Professional Documents
Culture Documents
1, FEBRUARY 1969
Aeronautical Lab., Inc., B~ffalo,N. Y., Rept. UA-1420-S-1, Vasant K. Prabhu (S’64-M’66)
Detect Memo. 2A, January 1961. was born in
Kumta,India, on
J . H. Halton and A. D. Spauldiug, “Error rates in differentially May 13, 1939. He received the
coherent phase systems in non-Ga,rmsian noise,” IEEE Trans. B.Sc. (Hons.) degree from Kar-
Communication Technology, vol. COM-14, pp. 594-601, October
1966. natak University, Dhanvar, India,
W. 1V.Hubbard, “The effect of a finite-width decision thresh- in 1958, and the B.E. (Dist.)
old on binary differentially coherent PSK systems,” Bell degree in electrical communication
Sys. Tech. J., vol. 45, pp. 307-319, February 1966. engineering from theIndianIn-
-,“The effect of noise correlationon binary differentially stitute of Science, Bangalore, India,
coherent PSK communication sy&ms,” Bell Sys. Tech. J., in 1962. He received the S.M. and
vol. 46, pp. 277-280, January 1967. Sc.D. degrees from the Mas-
-,“The effect of intersymbol interference on error rate in sachusetts Institute of Technology,
binary differentially coherent phase-shift-keyed systems,”
Bell Sys. Tech. J., vol. 46, pp. 1169-1172, July-August 1967. Cambridge,1963
in and 1966,
E. Bedrosian and S. 0. Rice, “13istortion and crosstalk of respectively.
linearly filtered, angle-modulated signals,” Proc. IEEE, vol. At M.I.T. he was amember of
56, pp. 2-13, January 1968. the Research Laboratory of Elec-
G. F. Montgomery, “A comparison of amplitudeand angle tronics, first as a Research Assistant and then as an Instrllctor in
modulation for narrow-bandcommunication of binary-coded electrical engineering. Presently he is a Member of the Technical
messages in fluctuation noise,” P ~ o c IfZE,
. vol. 42, pp. 447- Staff of Bell Telephone Laboratories, Inc., Holmdel, N. J. His areas
4.54, February 1954.
W. R. Bennett,“Methods of solving noise problems,” Proc. of interestare nonlinear systemstheory,communication theory,
IRE, vol. 44, pp. 609-638, May 1956. network theory, solid-state microwave devices, noise theory,and
W. B. Davenport and W. L. Root, Introduction to the Theory optical communication systems.
of RandomSignalsandNoise. New York: McGraw-Hill, Dr. Prabhu is a member of Tau Beta Pi, Eta Kappa Nu, Sigma
1958. S i , and the American Association for the Advancement of Science.
This procedure assumes that it is of greater importance Since Gl(x) has degree g1 = 12 - kl and G2(z) degree
that the processing of a code word with an invalid M I be gz = n - kz, the degree of K ( x ) is g1 - g2 - I = k2 - k, -
discovered than thatsuch a n invalid code word be decoded 1 or less where K ( x ) is given by
incorrectly because of additive noise. The141written
s1(x) = K(x)G,(z)
explicitly in thefirst code word of the message is protected
from noise by the paritycheck bits of that word. Therefore and s1(x) is a member of S. This leads to the following
the first code word should yield the correct MI. corollary.
Note that if the usual method of message identifying is
used, where the MI is explicitly displayed as part of the Corollary 1
information-bits portion of each code word, then memory The set X for code C1 has 2g1-gz= 2k2-k1members. (We
is also required by theportion of the systemprocessing the include K(z) = 0.)
message so that the 1\41of a block can be checked against Example: Consider the BCH codes C1,C2 formed by
the MIof preceding blocks. Thus thesame amount of stor- generators
age is needed with the method presented here as with the
usual method.
Gl(x) (xs x2 + +l ) ( ~x4 ~ x3+ + + x2
The following sections will examinecertain classes of + + + + +
1) (xS x4 x 2 x: 1)
cyclic codes to discover possible sets U and S. The use of
these sets is in some cases similar to theuse of unused syn-
Ga(z) = (x5 x 2 + + + + + +
I) (x5 x4 x 3 x 2 1)
dromes in cyclic codes that correct, synchronization errors where n = 31. Then tl = 3, and t a = 2; so 2t2 > tl. There-
PI. fore sI(x) = (Adz4 &x3 + A2x2+ Alx + +
Ao)G2(x)for
Ai = l , O . There are therefore 2j members in the set S.
111. FULL-LENGTH RANDOM E:RRORCORRECTING It should also be noted that if a code word having an
CYCLICCODES A41 syndrome s?(x) is substituted for acode word havingan
The length n of a cyclic code is such that the code gen- M I syndrome s1(z), then the result (if no noise errors are
+ +
erator G(z) divides z n 1 and no expression xe 1 where present) will be the syndrome sl(x) +
sz(z) = s(z). Know-
e < n. If all n bits are used for transmitting inform at'1011 ing sl(z) we can then derive the 141 of the wrongly sub-
thcn the code is said to be full length. stituted message by adding sl(z) to s ( ~ ) .
A. Bose-Chauclhuri-Hocquela(7hem ( B C H ) Codes [1J B. Adding Parity Bit to a Code of Odd Weight
These codes can correct t random errors with a word It is well known that any code can correct errors having
length of 2" - 1 bits. The redundancy is a t most mt bits. The weight t and simultaneously detect errors of magnitude
code generator polynomial is the product of t polynomials d 2 t if and only if the minimum weight of the code is a t
each of degree m - 1 or less. + +
least t d +
1. Let d = t 1 ;then theweight of the code
Suppose code Cz is generated by (&(x) and code C1 is is 2t+ 2. This code can be constructed by addingan over-
generated by Gl(x)where Gl(z) = F(z)G2(z).Cz corrects alleven parity check to a code correcting t errors and
tz errors and C1 corrects tl errors where tl > tz. Let the de- having weight 2t + 1. This will add one more bit of re-
gree of G 2 ( x ) = n - k2, and the degree of Gl(x)= 12 - kl. dundancy. Thus theresulting code will correct t errors and
Suppose sl(x) is a syndrome of C 1 which is a multiple of detect all errors of weight t + +
1. If t is odd, thent 1 will
G2(x).This is possible since the degree of G2(x)is less than be even. Let the length of this code C be n and have gen-
the degree of G'l(x). Then s1(x) = K ( z ) G 2 ( x ) .Anyerror erator G(x).
pattern giving this syndrome in thecode C1 has the form U is the setof different syndromes obtained fromthe set
K(x)G2(2) + A(x)G,(z) = G'Z(X)(K(X) + A(x)F(x)). (2)
of error patterns of weight t + 1.
S will be a set of syndromes resulting from patterns of
This expression is a code word of Cy2 and therefore has the weight t + 1 such that if sl(x), s2(x) are in S, then a(z)+
weight a t least 2ta+ 1. This leads to t,he following theorem s2(x) are in U . Let Y be a linear code of length n or less
having constant weight t + 1. The code words of Y form
due to Tong [2].
an additive group. Therefore the remainders after division
Thevrem I of each word of Y by G(x) will form an additive group.
+
If tl < 2t2 1, then the syndrome sl(z) = K(x)Ga(z)is This isshown by thefollowing. Let yl(s),y2(z)be members
unused for additive error correction by the code C1. As a of Y.Then
direct result we have the following;theorem. ?Jl(Z)= A(x)G(x) + a(s)
Theolem 2 +
Y~(Z)= B ( x ) G ( x ) s ~ ( x ) ..
For the BCH code C1 generated by Gl(x) = F(x)G2(x), Then
where t1 5 2t2, the set S is the set of all multiples of Gz(x)
having degree less than the degree of G1(x).The proof is + + + +
YI(~) ~ 2 ( z )= ( A ( x ) B ( x ) ) G ( x ) ( s 1 ( ~ ) ~ a ( z ) )
obvious since the sumof any two such multiples is another and the syndrome of the sumof yl(x), y2(z)is equal to the
multiple; thns. S = IT. sum of the syndromes sl(x), s2(z).
IDENTIFICATION
MANDELBAUM:
MESSAGE
CYCLIC CODING AND 45
Therefore Y can be used for S. Weiss [3] shows that any IV. BURSTERROR CODES
CORRECTING
linear code of constant weight is derivable from a maximal No known cyclic burst error connecting codes are per-
lengthshift-register sequence code. Anylinear code of fect, i.e., some syndromes do not correspond to an error
constant weight must have length m(2&- 1) and weight burst pattern thatis to be corrected (except for the single
m2"-' with k information bits, where m is any integer. error correcting Hamming codes).
Thus t + 1 must equal m2k-1 for some m and k , where A cyclic burst connecting code is generated b y a poly-
m(2k - 1) 5 n. The number of code words in this code Y nomial G(x) of degree n - k which divides x" - 1 and no
is 2k - 1. other polynomial of the form xe - 1, e < n. G(0) = 1,
Suppose t +1 = 2dp where p is odd. Then itfollows: and n is the length of the code. Every polynomial of degree
m2'c-l = 2dp I n - 1 is equivalent modulo G(s) to an element of a
finite set P, which has 2n-* elements.
or
. . ., Ezn-k(x).
nz = p, lc = d +1 &(x), Ez(x),
P is partitioned into subclasses such that if E,(z) is an
and the numberof members of Y is 2d+1 - 1. Since it was element of any subclass and E,($) is an element of any
assumed t is odd, then d 2 1, and 2d+1 - 1 2 3. For d = other subclass, then
1, 2, 3, 4, 5, . ., thenumber of nonzero code words of
weight t + 1 in Y is 3, 7, 15, 31, 63, . . ., respectively. xiEe(x) # xjE,(z)mod G(z)
Thus byadding one parity bit, a t least 4 M I syndromes are
for any i, j . This partition will give: a first subset of nl
obtained (including zero.) However, if d is small, then
elements, a second subset of n2 elements, . a bth subset
e ,
y;]
rectly comes from a Hadamard matrix
0 0 1 1 '
ber and length of cycles produced by Gi(x) individually.
For example, a primitive polynomial of degree m possesses
two cycles: the trivial zero cycle of length one and the
maximallength cycle of 2" - 1. The cycle structure is
formally represented by the expression
cycles of equal length n. The cycle structure of the product erated by Q(x) alone. Since Gl(x) is aprimitive poly-
G(x) of two distinct irreducible pokynomials is nomial of degree 2, it will produce a cycle of length 3.
l1 + Pl(nl)l[l + P2(n2)1 = [1 + /~!1(711) + PZ(n2) + P(n) 1 Thus theurlused syndromes form asingle cycle. As a result
thereare only4 MI syndromes: {0,G2(x), xGz(x), (x +
where 1)G2]. This is due to thefollowing argument.
then the type of cycles formed by the code generated by where the Gi(x) are irreducible and P ( x ) is primitive. The
Gl(z)G2(x) is given by - maximal cycle length of P ( x ) is a multiple of the maximal
[l +c z
Pi(%'] [1 + P(m2)1 =
[
1 + c Pit%) + P(md
a
cycle lengths of the G',(x). P ( x ) has degree greater than the
sum of the degrees of the Gi(x).
+C a
P ~ Pgcd (I*&
Lemma 1 (lcm(wm))].
For an optimumcode, a cycle is unused for burst correc-
Theoptimumburst correcting codes considered by tion if and only if its characteristicis a multiple of P ( x ) .
Abramson [4] are generated byG1(:c)G2(x) where GI(%)and Proof: If the characteristic of a cycle is xjni,Gi,(x)P(x),
G2(x)are primitivepolynomials of degree 2 and 77, - k - 2, the cycles formed are equivalent to the cycles formed by
respectively. The length of the code is n and it corrects x j n i r C i , ( x )with respect to thecode generated byH i G i ( x ) ,
bursts of length 3 or less. There axe n - k parity check since P ( x ) cancels out in the division process giving the
bits. These codes then have the cycle form remainder. By the definition of the G i ( s ) , none of these
[I + +
l ( 2 d - ' - 1)][1 1(2n--8-@--1) - l ) ] cycles areequal to the maximal cycle lengthandare
therefore unused for burst correction.
= [I + 1(2d-1 - 1) + 2d-1(2n--8-(d-1) - I)]
Conversely, suppose characteristic
a is
formed by
I
Proof: By Lemma 1,any syndrome of the form Q ( x ) P ( z ) candidates for the set X of MIS. Of course the sum of any
where deg &(x) < u isunused. Since &(x) is any poly- two such MI syndromes must again be another unused
nomial of degree u - 1 or less (including zero) and the sum syndrome.
of any two such polynomials has degree less than u,the
result follows. A . Random Error Correcting Codes
Example: Consider the following optimum burst code as Assume that the shortened cyclic code C generated by
given by Elspas and Short [5]. This code corrects length 4 G(x) corrects for t or fewer errors. There are y unused and
bursts, has a lengthof 1023 bits, and is generated by omitted leading positions in each code word. Let the setof
G,(z) = 5 + 1, G‘~(z) = x 2 + z + I error patterns that containa total of t or less errors and a t
least one error in the omitted y bits be U . The number of
P(z) = xlo + x4 + + + x + 1.
2 3 x2 syndromes in U is given by
terns is a pattern having t or fewer 1’s. The number of to each code word except the first. (The first code word
such patterns (including the zero pattern) is carries the MI as part of the information bits which are
protected from noise by the parity check bits.) BCH and
E
2=0 a);(
burst correcting cyclic codes have syndromes which are
not used for error correction. These canbe used by the MI
Now consider the situation when! t = 2u +
1. All pat- syndromes. The amount of added logic needed to imple-
terns containing u 1’s are in the set, but not all patterns ment this system is quite small. By use of this method
having u + 1 1’s can be used. This is because the sum of rather than theuse of information bits for the MI, the in-
two patterns each having u +
1 1’s in totally different formation efficiency of the system may besignificantly in-
positions will sum toa pattern having2u +
2 bits which is creased since these information bits arenow free for other
not correctable. Thus only those
(9
patterns of u + 1 1’s use.