18 1EEE TRANSACTIONS O N CIRCUITS AND SYSTEMS -11: ANALOG AND DIGITAL SIGNAL PROCESSING. VOL. 39. NO.

9. NO. I , JANUARY 1992

A Coding Theory Approach to Error Control

in Redundant Residue Number Systems -
Part 11: Multiple Error Detection
and Correction
Jenn-Dong Sun and Hari Krishna

Abstract-In this work, we study and extend the coding table replaced by appropriate computations. Consequently,
theory approach to error control in redundant residue number their implementation needs a memory space that is much
systems (RRNS). We derive new computationally efficient algo-
rithms for correcting multiple errors, single-burst error, and smaller than that required in [2]. Ramachandran [7] proposed
detecting multiple errors. These algorithms reduce the computa- a method to correct single errors that establishes a trade-off
tional complexity of the previously known algorithms by at least between computational complexity and extra redundant mod-
an order of magnitude. uli. A number of papers by Jenkins and his associates [8]-[ 111
Index terms: Algorithms, multiple errors detection and cor- applied mixed radix conversion (MRC) to digital filters and
rection, burst residue errors, redundant residue number systems,
coding theory, mixed radix conversion. residue number error checkers. Su and Lo [ 121 have used the
redundant digits of MRC as the entries to construct a lookup
table for single residue-error correction.
I. INTRODUCTION In our previous paper [ 141, we developed a coding theory
approach to error control in RRNS. The concepts of Ham-
A MONG the earliest researchers, Szabo and Tanaka [ I ]
have briefly sketched a method for single-error detec-
tion or single-error correction for use with a residue number
ming weight, minimum distance, weight distribution, and
error detection and correction capabilities in RRNS were
introduced. The necessary and sufficient conditions for the
system (RNS). However, the error correction procedure given
desired error control capability were derived from the mini-
in [ l ] is computationally inefficient. Also, it appears to be
mum distance point of view. A special case generated the
quite complicated for implementation. Watson and Hastings
maximum distance separable (MDS) RRNS. A computation-
[2] have constructed a redundant residue number system
ally efficient procedure was described for correcting single
(RRNS) to detect or correct single errors. However, their
residue error.
method for error correction needs a correction table, which
In this paper, we will extend the theory and present new
may require large memory space, thereby making it impracti-
procedures for simultaneously correcting single error and
cal for the correction of more than single residue errors.
detecting multiple errors, and simultaneously correcting dou-
Therefore, multiple-residue-error correction has not been
ble errors and detecting multiple errors. In addition, we
investigated by them. Mandelbaum [3] showed how single-
present a procedure for correcting a single-burst error.
error correction can be accomplished in an RRNS code and
This paper consists of eight sections. In Section 11, we
established that two redundant moduli with redundancy less
cover the definitions and basic coding theory for RRNS. In
than the redundancy in [2] are necessary for single residue
Section 111, we generalize the property of consistency check-
digit error correction. Later, necessary and sufficient condi-
ing for RRNS. Three computationally efficient procedures for
tions for minimal redundancy allowing the correction of the
i) simultaneously correcting single error and detecting multi-
whole class of single residue errors were derived by Barsi
ple errors, ii) simultaneously correcting double errors and
and Maestrini [4], who also developed the concept of an RNS
detecting multiple errors, and iii) correcting single-burst er-
product code (31, a concept that was earlier suggested by
ror, are presented in Sections IV. V, and VI, respectively. In
Mandelbaum. Yau and Liu [6] designed two error-correction
Section VII, we extend the previous algorithms in [4] and [9]
algorithms, one for single residue-error correction and the
for multiple-error correction and detection. Finally, we dis-
other for burst residue-error correction. Basically, the method
cuss and compare the various results in Section VIII. For
of Yau and Liu is Watson’s method with the error-correcting
reasons explained in 1141, we focus exclusively on MDS-
Manuscript received December 4, 1990: revised July 31, 1991. The work RRNS in this paper.
of J.-D. Sun was supported by a fellowahip from Chung-Shan Institute of
Science and Technology, ROC. This paper was recommended by Associate
A N D BASICCODING
11. DEFINITIONS THEORY
FOR RRNS
Editor E. J . Coyle.
The authors are with the Department of Electrical and Computer Engi-
neering, Syracuse University, Syracuse, NY 13244.1240. Let { rn,, m z , .. . , m k } be a set of k positive relatively
IEEE Log Number 9104716. prime integers called nonredundant moduli and let their

1057-7130,9?$03.00 1992 IEEE

SUN AND KRISHNA: CODING THEORY APPROACH-PAKT 11 19

product be Definition 2: The Hamming distance between two code-

vectors xi and xi, d(x,, xi) is the number of places in
which x i and xi differ.
Definition 3: The minimum distance d of the RRNS is
Any integer X in the range [0, M ) can be uniquely repre- defined as
sented by a k-tuple x = [ x,,x2,.. . , x,],where
d = min{d(x,, x i ) : x i , x j € L ? , x i # x i } . (6)
x,= X (mod m , ) . (2) For an RRNS code, the minimum distance d is the same as
The integer x,, called the ith residue digit, is the least the Hamming weight of the codevector in fl having the
non-negative remainder obtained upon dividing X by m , . smallest positive Hamming weight,
The set of k-tuples and the interpretation function assigning
d=min{wt(x): XEQ, x#O}. (7)
to each k-tuple an integer in the range [0, M ) , and vice
versa, define the residue number system (RNS) of moduli Theorem I : The minimum distance of an RRNS code is d
m , ,m 2 , m k . From the Chinese remainder theorem
a ,
if and only if the product of redundant moduli satisfies the
(CRT), for any given k-tuple [ x,,x2,.* . , x,]. where 0 5 following relation:
x, < m,, there exists one and only one integer X such that
0 5 X < M and x,3 X (mod m,). The numerical value of
X is computed using
k For MDS-RRNS, d - 1 = n - k , and the moduli satisfy
X= n x,7',M, (mod M )
I= I
(3) M , = max{HYL,'m,,}, where 1 I j, I n.
Theorem 2: The error detecting capability, I , of an RRNS
where Q is d - 1.
Theorem 3: The error correcting capability, t , of an
M RRNS Q is [ ( d - 1/2)], where [ a ] denotes the largest
M, = - (4)
mi integer less than or equal to a .
Theorem 4: An RRNS is capable of correcting h or fewer
7;Mi = 1 (mod m i ) . (5) errors and simultaneously detecting /3 (/3 > h) or fewer
For error control (both correction and detection are included), errors if, d 2 h + p + 1.
we are concerned with the redundant residue number system It will be seen that for an ( n , k ) MDS-RRNS code, any
(RRNS) obtained by appending n - k additional moduli, single or double-residue errors can be corrected, and /3 or
m k f , , m k + Z , * m,, called the redundant moduli, to the fewer errors can be detected if the RRNS satisfies the follow-
RNS to form an RRNS of n positive pairwise relatively ing two conditions: 1) (Theorem 4) d - 1 = n - k = h + /3
prime moduli. The product n:= ,+ ,
m ,is denoted by M,. An where h = 1 or 2 and /3 > A; and 2 )
integer X in the range [0, M ) is represented as an n-tuple. min {m,.,mr,} > max {2m,,m,, - m,, mil}
;,
x = [ x,,x2; . x k , xk+ * . , x,], corresponding to n
e ,
-

moduli. The integers in the range [0, M ) in the RRNS are where k < r l , rz I n and 1 5 i , , i , 5 k. It is worthwhile
called legitimate integers and the corresponding n-tuples are t o note here that the condition in 2) above is a sufficient
called legitimate, while the n-tuples associated with the condition.
integers in the range [ M , MM,) are called illegitimate. The
integers in the range [ M , MM,) are defined as the illegiti- 111. CONSISTENCY
CHECKING
FOR RRNS

mate numbers. In [14], we define the RRNS as an ( n , k ) For an ( n , k ) MDS-RRNS code. assume that there are c
semilinear code L?, as its codevectors satisfy the property of and g errors (c + g 5 p ) in the information and parity
linearity under certain appropriately predefined conditions. digits, respectively. Then the received residue vector can be
All the legitimate numbers are valid, and the corresponding represented as
residue vectors are said to constitute the k-dimensional code
yuy = [Yl,Y?""? Yh-9 y,+[,'*',Yn]
space. Note that all the n-tuple residue representations form
an n-dimensional vector space. Every residue representation y , = x,,I r i s n , i # i a , jz,,
a = l ,2 , . . . , c ,
in the code space is a codevector that can be divided into two
v=1,2;.., g
parts; the first k residue digits corresponding to the k
nonredundant moduli are called the information digits, and Y.,,? + e,,,(mod mi,,), 0 < e,,b < m,,??
= x,,,
the remaining n - k residue digits corresponding to the l ~ i ~ ~ s k 1,. 2a;...= c
n - k redundant moduli are called the parity digits.
Some definitions and coding theorems in the RRNS [I41 + 0 < eJJ
y,,, = xJ, ejl(mod m,&,), m,",k + 1 Ij uIn ,
will now be given in order to make this paper self-contained.
v = I , 2;.., g (9)
Definition 1: The Hamming weight o f a vector x , wt(x ) .
in an RNS is defined as the number of nonzero components where i , , i2; . ., i,. are the positions of errors in the infor-
of x. mation digits, the corresponding error values being e,,,
20 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-11: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL.. 39, NO. I , JANUARY 1992

e;*, - eic respectively; and j , , j,, . . ., j , are the positions

a , and
of errors in the parity, the corresponding error values being c
e,,, e,*,-..., ejs, respectively. The altered information A, = (mod mk+,.),k +r # j,
number Y can be represented as

o , . . . , ~ elc,
, o,...,~] (11) where r = 1, 2;.*, n - k , and v = 1, 2 ; * * , g.
A fundamental property of the syndrome digits is stated in
where the following theorem.
Theorem 5: Under the assumption that no more than 0
X - [XI, X,,"'> Xk]
residues are in error, for the RRNS with d = h /3 1 + +
E tf [O; * e , 0, e,, , 0 ; * * , 0, e,>,0 ; * a , 0 , elc,0 , . . . , 0 ] . ( h < 0) one of the following four cases occurs.

Since E = 0 (mod m,)for all i # i,, Q! = 1, 2; * c, it is

-
Case I : If all the syndromes A I , . , A n - are zero, then
a ,
no residue is in error, and vice versa,
a Of all information except m , l , m ~ z ,' '' For the following cases, let p be the number of nonzero
and m,'.Therefore, E is an integer of the type syndromes.
M Case 2: If p I A, then exactly p corresponding parity
E = e' digits are in error, and all other residue digits are correct.
mi,miz mjc Case 3: If X + 1 Ip 5 0, then more than h residue
eia(mod mi=) (12) digits are in error.
3
+
Case 4: If p 1 I p In - k , then at least one of the
where 0 < e' < m ;,mi2 . . . m ;,. Based on [ yl, information digits is in error.
y,, . . * , yk], and the base extension (BEX) operation [2], the proof: The Proofs of Cases 1, 2, and 4 are given in [61
purpose of which is to avoid processing large valued inte- and [141. Therefore, they are omitted here. For Case 3, since
gers, the parity digits are recomputed to get k correct information residues uniquely determine the correct
- +
redundant residues, it is clear that if h 1 Ip 5 /3 and all
J'L+r = Y(mod mk+r) 3 = 3 2 * * * 3 n - k. (I3) the information residues are correct, then more than X parity
residues are in error. Now, suppose that c information
We define the test quantities A, called the syndromes as
residues are in error. We have c 2 1, i.e., at least one
A,. = yi+,. - y,+,.(mod mk+,),
r= 1 , 2; ..,n - k. information residue is in error. Define A> = yi+,. -
x,+,.(mod mk+,).Then, from (9) and ( l l ) , we have
(14)
Xtt [ X I >x ~ , ' ' x 'k ~? x k + l ' " ' ? XnJ
Since both X and E in (10) are less than M , we consider
the following two cases. rtt [ Y , , J'2,"',Yk, Y;+,~'.*>Y~J
+
i) X E < M . In this case, F-x~[o;..,o,~,,,o;..,o,~,~,o;..,o,
- M e l C , 0 , - . . , 0A',,
, A;;..,A',-,].
Y =X E =X e ' v + (15) + It is obvious that if the received redundant residue digiLyk+,.
na=lm,,
is erroneous, i.e., y k + , # x k + , . ,then A> # A r . For Y > X
and (the case for r < X can be analyzed in an analogous man-
M
r
ner), since M > - X , from (7) we know that at least
A,. e-' c (mod mk+r), + # j y
(16) d - c = X + P + l -cofA',,A',;*.,A',-,arenonzero.
Llmr, Therefore, we have the following three assertions.
M 1) If at most /3 - c redundant residues are in error, then
A,. e'- - e,+,.(mod rnk+,.) 7 k +r = J ', (17) at least h /3 + + 1 - c - ( p - e) = h 1 syndromes +
%= 1 m,,
are nonzero.
where r = 1, 2 ; * * , n - k , and Y = 1, 2 ; * * , g . +
2) If at most h 1 - c redundant residues are in error,
ii) M 5 X +
E < 2 M . In this case, then at least /3 syndromes are nonzero.
3) If at most h - c redundant residues are in error, then
- M at least +1 syndromes are nonzero.
Y = X + E - M = X + e'? -M
k=lmfn Now, for Case 3, if h + 1 to p syndromes are nonzero
c and no more than h residue digits are in error, then there is a
e'- nm,, (18) contradiction with the assertion 3). This gives the proof for
Case 3. This proves the theorem.
SUN AND KRISHNA: CODING THEORY APPROACH-PART 11 21

If Case 2 of Theorem 5 takes place, the new parity digit and

y i + r is the correct value of the erroneous parity digit y k + r .
If Case 4 of Theorem 5 takes place, a procedure to determine M
A, E (e?,.) - m,) -(mod rnk+'), r = 1, 2,.. e , n - k
error locations and error values for singleldouble errors is
mj
described in the following sections. This procedure is based
on an efficient search and computation of the error location (26)
and their values from the given set of syndromes for the
to obtain the values e:"" and where r = 1, 2 , . . , n
specified error correction/detection capability of the code.
- k . Compare to check if the solutions to (25) or (26) are
This procedure will give rise to a unique solution if and only e(I.1)= (1.21 = . . . = e y l or e(2.11 =
if there is a one-to-one correspondence between the errors e, J

that take place and the associated syndromes. In this regard,

,:2.2) = .. . = ,!,n - k l . If either of the two conditions is
satisfied, for J = I, then the Ith information digit is declared
necessary and sufficient conditions (as well as simplified
to be in error, the value of the error e, = e(M/m,)(mod
if e:l.l) = e,(1.2) = . . . = e ( 1 3 n - k ) ,and
sufficient conditions) are described, which the moduli consti-
tuting the RRNS must satisfy for this procedure to work. In
m,), where e = lI.I:, , .
e = ,:,.I) if e(2.1) = ( 2 . 2 ) = . . . = e f . n - k ) . The correct
the following sections, these conditions are described for the J el
value of the Ith residue digit is ( y , - e,) mod m,, and y is
specific case being analyzed. One of the reviewers presented
decoded to the codevector f , where
a proof of the sufficient condition for the above mentioned
-,
one-to-one correspondence for the general case of h random
x, = y J , j = 1 , 2 ; . . , n , j # 1 (27)
error correction and /3 random error detection in the review.
This proof is included in Appendix A. -,
x, = y, - e,(mod m,) . (28)
IV. A PROCEDURE FOR SINGLE-ERROR CORRECTION AND
It is clear from (21), ( 2 2 ) , and (25) that if one error takes
MULTIPLE-ERROR DETECTION (d = h + +
1 , h = 1,
place in the p t h information digit such that X E <M, +
P > A) then for ;= p, e = e' = e (Jl . l ) - (1,2)- . . . = e:l.n-kl
-
By Theorem 4, the RRNS with d = h /3 + +
1 , h = 1, and I = p. Similarly, if one error takes place in the p t h
and fi > h can simultaneously correct single residue error information digit such that M IX +
E < 2 M , then for
and detect P ( P = d - 2 ) errors. Assume that only the p t h j p, e = e' = = (2.21 = . . . = e ? , n - k ) and I =
~

e,
information digit is in error; then (15)-(20) become the from (23), (24), and (26). However, it remains to be shown
following. that if the following two cases occur: 1) only one error takes
i) X +
E < M . In this case, place in the pth information digit and j # p ; and 2 ) more
than one but less than d - 1 errors take place, then at least
-
Y = X + e'-M two of {e:'.r);r = 1, 2 ; - . , n - k } and at least two of
m, { r = 1, 2 ; * . , n - k } are unequal, i.e., (25) and (26)
do not give consistent solutions. Under an additional con-
and straint on the form of the moduli, the result is established in
the following theorem.
M
A e'-( m o d m , + , ) , r = 1 , 2 ; - . , n - k . (22) Theorem 6: If the moduli of an ( n , k ) RRNS code, are
mP such that there do not exist integers n,, n,; 0 In, < m,,
0 Inc < M , = n:=
m,e, 1 Ii , 5 k , 1 5 c 5 /3 that sat-
ii) M 5 X + E < 2 M . In this case, isfy
c+ I
(23) n,Mc + n,m, = n m , , k < r,
i= I
5 n,1 5j Ik (29)

and then for either of two cases, that is, Case 1, only the p t h
information digit is received in error and J # p; and Case 2 ,
M more than one (I 6 ) residue digits are received in error, the
Ar ( e ' - m )-( mod m k + r ) , r = 1 , 2 ; . . , n - k
solutions to (25) and (26) are inconsistent.
m,
Proof: Case 1: for X +
E < M , and J # p , if the
(24) solutions to (25) are consistent, i.e., e:'.') - ('a= . . -
-
e Y k ) = e, then comparing (25) to (22), we obtain
where 0 < e' < mp. Given A I , A 2 ; . . , A n - k , based on
(21)-(24) a procedure to correct single error and detect the M
presence of up to d - 2 errors can be outlined as follows.
For J = 1, 2 , . . k , solve the congruences
a ,
-(em,
mPmJ
- e'm,) = 0 mod
( n r=t
mk+r

that is, emp - e'm, is a multiple of IIFIFmk+r.This is not

A,
, M
e(1-r)-
(mod m k + r ) , r = 1 , 2 ; - . , n - k (25) possible as 1 emp - e'm, 1 < II;::mk+r. Similarly, for
mJ X + E < M , and J # p, if J = (2.2) = . = -
22 IEEE TRANSACTIONS O N CIRCUITS A N D SYSTEMS-11: ANALOG A N D DIGITAL SIGNAL PROCESSING. VOL 39. NO. I . JANUARY 1992

e y , n - k )= e, then comparing (26) to (22) we obtain algorithm to correct a single error and simultaneously detect
presence of multiple error in RRNS can be described as
follows.

Step I : According to the received vector, we compute the

This is not possible as (e'm, +
( m , - e)m,,) < 2 m p m , < syndromes.
n,=,
n-k
mk+,. The case for M IX +
E < 2 M , can be ana- Step 2: Check how many syndromes are zeros:
lyzed in an analogous manner. 1) If all the syndromes are zero, then no error
Case 2: for X +E < M , if e:'." = e:'.') = . . - -
occurs. Stop.
e Y k J= e, then comparing (25) to ( I 6) we obtain 2 ) If only one syndrome is nonzero, then only
one error in corresponding parity. Correct it
and stop.
3 ) If d - 3 to 1 syndromes are zero, then go to
Step 6.
where k < r, In , v = 1, 2, * . . , g. This congruence cannot 4) If all the syndromes are nonzero, then go to
hold as l e I I ~ = , m I - m e'm,I < m , I I : ~ = , m , <
~y
Step 3.
n8+1- g
,=, m , . For example, assume that the p t h information Step 3: Let j = 1.
digit and the first g parity digits are in error, and 0 = c g , + Step 4: Perform single-error consistency-checking for the
where c = 1 , then the above congruence is
nonredundant moduli m,. Check the consistency
M of the solutions. If it is consistent, then go to Step
-(emp - e'm,) = O(mod m,-,m,,) 5 ; if it is in consistent, then j = j + 1. Go to step
mJmP 4 for j Ik . For j = k + 1, go to Step 6.
This is not possible as 1 emp - elm, I < m n - , m n . In other Step 5: Only one error in the j t h position. Correct it and
words, at least two of { e:'.r); r = 1 , . . . , n - k ) are un- stop.
equal. Similarly, for X + E < M , if e:2." = = ... Step 6: Declare more than one error detected and stop.
= e ; 2 . n - k ) = e, comparing (26) to (16), we obtain
A flowchart for this algorithm is given in Fig. 1.
Example I : Consider a (10, 6) RRNS code based on the
moduli m , = 23, m 2 = 25, m 3 = 27, m4 = 29, ms = 31,
m6 = 3 2 , m, = 67, m 8 = 71. m, = 73, m,, = 79, where
n-k-g m,, m8.m,, and m,, are the redundant moduli. Clearly,
m,m8 > 2m,m, - m5 - m,, therefore, the condition in
Theorem 6 and Lemma 1 is satisfied. M = n;=,
rn, = 446,
623, 200. Let X = 400, 000, 000. then x = [8, 0, 22, 13,
where k < r , In , v = 1, 2;. g. The above congruence
e , 2.5, 0, 17, 58, 4, 1 I ] . Since 0 IX < M , x is a codevector.
cannot hold as ( e l m j +( m i - e)13',=,mi,t)< 2 m j Assume that one error takes place in the first residue digit,
n : = l m i < v< Elf=+l'-grnr,, +
for c g < 0;and (29) for c g + and the received vector is [O, 0, 22, 13, 25, 0, 17, 58, 4, 1 I].
= 0.The case for M 5 X +E < 2 M can also be analyzed Based on the information part [O, 0, 22, 13, 2.5, 01 and BEX
in an analogous manner. This proves the theorem. operation, we compute the syndrome digits A , E 14(mod
Lemma I : If the moduli of RRNS satisfy the condition 67), A , E 48(mod 71), A 3 = 50(mod 73), A 4 = 17(mod
79). Following the decoding algorithm, we check the consis-
min ( m , , m r 2 } > max { 2 m i , m i 2- m i , - m , , ] (30) tency for j = 1 , 2, . . , k . k = 6 in this example. Based on
computing e:"'" and r = 1 , 2, 3, 4, it is seen that
where k < r l , r 2 I n and 1 5 i , , i, 5 k , then for either of when j = 1, e ( 2 . 1 ) - e ( 2 . 2 ) - e (I? . 3 ) -
- = e = 14 is a
I - I -
the two cases, that is, Case 1 , only the pth information digit
consistent solution to (26). So, declare that one error oc-
is received in error and j # p ; and Case 2 , more than one
curred in the first residue digit, the value of the error being
residue digit is received in error, the solutions of either (25)
e , = e ( M / m , ) ( m o d m , ) = 15. From (28), the estimated
or (26) are inconsistent.
value of the erroneous digit is given by jc? = 8.
Proof: The condition is sufficient since, if the moduli
Now, assume that two errors take place in the first and
satisfy (30),then (29) is satisfied trivially. This completes the
third residue digits, and the received vector is [O, 0, 23, 13,
proof.
25, 0, 17, 58, 4, 111. Based on the information part [O, 0,
It is interesting to note that this sufficient condition on the 23, 13, 25, 01 and BEX operation, we compute the syndrome
moduli for single error correction and multiple error detec- digits A , = 23(mod 67), A 2 = 50(mod 71), A 3 = 63(mod
tion is the same as the sufficient condition on the moduli for 73), and A4 = 6S(mod 79). Following the decoding algo-
single error correction [14]. Based on the concept developed rithm, we check the consistency for i = 1, 2;.., 6, it is
above, and under the assumption that the moduli rnl, i = 1 , found that there is no consistent solution, so declare more
2 , . * , n , satisfy the condition in Theorem 6 or Lemma 1, the than one error detected.
SUN AND KRISHNA: CODING THEORY APPROACH-PART 11 23

ACCORDING TO THE d = Atfi+l where 0 < e' < m P m 4 . Similar to the analysis in Section
RECEIVED VECTOR. dL4,h=l,P>h
COMPUTE THE
Note Under the assumption
IV, given A , , A , ; . . , A n P k , based on (31)-(34) a proce-
SYNDROMES
dure to determine error locations and error values can be
outlined as follows.
For i = 1, 2 ; * . , k - 1, j = i + 1, i + 2 ; * * , k , solve
the congruences

and

I CONSISTE NCY-CHECK1 NG
AT THE NONREDUNDANT
MODULUS ml
CHECK HOW MANY OF

r = 1,2;..,n-k (36)
to get the values
AND STOP

ARE DETECTED
STOP (37)

Fig. 1. A decoder flowchart for single-error correction and multiple-error

detection.

where r = 1, 2, 3, 4. Based on {e".'), e ( i ~ 2 ) }and {e(23'),

V. A PROCEDURE FOR DOUBLE-ERROR CORRECTION A VD and by the CRT, we can write
MULTIPLE-ERROR DETECTION ( d = h 0 1, h = 2 , + +
P > A)
If Case 4 of Theorem 5 happens, i.e., exactly two residue
digits are received in error, there are two possible cases; that
is, Case a): both errors in information part; Case b): one
error in information part and the other in parity part. By
Theorem 4, the RRNS with d = h 0 1, h = 2 , and + + where
P > h can simultaneously correct two errors and detect P
errors. Note that for correcting two errors only, i.e., for
d = 5, the analysis of the error correcting procedure is the
same as in this section.
Consider Case a): if the p t h and qth information digits are
received in error, (15)-(20) become the following.
i) X +
E < M . In this case, Then check if the following two conditions are satisfied: i)
- M el:) < m,m, and (35) holds by substituting el;.) for e(',r),
Y = X + E = X + e'- (31) i.e., the solutions to (35) are consistent, and the consistent
mPm4 solution is e!)); ii) e;:) < mim, and (36) holds by substitut-
and ing e::) for e"," , i.e.,
' the solutions to (36) are consistent,
and the consistent solution is e$). Note that { e('*3),e(',4)}
and { e",')} will be used for obtaining the consistent
solution for Case b). Let us say either of the two conditions is
satisfied for i = f and j = h . Then the fth and hth informa-
ii) M 5 X + < 2 M . In this case, tion digits are declared to be in error, the value of errors
- M M being
Y =X + e'- -M = X + (e' - m P m q )__- M
mPm4 mPmQ
e.f =
- e-(mod mf) (43)
(33) mJmh

M M
A, = (e' - m,m,) -(mod mk+r), eh = e-(mod mh) (44)
mPm4 mfmh

= 2 7* ' ' 9 - (34) where e = e:;), if e:).' is the consistent solution to (35);
24 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-11: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 39, NO. 1 , JANUARY 1992

e = e!;), if ej;) is the consistent solution to (36). Note that if and

e = O(mod m,) or e E O(mod m,), then only the hth or fth M
information digit is in error. The correct values of the fth A = e'-( mod m k + r ) , r # U
and hth residue digits are yf - ef(mod m,) and y , - mP
e,(mod m,), respectively, and y is decoded to the codevec-
tor x, where
/,
x j = y,, j = 1 , 2;.., n, j f f,h (45)
A

x, = y, - ef (mod m,) (46)

A
ii) For M r X + E < 2 M , r = 1, 2 ; . . , n - k, and
x h = y h - e,(mod m,). (47) Isurn-k,

It is clear from (31), ( 3 2 ) , and (35) that if two errors take

place in the p t h and 9th information digits such that X E+
< M , then for i = p and j = q , e = e' = e:;),f = p , and and
h = q . Similarly, if two errors take place in the p t h and qth
information digits such that M I X + E < 2 M , then for
i = p and j = q , e = e' = e!;), f = p, and h = q from
(33), (34), and (36). However, it remains to be shown that
the solutions to (35) and (36) cannot be consistent, if any one
of the following cases occurs: 1) only one error takes place in
the pth ( p # i, j ) information digit, 2) exactly two errors
take place in the p t h and qth information digits, where M
A, = ( e' - mp)- - ek+u(mod mk+u) (54)
p # i, j and/or q # i, j , 3) exactly two errors occur: one in mP
the information part and the other in the parity part, 4) more
+
than two (less than 0 1) errors occur. Under an additional where 0 < e' < mp, and m 4 can be any one of the informa-
constraint on the form of the moduli, the result is established tion moduli. Based on (49)-(54), the procedure to determine
in the following theorem. error locations and error values is the same as that for Case
Theorem 7: If the moduli of an ( n , k) RRNS code are a), except that when no consistent solution is found in Case
such that there do not exist integers n,,, n,; 0 5 n,, < a), then we check if any n - k - 1 congruences in either
m,m,, 0 I n , < Me = II:=,rnlu, 1 Ii, I k, 1 Ic I0; (35) or (36) have a consistent solution, say e, which satisfies
that satisfy e < m i m j and e E O(mod m,) or (mod mj). Note that this
consistent solution may be obtained from the combination
Cf2 { or { e(2,3),e(2,4)),
since one error may occur in
nijMe + n,mimj = n m , , k < ri I n , 1
i= 1
Ii , jIk +
the ( k + 1)th or ( k 2)th digit. Let us say the above
condition is satisfied for i = f and j = h , and the only
(48) inconsistent congruence is corresponding to the ( k s)th +
residue digit, then the fth (if e = O(mod m,)) and ( k s)th +
then for any one of the following three cases: 1) only one residue digits are declared to be in error. The correct value
error takes place in the p t h ( p # i, j ) information digit, 2) of the fth and hth residue digit can be obtained from (43)
exactly two errors take place in the p t h and qth digits, +
and (44), and the correct value of the ( k s)th residue digit
where p # i, j and/or q # i, j , 3) more than two (less than is as follows:
+
0 1) errors occur, the solutions to (35) and (36) are
/-.. M
inconsistent.
The proof of this theorem is similar to the proof of
xk+s E Yk+$ + As - mk+s) (")
mfmh
Theorem 6, and therefore, is omitted here. The sufficient
condition for Theorem 7 can be shown to be exactly the same if n - k - 1 congruences in ( 3 5 ) have a consistent solution
as the sufficient condition stated in Lemma 1. e; and
Now, consider Case b): assume that one error takes place
in the pth information digit and the other takes place in the
+
( k u)th digit (the uth parity digit); then (15)-(20) become
the following.
i) For X +E < M , r = 1, 2 ; - . , n - k, and 1 5 U I
if n - k - 1 congruences in (36) have a consistent solution
n - k,
e. It is clear from (49)-(51) and (35) that if one error takes
- M place in the pth information digit and the other takes place in
Y=X + E = X + e'- (49) the ( k + u)th digit such that X + E < M , then for i or
mP j = p , say i = p , e = = e'm, = O(mod m,), f = p ,
SUN AND KRISHNA: CODING THEORY APPROACH-PART 11 2s

and s = U , where r = 1, 2;.., n - k , and r # U . Simi- it is consistent, then go to Step 6; if it is inconsis-

larly, from (52)-(54) and (36) if one error takes place in the tent, then go to Step 7.
p t h information digit and the other takes place in the ( k + Step 6: If the consistent solution e = O(mod m i ) (or e =
u)th digit such that M IX +
E < 2 M , then for i or j = p , O(mod m j ) ) ,then only one error in the jth (or
say i = p , e = e(2")= elmj = O(mod mi), f = p , and s ith) position; otherwise, exactly two errors in the
= U , where r = 1, 2;-., n - k , and r # U . However, it ith and jth positions. Correct them and stop.
remains to be shown that neither (35) nor (36) has exactly Step 7: If exactly d - 2 solutions are consistent and the
n - k - 1 consistent solutions that are equivalent to zero consistent solution e = O(mod m,) (or e = O(mod
mod mi or mod m j , if any one of the following cases m j ) ) ,then exactly two errors: one in the jth (or
occurs: i) only one error takes place in the p t h ( p # i, j ) ith) position and the other in the parity for which
information digit, ii) exactly two errors take place in the pth the solution is inconsistent. Correct them and
and qth information digit ( p # i, j , and/or q # i, j ) , iii) stop. Otherwise, go to Step 8.
+
exactly two errors take place in the p t h and ( k u)th digit Step 8: j = j + 1, go to Step 5 for j I k . For j > k ,
( p # i, j ) , iv) more than two (I /3) errors take place. +
i = i 1 go to Step 4 for i 5 k - 1. For i = k ,
Under the same constraint on the form of the moduli as stated go to Step 9.
in Theorem 7, the above result is established in the following Step 9: Declare more than two errors detected and stop.
theorem.
A flowchart for this algorithm is given in Fig. 2.
Theorem 8: If the moduli of an ( n , k ) RRNS code are
Example 2: Consider the same RRNS as in Example 1,
such that there do not exist integers n i j , n,; 0 5 n i j <
based on the moduli m , = 23, m2 = 25, m3 = 27, m4 =
m i m j , 0 In , < M , = II:=,rnjm, 1 Ii, i k , 1 Ic I/3
29, ms = 31, m6 = 32, m7 = 67, m , = 71, m , = 73, m,,
that satisfy
= 79, where m7, m a , m Q ,and m ,._
- ,
, are the redundant
c+2 moduli. Clearly, m 7 m , > 2m,m6 - m5 - m,; therefore,
nlJMc+ n,m,m, = n m , ,k <
I= 1
TI 5 n, 1 5 i, j Ik the condition in Theorems 7 and 8 is satisfied. M = rI;=,rn,
= 446, 623, 200. Let X = 400, 000, 000, then x = [8, 0,
(57) 22, 13, 25, 0, 17, 58, 4, 111. Since 0 5 X < M , x is a
codevector. Assume that two errors take place in the first and
then when (35) or (36) has exactly n - k - 1 consistent
third residue digits, and the received vector is [0, 0, 23, 13,
solutions equivalent to zero mod m j or mod m i , the ith or
25, 0, 17, 58, 4, 111. Based on the information part [0, 0,
j t h information digit and the parity digit corresponding to the
23, 13, 25, 01 and BEX operation, we compute the syndrome
only inconsistent solution are received in error.
digits A , = 23(mod 67), A, = 50(mod 71), A 3 = 63(mod
The proof of this theorem is similar to the proof of
73), and A4 = 65(mod 79). Following the decoding algo-
Theorem 6, and therefore is omitted here.
Based on the concept developed above, and under the
rithm, we check the consistency for i = 1, 2; * 5, j = i
a , +
assumption that the moduli of MDS-RRNS satisfy the neces-
+
1, i 2, , 6, based on computing {e".'), e(',,), e(133),
{ e(*,'), from (37) and (38) and
sary and sufficient condition in Theorem 7 or 8, or the
{el;),el;)} from (39) and (40). It shows that when i = 1 and
sufficient condition in Lemma 1, the algorithm to correct
j = 3, er:) = 217( < m l m 3= 621) is a consistent solution to
double-error and simultaneously detect multiple errors can be
(36), and e!:) f O(mod 23) or (mod 27). So, declare two
described as follows.
errors occur in the first and third residue digits, and from
Step 1: According to the received vector, we compute the (43)-(47) the estimated values of the erroneous digits are
syndromes.
XI = y
+. , - e(2)-M (mod m , ) =8
Step 2: Check how many syndromes are zero: 13 m l m3
1) If all the syndromes are zero, then no error
occurs. Stop.
2) If only one syndrome is nonzero, then only
one error in corresponding parity. Correct it
and stop.
Now, assume that two errors take place in the first and
3) If two syndromes are nonzero, then two errors
in parity. Correct them and stop.
+
( k 4)th residue digits, and the received vector is [0, 0, 22,
13, 25, 0, 17, 58, 4, 121. Based on the information part [0,
4) If d - 4 to 2 syndromes are zero, then go to
0, 22, 13, 25, 01 and BEX operation, we compute the
Step 9.
syndrome digits A , = 14(mod 67), A, = 48(mod 71), A 3 =
5) If none or one syndrome is zero, then go to
50(mod 73), A4 = 16(mod 79). Following the decoding algo-
Step 3.
rithm, we check the consistency for i = 1, 2,. . . , 5 , j = i +
Step 3: i = 1. +
1, i 2,. . ., 6, based on computing {e''.'), e(',,), e(193),
Step 4: j = i 1.+ { from (37) and (38) and
Step 5: Perform double-error consistency-checking for all {ell.),el:)} from (39) and (40). It shows that when i = 1 and
the combination of the nonredundant moduli mi j = 2, e!:) = 350 = O(mod m,) < m , m , is a consistent so-
and m j . Check the consistency of the solutions. If lution to the first three congruences of (36), i.e., the only
26 IEEE TRANSACTIONS ON CIRCUITS A N D SYSTEMS- U: ANAI.0G AND DIGITAL SIGNAL PROCESSING. VOL 39. NO. I . JANUARY 1992

d = A+P+I
ACCORDING TO THE d 25. A=2, P>h this decoding procedure in our subsequent analysis. Here
R ~ ~ Note
~ Under
~ the~ assumption
~ ~ f P(A)~ denotes
~ R the ~probability of occurrence of event A. The
SYNDROMES
that no more than d-3
residues a r e i n e r r o r proof of Theorem 9 is similar to the proof of lemma 3 in
NO ERROR
TWO ERRORS [ 141, and therefore, is omitted here.
CHECK STOP
Theorem 9: For detecting all single-burst errors of length
HOW MANY
b or less, an RRNS code must have d >_ b + 1.
Theorem IO: For correcting all single-burst errors of
length b or less, an RRNS code must have d 2 2 b + 1.
Proof: Let x, be a codevector other than x in !d. The
Hamming distance among x , x,, and y satisfies the triangu-
lar inequality
PERFORM THE DOUBLE-ERROR
CONSISTENCY-CHECKING FOR THE
COMBINATION OF mi AND m,
CHECK HOW MbNV S O L U T l O N S I R E
d ( Y , x) + d(Y, x,) 2 d ( x , X I ) .
CONSISTENT
EXACTLY TWO
Since d ( x , x,) 2 d and d ( y . x ) = w t ( e ) 5 b , ( e = y -
EXACTLY TWO ERRORS ONE
X),
ithANDjth i t h POSlTlON
POSIT ION d ( y , x) 2 d - b.
CORRECT T H E M
STOP

I ,-,+I I
CORRECT THEM
STOP
.. If d 2 2 6 + 1 , then d ( y . x,) 2 b +
1. thereby implying
that the length of burst-error vector e is smaller than the
length of any other burst-error vector e , , e, = y - x,. On

Denotee=O(mod m,)
by lelmi-o, where
e i s the consistent
1=1+1 4
MORE THAN TWO
ERRORS A R E
I the other hand, if the codevector x is such that wt(x) = d ,
and the nonzero digits of x are confined to d consecutive
places (examples of two such codevectors are codevectors
solution DElECTED STOP corresponding to the integers X , = n:l,’m, and X , =
2nf;=-,’in,, respectively), then for d 5 2 6 , we can show that
Fig. 2. A decoder flowchart for double-error correction and multiple-error there exists at least one codevector x, such that d ( y , x ) =
detection. w t ( e ) 2 w t ( e , ) . e , = y - x,. In this case, incorrect decod-
ing will take place. This proves the theorem.
inconsistent congruence is corresponding to the modulus The proof of‘ Theorem 11 is similar to the proof of
m l 0 .So, declare that two errors occur in the first and tenth Theorem IO, and therefore. is omitted here.
residue digits, the estimated values of erroneous digits being Theorern 11: For correcting all single-burst errors of
given by (46) and (56), that is, length b or less and simultaneously detecting all burst errors
of length b’ ( b ’ z b ) or less, an RRNS code must have
/\ M drb+b’+l.
x, E y , - e!;)- (mod m , ) = 8
mlm2 A fundamental property of the syndrome digits is stated in
the following theorem.
,’\ M
x , =~ y,, + A 4 + ( m , m , - e!<’)-- (mod m,”) 11
k +
Theorem 12: For an MDS-RRNS ( n , k ) code, ( d = n -
1=b + + b‘ l), under the assumption that no more
mlm2
than one single-burst residue error of length 5 b’ occurs,
Note that i is always equal to 1, if two errors occur in a way one of the following four cases occurs.
that one error is in information part and one error is in parity
part.
,
Case I : If all the syndromes A , A 2 , * . an^ are zero,
a ,

then no residue is in error, and vice versa.

VI. A PROCEDURE FOR SINGLE-BURST
CORRECTION
A single-burst residue error vector of length b is defined
as an RNS vector whose nonzero residue digits are confined
to b consecutive digits, the first and last of which must be
nonzero. The decoding procedure for performing burst-error
correction depends on the criterion of performance. If residue
errors occur in burst such that
P(burst error of length a ) < P(burst error of length b )
a > b, and the criterion of performance is to minimize the
probability of decoding error (maximum likelihood decoding
(MLD)), then the decoding procedure can be summarized as
follows. Decode y to a codevector x for which the length of
burst-error vector e, e = y - f , is minimum. We will use
ERROR For the following cases, let p be the number of nonzero
syndromes.
Case 2: If p 5 b syndromes are nonzero, then exactly p
corresponding parity digits are in error, and all other residue
digits are correct.
+
Case 3: If b 1 5 p 5 6‘syndromes are nonzero, then
more than b residue digits are in error.
+
Case 4: If 6‘ 1 5 p 5 n - k syndromes are nonzero,
then at least one of the information digits is in error, and all
the last b + 1 redundant residue digits are correct.
The proof of this theorem is similar to the proof of
Theorem 5 , and therefore. is omitted here. If Case 4 of
Theorcm 12 happens, and exactly one single-burst residue
error of length Ib occurs, there are two possible cases,
that is, Case a): all the errors are in the information part;
Case b): the errors are in both information and parity part.
SUN A N D KRISHNA: CODING THEORY APPROACH-PART I1 27

Consider Case a). Assume that a single-burst error of a consistent solution to (63). In other words, check if e:') <
length c (1 Ic I6 ) takes place between the ith and ( i c + nfi=Imi+a- and (62) holds by substituting e:') for or
- 1)th information digits, i.e., in the position interval [i, e?) < IIZ=,m,+,- I , and (63) holds by substituting e?) for
+
i c - 11, then (15)-(20) become the following. e?,r). Let us say either of the two conditions is satisfied for
+
i) X E < M . In this case, j = p. Then declare that a single-burst error of length Ib
M +
occurs in the position interval [ p , p b - I]. The values
Y=X+E=X+e' (58) of the errors are
ne=1m;+a- I

and
M
Ar = e' (mod m k + r ) ,r = 1, 2 ; * - , n - k.
=
:' Imi+a- 1

(59)
where e = e'') if e$') is the consistent solution to (62);
ii) M IX + E < 2 M . In this case, e = e?) if e$) is the consistent solution to (63). The correct
- M values of the erroneous residue digits are ( y , - e,) mod
Y=X+E-M=X+ m,,whereq=p, p + l ; * * , p + b - l . T h e n yisde-
coded to the codevector x,where
(60)
and
A
xq = y q , q = 1,2;.., + l;**,p + b - 1
n, q f p , p
M m , ) , q = p , p + l ; . . , p + b 1.
A
x, = y q - e,(mod -
(mod mk+r)j
(70)
r = 1,2;..,n-k (61)
Note that e , = O(mod m,) implies that no error occurs in
where 0 < e' < n : = l m i + a - l . Given A , , A , ; . . , A n - k ,
the qth digit. It is clear from (58), (59), and (62) that if a
based on (58)-(61) a procedure to determine error locations single-burst error of length c ( c 5 b) takes place in the
and error values can be outlined as follows.
For j = 1, 2,. . . , k - b + position interval [i, i c - 11 such that X +
E < M , then +
1, solve the congruences for [ j , j +
b - I] including [i, i c - 11, e'm = e:'), +
where m is the product of the moduli corresponding to the
position interval [ p, p b - 13 exclusive of [ i, i c - 11.+ +
Note that if c = b, then e' = e$') and p = i. Similarly, if a
r = 1 , 2 ; * . , n - k (62) single-burst error of length c ( c Ib ) takes place in the
position interval [i, i c - 11 such that M IX E < + +
2 M , then for [ j , j +
b - 11, including [i, i c - 11, +
e'm = e?) from (60), (61), and (63). For this case, if
c = 6 , then e' = e?) and p = i. However, it remains to be
r = 1 , 2 ; . - , n - k (63)
shown that neither (62) nor (63) has a consistent solution, if a
to get the values and e?,'). Using the last b digits of single-burst error of length c (c Ib') takes place in the
r = 1, 2, * * n - k } and the last b digits of {
{ ejl,r); e , +
position interval [i, i c - 11 which is not included in [ j ,
r = 1, 2; * * , n - k } , and by the CRT we compute the j +
b - 11. Under an additional constraint on the form of the
following values: moduli, the result is established in the following theorem.
n-k
Theorem 13: If the moduli of an ( n , k ) RRNS code are
e:) = ey*r)7''Mi (mod M i ) (64) such that there do not exist integers n B , n,; 0 In B < M B
- nR=lm,e, o In , < M , = n z = l m i a ,1 s j a , i, < k ,
r=n-k-b+ 1
n-k 1 Ic Ib'; that satisfy
e?) e?.')T;M; (mod M i ) (65)
r=n-k-b+ 1 c+ b

where
n B M c + ncMB = n m r , ,k < r,
I= 1
In (71)

+
then for [i, i c - 11 not included in [ j , j +
b - 11,
neither (64) nor (65) has a consistent solution, where a
M ' = - Mi single-burst error of length c (1 Ic I6') occurs in the
mk+r +
position interval [ i , i c - 11.
T M ; = l(mod m k + r )
Proof: Consider the extreme case, that is, [ i, i c - 11 +
(68) and [ j , j +
b - 11 do not overlap and c = b'. For X E +
and then check if ejl:' is a consistent solution to (62) or is < M , if (62) has a consistent solution e$'),then comparing
28 IEEE TRANSACTIONS ON CIRCUITS A N D SYSTEMS-11: ANALOG A N D DIGITAL SIGNAL PROCESSING, VOL. 39, NO. I , JANUARY 1992

(62) to (59) we get

i
b b'
e' II m j + , - l - el1)aII
a=1 =l

where e' < II%= , m i + , - This cannot hold due to

r = c2 + l ; . . , n -k (77)

i II
b b'
e' m j + a - l - e;.')II m i + a - l where e' < II:= k - cl + I m ;. A procedure to determine error
a=l a= 1 locations and error values is described in the following.
b 6' n-k For j = k - b + 1, check if there exists the largest num-
< n
a= I
mj+ol-l
a= 1
mr+u-l < II m k + r .
r= I
ber c, in the range [I, b ) such that either of the following
two conditions is satisfied.
Similarly, for X +
E < M , if (63) has a consistent solution +
Condition 1): The last (b' c,) consecutive congruences
e?), then comparing (63) to (59) we get of (62) have a consistent solution, i.e..

where e' < n ab' = , m i + a - lThis

. cannot hold due to (71). The
case for M IX +
E < 2 M , can also be analyzed in an
analogous manner. This proves the theorem.
A sufficient condition for Theorem 13 can be shown and
to be exactly the same as the sufficient condition stated in
Lemma 1.
Now, consider Case b): In Case a), when j = k - b 1 +
if we do not find the consistent solution, it implies that a
Condition 2 ) : The last (b' + c3) consecutive congruences
burst error of length > b occurs or a burst error of length
of (63) have a consistent solution, i.e.,
5 b occurs in both the information and parity digits. Assume
that a single-burst error of length c (1 Ic Ib ) takes place
in the position interval [ k - c1 1 , k +
c 2 ] , where c l , +
+
c2 2 1 and c1 c2 5 b , then (15)-(20) become the follow-
ing.
i) X +
E < M . In this case,

and and

Let us say there exists a number c, such that either of the

M above two conditions is satisfied; then declare that a single-
burst error of length Ib takes place in the position interval
A e'
=: k -c, +I i
(mod mk+r),
+
[ k - c3 1, k + b - c 3 ] .The estimated values of the er-
rors and erroneous information digits can be obtained by (69)
r c2 + I ; . . , n - k.
= (74) and (70) for p = k - b + 1. The inconsistent congruences
ii) M 5 X + E < 2 M . In this case, correspond to the erroneous parity digits, y k + s , s = 1,
2,. . b - c3,the estimated values of erroneous parity digits
F=X
e ,

+ E - M= X being:
SUN AND KRISHNA: CODING THEORY APPROACH-PART I1 29

if e:’) is the consistent solution to (62); and +

4) If 6’ I to all syndromes are nonzero, then
go to Step 3.
Step 3: Let j = I .
M Step 4: Perform b-error consistency-checking for the
(mod (79) combination of the nonredundant moduli m,,
‘:=lmk-b+a mj+ I , . . . , mjsb- ,. Check the consistency of the
if e?’ is the consistent solution to (63). solutions. If it is consistent, then go to Step 5; if
It is clear from (72)-(74) and (62) that if a single-burst it is inconsistent, then for j < k - b +
1 go to
error of length c (1 5 c 5 6) takes place in the position Step 6 , go to Step 7 for j = k - b 1. +
interval [ k - c, + +
1, k +
c,] such that X E < M , where Step 5: A single-burst error in the position interval [ j ,
+
c,, c2 L 1 and c, c2 = c, then for j = k - b 1, (62) + j +
b - 11, correct it and stop.
+
has 6’ b - c, solutions, which are consistent and equiva- Step 6: j =j + 1, go to Step 4.
lent to zero mod m k - b + o i , i.e., = e$’’ = Step 7: +
If the last b’ c consecutive congruences of (62)
b-c,
e n , = , m k - b + a , where CY = 1, 2; * b - C , . r = c, 1,
e , + or (63) have a consistent solution which is equiva-
+
c, 2 ; * * , n - k , b - c2 = c 3 , and c, = c4. Similarly, lent to zero modrIi:Cmk-b+a, then a single-
from (75)-(77) and (63), if a single-burst error of length c burst error exists in the position interval [ k - c
(1 5 c Ib ) takes place in the position interval [ k - cl 1, + + 1, k + b - c]. Correct it and stop. Otherwise
+
k c,] such that M I X + E < 2 M , where c , , c2 2 1 go to Step 8.
and C , + +
c, = c, then for j = k - b 1, (63) has b’ b + Step 8: Declare that a burst-error of length > b is de-
- c, solutions, which are consistent and equivalent to zero tected. Stop.
b-c,
mod m k - b + a , i.e., e?..) = e?’ = e n , = , m k - b + a , where A flowchart for this algorithm is given in Fig. 3.
CY = 1 , 2 ; - * , b - c,, r = c2 + 1, c, +
2;*., n - k, b -
c2 = c 3 , and c, = c4. However, it remains to be shown that OF PREVIOUS
VII. EXTENSIONS ALGORITHMS
no c3 and c, exist to satisfy Condition 1) and 2 ) , if a
In [4] and [8]-[lo], the algorithms for locating a single
single-burst error of length > b (but Ib’) occurs or a residue digit error are based on the properties of modulus
single-burst error of length Ib occurs not in the position projection and MRC. MRC is an operation to represent the
+
interval [ k - b 2 , k + b - 11. The result is established integer X in the form
in the following theorem.
1-1
Theorem 14: If the moduli of ( n , k ) RRNS code are such fl

that there do not exist integers n B , n,; 0 In B < M B = x = IC= ]a , rr =Il m r (81)
IItzIrnjm,0 5 n , < M , = H:=lmicy, 1 Ij a , i, < k , 1 I
c Ib’; that satisfy w h e r e O I a , < m , , I = 1, 2 ; . . , n , a n d n ; = , m , = 1. In
c+ b
an ( n , k ) RRNS code, the first k mixed radix digits a , ,
nBMc + ncMB = n m , , k < r,
;= I
In (80) a 2 ,* ak are called the nonredundant mixed radix digits,
a ,

and the rest a k + a k + , , . . * , a, will be called the redundant

then when a single-burst error of length > b (but I6’) mixed radix digits. It is obvious that if the redundant mixed
occurs or a single-burst error of length Ib occurs not in the radix digits are all zero, the number X is a legitimate
+
position interval [ k - b 2, k +
b - 11, c3 and c4 do not number; otherwise, X is an illegitimate number. The modu-
exist to satisfy Conditions 1) and 2). lus m, - projection of X in an ( n , k ) RRNS code, denoted
The proof of this theorem is similar to the proof of by X , , is defined as
Theorem 13, and therefore, is omitted here. Based on the
concept developed above and under the assumption that the
moduli satisfy the condition in Theorem 13 or the sufficient
X, =X [ mod- M m T ]

condition in Lemma 1, the algorithm to correct single-burst

that is, X , can be represented as [x,,xZ;’ * , xi-,,
error of length b or less and simultaneously detect the
presence of single-burst error of length b’(b’ > b ) in RRNS x,, ,, x,] which is the residue representation of X in a
a ,

can be described as follows. reduced RRNS with the ith residue digit xi deleted. The
mixed radix representation of Xiis
Step 1: According to the received vector, we compute the
fl 1-1
syndromes.
Step 2: Check how many syndromes are zero.
xi = aln mr
r=l
I#; r#i
1) If all the syndromes are zero, then no error
occurs. Stop. where the new first k mixed radix digits are still called the
2 ) If p (1 Ip 5 b ) syndromes are nonzero, then nonredundant mixed radix digits and the rest are called the
exactly p corresponding parity digits are in redundant mixed radix digits, e.g., if i 5 k , a , ,
error. Correct them and stop. ,
a 2 , . . , ai- , ai+,,. . . , a k , ak ,
are the nonredundant
+

+
3) If b 1 to b‘ syndromes are nonzero, then go -
mixed radix digits, and a k + * , a k f 3 , . , a, are the redun-
to Step 8. dant mixed radix digits. It is obvious that if X,is a legiti-
30 IEEE TRANSACTIONS O N CIRCUITS A N D SYSTEMS-11: ANALOG A N D DIGITAL SIGNAL PROCESSING, VOL. 39, NO. 1 , JANUARY 1992

d=b+brt
According to the b ,b i.e.,
received vector
compute the Note Under the assumptionthat k+c
no more than one single-
burst residue error of length
< b occurs
M' = n m,, i ,
r= I
5 k + c , i,,, >k + c , i , Ik
r@
k
M' = M = n m,, i, > k .
r= 1
(86)

It is obvious that M' 2 M . M' is also the smallest nonzero

number represented by (85) with all the nonredundant mixed
for the combination radix digits zero. It follows from (84) that the MA-projection
of the nonredundant
of any legitimate number X in RRNS is still a legitimate
Check how manyof
number in [0, M ) , that is X A = X . The proof of the
following theorem has been given in [4], and therefore, is
omitted here.
Theorem 15: Let x
be a number in RRNS and TAbe
the MA-projection of 2, where MA = IIk=,mIu.If TA#
.L
x, the residue representation of x,
uniquely differs from x
jtk-b+l
Y $? +I A burst residue error
of length > b
1 in one or more of the residue digits corresponding to the
moduli m i , , m,?; . mlA.
I is detected Stop
e,

I
Y .L
Before considering the range of the MA-projection 2, of
any illegitimate number x
in RRNS, we give the following
9 7 Denote e= 0 (mod mj ) by leimr=O.
where e is the consistent solution theorems. Theorem 16 is based on the fact that all codevec-
tors differ in at least d places in an RRNS with minimum
Fig. 3 . A decoder flowchart for single-burst error correction.
distance d.
Theorem 16: If X is a legitimate number in the RRNS
mate number, the redundant mixed radix digits are all zero. having minimum distance d , then any integer difiering x
However, if the redundant mixed radix digits are all zero, from X in at least one and no more than d - 1 residue
then mathematically, X i could still be an illegitimate num- digits is an illegitimate number.
ber. In [8], it was shown that an illegitimate projection Theorem 17: Let 2 be an illegitimate number in the
resulting from a single error cannot be smaller than the RRNS. If there exists a legitimate number X differing from
smallest nonzero number represented by (83) with all the x in the i,th, i,th; * i,th residue digits, then the MA-
e ,

nonredundant mixed radix digits zero. Therefore, Jenkins's projection X, is a legitimate number, where MA =
algorithm [SI-[lo] for locating a single residue digit error nLlmIu.
consists in checking whether the redundant mixed radix digits Proof: Since 2 differs from X in the i,th,
are all zero for each modulus m,-projection, i = 1, 2, * ., n. i,th, . , i,th residue digits, x
can be expressed as follows:
In the following analysis, we will show, from the coding
theory point of view, that their algorithm can be extended for
detecting and correcting multiple residue digit errors. The
MA-projection of X , denoted by X,,, is defined by where 0 < e' < n L = , m j mBy
. definition,

XA= X
i 21mod- (84)

where MA = rIk= , m I mA , = { i , , i2; . . , i,, . . * , i,; i , < i,

< ... < i , < < ih}, and hr n - k = d - 1. X,
can also be represented as a reduced residue representation of
X with the residue digits x j l , x i 2 , . xi, deleted. Then, the
a ,

mixed radix representation of X , is = x, < M .

n 1-1
Therefore, FAis a legitimate number. This proves the
theorem.
Theorem 18: Let x
be an illegitimate number in the
RRNS having minimum distance d. If the MA-projection
TA,(MA = I I ~ = , m l 1u ,Ih < d ) , is a legitimate number,
The legitimate and illegitimate range of the reduced RRNS than there exists only one legitimate number X differing
are [0, M ' ) and [ M', MM, / M A ) ,respectively, where M' from x
in one or more of the residue digits corresponding to
is the product of the first k moduli of the reduced RRNS, the moduli m,,, m r 2 , m,,. - a ,
SUN AND KRISHNA: CODING THEORY APPROACH-PART I1 31

Proof: Since # x
that the legitimate number X =
x,, x,
it is obvious from Theorem 15
is a solution to the
is affected by
m,,,
p or fewer errors corresponding to the moduli
m,?, . . . , m],,.
From the definition of the reduced RRNS,
problem. Now assume that there exist two different legitimate
numbers X and X ' , both differing from 2 in one or more
i.e., (86).x., can be expressed as follows:
- MMR
of the residue digits corresponding to the moduli m,,, X,,= X,,+ e' .
d
miz,-.., Then, ~m=,~l,,~cy=lm,,,

. (mod where e'< nB mi,,

a= I

where 0 5 e, e' < n:=lm,,x. Since 0 5 X , X' < M , 0 I MR

X, = X ' , < M , and M M R / I l k ; l m , ~ 2
x M , it follows that
e = e' = 0 and X = X ' . This contradicts the original as-
sumption X # X ' . This proves the theorem.
Theorem 19: In an RRNS with minimum distance d = h
+ + p 1, an illegitimate number x
is originated from a
legitimate number X affected by h errors corresponding to where i,. 5 k + c , i,.,, > k c. +
the moduli m,,, m,,, . mlA,if an only if the M.,-pl_qjection
a ,

FAis a legitimate number and the M,-projection X, is an By Theorem 1, M R 2 ~ ~ _ , m , + m n k = c + l m3 , ~ ~ n , = l m , , ~ ,

illegitimate number, where M A = FIk= l m l , , i M , = it follows that x,
2 X , + M' 2 M'. This proves the

n:=lmjo, 1 5 P 5 /3, and gcd(Mp. M.,) f n c y = l m l= , t theorem.

M*.I Therefore, by checking only the redundant mixed radix
Proof: By Theorem 17, the M,-projection J7',
is-a digits, it can be determined whether or not the projection is
legitimate number. Since gcd(M,, M,) # n k = l m l , r , X, legitimate. If all the redundant mixed radix digits are zero,
can be treated as a number originating from X, affected by the projection is legitimate; otherwise, it is illegitimate.
h or fewer errors corresponding to the moduli m , , , Based on the above analysis, a procedure for correcting
m i Z ; ~ ~m,,,
, that is, double errors and detecting multiple errors can be outlined as
follows.
Step 1: Based on the received residue vector, compute the
mixed radix digits and check if all the redundant
where 0 < e' < rI:=lm,cx. By Theorem 1, M , 2 mixed radix digits are zero. If yes, then declare
FI:=lmjar12= lm,cy, it follows that xp
is an illegitimate no error and stop. Otherwise, go to next step.
number. This proves the necessity. The sufficiency is proved Step 2: Compute the mixed radix digits of the m,mj-
by contradiction. Since FAis a legitimate number, by Theo- projection for i = 1, 2 ; . ., n , j = 1, 2; n,
e . ,

rem 18 there exists only one legitimate number X differing and i # j . Check if all the redundant mixed radix
from x in one or more of residue digits corresponding to the digits are zero. If yes, then declare the ith and
j t h residue digits are received in error, correct
moduli m,,, m,?, . . . Lm,A. However, if the legitimate num-
ber X differs from X in less than X residue digits (say p them by base extension, and stop. Otherwise, go
( p < A) residue digits corresponding to the moduli mL, to next step.
m,,; mjp) then, by Theorem 17, the Mp-projection X,
a ,
Step 3: Declare more than two errors detected. Stop.
is a legitimate number, where M p = rI:- I mlczand gcd( M,, VIII. DISCUSSION
M,) # M,,. This contradicts the original assumption. This
proves the theorem. The present-day practical algorithms for residue-error cor-
The range of the illegitimate projection Fa,of any illegiti- rection and detection mostly focus on single-error correction
mate number x
is given in the following theorem. because of the considerations of large memory space require-
ment or computational inefficiency for multiple error correc-
Theorem 20: Under the assumption that no more than /3
residue errors occur, for an RRNS with d = h + 1 and + tion. Our algorithms developed above do not require large
P_> h, if the M,,-pojection x,,
of any illegitimate number memory space as required for table lookup. They also seem
to be much superior to the algorithms in [6] and [9] from a
X is illegitimate, X, will be in the range [ M ' , M M , /M.,),
where MA = rIkzlm,,x, 1 Ii, 5 n , and is the lower
MI
computational efficiency point of view, i.e., the requirement
bound of the illegitimate range of the reduced RRNS. of multiplications (MULT) and additions (ADD) [13]. The
Proof: The illegitimate projection x,
can be treated as
a number originating from X,,in the reduced RRNS which
comparison of our algorithms derived in Sections IV and V
with those in [6] and 191 in terms of the requirement of ADD
and MULT is shown in Table I. We should point out that for
' gcd ( a , b ) denotes the greatest common divisor of a and b double-error correction the expression in column 2 of Table I
32 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-11: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 39, NO. 1, JANUARY 1992

TABLE I
THEREQUIREMENT
OF MULT OR ADD FOR SINGLE/DOUBLE
ERRORS
CORRECTION
~ ~ ~ ~~~~~~~~~

# Error Yau and Liu’s Jenluns’s New

Correction Algorithm [6] Algorithm [9] Algorithm

1 +
$k3 $ k 2 + ik - 2(MULT) ik3 + : k 2 + 2k(MULT) ik’ + j k - I(MULT)
d = 3 +
$k3 f k 2 + +k(ADD) $k3 + t k z + 2k(ADD) ik2+ i k + ](ADD)
2 Not applicable $ k 4 + $ k 3 + Yk’ + T k + 8(MULT) +
F k 2 T k - 2(MULT)
d=5 ak4 + T k 3 + T k 2 + T k + 8(ADD) $k2 - yk 2(ADD)+

is obtained by extending Jenkins’s algorithm [9] as described it is easy to show that

in the previous section. For single-burst error correction, our
method only needs to find n - k syndromes, whereas Yau
and Liu’s method has to find n syndromes. The comparison
is shown in Table 11. Mandelbaum proposed a decoding (arbitrary, else.
procedure for multiple-error correction based on continued
fraction expansion and Euclid’s algorithm [ 151- [ 171. Since Let 9 and 9’ be any (possibly overlapping) sets of indices
this procedure needs to process large valued integers and use in the range 1,. . ., k . Suppose 9 has c elements and 9’ has
an iterative process, it appears to be more suitable for a c’ elements. Let = n a’ be the intersection, with
general purpose computer. Also, it is difficult to compare C” Imin ( c , c’) elements. Let 9’ = CP - be the ele-
Mandelbaum’s algorithms to the algorithms presented in this ments of that are not in a’, and let a* = a’ - a’,*be the
paper as the approaches are entirely different. elements of 9’ that are not in a. Now we assume that
211m,nm,IImk- nm,- nmk<IIm,.
APPENDIX E@’ ’ JEQ’ kc@’ JEQ’ kcQ ret
PROOFOF GENERALFAULT-DETECTION,
FAULT-CORRECTIONCASE (A.3)

In this Appendix we give an alternate proof of conditions
that allow detection of up to ,6 faults and correction of up to
h faults in a maximally redundant RRNS. We follow this
with the general procedure for finding and correcting these
faults.
Let i,, . , i, be the actual locations of the faults in the
-
information digits, and let j , ,. . ,j , be the locations of the
+
faults in the parity digits, where c g IA. Let I; be the
reconstruction of the value modulo M using only the infor-
mation digits, but let us define the error E differently than in
Section 111. In particular, define
E = F-X.
The value E can take one of M possible values. Without
computing F or X directly, however, we can definitely state
that E is confined to the range
-M < E < M .
Now let e, = E(mod m;). Clearly all the digits e, are zero
except for digits i = i ,, * i,. We can therefore state that E
e ,
can be written in the form:
M
where \E is any subset of at least c c’ of the redundant +
moduli k +
1, * , n. Note that this assumption is trivially
satisfied if
2m,m, < m,ms (A.4)
for any 1 Ii , j 5 k and k + 1 5 r , s 5 n.
Now for the main result. Assume (A.3) holds, and assume
the RRNS is maximally redundant. Now hypothesize that the
syndromes we observe might have been caused by some
alternative “different” combination of up to 0 digit failures
in information digit positions i;, . , i‘,,, and parity digit
+
positions j ; ,. . , f,!,where c’ g’ 5 0, and h ,6 = n - +
k . (“Different” implies that at least one error information
digit in the alternative hypothesis is assumed to have a
different value than in the correct set of error information
digits.) We show by contradiction that the syndromes of this
alternative hypothesis cannot equal those from the correct
hypothesis. Under the alternative assumption, we hypothesize
an error - M < E’ < M , and can write this error in the
form:
M
E =e’-
n:=lm,;
where

where -
Let @ = { i , , * , i,] be the set of information digits that
actually failed. Let CP‘ = { i ; , . * . , i;,} be the set of informa-
tion digits assumed to have failed in the alternative hypothe-
sis. Let 9 = { j , , . . . , j , } be the set of parity digits that have
actually failed. Let P’= { j ; , * * j b . be } the set of parity
a ,

Let us define syndromes in the same way as before. Then digits assumed to have failed in the alternative hypothesis.
SUN AND KRISHNA: CODING THEORY APPROACH-PART I1 33

TABLE I1
THEREQUIREMENT
OF MULT OR CORRECTION
ADD FOR SINGLE BURSTERROR
Yau and Liu’s Algorithm [6] New Algorithm

k2 + (4b2 - 3 b ) k - 4 b 2 k - (5b’ - 4b)(MULT)

2
k2 + (4b’ -
1
b ) k (ADD)
1
-k2
2
+ ~
12b-5
2
k - (46’ - 76 + 2)(ADD)
k = 8 2 268 (MULT) 188 (MULT)
b = 4 2 294 (ADD) 166 (ADD)
k = 12 2 616 (MULT) 338 (MULT)
b = 4 2 648 (ADD) 292 (ADD)
k = 16 2 1280 (MULT) 728 (MULT)
b = 8 2 1336 (ADD) 654 (ADD)

Let e; = E’(mod m,)be the error digits in the alternative redundant moduli in index sets and q 2to get
9’,
hypothesis, with e: # 0 for all i E W . Note that this implies
+
that e’(mod m i ) 0 for i E a’. Define e n m, - e ’ n mi)
ic@’ re*‘.’
n+ I‘+ ‘k2
rn, = 6 M R . (A.6)
aI.2 =
a n a‘ Now
@I @ -
n m , - e’ n m ,
7 f$’J

a2 = @I’ - @1,2
e
i€@ I

91.2 = 9 n 9’
9’= lp - q1.2

9 2 = 9‘- q1.2.

Note that
= @I92 + @I
= 2
+ @ * +@
,€@I
m, - n
i€**
m, -
I€*’
m,. n n
@’ = @ I 9 2 +r p Suppose set @’,’ has e” elements, where c” 5 min ( e , e’),
= 91.2 + lpl
and suppose set has g ” elements, where g ” Imin ( g ,
g ’ ) . By our assumption (A.3), the expression above is guar-
9‘ = 91.2 +9 2 .
anteed to be less than the product of any set of c e‘ or +
Now if the alternative hypothesis could indeed explain the more redundant moduli. Note that the set 9’ 9* + +
observed syndromes, then we would have +
has g g’ - g ” redundant moduli in it, which leaves X p +
- (g + g’ - g ” ) 2 c + e’ + g” redundant moduli not in
that set. Choose any c + e’ of these, and call this set q.
Then

Equating this with the actual value of the syndromes (A.2) in en m , - e’n mi mi-
terms of e, and using our index set definitions, we get
<
re%
n
+ *I.’+ +q2
m, IM R .

Combining this with (A.6) shows that the constant 6 = 0,

and we have
which implies that
en m , - e ‘ n m, (A.7)
[(enmi e’n m i ) n
EO2
-
re@’ ,€*I
M
* min , € * I m i n , € * 2 m , 1 But this is only possible if
t?nm,,
i€@

(mod
r e { k + 1;.
n
, n} -‘?I -*I- q2
m,) = 0 .
e =
i€QI
e‘ =
ic*’
m,

which would imply that

But since the moduli are relatively prime, this implies that

(en mi - e’n m,)

I€@ i€@l
= 6
r € { k + 1 ;. .,n }
n - ‘k’ ’ - * I - q2
mr
Also, this implies that e, = 0 for all i E @ ’ , and that e; = 0
for some integer 6. Multiply both sides by the product of the for all i E a 2 .
34 IEEE TRANSACTIONS ON CIRCUITS AND SYSTtMS-I1 ANALOG AND DIGITAI. SIGNAL PROCI-SING. VOL 39. NO. I . JANUARY 1992
Now, we originally assumed that there were nonzero er-
rors in all the index positions in the set+, including those in
the set 9 ' . We must conclude that 9 ' must be a null set. and
so 9 = Our alternative hypothesis assumed that there
were nonzero errors in all the index positions in the set +',
including those in the set + I . We must conclude that if the
syndromes match, then 9' is a null set, and so a' =
Putting this together, 9 = a', and so if the syndromes match
then the alternative hypothesis must have correctly identified
the faulty information digits. Also note that (A.7) implies that
if 9 ' and (P2 are null, then e = e' and E = E'. Thus the
error information digits e, and e; must also match, and the
alternative hypothesis exactly matches the actual errors. We
conclude that with Assumption (A.3) and with up to X faults.
the syndromes uniquely specify the erroneous information
error digits.
A procedure to correct up to X errors is as follows: Check
the number of nonzero syndromes. If there are none, then no
errors occurred. If there are fewer than A, then all errors are
in the parity digits. If there are greater than X but fewer than
/3 errors, then more than X errors occurred. If there are 0 or
more nonzero syndromes, then consider all possible sets of
zero up to X errors in information digits. For each hypothe-
sis, try to solve (A.2) for a consistent value e in the range
(A.1). At most one of these combinations can yield a consis-
tent solution e (either positive or negative) which would
explain the non-faulty syndromes. If none yield a consistent
solution e, there must be more than X faults. Total complex-
ity is the time required to compute the base extension, plus
the time to evaluate every possible hypothesis involving X
failures.
ACKNOWLEDGMENT
The authors wish to express their thanks and appreciation
towards the reviewers for their thoughtfulness, thoroughness,
and for providing constructive criticism and insight into
many aspects of the error control problem in RRNS. We
have a greatly improved and complete paper due to their
efforts and time. The associate editor is also to be thanked for
his assistance and guidance during the review process. The
proof in Appendix A is taken from the review of one of the
reviewers.
REFERENCES
N . S. Szabo and R. I. Tanaka. Residue Arithmetic and its Applica-
tion to Computer Technology. New York: McCraw-Hill, 1967.
R. W. Watson and C. W. Hastings. "Self-checked computation using
residue arithmetic." Proc. IEEE. pp. 1920- 1931, Dec. 1966.
D. Mandelbaum. "Error correction in residue arithmetic," IEEE
Trans. Cornput., pp. 538-545, June 1972.
F. Barsi and P. Maeutrini, "Error correcting properties of redundant
residue number system$." IEEE Trans. Cornput.. pp. 307-315.
Mar. 1973.
-. "Error detection and correction by product codes in residue
number system." I€€€ Trans. Comput.. vol. C-23, pp, 915-924,
Sept. 1974.
S . S-S Yau and Y-C Liu. "Error correction in redundant residue
number systems." IEEE Trans. Cornput.. vol. C-22. pp. 5-1 I , Jan.
1973.
V . Ramachandran. "Single residue error correction in residue number
systems." IE€€ Trans. Cornput.. pp. 504-507. May 1983.
M. H . Etzel and W. K . Jenkins. "Redundant residue number systems
for error detection and correction in digital filters," I€€€ Trans.
Acousi.. Speech. Signal Processing, pp. 538-544, Oct. 1980.
W . K . Jenkins and E. J. Altman. "Self-checking properties of residue
number error checkers based on mixed radix conversion." IEEE
Trans. Circuits Sysr.. pp. 159-167. Feb. 1988.
W . K . Jenkins. "The design of error checkers for self-checking
residue number arithmetic." / E € € Trans. Conipui.. pp. 388-396.
Apr. 1983.
M. J. Bell, Jr. and W . K . Jcnkin,. "A residue to mixed radix
convener and error checker for a five-moduli residue number system."
in Proc. IE€€ I n t . Conf. A S S P . vol. 3 . 1984.
C:C. Su and H.-Y. Lo. "An algorithm for scaling and single reaidue
error correction in residue number systems." IEEE Trans. Conzput ..
pp. 1053-1064. Aug 1990.
A. P . Shenoy and R. Kumaresan. "Fast base extension using a
redundant modulus in RNS." I€€E Trans. Cornput., pp. 292-297,
Feb. 1989.
H . Krishna. K. Y. Lin. and JLD. Sun. "A coding theory approach to
error control in redundant residue number systems-Part I: Theory
and \ingle error correction." pp. 8- 17. this issue.
D. M. Mandelhaum. "On a class ot' arithmetic codes and a decoding
algorithm." / € € E 7rans. Inform Theory. vol. IT-22, pp. 85-88.
Jan. 1976.
-. "Further results o n decoding arithmetic residue codes." I€€€
Trans Inform. Theory. vol. IT-24. pp. 643-644. Sept. 1978.
-. "An approach to an arithmetic analog of Berlekamp's algo-
rithm." /E€€ Trans. Inforr??. Theory. vol. IT-30, pp. 758-762,
Sept. 1984.
Jenn-Dong Sun. for a photograph and biography, please see page 17 of this
i\we.
Hari Kriahna. for a photogrdph and biography. please see page 17 of this
i\\ue