignments are provided as buffer zones to mitigate
the adjacent channel interference. For a fixed-assignment Time-Division Multiple Access
FDMA system, a user is assigned to a fixed subchannel for
Time-division multiple access (TDMA) uses the full spectrum
transmission, and the subchannel is retained until released
occupancy that is allocated to the system for a short duration
by the assigned user. The receiver terminal has precise
of time called the time slot, as shown in Fig. 2 (1, Fig. 9.7, p.
knowledge of the transmission subchannel, and a filter is
484). Note that the guard band is provided here for crosstalk
used to extract the designated signal out of the received com-
avoidance. TDMA employs the time-division multiplexing
posite signal.
method in which the system time is divided into multiple time
The advantages of the fixed-assignment FDMA are (2):
slots used by different users. Several time slots make up a
• Channel capacity increases as information bit rate de- frame. Each slot is made up of a preamble plus information
creases. To reduce the information bit rate one can use bits addressed to various terminal users as shown in Fig. 3
an efficient modulation method such as M-ary phase- (1, Fig. 9.9, p. 485). In a fixed-assignment TDMA system, a
shift keying (PSK) (3) or the continuous phase modula- transmit controller assigns different users to different time
tion (CPM) technique (4). slots, and the assigned time slot is retained by that user until
the user releases it. At the receiving end, a user terminal syn-
• Implementation is simple due to technological advances.
chronizes to the TDMA signal frame and extracts the time
The disadvantages associated with a fixed-assignment
FDMA are:
1 S1
• It needs to back-off the transmitter high power amplifier
i Slotting Channel Deslotting Si
(HPA) from saturation point to avoid intermodulation
M
(Satellite repeater) SM
. Multiplexing Demultiplexing
.
.
f5 i = source i
Si = sink i
Frequency band 3
Frequency
f4
Guard band
f3
Frame k Frame k + 1
Frequency band 2
f2 1 2 i M 1 2 i M Time
Guard band
f1
Frequency band 1
Preamble
f0
Time One time slot
Figure 1. Illustration of frequency-division multiple access tech- Figure 3. Illustration of fixed-assignment time-division multiple ac-
nique. cess technique.
Frequency cept (1, Fig. 9.14, p. 491). For CDMA, the system operates
. . .
. . . simultaneously over the entire system frequency bandwidth
. . . and system time. In CDMA systems the users are kept sepa-
Signal Signal Signal rate by assigning each of them a distinct user-signal code.
Band 3 ...
3 1 3 The design of these codes is usually based on spread-spectrum
(SS) signaling to provide sufficient degrees of freedom to sepa-
rate different users in both time and frequency domains (al-
though the use of SS does not imply CDMA). SS technique
Band 2 Signal Signal Signal ...
2 3 2 can be classified into two categories, namely, direct-sequence
SS (DS-SS) and frequency-hopping SS (FH-SS) (7). Hence,
CDMA can also categorize into DS-CDMA and FH-CDMA. In
CDMA systems a hybrid combination of DS and FH for
Band 1 Signal Signal Signal ... CDMA is also allowed. In the following, brief descriptions of
1 2 1
Time the DS-CDMA and FH-CDMA are given.
Slot 1 Slot 2 Slot 3

Direct-Sequence CDMA. In DS-CDMA systems each of N
Figure 4. Illustration of code-division multiplexing. users is preassigned its own code, PNi(t), where i ⫽ 1, 2, 3,
. . ., N. The user codes are selected such that they are ap-
proximately orthogonal to each other. This means that the
slot assigned to that user. Figure 3 illustrates the demul- cross-correlation of two different codes is approximately zero;
tiplexing procedure for a fixed-assignment TDMA system. that is
The advantages of a fixed-assignment TDMA include:

Tc 1, for i = j
• When used with a constant modulation scheme, the PNi (t)PN j (t) dt ≈ (1)
transmitter HPA can operate at saturation. This means 0 0, for i = j
it is power efficient.
• It is flexible due to variable bit rates allowed for users. where TC denotes the time duration of the code and usually is
referred to as the chip duration. Since the assigned codes are
• VLSI technology can be used for low cost in volume pro-
orthogonal to each other, they can be spread over the entire
duction.
spectrum of the communication resource simultaneously.
• TDMA utilizes bandwidth more efficiently because no The modulated signal for user 1 is denoted as
frequency guard band is required between the channels.
S1 (t) = A1 (t) cos[ω0t + φ1 (t)] (2)
The disadvantages associated with fixed-assignment
TDMA are (2):
where A1(t), 웆o, and ␾1(t) are the amplitude, angular fre-
• TDMA requires higher peak power than FDMA. This quency, and phase, respectively, of the signal specified for
may cause significant drawback for mobile applications user 1. Note that the modulated waveform presented in Eq.
due to the shortening of battery life. (2) is expressed in general form, without any restriction
placed on modulation type. Then the spread signal is obtained
• Complicated signal processing is used in the detection
by multiplying signal S1(t) with the code PN1(t), and the resul-
and synchronization with a time slot.
tant signal, S1(t)PN1(t), is then transmitted over the channel.
Figure 5 shows a simplified block diagram for a typical CDMA
Code-Division Multiple Access
system (1, Fig. 10.25, p. 572). Here the bandwidth of the code
Code-division multiple access (CDMA) is a hybrid combina- PN1(t) is much larger than the unspread signal S1(t). If one
tion of FDMA and TDMA (1,6). Figure 4 illustrates this con- denotes the code rate for PN1(t) as Rc and the signal data rate
T gi(t)gj(t) = 1 for i = j
0 0 for i ≠ j
Carrier g12(t)s1(t)
A cos ω 0t gN(t)sN(t) g1(t)g2(t)s2(t)
g2(t)s2(t) .
. .
Modulated signal .. .
s1(t) = A1(t) cos [ω 0t + φ 1(t)] g1(t)s1(t) g1(t)gN(t)sN(t)
Modulator × Σ ×
To conventional
demodulator
Information Code Code

g1(t) g1(t)
Figure 5. Illustration of code-division

Transmitter Receiver
multiple access technique.
as Rs, then the processing gain Gp of the system is given by • There is greater resistance to interference effects in a fre-
(1,7) quency reuse situation.
R • More flexibility is possible because there is no require-
c ment on time and frequency coordination among the var-
Gp (dB) = 10 log (3)
Rs ious transmitters.
The processing gain provides an indication of how well the The disadvantages of DS-CDMA are:
signal S1(t) is being protected from interfering signals (inten-
tional or unintentional). The larger the value of Gp, the better • It requires power control algorithms due to the near–far
the protection the code can provide. problem.
The spread signal S1(t)PN1(t) is received in the presence • Timing alignments must be within a fraction of a coded
of other spread signals, S2(t)PN2(t), S3(t)PN3(t), . . ., sequence chip.
SN(t)PNN(t). Assuming that the noise at the receiver is zero
• Performance degradation increases as the number of us-
and the signal delays are negligible, we can write the received
ers increases.
signal R(t) as

N Frequency-Hopping CDMA. An alternative to DS-CDMA is
R(t) = S1 (t)PN1 (t) + Si (t)PNi (t) (4) FH-CDMA (1,7). In FH-CDMA systems each user is assigned
i=2 a specific hopping pattern, and if all hopping patterns as-
signed are orthogonal, the near–far problem will be solved
Here we will also assume that the receiver is configured to (except for possible spectral spillover from a specified fre-
receive messages from user 1 so that the second term shown quency slot into adjacent slots). In practice, the codes as-
in Eq. (4) is an interference signal. To recover the signal signed for these hopping patterns are not truly orthogonal;
S1(t), the received signal R(t) is despread by multiplying R(t) thus, interference will result when more than one signal uses
with the code PN1(t) stored at the receiver, the same frequency at a given instant of time. A simplified
block diagram for a typical FH-CDMA modulator is shown in

N
Fig. 6 (1, Fig, 9.15, p. 492).
R(t)PN1 (t) = S1 (t) + Si (t)PNi (t)PN1 (t) (5)
i=2
FH-CDMA can be classified as fast FH-CDMA or slow FH-
CDMA. Fast FH systems use a single frequency hop for each
Here we have used the property PN12(t) ⫽ 1. If we chose the transmitted symbol. This means that, for fast FH systems,
code to have the orthogonal property, that is, the codes are the hopping rate equals or exceeds the information symbol
chosen to satisfy the condition expressed in Eq. (1), then it rate. On the other hand, slow FH systems transmit two or
can be shown that the undesired signal expressed in the sec- more symbols in the time interval between frequency hops.
ond term of Eq. (5) is negligible (7,8). Since the codes are not The advantages associated with FH-CDMA include:
perfectly orthogonal, the second term of Eq. (5) is negligible
for a limited number of users. The performance degradation • Multiple users can share the communication resources,
caused by the crosscorrelation in the second term sets the both frequency and time, simultaneously.
maximum number of simultaneous users. A rule of thumb for • Communication privacy is possible due to assigned codes
determining the maximum number of users N appears to be being known only to the users.
that (7) • There is an inherent robustness against mobile channel
degradations such as fading and multipath (7–10).
10G p (dB)/10
N≈ (6)
10
While the code design is of paramount importance, of po-

tential greater importance in DS-CDMA is the so-called
near–far problem (7,9,10). Since the N users are usually geo- Data
pulse stream Modulator
graphically separated, a receiver is trying to detect the ith x(t) sx(t) = A cos [ ω 0(t) + ∆ ω 0(t)]t
user, which is much farther than the jth user. Therefore, if
each user transmits with equal power, the power received by A cos ω 0(t)t
the jth user would be much stronger than that received by
the ith user. This particular problem is often so severe that Frequency
hopper
DS-CDMA systems will not work without appropriate power
control algorithms.
...
Advantages associated with DS-CDMA include:
PN code
• Multiple users can share the communication resources, generator
both frequency and time, simultaneously.
• Communication privacy is possible due to assigned codes
being known only to the users. Clock
• There is an inherent robustness against mobile channel Figure 6. Illustration of code-division multiple access frequency
degradations such as fading and multipath (7–10). hopping.
• There is an inherent robustness against interference.

• The near–far problem does not exist.
• Network implementation for FH-CDMA is simpler than
DS-CDMA systems because the required timing align-
ments must be within a fraction of a hop duration as
compared to a fraction of a chip duration.
• It performs best when a limited number of signals are
sent in the presence of nonhopped signals.
The disadvantages are:
• Performance degradation is possible due to spectral spill-

over from a specified frequency slot into adjacent slots.
• Frequency synthesizer can be very costly.
• As the hopping rate increases the reliability decreases
and synchronization becomes more difficult.
Space-Division Multiple Access Figure 8. Illustration of beamforming-based space-division multi-

ple access.
For wireless applications, space-division multiple access
(SDMA) can be classified into cell-based and beamforming-
based SDMA. The difference between the two approaches can where M is the total number of frequency channels, K is the
best be illustrated in Fig. 7 (11, Fig. 1.1, p. 4) for cell-based cell reuse factor and S is the number of sectors in a cell. K
SDMA and Fig. 8 (11, Fig. 1.2, p. 5) for beamforming-based can be expressed as
SDMA.
A primitive form of SDMA exists when frequency carriers 1
D 2
are reused in different cells separated by a special distance K= (8)
3 R
to reduce the level of co-channel interference. The larger the
number of cells the higher the level of frequency reuse and where D is the distance between two co-channel cells and R
thus the higher capacity that can be attained. This has re- is the cell radius. The corresponding average signal-to-inter-
sulted in cell-based SDMA, which has been predominant for ferer ratio (SIR) can be calculated for different types of sec-
quite a long time. toring systems, including adaptive beamforming with several
In a frequency reuse system, the term radio capacity is beams in beamforming-based SDMA.
used to measure the traffic capacity, and is defined as The system benefits of beamforming-based SDMA include:
M
Cr = (7) • Improvement of multipath fading problems since nar-
K ·S rower beams are used and the implicit optimal diversity
combining performed by the beamformer
• More flexible coverage of each base station to match the
local propagation conditions
Table 1 lists the capacity and SIR for several SDMA configu-
rations (12).
Adaptive beamforming algorithms require a certain refer-
ence signal in the optimization process. If the reference signal
is not explicit in the received data, blind adaptive beamform-
ing (BAF) can be used instead. For digital communication sig-
nals, one can vary certain signal properties such as constant
modulus applicable to FSK or PSK signals to result in the
Table 1. Radio Capacity and Signal-to-Noise

Ratio for Different Cells
Capacity
K S (Channels/Cell) SIR (dB)
Omnicells 7 1 M/7 18
120⬚ sectorial cells 7 3 M/21 24.5
Figure 7. Illustration of the cell-based space-division multiple ac- 60⬚ sectorial cells 4 6 M/24 26
cess. A different set of carrier frequencies is used in each of the sec- 60⬚ sectorial beams 7 6 3M/7 20
tors. These frequencies are used in other sectors of other cell sites. N adaptive beams 7 1 MN/7 18
The frequency reuse pattern is selected to minimize the interference.
;;;;
ting short messages. In this case, the random-access methods
are more flexible and efficient than the fixed-access methods.
This section discusses the three basic random-access schemes,
Whip namely, pure ALOHA, modified ALOHA (slotted and reserva-
Medium Loop tion), and carrier-sense multiple access with collision de-
tection.
Pure ALOHA
Pure ALOHA (P-ALOHA), or basic ALOHA, was developed at
Figure 9. Illustration of horizontal and vertical polarization diver- the University of Hawaii in 1971 with the goal of connecting
sity signals. several university computers by the use of random-access pro-
tocol (17). The system concept is very simple and has been
summarized by Sklar (1). The algorithm is listed below for
constant modulus adaptive beamforming algorithm (13), or future comparison with the enhanced version, the so-called
the cyclostationary properties of bauded signals to suggest slotted ALOHA.
the spectral self-coherence restoral (SCORE) algorithm (14).
Another method (15) that can be considered blind adaptive
• Mode 1: Transmission Mode. Users transmit at any time
beamforming is based on decision-directed equalization to
they desire, encoding their transmissions with an error
combat intersymbol interference (ISI) in digital communica-
detection code.
tions. Using this concept, a BAF demodulates the beamformer
output and uses it to make a decision in favor of a particular • Mode 2: Listening Mode. After a message transmission,
value in the known alphabet of the transmit sequence. A ref- a user listens for an acknowledgment (ACK) from the re-
erence signal is then generated based on the modulated out- ceiver. Transmissions from other users will sometimes
put of this decided demodulated beamformer output. overlap in time, causing reception errors in the data in
each of the contending messages. We say the messages
Polarization-Division Multiple Access have collided. In such cases, the errors are detected, and
the users receive a negative acknowledgment (NACK).
Signals transmitted in either horizontal or vertical electric
field are uncorrelated at both the mobile and base station’s • Mode 3: Retransmission Mode. When a NACK is re-
receiver. Suppose that a received vertically polarized signal is ceived, the messages are simply retransmitted. Of
course, if the colliding users were retransmitted immedi-

N ately, they would collide again. Therefore, the users re-
11 = ai e jψ i e− jβV t cos φ i (9) transmit after a random delay.
i−1
• Mode 4: Timeout Mode. If after a transmission, the user
and the received horizontally polarized signal is does not receive either an ACK or NACK within a speci-
fied time, the user retransmits the message.

N

22 = a i e jψ i e− jβV t cos φ i (10)
i=1 Figure 10 shows the concept of the pure ALOHA algorithm
(6, Fig. 11.15, p. 465).
where ai and ␺i are the amplitude and phase, respectively, for
each wave path and a⬘i and ␺i⬘ are their counterparts in Eq.
(9), V is the vehicle velocity, and ␾i is the angle of arrival of Modified ALOHA
the ith wave. Although these two polarized waves arrived at
In order to improve the pure ALOHA algorithm, the slotted
the receiver from the same number of incoming waves, it is
(18) and reservation ALOHA algorithms (19) have been pro-
not difficult to see that ⌫11 and ⌫22 are uncorrelated because
posed. Based on the summary described in Sklar (1), a brief
of their different amplitudes and phases. Thus, a PDMA sys-
tem can be illustrated as in Fig. 9 (16, Fig. 9-6, p. 281). In description of these algorithms will be given here.
this system, the base station can be two vertical and hori-
zontal dipoles and the antenna at the mobile can be a pair of
whip and loop antennas.
Packet B Packet C
RANDOM-ASSIGNMENT MULTIPLE ACCESS
Tp Tp
Fixed-assignment multiple access is most efficient when each
user has a steady flow of information for transmission. How- Packet A
t
ever, this method becomes very inefficient when the informa- t0
tion to be transmitted is intermittent or bursty in nature. As
an example, for mobile cellular systems, where the subscrib- Vulnerable period
for pure ALOHA
ers pay for service as a function of channel connection time,
fixed-assignment access can be very expensive for transmit- Figure 10. Illustration of collision mechanism in pure ALOHA.
Satellite time • Users send message packets only in their assigned por-
tion of the M slots.
Figure 12 shows an example of the R-ALOHA system (1, Fig.
m retransmits
n retransmits
Collision 9.22, p. 503). In this example, the users seek to reserve 3 slots
Send packet
with M ⫽ 5 slots and V ⫽ 6 subslots. Compared with S-

ACK
NAK
ALOHA, R-ALOHA is very efficient for high traffic intensity.
User time Carrier-Sense Multiple Access

User k Users m, n User m User n To improve the previous algorithms and to make efficient use
of the communications resources, the user terminal listens to
Figure 11. Illustration of slotted ALOHA.
the channel before attempting to transmit a packet. This pro-
tocol is called listen-before-talk and usually is referred to as
carrier-sense multiple access (CSMA) protocol (12). This algo-
Slotted ALOHA. The operation of the slotted ALOHA (S- rithm is widely used in both wired and wireless local area
ALOHA) is illustrated in Fig. 11 (1, Fig. 9.21, p. 501). A se- networks (LANs), where the transmission delays are low.
quence of synchronization pulses is broadcast to all users for There are several modified versions of CSMA, namely, CSMA
coordination among the users. Messages are sent through with busy-tone signaling, CSMA with collision detection, and
data packets with constant length between the synchroniza- CSMA with collision avoidance. In addition, there is another
tion pulses and can be started only at the beginning of a time modified version of CSMA, called data-sense multiple access
slot. This modification reduces the rate of collisions by half, (DSMA), which has been developed and adopted for use in
since only a packet transmitted in the same slot can interfere wireless packet data networks such as cellular digital packet
with one another (1). In S-ALOHA systems the users retrans- data (CDPD).
mit after a random delay of an integer number of slot times This section describes the three basic CSMA schemes,
when a NACK occurs. namely, 1-persistent CSMA, nonpersistent CSMA, and p-per-
sistent CSMA. Modified versions of CSMA will also be de-
Reservation ALOHA. Significant improvement can be scribed briefly.
achieved with the reservation ALOHA (R-ALOHA) scheme.
This scheme has two modes, namely, an unreserved mode and 1-Persistent Carrier-Sense Multiple Access. 1-Persistent car-
reserved mode. rier-sense multiple access (1-P CSMA) is the simplest form
The unreserved or quiescent mode, mode has three stages: of CSMA. In the basic form, 1-P CSMA is unslotted. The ‘‘1-
persistent’’ signifies the strategy in which the message is sent
• A time frame is formed and divided into several reserva- with probability 1 as soon as the channel is available. After
tion subslots. sending the packet, the user station waits for an ACK, and if
• Users employ these small subslots to reserve message none is received in a specified amount of time, the user will
slots. wait for a random amount of time and then resume listening
• After requesting a reservation, the users listen for an to the channel. When the channel is sensed idle, the packet
ACK and slot assignment. is retransmitted immediately. In unslotted form, the system
does not require synchronization between the user stations
and all transmissions are synchronized to the time slots. In
The reserved mode has four stages:
contrast with the unslotted form, the slotted 1-P CSMA re-
quires synchronization among all user stations and all trans-
• The time frame is divided into M ⫹ 1 slots whenever a
missions, whether initial transmissions or retransmissions,
reservation is made.
are synchronized to the time slots (1,6).
• The first M slots are used for message transmissions.
• The last slot is subdivided into subslots to be used for Nonpersistent Carrier-Sense Multiple Access. The main dif-
reservation/requests. ference between the 1-P CSMA and nonpersistent carrier-
M V First available slot

for reserved packets
Slots Subslots
Satellite time
t
t AC lo
u es K s ts Quiescent
eq fir state
R d
S en
0 5 10 15 20 25 30 35
Figure 12. Illustration of reservation
ALOHA. Station seeks to reserve 3 slots 1 round trip User time
(M ⫽ 5 slots, V ⫽ 6 subslots).
sense multiple access (NP CSMA) is that a user station does Carrier-Sense Multiple Access with Collision Avoidance. The
not sense the channel continuously while it is busy. Instead, carrier-sense multiple access with collision avoidance (CSMA/
after sensing the busy condition, the NP CSMA system waits CA) technique is widely used in many WLANs. The specific
a randomly selected interval of time before sensing again. collision avoidance strategy for CSMA/CA is different from
This random waiting time associated with NP CSMA could one manufacturer to another. In one system, CSMA/CA is re-
eliminate most of the collisions that would result from multi- ferred to as CSMA with an exponential back-off strategy and
ple users transmitting simultaneously upon sensing the tran- an acknowledgment scheme. Note that the exponential back-
sition from the busy to idle condition. off strategy is referred to as a collision avoidance mechanism.
Other systems can employ R-ALOHA as the collision avoid-
p-Persistent Carrier-Sense Multiple Access. The p-persistent ance strategy.
carrier-sense multiple access (pP CSMA) is a generalization
of the 1-P CSMA scheme, which is applicable to slotted chan- Data-Sense Multiple Access. Digital or data sense multiplex
nels. In this scheme, the slotted length is chosen to be the access (DSMA) is commonly used in full-duplex wireless data
maximum propagation delay. In this system, a message is communication networks such as CDPD and trans-European
sent from a station with probability p when the channel is trunked radio (TETRA) (6). In these systems, communications
sensed to be idle. With probability q ⫽ 1 ⫺ p the station de- from the mobile to base (also referred to as reverse channel
fers action to the next slot, where the station senses the chan- or uplink) and from base to mobile (also referred to as forward
nel again. If the next slot is idle, the station transmits with channel or downlink) are performed on different frequency
probability p or defers with probability q. This procedure is channels using different access techniques. The downlink
repeated until either the whole frame has been transmitted uses TDMA, while the uplink uses DSMA. Interleaved among
or the channel is sensed to be busy. If the channel is busy, other signals broadcast on the downlink, the base station
the station monitors the channel continuously until it be- transmits a busy–idle bit in each time frame to report the
comes free; then it starts the above procedure again (6). status of the uplink channel. A mobile terminal will check
this flag bit before transmission. If this bit indicates idle
Carrier-Sense Multiple Access with Busy-Tone Signaling. In channel, the terminal proceeds to send its packet in the fol-
wireless networks, the user terminals are not always within lowing time slot. As soon as the transmission starts, the base
the range and line-of-sight of each other, and when this situa- station switches the flag bit to busy state until the transmis-
tion occurs, it is referred to as ‘‘hidden terminal problem.’’ sion from the mobile terminal is completed.
This problem can be solved by using the carrier-sense multi-
ple access with busy-tone signal (CSMA/BTS) technique (6). Polling Technique
This technique divides the system bandwidth into two chan-
The polling technique is a form of ‘‘control’’ random-assign-
nels: a message channel and a busy-tone channel. The scheme
ment multiple access. In systems using this technique one
works as follows. Whenever the central station senses signal
station is used as a controller that periodically polls all the
energy on the message channel, it transmits a simple busy-
other stations to determine if they have data to transmit (17).
tone signal on the busy-tone channel, and this tone is detect-
Note that in R-ALOHA the control is distributed among all
able by all the user stations. With this technique, a user sta-
user terminals, while the polling technique utilizes central-
tion first senses the channel by detecting the busy-tone signal
ized control.
to determine if the network is busy. The procedure the user
Based on Refs. 6 and 20, a brief description of this tech-
station then follows in transmission of the message depends
nique is given here. Usually the controller station in the sys-
on the particular version of CSMA being used in the network,
tem is given a polling, instructing the order in which the ter-
and any of the CSMA techniques described earlier can be
minals are poled. If the polled station has something to
chosen.
transmit, it starts transmission. If not, a ‘‘negative reply’’ or
‘‘no reply’’ is detected by the controller, which then polls the
Carrier-Sense Multiple Access with Collision Detection. The
next terminal in the sequence. This technique is efficient only
carrier-sense multiple access with collision detection (CSMA/
if (1) the round-trip propagation delay is small (due to con-
CD) technique, also referred to as the ‘‘listen-while-talk’’
stant exchange of control messages between the controller
(LWT) technique, can be used with 1-P CSMA, NP CSMA, or
and terminals), (2) the overhead due to the polling message
pP CSMA, each with a slotted or unslotted version (6). In the
is low, and (3) the user population is not large and bursty.
operation of CSMA/CD, if the channel is detected to be idle
or busy, a user station first sends the message (in the form
Token Passing
of data packets) using the procedure dictated by the selected
protocol in use. While sending the packets, the user station This technique is another form of controlled random-assign-
keeps monitoring the transmission; it stops transmission, ment multiple access and it has been used widely in wired
aborting the collided packets and sending out a jamming sig- local area networks (LANs) for connecting computers. How-
nal, as soon as it detects a collision. The retransmission back- ever, this scheme is not very popular in wireless networks. In
off procedure is initiated immediately after detecting a colli- this system a ring or loop topology is used. Figure 13 illus-
sion. The purpose of the jamming signal is to force consensus trates a typical token-ring network (1, Fig. 9.42, p. 529). As
among users as to the state of the network, in that it ensures shown in Fig. 13, in the token-ring network, messages are
that all other stations know of the collision and go into back- passed from station to station along unidirectional links, until
off condition. Design of a proper back-off algorithm to ensure they return to the original station. This scheme passes the
stable operation of the network is an important topic for com- access privilege sequentially from station to station around
munications design engineers. the ring. Any station with data to send may, upon receiving
Interface by (6)
Ring 1-bit listen mode
interface delay DelayFDMA = T (12)
T
1

Station DelayTDMA = DelayFDMA − 1− (13)
2 M
To From
station station Therefore, based on Eq. (13), TDMA is superior to FDMA with
Unidirectional respect to the average delay packet when there are two or
ring Interface more users. It is interesting to note that for larger numbers
transmit mode
of users the average delay packet of TDMA is half that of
FDMA.
Spurious Narrowband Interference. An FDMA system out-

To From performs a TDMA system in the presence of spurious nar-
station station rowband interference. In an FDMA system, where the format
is a single user per channel per carrier, the narrowband inter-
Figure 13. Illustration of the token-ring network.
ference can impair the performance of only one user channel.
On the other hand, in a TDMA system, a narrowband inter-
ference signal can cause performance degradation to all user
the token, remove the token from the ring, send its message, channels in the TDMA data stream.
and then pass on the control token.
Pure ALOHA versus Slotted ALOHA
If one defines the throughput S as the number of successfully
PERFORMANCE COMPARISON OF delivered packets per packet transmission time Tp, and G is
MULTIPLE ACCESS TECHNIQUES the offered traffic load in packets per packet time, then the
throughputs for P-ALOHA and S-ALOHA, respectively, are
This section compares various standard mutiple access tech- given by (1,6)
niques such as FDMA, TDMA, pure ALOHA, slotted ALOHA,
SP−ALOHA = Ge−2G (14)
unslotted/slotted 1-P CSMA, and unslotted/slotted NP
CSMA. SS−ALOHA = Ge−G (15)
FDMA versus TDMA The maximum throughput S occurs at
Comparison between FDMA and TDMA schemes is not 1

SP−ALOHA(max) = = 0.18 (16)
straightforward. It involves several issues such as bit error 2e
rate (BER) performance, throughput performance, system de- 1
lay, and implementation. Some of these issues have greater SS−ALOHA(max) = = 0.37 (17)
e
or lesser performance depending on the type of system in
which the access method is to be employed. In this section we at the values of G ⫽ 0.5 and 1, for P-ALOHA and S-ALOHA,
briefly describe some of the major issues of comparison be- respectively. This means that for a P-ALOHA channel, only
tween FDMA and CDMA. 18% of the communication resource can be utilized. Compar-
ing Eqs. (16) and (17), we see that, for S-ALOHA, there is an
improvement of two times the P-ALOHA. A plot of P-ALOHA
Bit Rate Capability. If one neglects all overhead elements
(or ALOHA) and S-ALOHA is shown in Fig. 14 (6, Fig. 11.19,
such as guard bands in FDMA and guard time in TDMA, then
p. 473).
the data rate capability for both systems is identical. The ef-
fective data rate is given by
1.0
Mb
R= (11) 0.8 Slotted ALOHA
T
1-persistent CSMA,
Throughput, S
both slotted
where M is the number of disjoint channels and b is the num- 0.6 and nonslotted Slotted nonpersistent
ber of data bits transmitted over T seconds. CSMA
0.4
ALOHA
Message Packet Delay. The message packet delay is defined Nonpersistent
0.2 CSMA
as the packet waiting time before transmission plus packet
transmission time. If we let M be the number of users gener- 0
ating data at a constant uniform rate of R/M bits/s, and use 0.1 1.0 10 100
FDMA and TDMA systems that each transmit a packet of N Offered load, G
bits every T seconds, then one can show that the average Figure 14. Throughput performance comparison of multiple access
packet delays for FDMA and TDMA, respectively, are given techniques.
1-P CSMA versus NP CSMA

Again, using the same definition for the throughput, and let- B
ting a be the normalized propagation delay, we have G C
τ
a= (18) A
Tp
F D
The parameter a described here corresponds to the time inter- B E
val, which is normalized to the packet duration, during which
a transmitted packet can suffer a collision in the CSMA G C B
schemes. Note that practical values of a on the order of 0.01
A G C
are usually of interest. The throughput for unslotted 1-P
CSMA is found to be (6) F D A
aG
E F D
G 1 + G + aG 1 + G + e−G(1+2a)
2
SUnslot−1P = (19) E
G(1 + 2a) − (1 − e−aG ) + (1 + aG)e−G(1+a)
For slotted 1-P CSMA, Figure 15. Illustration of the cellular frequency reuse concept. Cells
with the same letter use the same set of frequencies. A cell cluster is
G[1 + a − eaG ]e−G(1+a) outlined in bold and replicated over the coverage area.
SSlot−1P = (20)
(1 + a)(1 − e−aG ) + ae−G(1+a)
replacing the single high power transmitter representing a
For unslotted NP-CSMA, large cell with several small low powered transmitters as in
small cells. Figure 15 (21, Fig. 2.1, p. 27) illustrates the ar-
Ge−aG
SUnslot−NP = (21) rangement of the smaller cells to achieve frequency reuse in
G(1 + 2a) + e−aG the allocated frequency band where cells labeled with the
same letter use the same group of channels. The hexagonal
For slotted NP-CSMA, shape of each cell serves to model the conceptual and idealis-
tic boundary of each cell in terms of coverage and would be
aGe−aG much more irregular in a real environment due to differing
SSlot−NP = (22)
1 + a − e−aG propagation effects and practical consideration in base sta-
tion placement.
The plots of Eqs. (19), (20), (21), and (22), for a ⫽ 0.01, are
shown in Fig. 14 (6, Fig. 11.19, p. 473). This figure shows Cellular Telephone System Terminology. Figure 16 (22, Fig.
that, for low levels of offered traffic, the persistent protocols 1.5, p. 15) shows a basic cellular telephone system consisting
provide the best throughput, but for higher load levels, the of mobile stations, base stations, and a mobile switching service
nonpersistent protocols are by far the best. The figure also center (MSC), sometimes called a mobile telephone switching
shows that the slotted NP-CSMA protocol has a peak office (MTSO). The function of the MSC is to provide connec-
throughput almost twice that of 1-P CSMA schemes. tivity to all mobile units to the public switched telephone net-
work (PSTN) in a cellular system. Each mobile unit communi-
APPLICATIONS OF RANDOM-ACCESS TECHNIQUES
IN CELLULAR TELEPHONY
The objective for earlier mobile radio systems was to achieve

a large coverage area by using a high-powered transmitter
with antenna on a tall tower to extend the receiving area. The
extensive coverage from this approach has also resulted in
limited user capacity capability, since increasing frequency
reuse would certainly increase interference for the users of
the system. At the same time, government regulatory agen-
cies are not able to allocate frequency bands in a timely man-
ner to keep up with demand for wireless services. It is there-
fore necessary to construct a mobile radio system to achieve
both high capacity and large coverage area with the con-
straint of a crowded radio spectrum. MSC PSTN
Cellular Communications Concept

Figure 16. Illustration of a cellular system. The towers represent
The cellular communications concept was developed to pro- base stations, which provide radio access between mobile users and
vide the solution for spectral congestion and user capacity by the mobile switching center.
cates with the base and may be handed off to any other base Overview of Cellular Systems
stations during the call.
Since the world’s first cellular system was implemented by
The mobile unit handset contains a transceiver, antenna,
Nippon Telephone and Telegraph (NTT) and deployed in Ja-
and control unit, whereas a base station consists of several
pan in 1979, many other systems have been developed in
transmitters and receivers to handle simultaneous full duplex
other countries. Tables 2–4 list the cellular systems in three
calls. The base station typically consists of a tower, with mul- major geographical areas of the world—North America, Eu-
tiple antennas for receiving transmitting RF signals, and as- rope, and Japan.
sociated electronics at the base. The communication lines be- In the United States, the Advance Mobile Phone System
tween the base station and the MSC can be regular telephone (AMPS) was introduced in 1983 by AT&T as the first major
and point-to-point microwave links. The typical MSC handles analog cellular system based in FDMA technology. By 1991,
the routing, billing, and system maintenance functions of the TIA (Telecommunication Industry Standard) IS-54B digi-
calls to and from the mobile units, and multiple MSCs can be tal standard was developed to allow US cellular operators to
used together by a wireless operator. ease the transition of analog cellular phone to an all digital
The communication between the base station and mobile system using TDMA. To increase capacity in large AMPS
units is defined by a standard common air interface (CAI). markets, Motorola developed the narrowband AMPS (N-
The CAI typically specifies the communication parameters, AMPS) that essentially provides three users in the 30 kHz
such as multiple access methods and modulation type, and bandwidth AMPS standard and thus reduces voice quality.
the use of four different channels for data transmission. From By 1993, a cellular system based on CDMA was developed by
the base station to the mobile unit, the forward voice channel Qualcomm Inc. and standardized as TIA IS-95. At the same
(FVC) is used for voice transmission and the forward control time as IS-95, cellular digital packet data (CDPD) was intro-
channel (FCC) is used for initiating and controlling mobile duced as the first data packet switching service that uses a
calls. From the mobile unit to the base station, the reverse full 30 kHz AMPS channel on a shared basis and utilizes slot-
voice channel (RVC) and the reverse control channel (RCC) ted CSMA/CD as the channel access method. The auction of
accomplish the same functionality as the forward channel, the 1900 MHz PCS band by the US government in 1995 opens
only in the other direction to ensure full duplex communica- the market for other competing cellular standards, such as
tions. the popular European GSM standard, which is implemented
All cellular systems provide roaming service for a cellular in the DCS-1900 standard.
subscriber who uses the mobile unit in a service area other In the United Kingdom, the E-TACS (Extended European
than the one area subscribed to by the mobile user. The regis- Total Access Cellular System) was developed in 1985 and is
tration of a roamer is accomplished by the MSC using the virtually identical to the US AMPS system except for the
FCC to ask for all mobile units, which are not registered to smaller voice channel bandwidth. The Nordic Mobile Tele-
report their MIN, and ESN reported over the RCC. This infor- phone (NMT) system in the 450 MHz and 900 MHz bands
mation is then used for validation as well as billing purposes. was developed in 1981 and 1986 using FDMA technology and
was deployed in the Scandinavian countries. In Germany, a
cellular standard called C-450 was introduced in 1985. Be-
The Process of a Cellular Call. When a mobile unit is first cause of the need to standardize over these different cellular
powered up, it scans for a group of forward control channels systems in Europe, the GSM (Global System for Mobile) was
to find the strongest available one to lock on and changes to first deployed in 1991 in a new 900 MHz band dedicated as
another channel when the signal level drops below a specified the cellular frequency band throughout Europe.
level. The control channels are standardized over a geo- In Japan, JTACS and NTACS (Narrowband and Japanese
graphic area. The standard ensures that the mobile unit will Total Access Communications System) are analog cellular
be using the same control channel when ready to make a systems similar to AMPS and NAMPS. The Pacific Digital
phone call. Cellular (PDC) standard provides digital cellular coverage us-
Upon initiating a phone call on the reverse control channel ing a system similar to North America’s IS-54.
using the subscriber’s telephone number (mobile identification
number or MIN), electronic serial number (ESN), called tele- Major Cellular Systems
phone number, and other control information, the base sta-
tion relays this information to the MSC, which validates the Currently, only a few of the cellular standards have survived
request and makes the connection to the called party through or been developed into major systems around the world in
the PTSN or through another MSC in the case of a called terms of the number of users. These major systems are briefly
mobile unit. Once the appropriate full duplex voice channels described in this section.
are allocated, the connection is established as a phone call.
For a call to a mobile from a PSTN phone, the MSC dis- AMPS and ETACS. In AMPS and ETACS, the FCC (forward
patches the request to all base stations in the cellular system. control channel) continuously transmits control messages
Then the base stations, using a paging message, broadcast data at 10 kbit/s (8 kbit/s for ETACS) using binary FSK with
the called telephone number (or MIN) over the forward con- a spectral efficiency of 0.33 bit/s/Hz. When a voice call is in
trol channel. When the mobile unit receives the paging mes- progress, three in-band SATs (supervisory signal tones) at
sage, it responds to the base station by identifying itself over 5970 Hz, 6000 Hz, or 6030 Hz serve to provide a handshake
the reverse control channel. The base station relays this infor- between the mobile unit and base station. Other control sig-
mation to the MSC, which then sets up the appropriate voice nals are bursty signaling tone (ST) on the RVC to indicate
channels and connection for the call. end of call, and blank-and-burst transmission in the voice
Table 2. Cellular Standards in North America

Multiple Frequency Band Channel
Year of Access (MHz), Data/Control Bandwidth
Standard Introduction Technique Reverse/Forward Parameters (kHz)
AMPS 1983 FDMA 824–849/869–894 FM/10 kbps FSK 30
IS-54 1991 TDMA 824–849/869–894 48.6 kbps 앟/4DQPSK/ 30
10 kbps FSK
NAMPS 1992 FDMA 824–849/869–894 FM/10 kbps FSK 10
CDPD 1993 FH/Packet 824–894 GMSK (BT ⫽ 0.5) 30
19.2 kbps
IS-95 1993 CDMA 824–894, 1.8–2.0 GHz QPSK/BPSK 1.25
DCS-1900 (GSM) 1994 TDMA 1.85–1.99 GHz GMSK 200
band having a duration less than 100 ms so as not to affect (FACCH), which carry various control messages to effect
voice quality. power control and call processing.
Prior to frequency modulation, voice signals are processed The USDC voice channel occupies the 30 kHz bandwidth
using a compander, a pre-emphasis filter, a deviation limiter, in each of the forward and reverse links and uses a TDMA
and a postdeviation limiter filter. These steps are taken to scheme with six time slots to support a maximum of three
accommodate a large speech dynamic range, to prevent spuri- users. For full-rate speech, each user is assigned two time
ous emission, and to minimize interference with the in-band slots in an equally spaced fashion as compared to one slot per
SAT signal. The channel coding on the forward and reverse user for half-rate speech.
control channels is BCH(40, 28) on FCC and BCH(48,36) on The speech coder used in IS-54 is called the vector sum
RCC. The line code used is Manchester. excited linear predictive (VSELP) code and is based on a code
book that determines how to quantize the residual excitation
IS-54. The analog AMPS system was not designed to sup- signal. The VSELP coder has an output bit rate of 7950 bps
port the demand for large capacity in large cities. Cellular and can produce a speech frame every 20 ms. The 159 bits
systems using digital modulation techniques potentially offer of speech within a speech frame are divided into two classes
large improvements in capacity and system performance. The according to their perceptual importance. Class 1 of 77 bits,
IS-54 standard, also known as the USDC (US Digital Cellu- being more important, are error protected using a rate con-
lar) was set up to share the same frequencies, the frequency volutional code of constraint length K ⫽ 6, in addition to us-
reuse plan, and base stations as AMPS so that both base sta- ing a 7 bit CRC error detection code on the 12 most significant
bits. Before transmission, the encoded speech data are inter-
tions and subscriber units can be provided with both AMPS
leaved over two time slots with the speech data from adjacent
and USDC channels within the same equipment. This way,
frames. For demodulation, differential detection may be per-
US cellular carriers would be able to gradually replace analog
formed at IF or base band, and equalization is needed based
phone and base stations with digital ones.
on training pattern imbedded in the data.
To maintain compatibility with AMPS phones, USDC for-
The IS-136 standard (formerly known as IS-54 Rev. C), re-
ward and reverse control channels use exactly the same sig-
cently introduced, is an improved version of IS-54. This stan-
naling techniques as AMPS while USDC voice channels use dard comes with the addition of a DQPSK digital control
앟/4 DQPSK at a rate of 48.6 kbit/s and spectral efficiency of channel to the existing FSK control channel, a greatly im-
1.62 bit/s/Hz. proved digital speech coder, new cellular features, and proto-
The numbers of USDC control channels are doubled from col additions to allow greater mobility management and bet-
AMPS to provide flexibility to service offerings such as pag- ter cellular service. The IS-136 protocol can be used in both
ing. There are three types of supervisory channels: the coded the 800 MHz cellular band and the 1900 MHz PCS.
digital verification color code (CDVCC), whose function is sim-
ilar to the SAT in AMPS, and the slow associated control Global Mobile System. Global Mobile System or GSM uti-
channel (SACCH) and fast associated control channel lizes two bands of 25 MHz set aside for system use in all
Table 3. Cellular Standards in Europe

Frequency Band Channel
Year of Multiple Access (MHz), Data/Control Bandwidth
Standard Introduction Technique Reverse/Forward Parameters (kHz)
NMT-450 1981 FDMA 453–457.5/463–467.5 FM/10 kbps FSK 25
E-TACS (UK) 1985 FDMA 872–905/917–950 FM/10 kbps FSK 25
C-450 (Germany, 1985 FDMA 450–455.74/460–465.74 FM 20/10
Portugal)
NMT-900 1986 FDMA 890–915/935–960 FM/10 kbps FSK 12.5
GSM 1990 TDMA and slow FH 890–915/935–960 GMSK (BT ⫽ 0.3) 200
DCS-1800 1993 TDMA 1710–1785/1805–1880 GMSK 200
Table 4. Cellular Standards in Japan

Frequency Band Channel
Year of Multiple Access (MHz), Bandwidth
Standard Introduction Technique Reverse/Forward Data/Control Parameters (kHz)
NTT 1979 FDMA 400/800 FM 25
JTACS 1988 FDMA 860–925 FM/10 kbps FSK 25
PDC 1993 TDMA 810–830 앟/4 DQPSK/10 kbps FSK 25
1429–1453/940–960
1477–1501
NTACS 1993 FDMA 843–925 FM 12.5
member countries. The multiaccess method is a combination ror correction coding, this 2400 bit/s data is sent at 1.4
of TDMA and slow FH. The use of FH combined with in- kbit/s.
terleaving is for mitigation of fading caused by multipath
transmission or interference effects. Frequency hopping is The GSM speech code is based on the residually excited
carried out on a frame-by-frame basis, and as many as 64 linear predictive (RELP) coding, which is enhanced by includ-
different channels may be used before the hopping sequence ing a long-term predictor. The GSM coder takes advantage of
is repeated. the fact that, in a normal conversation, a person speaks less
The available forward and reverse frequency bands are di- than 40% of the time on average. By incorporating a voice
vided into 200 kHz wide channels. There are two types of activity detector (VAD), the GSM system operates in a discon-
GSM channels—traffic channels (TCH), carrying digitally en- tinuous transmission mode, thus providing longer subscriber
coded user speech or data, and control channels (CCH), car- battery life and reduced radio interference when the trans-
rying signaling and synchronizing commands between the mitter is not active during the speech silent period. Channel
base stations and subscriber units. There are three main con- coding for speech and control channels is based on a rate
trol channels in the GSM system—broadcast channel (BCH), convolutional encoder with constraint length K ⫽ 5, whereas
common control channel (CCCH), and dedicated control chan- channel coding for data channels is based on a modified
nel (DCCH). Each control channel consists of several logical CCITT V.110 modem standard.
channels distributed in time to provide the necessary GSM Security is built into GSM by ciphering the contents of the
control function (22). data block with encryption keys known only to the base sta-
Each TDMA frame has 8 time slots for up to eight users tion and the subscriber unit and is further enhanced by
with an aggregate bit rate of up to 24.7 kbit/s per user. The changing encryption algorithm from call to call.
modulation used is 0.3 GMSK. The following full-rate speech
and data channels are supported: IS-95. Similar to IS-54, TIA IS-95 is designed to be com-
patible with the existing US analog system, where base sta-
• Full-rate speech channel (TCH/FS) carries the user tions and mobile units can work in the dual mode operation.
speech digitized at the raw rate of 13 kbit/s. With GSM Since this is a direct-sequence CDMA system, the need for
channel coding applied, the full-rate speech channel is frequency planning within a region is virtually eliminated.
sent at 22.8 kbit/s. Specification for IS-95 reverse link operation is in the 824–
• Full-rate data channel for 9600 bit/s (TCH/F9.6) carries 849 MHz band and forward link operation is in the 869–894
raw user data sent at 9600 bit/s. With GSM forward er- MHz band. The maximum user data rate is 9600 bps and is
ror correction coding, this 9600 bit/s data is sent at 22.8 spread to a channel chip rate of 1.2288 Mchips/s using a com-
kbit/s. bination of techniques. Each mobile subscriber is assigned a
• Full-rate data channel for 4800 bit/s (TCH/F4.8) carries different spreading sequence to provide perfect signal separa-
raw user data sent at 4800 bit/s. With GSM forward er- tion from other users.
ror correction coding, this 4800 bit/s data is sent at 22.8 Unlike other cellular standards, the user data rate but not
kbit/s. the channel chip rate changes in real-time depending on the
voice activity and network requirement. On the forward link,
• Full-rate data channel for 2400 bit/s (TCH/F2.4) carries
the base station simultaneously transmits user data for (or
raw user data sent at 2400 bit/s. With GSM forward er-
broadcast to) all mobile subscribers by using a different
ror correction coding, this 2400 bit/s data is sent at 22.8
spreading code (Walsh functions) for each subscriber. A pilot
kbit/s.
code is also broadcast at a high power level to allow all mo-
• Half-rate speech channel (TCH/HS) carries the user biles to perform coherent carrier detection while estimating
speech digitized at half the rate of full-rate speech. With the channel condition. On the reverse link, all mobiles would
GSM channel coding applied, the full-rate speech chan- respond asynchronously and have a constant signal level due
nel is sent at 11.4 kbit/s. to power control exercised by the base station to avoid the
• Half-rate data channel for 4800 bit/s (TCH/H4.8) carries ‘‘near–far problem’’ arising from different received power
raw user data sent at 4800 bit/s. With GSM forward er- levels.
ror correction coding, this 4800 bit/s data is sent at 11.4 The user data stream on the reverse link is first convolu-
kbit/s. tionally coded with a rate code. After interleaving, each
• Half-rate data channel for 2400 bit/s (TCH/H2.4) carries block of six encoded symbols is mapped to one of the 64 or-
raw user data sent at 2400 bit/s. With GSM forward er- thogonal Walsh functions to provide 64-ary orthogonal signal-
ing. A final fourfold spreading, giving a data rate of 1.2288

Mchips/s, is achieved by user specific codes having periods of F-ES
A-Interface
242 ⫺ 1 chips and base station specific codes having period of E-Interface
215. For the forward traffic channels, Table 5 summarizes the M-ES External
network
modulation parameters for different data rates.
Note that Walsh functions are used for different purposes
on the forward and reverse channels. On the forward chan- MDBS MD-IS IS IS
nels, Walsh functions are used for spreading to indicate a par-
ticular user channel, whereas on the reverse channel, Walsh
MDBS
functions are used for data modulation.
The speech encoder exploits gaps and pauses to reduce the
output from 9600 bps to 1200 bps during the silent period. MDBS MD-IS IS IS
Rates lower than 9600 bps are repeated to achieve a constant I-Interface
coded rate of 19,200 symbols per second for all possible infor-
M-ES
mation data rates.
At both the base station and subscriber unit, RAKE receiv- M-ES: Mobile end system
ers (1) are used to combine the delayed replica of the trans- F-ES: Fixed end system
mitted signal and therefore reduce the degree of fading. In IS- MD-IS: Mobile-data intermediate system
MDBS: Mobile-data base station
95, a three finger RAKE receiver is used at the base station.
IS: Intermediate system
Cellular Digital Packet Data. There are a number of wide- Figure 17. Cellular digital packet data network.
area packet-switched data services being offered over a dedi-
cated network using the specialized mobile radio (SMR) fre-
quency band near 800/900 MHz, for example, ARDIS (Ad- At the physical layer, CDPD transmissions are carried out
vanced Radio Data Information Service) and RAM Mobile using fixed-length blocks. The channel coding used is Reed–
Data System. However, CDPD is the packet-switched net- Salomon (63,47) block code with 6 bit symbols. For each
work that uses the existing analog cellular network such as packet, 282 bits are encoded into 378 bit blocks and provide
AMPS. CDPD occupies the voice channel on a secondary, non- correction for up to eight symbols. At the OSI layer 2, the
interfering basis by utilizing the unused airtime between mobile data link protocol (MDLP) is used to convey informa-
channel assignment by the MSC, which is estimated to be tion between the data link layer across the common air inter-
30% of the time. CDPD supports broadcast, dispatch, elec- face. The MDLP also provides logical data link connection,
tronic mail, and field monitoring applications. Figure 17 (2, sequence control, error detection, and flow control. The radio
Fig. 14.2, p. 361) illustrates a typical CDPD network. resource management protocol (RRMP) is a layer 3 function
The CDPD network has three interfaces: the air link inter- used for the management of radio resources, base station
face (A-Interface), the external interface (E-Interface) for exter- identification and configuration, channel hopping, and hand-
nal network interface, and the inter-service provider interface offs.
(I-Interface) for cooperating CDPD service providers. The mo-
bile subscribers (M-ES) are able to connect through the mo- BIBLIOGRAPHY
bile data base stations (MDBS) to the Internet via the interme-
diate systems (MD-IS and IS), which act as servers and 1. B. Sklar, Digital Communications—Fundamentals and Applica-
routers for the subscribers. Through the I-Interface, CDPD tions, Englewood Cliffs, NJ: Prentice-Hall, 1988.
can carry either Internet protocol (IP) or OSI connectionless 2. V. K. Garg and J. E. Wilkes, Wireless and Personal Communica-
protocol traffic. tions Systems, Upper Saddle River, NJ: Prentice-Hall, 1996.
In CDPD, the forward channel serves as a beacon and 3. K. Feher, Digital Communications—Satellite/Earth Station En-
transmits data from the PSTN side of the network while the gineering, Englewood Cliffs, NJ: Prentice-Hall, 1981.
reverse channel serves as the access channel and links all 4. J. B. Anderson, T. Aulin, and C.-E. Sunberg, Digital Phase Modu-
the mobile subscribers to the CDPD network. Collisions result lation, New York: Plenum, 1986.
when many mobile subscribers attempt to access the channel 5. A. A. M. Saleh, Frequency-independent and frequency-dependent
simultaneously and are resolved by slotted DSMA/CD. nonlinear models of TWT amplifiers. IEEE Trans. Commun.,
COM-29: 1715–1720, 1981.
6. K. Pahlavan and A. H. Levesque, Wireless Information Networks,
Table 5. Summary of Forward Traffic Channel Modulation New York: Wiley, 1995.
Parameters 7. C. E. Cook et al. (eds.), Spread-Spectrum Communications, Pisca-
taway, NJ: IEEE Press, 1983.
Parameter Data Rate (bps)
8. T. M. Nguyen, Optimize PSK systems through direct-sequence
User data rate 9600 4800 2400 1200 techniques. Microwaves RF, 24 (1): 118–126, 1985.
Coding data rate 1/2 1/2 1/2 1/2 9. G. C. Hess, HandBook of Land-Mobile Radio System Coverage,
Data repetition period 1 2 4 8 Norwood, MA: Artech House, 1998.
Baseband coded data rate 19,200 19,200 19,200 19,200
10. W. C. Lee, Mobile Communications Design Fundamentals, New
PN chips/coded data bit 64 64 64 64
York: Wiley, 1993.
PN chip rate (Mcps) 1.2288 1.2288 1.2288 1.2288
PN chips/bit 128 256 512 1024 11. J. Litva and T. Lo, Digital Beamforming in Wireless Communica-
tions, Norwood, MA: Artech House, 1996.
180 INFORMATION THEORY OF RADAR AND SONAR WAVEFORMS
12. L. Kleinrock and F. A. Tobagi, Carrier sense multiple access for

packet switched radio channels, Proc. IEEE ICC’74, 1974.
13. B. Agee, Blind separation and capture of communication signals
using a multitarget constant modulus beamformer, Proc. 1989
IEEE Military Commun. Conf., 1989, pp. 19.2.1–19.2.7.
14. B. Agee, S. V. Schell, and W. A. Gardner, Spectral self coherence
restoral: A new approach to blind adaptive signal extraction us-
ing antenna arrays, Proc. IEEE, 78: 753–767, 1990.
15. R. Gooch and B. Sublett, Joint spatial temporal in a decision di-
rected adaptive antenna system, Proc. 23rd Asilomar Conf.: Sig-
nals, Systems, and Computers, Noorwijk, The Netherlands, 1989.
16. W. C. Y. Lee, Mobile Communications Engineering, New York:
McGraw-Hill, 1982.
17. N. Abramson, The ALOHA system—another alternative for com-
puter communications, Proc. Fall Joint Comput. Conf. AFIPF,
37: 1970.
18. N. Abramson, Packet switching with satellites, Proc. Fall Joint
Comput. Conf. AFIPF, 42: 1973.
19. W. Crowther et al., A system for broadcast communication: reser-
vation ALOHA, Proc. 6th Hawaii Int. Conf. Syst. Sci., 1973.
20. F. A. Tobagi, Multiaccess protocols in packet communication sys-
tems, IEEE Trans. Commun., COM-28: 468–488, 1980.
21. T. S. Rappaport, Wireless Communications Principles and Prac-
tice, Upper Saddle River, NJ: Prentice-Hall, 1996.
22. L. Hanzo, The Pan-European Cellular System, in J. D. Gibson
(ed.), The Mobile Communications Handbook, Boca Raton, FL:
CRC Press, 1996.
TIEN M. NGUYEN
The Aerospace Corporation
HUNG NGUYEN
Mountain Technology Inc.
BOI N. TRAN
The Boeing Company
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Information Theory of Radar and Sonar Waveforms ● Email this page Enter words or
● Print this page
Mark R. Bell1
1Purdue University, West Lafayette, IN
reserved. Search
Abstract
Matched Filter Processing
The Ambiguity Function
Radar Waveform Design
Current and Future Directions
Keywords: ambiguity function; synthetic aperture; uncertainty ellipse; resolution; range-doppler; coherent; noncoherent

rowband signals normally encountered in radar and nar-

rowband active sonar systems, this is well approximated by a
shift in the scattered waveform’s center or carrier frequency
proportional to the carrier frequency and the closing radial
velocity between the target and scatterer (1). For wideband
signals encountered in impulse radar and wideband sonar
systems, this approximation is not accurate, and the Doppler
effect must be modeled explicitly as a contraction or dilation
of the time axis of the received signal.
One of the chief functions of a radar or sonar system is to
distinguish, resolve, or separate the scattered returns from
targets in the illuminated environment. This can be done by
resolving the scatterers in delay, Doppler, or both delay and
Doppler. In many problems of practical importance, resolu-
tion in delay or Doppler alone is not sufficient to achieve the
desired resolution requirements for the pulse-echo measure-
ment system. In these cases, joint delay-Doppler resolution is
essential. The resolution capabilities of any pulse-echo system
are a strong function of the shape of the transmitted wave-
forms employed by the system.
In the course of early radar development, radar systems
were designed to measure the delay—and hence range—to
the target, or they were designed to measure the Doppler fre-
quency shift—and hence radial velocity—of the target with
respect to the radar. The waveforms used for range measure-
ment systems consisted of very narrow pulses for which the
time delay between transmission and reception could easily
be measured; these systems are referred to as pulsed radar
systems. The waveforms used in the pulsed delay measure-
ment radars were narrow pulses, with the ability to resolve
closely spaced targets determined by the narrowness of the
the pulses. If the returns from two pulses overlapped because
two targets were too close to each other in range, the targets
could not be resolved. So from a range resolution point of
view, narrow pulses were considered very desirable. However,
because the ability to detect small targets at a distance de-
INFORMATION THEORY OF RADAR pends on the total energy in a pulse, it is not generally possi-
AND SONAR WAVEFORMS ble to make the pulses arbitrarily narrow and still achieve
the necessary pulse energy without requiring unrealistic in-
Radar and active sonar systems extract information about an stantaneous power from the transmitter.
environment by illuminating it with electromagnetic or acous- As radar systems theory and development progressed, it
tic radiation. The illuminating field is scattered by objects in became clear that it was not pulse width per se that deter-
the environment, and the scattered field is collected by a re- mined the delay resolution characteristics of a radar wave-
ceiver, which processes it to determine the presence, posi- form, but rather the bandwidth of the transmitted radar sig-
tions, velocities, and scattering characteristics of these ob- nal. As a result, waveforms of longer duration—but
jects. These active pulse-echo systems provide us with tools appropriately modulated to achieve the necessary bandwidth
for observing environments not easily perceived using our to meet the desired delay resolution requirements—could be
senses alone. The key idea in any pulse-echo measurement employed, which would allow for both sufficient energy to
system is to transmit a pulse or waveform and listen for the meet detection requirements and sufficient bandwidth to
echo. Information about the scattering objects is extracted by meet delay resolution requirements. The first detailed studies
comparing the transmitted pulse or waveform with the re- of waveforms with these properties were conducted by Wood-
ceived waveform scattered by the object. Many characteris- ward and Davies (2).
tics, including the delay between transmission and reception,
the amplitude of the echo, and changes in the shape of the
transmitted waveform, are useful in providing information MATCHED FILTER PROCESSING
about the scattering objects.
Two primary attributes characterizing the echo return in Radar systems typically process scattered target returns for
a pulse-echo system are the round-trip propagation delay and detection by filtering of the received signal with a bank of
the change in the received waveform resulting from the Dopp- matched filters matched to various time delayed and Doppler
ler effect. The Doppler effect induces a compression or dila- shifted versions of the transmitted signal. It is well known
tion in time for the scattered signal as a result of radial target that a matched filter—or the corresponding correlation re-
motion toward or away from the pulse-echo sensor. For nar- ceiver—provides the maximum signal-to-noise ratio of all lin-
INFORMATION THEORY OF RADAR AND SONAR WAVEFORMS 181
ear time-invariant receiver filters when the signal is being more, if 애(␶, ␯) describes a continuous scattering density, the
detected in additive white noise. Of course, if the filter is mis- response of the matched filter h␶,␯(t) to this scattering density
matched in delay or Doppler, the response, and hence signal- is
to-noise ratio, of the output will no longer be maximum. While
∞ ∞
this suboptimality of mismatched filters can in some cases be
OT (τ , ν) = µ(t, v)e − jφ (t ) χs (τ − t, ν − v) dt dv
detrimental (e.g., where processing constraints only allow for −∞ −∞
a small number of Doppler filters), it provides the basis for
target resolution in matched filter radar. We will now see how Here, ␾(␶) ⫽ ej2앟f0␶ is the carrier phase shift caused by the
this gives rise to the notion of the ambiguity function—a key propagation delay ␶. If we define 웂(␶, ␯) ⫽ 애(␶, ␯)e⫺j2앟f0␶, this
tool in radar resolution and accuracy assessment. becomes
Let s(t) be the baseband analytic signal transmitted by the
∞ ∞
radar system. After being demodulated down to baseband, the
received signal due to a scatterer with round-trip delay ␶0 and OT (τ , φ) = γ (t, v)χs (τ − t, ν − v) dt dv
−∞ −∞
Doppler frequency shift ␯0 is
which is the two-dimensional convolution of 웂(␶, ␯) with
r(t) = s(t − τ0 )e j2π ν 0 t e jφ
␹s(␶, ␯), and can be thought of as the image of 웂(␶, ␯) obtained
using an imaging aperture with point-spread function ␹(␶, ␯)
where ej␾ is the phase shift in the received carrier due to the (3, Chap. 4), as shown in Fig 1.
propagation delay ␶0; hence, ␾ ⫽ 2앟f 0␶0. If we process this sig-
nal with a matched filter
THE AMBIGUITY FUNCTION
hτ ,ν (t) = s∗ (T − t + τ )e− j2π ν (T −t )
As we have seen, the ambiguity function plays a significant
matched to the signal role in determining the delay-Doppler resolution of a radar
system. The ambiguity function was originally introduced by
q(t) = s(t − τ )e j2π νt Woodward (2), and several related but functionally equivalent
forms have been used since that time. Two common forms
and designed to maximize the signal output at time T, the currently used are the asymmetric ambiguity function and the
matched filter output at time T is given by symmetric ambiguity function, and they are defined as follows.
The asymmetric ambiguity function of a signal s(t) is defined
∞
OT (τ , ν) = r(t)hτ ,ν (T − t) dt as

−∞ ∞
∞
= s(t − τ0 )e j2π ν 0 t e jφ s∗ (t − τ )e − j2π νt dt χs (τ , ν) = s(t)s∗ (t − τ )e j2π νt dt (2)
−∞ −∞
∞
= e jφ s(u)e j2π ν 0 (u+τ 0 ) s∗ (u − (τ −τ0 ))e − j2π ν (u+τ 0 ) du and the symmetric ambiguity function of s(t) is defined as
−∞
∞ ∞
= e jφ e− j2π (ν −ν 0 )τ 0 s(u)s∗ (u − (τ −τ0 ))e− j2π (ν−ν 0 )u du s (τ , ν) = s(t + τ /2)s∗ (t − τ /2)e − j2π νt dt (3)
−∞
−∞
jφ − j2π (ν −ν 0 )τ 0
=e e χs (τ − τ0 , ν − ν0 )
The notation ‘‘ⴱ’’ denotes complex conjugation. The asymmet-
where ␹s(␶, ␯) is the ambiguity function of s(t), defined as ric ambiguity function is the form typically used by radar en-
∞
gineers and most closely related to the form introduced by
χs (τ , ν) = s(t)s∗ (t − τ )e j2π νt dt Woodward (2). The symmetric ambiguity function is more
−∞ widely used in signal theory because its symmetry is mathe-
matically convenient and it is consistent with the general the-
For narrowband signals, ␯␶0 Ⰶ 1 and ␯0␶0 Ⰶ 1 for all ␯, ␯0, ory of time–frequency distributions (4).
and ␶0 of interest, however, f 0␶0 Ⰷ 1. Hence, we can write The asymmetric ambiguity function ␹s(␶, ␯) and the sym-
metric ambiguity function ⌫s(␶, ␯) are related by
OT (τ , ν) = e− jφ χs (τ − τ0 , ν − ν0 ) (1)
s (τ , ν) = e jπ νt χs (τ , −ν)
Because h␶,␯(t) is a linear time-invariant filter, if we have N
scatterers with scattering strengths 애1, . . ., 애N, delays ␶1, and
. . ., ␶N, and Doppler shifts ␯1, . . ., ␯N, the response of h␶,␯(t)
to the collection of scatterers is χs (τ , ν) = e jπ νt s (τ , −ν)

N
so knowledge of one form implies knowledge of the other. In
OT (τ , ν) = µi e − jφ i χs (τ − τi , ν − νi )
i=1
practice, the ambiguity surface As(␶, ␯), given by the modulus
of the symmetric ambiguity function,
where ␾i is the carrier phase shift in the return from the ith
scatterer resulting from the propagation delay ␶i. Further- As (τ , ν) = |s (τ , ν)| = |χs (τ , −ν)|
ν ν ν
γ (τ,ν ) χ (τ,ν ) OT (τ,ν )
Figure 1. Imaging interpretation of

a delay-Doppler pulse-echo system. A τ τ τ
waveform s(t) with ambiguity function
␹(␶, ␯) gives rise to a delay-Doppler image
O T(␶, ␯) that is the convolution of the ideal
image 웂(␶, ␯) with the point-spread func- Target environment Aperture point-spread Observed delay-Doppler
tion ␹(␶, ␯). image function image
is usually sufficient to characterize a waveform’s delay-Dopp- These figures illustrate the very different delay-Doppler
ler resolution characteristics, as it gives the magnitude of the resolution characteristics provided by these signals when they
matched filter response for a delay-Doppler mismatch of (␶, ␯). are processed using a matched filter.
Figures 2 and 3 show ambiguity surfaces of a simple pulse The shape or properties of the main lobe of the ambiguity
surface 兩⌫s(␶, ␯)兩 centered about the origin determine the abil-
1, for |t| < 1/2 ity of the corresponding waveform to resolve two scatterers
s1 (t) =
0, elsewhere close together in both delay and Doppler. The ambiguity sur-
face squared 兩⌫s(␶, ␯)兩2 close to the origin can be expanded as
and a linear FM ‘‘chirp’’ a two-dimensional Taylor series about (␶, ␯) ⫽ (0, 0). From
this it follows that the ambiguity surface itself may be ap-
2
e jπ αt , for |t| < 1/2 proximated by (8, pp. 21–22)
s2 (t) =
0, elsewhere
|s (τ , ν)| ≈ (0, 0)[1 − 2π 2 TG2 ν 2 − 4πρTG BG τ ν − −2π 2 B2G τ 2 ]
(4)
(with 움 ⫽ 8), respectively. The ambiguity function of s1(t) is
where
s 1 (τ , ν) =
(1 − |τ |)sinc[ν(1 − |τ |)], for |τ | ≤ 1
0, elsewhere

BG = f2 − f 2
The ambiguity function of s2(t) is
is the Gabor bandwidth of the signal,
s 2 (τ , ν) =

(1 − |τ |)sinc[(ν − ατ )(1 − |τ |)], for |τ | ≤ 1
TG = t2 − t 2
0, elsewhere
is the Gabor timewidth of the signal, the frequency and time
1
1
0.5 10
Γ2(τ ,ν ) 0.5 10
0
5 0
5
–1
0 –1
–0.5 ν 0
–0.5 ν
0 –5 0 –5
τ τ
0.5 0.5
1 –10 1 –10
Figure 2. Symmetric ambiguity function ⌫1(␶, ␯) of a rectangular Figure 3. Symmetric ambiguity function ⌫2(␶, ␯) of a linear FM chirp
pulse of duration 1. of duration 1.
ν While the uncertainty ellipse provides a rough means of

determining the resolution performance of a waveform for re-
sb(t) solving closely spaced targets in isolation from other interfer-
ing scatterers, it is not sufficient to completely characterize a
waveform’s measurement characteristics. Target returns with
delay-Doppler coordinates falling in the sidelobes of the ambi-
τ guity function can have a significant effect on a radar’s mea-
surement and resolution capabilities. For this reason, in order
to effectively design radar waveforms for specific measure-
sa(t) ment tasks, it is important to have a thorough understanding
of the properties of ambiguity functions.
Properties of Ambiguity Functions

2
Figure 4. Uncertainty ellipses corresponding to sa(t) ⫽ e ⫺웁t
and In order to gain a thorough understanding of the delay-Dopp-
2 2
sb(t) ⫽ e⫺웁t ej앟움t . ler resolution characteristics of various signals under
matched filter processing, it is necessary to understand the
general properties of ambiguity functions. With this in mind,
moments of the signal s(t) are we now consider the properties of ambiguity functions. Proofs
∞
of these properties may be found in Refs. 5 (Chap. 9), 6
1 (Chaps. 5–7), 7 (Chap. 4), 8, and 9 (Chap. 10).
fn = f n |S( f )|2 d f
Es −∞
Property 1. The energy in the signal s(t) is given by
and
∞
∞ Es = s (0, 0) = |s(t)|2 dt
1
tn = t n |s(t)|2 dt −∞
Es −∞
Property 2 (Volume).
respectively, and the skew parameter ␳ is
∞ ∞
∞
1 j |s (τ , ν)|2 dτ dν = |s (0, 0)|2 = Es2
ρ= Re t ṡ(t)s∗ (t) dt − t f −∞ −∞
TB 2πEs −∞
Property 3. The time autocorrelation function ␾s(␶) of the
where ṡ(t) is the derivative of s(t).
signal s(t) is given by
The shape of the main lobe about the origin of the ambigu-
ity function can be determined by intersecting a plane paral- ∞
lel to the (␶, ␯) plane with the main lobe near the peak value. φs (τ ) = s (τ , 0) = s(t + τ /2)s∗ (t − τ /2) dt
−∞
Using the approximation of Eq. (4) and setting it equal to the
constant ambiguity surface height 웂0 specified by the inter-
secting plane, we have Property 4. The energy spectrum of the signal s(t) is given
by
(0, 0)[1 − 2π 2 TG2 ν 2 − 4πρTG BG τ ν − 2π 2 B2G τ 2 ] = γ0 ∞
s (0, ν) = |s(t)|2 e − j2π νt dt
which we can rewrite as −∞
B2G τ 2 + 2ρBG TG τ ν + TG2 ν 2 = C (5) Property 5. The symmetric ambiguity function of the sig-
nal s(t) can be written as
where C is a positive constant. This is the equation of an el- ∞
lipse in ␶ and ␯, and this ellipse is known as the uncertainty s (τ , ν) = S( f + ν/2)S∗ ( f − ν/2)e j2π f τ d f
ellipse of the waveform s(t). The uncertainty ellipse describes −∞
the shape of the main lobe of 兩⌫s(␶, ␯)兩 in the region around its
peak and hence provides a concise description of the capabil- where
ity of s(t) to resolve closely spaced targets concentrated in the ∞
main lobe region. The value of C itself is not critical, since the S( f ) = s(t)e − j2π f t dt
shape of the uncertainty ellipse is what is of primary interest. −∞
Figure 4 shows the uncertainty ellipses of a Gaussian pulse
is the Fourier transform of s(t).
2
sa (t) = e−βt Property 6. If s(0) ⬆ 0, s(t) can be recovered from ⌫s(␶, ␯)
using the relationship
and a linear FM chirp modulated Gaussian pulse
∞
1
−βt 2 jπ αt 2 s(t) = s (t, ν) jπνt dt
sb (t) = e e s∗ (0) −∞
where 2 of the ambiguity function. Property 1 states that the height

of 兩⌫s(␶, ␯)兩2 at the origin is 兩⌫s(0, 0)兩2 ⫽ Es2. Property 2 states
∞
|s(0)|2 = s (0, ν) dν that the total volume under 兩⌫s(␶, ␯)兩2 is Es2. So if we try to
−∞ construct a thumbtack-like 兩⌫s(␶, ␯)兩2 approximating an ideal
delta function, we run into the problem that as the height
Property 7 (Time Shift). Let s⬘(t) ⫽ s(t ⫺ ⌬). Then 兩⌫s(0, 0)兩 increases, so does the volume under 兩⌫s(␶, ␯)兩2. This
means that for a signal with a given energy, if we try to push
s (τ , ν) = e − j2π ν s (τ , ν) the volume of the ambiguity function down in one region of
the delay-Doppler plane, it must pop up somewhere else. So
Property 8 (Frequency Shift). Let s⬘(t) ⫽ s(t)ej2앟ft. Then there are limitations on just how well any waveform can do
in terms of overall delay-Doppler ambiguity performance. In
s (τ , ν) = e j2π f τ s (τ , ν) fact, the radar waveform design problem corresponds to de-
signing waveforms that distribute the ambiguity volume in
Property 9 (Symmetry). ⌫s(␶, ␯) ⫽ ⌫*s (⫺␶, ⫺␯). the (␶, ␯) plane in a way appropriate for the delay-Doppler
Property 10 (Maximum). The largest magnitude of the am- measurement problem at hand. We now investigate some of
biguity function is always at the origin: these techniques.
|s (τ , ν)| ≤ s (0, 0) = Es The Wideband Ambiguity Function

In the situation that the waveforms being considered are not
This follows directly from the Schwarz inequality.
narrowband or the target velocity is not small compared with
Property 11 (Time Scaling). Let s⬘(t) ⫽ s(at), where a ⬆ 0. the velocity of wave propagation, the Doppler effect cannot be
Then modeled accurately as a frequency shift. In this case, it must
be modeled as a contraction or dilation of the time axis. When
1
s (τ , ν) = s (aτ , ν/a) this is the case, the ambiguity functions ␹s(␶, ␯) and ⌫s(␶, ␯)
|a|
defined in Eqs. (2) and (3) can no longer be used to model the
2 output response of the delay and Doppler (velocity) mis-
Property 12 (Quadratic Phase Shift). Let s⬘(t) ⫽ s(t)ej앟움t . matched matched filter. In this case, the wideband ambiguity
Then function must be used (10–13). Several slightly different but
s (τ , ν) = s (τ , ν − ατ ) mathematically equivalent forms of the wideband ambiguity
function have been introduced. One commonly used form (13)
is
Property 13 (Self-transform). 兩⌫s(␶, ␯)兩2 is its own Fourier
transform in the sense that
s (τ , γ ) =
p|γ | ∞
s(t)s∗ (γ (t − τ )) dt (6)
∞ ∞ −∞
2 − j2π f τ
|s (τ , ν)| e e j2π tν
dτ dν = |s (t, f )| 2
−∞ −∞
where 웂 is the scale factor arising from the contraction or dila-
tion of the time axis as a result of the Doppler effect. Specifi-
Property 14 (Wigner Distribution). The two-dimensional in-
cally,
verse Fourier transform of the ambiguity function ⌫s(␶,
␯) of a signal s(t) is its Wigner distribution Ws(t, f): 1 − v/c
γ =
∞ ∞ 1 + v/c
s (τ , ν)e j2π f τ e j2π tν dτ dν = Ws (t, f )
−∞ −∞
where v is the radial velocity of the target with respect to the
sensor (motion away from the sensor positive), and c is the
where the Wigner distribution of s(t) is defined as (4,8)
velocity of wave propagation in the medium. While the theory
∞ of wideband ambiguity functions is not as well developed as
Ws (t, f ) = s(t + τ /2)s∗ (t − τ /2)e j2 pi f τ dτ for the case of narrowband ambiguity functions, a significant
−∞
amount of work has been done in this area. See Ref. 13 for a
readable survey of current results. We will focus primarily
These properties of the ambiguity function have immediate
on the narrowband ambiguity function throughout the rest of
implications for the design of radar waveforms. From the im-
this article.
aging analogy of delay-Doppler measurement, where the am-
biguity function plays the role of the imaging aperture, it is
clear that an ideal ambiguity function would behave much RADAR WAVEFORM DESIGN
like a pinhole aperture—a two-dimensional Dirac delta func-
tion centered at the origin of the delay-Doppler plane. Such The problem of designing radar waveforms with good delay-
an ambiguity function would yield a radar system giving a Doppler resolution has received considerable attention (14–
response of unity if the return had the assumed delay and 24). Waveforms developed for this purpose have generally
Doppler, but a response of zero if it did not. Such a system fallen into three broad categories:
would in fact have perfect delay-Doppler resolution proper-
ties. Unfortunately, such an ambiguity function does not ex- 1. Phase and frequency modulation of individual radar
ist. This can be seen by considering Property 1 and Property pulses
2. Pulse train waveforms φ (τ ;α )

3. Coded waveforms 1
We will now investigate these techniques and consider how

α=0
each can be used to improve radar delay-Doppler resolution
characteristics and shape the ambiguity functions of radar
waveforms in desirable ways. α=4
Phase and Frequency Modulation of Radar Pulses α = 16

The fundamental observation that led to the development of τ
phase and frequency modulation of radar pulses was that it –1 –0.5 0.5 1
is not the duration of a pulse, but rather its bandwidth, that
determines its range resolution characteristics. Early range
measurement systems used short duration pulses to make Figure 5. Time correlation ␾(␶, 움) ⫽ ⌫s(␶, 0) for linear FM chirp
range measurements, and narrow pulses were used to obtain pulses of duration 1 and modulation indices 움 of 0, 4, and 16.
good range resolution, but this put a severe limitation on the
detection range of these systems, because detection perfor-
mance is a function of the total energy in the transmitted shown in Fig. 5, we see that, although pulse durations are
pulse, and with the peak power limitations present in most equivalent (in this case we take T ⫽ 1), there is a significant
real radar systems, the only way to increase total energy is difference in range resolution. With increasing 움, we also
to increase the pulse duration. However, if the pulse used is have increasing bandwidth. Looking at the ambiguity func-
simply gating a constant frequency sinusoidal carrier, in- tion of the linear FM chirp shown in Fig. 3, and comparing
creasing the duration decreases the bandwidth of the trans- the ambiguity function of the simple rectangular pulse in Fig.
mitted signal. This observation led to the conjecture that per- 2, it is clear that the broadening of the pulse bandwidth has
haps it is large bandwidth instead of short pulse duration brought about increased delay resolution–however, not with-
that leads to good range resolution. This conjecture was in out cost.
fact shown to be true (14). We now investigate this using am- From Property 12, the quadratic phase shift property, we
biguity functions. see that the matched filter will not only have a large response
The ambiguity function of the simple rectangular pulse to the signal with the desired delay ␶ and Doppler ␯, but also
to any signal with delay ␶ ⫹ ⌬␶ and Doppler ␯ ⫹ ⌬␯, where
⌬␯ ⫺ 움 ⌬␶ ⫽ 0. This locus of peak response for the chirp is
1, for |t| ≤ T oriented along the line of slope 움 in the (␶, ␯) plane. So when
s1 (t) = matched filtering for a chirp with some desired delay and
0, elswhere
Doppler shift imposed on it, we are never certain if a large
response is the result of a scatterer at the desired delay and
of duration T is Doppler, or a scatterer with a delay-Doppler offset lying near
the locus of maximal delay-Doppler response. While for a sin-
gle scatterer the actual delay and Doppler can be determined
(T − |τ |)sinc[ν(T − |τ |)], for |t| ≤ T by processing with a sufficiently dense band of matched filters
1 (τ , ν) =
0, elswhere in delay and Doppler, scatterers lying along this maximal re-
sponse locus are hard to resolve if they are too close in delay
and Doppler. From the point of view of detection, however,
and the ambiguity function of the linear FM ‘‘chirp’’ pulse there is a benefit to this ‘‘Doppler tolerance’’ of the chirp
waveform. It is not necessary to have a bank of Doppler filters
2 as densely located in Doppler frequency in order to detect the
e jπ αt , for |t| ≤ T
s2 (t) = presence of targets (25, Chap. 9).
0, elswhere
Coherent Pulse Train Waveforms
of the same duration is Another way to increase the delay-Doppler resolution and
ambiguity characteristics of radar waveforms is through the
use of pulse trains—waveforms synthesized by repeating a
(T − |τ |)sinc[(ν − ατ )(T − |τ |)], for |t| ≤ T simple pulse shape over and over. An extension of this basic
2 (τ , ν) =
0, elswhere idea involves constructing the pulse train as a sequence of
shorter waveforms—not all the same—from a prescribed set
of waveforms (26). Most modern radar systems employ pulse
[Note that ⌫2(␶, ␯) is easily obtained from ⌫1(␶, ␯) using Prop- trains instead of single pulses for a number of reasons. Re-
erty 12 of the ambiguity function.] If we compare the time gardless of whether the pulse train returns are processed
autocorrelation functions ␾1(␶) ⫽ ⌫1(␶, 0) and ␾2(␶) ⫽ ⌫2(␶, 0) coherently (keeping track of the phase reference from pulse-
for various values of the linear FM modulation index 움 as to-pulse and using it to construct a matched filter) or nonco-
herently (simply summing the pulse-to-pulse amplitude of the

matched filter output without reference to phase), a pulse
train increases receiver output signal-to-noise ratio, and
hence increases detection range [e.g., see Ref. 25 (Chaps. 6
and 8)]. Furthermore, when processed coherently in a pulse-
Doppler processor, flexible, high-resolution delay-Doppler pro-
cessing is possible. In discussing pulse trains, we will focus
on coherent pulse-Doppler waveforms, as pulse-Doppler radar 4
systems have become the dominant form of radar for both sur- 2
2
veillance and synthetic aperture radar (SAR) applications.
A pulse train is constructed by repeating a single pulse 0
1
p(t) regularly at uniform intervals Tr; Tr is called the pulse –2
repetition interval (PRI). The frequency f r ⫽ 1/Tr is called the –10
pulse repetition frequency (PRF) of the pulse train. Typically, 0
–5 ν
the duration ␶p of the pulse v(t) is much less than Tr. A uni-
form pulse train s(t) made up of N repeated pulses and having 0 –1
PRI Tr can be written as τ
5

N−1
10 –2
x(t) = p(t − nTr )
n=0 Figure 7. The ambiguity function for a uniform pulse train of rectan-
gular pulses.
A typical example of such a pulse train in which the pulse
p(t) repeated is a simple rectangular pulse is shown in Fig. 6.
Centering this pulse train about the origin of the time axis,
we can write it as lobes’’ centered at (␶, ␯) pairs given by

N−1
(τ , ν) = (nTr , k/Tr )
s(t) = p(t − nTr + (N − 1)Tr /2) (7)
n=0
where n is any integer with 兩n兩 ⱕ N ⫺ 1, and k is any integer.
The symmetric ambiguity function of this pulse train is (6,8) From the behavior of the Dirichlet function

N−1
sin πνTr (N − |n|) sin πνTr (N − |n|)
s (τ , ν) = · p (τ − nTr , ν) (8)
n=−(N−1)
sin πνTr sin πνTr
where ⌫p(␶, ␯) is the ambiguity function of the elementary

weighting the delayed copies of ⌫p(␶, ␯) in Eq. (8), it is clear
pulse p(t) used to construct the pulse train.
that the peak amplitudes of these grating lobes fall off as we
In order to gain an understanding for the behavior of the
move farther away from the main lobe (n ⫽ 0 and k ⫽ 0).
ambiguity function of the pulse train, consider the special
case of a uniform pulse train of N ⫽ 5 rectangular pulses,
each of length ␶p ⫽ 1 with a PRI of Tr ⫽ 5. The plot of this
ambiguity function is shown in Fig. 7. A similar plot in which
p(t) is a linear FM chirp of the form
2
e jπ αt , for |t| ≤ 1
p(t) =
0, elsewhere
and 움 ⫽ 8 is shown in Fig. 8. From the form of Eq. (8), we see
;;;
that the ambiguity function of the pulse train has ‘‘grating 4
2 10
s(t) 0
5
(N–1)Tr
2Tr –1
Tr 0
–0.5 ν
τp
0 –5
... τ
0.5
t
1 –10
Figure 6. Uniform pulse train waveform s(t) constructed by re-
peating a basic pulse shape p(t) N times with a pulse repetition inter- Figure 8. The ambiguity function for a uniform pulse train of linear
val of Tr. FM chirp pulses.
ν however, this approach is only successful in sparse target en-

vironments. When there are many targets in proximity in
both delay and Doppler, sorting out the ambiguity becomes
unwieldy. Another disadvantage of these low PRF pulse trains
is that they have lower duty cycles for a given pulse width,
resulting in a significant decrease in average transmitted
power (and hence detection range) for a given elemental pulse
width ␶p and peak power constraint.
At the other extreme, if one makes Tr very small, the ef-
fects of Doppler ambiguity can be minimized. In fact, if 1/Tr
Tr
is greater than the maximum Doppler frequency shift we ex-
1/Tr
pect to encounter, there is no Doppler ambiguity. However,
τ there will most likely be severe range ambiguities if such high
PRF pulse trains are used.
For most radar surveillance problems involving the detec-
tion of aircraft and missiles, the size of the surveillance vol-
ume and the target velocities involved dictate that there will
be ambiguities in both delay and Doppler, and most often a
medium PRF pulse train is employed. In this case the PRF
is usually selected to meet the energy efficiency (duty-cycle)
constraints to ensure reliable detection and to make the na-
ture of the delay-Doppler ambiguities such that they are not
extreme in either the delay or Doppler dimension. In this
case, delay-Doppler ambiguities can be resolved by changing
the PRF from one coherent N-pulse train to the next by
Figure 9. Locations of grating sidelobes in the ambiguity function of
a uniform pulse train.
changing Tr from pulse train to pulse train. This technique is
sometimes called PRF staggering, and is effective in sparse
environments. As can be seen from Eq. (8), proper selection
of the Tr from pulse train to pulse train makes this feasible,
Figure 9 shows this grating lobe behavior for a uniform because in general, with proper selection of the PRIs Tr used,
pulse train. only the true delay-Doppler (␶, ␯) will be a feasible solution
By observing the main lobe of the uniform pulse train, we for all Tr. An additional benefit of changing Tr from pulse
see that its delay resolution is approximately ␶p —the range train to pulse train is that it alleviates the ‘‘blind range’’ prob-
resolution of the elementary pulse p(t)—while the Doppler lem in monostatic radars. These radars cannot transmit and
resolution is approximately 1/NTr, a value that can be made receive simultaneously. When they transmit a pulse train, the
arbitrarily small by making N sufficiently large, limited only receiver is turned off during pulse transmission and is turned
by practical considerations in coherently processing the re- on to listen for target returns in the periods between pulses.
ceived signal. However, the ambiguities introduced through Hence target returns having delays corresponding to the time
the grating lobes at (nTr, k/Tr) can result in uncertainty in intervals of successive pulse transmissions are not seen by
the actual delay and Doppler of the target. As a result, both the radar. Changing Tr from pulse train to pulse train moves
the range and Doppler determined radial velocity of the tar- the blind ranges around, ensuring nearly uniform surveil-
get can be ambiguous. While in principle this ambiguity can lance coverage at all ranges.
be resolved in the case of a small number of targets using the
fact that the sidelobes have successively smaller amplitude as
Phase and Frequency Coded Waveforms
we move away from the main lobe, this approach is not practi-
cal because of the way in which the bank of matched filters Another highly successful approach to designing waveforms
is actually implemented in a pulse-Doppler processor. Hence, with desirable ambiguity functions has been to use phase
another approach to resolving (nTr, k/Tr) ambiguity is needed. and/or frequency coding. The general form of a coded wave-
We will briefly discuss approaches that can be taken. form (with coding in both phase and frequency) is
One way to reduce the effects of the range ambiguity is to
make Tr large. This makes the delay ambiguity large, and
N−1
often the delay ambiguity (and hence unambiguous measure- s(t) = pT (t − nT ) exp{ j2πdnt/T} exp{ jφn } (9)
ment range) can be made sufficiently large so that range am- n=0
biguity is no longer a problem for ranges of interest. Of
course, this complicates the Doppler ambiguity problem, be- The coded waveform s(t) consists of a sequence of N identical
cause the pulse repetition frequency (PRF) 1/Tr is the effec- baseband pulses pT(t) of length T; these pulses pT(t ⫺ nT) are
tive sampling rate of the pulse-Doppler processor. A large usually referred to as the chips making up the waveform
value of Tp results in a low PRF and hence low sampling rate, s(t). Usually, the chip pulse pT(t) has the form
and there is significant aliasing of the Doppler signal. Some

systems do use this approach to deal with the ambiguity prob- 1, for 0 ≤ t < T
lem, using range differences (often called range rate measure- pT (t) =
ments) from pulse to pulse to resolve the Doppler ambiguity; 0, elsewhere
Note that each chip pulse pT(t ⫺ nT) is of duration T and each 6 1
successive pulse is delayed by T, so there are no empty spaces
5 2
in the resulting coded waveform s(t) of duration NT. In fact,
for the rectangular pT(t) specified above, 兩s(t)兩 ⫽ 1 for all t 僆 4 3
[0, NT). However, each pulse in the sequence is modulated by 3 4
an integral frequency modulating index dn and a phase ␾n
Doppler offset, k
2 5
that can take on any real number value. To specify the modu- 1 6
lating frequency and phase patterns of a coded waveform, we 0 7
must specify a length N sequence of frequency indices 兵d0,
–1 6
. . ., dN⫺1其 and a length N sequence of phases 兵␾0, . . ., ␾N⫺1其.
–2 5
If dn ⫽ 0 for n ⫽ 0, . . ., N ⫺ 1, then the coded waveform is
strictly phase modulated. If ␾n ⫽ 0 for n ⫽ 0, . . ., N ⫺ 1, –3 4
then the coded waveform is strictly frequency modulated. The –4 3
asymmetric ambiguity function of s(t) as given in Eq. (9) is –5 2
given by (26) –6 1
–6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6
N−1
N−1
χs (τ , ν) = e j(φ n −φ m ) e j2π (d m /T )τ e− j2π ν nT χ p T Delay offset, m

n=0 m=0 Figure 10. Ambiguity matrix of the 7-chip stepped frequency chirp
(dn − dm ) having frequency index sequence (0, 1, 2, 3, 4, 5, 6).
τ − (n − m)T, ν − (10)
T
There are many families of coded phase and frequency The sidelobe matrix gives the heights of the major side-
modulated waveforms. We will consider a few of the most in- lobes of a frequency coded waveform. These can be shown to
teresting of these. For a more thorough treatment of coded occur at locations (␶, ␯) ⫽ (mT, k/T), where m and k are inte-
waveforms, see Refs. 5 (Chap. 6), 6 (Chap. 8), and 7 (Chap. 8). gers. The sidelobe matrix is a table of the relative heights of
兩⌫s(mT, k/T)兩 ⫽ 兩␹s(mT, k/T)兩 for integer values of m and k in
Frequency Coded Waveforms. Consider an N-chip frequency the range of interest. So, for example, the sidelobe matrix of
coded waveform with the rectangular pT(t) defined above (here a 7-chip stepped linear FM chirp having frequency index se-
we assume ␾0 ⫽ ⭈ ⭈ ⭈ ⫽ ␾N⫺1 ⫽ 0): quence (0, 1, 2, 3, 4, 5, 6) is shown in Fig. 10, whereas that
for a 7-chip Costas waveform with frequency index sequence

N−1 (3, 6, 0, 5, 4, 1, 2) is shown in Fig. 11. Blank entries in the
s(t) = pT (t − nT ) exp{ j2πdnt/T} (11) sidelobe matrix correspond to zero. Clearly, there is a signifi-
n=0 cant difference between the ambiguity matrices (and hence
ambiguity functions) of these two frequency coded waveforms,
Waveforms of this kind are sometimes referred to as fre- despite the fact that they have the same duration, same num-
quency hopping waveforms, because the frequency of the ber of chips, and same set of modulating frequencies. It is
waveform ‘‘hops’’ to a new frequency when transitioning from only the order in which the modulating frequencies are used
chip to chip. Now suppose we take the sequence of frequency that determines their ambiguity behavior.
modulation indices to be
φn = n, n = 0, . . ., N − 1
6 1
Then the resulting s(t) is a stepped frequency approximation 5 1 1
to a linear FM chirp. Here we have used each of the frequency 4 1 1 1
modulation indices in the set 兵0, . . ., N ⫺ 1其 once and only 3 1 1 1 1
once. In general, we can describe the order in which the indi-
Doppler offset, k
2 1 1 1 1 1
ces are used to construct the waveform using a frequency in-
1 1 1 1 1 1 1
dex sequence of the form (d0, . . ., dN⫺1). So, for example, the
0 7
stepped linear FM sequence has frequency index sequence (0,
1, 2, . . ., N ⫺ 1). There are of course N! possible frequency –1 1 1 1 1 1 1
coded waveforms that use each of these indices once and only –2 1 1 1 1 1
once, since there are N! permutations of the N elements or, –3 1 1 1 1
equivalently, N! distinct frequency index sequences. Some of –4 1 1 1
these permutations give rise to waveforms with ambiguity
–5 1 1
functions that are very different from that of the stepped fre-
quency approximation to the linear FM chirp. For the purpose –6 1
of comparison, we consider two such waveforms, the 16-chip –6 –5 –4 –3 –2 –1 0 1 2 3 4 5 6
stepped linear FM waveform, and the 16-chip Costas wave- Delay offset, m
form (20). Before we do this, we introduce the notion of the Figure 11. Ambiguity matrix of the 7-chip stepped frequency coded
sidelobe matrix. Costas waveform having frequency index sequence.
In looking at the ambiguity matrix of the Costas waveform that the US B-2 Stealth bomber employs a high-resolution
in Fig. 11, it is apparent that from the point of view of both radar system using PN sequences of this kind (30).
mainlobe delay-Doppler resolution and sidelobe delay-Dopp- There are many specialized families of phase coded wave-
ler ambiguity, the Costas waveform is nearly ideal. All of the forms, most of which have the property that they have excel-
main sidelobes have a height of 1, while the mainlobe has a lent delay (range) resolution and ambiguity properties along
height of 7. In fact, by definition, an N-chip Costas waveform the ␯ ⫽ 0 axis. Many of these waveforms also have fairly good
is a frequency coded waveform with a frequency index se- ambiguity and resolution properties off the zero-Doppler axis
quence that is a permutation of the numbers 0, 1, 2, . . ., N as well. Examples of these waveforms include those generated
⫺ 1 such that the mainlobe entry of the ambiguity matrix is by Barker codes, Frank codes, and Gold Codes (see Ref. 27 for
N, while the maximum sidelobe entry is 1 (20). Sequences details on these and other related families of waveforms).
(d0, . . ., dN⫺1) yielding Costas waveforms can be found for One final family of phase codes worth mentioning are the
arbitrary N by exhaustive search; however, this becomes a complementary codes originally introduced by Marcel Golay
computationally intense task, because the number of N-chip (31) for use in optical spectroscopy, but later adapted to radar
Costas sequences grows much more slowly in N, than N!, the measurement problems as well. Complementary codes are ac-
number of N-chip frequency coded waveforms. For large N, tually families of phase coded codewords. Golay originally in-
this approach becomes impractical. More efficient techniques troduced complementary codes having two codewords of equal
for constructing Costas waveforms are discussed in Refs. 21 length, with each chip taking on a value of either ⫹1 or ⫺1.
and 22. One very efficient technique for constructing Costas The two codewords had the property that their delay side-
waveforms of length N ⫽ p ⫺ 1, where p is a prime number, is lobes along the zero-Doppler axis exactly negatives each
the Welch algorithm, which involves a simple iteration having other, while their main lobes are identical. As a result, if
computational complexity proportional to N. there is no Doppler offset and two measurements of the same
target scenario can be made independently, the properly de-
layed matched filter outputs can be added, and the result is a
Phase Coded Waveforms. Consider an N-chip phase coded response in which the delay sidelobes are completely can-
waveform with the rectangular pT(t) [here we assume d0 ⫽ celed. This results in excellent ambiguity function sidelobe
⭈ ⭈ ⭈ ⫽ dN⫺1 ⫽ 0 in Eq. (9)]: cancellation along the zero-Doppler axis (32). Golay’s basic
idea has been extended to nonbinary waveforms, complemen-

N−1
tary waveform sets with more than two waveforms, and non-
s(t) = pT (t − nT ) exp{ jφn } (12)
n=0
zero-Doppler offsets (18,19,26).
The sequence of phases (␾0, . . ., ␾N⫺1) specifies the phase CURRENT AND FUTURE DIRECTIONS
angle to be applied to each of the N chips making up the
waveform s(t). While the classical theory of radar and sonar signals is in
These waveforms are very similar to the types of wave- many ways mature, there are a number of interesting efforts
forms used in direct-sequence spread-spectrum communica- to extend the theory and practice of radar and sonar signal
tions and hence are often referred to as direct-sequence wave- design. We briefly outline a few of these.
forms. Most often, the set of phases considered is a finite set, One area that has received significant attention is the de-
such as 兵0, 앟其, 兵0, 앟/2, 앟, 3앟/2其, or more generally 兵0, 앟/L, sign of sets of multiple radar waveforms for use together. The
2앟/L, . . ., (L ⫺ 1)앟/L其, where the phases ␾n take on values simplest examples of these waveform sets are Golay’s comple-
from these sets, often repeating values unlike the frequency mentary sequence waveforms (31), which we have already
coded waveforms we considered in the last section. considered, as well as their extensions (18,19,26), which we
One family of phase coded waveforms that have been ap- discussed in the last section. The basic idea is to make com-
plied to radar problems are the pseudonoise (PN) sequences plementary diverse measurements that allow for extraction of
or m-sequences commonly used in spread-spectrum communi- greater information about the target environment than can
cations (27–29). These waveforms take on values of either ⫹1 be obtained with a single waveform. Another reason for de-
or ⫺1 on each chip, and hence the phases are taken from the signing sets of waveforms for use together is for use in
set 兵0, 앟其. These waveforms are useful for generating very multistatic radar and sonar systems, where there may be sev-
wide bandwidth signals by taking N large and T small. These eral transmitters and receivers in different locations. By
sequences have excellent correlation properties and are easily allowing each receiver to listen to the returns from all trans-
generated using linear and nonlinear feedback shift register mitters, it is possible to extract much more information about
circuits. Their correlation properties give rise to sharp thumb- the environment than is possible with a single—or even mul-
tack-like responses when evaluated on the zero-Doppler (␯ ⫽ tiple—monostatic systems. For these systems to be feasible,
0) axis. As a result, high resolution and low range ambiguity it is important that the waveforms in the set have low cross-
measurements can be made using these waveforms. These correlation, as well as envelope and spectral characteristics
waveforms have the appearance of wideband noise when ob- that allow for efficient amplification and transmission in real
served with a spectral analyzer and hence are hard to detect systems. In Refs. 33 and 34, designs for a family of waveforms
without detailed knowledge of the phase sequence (␾0, ␾1, ␾2, of this type for sonar applications are considered. Another
. . ., ␾N⫺1) and have thus been used for low probability of in- novel approach to multiple waveform imaging is Bernfeld’s
tercept (LPI) ‘‘quiet radar’’ systems, where it is not desired to chirp-Doppler radar (35,36), which uses a mathematical anal-
give away the fact that the radar is in operation. It is rumored ogy between measurement using a chirp and transmission to-
mography to obtain ‘‘projections’’ of a delay-Doppler scatter- 15. C. H. Wilcox, The synthesis problem for radar ambiguity functions,
ing profile. These projections are then used to form a MRC Tech. Summary Rep. 157, Mathematics Research Center,
reconstruction of the delay-Doppler profile using the inverse US Army, Univ. Wisconsin, Madison, WI, Apr. 1960.
Radon transform techniques typically employed in projection 16. S. M. Sussman, Least-squares synthesis of radar ambiguity func-
tomography. tions, IRE Trans. Inf. Theory, Apr. 1962.
When making measurements using sets of waveforms, the 17. W. L. Root, Radar resolution of closly spaced targets, IRE Trans.
question of which waveforms from the set to transmit and in Mil. Electron., MIL-6 (2): 197–204, 1962.
what order they should be transmitted naturally arises. This 18. C. C. Tseng and C. L. Liu, Complementary sets of sequences,
gives rise to the notion of adaptive waveform radar (37). In IEEE Trans. Inf. Theory, IT-18: 644–652, 1972.
Ref. 38, the problem of designing and adaptively selecting 19. R. Sivaswami, Multiphase complementary codes, IEEE Trans.
waveforms for transmission to effect target recognition is con- Inf. Theory, 24: 546–552, 1978.
sidered. The approach used selects waveforms from a fixed 20. J. P. Costas, A study of a class of detection waveforms having
set (designed for a particular ensemble of targets to be classi- nearly ideal range-Doppler ambiguity properties, Proc. IEEE, 72:
fied) in such a way that the Kullback–Leibler information 996–1009, 1984.
measure is maximized by each selection. 21. S. W. Golomb and H. Taylor, Constructions and properties of Cos-
The idea of designing radar waveforms matched to specific tas arrays, Proc. IEEE, 72: 1143–1163, 1984.
target tasks has also been considered. In Ref. 39, the prob- 22. S. W. Golomb, Algebraic constructions for Costas arrays, J. Com-
binatorial Theory Ser. A, 37: 13–21, 1984.
lems of wideband radar waveform design for detection and
information extraction for targets with resonant scattering 23. O. Moreno, R. A. Games, and H. Taylor, Sonar sequences from
are considered. It is noted that waveforms for target detection Costas arrays and the best known sonar sequences with up to
100 symbols, IEEE Trans. Inf. Theory, 39: 1985–1987, 1993.
versus information extraction have very different characteris-
24. S. W. Golomb and O. Moreno, On Periodicity Properties of Costas
tics. It is shown that waveforms for target detection should
Arrays and a Conjecture on Permutation Polynomials, Proc.
have as much energy as possible in the target’s largest scat-
IEEE Int. Symp. Inf. Theory, Trondheim, Norway, 1994, p. 361.
tering modes, under the energy and time–bandwidth con-
25. J. Minkoff, Signals, Noise, and Active Sensors: Radar, Sonar, Laser
straints imposed on the system, while waveforms for informa-
Radar, New York: Wiley, 1992.
tion extraction (e.g., target recognition) should have their
26. J. C. Guey and M. R. Bell, Diversity waveform sets for delay-
energy distributed among the target’s scattering modes in
Doppler imaging, IEEE Trans. Inf. Theory, 44: 1504–1522, 1998.
such a way that the information about the target is max-
27. D. V. Sarwate and M. B. Pusley, Crosscorrelation properties of
imized.
pseudorandom and related sequences, Proc. IEEE, 68: 593–619,
1980.
BIBLIOGRAPHY 28. S. W. Golomb, Shift Register Sequences, San Francisco: Holden-
Day, 1967.
1. T. P. Gill, The Doppler Effect, New York: Academic Press, 1965. 29. R. J. McEliece, Finite Fields for Computer Scientists and Engi-
2. P. M. Woodward, Probability and Information Theory, with Appli- neers, Norwell, MA: Kluwer, 1987.
cations to Radar, London: Pergamon, 1953. 30. R. Vartabedian, Unmasking the Stealth Radar, The Los Angeles
3. J. W. Goodman, Fourier Optics, New York: McGraw-Hill, 1968. Times, Sec. D, 1–2, Sun., July 28, 1991.
4. L. Cohen, Time-Frequency Analysis, Upper Saddle River, NJ: 31. M. J. E. Golay, Complementary series, IRE Trans. Inf. Theory, 6:
Prentice-Hall, 1995. 400–408, 1960.
5. A. W. Rihaczek, Principles of High Resolution Radar, New York: 32. R. Turyn, Ambiguity functions of complementary sequences,
McGraw-Hill, 1969; Santa Monica, CA: Mark Resources, 1977. IEEE Trans. Inf. Theory, 9: 46–47, 1963.
6. N. Levanon, Radar Principles, New York: Wiley-Interscience, 33. G. Chandran and J. S. Jaffe, Signal set design with constrained
1988. amplitude spectrum and specified time-bandwidth product, IEEE
7. C. E. Cook and M. Bernfeld, Radar Signals, New York: Academic Trans. Commun., 44: 725–732, 1996.
Press, 1967. 34. J. S. Jaffe and J. M. Richardson, Code-Division Multiple Beam
8. R. E. Blahut, Theory of Remote Surveillance Algorithms, in R. E. Imaging, IEEE Oceans ’89, part 4: Acoustics, Seattle: Arctic Stud-
Blahut, W. Miller, C. H. Wilcox (eds.), Radar and Sonar, part I, ies, 1989, pp. 1015–1020.
New York: Springer-Verlag, 1991. 35. M. Bernfeld, Chirp Doppler radar, Proc. IEEE, 72: 540–541,
9. C. W. Helstrom, Elements of Signal Detection and Estimation, 1984.
Upper Saddle River, NJ: Prentice-Hall, 1995. 36. M. Bernfeld, Tomography in Radar, in F. A. Grünbaum, J. W.
10. R. A. Altes, Target position estimation in radar and sonar, gener- Helton, and P. Khargonekar (eds.), Signal Processing, part II:
alized ambiguity analysis for maximum likelihood parameter es- Control Theory and Applications, New York: Springer-Verlag,
timation, Proc. IEEE, 67: 920–930, 1979. 1990.
11. L. H. Sibul and E. L. Titlebaum, Volume properties for the wide- 37. D. T. Gjessing, Target Adaptive Matched Illumination Radar, Lon-
band ambiguity function, IEEE Trans. Aerosp. Electron. Syst., 17: don: Peregrinus, 1986.
83–86, 1981. 38. S. Sowelam and A. H. Tewfik, Adaptive Waveform Selection for
12. H. Naparst, Dense target signal processing, IEEE Trans. Inf. The- Target Classification, Proc. of EUSIP-96, VIII Eur. Signal Process.
ory, 37: 317–327, 1991. Conf., Trieste, Italy, 1996.
13. L. G. Weiss, Wavelets and wideband correlation processing, IEEE 39. M. R. Bell, Information theory and radar waveform design, IEEE
Sig. Proc. Mag., 11 (4): 13–32, 1994. Trans. Inf. Theory, 39: 1578–1597, 1993.
14. J. R. Klauder, The design of radar signals having both high range
resolution and high velocity resolution, Bell Syst. Tech. J., 808– MARK R. BELL
819, July 1960. Purdue University
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Information Theory of Spread-Spectrum Communication ● Email this page Enter words or
● Print this page
Michael J. Medley1
1Air Force Research Laboratory, Rome, NY
reserved. Search
Abstract
Spread Spectrum Systems
Spreading the Spectrum
Applications

INFORMATION THEORY OF SPREAD-SPECTRUM COMMUNICATION 191
INFORMATION THEORY OF ited, nominal-bandwidth information signal over a much

SPREAD-SPECTRUM COMMUNICATION greater bandwidth. LPI and LPD, combined with appropriate
encryption/decryption techniques, effectively establish the
SPREAD SPECTRUM SYSTEMS basis of military and civilian covert communications.
The transition of interest in SS communications from pri-
Since its inception in the mid-1950s, the term spread spec- marily defense-oriented markets to commercial products and
trum (SS) has been used to characterize a class of digital mod- enterprises has been due in part to two commercially recog-
ulation techniques which satisfy the following criteria (1): nized deficiencies: (1) a way in which to support multiple us-
ers while simultaneously using bandwidth efficiently, and (2)
1. The transmitted signal is characterized by a bandwidth a means of combating multipath fading in mobile communica-
that is much greater than the minimum bandwidth nec- tions. As discussed in the following sections, the autocorrela-
essary to send the information. tion of the SS waveform closely approximates that of an im-
2. Spreading is accomplished prior to transmission using pulse function. This noiselike quality of the spread signal
a spreading sequence, or code, that is independent of facilitates the design and implementation of multiuser/multi-
the information being sent. ple-access communications systems in which several users
3. Detection/demodulation at the receiver is performed by are assigned unique signature codes and are allowed to trans-
correlating the received signal with a synchronously mit simultaneously. At the receiver, each user’s desired signal
generated replica of the spreading code used at the is extracted from the composite sum of all user signals plus
transmitter. noise through correlation with the appropriate signature se-
quence—this description delineates the basis of code-division
Despite what might seem to be an inefficient utilization of multiple-access (CDMA) systems in use today. In light of the
resources, that is, increasing bandwidth without gain over fact that bandwidth is a physically limited commodity, CDMA
noise, the combined process of ‘‘spreading’’ and ‘‘despreading’’ essentially allows the number of users supported by existing
the information-bearing signal offers potential improvement channels to increase independently of bandwidth at the cost
in communications capability that more than offsets the cost of lower performance, that is, higher error rate. Accordingly,
incurred in using additional bandwidth. Indeed, SS offers bandwidth is conserved and, thus, utilized more efficiently.
such benefits as Robustness to multipath is also realized as a result of the SS
waveform’s similarity to white noise. Due to the similarity
between the autocorrelation response of the SS waveform and
• Interference suppression
an impulse function, multiple time-delayed replicas of the
• Power spectral density reduction original signal plus noise can be resolved and coherently com-
• Selective addressing capability bined at the receiver to effectively raise the signal-to-noise
• Resistance to multipath fading ratio (SNR).
Interference suppression refers to the SS system’s ability to Spreading Codes

operate reliably in an environment corrupted or congested by
a level of interference that would compromise the utility of Based on the previous definition of SS systems, it is apparent
conventional digital modulation techniques. In general, SS that some type of code, or sequence, capable of spreading the
signaling is considered robust with respect to interference in information bandwidth must be identified. Here, such codes
the sense that the received signal-to-interference power ratio are discussed—the actual mechanisms by which they effect
is independent of the time-frequency distribution of the inter- bandwidth spreading are the focus of subsequent sections.
ference energy (2). Accordingly, SS systems have found appli- In practice, data-independent pseudorandom, or pseu-
cation in military communications in which hostile sources donoise (PN), sequences govern the spreading and despread-
intentionally jam the channel as well as in civilian settings ing processes. As their name implies, pseudonoise spreading
wherein other users or wireless services inadvertently hinder codes have statistical properties closely approximating those
data transmission through the channel. Due to its effective- of sampled white noise; in fact, to the unintended listener,
ness against a variety of interference sources, including nar- such sequences appear as random binary sequences. Al-
rowband, wideband, multiple-access and multipath interfer- though spreading codes can be generally categorized into two
ence, interference suppression has long been considered the classes, periodic and aperiodic, the most often used spreading
primary advantage of SS communications. codes in contemporary communications systems are periodic
The combined advantages of interference suppression and in nature. This is in part due to the limited number of aperi-
power spectral density reduction go a long way to explain the odic, or Barker, sequences with sufficient length for practical
military’s historical involvement in and application of SS re- applications as well as the availability of simple shift register
search since World War II [although this historical marker structures capable of producing pseudorandom periodic se-
contradicts the mid-1950s date previously espoused, the exact quences (3).
origins of SS communications are rather nebulous and defy In many applications, maximal length sequences, or m-se-
precise attribution regarding date and source of origin (1)]. quences, are often used because of their ease of generation
While interference suppression facilitates reliable operation and good randomness properties. These binary-valued, shift
in hostile environments, power spectral density reduction is register sequences are generated as the output of an n-stage
often exploited to produce low probability of intercept (LPI) maximum-length shift register (MLSR) with a feedback net-
or low probability of detect (LPD) waveforms. Low power work consisting of modulo-2 adders. Due to its cyclic nature,
spectral density is a direct result of spreading the power-lim- an n-stage MLSR produces a periodic sequence with period
192 INFORMATION THEORY OF SPREAD-SPECTRUM COMMUNICATION
X1 X2 X3 function. In this case, the spreading sequence, (0, 0, 1, 1, 1, 0,

Z –1 Z –1 Z –1 Output 1), which is converted to (⫺1, ⫺1, 1, 1, 1, ⫺1, 1) for transmis-
sion, produces the cyclic autocorrelation response given in Eq.
(1) with L ⫽ 7. Further details regarding the origin and im-
+ plementation of m-sequences as well as other potential
mod-2 adder
spreading codes, including Barker, Gold and Kasami se-
quences, are found in the literature (2–7).
Figure 1. Maximum-length shift register with n ⫽ 3 and period,
L ⫽ 2n ⫺ 1 ⫽ 7.
SPREADING THE SPECTRUM
L ⫽ 2n ⫺ 1; L is also the length of the corresponding m-se- As stated in the definition, spread spectrum is accomplished
quence. Each sample in the sequence is called a chip, mean- using a spreading code that is independent of the information
ing that a given m-sequence is 2n ⫺ 1 chips long. Specific being sent. The nature and properties of a common class of
properties of m-sequences include (3): spreading waveforms has been addressed in the preceding
section. Here, the physical mechanisms by which spectrum
Balance Property. In each period of a maximal length se- spreading is accomplished are discussed. Although there are
quence, the number of 1s is always one more than the various approaches to generating spread spectrum wave-
number of 0s. forms, such as direct-sequence (DS) modulation, frequency-
Run Property. Among the runs of 1s and 0s in each period hop (FH) and time-hop (TH) as well as hybrid variants incor-
of an m-sequence, one-half of the runs of each kind are porating aspects of each of these, each approach is fundamen-
of length one, one-fourth are of length two, one-eighth tally based on the underlying spreading code and endeavors
are of length three, and so on, provided these fractions to create a wideband signal from the given information data.
represent meaningful numbers of runs. Of these techniques, DS and FH spread spectrum are most
Correlation Property. The autocorrelation function of an commonly employed. Information regarding other spread
m-sequence is periodic and binary-valued, that is, spectrum formats is presented in (5,7).
Direct-Sequence Spread Spectrum
+L k = iL
R[k] = (1)
−1 k = iL In direct-sequence spread spectrum (DS-SS), the wideband
transmitted signal is obtained by multiplying the binary base-
where i is any integer and L is the length of the code. band information signal, b(t), by the spreading code as shown
Note that this expression is valid for all m-sequences in Fig. 3. Note that this figure inherently incorporates the use
regardless of the value of L. of binary phase-shift keying (BPSK) modulation and is thus
representative of a DS/BPSK spread spectrum system; more
Figure 1 illustrates an n ⫽ 3-stage MLSR as an example. generally, the combination of DS-SS with M-ary PSK modula-
Assuming that the MLSR initial state is set to [X1, X2, X3] ⫽ tion is referred to as DS/MPSK-SS. Although practical DS/
[1, 0, 0], the resulting binary output sequence, determined by MPSK-SS systems often modulate individual data bits using
cycling through successive register states [X1, X2, X3] ⫽ [1, 0, only a portion of the total m-sequence, it is assumed here, for
0], [1, 1, 0], [1, 1, 1], [0, 1, 1], [1, 0, 1], [0, 1, 0] and [0, 0, 1], convenience, that each bit is modulated by a single, full-
is (0, 0, 1, 1, 1, 0, 1). Successive iterations return the MLSR length m-sequence with the number of chips in the spreading
state to its initial value, [1, 0, 0], wherein the process as well code equal to its length, L. With the bit period defined as Tb,
as the resulting output sequence begin to repeat. The MLSR the number of chips per bit is given by the ratio Tb /Tc ⫽ L
output sequence is thus periodic in the sense that the L ⫽ 7- where Tc is the chip duration. The rate of the DS-SS wave-
chip m-sequence, (0, 0, 1, 1, 1, 0, 1), is repeated every seven form, also called the chip rate of the system, is Rc ⫽ 1/Tc
iterations as long as the MLSR is allowed to run. Clearly, the chips/s.
spreading sequence (0, 0, 1, 1, 1, 0, 1) contains four ones and In practice, the bit duration, Tb, is typically much greater
three zeros as is consistent with the balance property. Like- than Tc. Consequently, the chip rate is often orders of magni-
wise, the total presence of four runs—two of length one, one tude larger than the original bit rate Rb ⫽ 1/Tb, thus necessi-
of length two, and one of length three essentially meets the tating the increase, or spread, in transmission bandwidth. As
specifications of the run property. shown in Fig. 4, the frequency response of the spread wave-
As an illustration of the correlation property, Fig. 2 depicts form has a sinc(x) ⫽ sin(x)/x shape with main lobe bandwidth
a 7-chip m-sequence and its corresponding autocorrelation equal to 2Rc. Pulse shaping can be used to minimize the side-
lobes and effectively reduce the DS-SS bandwidth if nec-
essary.
7 Given the baseband information signal, b(t), and the
1 spreading code, c(t), the DS-SS waveform is given by
m(t) = c(t)b(t) (2)

Time –1
–1 Time-lag (k)
Subsequent transmission over a noisy channel corrupted by
7-chip pseudonoise sequence Autocorrelation sequence
interference produces the receiver input
Figure 2. Length L ⫽ 7 m-sequence and corresponding cyclic auto-
correlation response. r(t) = m(t) + i(t) + n(t) (3)
At the transmitter:
Tb
Time
–1
Binary input data: b(t)
Tc
1 ...
... Time
–1
7-chip pseudo noise sequence: c(t)
Time
–1
DS modulated baseband waveform: m(t)
At the receiver (assuming no channel noise and perfect synchronization):
Time
–1
DS modulated baseband waveform: m(t)
1 ...
Time
–1 ...
Local replica of 7-chip pseudo noise sequence: c(t)
7
sample
1
–1 Time
sample sample
–7 Correlator output sequence: R(k)
Time
–1 Figure 3. Direct-sequence spread spec-
Detected data sequence : b(t)
trum modulation and demodulation.
where i(t) and n(t) denote interference and white noise, re- At the receiver, demodulation, or despreading, as depicted in
spectively. Often when using SS signaling, the interference Fig. 3 in the absence of noise and interference, is accom-
power is assumed to be much greater than that of the noise plished by multiplying r(t) with a synchronized replica of the
and this expression is simplified as spreading code, that is,
r(t) = m(t) + i(t) (4) u(t) = c(t)r(t) (6)
= c(t)b(t) + i(t) (5) = c (t)b(t) + c(t)i(t)
2
(7)
= b(t) + c(t)i(t) (8)
with the final equality a result of the relationship, c2(t) ⫽ 1

Original spectrum
for all t. Subsequent integration of u(t) over each symbol pro-
duces the correlator output which, when sampled at the ap-
Spread spectrum propriate instances, yields the detected data sequence. The
2Rc preceding steps demonstrate that multiplying a signal once
f
by the spreading code spreads its energy across a much larger
Figure 4. Magnitude-squared frequency response of a DS-SS bandwidth while a second multiplication reverses this process
waveform. by despreading the energy and restoring the spread waveform
DS/BPSK transmitter DS/BPSK receiver
m(t) ^
b(t) BPSK x(t) y(t) BPSK b(t)
BPF
modulator demodulator
Correlator
Figure 5. Synchronized DS/BPSK ^
transmitter/receiver structures. c(t) Carrier c(t – Td) Local carrier
to its original, prespread condition. Equation (8) verifies that Rb bits/s, Gp can be approximated in DS-SS systems by the
the information signal, b(t), which is multiplied twice by the ratio of the chip rate to the data rate,
spreading code, is recovered, and returned to its initial state,
Rc T
while the interference, which is multiplied only once, under- Gp ≈ = b =N (12)
goes spreading due to c(t). Rb Tc
Whereas the previous discussion has focused on baseband
signals, practical implementations typically modulate the where N corresponds to the number of chips per spread data
baseband SS waveform onto a sinusoidal carrier as dia- bit; N ⫽ L when individual data bits are modulated by the
grammed in Fig. 5. Here, sinusoidal modulation produces the entire spreading sequence. Note that, in practice, the entire
spreading code may not be used to modulate each data bit
DS/BPSK SS signal,
(depending on the application, a subset of K ⬍ L chips may
√ be used). In essence, Gp roughly gauges the antijam capability
x(t) = 2Pm(t) cos 2π f c t (9)
and LPI/D quality of the SS system.
System performance is ultimately a function of SNRo,
where P denotes the average power and f c is the carrier fre-
which determines the bit-error-rate (BER) experienced by the
quency. The receiver input is thus the bandpass waveform
communication link. For a given data rate, spreading the
transmitted signal energy over a larger bandwidth allows the
y(t) = x(t) + i(t) + n(t) (10)
receiver to operate at a lower value of SNRi. The range of
SNRi for which the receiver can provide acceptable perfor-
Figure 5 illustrates correlation as performed at the receiver
mance is determined by the jamming margin, MJ, which is
by multiplying the received signal with a synchronized copy
expressed in decibels (dB) as
of the spreading code, c(t ⫺ T̂d), where T̂d represents the esti-
mated propagation delay of the transmitted signal, and band- MJ = G p − [SNRo min + Lsys] (13)
pass filtering to remove out-of-band components. Subsequent
BPSK demodulation produces the estimate of the transmitted where SNRomin is the minimum SNRo required to support the
data sequence, b̂(t). maximum allowable BER, and Lsys accounts for any losses due
Synchronization between the received signal and the to receiver implementation. Hence, in addition to Gp, MJ rep-
spreading sequence is typically performed in two stages: (1) resents another metric available to system designers indicat-
an acquisition stage, which provides coarse alignment being how much interference can be tolerated while still main-
tween the two waveforms, typically to within a fraction of a taining a prescribed level of reliability.
chip, and (2) a tracking stage, which maintains fine synchro-
nization and, essentially, the best possible alignment be- Frequency-Hop Spread Spectrum
tween y(t) and c(t). Rudimentary discussions of synchroniza- In contrast to DS-SS, which directly employs the spreading
tion techniques for SS systems are presented in (4,5), while sequence to modulate a phase-shift-keyed version of the infor-
more in-depth expositions are found in (5,8). mation bearing waveform, frequency-hop spread spectrum
As demonstrated in Eq. (8), multiplication of the received (FH-SS) utilizes the spreading code to determine the carrier
signal with a locally generated, synchronized copy of the frequency, or frequency slot, used to transmit data over a spe-
spreading code simultaneously collapses the spread data sig- cific period of time. In this manner, a broadband signal is
nal back to its original bandwidth while spreading any addi- generated by sequentially moving, or hopping, a fixed-fre-
tive noise or interference to the full SS bandwidth or greater. quency data-modulated carrier throughout the frequency
As shown in Fig. 5, a bandpass filter with bandwidth matched spectrum as directed by a pseudorandom pattern known (ide-
to that of the original data is subsequently used to recover ally) only to the transmitter and its intended receivers. Fig-
the data and reject a large fraction of the spread interference ure 6 shows the idealized frequency spectrum of a FH-SS sig-
energy. The ratio of the signal-to-noise ratio (SNR) after de-
spreading, SNRo, to the input signal-to-noise ratio, SNRi, is
defined as the processing gain, Gp, that is, fh
SNRo
Gp = (11)
SNRi
Note that in both SNRi and SNRo the noise term implicitly
Nfh f
denotes the sum of additive white Gaussian noise (AWGN)
plus any additional interference. Given an input data rate of Figure 6. Idealized frequency spectrum of a FH-SS waveform.
nal in which the N hop frequencies are equally spaced at several MFSK symbols are transmitted per hop with the chip
intervals of f h Hz—the spread bandwidth of this signal is rate, Rc, equal to the MFSK symbol rate, Rs. The converse is
thus Nf h Hz; in practice, each of the illustrated tones is re- true in FFH/MFSK-SS, wherein several hops are performed
placed by the actual narrowband signal spectrum associated per MFSK symbol, with the resulting chip rate equal to the
with the underlying narrowband modulation scheme em- hopping rate, Rh.
ployed. In the example of SFH/MFSK-SS shown in Fig. 8, the in-
The modulation format most often used in conjunction formation signal b(t), whose bit rate, Rb, is related to the bit
with FH-SS is M-ary FSK (MFSK); this combination is simply duration, Tb, via Rb ⫽ 1/Tb, is segmented into two-bit pairs
referred to as FH/MFSK. Figure 7 depicts typical FH/MFSK which select the frequency (one out of four possible frequen-
transmitter and receiver block diagrams. In the FH/MFSK cies assuming M ⫽ 4-FSK modulation) to be transmitted.
transmitter, k ⫽ log2M information bits determine which of Since two bits are required per MFSK output, the duration
the M frequencies of the MFSK modulator is to be transmit- of each symbol is Ts ⫽ 2Tb yielding the symbol rate, Rs ⫽
ted. The function of the frequency synthesizer is to produce a 1/Ts ⫽ Tb (note that Rs is equivalent to f h of Fig. 6). Using
sinusoidal waveform, or tone, which when mixed with the the periodically repeated m-sequence generated by the MLSR
MFSK modulator output effectively shifts its position in fre- in Fig. 1, that is, the sequence (0, 0, 1, 1, 1, 0, 1, 0, . . .) with
quency. Note that the mixing operation as well as the re- boldface type denoting the first period, the output of the
quired bandpass filtering is performed by the up-converter. MFSK modulator is hopped through N ⫽ 8 different fre-
As might be surmised, the frequency of the synthesizer output quency slots. To determine the hopping pattern, the PN se-
is pseudorandomly determined by the PN generator driving quence is divided into successive (not necessarily disjoint)
it. Typically, j ⫽ log2 N chips of the spreading code are fed k ⫽ 3 bit segments, each indicating the particular frequency
into the frequency synthesizer to select one of N possible slot to be used; in this case, frequency assignment is unique
tones. The FH/MFSK receiver shown in Fig. 7 simply re- since N ⫽ 2k. Below the resulting SFH/MFSK waveform dia-
verses the processes performed in the transmitter by down- gram in Fig. 8 are the corresponding 3-bit segments, 001,
converting the received signal with a locally generated copy of
110, 100, 111, 010, . . . governing the hopping pattern. Note
the tone used at the transmitter and subsequently performing
that in this example the 000 sequence never appears and thus
conventional MFSK demodulation to produce the estimated
N is effectively only seven; such an aberration is seldom en-
information signal, b̂(t).
countered in practice and, even if it were, the general princi-
As discussed above, at any instant in time, the actual
ple illustrated here would still be valid. In this example,
amount of bandwidth used in FH/MFSK signaling is identical
two symbols are transmitted per hop. Thus, the hop dura-
to that of conventional MFSK. This bandwidth is much less
than the effective FH-SS bandwidth realized by averaging tion, Th ⫽ 2Ts with the corresponding hop rate given by Rh ⫽
over many hops. Recognizing that the total number of possi- 1/Th ⫽ Rs /2. The effective FH-SS bandwidth is Bss ⫽ NRs.
ble tones is N ⫽ 2j, the FH/MFSK-SS bandwidth is roughly Figure 9 illustrates FFH/MFSK-SS signaling. As in the
Nf h and, in practice, is limited primarily by the operational SFH/MFSK example, two-bit pairs from b(t) drive the MFSK
range of the frequency synthesizer. Frequency hopping over modulator thus again yielding the symbol duration, Ts ⫽
very large bandwidths typically precludes the use of coherent 2Tb. In contrast to the previous example, however, multiple
demodulation techniques due to the inability of most fre- hops in frequency occur per MFSK symbol. Although fre-
quency synthesizers to maintain phase coherence over succes- quency hop assignment is again governed by the periodic PN
sive hops. Consequently, noncoherent demodulation is usually sequence segmented into the 3-bit patterns, 001, 110, 100,
performed at the receiver (5). 111, 010, 011, 101, 001, . . . (boldface denotes initial register
Whereas the term chip in DS-SS corresponds to the sam- states associated with the 7-chip m-sequence), in this exam-
ples of the spreading sequence, in FH-SS, it refers to the FH/ ple, two 3-bit patterns are used per symbol; the actual 3-bit
MFSK tone with the shortest duration. The amount of time pairs used per symbol are listed below the FFH/MFSK wave-
spent at each hop determines whether the FH/MFSK system form diagram. Accordingly, the hop duration, Th ⫽ Ts /2, with
is classified as slow frequency-hopping (SFH/MFSK) or fast Rh ⫽ Rb. The overall FH-SS bandwidth, which is independent
frequency-hopping (FFH/MFSK). In SFH/MFSK systems, of the hop rate, is again Bss ⫽ NRs.
FH/MFSK transmitter FH/MFSK receiver
MFSK Up Down MFSK ^

b(t) converter x(t) y(t) converter modulator b(t)
modulator
Frequency Frequency
synthesizer synthesizer
PN PN
generator generator Figure 7. Synchronized FH/MFSK
transmitter/receiver structures.
b(t) Tb
1 0 0 1 1 1 0 1 0 0 1 0 0 0 1 1 0 1 1 1
Time
MFSK modulator output

11
10
Rs
01
00
Ts Time
SFH/MFSK waveform Time
111 Rs
Th
110
101
100
Frequency
Bss = NRs
011
010
001
000
Figure 8. SFH/MFSK modulation with

M ⫽ 4, N ⫽ 8, and a dwell time of 2Tb s. PN: 001 110 100 111 010
In general, FFH/MFSK-SS is an effective means of com- Bss

Gp ≈ (14)
bating certain types of jammers called follower and repeat- Rb
back jammers which attempt to intercept the frequency of the
transmitted waveform and retransmit it along with addi- Assuming that the interference energy is spread over the en-
tional frequencies so as to degrade receiver performance (4). tire FH bandwidth and that the original data rate is approxi-
When using FFH/MFSK-SS, the jammers do not typically mately equal to the symbol rate, Rb ⫽ Rs, the processing gain
have sufficient time to intercept and jam the spread waveform for either FH-SS system shown in Fig. 8 and Fig. 9 is approxi-
before it hops to another frequency. The price paid for such mately equal to N, the number of different frequencies over
evasion, however, is the need for fast-frequency synthesizers which the MFSK modulator output is hopped. The expression
capable of changing frequency at the required hopping rates. for the jamming margin, MJ, as given in Eq. (13), holds.
As in DS-SS, the processing gain, Gp, serves as a metric
indicating the signaling scheme’s robustness with respect to APPLICATIONS
interference. For either fast-FH or slow-FH, the effective pro-
cessing gain can be approximated as the ratio of the spread Primary applications of spread spectrum in contemporary
spectrum bandwidth, Bss, to the original data rate, Rb, that is, communications include antijam (AJ) communications, code-
division multiple-access (CDMA), and multipath interfer- other hand, is limited in spread bandwidth and, thus, pro-
ence rejection. Not surprisingly, each of these applications cessing gain, only by the operational limits of the frequency
has been directly foreshadowed by the list of attributes synthesizer. In practice, physical implementations of FH-SS
associated with SS signaling presented at the beginning of are typically capable of sustaining wider bandwidth signals
this topic. than practical DS-SS systems.
Even though SS systems possess a fundamental level of
inherent interference immunity, different types of interfer-
AntiJam Communications ence pose threats to DS-SS and FH-SS systems. In particular,
pulsed-noise jammers are especially effective against DS/
As previously discussed, the AJ capability of a SS system is MPSK systems, while partial-band and multitone interfer-
directly related to its overall processing gain, Gp. Although in ence, perhaps due to CDMA overlay and/or narrowband ser-
theory the processing gain associated with a DS-SS waveform vices present within the SS bandwidth, represent significant
can be arbitrarily increased by using longer spreading codes, threats to reliable FH/MFSK systems. Additional sources of
stringent synchronization requirements and practical band- interference, as well as their effects on SS communications,
width considerations limit its availability. FH-SS, on the are found throughout the literature (4–7).
Tb
1 0 0 1 1 1 0 1 0 0 1 0 0 0 1 1 0 1 1 1
Time
MFSK modulator output

11
10
Rs
01
00
Ts Time
FFH/MFSK waveform Time
Th
111 Rs
110
101
100
Frequency
Bss = NRs
011
010
001
000
Figure 9. FFH/MFSK modulation with

PN: 001/110 100/111 010/011101/001110/100 111/010101/001110/100111/010010/011 M ⫽ 4, N ⫽ 8, and a dwell time of Tb s.
198 INFRARED DETECTOR ARRAYS, UNCOOLED
Code-Division Multiple-Access FH-SS can also be used to combat multipath interference

provided that the transmitted signal hops fast enough rela-
Prior to the introduction of code-division multiple-access
tive to the differential time delay between the direct path and
(CDMA), conventional multiple-access techniques focused on
multipath signal components. In this case, much of the
dividing the available time or frequency space into disjoint
multipath energy falls into frequency slots vacated by the FH-
partitions and assigning them to individual users. In time-
SS waveform and, thus, its effect on the demodulated signal
division multiple-access (TDMA), users are multiplexed in
is minimized (3).
time and allowed to transmit sequentially over a given chan-
nel. In contrast, in frequency-division multiple-access
(FDMA), each user is assigned a portion of the channel band- BIBLIOGRAPHY
width, separated from other users by a guard band, and al-
lowed to use the channel simultaneously without interfering 1. R. A. Scholtz, The origins of spread-spectrum communications,
with one another. As opposed to partitioning either the time IEEE Trans Commun., COM-30: 822–854, 1982.
or frequency plane, CDMA provides both time and frequency 2. R. L. Pickholtz, D. L. Schilling, and L. B. Milstein, Theory of
diversity to its users through the use of spread spectrum mod- spread-spectrum communications—A tutorial, IEEE Trans. Com-
ulation techniques. mun., COM-30: 855–884, 1982.
In CDMA, each user is assigned a pseudorandom signature 3. S. Haykin, Digital Communications, New York: Wiley, 1988.
code, or sequence, similar in structure to the m-sequences dis- 4. J. G. Proakis, Digital Communications, 3rd ed., New York:
cussed earlier. Gold codes and Kasami sequences, like m-se- McGraw-Hill, 1995.
quences, have impulselike autocorrelation responses and are 5. M. K. Simon et al., Spread Spectrum Communications Handbook,
frequently used in such applications. Unlike m-sequences, New York: McGraw-Hill, 1994.
however, these codes are generated as a set of spreading 6. B. Sklar, Digital Communications Fundamentals and Applications,
codes whose members possess minimal cross-correlation prop- Englewood Cliffs, NJ: Prentice-Hall, 1988.
erties (4). Low cross-correlation among multiple users allows 7. R. E. Ziemer and R. L. Peterson, Digital Communications and
them to communicate simultaneously without significantly Spread Spectrum Systems, New York: Macmillan, 1985.
degrading each other’s performance. In contrast to TDMA, 8. R. C. Dixon, Spread Spectrum Systems with Commercial Applica-
CDMA does not require an external synchronization network tions, 3rd ed., New York: Wiley-Interscience, 1994.
and it offers graceful degradation as more users are added to
the channel (due to the fact that since the spreading codes Reading List
approximate wideband noise, each additional CDMA user ap-
C. E. Cook and H. S. Marsh, An introduction to spread-spectrum,
pears as an additional noise source which incrementally IEEE Comm. Mag., 21 (2): 8–16, 1983.
raises the noise floor of the channel). In addition, CDMA also
J. K. Holmes, Coherent Spread Spectrum Systems, New York: Wiley,
offers the benefits of SS communications including resistance 1982.
to multipath as well as jamming.
A. J. Viterbi, Spread spectrum communications—Myths and realities,
IEEE Commun. Mag., 17 (3): 11–18, 1979.
Multipath Suppression
MICHAEL J. MEDLEY
In many communications systems, actual data transmission Air Force Research Laboratory
occurs along direct, line-of-sight paths, as well as from a num-
ber of physical paths which are the result of reflections of the
transmitted signal off of various scatterers such as buildings,
trees, and mobile vehicles. Multipath interference is a result INFORMATION VISUALIZATION. See DATA VISUAL-
of the combination of these direct and indirect signal trans- IZATION.
missions arriving at the receiver at a slight delay relative to INFRARED. See PHOTODETECTORS QUANTUM WELL.
each other. When the direct path signal is substantially
stronger than the reflected components, multipath does not
represent much of a challenge, if any, to reliable communica-
tions. When the direct path signal is either nonexistent or,
more likely, comparable in strength to the indirect, delayed
components, however, multipath interference results in varia-
tions in the received signal’s amplitude, which is called
fading.
Under slow fading conditions, multipath can be combatted
directly through the use of DS-SS. Due to the noiselike prop-
erty of the DS-SS waveform, multipath signal components,
when correlated with the local reference code, can be resolved
in time (provided the multipath spread is greater than a chip
duration) and combined coherently to improve data detection.
Under these conditions, the degradation in receiver perfor-
mance due to multipath is directly related to the chip rate
associated with DS modulation—the greater the chip rate,
the less effect multipath will have on performance.
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Information Theory of Stochastic Processes ● Email this page Enter words or
● Print this page
John C. Kieffer1
1University of Minnesota
reserved. Search
Abstract
Asymptotic Equipartition Property
Information Stability Property
Application to Source Coding Theory
Application to Channel Coding Theory
Final Remarks

J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering
Copyright c 1999 John Wiley & Sons, Inc.
INFORMATION THEORY OF STOCHASTIC PROCESSES

This article starts by acquainting the reader with the basic features in the design of a data communication
system and discusses, in general terms, how the information theory of stochastic processes can aid in this
design process. At the start of the data communication system design process, the communication engineer
is given a source, which generates information, and a noisy channel through which this information must be
transmitted to the end user. The communication engineer must then design a data communication system
so that the information generated by the given source can be reliably transmitted to the user via the given
channel. System design consists in finding an encoder and decoder through which the source, channel, and end
user can be linked as illustrated in Fig. 1.
To achieve the goal of reliable transmission, the communication engineer can use discrete-time stochastic
processes to model the sequence of source outputs, the sequence of channel inputs, and the sequence of channel
outputs in response to the channel inputs. The probabilistic behavior of these processes can then be studied over
time. These behaviors will indicate what level of system performance can be achieved by proper encoder/decoder
design. Denoting the source in Fig. 1 by (S and denoting the channel in Fig. 1 by C, one would like to know the
rate R(S) at which the source generates information, and one would like to know the maximum rate R(C) at
which the channel can reliably transmit information. If R(S) ≤ R(C), the design goal of reliable transmission of
the source information through the given channel can be achieved.
Information theory enables one to determine the rates R(S) and R (C). Information theory consists of
two subareas—source coding theory and channel coding theory. Source coding theory concerns itself with the
computation of R (S) for a given source model S, and channel coding theory concerns itself with the computation
of R(C) for a given channel model C.
Suppose that the source generates an output U i at each discrete instant of time i = 1, 2, 3, . . .. The discrete-
time stochastic process {U i : i ≥ 1} formed by these outputs may obey an information-theoretic property called
the asymptotic equipartition property, which will be discussed in the section entitled “Asymptotic Equipartition
Property.” The asymptotic equipartition property will be applied to source coding theory in the section entitled
“Application to Source Coding Theory.” If the asymptotic equipartition property is satisfied, there is a nice way
to characterize the rate R(S) at which the source S generates information over time.
Suppose that the channel generates a random output Y i at time i in response to a random input X i at
time i, where i = 1, 2, 3, . . .. The discrete-time stochastic process {(X i , Y i ): i ≥ 1} consisting of the channel
input–output pairs (called a channel pair process) may obey an information-theoretic property called the
information stability property, which shall be discussed in the section entitled “Information Stability Property.”
The information stability property will be applied to channel coding theory in the section entitled “Application
to Channel Coding Theory.” If sufficiently many channel pair processes obey the information stability property,
there will be a nice way to characterize the rate R (C) at which the channel C can reliably transmit information.
In conclusion, the information theory of stochastic processes consists of the development of the asymptotic
equipartition property and the information stability property. In this article we discuss these properties, along
with their applications to source coding theory and channel coding theory.
1
2 INFORMATION THEORY OF STOCHASTIC PROCESSES
Fig. 1. Block diagram of data communication system.
Asymptotic Equipartition Property
If the asymptotic equipartition property holds for a random sequence {Ui: i ≥ 1}, then, for large n, the random
vector (U 1 , U 2 , . . ., U n ) will be approximately uniformly distributed. In order to make this idea precise, we
must first discuss the concept of entropy.
Entropy. Let U be a discrete random variable. We define a nonnegative random variable h(U), which is
a function of U, so that
whenever U = u. The logarithm is taken to base two (as are all logarithms in this article). Also, we adopt the
convention that h(U) is defined to be zero, whenever Pr[ U = u] = 0. The random variable h(U) is called the
self-information of U.
The expected value of h(U) is called the entropy of U and is denoted H(U). In other words,
where E (here and elsewhere) denotes the expected value operator. Certainly, H(U) satisfies
We shall only be interested in the finite entropy case in which H(U) < ∞. One can deduce that U has
finite entropy if U takes only finitely many values. Moreover, the bound
holds in this case, where N is the number of values of U. To see why Eq. (1) is true, we exploit Shannon’s
inequality, which says
whenever {p(u)} and {q(u)} are probability distributions on the space in which U takes its values. In Shannon’s
inequality, take
for each value u of U, thereby obtaining Eq. (1). If the discrete random variable U takes on a countably infinite
number of values, then H(U) may or may not be finite, as the following examples show.
INFORMATION THEORY OF STOCHASTIC PROCESSES 3
Example 1. Let the set of values of U be 2, 3, 4, . . .}, and let
for every value u of U, where C is the normalization constant that makes these probabilities sum to one. It can
be verified that H(U) = ∞.
Example 2. Let U follow a geometric distribution
where p is a parameter satisfying 0 < p < 1. It can be verified that
We are now ready to discuss the asymptotic equipartition property. Let {U i : i ≥ 1} be a discrete-time
stochastic process, in which each random variable U i is discrete. For each positive integer n, let U n denote
the random vector (U 1 , U 2 , . . ., U n ). (This notational convention shall be in effect throughout this article.) We
assume that the process {U i : i ≥ 1} obeys the following two properties:
(1) H(U n ) < ∞, n ≥ 1.

(2) The sequence {H(U n )/n: n ≥ 1} has a finite limit.
Under this assumption, we can define a nonnegative real number by
The number is called the entropy rate of the process {U i : i ≥ 1}. Going further, we say that the process
{U i : i ≥ 1} obeys the asymptotic equipartition property (AEP) if
What does the AEP tell us? Let ε be a fixed, but arbitrary, positive real number. The AEP implies that we
may find, for each positive integer n, a set En consisting of certain n-tuples in the range of the random vector
U n , such that the sets {En } obey the following properties:
(2.3) lim n→∞ Pr[U N ∈ EN ] = 1. For each n, if un is an n-tuple in En , then (2.5) For sufficiently large n,
if |En | is the number of n-tuples in En , then
In loose terms, the AEP says that for large n, U n can be modeled approximately as a random vector taking
roughly 2nH equally probable values. We will apply the AEP to source coding theory in the section entitled
“Application to Source Coding Theory.”
Example 3. Let {U i : i ≥ 1} consist of independent and identically distributed (IID) discrete random
variables. Letting H(U 1 ) < ∞, assumptions (2.1) and (2.2) hold, and the entropy rate is = H(U 1 ). By the law
of large numbers, the AEP holds.
Example 4. Let {U i : i ≥ 1} be a stationary, ergodic homogeneous Markov chain with finite state space.
Assumptions (2.1) and (2.2) hold, and the entropy rate is given by = H(U 2 ) − H(U 1 ). Shannon (1) proved that
the AEP holds in this case.
Extensions. McMillan (2) established the AEP for a stationary ergodic process {U i : i ≥ 1} with finite
alphabet. He established L1 convergence, namely, he proved that
which is a stronger notion of convergence than the notion of convergence in Eq. (3). In the literature, McMillan’s
result is often referred to as the Shannon–McMillan Theorem. Breiman (3) proved almost sure convergence of
the sequence {n − 1 h(U n ): n ≥ 1} to the entropy rate :, for a stationary ergodic finite alphabet process {U i : i
≥ 1}. This is also a notion of convergence that is stronger than Eq. (3). Breiman’s result is often referred to as
the Shannon–McMillan–Breiman Theorem. Gray and Kieffer (4) proved that a type of nonstationary process
called an asymptotically mean stationary process obeys the AEP. Verdú and Han (5) extended the AEP to a
class of information sources called flat-top sources. Many other extensions of the AEP are known. Most of these
results fall into one of the three categories described below.
(1) AEP for Random Fields. A random field {U g : g ∈ G} is given in which G is a countable group, and there is
a finite set A such that each random variable U g takes its values in A. A sequence {F n : n ≥ 1} of growing
finite subsets of G is given in which, for each n, the number of elements of F n is denoted by |F n |. For each
n, let U n denote the random vector
One tries to determine conditions on {U g } and {F n } under which the sequence of random variables {|F n | − 1 h
(U Fn ): n ≥ 1} converges to a constant. Results of this type are contained in Refs. (6) (L1 convergence) and
(7) (almost sure convergence).
(2) Entropy Stability for Stochastic Processes. Let {U i : i ≥ 1} be a stochastic process in which each random
variable U i is real-valued. For each n = 1, 2, . . ., suppose that the distribution of the random vector U n is
absolutely continuous, and let F n be its probability density function. For each n, let gn be an n-dimensional
probability density function different from F n . One tries to determine conditions on {U i } and {gn } under
which the sequence of random variables
converges to a constant. A process {U i : i ≥ 1} for which such convergence holds is said to exhibit the entropy
stability property (with respect to the sequence of densities {gn }). Perez (8) and Pinsker [(9), Sections 7.6,
8.4, 9.7, 10.5, 11.3] were the first to prove theorems showing that certain types of processes {U i : i ≥ 1}
exhibit the entropy stability property. Entropy stability has been studied further (10 11 12 13 14,15. In the
textbook (16), Chapters 7 and 8 are chiefly devoted to entropy stability.
(3) Entropy Stability for Random Fields. Here, we describe a type of result that combines types (i) and (ii). As
in (i), a random field {U g : g ∈ G} and subsets {F n : n ≥ 1} are given, except that it is now assumed that
each random variable U g is real-valued. It is desired to find conditions under which the sequence of random
variables
converges to a constant, where, for each n, F n is the probability density function of the |F n |-dimensional
random vector U F n and gn is some other |F n |-dimensional probability density function. Tempelman (17)
gave a result of this type.
Further Reading. In this article, we have focused on the application of the AEP to communication
engineering. It should be mentioned that the AEP and its extensions have been exploited in many other areas
as well. Some of these areas are ergodic theory (18,19), differentiable dynamics (20), quantum systems (21),
statistical thermodynamics (22), statistics (23), and investment theory (24).
Information Stability Property
The information stability property is concerned with the asymptotic information-theoretic behavior of a pair
process, that is, a stochastic process {(X i , Y i ): i ≥ 1} consisting of pairs of random variables. In order to discuss
the information stability property, we must first define the concepts of mutual information and information
density.
Mutual Information. Let X, Y be discrete random variables. The mutual information between X and Y,
written I(X; Y), is defined by
where we adopt the convention that all terms of the summation in which Pr[X = x, Y = y] = 0 are taken to
be zero. Suppose that X, Y are random variables that are not necessarily discrete. In this case, the mutual
information I(X; Y) is defined as
where the supremum is taken over all pairs of random variables (X d , Y d ) in which X d , Y d are discrete functions
of X, Y, respectively. From Shannon’s inequality, Eq. (2), I(X; Y) is either a nonnegative real number or is +∞.
We shall only be interested in mutual information when it is finite.
Example 5. Suppose X and Y are independent random variables. Then I(X;Y) = 0. The converse is also
true.
Example 6. Suppose X is a discrete random variable. The inequality
always holds. From this inequality, we see that if H (X) or H (Y) is finite, then I (X;Y) is finite. In particular,
we see that I (X;Y) is finite if either X or Y take finitely many values.
Example 7. Suppose X, Y are real-valued random variables, with variances σ2 x > 0, σ2 y > 0, respectively.
Let (X, Y) have a bivariate Gaussian distribution, and let ρxy be the correlation coefficient, defined by
It is known (9, p. 123) that
In this case, we conclude that I(X;Y) < ∞ if and only if −1 < ρxy < 1.
Example 8. Suppose X and Y are real-valued random variables, and that (X, Y) has an absolutely
continuous distribution. Let f (X, Y) be the density function of (X, Y), and let f (X) and g(Y) be the marginal
densities of X, Y, respectively. It is known (9, p. 10) that
Information Density. We assume in this discussion that X, Y are random variables for which I(X;Y) < ∞.
The information density i(X;Y) of the pair (X, Y) shall be defined to be a random variable, which is a function
of (X, Y) and for which
In other words, the expected value of the information density is the mutual information. Let us first define
the information density for the case in which X and Y are both discrete random variables. If X = X and Y = Y,
we define
Now suppose that X, Y are not necessarily discrete random variables. The information density of the pair
(X, Y) can be defined (16, Chap. 5) as the unique random variable I(X;Y) such that, for any ε > 0, there exist
discrete random variables X ε , Y ε , functions of X, Y, respectively, such that
whenever X , Y are discrete random variables such that
• X ε is a function of X and X is a function of X.

• Y ε is a function of Y and Y is a function of Y.
Example 9. In Example 8, if I(X; Y) < ∞, then
Example 10. If X is a discrete random variable with finite entropy, then
We are now ready to discuss the information stability property. Let {(X i , Y i ): i ≥ 1} be a pair process
satisfying the following two properties:
(1) (10.1)I(X n ; Y n ) < ∞, n ≥ 1.

(2) (10.2)The sequence {n − 1 I(X n ; Y n ): n ≥ 1} has a finite limit.
We define the information rate of the pair process [(X i , Y i ): I ≥ 1 ] to be the nonnegative real number
A pair process [(X i , Y i ): I ≥ 1 ] satisfying (10.1) and (10.2) is said to obey the information stability property
(ISP) if
We give some examples of pair processes obeying the ISP.

Example 11. Let the stochastic process [X i : I ≥ 1 ] and the stochastic process [Y i : I ≥ 1 ] be statistically
independent. For every positive integer n, we have I(X n ; Y n ) = 0. It follows that the pair process [(X i , Y i ): I ≥
1 ] obeys the ISP and that the information rate is zero.
Example 12. Let us be given a semicontinuous stationary ergodic channel through which we must
transmit information. “Semicontinuous channel” refers to the fact that the channel generates an infinite
sequence of random outputs [Y i } from a continuous alphabet in response to an infinite sequence of random
inputs {X i } from a discrete alphabet. “Stationary ergodic channel” refers to the fact that the channel pair
process {(X i , Y i )} will be stationary and ergodic whenever the sequence of channel inputs {X i } is stationary
and ergodic. Suppose that {X i } is a stationary ergodic discrete-alphabet process, which we apply as input to
our given channel. Let [Y i ] be the resulting channel output process. In proving a channel coding theorem (see
the section entitled “Application to Channel Coding Theory”), it could be useful to know whether the stationary
and ergodic pair process {(X i , Y i ): I ≥ 1} obeys the information stability property. We quote a result that allows
us to conclude that the ISP holds in this type of situation. Appealing to Theorems 7.4.2 and 8.2.1 of (9), it is
known that a stationary and ergodic pair process [(X i , Y i ): I ≥ 1 ] will obey the ISP provided that X 1 is discrete
with H(X 1 ) < ∞. The proof of this fact in (9) is too complicated to discuss here. Instead, let us deal with the
special case in which we assume that Y 1 is also discrete with H(Y 1 ) < ∞. We easily deduce that [(X i ,Y i ): I ≥ 1
] obeys the ISP. For we can write
for each positive integer n. Due to the fact that each of the processes {X i }, {Y i }, {(X i ,Y i )} obeys the AEP, we
conclude that each of the three terms on the right hand side of Eq. (6) converges to a constant as n → ∞. The
left side of Eq. (6) therefore must also converge to a constant as n → ∞.
Example 13. An IID pair process [(X i ,Y i ): I ≥ 1 ] obeys the ISP provided that I(X 1 ; Y 1 ) < ∞. In this case,
the information rate is given by Ĩ = I (X 1 ; Y 1 ). This result is evident from an application of the law of large
numbers to the equation
This result is important because this is the type of channel pair process that results when an IID process
is applied as input to a memoryless channel. (The memoryless channel model is the simplest type of channel
model—it is discussed in Example 21.)
Example 14. Let [(X i ,Y i ): I ≥ 1 ] be a Gaussian process satisfying (10.1) and (10.2). Suppose that the
information rate of this pair process satisfies > 0. It is known that the pair process obeys the ISP (9, Theorem
9.6.1).
Example 15. We assume that [(X i ,Y i ): I ≥ 1 ] is a stationary Gaussian process in which, for each I, the
random variables X i and Y i are real-valued and have expected value equal to zero. For each integer k ≥ 0,
define the trix
Assume that
Following (25, p. 85), we define the spectral densities
where in Eq. (7), for k < 0, we take R(k) = R(−k)T . Suppose that
where the ratio |S1,2 (ω)|2 /S1,1 (ω)S2,2 (ω) is taken to be zero whenever S1,2 (ω) = 0. It is known (9, Theorem 10.2.1)
that the pair process [(X i ,Y i ): I ≥ 1 ] satisfies (10.1) and (10.2), and that the information rate Ĩ is expressible as
Furthermore, we can deduce that [(X i , Y i ): I ≥ 1 ] obeys the ISP. For, if Ĩ > 0, we can appeal to Example
14. On the other hand, if Ĩ = 0, Eq. (8) tells us that the processes {X i } and {Y i } are statistically independent,
upon which we can appeal to Example 11.
Example 16. Let {(X i ,Y i )}: I ≥ 1 ] be a stationary ergodic process such that, for each positive integer n,
holds almost surely for every choice of measurable events A1 , A2 , . . ., An . [The reader not familiar with the
types of conditional probability functions on the two sides of Eq. (9) can consult (26, Chap. 6).] In the context
of communication engineering, the stochastic process [ Y i : i ≥ 1 ] may be interpreted to be the process that is
obtained by passing the process [ X i : i ≥ 1 ] through a memoryless channel (see Example 21). Suppose that
I(X 1 ; Y 1 ) < ∞. Then, properties (10.1) and (10.2) hold and the information stability property holds for the pair
process [(X i ,Y i ): i ≥ 1 ] (14, 27).
Example 17. Let [(X i ,Y i ): i ≥ 1 ] be a stationary ergodic process in which each random variable X i is
real-valued and each random variable Y i is real-valued. We suppose that (10.1) and (10.2) hold and we let Ĩ
denote the information rate of the process [(X i ,Y i ): i ≥ 1 ]. A quantizer is a mapping Q from the real line into a
finite subset of the real line, such that for each value q of Q, the set [ r: Q(r) = q ] is a subinterval of the real
line. Suppose that Q is any quantizer. By Example 12, the pair process [(Q(X i ),Q(Y i )): i ≥ 1 ] obeys the ISP; we
will denote the information rate of this process by I˜Q . It is known that [(X i ,Y i ): I ≥ 1 ] satisfies the information
stability property if
where the supremum is taken over all quantizers Q. This result was first proved in (9, Theorem 8.2.1). Another
proof of the result may be found in (28), where the result is used to prove a source coding theorem. Theorem
7.4.2 of (9) gives numerous conditions under which Eq. (10) will hold.
Example 18. This example points out a way in which the AEP and the ISP are related. Let [ X i : I ≥ 1 ]
be any process satisfying (2.1) and (2.2). Then the pair process {(X i ,X i ): i ≥ 1} satisfies (10.1) and (10.2). The
entropy rate of the process [ X i : i ≥ 1 ] coincides with the information rate of the process (X i ,X i ): i ≥ 1 ]. The
AEP holds for the process [ X i : i ≥ 1 ] if and only if the ISP holds for the pair process [(X i ,X i ): i ≥ 1 ]. To see
that these statements are true, the reader is referred to Example 10.
Further Reading. The exhaustive text by Pinsker (9) contains many more results on information stability
than were discussed in this article. The text by Gray (16) makes the information stability results for stationary
pair processes in (9) more accessible and also extends these results to the bigger class of asymptotically mean
stationary pair processes. The text (9) still remains unparalleled for its coverage of the information stability of
Gaussian pair processes. The paper by Barron (14) contains some interesting results on information stability,
presented in a self-contained manner.
Application to Source Coding Theory
As explained at the start of this article, source coding theory is one of two principal subareas of information
theory (channel coding theory being the other). In this section, explanations are given of the operational
significance of the AEP and the ISP to source coding theory.
Fig. 2. Lossless source coding system.
An information source generates data samples sequentially in time. A fixed abstract information source
is considered, in which the sequence of data samples generated by the source over time is modeled abstractly
as a stochastic process [ U i : i ≥ 1 ]. Two coding problems regarding the given abstract information source shall
be considered. In the problem of lossless source coding, one wishes to assign a binary codeword to each block of
source data, so that the source block can be perfectly reconstructed from its codeword. In the problem of lossy
source coding, one wishes to assign a binary codeword to each block of source data, so that the source block can
be approximately reconstructed from its codeword.
Lossless Source Coding. The problem of lossless source coding for the given abstract information
source is considered first. In lossless source coding, it is assumed that there is a finite set A (called the source
alphabet) such that each random data sample U i generated by the given abstract information source takes its
values in A. The diagram in Fig. 2 depicts a lossless source coding system for the block U n = (U 1 , U 2 , . . ., U n ),
consisting of the first n data samples generated by the given abstract information source.
As depicted in Fig. 2, the lossless source coding system consists of encoder and decoder. The encoder
accepts as input the random source block U n and generates as output a random binary codeword B(U n ). The
decoder perfectly reconstructs the source block U n from the codeword B(U n ). A nonnegative real number R is
called an admissible lossless compression rate for the given information source if, for each δ > 0, a Fig. 2 te
system can be designed for fficiently large n so that
where | B(U n )| denotes the length of the codeword B(U n ).

Let us now refer back to the start of this article, where we talked about the rate R(S) at which the
information source S in a data communication system generates information over time (assuming that the
information must be losslessly transmitted). We were not precise in the beginning concerning how R(S) should
be defined. We now define R(S) to be the minimum of all admissible lossless compression rates for the given
information source S.
As discussed earlier, if the communication engineer must incorporate a given information source S into
the design of a data communication system, it would be advantageous for the engineer to be able to determine
the rate R(S). Let us assume that the process {U i : i ≥ 1} modeling our source S obeys the AEP. In this case, it
can be shown that
where is the entropy rate of the process {U i }. We give here a simple argument that is an admissible lossless
compression rate for the given source, using the AEP. [This will prove that R(S) ≤ . Using the AEP, a proof
can also be given that R(S) ≥ , thereby completing the demonstration of Eq. (12), but we omit this proof.] Let
An be the set of all n-tuples from the source alphabet A. For each n ≥ 1, we may pick a subset En of An so that
properties (2.3) to (2.5) hold. [The ε in (2.4) and (2.5) is a fixed, but arbitrary, positive real number.] Let F n be
the set of all n-tuples in An , which are not contained in En . Because of property (2.5), for sufficiently large n,
we may assign each n-tuple in En a unique binary codeword of length 1 + n( + ε), so that each codeword
begins with 0. Letting |A| denote the number of symbols in A, we may assign each n-tuple in F n a unique binary
codeword of length 1 + CRn log |A|, so that each codeword begins with 1. In this way, we have a lossless
Fig. 3. Lossy source coding system.
codeword assignment for all of An , which gives us an encoder and decoder for a Fig. 2 lossless source coding
system. Because of property (2.3), Eq. (11) holds with R = and δ = 2ε. Since ε (and therefore δ ) is arbitrary,
we can conclude that is an admissible lossless compression rate for our given information source.
In view of Eq. (12), we see that for an abstract information source modeled by a process {U i : i ≥ 1}
satisfying the AEP, the entropy rate has the following operational significance:
• No R < is an admissible lossless compression rate for the given source.

• Every R ≥ is an admissible lossless compression rate for the given source.
If the process {U i : i ≥ 1} does not obey the AEP, then Eq. (12) can fail, even when properties (2.1) and
(2.2) are true and thereby ensure the existence of the entropy rate . Here is an example illustrating this
phenomenon.
Example 19. Let the process {U i : i ≥ 1} modeling the source S have alphabet A = {0, 1} and satisfy, for
each positive integer n, the following properties:
Properties (2.1) and (2.2) are satisfied and the entropy rate is = 12 . Reference 29 shows that R(S) = 1.
Extensions. The determination of the minimum admissible lossless compression rate R(S), when the
AEP does not hold for the process [ U i : I ≥ 1 ] modeling the abstract source S, is a problem that is beyond
the scope of this article. This problem was solved by Parthasarathy (29) for the case in which [ U i : I ≥ 1 ] is a
stationary process. For the case in which [ U i : I ≥ 1 ] is nonstationary, the problem has been solved by Han
and Verdú (30, Theorem 3).
Lossy Source Coding. The problem of lossy coding of a given abstract information source is now
considered. The stochastic process [U i : I ≥ 1] is again used to model the sequence of data samples generated by
the given information source, except that the source alphabet A is now allowed to be infinite. Figure 3 depicts
a lossy source coding system for the source block U n = (U 1 , U 2 , . . ., U n ).
Comparing Fig. 3 to Fig. 2, we see that what distinguishes the lossy system from the lossless system is the
presence of the quantizer in the lossy system. The quantizer in Fig. 3 is a mapping Q from the set of n-tuples
An into a finite subset Q(An ) of An . The quantizer Q assigns to the random source block U n a block
The encoder in Fig. 3 assigns to the quantized source block Û n a binary codeword B from which the
decoder can perfectly reconstruct . Thus the system in Fig. 3 reconstructs not the original source block U n ,
but , a quantized version of U n .
In order to evaluate how well lossy source coding can be done, one must specify for each positive integer n
a nonnegative real-valued function ρn on the product space An × An (called a distortion measure). The quantity
ρn (U n , ) measures how closely the reconstructed block in Fig. 3 resembles the source block U n . Assuming
that ρn is a jointly continuous function of its two arguments, which vanishes whenever the arguments are
equal, one goal in the design of the lossy source coding system in Fig. 3 would be:
• Goal 1. Ensure that ρn (U n , ) is sufficiently close to zero.
However, another goal would be:
• Goal 2. Ensure that the length | B | of the codeword B is sufficiently small.
These are conflicting goals. The more closely one wishes to resemble U n [corresponding to a sufficiently
small value of ρn (U n , ) ], the more finely one must quantize U n , meaning an increase in the size of the set
Q(An ), and therefore an increase in the length of the codewords used to encode the blocks in Q(An ). There must
be a trade-off in the accomplishment of Goals 1 and 2. To reflect this trade-off, two figures of merit are used
in lossy source coding. Accordingly, we define a pair (R, D) of nonnegative real numbers to be an admissible
rate-distortion pair for lossy coding of the given abstract information source, if, for any ε > 0, the Fig. 3 system
can be designed for sufficiently large n so that
We now describe how the information stability property can allow one to determine admissible rate-
distortion pairs for lossy coding of the given source. For simplicity, we assume that the process [ U i : I ≥ 1 ]
modeling the source outputs is stationary and ergodic. Suppose we can find another process {V i : I ≥ 1 ] such
that
• The pair process [(U i ,V i ) : I ≥ 1 ] is stationary and ergodic.

• There is a finite set Â ⊂ A such that each V i takes its values in Â.
Appealing to Example 12, the pair process [(U i ,V i ): I ≥ 1 ] satisfies the information stability property. Let
Ĩ be the information rate of this process. Assume that the distortion measures [ ρn ] satisfy
for any pair of n-tuples (u1 , . . ., U n ), (û1 , . . ., ûn ) from An . (In this case, the sequence of distortion measures [
ρn ] is called a single letter fidelity criterion.) Let D = E[ρ1 (U 1 , V 1 ) ]. Via a standard argument (omitted here)
called a random coding argument [see proof of Theorem 7.2.2 of (31)], information stability can be exploited
to show that the pair (Ĩ, D) is an admissible rate-distortion pair for our given abstract information source. [It
should be pointed out that the random coding argument not only exploits the information stability property
but also exploits the property that
which is a consequence of the ergodic theorem [(32), Chap. 3]].

Example 20. Consider an abstract information source whose outputs are modeled as an IID sequence
of real-valued random variables [ U i : I ≥ 1 ]. This is called the memoryless source model. The squared-error
single letter fidelity criterion [ ρn ] is employed, in which
It is assumed that E[U 2 1 ] < ∞. For each D > 0, let R(D) be the class of all pairs of random variables (U,
V) in which
• U has the same distribution as U 1 .

• V is real-valued.
• E[(U − V)2 ] ≤ D.
The rate distortion function of the given memoryless source is defined by
Shannon (33) showed that any (R, D) satisfying R ≥ r(D) is an admissible rate-distortion pair for lossy
coding of our memoryless source model. A proof of this can go in the following way. Given the pair (R, D)
satisfying R ≥ r(D), one argues that there is a process [ V i : I ≥ 1 ] for which the pair process [ U i ,V i ): I ≥ 1
] is independent and identically distributed, with information rate no bigger than R and with E[(U 1 − V 1 )2 ]
≤ D. A random coding argument exploiting the fact that [(U i , V i ): I ≥ 1 ] obeys the ISP (see Example 13) can
then be given to conclude that (R, D) is indeed an admissible rate-distortion pair. Shannon (33) also proved the
converse statement, namely, that any admissible rate-distortion pair (R, D) for the given memoryless source
model must satisfy R ≥ r(D). Therefore the set of admissible rate-distortion pairs for the memoryless source
model is the set
Extensions. The argument in Example 20 exploiting the ISP can be extended [(31), Theorem 7.2.2] to
show that for any abstract source whose outputs are modeled by a stationary ergodic process, the set in Eq. (15)
coincides with the set of all admissible rate-distortion pairs, provided that a single letter fidelity criterion is
used, and provided that the rate-distortion function r(D) satisfies r(D) < ∞ for each D > 0. [The rate-distortion
function for this type of source must be defined a little differently than for the memoryless source in Example
20; see (31) for the details.] Source coding theory for an abstract source whose outputs are modeled by a
stationary nonergodic process has also been developed. For this type of source model, it is customary to replace
the condition in Eq. (13) in the definition of an admissible rate-distortion pair with the condition
A source coding theorem for the stationary nonergodic source model can be proved by exploiting the
information stability property, provided that the definition of the ISP is weakened to include pair processes
[(U i , V i ): I ≥ 1 ] for which the sequence [ n − 1 I(U n ; V n ): n ≥ 1 ] converges to a nonconstant random variable.
However, for this source model, it is difficult to characterize the set of admissible rate-distortion pairs by use
of the ISP. Instead, Gray and Davisson (34) used the ergodic decomposition theorem (35) to characterize this
set. Subsequently, source coding theorems were obtained for abstract sources whose outputs are modeled by
asymptotically mean stationary processes; an account of this work can be found in Gray (16).
Further Reading. The theory of lossy source coding is called rate-distortion theory. Reference (31) pro-
vides excellent coverage of rate-distortion theory up to 1970. For an account of developments in rate-distortion
theory since 1970, the reader can consult (36,37).
Application to Channel Coding Theory
In this section, explanations are given of the operational significance of the ISP to channel coding theory. To
accomplish this goal, the notion of an abstract channel needs to be defined. The description of a completely
general abstract channel model would be unnecessarily complicated for the purposes of this article. Instead,
an abstract channel model is chosen that will be simple to understand, while of sufficient generality to give the
reader an appreciation for the concepts that shall be discussed.
We shall deal with a semicontinuous channel model (see Example 12) in which the channel input phabet is
finite and the channel output alphabet is the real line. We proceed to give a precise formulation of this channel
model. We fix a finite set A, from which inputs to our abstract channel are to be drawn. For each positive integer
n, let An denote the set of all n-tuples X n = (X 1 , X 2 , . . ., X n ) in which each X i ∈ A, and let Rn denote the set
of all n-tuples Y n = (Y 1 , Y 2 , . . ., Y n ) in which each Y i ∈ R, the set of real numbers. For each n ≥ 1, a function
F n is given that maps each 2 n-tuple (X n , Y n ) ∈ An × Rn into a nonnegative real number F n (Y n |X n ) so that the
following rules are satisfied:
• For each X n ∈ An , the mapping Y n → F n (Y n |X n ) is a jointly measurable function of n variables.

• For each X n ∈ An ,
For each n ≥ 2, each (x1 , x2 , . . ., xn ) ∈ An , and each (y1 , . . ., yn − 1) ∈ Rn − 1,
We are now able to describe how our abstract channel operates. Fix a positive integer n. Let X n ∈ An
be any n-tuple of channel inputs. In response to X n , our abstract channel will generate a random n-tuple of
outputs from Rn . For each measurable subset En of Rn , let Pr[ En|xn ] denote the conditional probability that
the channel output n-tuple will lie in En, given that the channel input is X n . This conditional probability is
computable via the formula
We now need to define the notion of a channel code for our abstract channel model. A channel code for our
given channel is a collection of pairs [(x(i), E(i)): i = 1, 2, . . ., 2k ] in which
Fig. 4. Implementation of a (k, n) channel code.
(1) k is a positive integer.

(2) For some positive integer n,
• x (1),x(2), . . ., X(2k ) are n-tuples from An .

• E(1), E(2), . . ., E(2k ) are subsets of Rn , which form a partition of Rn .
The positive integer n given by (ii) is called the number of channel uses of the channel code, and the
positive integer k given by (i) is called the number of information bits of the channel code. We shall use the
notation cn as a generic notation to denote a channel code with n channel uses. Also, a channel code shall be
referred to as a (k,n) channel code if the number of channel uses is n and the number of information bits is k.
In a channel code {(x(i), E(i))}, the sequences {x(i)} are called the channel codewords, and the sets {E(i)} are
called the decoding sets.
A (k,n) channel code {(x(i), E(i)): i = 1, 2, . . ., 2k } is used in the following way to transmit data over
our given channel. Let {0, 1}k denote the set of all binary k-tuples. Suppose that the data that one wants to
transmit over the channel consists of the k-tuples in {0, 1}k . One can assign each k-tuple B ∈ [0, 1] k an integer
index I = I(B) satisfying 1 ≤ I ≤ 2k , which uniquely identifies that k-tuple. If the k-tuple B is to be transmitted
over the channel, then the channel encoder encodes B into the channel codeword X(I) in which I = I(B), and
then x(i) is applied as input to the channel. At the receiving end of the channel, the channel decoder examines
the resulting random channel output n-tuple Y n that was received in response to the channel codeword x(i).
The decoder determines the unique random integer J such that Y n ∈ E(J) and decodes Y n into the random
k-tuple B̂∈{0,1}k whose index is J. The transmission process is depicted in Fig. 4.
There are two figures of merit that tell us the performance of the (k, n) channel code cn depicted in Fig. 4,
namely, the transmission rateR(cn ) and the error probabilitye(cn ). The transmission rate measures how many
information bits are transmitted per channel use and is defined by
The error probability gives the worst case probability that B̂ in Fig. 4 will not be equal to B, over all
possible B ∈ {0, 1}k . It is defined by
It is desirable to find channel codes that simultaneously achieve a large transmission rate and a small
error probability. Unfortunately, these are conflicting goals. It is customary to see how large a transmission
rate can be achieved for sequences of channel codes whose error probabilities → 0. Accordingly, an admissible
transmission rate for the given channel model is defined to be a nonnegative number R for which there exists
a sequence of channel codes [ cn : n = 1, 2, . . . ] satisfying both of the following:
We now describe how the notion of information stability can tell us about admissible transmission rates
for our channel model. Let [ X i : i ≥ 1 ] be a sequence of random variables taking their values in the set A,
which we apply as inputs to our abstract channel. Because of the consistency criterion, Eq. (16), the abstract
channel generates, in response to [ X i : i ≥ 1 ], a sequence of real-valued random outputs [ Y i : i ≥ 1 ] for which
the distribution of the pair process [(X i ,Y i ): i ≥ 1 ] is uniquely specified by
for every positive integer n, every n-tuple X n ∈ An , and every measurable set En ⊂ Rn . Suppose the pair process
[(X i ,Y i ): i ≥ 1 ] obeys the ISP with information rate Ĩ. Then a standard argument [see (38), proof of Lemma
3.5.2] can be given to show that Ĩ is an admissible transmission rate for the given channel model.
Using the notation introduced earlier, the capacityR(C) of an abstract channel C is defined to be the
maximum of all admissible transmission rates. For a given channel C, it is useful to determine the capacity
R(C). (For example, as discussed at the start of this article, if a data communication system is to be designed
using a given channel, then the channel capacity must be at least as large as the rate at which the information
source in the system generates information.) Suppose that an abstract channel C possesses at least one input
process [ X i : i ≥ 1 ] for which the corresponding channel pair process [(X i ,Y i ): i ≥ 1 ] obeys the ISP. Define RISP
(C) to be the supremum of all information rates of such processes [(X i ,Y i ): I ≥ 1 ]. By our discussion in the
preceding paragraph, we have
For some channels C, one has R C (R ) = R ISP (C). For such a channel, an examination of channel pair
processes satisfying the ISP will allow one to determine the capacity.
Examples of channels for which this is true are the memoryless channel (see Example 21 below), the finite-
memory channel (39), and the finite-state indecomposable channel (40). On the other hand, if R (C) > R ISP (C)
for a channel C, the concept of information stability cannot be helpful in determining the channel capacity—
some other concept must be used. Examples of channels for which R (C) > R ISP (C) holds, and for which the
capacity R (C) has been determined, are the Ĩ continuous channels (41), the weakly continuous channels (42),
and the historyless channels (43). The authors of these papers could not use information stability to determine
capacity. They used instead the concept of “information quantiles,” a concept beyond the scope of this article.
The reader is referred to Refs. 41–43 to see what the information quantile concept is and how it is used.
Example 21. Suppose that the conditional density functions [ Fn : n = 1, 2, . . . ] describing our channel
satisfy
for every positive integer n, every n-tuple xn = (x1 , . . ., xn ) from An , and every n-tuple Y n = (Y 1 , . . ., Y n ) from
Rn . The channel is then said to be memoryless. Let R∗ be the nonnegative real number defined by
where the supremum is over all pairs (X, Y) in which X is a random variable taking values in A, and Y is a
real-valued random variable whose conditional distribution given X is governed by the function f 1 . (In other
words, we may think of Y as the channel output in response to the single channel input X.) We can argue that
R∗ is an admissible transmission rate for the memoryless channel as follows. Pick a sequence of IID channel
inputs [ X i : I ≥ 1 ] such that if [ Y i : I ≥ 1 ] is the corresponding sequence of random channel outputs, then
I(X 1 ; Y 1 ) = R∗ . The pairs [(X i ,Y i ): i ≥ 1 ] are IID, and the process [(X i ,Y i ): i ≥ 1 ] obeys the ISP with information
rate Ĩ = R∗ (see Example 13). Therefore R∗ is an admissible transmission rate. By a separate argument, it
is well known that the converse is also true; namely, every admissible transmission rate for the memoryless
channel is less than or equal to R∗ (1). Thus the number R∗ given by Eq. (17) is the capacity of the memoryless
channel.
Final Remarks
It is appropriate to conclude this article with some remarks concerning the manner in which the separate
theories of source coding and channel coding tie together in the design of data communication systems. In
the section entitled “Lossless Source Coding,” it was explained how the AEP can sometimes be helpful in
determining the minimum rate R(S) at which an information source S can be losslessly compressed. In the
section entitled “Application to Channel Coding Theory,” it was indicated how the ISP can sometimes be used
in determining the capacity R(C) of a channel C, with the capacity giving the maximum rate at which data
can reliably be transmitted over the channel. If the inequality R(S) ≤ R(C) holds, it is clear from this article
that reliable transmission of data generated by the given source S is possible over the given channel C. Indeed,
the reader can see that reliable txsransmission will take place for the data communication system in Fig. 1 by
taking the encoder to be a two-stage encoder, in which a good source encoder achieving a compression rate close
to R(S) is followed by a good channel encoder achieving a transmission rate close to R(C). On the other hand,
if R(S) > R(C), there is no encoder that can be found in Fig. 1 via which data from the source S can reliably
be transmitted over the channel C [see any basic text on information theory, such as (44), for a proof of this
result]. One concludes from these statements that in designing a reliable encoder for the data communication
system in Fig. 1, one need onlyonsider the two-stage encoders consisting of a good source encoder followed
by a good channel encoder. This principle, which allows one to break down the problem of encoder design in
communication systems into the two separate simpler problems of source encoder design and channel encoder
design, has come to be called “Shannon’s separation principle,” after its originator, Claude Shannon.
Shannon’s separation principle also extends to lossy transmission of source data over a channel in a data
communication system. In Fig. 1, suppose that the data communication system is to be designed so that the
data delivered to the user through the channel C must be within a certain distance D of the original data
generated by the source S. The system can be designed if and only if there is a positive real number R such that
(1) (R, D) is an admissible rate-distortion pair for lossy coding of the source S in the sense of the “Lossy Source
Coding” section, and (2) R ≤ R(C). If R is a positive real number satisfying (1) and (2), Shannon’s separation
principle tells us that the encoder in Fig. 1 can be designed as a two-stage encoder consisting of source encoder
followed by channel encoder in which:
• The source encoder is designed to achieve the compression rate R and to generate blocks of encoded data
that are within distance D of the original source blocks.
• The channel encoder is designed to achieve a transmission rate close to R(C).
It should be pointed out that Shannon’s separation principle holds only if one is willing to consider
arbitrarily complex encoders in communication systems. [In defining the quantities R(S) and R(C) in this
article, recall that no constraints were placed on how complex the source encoder and channel encoder could
be.] It would be more realistic to impose a complexity constraint specifying how complex an encoder one is
willing to use in the design of a communication system. With a complexity constraint, there could be an
advantage in designing a “combined source–channel encoder” which combines data compression and channel
error correction capability in its operation. Such an encoder for the communication system could have the
same complexity as two-stage encoders designed according to the separation principle but could afford one a
better data transmission capability than the two-stage encoders. There has been much work in recent years on
“combined source–channel coding,” but a general theory of combined source–channel coding has not yet been
put forth.
BIBLIOGRAPHY
1. C. Shannon, A mathematical theory of communication, Bell Syst. Tech. J., 27: 379–423, 623–656, 1948.
2. B. McMillan, The basic theorems of information theory, Ann. Math. Stat., 24: 196–219, 1953.
3. L. Breiman, The individual ergodic theorem of information theory, Ann. Math. Stat., 28: 809–811, 1957.
4. R. Gray J. Kieffer, Asymptotically mean stationary measures, Ann. Probability, 8: 962–973, 1980.
5. S. Verdú T. Han, The role of the asymptotic equipartition property in noiseless source coding, IEEE Trans. Inf. Theory, 43:
847–857, 1997.
6. J. Kieffer, A generalized Shannon-McMillan theorem for the action of an amenable group on a probability space, Ann.
Probability, 3: 1031–1037, 1975.
7. D. Ornstein B. Weiss, The Shannon-McMillan-Breiman theorem for a class of amenable groups, Isr. J., Math., 44:
53–60, 1983.
8. A. Perez, Notions généralisées d incertitude, d entropie et d information du point de vue de la théorie de martin-
gales, Trans. 1st Prague Conf. Inf. Theory, Stat. Decision Funct., Random Process., pp. 183–208, 1957.
9. M. Pinsker, Information and Information Stability of Random Variables and Processes, San Francisco: Holden-Day,
1964.
10. A. Ionescu Tulcea Contributions to information theory for abstract alphabets, Ark. Math., 4: 235–247, 1960.
11. A. Perez, Extensions of Shannon-McMillan’s limit theorem to more general stochastic processes, Trans. 3rd Prague
Conf. Inf. Theory, pp. 545–574, 1964.
12. S. Moy, Generalizations of Shannon-McMillan theorem, Pac. J. Math., 11: 705–714, 1961.
13. S. Orey, On the Shannon-Perez-Moy theorem, Contemp. Math., 41: 319–327, 1985.
14. A. Barron, The strong ergodic theorem for densities: Generalized Shannon-McMillan-Breiman theorem, Ann. Proba-
bility, 13: 1292–1303, 1985.
15. P. Algoet T. Cover, A sandwich proof of the Shannon- McMillan-Breiman theorem, Ann. Probability, 16: 899–909, 1988.
16. R. Gray, Entropy and Information Theory, New York: Springer-Verlag, 1990.
17. A. Tempelman, Specific characteristics and variational principle for homogeneous random fields, Z. Wahrschein. Verw.
Geb., 65: 341–365, 1984.
18. D. Ornstein, Ergodic Theory, Randomness, and Dynamical Systems, Yale Math. Monogr. 5, New Haven, CT: Yale
University Press, 1974.
19. D. Ornstein B. Weiss, Entropy and isomorphism theorems for actions of amenable groups, J. Anal. Math., 48: 1–
141, 1987.
20. R. Mañé Ergodic Theory and Differentiable Dynamics, Berlin and New York: Springer-Verlag, 1987.
21. M. Ohya, Entropy operators and McMillan type convergence theorems in a noncommutative dynamical system, Lect.
Notes Math., 1299, 384–390, 1988.
22. J. Fritz, Generalization of McMillan’s theorem to random set functions, Stud. Sci. Math. Hung., 5: 369–394, 1970.
23. A. Perez, Generalization of Chernoff’s result on the asymptotic discernability of two random processes, Colloq. Math.
Soc. J. Bolyai, No. 9, pp. 619–632, 1974.
24. P. Algoet T. Cover, Asymptotic optimality and asymptotic equipartition properties of log-optimum investme, Ann.
Probability, 16: 876–898, 1988.
25. A. Balakrishnan, Introduction to Random Processes in Engineering, New York: Wiley, 1995.
26. R. Ash, Real Analysis and Probability, New York: Academic Press, 1972.
27. M. Pinsker, Sources of messages, Probl. Peredachi Inf., 14, 5–20, 1963.
28. R. Gray J. Kieffer, Mutual information rate, distortion, and quantization in metric spaces, IEEE Trans. Inf. Theory, 26:
412–422, 1980.
29. K. Parthasarathy, Effective entropy rate and transmission of information through channels with additive random
noise, Sankhyā, Ser. A, 25: 75–84, 1963.
30. T. Han S. Verdú, Approximation theory of output statistics, IEEE Trans. Inf. Theory, 39: 752–772, 1993.
31. T. Berger, Rate Distortion Theory: A Mathematical Basis for Data Compression, Englewood Cliffs, NJ: Prentice–Hall,
1971.
32. W. Stout, Almost Sure Convergence, New York: Academic Press, 1974.
33. C. Shannon, Coding theorems for a discrete source with a fidelity criterion, IRE Natl. Conv. Rec., Part 4, pp. 142–163,
1959.
34. R. Gray L. Davisson, Source coding theorems without the ergodic assumption, IEEE Trans. Inf. Theory, 20: 502–
516, 1974.
35. R. Gray L. Davisson, The ergodic decomposition of stationary discrete random processes, IEEE Trans. Inf. Theory, 20:
625–636, 1974.
36. J. Kieffer, A survey of the theory of source coding with a fidelity criterion, IEEE Trans. Inf. Theory, 39: 1473–1490, 1993.
37. T. Berger J. Gibson, Lossy source coding, IEEE Trans. Inf. Theory, 44: 2693–2723, 1998.
38. R. Ash, Information Theory, New York: Interscience, 1965.
39. A. Feinstein, On the coding theorem and its converse for finite-memory channels, Inf. Control, 2: 25–44, 1959.
40. D. Blackwell, L. Breiman, A. Thomasian, Proof of Shannon’s transmission theorem for finite-state indecomposable
channels, Ann. Math. Stat., 29: 1209–1220, 1958.
41. R. Gray D. Ornstein, Block coding for discrete stationary d? continuous noisy channels, IEEE Trans. Inf. Theory, 25:
292-306, 1979.
42. J. Kieffer, Block coding for weakly continuous channels, IEEE Trans. Inf. Theory, 27, 721–727, 1981.
43. S. Verdú T. Han, A general formula for channel capacity, IEEE Trans. Inf. Theory, 40: 1147–1157, 1994.
44. T. Cover J. Thomas, Elements of Information Theory, New York: Wiley, 1991.
READING LIST
R. Gray L. Davisson, Ergodic and Information Theory, Benchmark Pap. Elect. Eng. Comput. Sci. Vol. 19, Stroudsburg, PA:
Dowden, Hutchinson, & Ross, 1977.
IEEE Transactions of Information Theory, Vol. 44, No. 6, October, 1998. (Special issue commemorating fifty years of
information theory.)
JOHN C. KIEFFER
University of Minnesota
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Maximum Likelihood Imaging ● Email this page Enter words or
● Print this page
Timothy J. Schulz1
1Michigan Technological University, Houghton, MI
reserved. Search
Abstract
Scalar Fields and Coherence
Incoherent Imaging
Coherent Imaging
Summary

438 MAXIMUM LIKELIHOOD IMAGING
MAXIMUM LIKELIHOOD IMAGING
Imaging science is a rich and vital branch of engineering in

which electromagnetic or acoustic signals are measured, pro-
cessed, analyzed, and interpreted in the form of multidimen-
sional images. Because these images often contain informa-
tion about the physical, biological, or operational properties of
remote objects, scenes, or materials, imaging science is justly
considered to be a fundamental component of that branch of
engineering and science known as remote sensing. Many sub-
jects benefit directly from advances in imaging science—these
range from astronomy and the study of very large and dis-
tance objects to microscopy and the study of very small and
nearby objects.
The photographic camera is probably the most widely
known imaging system in use today. The familiar imagery
recorded by this device usually encodes the spectral re-
flectance properties of an object or scene onto a two-dimen-
sional plane. The familiarity of this form of imagery has led
to a common definition of an image as ‘‘an optical reproduc-
tion of an object by a mirror or lens.’’ There are, however,
many other imaging systems in use and the object or scene
properties encoded in their imagery can be very different from
MAXIMUM LIKELIHOOD IMAGING 439
those recorded by a photographic camera. Temperature varia- sured data. In some situations structural restrictions such as
tions can, for instance, be ‘‘imaged’’ with infrared sensing, ve- these are acceptable, but in many others they are not and the
locity variations can be ‘‘imaged’’ with radar, geological for- advent of faster and more sophisticated computing resources
mations can be ‘‘imaged’’ with sonar, and the physiological has served to greatly lessen the need for and use of structural
function of the human brain can be ‘‘imaged’’ with positron constraints in imaging problems.
emission tomography (PET). Many criteria can be used to quantify image quality and
A photographic camera forms images in a manner very induce optimal signal-processing algorithms. One might ask,
similar to the human eye, and, because of this, photographic for example, that the processed imagery produce the ‘‘correct’’
images are easily interpreted by humans. The imagery re- image on average. This leads to an unbiased estimator, but
corded by an infrared camera might contain many of the fea- such an estimator may not exist, may not be unique, or may
tures common to visible imagery; however, the phenomena result in imagery whose quality is far from adequate. By re-
being sensed are different and some practice is required be- quiring that the estimated image also have, in some sense,
fore most people can faithfully interpret raw infrared imag- the smallest deviations from the correct image this criterion
ery. For both of these modalities, though, the sensor data is could be modified to induce the minimum variance, unbiased
often displayed as an image without the need for significant estimator (MVUE), whose imagery may have desirable quali-
signal processing. The data acquired by an X-ray tomograph ties, but whose processing structure can be difficult or impos-
or synthetic aperture radio telescope, however, are not easily sible to derive and implement. The maximum-likelihood
interpreted, and substantial signal processing is required to method for estimation leads to an alternative criterion
form an ‘‘image.’’ In these situations, the processing of raw whereby an image is selected to optimize a mathematical cost
sensor data to form imagery is often referred to as image re- function that is induced by the physical and statistical model
construction or image synthesis (1), and the importance of sig- for the acquired data. The relative simplicity of the maxi-
nal processing in these applications is great. To confirm this mum-likelihood estimation method, along with the fact that
importance, the 1979 Nobel prize in physiology and medicine maximum-likelihood estimates are often asymptotically unbi-
was awarded to Alan M. Cormack and Sir Godfrey N. Houns- ased with minimum variance, makes this a popular and
field for the development and application of the signal pro- widely studied method for statistical inference. It is largely
cessing methods used for X-ray computed tomography, and for this reason that the development and utilization of maxi-
the 1974 Nobel prize in physics was awarded to Sir Martin mum-likelihood estimation methods for imaging are the focus
Ryle for the development of aperture synthesis techniques of this article.
used to form imagery with radio telescope arrays. For both of One of the most important steps in the utilization of the
these modalities the resulting images are usually very differ- maximum-likelihood method for imaging is the development
ent from the visible images formed by photographic cameras, of a practical and faithful model that represents the relation-
and significant training is required for their interpretation. ship between the object or scene being sensed and the data
Imagery formed by photographic cameras, and similar in- recorded by the sensor. This modeling step usually requires a
struments such as telescopes and microscopes, can also be dif- solid understanding of the physical and statistical character-
ficult to interpret in their raw form. Focusing errors, for ex- istics of electromagnetic- or acoustic-wave propagation, along
ample, can make imagery appear blurred and distorted, as with an appreciation for the statistical characteristics of the
can significant flaws in the optical instrumentation. In these data acquired by real-world sensors. For these reasons, a
situations, a type of signal processing known as image resto- strong background in the fields of Fourier optics (10,11), sta-
ration (2,3) can be used to remove the distortions and restore tistical optics (12–14), basic probability and random-process
fidelity to the imagery. Processing such as this received na- theory (15,16), and estimation theory (5–9) is essential for
tional attention after the discovery of the Hubble Space Tele- one wishing to apply maximum-likelihood methods to the
scope aberrated primary mirror in 1990, and one of the most field of imaging science.
successful and widely used algorithms for restoring resolution Statistical inference problems such as those encountered
to Hubble imagery was based on the maximum-likelihood es- in imaging applications are frequently classified as ill-posed
timation method (4). The motivation for and derivation of this problems (17). An image-recovery or -restoration problem is
image-restoration algorithm will be discussed in great detail ill posed if it is not well posed, and a problem is well posed in
later in this article. the classical sense of Hadamard if the problem has a unique
When signal processing is required for the formation or solution and the solution varies continuously with the data.
improvement of imagery, the imaging problem can usually be Abstract formulations of image recovery and restoration prob-
posed as one of statistical inference. A large number of esti- lems on infinite-dimensional measurement and parameter
mation-theoretic methods are available for solving statistical- spaces are almost always ill posed, and their ill-posed nature
inference problems (5–9), and the method to be used for a is usually due to the discontinuity of the solution. Problems
particular application depends largely on three factors: (1) the that are formulated on finite-dimensional spaces are fre-
structure imposed on the processing; (2) the quantitative cri- quently well-posed in the classical sense—they have a unique
teria used to define image quality; and (3) the physical and solution and the solution is continuous in the data. These
statistical information available about the data collection problems, however, are often ill conditioned or badly behaved
process. and are frequently classified as ill posed even though they are
Structure can be imposed on processing schemes for a vari- technically well posed.
ety of reasons, but the most common is the need for fast and For problems that are ill posed or practically ill posed, the
inexpensive processing. The most common structure imposed original problem’s solution is often replaced by the solution to
for this reason is linear processing, whereby imagery is a well-posed (or well-behaved) problem. This process is re-
formed or improved through linear combinations of the mea- ferred to as regularization and the basic idea is to change the
problem in a manner such that the solution is still meaning- Coherence is an important concept in imaging that is used
ful but no longer badly behaved (18). The consequence for im- to describe properties of waveforms, sensors, and processing
aging problems is that we do not seek to form a ‘‘perfect’’ im- algorithms. Roughly speaking, coherence of a waveform refers
age, but instead settle for a more stable—but inherently to the degree to which a deterministic relationship exists be-
biased—image. Many methods are available for regularizing tween the complex envelope phase ␪ (x, y, z; t) at different time
maximum-likelihood estimation problems, and these include: instances or spatial locations. Temporal coherence at time
penalty methods, whereby the mathematical optimization delay ␶ quantifies the relationship between ␪ (x, y, z; t) and
problem is modified to include a term that penalizes un- ␪ (x, y, z; t ⫹ ␶), whereas the spatial coherence at spatial shift
wanted behavior in the object parameters (19); sieve methods, (⌬x, ⌬y, ⌬z) quantifies the relationship between ␪(x, y, z; t) and
whereby the allowable class of object parameters is reduced ␪ (x ⫹ ⌬x, y ⫹ ⌬y, z ⫹ ⌬z; t). A coherent sensor is one that
in some manner to exclude those with unwanted characteris- records information about the complex-envelope phase of a
tics (20); and stopping methods, whereby the numerical algo- waveform, and a coherent signal-processing algorithm is one
rithms used to solve a particular optimization problem are that processes this information. Waveforms that are coherent
prematurely terminated before convergence and before the only over vanishingly small time delays are called temporally
object estimate has obtained the unwanted features that are incoherent; waveforms that are coherent only over vanish-
characteristic of the unconstrained solution obtained at con- ingly small spatial shifts are called spatially incoherent. Sen-
vergence (21). Penalty methods can be mathematically, but sors and algorithms that neither record nor process phase in-
not always philosophically, equivalent to the maximum a pos- formation are called incoherent.
teriori (MAP) method, whereby an a priori statistical model Many phenomena in nature are difficult, if not impossible
for the object is incorporated into the estimation procedure. within our current understanding, to model in a deterministic
The MAP method is appealing and sound provided that a manner, and the statistical properties of acoustic and electro-
physically justified model is available for the object parame- magnetic fields play a fundamental role in modeling the out-
ters. Each of these regularization methods is effective at come of most remote sensing and imaging experiments. For
times, and the method used for a particular problem is often most applications an adequate description of the fields in-
a matter of personal taste. volved is captured through second-order averages known as
coherence functions. The most general of these is the mutual
coherence function, which is defined mathematically in terms
SCALAR FIELDS AND COHERENCE of the complex envelope for a field as
Because most imaging problems involve the processing of 12 (τ ) = E[u(x1 , y1 , z1 , t + τ )u∗ (x2 , y2 , z2 , t)] (4)
electromagnetic or acoustic fields that have been measured
after propagation from a remote object or scene, a good place The proper interpretation for the expectation in this defini-
to begin our technical discussion is with a review of scalar tion depends largely on the application, and much care must
waves and the concept of coherence. The scalar-wave theory be taken in forming this interpretation. For some applications
is widely used for two reasons: (1) acoustic wave propagation a definition involving time averages will be adequate,
is well-modeled as a scalar phenomenon; and (2) although whereas other applications will call for a definition involving
electromagnetic wave propagation is a vector phenomenon, ensemble averages.
the scalar theory is often appropriate, particularly when the The mutual coherence function is often normalized to form
dimensions of interest in a particular problem are large in the complex degree of coherence as
comparison to the electromagnetic field wavelength.
A scalar field is in general described by a function in four 12 (τ )
γ12 (τ ) = (5)
dimensions s(x, y, z; t), where x, y, and z are coordinates in [11 (0)22 (0)]1/2
three-dimensional space, and t is a coordinate in time. In
many situations, the field fluctuations in time are concen- and it is tempting to define a coherent field as one for which
trated about some center frequency f 0, so that the field can be 兩웂12(␶)兩 ⫽ 1 for all pairs of spatial locations, (x1, y1, z1) and (x2,
conveniently expressed as y2, z2), and for all time delays, ␶. Such a definition is overly
restrictive and a less restrictive condition, as discussed by
s(x, y, z; t) = a(x, y, z; t) cos [2π f 0 t + θ (x, y, z; t)] (1) Mandel and Wolf (22), is that
or, in complex notation, as max |γ12 (τ )| = 1 (6)

τ
s(x, y, z; t) = Re{u(x, y, z; t)e j2π f 0 t } (2) for all pairs of spatial locations, (x1, y1, z1) and (x2, y2, z2). Al-
though partial degrees of coherence are possible, fields that
where are not coherent are usually called incoherent. In some situa-
tions a field is referred to as being fully incoherent over a
u(x, y, z; t) = a(x, y, z; t)e jθ (x,y,z;t ) (3) particular region and its mutual coherence function is mod-
eled over this region as
is the complex envelope for the field. Properties of the field
amplitude a, phase ␪, or both are often linked to physical or 12 (τ ) κI(x1 , y1 , z1 )δ3 (x1 − x2 , y1 − y2 , z1 − z2 )δ1 (t − τ ) (7)
operational characteristics of a remote object or scene, and
the processing of remotely sensed data to determine these where I( ⭈ ) is the incoherent intensity for the field, 웃3( ⭈ , ⭈ , ⭈ )
properties is the main goal in most imaging applications. is the three-dimensional Dirac impulse, 웃1( ⭈ ) is the one-di-
mensional Dirac impulse, and ␬ is a constant with appro- When optical fields interact with a photodetector, the ab-
priate units. Most visible light used by the human eye to form sorption of a quantum of energy—a photon—results in the
images is fully incoherent and fits this model. Goodman (13) release of an excited electron. This interaction is referred to
and Mandel and Wolf (22) provide detailed discussions of the as a photoevent, and the number of photoevents occurring
coherence properties of electromagnetic fields. over a particular spatial region and time interval are referred
to as photocounts. Most detectors of light record photocounts,
and although the recorded data depend directly on the image
INCOHERENT IMAGING
intensity, the actual number of photocounts recorded is a fun-
damentally random quantity. The images shown in Fig. 1
Astronomical telescopes, computer assisted tomography
help to illustrate this effect. Here, an image of Simeon Pois-
(CAT) scanners, PET scanners, and many forms of light mi-
son (for whom the Poisson random variable is named) is
croscopes are all examples of incoherent imaging systems; the
shown as it might be acquired by a detector when 1 million,
waveforms, sensors, and algorithms used in these situations
10 million, and 100 million total photocounts are recorded.
are all incoherent. The desired image for these systems is typ-
ically related to the intensity distribution of a field that is
Statistical Model
transmitted through, reflected by, or emitted from an object
or scene of interest. For many of these modalities it is com- For many applications involving charge coupled devices
mon to acquire data over a variety of observing scenarios, and (CCD) and other detectors of optical radiation, the semiclassi-
the mathematical model for the signal acquired by these sys- cal theory leads to models for which the photocounts recorded
tems is of the form by each detector element are modeled as Poisson random
variables whose means are determined by the measurement
Ik (y) = hk ( y, x)I(x) dx, k = 1, 2, . . ., K (8) intensity Ik( ⭈ ). That is, the expected number of photocounts
acquired by the nth photodetector during the kth observation
where I( ⭈ ) is the object incoherent intensity function— interval is
usually related directly to the emissive, reflective, or trans-
missive properties of the object, hk( ⭈ , ⭈ ) is the measurement Ik [n] = γ Ik ( y) dy (9)
kernel or system point-spread function for the kth observa- Yn
tion, Ik( ⭈ ) is the incoherent measurement signal for the kth
observation, x is a spatial variable in two- or three-dimen- where n is a two-dimensional discrete index to the elements
sions, and y is usually a spatial variable in one-, two-, or of the detector array, Y n is the spatial region over which the
three-dimensions. The mathematical forms for the system nth detector element integrates the image intensity, and 웂 is
point-spread functions 兵hk( ⭈ , ⭈ )其 are induced by the physical a nonnegative scale factor that accounts for overall detector
properties of the measurement system, and much care should efficiency and integration time. Furthermore, the number of
be taken in their determination. In telescope and microscope photocounts acquired by different detector elements are usu-
imaging, for example, the instrument point-spread functions ally statistically independent, and the detector regions are of-
model the effects of diffraction, optical aberrations, and inho- ten small in size relative to the fluctuations in the image in-
mogeneities in the propagation medium; whereas for trans- tensity so that the integrating operation can be well-modeled
mission or emission tomographs, geometrical optics approxi- by the sampling operation
mations are often used and the point-spread functions model
the system geometry and detector uncertainties. Ik [n] γ |Yn |Ik (yn ) (10)
For situations such as astronomical imaging with ground-
based telescopes, each measurement is in the form of a two- where yn is the location of the nth detector element and 兩Y n兩
dimensional image, whereas for tomographic systems each is its integration area.
measurement may be in the form of a one-dimensional projec-
tion of a two-dimensional transmittance or emittance func- Other Detector Effects
tion. In either situation, the imaging task is to reconstruct
In addition to the quantum noise, imaging detectors introduce
the intensity function I( ⭈ ) from noisy measurements of Ik( ⭈ ),
other nonideal effects into the imagery that they record. The
k ⫽ 1, 2, . . ., K.
efficiency with which detectors convert electromagnetic en-
ergy into photoevents can vary across elements within a de-
Quantum Noise in Incoherent Imagery
tector array, and this nonuniform efficiency can be captured
Light and other forms of electromagnetic radiation interact by attaching a gain function to the photocount mean
with matter in a fundamentally random manner, and, be-
cause of this, statistical models are often used to describe the Ik [n] = a[n]γ |Yn |Ik (yn ) (11)
detection of optical waves. Quantum electrodynamics (QED)
is the most sophisticated theory available for describing this Seriously flawed detector elements that fail to record data are
phenomenon; however, a semiclassical theory for the detec- also accommodated with this model by simply setting the gain
tion of electromagnetic radiation is often sufficient for the de- to zero at the appropriate location. If different detectors are
velopment of sound and practical models for imaging applica- used for each observation the gain function may need to vary
tions. When using the semiclassical theory, electromagnetic with each frame and, therefore, be indexed by k.
energy is transported according to the classical theory of wave Because of internal shot noise, many detectors record pho-
propagation—it is only during the detection process that the toevents even when the external light intensity is zero. The
field energy is quantized. resulting photocounts are usually modeled as independent
Figure 1. Image of Simeon Poisson as it

might be acquired by a detector when 1
million, 10 million, and 100 million total
photocounts are recorded.
Poisson random variables, and this phenomenon is accommo- Maximum-Likelihood Image Restoration
dated by inserting a background term into the imaging equa-
Consistent with the noise models developed in the previous
tion
sections, the data recorded by each detector element in a pho-
Ik [n] a[n]γ |Yn |Ik (yn ) + Ib [n] (12) ton-counting camera are a mixture of Poisson and Gaussian
random variables. Accordingly, the probability of receiving N
As with the gain function, if different detectors are used photocounts in the nth detector element is
for each observation this background term may need to vary
with each frame and, therefore, be indexed by k. With the Pr{Nk [n] = N; I} = exp(−Ik [n])(Ik [n])N /N! (14)
inclusion of these background counts, the number of photo-
counts acquired by detector element n is a Poisson random where
variable with mean Ik[n] and is denoted by Nk[n].
The data recorded by many detectors are also corrupted by Ik [n] = a[n]γ |Yn |Ik (yn ) + Ib [n]
another form of noise that is induced by the electronics used (15)
= a[n]γ |Yn | hk (yn , x)I(x) dx + Ib [n]
for the data acquisition. For CCD detectors, this is read-out
noise and is often approximated as additive, zero-mean
Gaussian random variables so that the recorded data are contains the dependence on the unknown intensity function
I( ⭈ ). Furthermore, the probability density for the read-out
modeled as
noise is
dk [n] = Nk [n] + gk [n] (13)
pg (g) = (2πσ 2 [n])−1/2 exp[−g2 /(2σ [n])] (16)
k [n]
where gk[n] models the read-out noise at the nth detector for
the kth observation. The variance of the read-out noise ␴2[ ⭈ ] so that the density for the measured data is
may vary with each detector element, and the read-out noise
for different detectors is usually modeled as statistically inde-
∞
pendent. pd
k
[n] (d; I) = pg
k
[n] (d − N)Pr{Nk [n] = N; I}
The appropriate values for the gain function a[ ⭈ ], back- N=0
ground function Ib[ ⭈ ], and read noise variance ␴2[ ⭈ ] are usu- (2πσ 2 [n])−1/2 ∞
(17)
= exp[−(d − N)2 /(2σ [n])]
ally selected through a controlled study of the data acquisi- N N=0
tion system. A detailed discussion of these and other camera
effects for optical imaging is given in Ref. 23. exp(−Ik [n])(Ik [n])N
For a given data set 兵dk[ ⭈ ]其, the maximum-likelihood estimate the center of each pixel. Many other basis sets are possible
of I( ⭈ ) is the intensity function that maximizes the likelihood and a clever choice here can greatly affect estimator perfor-
mance, but the grid of two-dimensional impulses is probably

K
the most common. Using this basis, the data mean is ex-
l(I) = pd [n] (dk [n]; I) (18)
k pressed as
k=1 n
Ik [n] = a[n]γ |Yn |Ik ( yn ) + Ib [n]

or, as is commonly done, its logarithm (the log-likelihood)
= a[n]γ |Yn | hk ( yn , x) I[m]δ2 (x − xm ) dx + Ib [n]
L (I) = ln l(I) m (24)

K
(19) = a[n]γ |Yn | hk ( yn , xm )I[m] + Ib [n]
= ln pd [n] (dk [n]; I) m
k
k=1 n
where yn denotes the location of the nth measurement, xm de-
The complicated form for the measurement density pdk[n]( ⭈ ; I) notes the location of the mth object pixel, and 웃2( ⭈ ) is the two-
makes this an overly complicated optimization. When the dimensional Dirac impulse. The estimation problem, then, is
read-out noise variance is large (greater than 50 or so), how- one of estimating the discrete samples I[ ⭈ ] of the intensity
ever, ␴2[n] can be added to the measured data to form the function from the noisy data 兵dk[ ⭈ ]其. Because I[ ⭈ ] represents
modified data samples of an intensity function, this function is physically
constrained to be nonnegative.
d˜k [n] = dk [n] + σ 2 [n] Ignoring terms in the log-likelihood that do not depend
= Nk [n] + gk [n] + σ 2 [n] (20) upon the unknown object intensity, the optimization problem
required to solve for the maximum-likelihood object estimate
Nk [n] + Mk [n] is
where Mk[n] is a Poisson-distributed random variable whose

K
mean value is ␴2[n]. The modified data at each detector Î[n] = arg max − (Ik [n] + σ 2 [n])
I≥0
element are then similar (in distribution) to the sum of two k=1 n
(25)
Poisson-distributed random variables Nk[n] and Mk[n] and,
K
as such, are also Poisson-distributed with the mean value + d˜k [n] ln(Ik [n] + σ 2 [n])
Ik[n] ⫹ ␴2[n]. This approximation is discussed by Snyder et al. k=1 n
in Refs. 23 and 24. The probability mass function for the mod-
ified data is then modeled as where d̃k[n] ⫽ dk[n] ⫹ ␴2[n] is the modified data and

Pr[d˜k [n] = D; I] = exp{−(Ik [n] + σ 2 [n])}(Ik [n] + σ 2 [n])D /D! Ik [n] = a[n]γ |Yn | hk ( yn , xm )I [m] + Ib [n]
m
(21)
is the photocount mean. The solution to this problem gener-
so that the log-likelihood is ally requires the use of a numerical method, and a great num-

K ber of techniques are available for this purpose. General-pur-
L (I) = {−(Ik [n] + σ 2 [n]) pose techniques such as those described in popular texts on
k=1 n (22) optimization theory (25,26) can be applied. In addition, spe-
+ d˜k [n] ln(Ik [n] + σ 2 [n]) − ln dk [n]!} cialized numerical methods devised specifically for the solu-
tion of maximum-likelihood and related problems can be ap-
Two difficulties are encountered when attempting to find the plied (27,28)—a specific example is discussed in the
intensity function I( ⭈ ) that maximizes the log-likelihood following section.
L (I): (1) the recovery of an infinite-dimensional function I( ⭈ )
from finite data is a terribly ill-conditioned problem; and (2) The Expectation-Maximization Method. The expectation-
the functional form of the log-likelihood does not admit a maximization (EM) method is a numerical technique devised
closed form, analytic solution for the maximizer even after specifically for maximum-likelihood estimation problems. As
the dimension of the parameter function has been reduced. described in Ref. 27, the classical formulation of the EM pro-
To address the dimensionality problem, it is common to cedure requires one to augment the measured data—
approximate the parameter function in terms of a finite-di- commonly referred to as the incomplete data—with a set of
mensional basis set complete data which, if measured, would facilitate direct esti-
mation of the unknown parameters. The application of this

I(x) I[m]ψm (x) (23) procedure then requires one to alternately apply an E-step,
m wherein the conditional expectation of the complete-data log-
likelihood is determined, and an M-step, wherein all parame-
where the basis functions 兵␺m( ⭈ )其 are chosen in an appropriate ters are simultaneously updated by maximizing the expecta-
manner. When expressing the object function with a predeter- tion of the complete-data log-likelihood with respect to all of
mined grid of image pixels, for example, ␺m( ⭈ ) might be an the unknown parameters. In general, the application of the
indicator function that denotes the location of the mth pixel. EM procedure results in an iterative algorithm that produces
For the same situation, the basis functions might alterna- a sequence of parameter estimates that monotonically in-
tively be chosen as two-dimensional impulses co-located with creases the measured data likelihood.
The application of the EM procedure to the incoherent im- The intensity estimate is then updated in the M-step by max-
aging problems has been proposed and described for numer- imizing this conditional expectation over I
ous applications (29–32). The general application of this
method is outlined as follows. First, recall that the measured I new = arg max Q (I; I old ) (32)
(or incomplete) data d̃k[n] for each observation k and detector I≥0
element n are independent Poisson variables with the ex-

pected value It is straightforward to show that the object estimate is then
updated according to

E{d˜k [n]} = a[n]γ |Yn | hk ( yn , xm )I [m] + Ib [n] + σ 2 [n] (26)
m E [N ck [n, m]|{d˜k [n]}; I old ]
k n
I new [m] = (33)
Because the sum of Poisson random variables is still a Pois- a[n]γ |Yn |hk ( yn , xm )
son random variable (and the expected value is the sum of k n
the individual expected values), the incomplete data can be

statistically modeled as As described in Ref. 29, the conditional expectation is evalu-
ated as

d˜k [n] = Nkc [n, m] + Mkc [n] (27)
m E [N ck [n, m]|{d˜k [n]}; I old ]
a[n]γ |Yn hk ( yn , xm )I old [m]
where for all frames k, detector locations n, and object pixels = d˜k [n] (34)
a[n]γ |Yn hk ( yn , xm )I old [m ] + Ib [n] + σ 2 [n]
m, the data Nck[n, m] are Poisson random variables, each with
m
the expected value
so that the iterative formula for updating the object estimate
E{Nkc [n, m]} = a[n]γ |Yn |hk ( yn , xm )I [m] (28) is
and for all frames k and detector locations n, the data Mck[n] I new [m] = I old [m]
are Poisson random variables, each with the expected value
h k ( yn , xm )
E{M kc [n]} = Ib [n] + σ 2 [n] k n

(29) a[n]γ |Yn |d˜k [n]
 
a[n]γ |Yn hk ( yn , xm )I old [m ] + Ib [n] + σ 2 [n]
In the terminology of the EM method, these data 兵Nck[ ⭈ , ⭈ ], m
Mck[ ⭈ ]其 are the complete data, and although they cannot be (35)
a[n]γ |Yn |hk ( yn , xm )
observed directly, their measurement, if possible, would k n
greatly facilitate direct estimation of the underlying object in-
tensity. For the special case of uniform gain with no background or
Because the complete data are independent, Poisson ran- detector noise, the iterative algorithm proposed by Richard-
dom variables, the complete-data log-likelihood is son (33) and Lucy (34) has the same form as these iterations.
An excellent historical perspective of the application of the
L c (I ) = − a[n]γ |Yn |hk ( yn , xm )I [m] EM method to imaging problems is presented in Ref. 35, and
k n m detailed discussions of the convergence properties of this algo-

+ N ck [n, m] ln(a[n]γ |Yn |hk ( yn , xm )I [m]) rithm along with the pioneering derivations for applications
k n m in emission tomography can be found in Ref. 36.
(30) Figures 2 and 3 illustrate the use of this technique on im-
agery acquired by the Hubble Space Telescope (HST). Shortly
where terms not dependent upon the unknown object inten- after the launch of the HST with its aberrated primary mirror
sity I[ ⭈ ] have been omitted. Given an estimate for the object in 1990, the imagery acquired by this satellite became a focus
intensity Iold[ ⭈ ], the EM procedure makes use of the complete of national attention. Whereas microscopic flaws in the tele-
data and their corresponding log-likelihood to update the ob- scope’s mirror resulted in the severely distorted imagery, im-
ject intensity estimate in such a way that Inew[ ⭈ ] increases the age restoration methods were successful in restoring much of
measured data log-likelihood. The E-step of the EM procedure the lost resolution (4). Figure 2, for example, shows imagery
requires the expectation of the complete-data log-likelihood, of the star cluster R136 in a star formation called 30 Doradus
conditional on the measured (or incomplete) data and using as acquired by the telescope and as restored using the meth-
the old object intensity estimate Iold[ ⭈ ] ods described in this article. Also shown in this figure are
imagery acquired by the telescope after its aberrated mirror
Q (I; I old ) = E[L c (I)|{d˜k [n]}; I old ] was corrected, along with a processed image showing the po-
tential advantage of applying image restoration methods to
=− a[n]γ |Yn |hk ( yn , xm )I [m]
n m
imagery acquired after the correction. Figure 3 contains an
k
(31) image of Saturn along with restorations formed by simple in-
+ E [N ck [n, m]|{d˜k [n]}; I old ] verse filtering, Wiener filtering, and by the maximum-likeli-
k n m
hood method. According to scientific staff at the Space Tele-
ln(a[n]γ |Yn |hk ( yn , xm )I [m]) scope Science Institute, the maximum-likelihood restoration
ever, and because of this the maximum-likelihood image esti-

mates frequently exhibit severe noise artifacts. Common
methods for addressing this problem are discussed briefly in
this section.
Stopping Rules. Probably the simplest method to imple-
ment for overcoming the noise artifacts seen in maximum-
likelihood image estimates obtained by numerical procedures
is to terminate the iterative process before convergence. Im-
plementation of such a procedure is straightforward; however,
the construction of optimal ‘‘stopping rules’’ can be challeng-
ing. Criteria for developing these rules for problems in coher-
ent imaging are discussed in Refs. 21, 37, 38.
Sieve Methods. The basic idea behind the method of sieves
is to constrain the set of allowable image estimates to be in a
smooth subset called a sieve. The sieve is selected in a man-
ner that depends upon the degree to which the problem is
ill-conditioned and upon the noise level. Badly ill-conditioned
problems and noisy data require a ‘‘small’’ sieve set con-
taining only very smooth functions. Problems that are better
conditioned with little noise can accommodate ‘‘large’’ sieve
sets, and the sieve is ideally selected so that its ‘‘size’’ grows
with decreasing noise levels in such a manner that the con-
strained image estimate converges to the true image as the
noise level shrinks to zero. Establishing this consistency prop-
Figure 2. Imagery of the star cluster R136 in the star formation 30 erty for a sieve can, however, be a difficult task.
Doradus as acquired by the Hubble Space Telescope both before and The general method of sieves as a statistical inference tool
after its aberrated primary mirror was corrected. Upper left: raw data was introduced by Grenander (20). The application of this
acquired with the aberrated primary mirror; upper right: restored method to problems in incoherent imaging was proposed and
image obtained from imagery acquired with the aberrated primary investigated by Snyder et al. (39,40). The method is based on
mirror; lower left: raw data acquired after correction; lower right: re- a kernel sieve defined according to
stored image obtained from imagery acquired after the correction.

(Courtesy of R. J. Hanisch and R. L. White, Space Telescope Science
Institute and NASA.) S = I : I[m] = s[m, p]α[ p] (36)
p
provides the best trade-off between resolution and noise am- where intensity functions within the sieve set S are deter-
plification. mined by the nonnegative parameters 兵움[p]其. The sieve-con-
strained optimization problem then becomes one of maximiz-
Regularization. Under reasonably unrestrictive conditions, ing the likelihood subject to the additional constraint I 僆 S .
the EM method described in the previous section produces a The smoothness properties of the sieve are induced by the
sequences of images that converges to a maximum-likelihood sieve kernel s[ ⭈ , ⭈ ]. With a Gaussian kernel, for instance, the
solution (36). Imaging problems for which this method is ap- smoothness of the sieve set is determined by the variance pa-
plicable are often ill-conditioned or practically ill-posed, how- rameter ␴

1 (m − p)2
s[m, p] = √ exp − (37)
2πσ 2 2σ 2
This Gaussian kernel was investigated in Refs. 39, 40, but

kernels with other mathematical forms can be used. The EM
method can, with straightforward modifications, be applied to
problems in which kernel sieves are used for regularization.
Penalty and MAP Methods. Another method for regularizing
maximum-likelihood estimation problems is to augment the
likelihood with a penalty function
C (I ) = L (I ) − γ (I ) (38)
where ⌽ is a function that penalizes undesirable qualities (or

Figure 3. Raw imagery and restorations of Saturn as acquired by
the Hubble Space Telescope. From left to right: telescope imagery; rewards desirable ones) of the image estimate, and 웂 is a non-
restoration produced by simple inverse filtering; restoration produced negative scale factor that determines the relative contribution
by Wiener filtering; restoration produced by the maximum-likelihood of the penalty to the optimization problem. The penalized im-
method. (Courtesy of R. J. Hanisch and R. L. White, Space Telescope age estimate is then selected to maximize the cost function C ,
Science Institute and NASA. which involves a trade between maximizing the likelihood L
and minimizing the penalty ⌽. The choice of the penalty can where p is an index to sensor locations (either real or syn-
greatly influence the resulting image estimate, as can the se- thetic), up is the complex-amplitude measured by the pth sen-
lection of the scale factor 웂. A commonly used penalty is the sor, u(x) is the complex-amplitude of the field that is reflected
quadratic smoothness penalty from an object or scene of interest, hp(x) is a sensor response
function for the pth sensor measurement, and wp accounts for
(I ) = wnm (I [n] − I [m])2 (39) additive sensor noise. The response function accounts for both
n m∈Nn the sensor characteristics and for wave propagation from the
object or scene to the sensor; in the Fraunhofer approximation
where N n denotes a neighborhood of pixel locations about the for wave propagation, these functions take on the form of a
nth object pixel, and the coefficients wnm control the link be- Fourier-transform kernel (10).
tween pixel n and m. This penalty can also be induced by When the object or scene gives rise to diffuse reflections,
using a MAP formulation with Gaussian Markov random field the Gaussian speckle model (50) is often used as a statistical
(GMRF) prior model for the object. However, because the use model for the reflected field u( ⭈ ). That is, u( ⭈ ) is modeled as
of this penalty often results in excessive smoothing of the ob- a complex Gaussian random process (13,51,52) with zero-
ject edges, alternative penalties have been developed and in- mean and the covariance
vestigated (41–43). A particularly interesting penalty is in-
duced by using a MAP formulation with the generalized E [u(x)u∗ (x )] s(x)δ2 (x − x ) (42)
Gaussian Markov random field (GGMRF) model (43). The use
of this prior results in a penalty function of the form where s( ⭈ ) is the object incoherent scattering function. The
sensor noise is often modeled as zero-mean, independent com-
(I ) = γ q wnm |I [n] − I [m]|q (40) plex Gaussian variables with variance ␴2 so that the recorded
n m∈Nn data are complex Gaussian random variables with zero-mean
and the covariance
where q 僆 [1, 2] is a parameter that controls the smoothness
of the reconstruction. For q ⫽ 2 this is the common quadratic E [u p u∗p ] = h p (x)h∗p (x)s(x) dx + σ 2 δ[ p − p ] (43)
smoothing penalty, whereas smaller values of q will, in gen-
eral, allow for sharper edges in the object estimates. where 웃[ ⭈ ] is the Kronecker delta function. The maximum-
Although the EM method is directly applicable to problems likelihood estimation of the object scattering function s( ⭈ )
in which stopping rules or kernel sieves are used, the EM then becomes a problem of covariance estimation subject to
approach is less simple to use when penalty or MAP methods the linear structure constraint of Eq. (43).
are employed. The major difficulty arises because the maximi- Using vector-matrix notation the data covariance is, as a
zation step usually has no closed-form solution; however, ap- function of the unknown object scattering function
proximations and modifications can be used (41,44) to address
this problem. R (s) = E [uu
uu† ]
(44)
= h (x)hh† (x)s(x) dx + σ 2I
Alternative Numerical Approaches
A major difficulty encountered when using the EM method where u ⫽ [u1u2 ⭈ ⭈ ⭈ uP]T is the data vector, h(x) ⫽ [h1(x)h2(x)
for incoherent-imaging problems is its slow convergence (45). ⭈ ⭈ ⭈ hP(x)]T is the system response vector, [ ⭈ ]T denotes matrix
Many methods have been proposed to overcome this problem, transposition, [ ⭈ ]† denotes Hermitian matrix transposition,
and a few of these are summarized briefly here. Because of and I is the P ⫻ P identity matrix. Accordingly, the data log-
the similarities of the EM method to gradient ascent, line- likelihood is
search methods can be used to accelerate convergence (45), as
L(s) = − ln det[R R −1 (s)S
R (s)] − tr[R S] (45)
can other gradient-based optimization methods (46,47). Sub-
stantial improvements in convergence can also be obtained by where S ⫽ uu† is the data sample-covariance. Parameteriza-
using a generalization of the EM method—the space-alternat- tion of the parameter function as in Eq. (23) is a natural step
ing generalized expectation-maximization (SAGE) method before attempting to solve this problem, but direct maximiza-
(28,48)—whereby convergence is accelerated through a novel tion of the likelihood is still a difficult problem. Because of
choice for the complete data at each iteration. In addition, a this, the EM method has been proposed and discussed in Refs.
coordinate descent (or ascent) optimization method has been 53–55 for addressing this problem, and the resulting algo-
shown to provide for greatly reduced computational time (49). rithm has been shown to produce parameter estimates with
lower bias and variance than alternative methods (56). A ma-
COHERENT IMAGING jor problem with this method, though, is the high computa-
tional cost; however, the application of the SAGE method (28)
For synthetic aperture radar (SAR), ultrasound, and other to this problem has shown great promise for reducing the
forms of coherent array imaging, an object or scene is illumi- computational burden (57). The development and application
nated by a highly coherent source (such as a radar transmit- of regularization methods for problems in coherent imaging is
ter, laser, or acoustic transducer), and heterodyne, homodyne, an area of active research.
or holographic methods are used to record amplitude and
phase information about the reflected field. The basic signal SUMMARY
model for these problems is of the form:
Imaging science is a rich and vital area of science and tech-
u p = h p (x)u(x) dx + w p , p = 1, 2, . . ., P (41) nology in which information-theoretic methods can be and
have been applied with great benefit. Maximum-likelihood 22. L. Mandel and E. Wolf, Optical Coherence and Quantum Optics,
methods can be applied to a variety of problems in image res- New York: Cambridge University Press, 1995.
toration and synthesis, and their application to the restora- 23. D. L. Snyder, A. M. Hammoud, and R. L. White, Image recovery
tion problem for incoherent imaging has been discussed in from data acquired with a charge-coupled-device camera, J. Opt.
great detail in this article. To conclude, the future of this field Soc. Am., A, 10 (5): 1014–1023, 1993.
is best summarized by the following quote from Bracewell 24. D. L. Snyder et al., Compensation for readout noise in CCD im-
(58): ages, J. Opt. Soc. Am., A, 12 (2): 272–283, 1995.
25. D. G. Luenberger, Linear and Nonlinear Programming, Reading,
The study of imaging now embraces many major areas of modern
technology, especially the several disciplines within electrical en- 26. R. Fletcher, Practical Methods of Optimization, New York: Wi-
gineering, and will be both the stimulus for, and recipient of, new ley, 1987.
advances in information science, computer science, environmental 27. A. P. Dempster, N. M. Laird, and D. B. Rubin, Maximum likeli-
science, device and materials science, and just plain high-speed hood from incomplete data via the EM algorithm, J. R. Stat. Soc.,
computing. It can be confidently recommended as a fertile subject B, 39: 1–37, 1977.
area for students entering upon a career in engineering. 28. J. A. Fessler and A. O. Hero, Space-alternating generalized ex-
pectation-maximization algorithm, IEEE Trans. Signal Process.
42: 2664–2677, 1994.
BIBLIOGRAPHY 29. L. A. Shepp and Y. Vardi, Maximum-likelihood reconstruction for
emission tomography, IEEE Trans. Med. Imaging, MI-1: 113–
1. H. Stark (ed.), Image Recovery: Theory and Application, Orlando, 121, 1982.
FL: Academic Press, 1987. 30. D. L. Snyder and D. G. Politte, Image reconstruction from list-
2. A. K. Katsaggelos (ed.), Digital Image Restoration, Heidelberg: mode data in an emission tomography system having time-of-
Springer-Verlag, 1991. flight measurements, IEEE Trans. Nucl. Sci., NS-30: 1843–
3. R. L. Lagendijk and J. Biemond, Iterative Identification and Resto- 1849, 1983.
ration of Images, Boston: Kluwer, 1991. 31. K. Lange and R. Carson, EM reconstruction algorithms for emis-
4. R. J. Hanisch and R. L. White (eds.), The Restoration of HST sion and transmission tomography, J. Comput. Assisted Tomogra-
Images and Spectra—II, Baltimore, MD: Space Telescope Science phy, 8: 306–316, 1984.
Institute, 1993. 32. T. J. Holmes, Maximum-likelihood image restoration adapted for
noncoherent optical imaging, J. Opt. Soc. Am., A, 6: 666–673,
5. H. L. Van Trees, Detection, Estimation, and Modulation Theory,
1989.
Part I, New York: Wiley, 1968.
33. W. H. Richardson, Bayesian-based iterative method of image res-
6. H. V. Poor, An Introduction to Signal Detection and Estimation,
toration, J. Opt. Soc. Am., 62 (1): 55–59, 1972.
2nd ed., New York: Springer-Verlag, 1994.
34. L. B. Lucy, An iterative technique for the rectification of observed
7. B. Porat, Digital Processing of Random Signals: Theory and Meth-
distributions, Astronom. J., 79 (6): 745–754, 1974.
ods, Englewood Cliffs, NJ: Prentice-Hall, 1993.
35. Y. Vardi and D. Lee, From image deblurring to optimal invest-
8. L. L. Scharf, Statistical Signal Processing: Detection, Estima-
ments: Maximum likelihood solutions for positive linear inverse
tion, and Time Series Analysis, Reading, MA: Addison-Wesley,
problems, J. R. Stat. Soc. B, 55 (3): 569–612, 1993.
1991.
36. Y. Vardi, L. A. Shepp, and L. Kaufman, A statistical model for
9. S. M. Kay, Modern Spectral Estimation: Theory and Applications, positron emission tomography, J. Amer. Stat. Assoc., 80: 8–37,
Englewood Cliffs, NJ: Prentice-Hall, 1988. 1985.
10. J. W. Goodman, Introduction to Fourier Optics, 2nd edition, New 37. T. Hebert, R. Leahy, and M. Singh, Fast MLE for SPECT using
York: McGraw-Hill, 1996. an intermediate polar representation and a stopping criterion,
11. J. D. Gaskill, Linear Systems, Fourier Transforms, and Optics, IEEE Trans. Nucl. Sci., NS-34: 615–619, 1988.
New York: Wiley, 1978. 38. J. Llacer and E. Veklerov, Feasible images and practicle stopping
12. M. Born and E. Wolf, Principles of Optics, 6th edition, Elmsford, rules for iterative algorithms in emission tomography, IEEE
NY: Pergamon, 1980. Trans. Med. Imaging, MI-8: 186–193, 1989.
13. J. W. Goodman, Statistical Optics, New York: Wiley, 1985. 39. D. L. Snyder and M. I. Miller, The use of sieves to stabilize im-
14. B. E. A. Saleh and M. C. Teich, Fundamentals of Photonics, New ages produced with the EM algorithm for emission tomography,
York: Wiley, 1991. IEEE Trans. Nucl. Sci., NS-32: 3864–3871, 1985.
15. A. Papoulis, Probability, Random Variables, and Stochastic Pro- 40. D. L. Snyder et al., Noise and edge artifacts in maximum-likeli-
cesses, 3rd edition, New York: McGraw-Hill, 1991. hood reconstructions for emission tomography, IEEE Trans. Med.
Imaging, MI-6: 228–238, 1987.
16. R. M. Gray and L. D. Davisson, Random Processes: An Introduc-
tion for Engineers, Englewood Cliffs: Prentice-Hall, 1986. 41. P. J. Green, Bayesian reconstructions from emission tomography
data using a modified EM algorithm, IEEE Trans. Med. Imaging,
17. A. Tikhonov and V. Arsenin, Solutions of Ill-Posed Problems, 9: 84–93, 1990.
Washington, DC: Winston, 1977.
42. K. Lange, Convergence of EM image reconstruction algorithms
18. W. L. Root, Ill-posedness and precision in object-field reconstruc- with Gibbs priors, IEEE Trans. Med. Imaging, MI-9: 439–446,
tion problems, J. Opt. Soc. Am., A, 4 (1): 171–179, 1987. 1990.
19. J. R. Thompson and R. A. Tapia, Nonparametric Function Estima- 43. C. A. Bouman and K. Sauer, A generalized Gaussian image
tion, Modeling, and Simulation, Philadelphia: SIAM, 1990. model for edge-preserving MAP estimation, IEEE Trans. Image
20. U. Grenander, Abstract Inference, New York: Wiley, 1981. Process., 2: 296–310, 1993.
21. E. Veklerov and J. Llacer, Stopping rule for the MLE algorithm 44. T. Hebert and R. Leahy, A generalized EM algorithm for 3-D
based on statistical hypothesis testing, IEEE Trans. Med. Im- Bayesian reconstruction from Poisson data using Gibbs priors,
aging, 6: 313–319, 1987. IEEE Trans. Med. Imaging, MI-8: 194–202, 1989.
448 MEASUREMENT ERRORS
45. L. Kaufman, Implementing and accelerating the EM algorithm

for positron emission tomography, IEEE Trans. Med. Imaging,
MI-6: 37–51, 1987.
46. E. U. Mumcuoglu et al., Fast gradient-based methods for Bayes-
ian reconstruction of transmission and emission PET images,
IEEE Trans. Med. Imag., MI-13: 687–701, 1994.
47. K. Lange and J. A. Fessler, Globally convergent algorithm for
maximum a posteriori transmission tomography, IEEE Trans.
Image Process., 4: 1430–1438, 1995.
48. J. A. Fessler and A. O. Hero, Penalized maximum-likelihood im-
age reconstruction using space-alternating generalized EM algo-
rithms, IEEE Trans. Image Process., 4: 1417–1429, 1995.
49. C. A. Bouman and K. Sauer, A unified approach to statistical
tomography using coordinate descent optimization, IEEE Trans.
Image Process., 5: 480–492, 1996.
50. J. C. Dainty (ed.), Laser speckle and related phenomena. Topics
in Applied Physics, vol. 9, 2nd ed., Berlin: Springer-Verlag, 1984.
51. F. D. Neeser and J. L. Massey, Proper complex random processes
with applications to information theory, IEEE Trans. Inf. Theory,
39: 1293–1302, 1993.
52. K. S. Miller, Complex Stochastic Processes: An Introduction to The-
ory and Applications, Reading, MA: Addison-Wesley, 1974.
53. D. B. Rubin and T. H. Szatrowski, Finding maximum likelihood
estimates of patterned covariance matrices by the EM algorithm,
Biometrika, 69 (3): 657–660, 1982.
54. M. I. Miller and D. L. Snyder, The role of likelihood and entropy
in incomplete-data problems: Applications to estimating point-
process intensities and Toeplitz constrained covariances, Proc.
IEEE, 75: 892–907, 1987.
55. D. L. Snyder, J. A. O’Sullivan, and M. I. Miller, The use of maxi-
mum-likelihood estimation for forming images of diffuse radar-
targets from delay-doppler data, IEEE Trans. Inf. Theory, 35:
536–548, 1989.
56. M. J. Turmon and M. I. Miller, Maximum-likelihood estimation of
complex sinusoids and Toeplitz covariances, IEEE Trans. Signal
Process., 42: 1074–1086, 1994.
57. T. J. Schulz, Penalized maximum-likelihood estimation of struc-
tured covariance matrices with linear structure, IEEE Trans. Sig-
nal Process., 45: 3027–3038, 1997.
58. R. N. Bracewell, Two-Dimensional Imaging, Upper Saddle River,
TIMOTHY J. SCHULZ
Michigan Technological University
MEASUREMENT. See ACCELERATION MEASUREMENT; DEN-

SITY MEASUREMENT;
DISPLACEMENT MEASUREMENT; MAG-
NETIC FIELD MEASUREMENT;
MILLIMETER WAVE MEASURE-
MENT; Q-FACTOR MEASUREMENT.
MEASUREMENT, ATTENUATION. See ATTENUATION
MEASUREMENT.
MEASUREMENT, C-V. See C-V PROFILES.
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Queueing Theory ● Email this page Enter words or
● Print this page
Nader Mehravari1
1Lockheed Martin, Owego, NY
reserved. Search
Abstract
History of the Development of Queueing Theory
Applications of Queueing Theory
Specification and Characterization of Queueing Systems
Notions of Probability Theory of Importance to the Study of Queues
Modeling and Analysis of Elementary Queueing Systems
References To More Advanced Topics

QUEUEING THEORY tain limit? What are subjective and economical advantages
and disadvantages of modifying various parameters of the
systems such as the number of servers or the size of the
TELETRAFFIC THEORY waiting room? How often is the server busy? Queueing the-
ory attempts to answer these and other related questions
through detailed mathematical analysis and provides us
NETWORK OF QUEUES with the necessary tools to evaluate related performance
measures.
All of us, either directly or through the use of various ma- The purpose of this article is to provide an introductory
chines that we have become dependent upon, wait for ser- overview of the fundamental notions of queueing theory.
vice in a variety of lines on a regular basis. Customers The remaining sections of this article will discuss the fol-
wait in lines at banks to be served by a bank teller; drivers lowing topics: a brief history of the development of queue-
wait in their cars in traffic jams or at toll booths; patients ing theory; applications of queueing theory; specification
wait in doctors’ waiting rooms; electronic messages wait and characterization of queueing systems; notions of prob-
in personal computers to be delivered over communication ability theory of importance to queueing theory; modeling
networks; telephone calls are put on hold to be answered and analysis of elementary queueing systems; references
by operators; computer programs are stored in computer to more advanced topics; and a list of references.
memory to be executed by a time-sharing computer sys-
tem; and so on. In many situations, scarce resources are to
be shared among a collection of users who require the use HISTORY OF THE DEVELOPMENT OF QUEUEING
of these resources at unspecified times. They also require THEORY
the use of these resources for random periods of time. This
probabilistic nature of requests causes these requests to ar- The English word “queue” is borrowed from the French
rive while the resources are in use by other members of the word “queue” which itself is taken from the Latin word
user community. A mechanism must be put in place to pro- “cauda” meaning “tail.” Most researchers and scientists
vide an orderly access to the resources requested. The most in the field prefer the spelling “queueing” over “queuing.”
common mechanism is to put the user requests in a wait- However, many American dictionaries and software spell
ing line or “queue.” “Queueing theory” deals with the study checkers prefer the spelling “queuing.” For further discus-
of the behavior and the control of waiting lines. It provides sion of “queueing” vs. “queuing” spelling, see Refs. 27, 28.
us with the necessary mathematical structure and proba- Queueing theory has been under development since the
bility tools to model, analyze, study, evaluate, and simulate early years of this century. It has since progressed con-
systems involving waiting lines and queues. It is a branch siderably, and today it is based upon a vast collection of
of applied mathematics, applied probability theory, and op- results, methods, techniques, and voluminous literature. A
erations research. It is known under various names such good summary of the early history of queueing theory can
as: queueing theory, theory of stochastic server systems, be found in Ref. 6, pp. 20–25.
theory of systems of flows, traffic or teletraffic theory, con- Historically, queueing theory originated as a very prac-
gestion theory, and theory of mass service. Standard texts tical subject. It was developed to provide models to predict
on queueing theory include Refs. 1–31. For a summary of the behavior of systems that attempt to provide service for
many of the most important results in queueing theory, the randomly arising demands. Much of the early work was de-
reader is referred to a survey paper by Cooper (7). For a bib- veloped in relation with problems in telephone traffic en-
liography of books and survey papers on queueing theory gineering. The pioneering work of Agner Krarup Erlang,
see Refs. 8, 29. For nontechnical articles explaining queue- from 1909 to 1929, laid the foundations of modern teletraf-
ing theory for the layman the reader is referred to Refs. 9, fic and queueing theory. Erlang, a Danish mathematician
26. and engineer who worked for the Copenhagen Telephone
A typical queueing system can be described as one where Exchange, published his first article in 1909 on the applica-
customers arrive for service, wait for service, and, leave the tion of probability theory to telephone traffic problems (10).
system after being served. The service requests occur ac- Erlang’s work soon drew the attention of other probability
cording to some stochastic process, and the time required theorists such as T. C. Fry and E. C. Molina in the 1920s,
for the server(s) to service a request is also probabilistically who expanded much of Erlang’s work on the application of
distributed. In general, arrivals and departures (i.e., ser- the theory to telephone systems. Telephony remained one
vice completions) cannot be synchronized, so waiting time of the principal applications until about 1950.
may result. It is, therefore, critical to be able to character- In the years immediately following World War II, ac-
ize waiting time and many other important performance tivity in the fields of probability theory and operations re-
measures of a queueing system. For a typical queueing search (11, 12) grew rapidly, causing a new surge of in-
system, one is interested in answering questions such as: terest in the subject of queueing theory. In the late 1950s,
How long does a typical customer have to wait? What is the queueing theory became one of the most popular subjects
number of customers in the system at any given point in within the domains of applied mathematics and applied
time? How large should the waiting room be to accommo- probability theory. This popularity, however, was fueled
date certain percentage of potential customers? How many by its mathematical challenges and not by its applica-
servers are needed to keep the waiting time below a cer- tions. Clever and elegant mathematical techniques has en-
abled researchers (such as Pollaczek, Kolmogorov, Khin-
J. Webster (ed.), Wiley Encyclopedia of Electrical and Electronics Engineering. Copyright © 2007 John Wiley & Sons, Inc.
2 Network of Queues
chine, Crommelin, and Palm) to derive exact solutions for sonal preferences. Some customers often join the shortest
a large number of mathematical problems associated with queue, and some join the queue of a particular teller that
models of queueing systems. Regrettably, in the period of they personally know, whereas others may join the queue of
1950–1970, queueing theory, which was originated as a the teller that is perceived to be the fastest. On the other ex-
very practical subject, had become of little direct practical treme, some banks (via the use of various directional signs
value. and/or ropes) direct all the incoming customers into a sin-
Since the 1970s there has seen a rebirth and explosion gle waiting line that feeds all the tellers. The customer at
of queueing theory activities with an emphasis on practi- the head of this queue is then served by the next available
cal applications. The performance modeling and analysis of bank teller. The question now becomes which one of these
computer systems and data transmission networks opened two modes of operation is more appropriate. The answer
the way to investigate queues characterized by complex strongly depends on such parameters as the performance
service disciplines and interconnected systems. Most of measures that the bank management is interested in op-
the theoretical advances since the 1970s are directly at- timizing, the number and the speed of the tellers, the type
tributable to developments in the area of computer systems of banking transactions, and the number of incoming cus-
performance evaluation as represented in Refs. 13–16. tomers visiting the bank in a typical day. Similar issues
arise in other cases such as supermarket checkout coun-
ters, fast-food restaurants, airport landing and departure
APPLICATIONS OF QUEUEING THEORY schedules, and multiprocessor computer systems. Queue-
ing theory methods enable us to model, analyze, and decide
Interest in queueing theory has often been stimulated on the best strategy for such applications.
by practical problems and real world situations. Queue-
ing theory concepts have applications in many disciplines
SPECIFICATION AND CHARACTERIZATION OF
such as telephone systems traffic engineering, migration
QUEUEING SYSTEMS
and population models in biology, electrical and fluid flow
models, clustering models in chemistry, manufacturing sys-
Figure 1 represents the basic elements of a queueing sys-
tems, computer systems, digital data transmission sys-
tem. As shown in Fig. 1, a basic queueing system is one
tems, flow through communication networks, inventory
where members of a population (i.e., customers or entities
control, time sharing and processor sharing computer sys-
of some kind) arrive at a service station to receive service
tems, telecommunications, machine repair, taxi stands,
of some type. After receiving service, the units depart the
aircraft landing, loading and unloading ships, scheduling
service facility. A “queue” or waiting line is developed when-
patients in hospitals, factory production flow, intelligent
ever the service facility cannot service all the units requir-
transportation systems, call centers, and so on. There are
ing service. Although many queueing systems may be rep-
many other important applications of the queueing theory
resented by similar diagrams, an accurate representation
as presented in Refs. 1–6 and 13–16. We elaborate further
of such a system requires a detailed characterization of the
on only two of these applications in this section.
underlying parameters and processes.
Queueing theory has played a major role in the study
of both packet switching and circuit switching communica-
Key Parameters and Varieties of Queueing Systems
tion networks. Queueing arises naturally in packet switch-
ing networks where user messages are broken into small To fully describe a queueing system analytically, various
units of transmission called packets. Packets arriving at aspects and parameters of the system must be known. The
various intermediate network nodes, on the way to their fi- most important of them are presented here.
nal destination, are buffered in memory, processed to deter-
mine the appropriate outgoing route, and then are trans- The Arrival Pattern. Let the successive customers arrive
mitted on the chosen outgoing link when their time for to the system at times t1 , t2 , t3 , . . . , where 0 ≤ t1 < t2 <
transmission comes up. If, for example, the chosen outgo- t3 < ··· < tn < ···. Then we define yi = ti+1 − ti , where i =
ing link is in use when it is time for a given packet to be 1, 2, 3, . . . , as the interarrival times of the customers. We
transmitted, then that packet must be kept in the memory normally assume that arrival times form a stochastic pro-
(i.e., queued) until the link becomes available. The time cess and that the interarrival times, yi , are independent
spent in the buffer waiting for transmission is an impor- and identically distributed (iid) according to probability
tant measure of system performance. This waiting time distribution function A(·), where A(τ) = P(yi ≤ τ). Function
depends on various parameters such as nodal processing A(·) is then referred to as the interarrival time distribution
power, transmission link speed, packet lengths, traffic rates or simply the arrival distribution. Additional information
in terms of packets per second, and so on. Queueing theory such as whether each arrival event contains one or a group
provides the necessary mathematical tools to model and of customers of fixed or random size (i.e., “bulk arrivals”)
analyze such queueing configurations. can also be specified if applicable.
For another example of application of queueing theory
consider a typical bank and the mechanism that bank man- Customer Population and Behavior. The customer popu-
agement has put in place to direct incoming customers to lation, or the source of the customers, can either be finite or
the available bank tellers. In some banks, each teller has infinite. Infinite customer populations are normally easier
his or her own queue and incoming customers are free to to describe mathematically and analyze their performance
join the waiting line of any of the tellers based on some per- analytically. This is because in a finite population source
Network of Queues 3
Figure 1. Basic elements of a typical

queueing system.
model, the number of customers in the system affects the 3. Service in Random Order (SIRO) or Random Selec-
arrival rate which in turn makes the analysis more diffi- tion for Service (RSS) The customer to be served next
cult. In addition to the properties of the entire customer is chosen stochastically from the waiting customers
population, behavior of individual customers could also be according to a uniform probability distribution. In
of importance and, therefore, must be formally specified. general, the probability distribution used to choose
For example, if a customer decides not to join the system af- the next customer could be any discrete probability
ter seeing the size of the queue, it is said that the customer distribution.
has “balked.” Or, for example, a customer is said to have “re- 4. Priority (PR or PRI) There could also be some no-
neged” if, after having waited in the queue for some time, tion of priority in the queueing system where the
he or she becomes impatient and leaves the system before customer population is divided in two or more pri-
his service begins. Customers, if allowed, may “jockey” from ority classes. Any waiting member of a higher prior-
one queueing system to another (with a perceived shorter ity class is chosen to be served before any customer
waiting time, for example). from a lower priority class. Queueing systems with
priority classes are divided into two types. Under a
“preemptive priority” discipline, whenever a higher
The Service Mechanism. The queue’s service mechanism
priority customer arrives while a lower priority cus-
is described by specifying the number of servers, c, and
tomer is in service, the lower priority customer is pre-
the stochastic characterization of the service times. It is
empted and is taken out of service without having his
normally assumed that the service times of successive cus-
service completed. In this case, the preempted cus-
tomers, x1 , x2 , x3 , . . . , are iid with probability distribution
tomer is placed back in the queue ahead of all cus-
B(·), where B(τ) = P(xi ≤ τ), and are also independent of
tomers of the same class. Under the “non-preemptive
the interarrival times y1 , y2 , y3 , . . . . Additional information
priority” discipline, once the service of any customer
such as whether the customers are served individually or
is started, it is allowed to be completed regardless
in groups (of fixed or random size) can also be specified if
of arrivals from higher priority classes. Moreover,
applicable.
the preemptive priority queueing systems can fur-
ther be divided into two types. Under the discipline
The Queueing Discipline. The queueing discipline is the of “preemptive resume,” whenever a preempted cus-
description of the mechanism for determining which of the tomer reenters service he simply continues his ser-
waiting customers gets served next, along with the asso- vice where he left off. Under “preemptive repeat,” a
ciated rules regarding formation of the queue. The most preempted customer draws a new value of service
basic queueing disciplines are listed and described below: time from the service time distribution each time it
reenters service.
1. First-Come First Served (FCFS) or First-In First-Out
(FIFO) The waiting customers are served in the order Maximum Number of Customers Allowed. In many sys-
of their arrival times. tems the capacity of queueing system is assumed to be infi-
2. Last-Come First-Served (LCFS) or Last-In First-Out nite, which implies that every arriving customer is allowed
(LIFO) The customer who has arrived last is chosen to join the queue and wait until served. However, in many
as the one who gets served when a server becomes real-life situations, the queueing systems have either no
available. or only a finite amount of capacity for customers to wait.
4 Network of Queues
In a queueing system with no room for customers to wait, the number of customers in the queue proper (i.e.,
whenever all the servers are busy, any additional arriving not including the one or more customers who are be-
customer is turned away; this type of system is referred to ing served), and others use it to represent the total
as “loss systems.” Loss systems have been used to model the number of customers in the system. In the former
behavior of many dial-up telephone systems and telephone case it is often referred to as the “queue length,” and
switching equipment. Queueing systems with a positive in the latter case it is often referred to as the “number
but finite waiting room have been deployed to characterize in the system.”
the performance of various computing and telecommuni- 2. The Waiting Time This performance measure is re-
cations systems where the finite waiting room models the lated to the amount of time spent by a customer in
finite amount of memory or buffer present in such real- the system. This term is used in two different ways.
world systems. Some authors use the term to refer to the total time
spent by a customer in the queueing system, which
Number of Servers. In general a queueing system can is the sum of the time spent by the customer in the
have either one, or finitely many, or an infinite number waiting line before service and the service time itself.
of servers. “Single-server systems” are the simplest ones Others define it as only the time spent in the queue
where a maximum of one user can be served at any given before the service. In the former case it is often re-
point in time. A “multiserver system” contains c servers, ferred to as the “system time,” and in the latter case
where 0 < c < ∞, and can serve up to c simultaneous cus- it is often referred to as the “queueing time.”
tomers at any given point in time. An “infinite-server sys- 3. The Busy Period This is the length of time during
tem” is one in which each arriving customer is immediately which the server is continuously busy. Any busy pe-
provided with an idle server. riod begins when a customer arrives at an empty sys-
tem, and it ends when the number of customers in
Performance Measures the system reaches zero. The time period between
two successive busy periods is referred to as the “idle
In any queueing system there are many performance
period” for obvious reasons.
tradeoffs to be considered. For example, if the number of
servers in the system is so large that queues rarely form,
then the servers are likely to be idle a large fraction of time, Kendall’s Notation for Queueing Systems
resulting in wasting of resources and extra expense. On the It is a common practice to use a short-hand notation of
other hand, if almost all customers must join long queues, the form A/B/c/K/m/Z to denote various aspects of a queue-
and servers are rarely idle, there might be customer dissat- ing system. This notation is referred to as Kendall’s nota-
isfaction and possibly lost customers which again has neg- tion. This type of short-hand was first developed by Kendall
ative economical consequences. Queueing theory provides (17) and later extended by Lee (18). It defines some of the
the designer the necessary tools to analyze the system and basic parameters which must be known about a queue in
ensure that the proper level of resources are provided in the order to study its behavior and analyze its performance.
system while avoiding excessive cost. The designer can ac- In Kendall’s notation A/B/c/K/m/Z, A describes the interar-
complish this, for example, by considering several alterna- rival time distribution, B describes the service time distri-
tive system architectures and by evaluating each by queue- bution, c is the number of (parallel) servers, K is the maxi-
ing theory methods. In addition, the future performance of mum number of customers allowed in the system (waiting
an existing system can also be predicted so that upgrading plus in service), m is the size of the customer population,
of the system can be achieved in a timely and economical and Z describes the queue discipline. The traditional sym-
fashion. For example, an analytical queueing model of a bols used in the first and second positions of Kendall’s no-
computer communication network might indicate that, in tation, and their meanings, are:
its present configuration, it cannot adequately support the
expected traffic load two years in the future. The model
M
may make it possible to evaluate different alternatives for
increased capacity such as increasing the number of nodes D
in the network, increasing the computing power of exist- Ek
ing nodes, providing more memory and buffer space in the Hk
network nodes, increasing the transmission speeds of the G
communication links, or increasing the number of commu-
nication links. Determining the most appropriate solution
Exponentially distributed interarrival time or service
can be done through careful evaluation of various perfor-
time distribution
mance measures of the queueing systems.
The following performance measures represent some of Deterministic (i.e., constant) interarrival time or service
the most common and important aspects of queueing sys- time distribution
tems which are normally investigated: k-stage Erlangian (Erlang-k) interarrival time or ser-
vice time distribution
1. The Queue Length This performance measure is re- k-stage Hyperexponential interarrival time or service
lated to the number of customers waiting in the sys- time distribution
tem. Some authors use this term to represent only General interarrival or service time distribution
Network of Queues 5
The third, fourth, and fifth positions in Kendall’s nota- parameter λ > 0 if its density function f(·) is defined by
tion could be any positive integer. The traditional symbols
used in the last position of Kendall’s notation are: FCFS,
FIFO, LCFS, LIFO, SIRO, RSS, PR, and PRI, as described
earlier in this section; and also GD, which refers to a gen-
It distribution function is given by
eral queue discipline.
As an example of Kendall notation, an
M/D/2/50/∞/SIRO queueing system is one with expo-
nential interarrival time, constant service time, 2 parallel
servers, a system capacity of 50 (i.e., a maximum of 48 in Both its mean and its standard deviation are equal to 1/λ.
the queue and 2 in service), a customer population that is The exponential distribution is unique among the con-
infinitely large, and the waiting customers are served in a tinuous distributions because it has the so-called “mem-
random order. oryless property” or “Markov property.” The memoryless
Whenever the last three elements of Kendall’s notation property is that if we know that a random variable has an
are omitted, it is meant that K = ∞, m = ∞, and Z = FCFS exponential distribution, and we know that the value of
(i.e., there is no limit to the queue size, the customer source the random variable is at least some value, say t, then the
is infinite, and the queue discipline is FCFS). As an exam- distribution for the remaining value of the variable (i.e.,
ple of the shorter version of Kendall’s notation, an M/M/1 the difference between the total value and t) has the same
queue has Poisson arrivals, exponential service time, and exponential distribution as the total value. That is,
1 server, there is no limit to the queue size, the customer
source is infinite, and the queue discipline is FCFS.
It should be noted that although Kendall’s notation is Another interpretation of Eq. (3) is that, if X is the waiting
quite useful and very popular, it is not meant to charac- time until a particular event occurs and t units of time
terize all possible models and configurations of queueing have produced no event, then the distribution of further
systems. For example, Kendall’s notation is normally not waiting time is the same as it would be if no waiting time
used to indicate bulk arrivals, or queues in series, and so had passed; that is, the system does not “remember” that t
on. time units have produced no “arrival.”
NOTIONS OF PROBABILITY THEORY OF IMPORTANCE Poisson Probability Distribution and Poisson Random
TO THE STUDY OF QUEUES Process. Poisson random variable is used in many appli-
cations where we are interested in counting the number
Probability theory has a major and fundamental role in of occurrences of an event (such as arrivals to a queueing
the study and analysis of queueing models. As mentioned system) in a certain time period or in a region of space.
earlier, queueing theory is considered a branch of applied Poisson random variables also appear naturally in many
probability theory. It is assumed here that the reader is physical situations. For example, the Poisson probability
familiar with the basic notions of elementary probability mass function gives an accurate prediction for the relative
theory such as notions of events, probability, statistical in- frequencies of the number of particles emitted by a radioac-
dependence, distribution and density functions, and expec- tive mass during a fixed time period. A discrete random
tations or averages. The reader is referred to Ref. 19 for a variable X is said to have a Poisson distribution with pa-
complete treatment of probability theory. Here we discuss rameter λ > 0 if X has a probability mass function of the
a few aspects of probability notions which are of great im- form
portance to the study of queues.
Probability Distributions of Importance to Queueing

Both the mean and the standard deviation of the Poisson
Theory
random variable are equal to λ.
As is indicative of Kendall’s notation, queueing theory Now consider a situation in which events occur at ran-
deals with a large number of different types of probability dom instants of time at an average rate of λ events per sec-
distributions to mathematically model the behavior of cus- ond. For example, an event could represent the arrival of
tomer interarrival times and the customer service times. In a customer to a service station or the breakdown of a com-
the rest of this section, we briefly describe some of the most ponent in some system. Let N(t) be the number of event
important probability distributions that are used often in occurrences in the time interval [0, t]. N(t) is then a nonde-
various queueing theory analysis. creasing, integer-valued, continuous-time random process.
Such a random process is said to be a Poisson process if the
Exponential Probability Distribution. The probability dis- number of event occurrences in the time interval [0, t] has
tribution most commonly assumed for customer interar- a Poisson distribution with mean λt. That is,
rival time and for customer service times in queueing mod-
els is the exponential distribution. This popularity is due to
its pleasant mathematical properties which often result in
much simplification of the analytical work. A continuous Like the exponential distribution, Poisson process also has
random variable X has an exponential distribution with a number of unique properties which has made it very at-
6 Network of Queues
tractive for analytical studies of queueing systems. In par-

ticular, Poisson process has a “memoryless property”; oc-
currence of events during a current interval of time is in-
dependent of occurrences of events in previous intervals. In
other words, events occurring in nonoverlapping intervals The two-stage hyperexponential distribution described
of time are independent of each other. Furthermore, the above can be generalized to k stages for any positive in-
interevent times (i.e., interarrival times in case of queue- teger greater than 2.
ing system) in a Poisson process from an iid sequence of
exponential random variables with mean 1/λ. Notions of Transient and Steady State
Analysis of a queueing system often involves the study of
Erlang-k Probability Distribution. A. K. Erlang (10) used the system’s characteristics over time. A system is defined
a special class of gamma random variables (19), now of- to be in “transient state” if its behavior and associated per-
ten called “Erlang-k” or “k-stage Erlangian,” in his study formance measures are dependent on time. This usually
of delays in telephone traffic. A random variable, T, is said occurs at the early stages of the operation of the system
to be an Erlang-k random variable with parameter λ or to where its behavior is heavily dependent on the initial state
have an Erlang distribution with parameters k and λ, if of the system. A system is said to be in “steady state” or
T is gamma random variable with the density function f “equilibrium” when the behavior of the system becomes in-
given by dependent of time. This usually occurs after the system
has been in operation for a long time, and the influence of
initial conditions and of the time since start-up have di-
minished. In steady state, the number of customers in the
system and in the queue are independent of time.
The mean and variance of Erlang-k random variable are A necessary condition for a queueing system to reach
1/λ and 1/(kλ2 ), respectively. An Erlang-k random variable steady state is that the elapsed time since the start of the
can be obtained by adding k independent exponentially dis- operation is mathematically long enough (i.e., the limit as
tributed random variables each with parameter λk. The time tends to infinity). However, this condition is not suf-
physical model that Erlang had in mind was a service fa- ficient to guarantee that a queueing system is in steady
cility consisting of k identical independent service substa- state. In addition to elapsed time, particular parameters of
tions connected in series one after another, each with an the queueing system itself will have an effect on whether
exponential distribution of service time. He wanted this and when the system reaches steady state. For example,
special facility to have the same average service time as a if the average arrival rate of customers is higher than the
single service facility whose service time was exponential overall average service rate of the system, then the queue
with parameter λ. Thus the service time, T, for the facility length will continue to grow forever and steady state will
with k stages could be written as the sum of k exponential never be reached. Although many authors have studied the
random variables, each with parameter λk. transient behavior of queueing systems, the majority of the
key results and existing literature deal with steady-state
behavior of queueing systems.
Hyperexponential Probability Distribution. If the service
time of a queueing system has a large standard deviation
Random Variables of Interest
relative to the mean value, it can often be approximated by
a hyperexponential distribution. The model representing In this section we define and list the key random variables
the simplest hyperexponential distribution is one with two and associated notations used in queueing theory and in
parallel stages in the facility; the top one having exponen- the rest of this article. Some of the primary random vari-
tial service with parameter µ1 , and the bottom stage having ables and notations are graphically illustrated in Fig. 2 and
exponential service with parameter µ2 . A customer enter- many more are listed in Table 1. Clearly, there are some ob-
ing the service facility chooses the top stage with probabil- vious relationships between some of the random variables
ity α1 or the bottom stage with probability α2 , where α1 + listed in Fig. 2 and/or Table 1. For example, with respect to
α2 = 1. After receiving service at the chosen stage, with the the number of customers in the system, we must have
service time being exponentially distributed with average
service rate µi , the customer leaves the service facility. A
new customer is not allowed to enter the facility until the and
original customer has completed service. The probability
density function for the service time, the probability dis-
tribution function, mean, and variance are given by In Eq. (12), it is assumed that the queueing system has
reached the steady state. It should, however, be noted that
although the system is in steady state, quantities N, Nq ,
and Ns are random variables; that is, they are not constant
and have probability distributions associated with them.
In other words, “steady state” means that the probabilities
are independent of time but not that the system becomes
deterministic.
Network of Queues 7
Figure 2. Graphical representation of

some variables of importance to queueing
theory.
There are similar obvious relationships between some of

the random variables related to waiting times. For exam-
ple, the total time in the queueing system for any customer
is the sum of his waiting time in the queue and his service
time, that is,
We are clearly interested in studying relationships be-

tween other random variables and parameters of the in-
terest which might not be as obvious as those given in Eqs.
(11)–(14). Development of such relationships are a major
byproduct of modeling and analysis of queueing systems,
as will be discussed in the next section.
MODELING AND ANALYSIS OF ELEMENTARY

QUEUEING SYSTEMS
In this section we present, in some detail, some of the key

techniques used by queueing theory community to model
and analyze some of the elementary queueing models. In
particular, we will illustrate the application of birth-and-
death stochastic processes to the analysis of these models.
Little’s Formula
Little’s formula (which is also known as “Little’s result”
and “Little’s theorem”) is one of the most fundamental and
often used results in queueing theory. It provides a simple,
but very general, relationship between the average wait-
ing time and the average number of customers in a queue-
Applying expectations operation to both sides of Eq. (12), ing system. Its first rigorous proof in its general form was
we get given by J. D. C. Little (20). Its validity and proofs of some
special cases, however, were known to researchers prior to
Little’s proof. Consider an arbitrary queueing system in
8 Network of Queues
steady state. Let L, W, and λ be the average number of cus- Birth-and-Death Process
tomers in the system, average time spent by customers in
Most elementary queueing models assume that the inputs
the system, and average number of customer arrivals per
(i.e., arriving customers) and outputs (i.e., departing cus-
unit time, respectively. Little’s theorem states that
tomers) of the queueing system occur according to the so-
called “birth-and-death process.” This important process
in probability theory has application in other areas also.
regardless of the interarrival and service time distribu- However, in the context of queueing theory, the term “birth”
tions, the service discipline, and any dependencies within refers to the arrival of a new customer and the term “death”
the system. refers to the departure of a served customer. The state of
Rigorous proof of Little’s theorem is given in every stan- the system at time t, for t ≥ 0, is given by random variable
dard queueing theory text (1–6). What follows is an intu- N(t) defined as the number of customers in the system at
itive justification of Little’s result given in Ref. 12. Suppose time t. Thus the birth-and-death process describes proba-
that the system receives a reward (or penalty) of 1 for ev- bilistically how N(t) changes at t increases.
ery unit of time that a customer spends in it. Then the Formally speaking, a stochastic process is a birth-and-
total expected reward per unit time is equal to the average death process if it satisfies the following three assump-
number of customers in the system, L. On the other hand, tions: (1) Given N(t) = n, the current probability distri-
the average number of customers coming into the system bution of the remaining time until the next birth is ex-
per unit time is λ; the expected reward contributed by each ponentially distributed with parameter λn for n = 0, 1, 2,
customer is equal to his average residence time, W. Since it . . . ; (2) given N(t) = n, the current probability distribution
does not matter whether the reward is collected on arrival of the remaining time until the next death is exponen-
or continuously, we must have L = λW. A different inter- tially distributed with parameter µn for n = 0, 1, 2, . . . ;
pretation of Little’s result is obtained by rewriting it as λ and (3) only one birth or death can occur at a time. Figure
= L/W. Since a customer in the system remains there for 3, which shows the state transition diagram of a birth-and-
an average time of W, his average rate of departure is 1/W. death process, graphically summarizes the three assump-
The total average departure rate is, therefore, L/W. Thus, tions just described. The arrows in this diagram show the
the relation holds if the average arrival rate is equal to the only possible transitions in the state of the system, and the
average departure rate. But the latter is clearly the case label for each arrow gives the mean rate for the transition
since the system is in equilibrium. when the system is in the state at the base of the arrow.
It is important to note that we have not even speci- Except for a few special cases, analysis of the birth-and-
fied what constitutes “the system,” nor what customers do death process is very difficult when the system is in a tran-
there. It is just a place where customers (entities) arrive, sient condition. On the other hand, it is relatively easy to
remain for some time, and then depart after having re- derive the probability distribution of the number of cus-
ceived service. The only requirement is that the processes tomers in the system in steady state. In steady state, the
involved should be stationary (i.e., system should be in probability of finding the system in a given state does not
steady state). Therefore, we can apply Little’s theorem not change with time. In particular, the probability of there
only to the entire queueing system [as represented by Eq. being more than k customers in the system is constant.
(15)], but also to particular subsections of it. For example, The transition from state k to state k + 1 increases this
applying Little’s theorem to only the waiting line portion probability, and the transition from state k + 1 to state k
of a G/G/c queueing system, where 1 ≤ c ≤ ∞, results in decreases it. Therefore, these two transitions must occur at
the same rate. If this were not so, the system would not be
in steady state. This yields to the following key principle:
In equilibrium, the average rate into any state is equal to
where Lq and Wq are as defined in Table 1. Now consider the average rate out of that state. This basic principle can
another situation, where the “system” is defined as the “set be used to generate a set of equations called the “balance
of c servers” in a G/G/c queueing system, where 1 ≤ c ≤ equations.” After constructing the balance equations for all
∞. Since every incoming customer enters a server eventu- the states in terms of the unknown probabilities Pn , this
ally, the rate of arrivals into the “set of c servers” is also system of equations can then be solved to find these prob-
λ. The average time a customer spends in the system here abilities. As shown in Fig. 3, there are only two transitions
is simply 1/µ. According to Little’s theorem, the average associated with state zero which result in the following bal-
number of customers in the system is therefore λ/µ. Thus ance equation for that state:
in any G/G/c or G/G/∞ system in steady state, the average
number of busy servers is equal to the traffic intensity, ρ.
When c = 1, the average number of busy servers is equal
to the probability that the server is busy. Therefore, in any There are four transitions associated with state 1 resulting
single-server system in the steady state we have in the following balance equation for that state:
Balance equations for states n ≥ 2 are similar to that of

state 1 and can be easily be generated by inspecting the
associated transitions in Fig. 3. This collection of balance
Network of Queues 9
Figure 3. State transition diagram for a birth-and-death process.
equations along with the auxiliary equation Therefore, by using Eqs. (22)–(24), we get
where ρ = λ/µ. The mean number of customers in the sys-

can be solved for Pn , n = 0, 1, 2, 3, . . . , resulting in the tem can now be computed as
following set of steady-state probabilities for the number
of customers in the system:
Having found the mean number of customers in the system,

we can now use Little’s formula to determine the average
total waiting time, W, as follows:
where
Behavior of the average number of customers in the sys-
tem (i.e., L) and the normalized average waiting time (i.e.,
Wµ) for the M/M/1 queue as a function of traffic intensity,
Given these expressions for the steady-state probability ρ, has been graphically shown in Fig. 4. Note that the av-
of number of customers in the system, we can derive the erage waiting time and the queue length explode as traffic
average number of customers in the system by intensity approaches 1. Therefore, the M/M/1 queue is sta-
ble only if 0 ≤ ρ < 1.
Other Elementary Queueing Systems

These steady-state results have been derived under the as- There are a number of other single-queue models whose
sumption that the λn and µn parameters are such that the steady-state behavior can be determined via birth-and-
process actually can reach a steady-state condition. This death process techniques. We briefly mention the most im-
assumption always holds if λn = 0 for some value of n, so portant ones and refer the reader to standard texts on
that only a finite number of states (those less than n) are queueing theory (1–6) for detailed analysis and the associ-
possible. It also always holds when λn = λ and µn = µ for ated mathematical expressions. Lack of space prevents us
all n and when ρ = λ/µ < 1. from listing all the associated results and formulas in these
areas. The reader is referred to Ref. 3 (pp. 400–409) for a
M/M/1 Queue tabular listing of all the key formulas related to important
Consider the simplest model of a nontrivial queueing queueing models.
model. This model assumes a Poisson arrival process (i.e.,
exponentially distributed interarrival times), an exponen- M/M/1/K. The M/M/1 model is somewhat unrealistic in
tially distributed service time, a single server, infinite the sense that, for example, no communication link can
queue capacity, infinite population of customers, and FCFS have an unlimited number of buffers. The M/M/1/K system
discipline. If the state of the system at time t, for t ≥ 0, is is a more accurate model of this type of system in which a
given by the random variable N(t), defined as the number limit of K customers is allowed in the system. When the sys-
of customers in the system at time t, it represents a birth- tem contains K customers, arriving customers are turned
and-death process with rates away. This model can easily be analyzed by truncating the
birth-and-death state diagram of the M/M/1 queue to only
K states. This results in a birth-and-death process with
coefficients
10 Network of Queues
Figure 4. Performance characteristics of M/M/1 queue.
and
Historically, the expression for the probability that all
servers are busy in an M/M/c/c queueing system is referred
to as “Erlang’s B Formula” or “Erlang’s Loss Formula” (3,
p. 404). Tables of values of Erlang’s B Formula are often
given in standard queueing texts; see, for example, Ref. 1
M/M/c. For this model we assume exponential inter- (pp. 316–319).
arrival times, exponential service times, and c identical
servers. This system can be modeled as a birth-and-death M/M/∞ Queueing System. Mathematically speaking, an
process with the coefficients M/M/∞ queueing system has an infinite number of servers
which cannot be physically realized. M/M/∞ queueing sys-
tems are used to model situations where a server is always
and immediately provided for each arriving customer. The coef-
ficients of the associated birth-and-death process are given
by
Note that Eq. (33) agrees with Eq. (26) when c = 1; that is, and
for the M/M/1 queueing system, as it should. Historically,
the expression of the probability that an arriving customer
must wait is known as “Erlang’s C Formula” or “Erlang’s Solving the birth-and-death equations for the steady-state
Delay Formula” (3, p. 404). Tables of values of Erlang’s C probability of number of customers in the queue results in
Formula are often given in standard queueing texts; see,
for example, Ref. 1 (pp. 320–323).
M/M/c/c. This system is sometimes called the “M/M/c

loss system” because customers who arrive when all the Therefore, the number of customers in an M/M/∞ queue is
servers are busy are not allowed to wait for service and distributed according to a Poisson distribution with param-
are lost. Each newly arriving customer is given his private eter λ/µ. The average number of customers in the system
server; however, if a customer arrives when all servers are is simply L = λ/µ and the average waiting time is W = 1/µ.
occupied, that customer is lost; when modeling telephone This answer is obvious since if we provide each arriving
calls, it is said that this is a system where blocked calls are customer his own server, then his time in the system is
cleared. The birth-and-death coefficients for this model are equal to his service time. M/M/∞ models can be used to es-
timate the number of lines in use in large communications
networks or as an estimate of values of M/M/c or M/M/c/c
systems for large values of c.
and M/M/1/K/K and M/M/c/K/K Queueing Systems. These

queueing systems, with a limited source model in which
there are only K customers, is usually referred to as the
Network of Queues 11
“machine repair model” or “machine interference model.” Queueing Systems with Priority
One way to interpret these models is to assume that there
Queueing models with priority are those where the queue
is a collection of K machines, each of which has an up
discipline is based on a priority mechanism where the or-
time which is exponentially distributed. The operating ma-
der in which the waiting customers are selected for service
chines are outside of the system and enter the system only
is dependent on their assigned priorities. Many real queue-
when they break down and thus need repair. The one re-
ing systems fit these priority-discipline models. Rush jobs
pairman (or c repairmen) repairs the machines at an expo-
are taken ahead of other jobs, important customers may
nential rate. The coefficients of the associated birth-and-
be given precedence over others, and data units containing
death process are
voice and video signals may be given higher priority over
data units containing no real-time information in a packet
switched computer communication network. Therefore, the
and use of queueing models with priority often provides much
needed insight into such situations. The inclusion of pri-
ority makes the mathematical analysis of models much
more complicated. There are many ways in which notions
of priority can be integrated into queueing models. The
REFERENCES TO MORE ADVANCED TOPICS most popular ones were defined earlier in this article under
queue disciplines. They include such priority disciplines as
The discussion of previous sections has been limited to non-preemptive priority, preemptive resume priority, and
some of the more elementary, but important, queueing preemptive repeat priority (21).
models. However, the queueing theory literature currently
contains a vast amount of results dealing with much more
Networks of Queues
advanced and sophisticated queueing systems whose dis-
cussions are outside of the scope of this introductory article. Many queueing systems encountered in practice are queue-
The purpose of this section is to inform the reader of the ing networks consisting of a collection of service facilities
existence of such advanced and complex models and to re- where customers are routed from one service center to an-
fer the interested reader to appropriate sources for further other, and they receive service at some or all of these ser-
investigation. vice facilities. In such systems, it is necessary to study the
entire network in order to obtain information about the
Imbedded Markov Chain Queueing Models performance of a particular queue in the network. Such
models have become very important because of their ap-
Our discussion of queueing models in the previous section plicability to modeling computer communication networks.
was limited to those whose probabilistic characterization This is a current area of great research and application in-
could be captured by birth-and-death processes. When one terest with many difficult problems. Networks of queues
ventures beyond the birth-and-death models into the more can be described as a group of nodes (say n of them) where
general Markov processes, then the type of solution meth- each node represents a service center each with ci servers,
ods used previously no longer apply. In the preceding sec- where i = 1, 2, . . . , n. In the most general case, customers
tions we dealt mainly with queues with Poisson arrivals may arrive from outside the system to any node and may
and exponential service times. These assumptions imply depart the system from any node. The customers entering
that the future evolution of the system will depend only the system traverse the network by being routed from node
on the present state of the system and not on the past his- to node after being served at each node they visit. Not all
tory. In these systems, the state of the system was always customers enter and leave from the same nodes, or take the
defined as the number of customers in the system. same path through the network. Customers may return to
Consider the situation in which we like to study a queue- nodes previously visited, skip some nodes, or choose to re-
ing system for which the knowledge of the number of cus- main in the system forever. Analytical results on queueing
tomers in the system is not sufficient to fully character- networks have been limited because of the difficulty of the
ize its behavior. For example, consider a D/M/1 queue in problem. Most of the work has been confined to cases with
which the service times are exponentially distributed, but a Poisson input and exponential service times and proba-
the customer interarrival times are a constant. Then the bilistic routing between the nodes. The reader is referred
future evolution of the system from some time t would de- to Ref. 22 for a complete treatment of network of queues.
pend not only on the number of customers in the system at
time t, but also on the elapsed time since the last customer
Simulation of Queueing Systems
arrival. This is so because the arrival epoch of the next cus-
tomer in a D/M/1 queue is fully determined by the arrival Very often, analytical solutions to many practical queueing
time of the last customer. A different and powerful method models are not possible. This is often due to many factors
for the analysis of certain queueing models, such as the one such as the complexity of the system architecture, the na-
mentioned above, is referred to as the “imbedded Markov ture of the queue discipline, and the stochastic character-
chain” which was introduced by Kendall (17). The reader istics of the input arrival streams and service times. For
is referred to Refs. 1–6 for detailed discussion of imbedded example, it would be impractical to develop analytical so-
Markov chain techniques and its application for analyzing lutions to a multinode multiserver system where the cus-
such queueing systems as M/G/1, GI/M/c, M/D/c, Ek /M/c. tomers are allowed to recycle through the system, the ser-
12 Network of Queues
vice times are distributed according to truncated Gaussian timize certain performance measures of a queueing sys-
distribution, and each node has its own complex queueing tems. Examples of practical questions that deal with this
discipline. For analytically intractable models, it may be area of queueing theory include the following (22, Chap.
necessary to resort to analysis by simulation. Another area 8): When confronted with the choice of joining one waiting
that simulation could be used for is those models in which line among many (such a supermarket checkout counter or
analytical results are only available for steady state and highway tool booths), how does one choose the “best” queue?
one needs to study the transient behavior of the system. Should a bank direct the customers to form a single wait-
Generally speaking, simulation refers to the process of ing line, or should each bank teller have his or her own
using computers to imitate the operation of various kinds queue? Should a congested link in a communication net-
of real-world systems or processes. While simulation may work be replaced with another link twice as fast, or should
offer a mechanism for studying the performance of many it be augmented with a second identical link working in
analytically intractable models, it is not without its dis- parallel with the first one?
advantages. For example, since simulation can be consid-
ered analysis by experimentation, one has all the usual
BIBLIOGRAPHY
problems associated with running experiments in order to
make inferences concerning the real world, and one must
1. R. B. Cooper Introduction to Queueing Theory, 2nd ed., New
be concerned with such things as run length, number of York: Elsevier/North-Holland, 1981.
replications, and statistical significance. Although simula-
2. D. Gross C. H. Harris Fundamentals of Queueing Theory, New
tion can be a powerful tool, it is neither cheap nor easy to York: Wiley, 1985.
apply correctly and efficiently. In practice, there seems to be
3. L. Kleinrock Queueing Systems, Volume I: Theory, New York:
a strong tendency to resort to simulation from the outset. Wiley-Interscience, 1975.
The basic concept is easy to understand, it is relatively easy
4. L. Kleinrock Queueing Systems, Volume II: Computer Applica-
to justify to management, and many powerful simulation tions, New York: Wiley-Interscience, 1976.
tools are readily available. However, an inexperienced an-
5. E. Gelenbe G. Pujolle Introduction to Queueing Networks,
alyst will usually seriously underestimate the cost of many Paris: Wiley, 1987.
resources required for an accurate and efficient simulation
6. T. L. Saaty Elements of Queueing Theory, New York: McGraw-
study. Hill, 1961.
Viewing it from a high level, a simulation model
7. R. B. Cooper Queueing theory. In D. P. Heyman and M. J. Sobel
program consists of three phases. The data generation (eds.), Stochastic Models, Handbooks of Operations Research
phase involves the production of representative interar- and Management Science, Vol. 2, New York: North-Holland,
rival times and service times where needed throughout the 1990.
queueing system. This is normally achieved by employing 8. N. U. Prabhu A bibliography of books and survey papers on
one of the many random number generation schemes. The queueing systems: theory and applications, Queueing Systems,
so-called bookkeeping phase of a simulation program deals 1: 1–4, 1987.
with (a) keeping track of and updating the state of the sys- 9. M. A. Leibowitz Queues, Sci. Am., 219 (2): 96–103, 1968.
tem whenever a new event (such as arrival or departure) 10. A. K. Erlang The theory of probabilities and telephone conver-
occurs and (b) monitoring and recording quantities of inter- sations, Nyt Tidsskrift Matematik, Series B, 20: 33–39, 1909.
est such as various performance measures. The final phase
of a simulation study is normally the analysis of the output 11. F. S. Hillier G. J. Lieberman Introduction to Operations Re-
of the simulation run via appropriate statistical methods. search, 4th ed., Oakland, CA: Holden-Day, 1986.
The reader is referred to Refs. 23 and 24 for a comprehen- 12. H. A. Taha Operations Research: An Introduction, New York:
sive look at simulation techniques. Macmillan, 1971.
13. E. Gelenbe I. Mitrani Analysis and Synthesis of Computer Sys-
Cyclic Queueing Models tems, New York: Academic Press, 1980.
This area deals with situations where a collection of queues 14. J. F. Hayes Modeling and Analysis of Computer Communica-
tions Networks, New York: Plenum Press, 1984.
is served by a single server. The server visits each queue
according to some predetermined (or random) order and 15. C. H. Sauer K. M. Chandy Computer Systems Performance
Modeling, Englewood Cliffs, NJ: Prentice-Hall, 1981.
serves each queue visited for a certain amount of time (or
certain number of customers) before traversing to the next 16. J. N. Daigle Queueing Theory for Telecommunications, Read-
ing, MA: Addison-Wesley, 1992.
queue. Other terms used to refer to this area of queueing
theory are “round-robin queueing” or “queueing with va- 17. D. G. Kendall Stochastic processes occurring in the theory of
queues and their analysis by the method of imbedded markov
cations.” As an example, a time-shared computer system
chains, Ann. Math. Stat., 24: 338–354, 1953.
where the users access the central processing unit through
18. A. M. Lee Applied Queueing Theory, London: Macmillan, 1966.
terminals can be modeled as a cyclic queue. The reader is
referred to Ref. 25 and Section 5.13 of Ref. 1 for detailed
19. A. Leon-Garcia Probability and Random Processes for Electri-
discussion of cyclic queues.
cal Engineering, Reading, MA: Addison-Wesley, 1989.
20. J. D. C. Little A proof for the queueing formula L = λ W, Oper.
Control of Queues
Res., 9: 383–387, 1961.
This area of queueing theory deals with optimization tech- 21. N. K. Jaiswell Priority Queues, New York: Academic Press,
niques used to control the stochastic behavior and to op- 1968.
Network of Queues 13
22. J. Walrand An Introduction to Queueing Networks, Englewood

Cliffs, NJ: Prentice-Hall, 1988.
23. P. Bratley B. L. Fox L. E. Schrage Guide to Simulation, New
York: Springer-Verlag, 1983.
24. A. M. Law W. D. Kelton Simulation Modeling and Analysis,
2nd ed., New York: McGraw-Hill, 1991.
25. H. Takagi Analysis of Polling Systems, Cambridge, MA: MIT
Press, 1986.
26. “Queueing Theory,” http://en.wikipedia.org/wiki/Queuing theory.
27. Myron Hlynka,“What is the proper spelling — queueing or

queuing?,” http://www2.uwindsor.ca/ hlynka/qfaq.html.
28. Jeff Miller,“A Collection of Word Oddities and Trivia,”
http://members.aol.com/gulfhigh2/words6.html.
29. Myron Klynka,“Myron Hlynka’s Queueing Theory Page,”
http://www2.uwindsor.ca/ hlynka/queue.html.
30. John N. Daigle,“ Queueing Theory with Application to Packet
Telecommunications,” Springer, 2004.
31. N. U. Prabhu,“ Foundations of Queueing Theory,” Springer,
1997.
NADER MEHRAVARI
Lockheed Martin, Owego, NY
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Signal Detection Theory ● Email this page Enter words or
● Print this page
Brian L. Hughes1
1North Carolina State University, Raleigh, NC
reserved. Search
Abstract
Basic Principles
Detection of Known Signals in Noise
Detection of Signals with Unknown Parameters
Detection of Random Signals
Advanced Topics
Further Reading

SIGNAL DETECTION THEORY 269
SIGNAL DETECTION THEORY

In remote sensing and communications, we are often required
to decide whether a particular signal is present—or to distin-
guish among several possible signals—on the basis of noisy
observations. For example, a radar transmits a known elec-
tromagnetic signal pulse into space and detects the presence
of targets (e.g., airplanes or missiles) by the echoes which
they reflect. In digital communications, a transmitter sends
data symbols to a distant receiver by representing each sym-
bol by a distinct signal. In automatic speech recognition, an
electrical microphone signal is processed in order to extract a
270 SIGNAL DETECTION THEORY
sequence of phonemes, the elementary sounds that make up scribes an observation Y. For simplicity, consider the problem
spoken language. Similar problems arise in sonar, image proof deciding between two models, ‘‘target present’’ or ‘‘target
cessing, medical signal processing, automatic target recogni- absent.’’ Suppose the probability density function (pdf) of Y is
tion, radio astronomy, seismology, and many other applica- given by p1(y) when the target is present and by p0(y) when
tions. the target is absent. The problem of deciding whether Y is
In each of these applications, the signal received is typi- best modeled by p0(y) or p1(y) can be expressed as a choice
cally distorted and corrupted by spurious interference, which between two hypotheses:
complicates the task of deciding whether the desired signal is
present. In particular, the received signal in radar and com- H0 : Y has pdf p0 (y) (target absent)
(1)
munication systems inevitably contains random fluctuations, H1 : Y has pdf p1 (y) (target present)
or noise, due to the thermal motion of electrons in the receiver
and ions in the atmosphere, atmospheric disturbances, elec- where H0 is often called the null hypothesis and H1 is the al-
tromagnetic clutter, and other sources. This noise component ternative hypothesis. A detector (or decision rule or hypothesis
is inherently unpredictable; the best we can do is to describe test) is a procedure for deciding which hypothesis is true on
it statistically in terms of probability distributions. In some the basis of the observation Y. More precisely, a detector is a
situations, the desired signal may also be distorted in unpre- function that assigns to each possible observation Y ⫽ y a
dictable ways—by unknown time delays, constructive and de- decision d(y) ⫽ H0 or H1. There are two possible ways for the
structive interference of multiple signal reflections, and other detector to make an error: It may conclude that a target is
channel impairments—and may be best modeled as a ran- present when there is none (a false alarm), or it may decide
dom process. that no target is present when in fact there is one (a miss).
The task of the receiver is to decide whether the desired The performance of a detector d can therefore be measured
signal is present in an observation corrupted by random noise by two quantities, the probability of false alarm PF(d) and the
and other distortions. The mathematical framework for deal- probability of a miss PM(d). Ideally, we would like to make
ing with such problems comes from the field of hypothesis test- both error measures as small as possible; however, these are
ing, a branch of the theory of statistical inference. In engi- usually conflicting objectives in the sense that reducing one
neering circles, this is also called detection theory because of often increases the other. In order to determine which detec-
its early application to radar problems. Detection theory pro- tor is best for a particular application, we must strike a bal-
vides the basis for the design of receivers in communication ance between PF(d) and PM(d) which reflects the relative im-
and radar applications, algorithms for identifying edges and portance of these two types of errors.
other features in images, algorithms for parsing an electrical Several methods can be used to weigh the relative impor-
speech signal into words, and many other applications. tance of PF(d) and PM(d). If the prior probabilities of the
In addition to deciding whether a signal is present, we of- hypotheses are known, say 앟 ⫽ Pr兵H0其 ⫽ 1 ⫺ Pr兵H1其, it is
ten want to estimate real-valued parameters associated with natural to seek a minimum-probability-of-error detector—that
the signal, such as amplitude, frequency, phase, or relative is, one that minimizes the average error probability:
time delay. For example, once a target has been detected, a
radar will typically attempt to determine its range by esti- πPF (d) + (1 − π )PM (d)
mating the round-trip propagation delay of the pulse echo.
Problems of this type are the province of estimation theory, a Such detectors are appropriate in digital communication re-
field of statistical inference closely related to detection theory. ceivers where the hypotheses represent the possible transmit-
In essence, detection theory deals with the problem of decid- ted data symbols and where the goal is to minimize the aver-
ing among a finite number of alternatives, whereas estima- age number of errors that occur in a series of transmissions.
tion theory seeks to approximate real-valued signal param- More generally, when the two kinds of errors are not equally
eters. serious, we can assign a cost Cij to choosing hypothesis Hi
This article provides an overview of the basic principles when Hj is actually true (i, j ⫽ 0, 1). A detector that mini-
and selected results of signal detection theory. The subject mizes the average cost (or risk) is called a Bayes detector. It
of estimation theory is treated elsewhere in this volume (see sometimes happens that the prior probabilities of H0 and H1
ESTIMATION THEORY). A more complete and detailed treatment are not known, in which case the Bayes and minimum-proba-
of both topics can be found in Refs. 1 and 2. In the next sec- bility-of-error detectors cannot be applied. In this case, it of-
tion, we introduce some fundamental concepts that underlie ten makes sense to choose a detector that minimizes the aver-
the design of optimal detection procedures. In succeeding sec- age cost for the worst prior probability—for example, one that
tions, we apply these concepts to the problem of detecting sig- minimizes
nals in additive Gaussian noise. Finally, we close with a dis-
max πPF (d) + (1 − π )PM (d) = max{PF (d), PM (d)}
cussion of selected advanced topics in detection theory. 0≤π ≤1
The resulting detector is called a minimax detector. Finally,

BASIC PRINCIPLES
in other circumstances it may be difficult to assign costs or
prior probabilities. In radar, for example, what is the prior
Simple Hypotheses
probability of an incoming missile, and what numerical cost
The design of optimal receivers in radar and communications is incurred by failing to detect it? In situations like this, it
is based on principles from the theory of statistical hypothesis seems inappropriate to weigh the relative importance of false
testing. The fundamental problem of hypothesis testing is to alarms and misses in terms of numerical costs. An alternative
decide which of several possible statistical models best de- approach is to seek a detector that makes the probability of
a miss as small as possible for a given probability of false minimizes PM(d兩␪) for each ␪ over all detectors with a given
alarm: false-alarm probability, max␪ PF(d兩␪) ⱕ 움. A detector with this
property is said to be uniformly most powerful. When a uni-
minimize PM (d) subject to PF (d) ≤ α formly most powerful detector does not exist, it is natural to
use an estimate ␪ˆ of the unknown parameter derived from the
A detector of this type is called a Neyman–Pearson detector. observation Y ⫽ y. The most commonly used estimates are
Remarkably, all of the optimal detectors mentioned above maximum likelihood estimates, which are defined as the value
take the same general form. Each involves computing a likeli- of ␪ that maximizes the conditional probability density of the
hood ratio from the received observation observation:
p1 (y) pi (y|θ̂i ) = max pi (y|θ ), i = 0, 1

(y) = (2) θ
p0 (y)
Substituting the maximum likelihood estimates ␪ˆ0 and ␪ˆ1 into
and comparing it with a threshold ␶. When Y is observed, the
the likelihood ratio, we obtain the generalized likelihood ratio
detectors choose H1 if ⌳(Y) ⬎ ␶ and choose H0 if ⌳(Y) ⬍ ␶. This
(GLR):
detector can be expressed concisely as
maxθ p1 (y|θ )
H1 G (y) =
(Y ) ? τ (3) maxθ p0 (y|θ )
H0
Detectors based on the GLR take the same form as the likeli-
When ⌳(Y) ⫽ ␶, minimax and Neyman–Pearson detectors hood ratio detector [Eq. (3)], with ⌳G(y) substituted for ⌳(y).
may involve a random decision, such as choosing H0 or H1
based on the toss of a biased coin. The minimum-probability- Multiple Hypotheses and Observations
of-error, Bayes, minimax, and Neyman–Pearson detectors Each of the detectors described above extends in a straightfor-
mentioned earlier differ only in their choice of threshold ␶ and ward way to a sequence of observations Y ⫽ (Y1, . . ., Yn). In
behavior on the boundary ⌳(Y) ⫽ ␶. this case, the hypotheses become
Composite Hypotheses H0 : Y has pdf p0 (y1 , . . ., yn ) (target absent)

Thus far we have assumed that the probability distribution H1 : Y has pdf p1 (y1 , . . ., yn ) (target present)
of Y is known perfectly under both hypotheses. It is very com-
mon, however, for a signal to depend on parameters that are Again, the minimum-probability-of-error, Bayes, minimax,
not known precisely at the detector. In radar, for example, and Neyman–Pearson detectors are of the form shown in Eq.
since the distance to the target is not known at the outset, the (3), where the likelihood ratio is
radar pulse will experience an unknown propagation delay as
it travels to the target and back. In digital communications, p1 (y1 , . . ., yn )
(y1 , . . ., yn ) =
the phase of the carrier signal is often unknown to the re- p0 (y1 , . . ., yn )
ceiver. In such situations, the hypothesis ‘‘target present’’ cor-
responds to a collection of possible probability distributions, The generalized likelihood ratio detector also extends in an
rather than one. A hypothesis of this type is called a compos- analogous way.
ite hypothesis, in contrast to a simple hypothesis in which Y is We have so far considered only detection problems involv-
described by a single pdf. ing two hypotheses. In some situations there may be more
Let ␪ denote an unknown parameter associated with the than two possible models for the observed data. For example,
observation, and let p0(y兩␪) and p1(y兩␪) denote the conditional digital communication systems often use nonbinary signaling
probability densities of Y given ␪ under H0 and H1, respec- techniques in which one of M possible symbols is transmitted
tively. In some cases, it may be appropriate to model ␪ as a to the receiver in each unit of time. The receiver then has M
random variable with known probability densities q0(␪) and hypotheses from which to choose, one corresponding to each
q1(␪) under hypothesis H0 and H1, respectively. In such cases, possible transmitted symbol. The hypothesis-testing problem
the composite hypothesis testing problem is equivalent to a can then be expressed as
simple hypothesis testing problem with probability densities
Hi : Y has pdf pi (y), i = 0, . . ., M − 1

p0 (y) = p0 (y|θ )q0 (θ ) dθ, p1 (y) = p1 (y|θ )q1 (θ ) dθ (4) In such situations, we are usually interested in finding a min-
θ θ
imum-probability-of-error detector for some given prior proba-
and the optimal detectors are again of the form shown in Eq. bilities 앟i ⫽ Pr兵Hi其, i ⫽ 0, . . ., M ⫺ 1. The average probability
(3). If ␪ is a random variable with unknown probability densi- of error for a detector d is given by
ties under H0 and H1, we can follow a minimax-type approach

M−1
and look for the detector that minimizes the worst-case aver- Pr{d(Y ) = Hi |Hi is true}πi (5)
age cost over all probability densities q0(␪) and q1(␪). i=0
When ␪ cannot be modeled as a random variable, the situa-
tion is more complex. Occasionally, there exists a detector This error probability is minimized by the maximum a posteri-
that is simultaneously optimal for all ␪, in the sense that it ori probability (MAP) detector, which chooses the hypothesis
that is most probable given the observation Y ⫽ y. Mathemat- It is easy to verify that ⌳(y) is a monotonically increasing
ically, the MAP detector takes the form function of the test statistic
d(y) = Hi that maximizes q(Hi |y) (6)

n
1 n
yi s i − s2 (9)
i=1
2 i=1 i
where
pi (y)πi Thus, the likelihood ratio detector [Eq. (3)] can be expressed
q(Hi |y) = , p(y) = pi (y)πi in the equivalent form
p(y) i

n H1
is the conditional probability of hypothesis Hi given the obser- Yi Si ? τ (10)
vation Y ⫽ y. In digital communications, the possible trans- i=1 H0
mitted symbols are often equally likely (앟i ⫽ 1/M, i ⫽ 1, . . .,
M ⫺ 1), in which case the MAP detector reduces to the maxi- where the quadratic term in Eq. (9) has been merged with the
mum likelihood (ML) detector threshold ␶⬘. The optimal receiver thus consists of correlating
the received sequence against the desired signal and compar-
d(y) = Hi , where i maximizes pi (y) (7) ing the result to a threshold. A receiver of this type is called
a correlation receiver.
It is easy to check that the MAP and ML detectors reduce to This receiver extends in a natural way to continuous-time
likelihood ratio detectors when M ⫽ 2. detection. The correlation receiver extends in a natural way
The results presented in this section form the basis for the to continuous-time detection problems. A proof of this exten-
design of optimal receivers for a wide variety of communica- sion is nontrivial and requires generalizing the likelihood ra-
tions and remote sensing problems. In the following sections, tio to continuous-time observations (see Chapter 6 of Ref. 1
we apply these results to the problem of detecting signals in for details). Consider the signal detection problem
noise. In the process, we obtain several of the most important
and widely used receivers in communications and radar as H0 : Y (t) = N(t), 0≤t <T
particular instances of the likelihood ratio detector [Eq. (3)].
H1 : Y (t) = s(t) + N(t), 0≤t <T
DETECTION OF KNOWN SIGNALS IN NOISE where s(t) is a known deterministic signal and N(t) is a con-
tinuous-time white Gaussian noise process with two-sided
We now consider the problem of detecting the presence or ab- power spectral density N0 /2 (see KALMAN FILTERS AND OBSERV-
sence of a discrete-time signal observed in noise. A detector ERS). The likelihood ratio is again a monotonically increasing
for this purpose is also called a receiver in the terminology of function of a correlation statistic
radar and communications. We assume for now that both the
signal and the noise statistics are known precisely at the re- T T
1
ceiver, in which case the detection problem can be expressed y(t)s(t) dt − s2 (t) dt (11)
0 2 0
as a choice between the simple hypotheses:
H0 : Yi = Ni , i = 1, . . ., n Merging the second term with the threshold, we again find

that the likelihood ratio detector is a correlation receiver,
H1 : Yi = si + Ni , i = 1, . . ., n
which is illustrated in Fig. 1.
The correlation in Fig. 1 can also be expressed as a filter-
where si, i ⫽ 1, . . ., n is a deterministic signal and Ni, i ⫽ 1,
ing operation:
. . ., n is a white Gaussian noise sequence—that is, a se-
quence of independent and identically distributed (i.i.d.) T ∞
Gaussian random variables with mean zero and variance y(t)s(t) dt = h(T − t)y(t) dt
␴2 ⬎ 0. Thus, the probability densities of Y ⫽ (Y1, . . ., Yn) 0 −∞
under both hypotheses are multivariate Gaussian where

where h(t) ⫽ s(T ⫺ t), 0 ⱕ t ⱕ T. Here h(t) can be regarded
1 1 n as the impulse response of a linear time-invariant filter. The
p1 (yy ) = exp − (y − s ) 2
(8) frequency response of this filter is given by the Fourier trans-
(2πσ 2 )n/2 2σ 2 i=1 i i
form of h(t):
and p0(y) is given by the same formula, with the si’s set to
H( f ) = S∗ ( f )e−2π j f T
zero.
From the previous section, we know that each of the opti-
mal detectors (minimum probability of error, Bayes, minimax,
Neyman–Pearson) reduces to a likelihood ratio detector [Eq. T
(3)]. From Eq. (8), the likelihood ratio for this problem takes Y(t) ∫ (·) dt Threshold
the form 0

1 n
s(t)
(y ) = exp − 2
y [(yi − si ) − yi ]
2 2
2σ i=1
Figure 1. Correlation receiver.
is often unrealistic in practice. Unknown path losses, Doppler

Y(t) H(f) Threshold shifts, and propagation delays can lead to uncertainty about
Sampling the amplitude, phase, frequency, and delay of the signal.
at t = T
When signal parameters are unknown, the detection problem
Figure 2. Matched filter receiver. involves composite hypotheses. As discussed earlier, detection
procedures for composite-hypothesis testing depend on
whether the unknown parameter is modeled as random or
where S*( f) is the complex conjugate of the Fourier transform deterministic. In this section, we consider only the example of
of s(t). The correlation receiver can therefore be implemented an unknown random parameter.
in the form of a filter sampled at time t ⫽ T, as illustrated in Many radar and communication problems involve detec-
Fig. 2. tion of a sinusoidal signal with an unknown phase. The phase
Since the amplitude of the filter H( f) matches the signal is typically modeled as a random variable ⌰, uniformly dis-
spectrum S( f), this form of the detector is called a matched- tributed on [0, 2앟). For example, consider the discrete-time
filter receiver. The matched filter has the property that it max- binary detection problem:
imizes the signal-to-noise ratio (the ratio of signal power to
noise power) at the input to the threshold operation (see Ref. H0 : Yi = Ni , i = 1, · · ·, n
2). H1 : Yi = A cos(ωiTs + ) + Ni , i = 1, . . ., n
The receiver in Fig. 1 is optimal for deciding whether a
known signal is present or absent in white Gaussian noise. where A is a known constant, Ts is the sampling interval, 웆
Very similar correlation structures appear in receivers for is a frequency such that n웆Ts is an integer multiple of 2앟,
deciding among several possible signals. For a detection prob- and Ni is a discrete-time white Gaussian noise sequence
lem involving M possible signals, the minimum-probability- which is independent of ⌰. The likelihood ratio for this detec-
of-error detector will compare the outputs of a bank of M tion problem is given by Eqs. (3) and (4). Given ⌰ ⫽ ␪, the
correlation branches, one for each possible signal. For exam- conditional probability density of Y under H1 is
ple, consider the problem of deciding between two equally

likely (앟0 ⫽ 앟1 ⫽ ) signals in white Gaussian noise: 1 1 n

p1 (yy|θ ) = exp − 2 [y − A cos(ωiTs + θ )] 2
H0 : Y (t) = s0 (t) + N(t), 0≤t <T (2πσ 2 )n/2 2σ i=1 i
H1 : Y (t) = s1 (t) + N(t), 0≤t <T

and the unconditional pdf is given by
The receiver that minimizes the average probability of error 2π
1
[Eq. (5)] in this case is the maximum likelihood detector. p1 (yy ) = p1 (yy|θ ) dθ
When Y(t) ⫽ y(t) is received, the ML detector chooses the hy- 2π 0
pothesis Hi such that s(t) ⫽ si(t) maximizes the correlation

statistic [Eq. (11)]. Thus, the optimal receiver consists of a After some manipulation, the likelihood ratio reduces to
correlation receiver with a branch for each possible transmit-
p1 (yy ) A2 n Aq
ted signal, as illustrated in Fig. 3 [where Ei is the energy of (yy ) = = exp − 2 I0
si(t)]. p0 (yy ) 4σ σ2
As in the case of one signal, the correlation receiver in Fig.
3 can be implemented in the alternative form of a bank of where
matched filters, each sampled at t ⫽ T. 2 2
n n
q =
2
yi cos(ωiTs ) + yi sin(ωiTs ) (12)
DETECTION OF SIGNALS WITH UNKNOWN PARAMETERS i=1 i=1
In the preceding section, we assumed that the desired signal and

is known precisely at the receiver. However, this assumption 2π
1
I0 (x) = exp{x cos θ} dθ
2π 0
T
∫ (·) dt is a modified Bessel function of the first kind. Since I0(x) is
0 symmetric in x and monotonically increasing for x ⱖ 0, the
likelihood ratio is an increasing function of the quadrature
s0(t) –E0/2 statistic [Eq. (12)], and the likelihood ratio detector [Eq. (3)]
Y(t) Choose can be expressed in the alternate form:
largest
H1
s1(t) –E1/2
q2 ? τ
H0
T
∫ (·) dt
This detector is called a quadrature receiver. It consists of cor-
0
relating the received signal with two phase-shifted versions
Figure 3. Correlation receiver for binary signals. of the desired signal, cos(웆iTs) and sin(웆iTs). The two correla-
T
where S ⫽ (S1, . . ., Sn) is a zero-mean Gaussian random se-
∫ (·) dt (·)2 quence with known covariance E兵SST其 ⫽ ⌺, Ni is discrete-time
0 white Gaussian noise, E兵 ⭈ 其 denotes the expectation, and T
denotes transpose. Note that Y is a zero-mean Gaussian vec-
cos(ω t) tor under hypotheses H0 and H1, with respective covariances
Compare ␴ 2I and ␴ 2I ⫹ ⌺. The likelihood ratio is then
Y(t) to
threshold
sin(ω t) 1 1
(yy ) = |II + σ −2 |−1/2 exp − y T (σ 2I + )−1y + y T (σ 2I )−1 y
2 2
T
∫ (·) dt (·)2
Since this is a monotonically increasing function of the test
0
statistic yTQy, where
Figure 4. Quadrature receiver.
Q = I − (II + σ −2 )−1 = (σ 2I + )−1 (13)
tions are then squared, summed, and compared to a the likelihood ratio detector [Eq. (3)] can be expressed as
threshold. H1
This detector extends in a straightforward way to the de- Y T QY
Y ? τ
tection of continuous-time sinusoidal signals with random H0
phase. Consider the detection problem
This detector is called a quadratic receiver. In the particular
H0 : Y (t) = N(t), 0≤t <T
case when the desired signal is also a white noise process (i.e.,
H1 : Y (t) = A cos(ωt + ) + N(t), 0≤t<T ⌺ ⫽ 움2I), the quadratic receiver statistic yTQy is proportional
to 储 y储2 and the likelihood ratio detector reduces to
where ⌰ is a random phase uniformly distributed on [0, 2앟),
A is a constant, 웆 is an integer multiple of 2앟/T, and N(t) is H1
white Gaussian noise. The likelihood ratio detector reduces to Y

Y 2 ? τ
H0
a threshold test involving the quadrature statistic:
2 2 Since 储 y储2 is proportional to the average energy in the se-
T T
y(t) cos(ωt) dt + y(t) sin(ωt) dt quence y, this detector is called an energy detector or radi-
0 0 ometer.
In continuous time, the likelihood ratio detector takes a
The resulting continuous-time quadrature receiver is illus- more complex but analogous form. Consider the problem of
trated in Fig. 4. deciding among the hypotheses
H0 : Y (t) = N(t), 0≤t <T

DETECTION OF RANDOM SIGNALS
H1 : Y (t) = S(t) + N(t), 0≤t<T
So far we have assumed the receiver knows the desired signal
exactly, with the possible exception of specific parameters where S(t) is a zero-mean Gaussian noise process with autoco-
such as amplitude, phase, or frequency. However, sometimes variance function
the received signal may be so distorted by the channel that it
must be modeled by a more complex type of random process. C(t, u) = E{S(t)S(u)}
In certain situations, for example, the transmitted signal
propagates to the receiver by many different paths due to sig- and N(t) is a white Gaussian noise process with one-sided
nal reflection and scattering. In such cases, the received sig- power spectral density N0 /2. The likelihood ratio detector for
nal consists of many weak replicas of the original signal, this problem can also be expressed in terms of a quadratic
called multipath signals, with different amplitudes and rela- statistic
tive time delays. The superposition of these multipath signals
T T
can resemble a Gaussian random process statistically. Typical Q(t, u)Y (t)Y (u) dt du
examples include channels that use ionospheric reflection or 0 0
tropospheric scattering as a primary mode of propagation,
and land mobile radio, where scattering and reflection by where Q(t, u) is the solution to the integral equation
nearby ground structures can produce a similar effect.
In this section, we consider the problem of detecting sig- T
N0
nals that are described by random processes. We again begin C(t, u) = Q(t, ξ )C(ξ , u) dξ + Q(t, u), 0 ≤ t, u < T
0 2
by considering a discrete-time detection problem:
H0 : Yi = Ni , i = 1, . . ., n This equation is a continuous-time analog of Eq. (13), as can

be seen by writing Eq. (13) in the alternative form ⌺ ⫽ Q⌺ ⫹
H1 : Yi = Si + Ni , i = 1, . . ., n ␴ 2Q.
ADVANCED TOPICS While the receivers presented in this article can be used in
the presence of non-Gaussian noise, they are not optimal for
Detection in Colored Gaussian Noise this purpose and may perform poorly in comparison to the
likelihood ratio detector [Eq. (3)] based on the actual non-
In the preceding sections, we assumed white Gaussian noise
Gaussian noise statistics. In contrast to the simple linear cor-
models for both the discrete and continuous-time detection
relation operations that arise in Gaussian detection problems,
problems. When the noise is Gaussian but not white, we can
optimal detectors for non-Gaussian noise often involve more
transform the detection problem into an equivalent problem
complicated nonlinear operations. A thorough treatment of
involving white noise. For example, suppose we are interested
detection methods for i.i.d. non-Gaussian noise can be found
in detecting a known signal s ⫽ (s1, . . ., sn),
Kassam (3). A recent survey of detection techniques for de-
H0 : Y = N pendent non-Gaussian noise sequences is given in Poor and
Thomas (4).
H1 : Y = s + N
Nonparametric Detection
where N is a zero-mean Gaussian noise vector with positive-
definite covariance matrix ⌺N. Using the Cholesky decomposi- Throughout this article, we have assumed the receiver knows
tion (see p. 84 of Ref. 1), the noise covariance can be written the probability density of the observation under each hypoth-
in the form esis, with the possible exception of a real-valued parameter ␪.
Under this assumption, the detection problem is a choice be-
N = CCT tween composite hypotheses that each represents a collection
of possible densities, say
where C is an n ⫻ n nonsingular lower-triangular matrix. 0 = {p0 (y|θ ) : θ ∈ 0 }, 1 = {p1 (y|θ ) : θ ∈ 1 }
Since C is invertible, it is intuitive that no information is lost
by taking the observation to be Y⬘ ⫽ C⫺1Y instead of Y. The This is called a parametric model, because the set of possible
detection problem can then be expressed as probability distributions under both hypotheses can be in-
dexed by a finite number of real parameters.
H0 : Y = N In practice, however, precise models for the signal and the
H1 : Y = s + N underlying noise statistics are frequently not available. In
such cases, it is desirable to find detectors that perform well
where s⬘ ⫽ C⫺1s and N⬘ ⫽ C⫺1N. It is easy to verify that N⬘ is for a large class of possible probability densities. When the
a white Gaussian noise vector with covariance ⌺Nⴕ ⫽ I; thus, probability classes ⍀0 and ⍀1 are so broad that a parametric
the likelihood ratio detector is the correlation receiver [Eq. model cannot describe them, the model is said to be nonpara-
(10)]. Here, the overall approach is to prewhiten the original metric. In general, nonparametric detection methods may be
detection problem, by transforming it to an equivalent prob- classified as robust or simply nonparametric depending on the
lem involving white noise. After prewhitening, the detection breadth of the underlying probability classes ⍀0 and ⍀1.
problem can be solved by the methods described in the previ- In robust detection, the probability densities of the observa-
ous sections. tion are known approximately under each hypothesis and the
A similar prewhitening procedure can be performed for aim is to design detectors that perform well for small devia-
continuous-time detection problems. Let N(t) be a zero-mean tions from these densities. Usually, the probability classes
colored Gaussian noise process with known autocovariance ⍀0 and ⍀1 consist of small nonparametric neighborhoods of
function R(t, u) ⫽ E兵N(t)N(u)其. Under mild conditions on R(t, the nominal probability densities. One widely studied model
u), there is a whitening filter h(t, u) with the property that for these neighborhoods is the ⑀-contamination class
T i = {p(y) : p(y) = (1− ∈)pi (y)+ ∈ h(y)}, i = 0, 1

N (t) = h(t, u)N(u) du, 0≤t <T
0 where pi(y) is the nominal probability density under hypothe-
sis Hi, 0 ⱕ ⑀ ⬍ 1 is small enough so that ⍀0 and ⍀1 do not
is a white Gaussian noise process with unit power spectral overlap, and h(y) is an arbitrary probability density. In robust
density. This filter can be used to transform a detection prob- detection, the performance of a detector d is typically mea-
lem involving N(t) into an equivalent problem involving white sured by worst-case performance over all probability densities
Gaussian noise. in ⍀0 and ⍀1. Optimal detectors are those that yield the best
worst-case performance. For ⑀-contamination models, the op-
Detection in Non-Gaussian Noise timal robust detector consists of a likelihood ratio detector for
the nominal probability densities that includes some type of
We have thus far focused exclusively on Gaussian noise mod-
soft-limiting operation. For example, a robust form of the cor-
els. Gaussian processes can accurately model many important
relation receiver (appropriate for small deviations from the
noise sources encountered in practice, such as electrical noise
Gaussian noise model) is obtained by replacing Yisi with
due to thermal agitation of electrons in the receiver electron-
g(Yisi) in Eq. (10), where g is a soft-limiter of the form
ics, radio emissions from the motion of ions in the atmo-

sphere, and cosmic background radiation. However, other 
b if x > b
sources of noise are not well described by Gaussian distribu-
g(x) = x if a < x < b
tions, such as impulsive noise due to atmospheric distur- 

bances or radar clutter. a if x < a
276 SIGNALING
An extensive survey of the robust detection literature prior to and by Srinath, Rajasekaran, and Viswanathan (2). Further
1985 can be found in Kassam and Poor (5). information on the applications of detection theory in commu-
The term nonparametric detection is usually reserved for nications and radar is contained in the books by Proakis (7)
situations in which very little is known about the probability and by Nathanson (8).
distribution of the underlying noise, except perhaps that it is Current research in signal detection and its applications is
symmetric and possesses a probability density. In such situa- published in a wide variety of journals. Perhaps chief among
tions, the aim is to develop detectors that provide a guaran- these are the IEEE Transactions on Information Theory, IEEE
teed false-alarm probability over very wide classes of noise Transactions on Signal Processing, and the IEEE Transactions
distributions. The simplest nonparametric detector is the sign on Communications.
detector, which counts the number of positive observations in
a sequence and compares it to a threshold. It can be shown
BIBLIOGRAPHY
that this detector provides a constant false-alarm probability
for detecting the presence or absence of a constant positive
1. H. V. Poor, An Introduction to Signal Detection and Estimation,
signal in any i.i.d. zero-median additive noise sequence. A dis-
New York: Springer-Verlag, 1988.
cussion of further results in nonparametric detection may be
2. S. Srinath, P. K. Rajasekaran, and R. Viswanathan, Introduction
found in Gibson and Melsa (6).
to Statistical Signal Processing with Applications, Englewood Cliffs,
Sequential Detection 3. S. A. Kassam, Signal Detection in Non-Gaussian Noise, New York:
All of the discrete-time detection problems considered above Springer-Verlag, 1987.
involve a fixed number of observations. There are some situa- 4. H. V. Poor and J. B. Thomas, Signal Detection in Dependent Non-
tions, however, in which it may be advantageous to vary the Gaussian Noise, in H. V. Poor and J. B. Thomas, (eds.), Advances
number of observations. In radar systems, for example, a se- in Statistical Signal Processing, Vol. 2, Signal Detection, Green-
wich, CT: JAI Press, 1993.
ries of observations might correspond to repeated measure-
ments of a weak target. Naturally, we want to detect the 5. S. A. Kassam and H. V. Poor, Robust techniques for signal pro-
cessing: A survey, Proc. IEEE, 73: 433–481, 1985.
target as soon as possible—that is, using the fewest observa-
tions. Detection methods that permit a variable number of 6. J. D. Gibson and J. L. Melsa, Introduction to Nonparametric Detec-
tion with Applications, New York: Academic Press, 1975.
observations are the subject of sequential detection. Such
methods are applicable whenever each observation carries a 7. J. G. Proakis, Digital Communications, 2nd ed., New York:
McGraw-Hill, 1989.
cost, and we want to minimize the overall average cost of
making a reliable decision. 8. F. E. Nathanson, Radar Design Principles, New York: McGraw-
Hill, 1991.
One of the most important techniques in sequential detec-
tion is a Neyman–Pearson-type test called the sequential
probability ratio test (SPRT) (1). Suppose we want to decide BRIAN L. HUGHES
North Carolina State University
between the two hypotheses
H0 : Yi is i.i.d. with pdf p0 (y), i = 1, 2, . . .

H1 : Yi is i.i.d. with pdf p1 (y), i = 1, 2, . . .
using the smallest average number of observations necessary

to achieve a probability of false alarm PF and probability of
miss PM. The SPRT involves testing the accumulated data
after each observation time j ⫽ 1, 2, . . .. The test statistic at
time j consists of the likelihood ratio of all observations up to
that time, that is,
j
p1 (yi )
j (y1 , . . ., y j ) = i=1
j
i=1 p0 (yi )
At time j, we calculate ⌳j(Y1, . . ., Yj) and compare it to two

thresholds, ␶0 and ␶1. If ⌳j ⱖ ␶1 we decide in favor of H1, if
⌳j ⱕ ␶0 we decide in favor of H1, otherwise we take another
observation and repeat the test. The thresholds ␶0 and ␶1 are
chosen to provide the desired false-alarm and miss probabili-
ties. The SPRT minimizes the average number of observa-
tions under both H0 and H1, subject to constraints on PF and
PM.
FURTHER READING
A more complete and detailed treatment of most of the topics

covered in this article can be found in the books by Poor (1)
}{{}}
HOME
●
ABOUT US
●
CONTACT US
●
HELP

Your Librarian
Profile
Trellis-Coded Modulation ● Email this page Enter words or
● Print this page
Hamid Jafarkhani1 and Vahid Tarokh2
1AT&T Labs—Research, Red Bank, NJ
2AT&T Labs—Research, Red Bank, NJ ❍ Advanced Product
Copyright © 1999 by John Wiley & Sons, Inc. All rights Search
reserved. ❍ Search All Content
DOI: 10.1002/047134608X.W4216 ❍ Acronym Finder
Abstract
Historical Remarks
Overview
Trellises as Finite-State Machines
Mapping by Set-Partitioning
Decoding Trellis Codes: the Dynamic Programming Algorithm
Multidimensional Trellis Codes
Research Activities

TRELLIS-CODED MODULATION 567
TRELLIS-CODED MODULATION
Any communication in nature suffers from impairments such

as noise, which corrupts the data transmitted from the trans-
mitter to the receiver. In this article, we consider the princi-
ples behind trellis-coded modulation (TCM), which is an es-
tablished method to combat the aforementioned impairments.
TCM is one of the main components of the modern modulator-
demodulator (modem) systems for data transmission over
telephone lines.
HISTORICAL REMARKS
Trellis diagrams (or state transition diagrams) were origi-

nally introduced in communications by Forney (1) to describe
maximum likelihood sequence detection of convolutional
codes. They were employed to soft decode convolutional codes
using a dynamic programming algorithm (also known as the
Viterbi algorithm).
The concept of trellis was later extended by Bahl et al. (2)
to linear block codes where they were used as a natural
framework to implement the maximum a posteriori probabil-
ity (MAP) algorithm. Later, Forney unveiled the trellis struc-
ture of Euclidean Codes and Lattices.
Trellis-coded modulation is perhaps the most frequently
applied branch of trellis theory. Such an implementation com-
bines channel coding and modulation for transmission over
band-limited channels. Specifically, trellis-coded modulation
integrates the trellis of convolutional codes with M-ary linear
modulation schemes such as, for example, M-phase-shift key-
ing. Generally, modulation schemes containing larger Euclid-
ean distances between their signal sets provide more ro-
bustness against noise over Gaussian channels. On the other
hand, traditionally channel codes were designed so that dis-
tinct codewords have large Hamming distances (3). These two
criteria are not equivalent unless 2-amplitude modulation or
4-phase-shift keying (4-PSK) modulation is used. Combining
channel coding and modulation makes it possible to use a dis-
tance measure in coding which is equivalent to Euclidean dis-
tance in modulation. When the noise is additive white
Gaussian, trellis-coded modulation provides 3–6 dB improve-
ments over uncoded modulation schemes for the same band-
width efficiency. Although Massey had proposed the idea of
combining channel coding and modulation in 1974 (4), the
first trellis-coded modulation scheme was introduced by Un-
gerboeck and Csjaka in 1976 (5,6).
OVERVIEW
Figure 1 shows a block diagram of a communication system

in which binary data are transmitted over a noisy channel.
Since the signal transmitted over the physical channel is a
continuous electrical waveform, the modulation scheme con-
verts its binary (discrete) input to continuous signals which
are suitable for transmission over band-limited channels. If
the effects of noise on the transmitted signal can be modeled
by adding uncorrelated Gaussian noise samples, the channel
is called an additive Gaussian noise channel. The ratio of the
transmitted power to the noise power, signal-to-noise ratio
(SNR), is an important parameter which affects the perfor-
mance of the modulation scheme. For a given SNR and band-
568 TRELLIS-CODED MODULATION
Discrete Channel given rate and bandwidth, trellis-coded modulation uses a re-
dundant signal set at the modulator and a maximum likeli-
hood soft decoder at the demodulator. In trellis-coded modula-
Input Channel tion, the necessary redundancy of coding comes from
Modulator
(binary) Encoder expanding the signal sets not bandwidth expansion, as will
be discussed in the next section. Designing a good coded mod-
Physical ulation scheme is possible by maximizing the free Euclidean
(continuous) distance for the code. In fact, Ungerboeck and Csjaka’s point
Channel of departure from traditional coding is that the free distance
of a trellis-coded modulation can be significantly more than
Output Channel that of the corresponding uncoded modulation scheme.
Demodulator A trellis (state-transition diagram) can be used to describe
Decoder trellis-coded modulation. This trellis is similar to that of con-
volutional codes. However, the trellis branches in trellis-coded
modulation consist of modulation signals instead of binary
Figure 1. Block diagram of a communication system. codes. Since the invention of trellis-coded modulation, it has
been used in many practical applications. The use of trellis-
coded modulation in modulator–demodulators (modems) for
width, there is a theoretical limit for the maximum bit rate data transmission over telephone lines has resulted in tre-
which can be reliably transferred over a continuous channel mendous increased in the bit rate. International Telegraph
(Shannon capacity) (7). If the bit rate is less than the Shan- and Telephone Consultative Committee (CCITT) and its suc-
non capacity, the objective of a modulation scheme is to mini- cessor International Telecommunication Union (ITU) have
mize the bit error rate for a given SNR and a given band- widely utilized trellis-coded modulation in high-speed mo-
width. dems for data transmission over telephone lines (8–10).
The combination of modulation, continuous channel, and
demodulation can be considered as a discrete channel. Be-
cause of the hard-decision at the demodulator, the input and TRELLISES AS FINITE-STATE MACHINES
output of the discrete channel are binary. The effects of noise
in the physical channel translates into bit errors in the dis- Much of the existing literature (11–13) uses set partitioning
crete channel. The job of channel coding is to correct errors and trellis structure of convolutional codes to describe trellis-
by adding some redundancy to the bit stream. In other words, coded modulation. This may be attributed to the fact that this
error correcting codes systematically add new bits to the bit approach was taken by Ungerboeck and Csjaka in their semi-
stream such that the decoder can correct some of the bit er- nal paper where the foundation of coded modulation was laid.
rors by using the structure of the redundancy. Of course, the In this exposition, the goal is to present the results with the
adding redundancy reduces the effective bit rate per trans- required background kept as small as possible. In this light,
mission bandwidth. we pursue a different line of thought and approach the topic
Before the seminal work of Ungerboeck and Csjaka, chan- using finite-state machines.
nel codes and modulation schemes were designed separately.
Error correcting codes were designed to have codewords with Finite-State Machines
large Hamming distance from each other. Modulation
A finite-state machine can be thought of as a three-tuple
schemes utilize signal sets with maximum Euclidean dis-
M ⫽ (S , T , L ), where S , T , and L , respectively are referred
tance. Since Hamming distance and Euclidean distance are
to as the set of states, the set of transitions, and the defining
not equivalent for most modulation schemes, designing modu-
alphabet of M . Each element of the set T is a transition (si,
lation and coding scheme separately results in about 2 dB loss
se, l) with si, se 僆 S and l 僆 L . Such a transition is said to
in SNR. In contrast, trellis-coded modulation is designed to
start in si, end in se, and is labelled with l. All transitions
maximize Euclidean distance between the channel signal sets
starting from the same state, si, and ending at the same
by combining channel codes and modulation (Fig. 2). For a
state, se, are called parallel transitions. For each state s, the
number of transitions starting (respectively ending) in s is
called the out-degree (respectively the in-degree) of s.
TCM The finite-state machine M is said to be regular if the in-
Input
degrees and out-degrees of all the states of S are the same.
(binary) Encoder The machine M is binary if it is regular and if the out-de-
grees and in-degrees of elements of S as well as the number
Physical of states of S are powers of 2. In this article, we are only
(continuous) interested in binary machines.
Channel
The Trellis of a Binary Finite-State Machine
Output TCM
Every finite-state machine M has a trellis diagram T(M )
Decoder which is a graphical way to represent the evolution path of
M . Let M denote a binary finite-state machine having 2n
Figure 2. Using trellis-coded modulation to combine channel coding states. A trellis diagram T(M ) of M is defined as a labelled
and modulation. directed graph having levels 0, 1, 2, 3, . . . . Each level of M
merous other cases. A good general reference for trellis codes

0 0 is (14).
0 1
1 1 Trellis Codes for the Gaussian Channel
2 3 2 2 The design criterion (albeit an approximate one) for the

Gaussian channel is well established in the literature. In gen-
3 3 eral a code C is expected to perform well over a Gaussian
Figure 3. A four-state finite-state machine and the corresponding
channel if the codewords are as far from each other (in terms
trellis. Graphical equivalence between trellises and finite state ma- of Euclidean distance) as possible. The computation of Euclid-
chines is clearly visible. ean distance of two codewords of a code is not that difficult
and hence this criterion is tractable for design. To remove any
ambiguity, we mathematically define the distance between
two paths of T(M ) with the same starting and ending states.
has 2n states labelled 0, 1, . . ., 2n ⫺ 1 corresponding to, re-
Without loss of generality, let us assume that the two paths
spectively, s0, s1, . . ., s2n⫺1 elements of S . There is an edge
emerge at time t ⫽ 0 and remerge at time t ⫽ t⬘. Suppose
labelled with l between state i of level k and j of level k ⫹ 1
that the branches are labelled ct1 and ct2, t ⫽ 0, 1, . . ., t⬘, for
if and only if (si, sj, l) 僆 T where i, j ⫽ 1, 2, . . ., 2n, k ⫽ 1, 2,
the first and second path, respectively. Then, the distance be-
. . . and l 僆 L .
tween the two paths is defined by 兺t⫽0 兩ct1 ⫺ ct2兩2.
t⬘
Figure 3 shows an example of a finite-state machine M

For the design of a trellis code M , the minimum of dis-
containing four states and the corresponding trellis diagram
tances between any two paths of T(M ) that emerge from
T ⫽ T(M ). In Fig. 3, we only show the transitions between
some state at some time and remerge at another state of the
different states (not the labels). One can use different labels
trellis at a later time dominates the performance of the code.
on the transitions to construct different codes. This is the sub-
This quantity is called the free distance of the trellis code.
ject of the next section. It is clear that given a trellis diagram
Thus trellis codes that are useful for Gaussian channel must
T as defined, one can construct a finite-state machine M such
have large free distances.
that T ⫽ T(M ) and vice versa.
However, in pursuing such a design, we should take the
bandwidth requirements into account. Fixing the symbol du-
Trellis Codes ration (time to transmit a constellation symbols), the dimen-
A trellis code is the trellis of a binary finite-state machine sionality of the signal constellation directly relates to the
where the alphabet L comes from a signal constellation hav- bandwidth requirement for the channel. This is a fundamen-
ing unit average energy (we use unit average energy for all tal result known as the Landau–Pollak–Slepian Theorem
signal constellations in this article). Practical signal modula- (15,16). The consequence of this result is that a comparison
tion includes but is not restricted to the 4-PSK, 8-PSK, and between the free distances of two trellis codes is justified only
16-quadrature amplitude (16-QAM) constellations. In this if they use signal constellations of same dimensionality.
light, we only consider these signal constellations here.
Let M denote a trellis code with 2n states such that the in- An Ungerboek–Csjaka Idea
degree and out-degree of each state is 2R. Let T(M ) denote
the trellis of M and assume that at time zero the machine is Suppose that we would like to design a trellis code for the
at state zero. The trellis code M can be used to encode R bits transmission of R bits per channel use. One way of transmis-
of information at each time instance. At each time t ⫽ 0, 1, 2, sion is using a trellis code M that has one state and use a
signal constellation S C having 2R elements. The 2R edges be-
. . . a block of R bits of data denoted by B(t) arrives at the
tween the state of level t with that of t ⫹ 1 in T(M ) are la-
encoder. Depending on the 2R possible values of this block of
data and the state si(t) of the machine at time t, a transition belled with the different signal constellation symbols. This
beginning in that state such as (si(t), se(t ⫹ 1), l(t)) is chosen. trellis code is called the uncoded signal constellation S C . The
The trellis code then moves to the state se(t ⫹ 1) and outputs uncoded binary phase-shift keying (BPSK) constellation is
l(t) the label of the transition. Thus, B(0)B(1)B(2) . . . is given in Fig. 4. Clearly the free distance of the uncoded signal
mapped to the codeword l(0)l(1)l(2) . . . . We let C(M ) denote constellation S C is the minimum distance between the points
the set of all possible output sequences and also refer to it as of S C .
One way of obtaining larger free distances is to use a sig-
the code of M when there is no ambiguity.
nal constellation having more than 2R elements for transmis-
The alert reader notices that such an encoder may be com-
pletely useless. Indeed, if all the transitions are labelled with sion of R bits per channel use. In practice, it is good to double
the same signal constellation symbol, all bit sequences will be the constellation size while designing over the Gaussian chan-
nels. As the dimensionality of the signal constellation is fixed
mapped to the same codeword. Thus, it is important to design
the trellis code so that such a scenario is avoided. and the number of signals in the constellation is doubled, we
can expect a reduction in minimum distance of the new con-
The assignment of labels to transition in particular is what
determines the performance of a code over a transmission me- stellation.
dia. Thus, a performance criterion is needed before designing
a trellis code for real applications. In most of the situations,
an exact performance criterion is intractable for design and
a tractable approximate criterion is used instead. Tractable
approximate design criteria are known for the Gaussian chan- Figure 4. An uncoded BPSK constellation. Each point represents a
nel, rapidly fading channel, slowly fading channel, and nu- signal to be transmitted over the channel.
As an example to transmit 1 bit per channel use we will {A0, A1, A2, A3}
use a 4-PSK (Fig. 5) instead of BPSK constellation. The mini- {A0, A2} {A1, A3}
mum distance of the 4-PSK constellation is 兹2 while the min- {A0} {A1} {A2} {A3}
imum distance of the BPSK constellation is 2 (both have unit
average energy). Thus, there is a loss in minimum distance
by doubling the size of constellation. A Trellis code on 4-PSK
alphabet can only be useful as compared to the uncoded case
if it can compensate this loss by having a larger free distance
than 2.
Ungerboeck and Csjaka demonstrated that there exist trel-
lis codes that can outperform the uncoded signal constella-
tions. They also proposed mapping by set partitioning as the
machinery to construct these trellis codes.
MAPPING BY SET-PARTITIONING
Let S C be a signal set. Let S C 1 債 S C such that 兩S C 兩 the

Figure 6. Set partitioning for 4-PSK constellation. The partitioning
number of elements of S C be a multiple of 兩S C 1兩. A parti-
increases the minimum distance in each level.
tioning of S C based on S C 1 is a collection ⌺1 of disjoint sub-
sets of S C such that ⌺1 contains S C 1 and 傼X僆⌺1X ⫽ S C . Ele-
ments of ⌺1 are called the cosets of S C 1 in S C . The concept
(where j ⫽ 兹⫺1) are used to represent the 4-PSK, 8-PSK,
of partitioning can be extended to the nested chains of subsets
and 16-QAM constellations throughout this article.
of S C .
As can be seen from Fig. 8, the minimum distances of the
Specifically, consider a decreasing chain of subsets of a sig-
partitions in the 16-QAM case increase by a factor of 兹2 for
nal constellation S C
each level. By choosing appropriate signals from each parti-
S C = S C0 ⊇ S C1 ⊇ S C2 ⊇ . . . ⊇ S CJ tion level as the labels of transitions of a finite-state machine,
we could achieve very high free distances. This is the heart of
such that 兩S C i兩 is a multiple of 兩S C i⫹1兩 for i ⫽ 0, 1, . . ., J ⫺ Ungerboeck–Csjaka design and is called mapping by set parti-
1. Such a decreasing chain induces partitioning in each level. tioning.
First, S C is partitioned into a set ⌺1 of cosets of S C 1 in S C The general heuristic rules established for design by Un-
which in particular contains S C 1. Each element of ⌺1 con- gerboeck–Csjaka are
tains 兩S C 1兩 elements of S C . In a similar way, S C 1 can be
partitioned into cosets of S C 2 in S C 1 and the other elements • Parallel transitions (those starting from and ending in
of ⌺1 can be partitioned into sets of cardinality 兩S C 2兩. The re- the same states) are assigned to signal points with maxi-
sult is ⌺2, the collection of all the cosets of S C 2 in S C which mum Euclidean distance.
in particular includes S C 2. The process is then repeated for • The signal points should occur with the same frequency.
J times and all the cosets of S C i in S C j for 1 ⱕ j ⱕ i ⱕ J are • Transitions originating from and merging into any state
derived. In this article, we are only interested in partitions are assigned from elements of different cosets.
based on binary chains corresponding to the case when
兩S C i兩, i ⫽ 1, 2, . . ., J, are powers of two.
The central theme of the Ungerboeck–Csjaka paper (5) is
{B0, B1, B2, B3, B 4, B 5, B 6, B 7}
that given a binary set partitioning based on a decreasing {B1, B3, B5, B7} {B0, B2, B4, B6}
chain of subsets of S C as described, the minimum distance of {B3, B7} {B1, B7} {B4, B6} {B0, B4}
cosets of S C i in S C is a nondecreasing function of i. Indeed, {B7} {B3} {B1} {B5} {B2} {B6} {B0} {B4}
if the partitioning is done in a clever way, the distances can
substantially increase. Examples of such a set partitioning for
the 4-PSK, 8-PSK, and 16-QAM are given in Figs. 6, 7, and
8, respectively. The notations
Ak = cos(2πk/4) + sin(2πk/4)j, k = 0, 1, 2, 3
Bk = cos(2πk/8) + sin(2πk/8)j, k = 0, 1, 2, . . ., 7
√
Qk ,k = ((2k1 − 3) + (2k2 − 3)j)/ 10,
1 2
k1 = 0, 1, 2, 3, k2 = 0, 1, 2, 3
Figure 5. An uncoded 4-PSK constellation. Each point represents a Figure 7. Set partitioning for 8-PSK constellation. The partitioning
signal to be transmitted over the channel. increases the minimum distance in each level.
兵Q0,3 , Q2,1 , Q0,1 , Q2,3 , Q1,2 , Q3,0 , Q1,0 , Q3,2 , Q0,0 , Q2,2 , Q0,2 , Q2,0 , Q1,1 , Q3,3 , Q1,3 , Q3,1其
" '
兵Q0,3 , Q2,1 , Q0,1 , Q2,3 , Q1,2 , Q3,0 , Q1,0 , Q3,2其兵Q0,0 , Q2,2 , Q0,2 , Q2,0 , Q1,1 , Q3,3 , Q1,3 , Q3,1其
" ' " '
兵Q0,3 , Q2,1 , Q0,1 , Q2,3其兵Q1,2 , Q3,0 , Q1,0 , Q3,2其兵Q0,0 , Q2,2 , Q0,2 , Q2,0其兵Q1,1 , Q3,3 , Q1,3 , Q3,1其
" ' " ' " ' " '
兵Q0,3 , Q2,1其兵Q0,1 , Q2,3其兵Q1,2 , Q3,0其兵Q1,0 , Q3,2其兵Q0,0 , Q2,2其兵Q0,2 , Q2,0其兵Q1,1 , Q3,3其兵Q1,3 , Q3,1其
Figure 8. Set partitioning for 16-QAM con-
" ' " ' " ' " ' " ' " ' " ' " ' stellation. The partitioning increases the mini-
兵Q0,3其兵Q2,1其兵Q0,1其兵Q2,3其兵Q1,2其兵Q3,0其兵Q1,0其兵Q3,2其兵Q0,0其兵Q2,2其兵Q0,2其兵Q2,0其兵Q1,1其兵Q3,3其兵Q1,3其兵Q3,1其 mum distance in each level.
These rules follow the intuition that good codes should have To understand the implementation of the decoder, we first
symmetry and large free distances. Examples of 4-PSK, define the constraint length ␯ (C) of a trellis code C(M ) to be
8-PSK, and 16-QAM codes are given in Tables 1–5. the minimum t such that there exists two paths of time length
From these tables, it is clear that by increasing the num- t starting at the same state and remerging at another state.
ber of states in the trellis, the free distance (and hence the Practically, we choose a multiple of ␯ (C) depending on the
performance) can be improved. However, we will see that this decoding delay allowed in the application and refer to it as
has a penalty in terms of decoding complexity. the decoding depth ␪(C). We then proceed to execute the finite
Let us now consider an example. Consider the set parti- decoding depth Viterbi algorithm. At each stage of the algo-
tioning of the 8-PSK and the four-state trellis code given in rithm, for every possible state s of the encoder, a survivor
Table 3 based on the previous partitioning. As can be seen path Pt(s) of length ␪(C) and an accumulated metric mt(s) is
from the table, the labels of the transitions originating from preserved. We denote the possible states of the encoder by si,
each state of the trellis belong to the same coset while those i ⫽ 0, 1, . . ., 2n ⫺ 1, and the received signal at time t by rt.
of distinct states belong to different cosets. The design has We always follow the convention that the encoder is in the
a lot of symmetries as it is expected that good codes should zero state at time zero.
demonstrate a lot of symmetries. It can be easily shown that The decoder starts by setting m0(s0) ⫽ 0 and m0(si) ⫽ 앝 for
free distance of the previous trellis code is 兹2 times the mini- all i ⫽ 1, 2, . . ., 2n ⫺ 1. In practice, one can choose a large
mum distance of a 4-PSK constellation. This translates into number instead of 앝. Further, at the beginning of the decod-
3-dB asymptotic gain (in SNR). In general the asymptotic ing process, the decoder sets the survivor paths Pt(si), i ⫽ 0,
gain of a trellis code with rate R bits per channel use (2R⫹1 1, 2, . . ., 2n ⫺ 1, to be the void string. In other words, at the
elements in the constellation) over an uncoded constellation beginning of the decoding nothing is saved as the survivor
2 2
with the same rate is defined by 10 log dfree /dmin where dfree is paths of each state.
the minimum free distance of the code and dmin is the mini- The decoder then starts decoding by computing the branch
mum distance between the uncoded constellation elements. metrics of each branch at time t ⫽ 0, 1, 2, 3, . . . . Suppose
Figures 9 and 10 give information about the coding gain that a branch at time t is labelled with ct, then the metric of
versus the number of states of best 8-PSK and 16-QAM trellis this branch is 兩rt ⫺ ct兩2. The decoder computes for each state
codes known for transmission of 2 and 3 bits/channel use, re- si, the sum of the accumulated metric mt(sj) and the branch
spectively. metric of any state sj with any branch starting at state sj at
time t and ending in state si at time t ⫹ 1. The decoder then
DECODING TRELLIS CODES: THE DYNAMIC computes the minimum of all these possible sums and sets
PROGRAMMING ALGORITHM mt⫹1(sj) to be this minimum. If this minimum is given by the
state i at time t and some branch bt, the survivor path
Decoding trellis codes is usually done through the dynamic Pt⫹1(sj) is given by the path Pt(si) continued by the branch bt.
programming algorithm also known as the Viterbi algorithm. This process is then repeated at each time.
The Viterbi algorithm is in some sense an infinite algorithm The decoder starts outputting decision bits after time t ⱖ
that decides on the path taken by the encoder. This was ␪(C), where ␪(C) denotes the decoding depth. At each time
proved to be optimum for sequence estimation by Forney. t ⱖ ␪(C), the decoder looks at the survivor path of the state
However, in practice one has to implement a finite version of with the lowest accumulated metric. The decoder outputs the
the algorithm. Naturally, only practice is of interest here. sequence of bits corresponding to the branch of path at time
t ⫺ ␪(C). In this way, a decoding delay of ␪(C) must be tol-
erated.
Table 1. A 4-State 4-PSK Trellis Code
se ⫽ 0 se ⫽ 1 se ⫽ 2 se ⫽ 3
MULTIDIMENSIONAL TRELLIS CODES
si ⫽ 0 A0 A2
si ⫽ 1 A1 A3 The trellis codes constructed in the previous section use an
si ⫽ 2 A2 A0 element of a two-dimensional constellation for labels. It is nei-
si ⫽ 3 A3 A1
ther necessary to have a two-dimensional constellation nor
Note: The states si and se are, respectively, the beginning and ending states. only one symbol of the constellation per label of transitions.
The corresponding transition label is given in the table. Blank entries represent This gives rise to multidimensional trellis codes or M-TCM
transitions that are not allowed.
codes.
Table 2. An 8-State 4-PSK Trellis Code

se ⫽ 0 se ⫽ 1 se ⫽ 2 se ⫽ 3 se ⫽ 4 se ⫽ 5 se ⫽ 6 se ⫽ 7
si ⫽ 0 A0 A2
si ⫽ 1 A1 A3
si ⫽ 2 A2 A0
si ⫽ 3 A3 A1
si ⫽ 4 A0 A2
si ⫽ 5 A1 A3
si ⫽ 6 A2 A0
si ⫽ 7 A3 A1
Note: The states si and se are, respectively, the beginning and ending states. The corresponding
transition label is given in the table. Blank entries represent transitions that are not allowed.
Table 3. A 4-State 8-PSK Trellis Code

se ⫽ 0 se ⫽ 1 se ⫽ 2 se ⫽ 3
si ⫽ 0 B 0 , B4 B 2 , B6
si ⫽ 1 B 1 , B5 B3 , B 7
si ⫽ 2 B 2 , B6 B 0 , B4
si ⫽ 3 B 3 , B7 B1 , B 5
Note: The states si and se are, respectively, the beginning and ending states.
The corresponding possible transition labels are given in the table. Blank en-
tries represent transitions that are not allowed.
Table 4. An 8-State 8-PSK Trellis Code

se ⫽ 0 se ⫽ 1 se ⫽ 2 se ⫽ 3 se ⫽ 4 se ⫽ 5 se ⫽ 6 se ⫽ 7
si ⫽ 0 B0 B4 B2 B6
si ⫽ 1 B1 B5 B3 B7
si ⫽ 2 B4 B0 B6 B2
si ⫽ 3 B5 B1 B7 B3
si ⫽ 4 B2 B6 B0 B4
si ⫽ 5 B3 B7 B1 B5
si ⫽ 6 B6 B2 B4 B0
si ⫽ 7 B7 B3 B5 B1
Note: The states si and se are, respectively, the beginning and ending states. The corresponding
possible transition labels are given in the table. Blank entries represent transitions that are
not allowed.
Table 5. A 4-State 16-QAM Trellis Code

se ⫽ 0 se ⫽ 1 se ⫽ 2 se ⫽ 3
si ⫽ 0 Q1,3 , Q3,3 , Q1,1 , Q3,1 Q0,0 , Q0,2 , Q2,0 , Q2,2
si ⫽ 1 Q0,1 , Q0,3 , Q2,1 , Q2,3 Q1,0 , Q1,2 , Q3,0 , Q3,2
si ⫽ 2 Q0,0 , Q0,2 , Q2,0 , Q2,2 Q1,3 , Q3,3 , Q1,1 , Q3,1
si ⫽ 3 Q1,0 , Q1,2 , Q3,0 , Q3,2 Q0,1 , Q0,3 , Q2,1 , Q2,3
Note: The states si and se are, respectively, the beginning and ending states. The corresponding possible transition labels are given in the table. Blank entries
represent transitions that are not allowed.
6 achieve higher coding gains but have other implementation

* problems including the design of slicer and increased decod-
5.5 ing complexity.
*
Trellis ideas were also applied to quantization giving rise
5 *
to trellis-coded quantization which can be used to quantize
*
dB
4.5 various sources (19,20).

In general, we believe that a fruitful area of research may
4 *
be the study of implementation issues of trellis codes over
* channels with ISI and non-Gaussian channels in the presence
3.5
of various impairments due to practical situations. There is a
3* well-established body of literature on this topic (21,22) but we
2 3 4 5 6 7 8 believe that there is a lot more to be done.
n
Figure 9. Asymptotic coding gain of coded 8-PSK over uncoded 4-
PSK (number of states ⫽ 2n). Coding gain represents the improve- BIBLIOGRAPHY
ment in the performance of the coded system over that of the un-
coded system. 1. G. D. Forney, Trellises old and new, in Communications and
Cryptography: Two Sides of One Tapestry, R. E. Blahut et al.
(eds.). Dordrecht, The Netherlands: Kluwer, 1994.
2. L. Bahl, J. Cocke, F. Jelinek, and J. Raviv, Optimum decoding of
An example of an M-TCM code is given in Ref. 16, which linear codes for minimizing symbol error rates, IEEE Trans. Inf.
is a four-dimensional trellis code known as the Wei code (17). Theory, IT-20: 284–287, 1974.
3. F. J. MacWilliams and N. J. A. Sloane, The Theory of Error-Cor-
recting Codes, New York: Elsevier, 1996.
RESEARCH ACTIVITIES 4. J. L. Massey, Coding and modulation in digital communications,
Proc. 1974 Int. Zürich Seminar on Digital Commun., E2(1)–(4),
An active area of theoretical research is studying the trade- Mar. 1974.
off between the complexity and coding gain of trellis codes. In 5. G. Ungerboeck and I. Csajka, On improving data-link perfor-
essence, we would like to see trellis codes with lower complex- mance by increasing the channel alphabet and introducing se-
ity of decoding and higher coding gain. Much effort has been quence coding, Int. Symp. Inform. Theory, Ronneby, Sweden,
put into finding solutions to this problem, but only meager June 1976.
improvements have been observed over the codes constructed 6. G. Ungerboeck, Channel coding with multilevel/phase signals,
in the original paper of Ungerboeck–Csjaka. IEEE Trans. Inf. Theory, IT-28: 55–67, 1982.
A second active area is to find suboptimal algorithms for 7. C. E. Shannon, A mathematical theory of communication, Bell
decoding trellis codes which give performance close to that of Syst. Tech. J., 27: 379–423, 1948.
the optimum Viterbi algorithm. Numerous papers have been 8. CCITT, A family of 2-wire, duplex modems operating on the gen-
written on this topic proposing reduced complexity algorithms eral switched telephone network and on leased telephone-type
including the sequential decoding algorithm and the M-algo- circuits, Recommendation V.32, 1984.
rithm (18). These decoding algorithms perform close to opti- 9. CCITT, 14400 bits per second modem standardized for use on
mal but do not seem promising due to other implementation point-to-point 4-wire leased telephone-type circuits, Recommen-
problems including the problem with buffer overflow. dation V.33, 1988.
Another research area is to combine mathematical objects 10. ITU-T, A modem operating at data signaling rates of up to 33600
called lattices with trellis codes (14). These theoretically bit/s for use on the general switched telephone network and on
leased point-to-point 2-wire telephone-type circuits, Recommen-
dation V.34, 1996.
11. G. Ungerboeck, Trellis-coded modulation with redundant signal
7.5 sets part II: State of the art, IEEE Commun. Magazine, 25: 12–
* * 21, 1987.
7
* 12. E. A. Lee and D. G. Messerschmitt, Digital Communication, Bos-
6.5 ton: Kluwer, 1988.
* * 13. J. G. Proakis, Digital Communications, New York: McGraw-Hill,
6
Inc. 1989.
dB
5.5 14. E. Biglieri et al., Introduction to Trellis-Coded Modulation with

*
5 Applications, New York: Macmillan, 1991.
15. H. O. Pollak and H. J. Landau, Prolate spheroidal wave func-
4.5 tions, Fourier Analysis and uncertainty II, Bell Syst. Tech. J., 40:
*
4 65–84, 1961.
2 3 4 5 6 7 8 16. H. J. Landau and H. O. Pollak, Prolate spheroidal wave func-
n tions, fourier analysis and uncertainity III: The dimension of the
Figure 10. Asymptotic coding gain of coded 16-QAM over uncoded space of essentially time and band-limited signals, Bell Syst.
8-PSK (number of states ⫽ 2n). Coding gain represents the improve- Tech. J., 41: 1295–1366, 1962.
ment in the performance of the coded system over that of the un- 17. L. F. Wei, Trellis-coded modulation with multi-dimensional con-
coded system. stellations, IEEE Trans. Inf. Theory, IT-33: 483–501, 1987.
574 TRIBOELECTRICITY
18. J. M. Wozencraft and B. Reiffen, Sequential Decoding, Cam-

19. M. W. Marcellin and T. R. Fischer, Trellis-coded quantization of
memoryless and Gauss-Markov sources, IEEE Trans. Commun.,
38: 82–93, 1990.
20. M. Wang and T. R. Fischer, Trellis-coded quantization designed
for noisy channels, IEEE Trans. Inf. Theory, 40: 1792–1802, 1994.
21. D. Divsalar and M. K. Simon, The design of trellis-coded MPSK
for fading channel: Performance criteria, IEEE Trans. Comm., 36:
1004–1012, 1988.
22. D. Divsalar and M. K. Simon, The design of trellis-coded MPSK
for fading channel: Set partitioning for optimum code design,
IEEE Trans. Commun. 36: 1013–1021, 1988.
HAMID JAFARKHANI
VAHID TAROKH
AT&T Labs
TRELLIS CODES. See TRELLIS-CODED MODULATION.

TRENDS IN SYSTEMS ENGINEERING. See SYSTEMS
ENGINEERING TRENDS.
TRENDS, SYSTEMS ENGINEERING. See SYSTEMS ENGI-
NEERING TRENDS.
TRIANGLE WAVE GENERATION. See RAMP GEN-
ERATOR.

Algebraic Coding Theory PDF

Uploaded by

Copyright:

Available Formats

Algebraic Coding Theory PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Algebraic Coding Theory PDF

Uploaded by

Copyright:

Available Formats

29• Information Theory

29• Information Theory

Algebraic Coding Theory Abstract | Full Text: PDF (222K)

Cryptography Abstract | Full Text: PDF (116K)

Data Compression Codes, Lossy Abstract | Full Text: PDF (190K)

Estimation Theory Abstract | Full Text: PDF (161K)

Image Codes Abstract | Full Text: PDF (399K)

Information Theory Abstract | Full Text: PDF (131K)

Information Theory of Multiaccess Communications Abstract | Full Text: PDF (204K)

Information Theory of Spread-Spectrum Communication Abstract | Full Text: PDF (141K)

Information Theory of Stochastic Processes Abstract | Full Text: PDF (279K)

Maximum Likelihood Imaging Abstract | Full Text: PDF (290K)

Queueing Theory Abstract | Full Text: PDF (456K)

Signal Detection Theory Abstract | Full Text: PDF (152K)

Trellis-Coded Modulation Abstract | Full Text: PDF (143K)

Wiley Encyclopedia of Electrical and ● Recommend to Browse this title

Some Bounds on Codes

The Weight Distribution of a Code

The Binary Golay Code

Nonlinear Codes from Codes over Z

file:///N|/000000/0WILEY%20ENCYCLOPEDIA%20OF%20...NGINEERING/29.%20Information%20Theory/W4205.htm (1 of 2)17.06.2008 14:13:16

About Wiley InterScience | About Wiley | Privacy | Terms & Conditions

file:///N|/000000/0WILEY%20ENCYCLOPEDIA%20OF%20...NGINEERING/29.%20Information%20Theory/W4205.htm (2 of 2)17.06.2008 14:13:16