Professional Documents
Culture Documents
Our rst goal today is to give a very abstra
t view of the Wel
h-Berlekamp de
oding algorithm for
Reed-Solomon
odes. This abstra
tion will allow us to see its generality, and thus apply it to other
families of error-
orre
ting
odes. The algorithm we des
ribe here is from the works of Pellikaan [9℄,
Kotter [6℄, and Duursma [1℄.
1.1 Reed-Solomon De
oding Review
Re all the Reed-Solomon de oding algorithm from last le ture. Here we have distin t points n
n 2 Fq given impli
itly, and we are expli
itly given as input elements
x1 ; : : : ; x n 2 Fq . The
y1 ; : : : ; y
For all 2 [ ℄, ( i ) i = ( i ).
i n a x y b x
2. bx( )
Output:
ax( )
The above algorithm and its proof of
orre
tness used the properties of Reed-Solomon
odes in
several ways. Below we list the dierent aspe
ts that seem spe
i
to Reed-Solomon
odes.
1. We used the fa
t that the indi
es of the
odewords (i.e., n ) are eld elements.
x1 ; : : : ; x
2. We used the fa
t that two low degree polynomials
annot agree in very many pla
es, several
times (in the
onstru
tion of the
ode, as well as the analysis of the de
oding algorithm).
3. We used the fa
t that multipli
ation of two low degree polynomials is a low degree polynomial.
4. We also used the fa
t that under the right
onditions, the ratio of two polynomials is a poly-
nomial.
Not very many of these above fa
ts were really
riti
al to the proof | they were just the simplest
way to get to the end. Some of the above properties (like (2)) are just fa
ts that hold for any
error-
orre
ting
ode. Others, in parti
ular (3), are somewhat spe
ial, but
an still be abstra
ted
with
are.
11-1
1.3 Multipli
ation of ve
tors
One of the
riti
al operations in the de
oding algorithm is that of multiplying the error-lo
ator
polynomial ( ) with the message polynomial ( ) and
onsidering the evaluations of the produ
t
a x p x
Why are these operations interesting? To motivate them, let us look ba
k at the RS de
oding
algorithm:
Let be the set of ve
tors obtained by evaluations of polynomials of degree at most .
U t
Let be the set of ve
tors obtained by evaluations of polynomials of degree less than .
V k
Then is the set of evaluations of polynomials of degree less than + that fa
tor into a
U ?V k t
In parti
ular is a subset of the set of evaluations of polynomials of degree less than + .
U ?V k t
To see that this is spe
ial, note that is a ve
tor spa
e of dimension + 1 and is a ve
tor
U t V
spa
e of dimension . What we have noti
ed is that their produ
t is
ontained in a ve
tor
k
spa
e of dimension + . This is very spe
ial. In general if we take two arbitrary ve
tor spa
es
k t
of dimension and , their produ
t would not be
ontained in any ve
tor spa
e of dimension
k t
In what follows we will show that this is the only spe
iality of polynomial based
odes. We will
show that if any
ode ends up having ni
e properties with respe
t to produ
t with some other
odes,
then it
an be de
oded.
1.4 Error-Lo
ating Pairs
Let be a [ ?℄q
ode and suppose, we wish to de
ode up to errors with C. The following de-
C n; k; e
nition des
ribes a simple
ombinatorial obje
t, whose existen
e suÆ
es to give a de
oding algorithm
for .
C
Denition 3 A pair of linear
odes ( ), with Fnq , form an -error-
orre
ting pair for a
A; B A; B e
11-2
1.5 The Generalized Algorithm
In this algorithm, we assume that we are given generator matri es for linear odes , , and , A B C
Step 3: Output the result of performing erasure de oding on z for ode , if this results in a unique C
odeword.
As usual we argue eÆ
ien
y rst and
orre
tness later. EÆ
ien
y is obvious for Step 2. For Step 3,
re
all that we observed in the last le
ture that erasure-de
oding
an be done in ( ) time for every O n
3
linear
ode. So it suÆ
es to argue eÆ
ien
y of Step 1. We do so by
laiming this is a task of nding
a solution to a linear system. Note that we are sear
hing for unknowns n n 2 Fq . a1 ; : : : ; a ; b 1 ; : : : ; b
turns into linear
onstraints on n. Also, the
onstraints i i = i is also linear in i and i
b1 ; : : : ; b a y b a b
(sin
e i is xed). Lastly, the
ondition (a b) 6= (0 0) is just asking for a non-zero solution to this
y ; ;
linear system. So the task at hand is that of nding a non-trivial solution to a homogenous linear
system. Again this
an be done eÆ
iently, in ( ) time. It remains to prove
orre
tness of the
O n
3
The proof of
orre
tness goes through the usual steps. We assume below that there exists a
odeword
2 C that is
lose to (within a Hamming distan
e of e of) the re
eived ve
tor y. We x this
odeword
and use it to argue
orre
tness.
Below we argue that a solution pair (a b) to Step 1 does exist. We argue that any solution pair
;
(a0 b0) to Step0 1 satises a0
=0 b00. Next0 we show that for any pair a0 2 and b0 2 there is
; ? A B
at most one
2 su
h that a
= b . The
orre
tness follows by noti
ing that the solution
C ?
0 output by the algorithm satises a0
0 = b0 , if (a0 b0 ) is the solution found by Step 1. Details
? ;
below.
Claim 4 There exists a pair (a b) as required in Step 1 exists. Furthermore they satisfy a
= b.
; ?
Proof By peeking ba
k at the analogous
laim in the Reed-Solomon de
oder, we realize we want a
to be the \error-lo
ator", i.e., satisfying i = 0 if i 6= i . Can we nd su
h a ve
tor that is non-zero?
a y
Turns out we
an, if we know has suÆ
iently large dimension. In parti
ular, the
onstraints i = 0
A a
(2) of the denition of an error-
orre
ting pair), it
ontains a non-zero ve
tor that satises all these
onstraints!
Now take a to be any non-zero ve
tor satisfying i = 0 if i 6= i . Take b to be the ve
tor
a
y
11-3
Proof Sin
e a0 b0 are outputs of Step 1, they satisfy the
ondition, a0 y = b0. Let a0
= b.
; ? ?
b and b are two
odewords of that agree on every
oordinate for whi
h i = i ( 0i = i i
0 B i y
b a y
(Condition (3) of the denition of an error- orre ting pair), implying b0 = b as required.
Claim 6 For any (a0 ; b0 ) 2 A B f(0; 0)g there exists at most one
2 C su
h that a0 ?
= b0 .
Proof Let
;
0 2 C satisfy a0 ?
= b0 = a0 ?
0 . First, let us note that we a
tually have a0 6= 0.
(If not, then b0 = a0 ?
0 would also be 0 and this
ontradi
ts the
ondition that together then are
non-zero.)
To prove0 the
laim, we wish to show
=0
0 . Sin
e both are
odewords of , it suÆ
es to show C
as desired.
We
an now formally prove the
orre
tness of the de
oding algorithm.
Lemma 7 If ( ) form an -error-
orre
ting pair for and
2 and y 2 Fnq satisfy (
y) ,
A; B e C C ; e
on
lude, we need to ensure that
is the only solution to Step 3 of the algorithm. But this is also
lear, sin
e any solution
0 to this step must satisfy 0i 0i = 0i i = 0i for every , and thus must be a
a
a y b i
odeword of satisfying a0
0 = b0 and by Claim 6,
is the unique ve
tor with this property.
C ?
Exer
ise:Verify that the Wel
h-Berlekamp algorithm from the last le
ture is an instantiation of
the Abstra
t de
oding algorithm given today.
We now move on to a more interesting appli
ation. Re
all the
onstru
tion of algebrai
-geometry
odes. (To be more pre
ise, re
all that we know very little about them to re
all mu
h.)
Algebrai
-geometry
odes. These
odes were
onstru
ted by nding points in Fmq and eval- n
uating all polynomials of \order" at most at all pla
es. The following properties of order were
` n
2. Two distin
t polynomials of order at most
an agree on at most out of the evaluation
` ` n
points.
3. The produ
t of two polynomials of order and has order at most + . `1 `2 `1 `2
11-4
These properties suÆ
e to prove that
odes obtained by the evaluation of polynomials of order at
most n d +1 give an [ ℄q
ode for some
n; k; d +1. As we see below, the same properties
k n d g
Proof Let n g ` . Below all referen
es to the \Conditions" are to the four
onditions in
e
1
To get an -error-
orre
ting pair, we need dim( ) (to satisfy Condition 2). We will pi
k
e A > e A
to be the algebrai -geometry ode obtained by evaluations of all polynomials of order at most + . e g
need ( ) . We know from the properties of algebrai
-geometry
odes, that has distan
e at
B > e B
e is less than half the RHS. Thus ( ) as
hosen above give an -error-
orre
ting
ode for .
A; B e C
As a
orollary we see that we get a pretty de
ent de
oding algorithm for algebrai
-geometry
odes,
orre
ting about ( ) 2 errors. However, it does not de
ode up to half the minimum
n ` g =
By the Chinese Remainder Theorem, residues of modulo any of the primes, suÆ
es to spe
ify
m k n
m . Thus spe
ifying its residue modulo primes, makes the above a redundant en
oding. This is
n
not one of our usual algebrai
odes | in fa
t it is not even linear! However, one
an apply our
usual notions of distan
e, error-dete
tion and
orre
tion to this
ode. We note the
ode has distan
e
n k + 1 (sin
e spe
ifying the residues modulo primes, spe
ies the message). So one
an as
k
the question - is it possible to
orre
t ( ) 2 errors. The rst de
oding algorithm was given
n k =
by Mandelbaum [8℄. Later work of Goldrei
h, Ron, and Sudan [2℄ showed how to interpret this
algorithm as a relative of the abstra
t de
oding algorithm given here. Turns out both algorithms
orre
t slightly less than ( ) 2 errors in polynomial time | this was xed later by Guruswami,
n k =
Sahai, and Sudan [3℄ using the algorithm of [2℄ in
ombination with an algorithm known as the
\Generalized Minimum Distan
e Algorithm" that we will talk about in the next se
tion.
2 De
oding Con
atenated Codes
We now move on to an elegant solution for the unambiguous de
oding problem for some families of
on
atenated
odes. Let us start by re
alling
on
atenated
odes.
11-5
m
E1
x1 x2 x3 xn
E2 E2 E2 E2
y1 y2 y3 yn
r1 r2 r3 rn
Inner Decoding
u1 u2 u3 un
Outer Decoding
m’
odeword of and de
ode it using a de
oding algorithm for . (See Figure 2.)
C1 E1
First, let us note that for the
on
atenated
odes we have
onsidered so far (Forney
odes and
Justesen
odes), the above is a
tually eÆ
ient. Re
all that in these appli
ations the outer
ode
is a well-studied
ode su
h as the Reed-Solomon
ode with eÆ
ient de
oding algorithms. On the
other hand the inner
ode is not well-understood | the only thing we know about it is that it
has good minimum distan
e properties. So it is not reasonable to expe
t a sophisti
ated de
oding
algorithm for the inner
ode. But then the inner
odes are so small the brute-for
e de
oding only
takes polynomial time in the length of the
on
atenated
ode, so we don't need a sophisti
ated inner
de
oding algorithm. So the entire de
oding pro
ess des
ribed above takes time polynomial in the
length of the
on
atenated
ode, for the typi
al
odes.
However, this doesn't give us an algorithm de
oding upto half the minimum distan
e of the
on
atenated
ode! It may only de
ode dd2 errors. We won't prove that it
an
orre
t so many
4
11-6
errors, or that it
an't
orre
t more. But to see a plausibility argument, note that to get a de
oding
failure, the adversary only has to ensure 2 of the symbols i 6= i and to get any su
h de
oding
d= u x
error for the inner de
oder it may only need to
ip 2 symbols of the inner alphabet. Thus a total
d2 =
of 4 errors may lead to a de
oding failure. Below we will show a
lever algorithm whi
h gets
dd2 =
around this problem with relatively little extra information about the outer de
oding algorithm.
This algorithm is the Generalized Minimum Distan
e (GMD) De
oding Algorithm, due to
Forney [5℄ (see also [4℄).
Exer
ise: Prove that the de
oding algorithm outlined above does indeed
orre
t at least ( 1)( d d2
1) 4 errors.
=
Let us start by looking at the de
oding algorithm we have for the outer
ode. We already seem to
be making full use of it | we assume it
an
orre
t ( 1) 2 errors, and it
an't possibly
orre
t
d =
more errors unambiguously. Turns out, there is one additional feature of the de
oding algorithm for
the outer
ode that
omes in quite handy. This is the feature that it
an deal with erasures quite
naturally and benet from it.
Proposition 10 Let be an [ C + 1 ℄q Reed-Solomon
ode. Suppose r 2 (Fq [ f?g)n is
n; n d ;d
a ve
tor derived from a
odeword
2 by erasures and errors. I.e., jf j i =?gj = and
C s t i r s
Proof The proposition is straightforward, given the observation that the ode 0 , obtained by C
How
an we use the option to de
lare erasures? In the pro
ess of de
oding the inner
ode, we
ignored some obvious information that was available to us. We not only
ould nd out the en
oding
of whi
h message word i was
losest to the th re
eived blo
k ri , but we also know how many
u i
errors have (potentially) o
urred in ea
h blo
k. Somehow we should make use of this information.
A natural idea is to to de
lare blo
ks with large number of errors to be erasures. This works out
roughly
orre
t. We will see that a simple probabilisti
interpretation of the distan
es give a right
strategy for de
laring erasures.
2.2 A randomized de
oding algorithm
Let us x some notation, that we have been introdu
ing as we went along. The message ve
tor is
m, its en
oding under the outer
ode is x = h n i, where i 2 FQ
= Fkq 2 . The en
oding of
x1 ; : : : ; x x
xi under the inner
ode is yi and thus the nal
odeword is y = hy yn i. The noisy
hannel 1; : : : ;
orrupts y and a ve
tor r = hr rn i is re
eived. We now des
ribe the de
oding algorithm for
1; : : : ;
this
ode.
Random-Con
at-De
oder
Given: r = hr1 ; : : : ; rn i.
Step 1: For 2 [ ℄,
ompute i that minimizes ( ( i) ri ).
i n u E2 u ;
Step 3: For every 2 [ ℄ repeat the following: With probability 2 0i , set i =?, else set i = i .
i n e =d2 v v u
11-7
Step 5: If a ve
tor m0 2 FkQ is obtained in Step 4,
he
k to see if ( (m) r) ( ) 2 and E2 E1 ; dd2 =
if so output m0.
The above algorithm
learly works in time that is ( + ( )), where ( ) is the time taken
O nn2 Q T n T n
by the errors and erasures de
oding algorithm. The main issue is how many errors it
orre
ts. We
will
laim that it \essentially" solves the unambiguous de
oding algorithms for the
on
atenated
ode. We do so via a
laim that is weak in that it only shows that the expe
ted number of errors
and erasures fall within the de
oding
apability of the outer
ode.
;
P
Lemma 11 Let i = (ri yi ) and = ni i . When v is pi
ked at random by Random-Con
at-
e e
=1 e
De
oder, then we have:
Exp[# erasures in v + 2 (# errors in v)℄ 2 e=d2 :
Remark: Note that if 2, the RHS above is less than as one would hope for.
e < dd2 = d
and then the lemma will follow by linearity of expe
tations. We prove this in
ases.
Case 1: i = i : In this
ase, the probability of an error in the th
oordinate is 0 and the probability
u x i
Case 2: i 6= i . In this
ase, we
ould have an error in the th
oordinate | this happens with
u x i
and the probability of an erasure is at most 2 2 i . Plugging these into the LHS of (1), again
e =d2
The above lemma implies that there is positive probability asso
iated with the event that v has
small number of errors and erasures. However we did prove not a high probability result. We won't
do so; instead we will just derandomize the algorithm above to get a deterministi
algorithm for the
unambiguous de
oding problem.
2.3 Deterministi
De
oder for Con
atenated Codes
We will develop the deterministi
de
oder in two (simple) steps. First note that we didn't really
need the random
hoi
es in Step 3 of Random-Con
at-De
oder do not need to independent
and expe
tation bound (Lemma 11) also holds if these events are
ompletely dependent. So, in
parti
ular, the following algorithm would work as well.
Modied-Random-Con
at-De
oder
Given: r = hr1 ; : : : ; rn i.
Step 1: For 2 [ ℄,
ompute i that minimizes ( ( i) ri ).
i n u E2 u ;
11-8
Step 5: If a ve
tor m0 2 FkQ is obtained in Step 4,
he
k to see if ( (m) r) ( ) 2 and E2 E1 ; dd2 =
if so output m0.
As in the analysis of Random-Con
at-De
oder we see that the random variable v as dened
in Steps 3.1 and 3.2 above satises the
ondition:
Exp[# erasures in v + 2 (# errors in v)℄ 2 e=d2 :
In parti
ular, there exists a
hoi
e of in Step 3.1, su
h that for this
hoi
e of , the ve
tor v
p p
The only interesting
hoi
es of are from the set = f0 1g [ f2 0i j 2 [ ℄g sin
e for every other
p S ; e =d2 i n
hoi
e of , there is some 0 2 for whi
h the ve
tor v obtained is identi
al for 0 and . This gives
p p S p p
Step 3.1: Let = f0 1g [ f2 0i j 2 [ ℄g. Repeat Steps 3.2 to 5 for every
hoi
e of 2 .
S ; e =d2 i n p S
if so output m0.
The running time of this algorithm is ( + ( )) where ( ) is the running time of the
O nn2 Q nT n T n
errors and erasures de
oder of the outer
ode. Corre
tness follows from the arguments developed so
far. We summarize the dis
ussion with the following theorem:
Theorem 12 Let be the
on
atenation of an [
C ℄Q outer
ode and an [ n; k; d ℄q inner C1 n2 ; k2 ; d2
Again, the running times
an be improved with some eort. In parti
ular, Kotter [7℄ shows how
to
ut down the run time to just ( + ( )) for some families of
on
atenated
odes.
O nn2 Q T n
Referen es
[1℄ Iwan M. Duursma. De
oding Codes from Curves and Cy
li
Codes. PhD thesis, Eindhoven
University of Te
hnology, 1993.
[2℄ Oded Goldrei
h, Dana Ron, and Madhu Sudan. Chinese remaindering with errors. IEEE Trans-
a
tions on Information Theory, 46(5):1330{1338, July 2000. Extended version appears as ECCC
Te
hni
al Report TR98-062 (Revision 4), http://www.e
.uni-trier.de/e
.
11-9
[3℄ Venkatesan Guruswami, Amit Sahai, and Madhu Sudan. Soft-de
ision de
oding of Chinese
Remainder
odes. In Pro
eedings of the 41st IEEE Symposium on Foundations of Computer
S
ien
e, pages 159{168, Redondo Bea
h, California, 12-14 November 2000.
[4℄ G. David Forney Jr. Con
atenated Codes. MIT Press, Cambridge, MA, 1966.
[5℄ G. David Forney Jr. Generalized minimum distan
e de
oding. IEEE Transa
tions on Information
Theory, 12(2):125{131, April 1966.
[6℄ Ralf Kotter. A unied des
ription of an error lo
ating pro
edure for linear
odes. In Pro
eedings
of the International Workshop on Algebrai
and Combinatorial Coding Theory, pages 113{117,
Voneshta Voda, Bulgaria, 1992.
[7℄ Ralf Kotter. Fast generalized minimum distan
e de
oding of algebrai
geometry and Reed-
Solomon
odes. IEEE Transa
tions on Information Theory, 42(3):721{737, May 1996.
[8℄ David M. Mandelbaum. On a
lass of arithmeti
odes and a de
oding algorithm. IEEE Trans-
a
tions on Information Theory, 21(1):85{88, January 1976.
[9℄ Ruud Pellikaan. On de
oding linear
odes by error
orre
ting pairs. Preprint, Eindhoven Uni-
versity of Te
hnology, 1988.
11-10