Académique Documents
Professionnel Documents
Culture Documents
Talk #$er$ie%
LDPC Codes Message Passing Decoding &nalysis of Message Passing Decoding 'Density ($ol)tion* &ppro+imations to Density ($ol)tion Design of LDPC Codes )sing D,(,
xC X n
Linear Codes
& linear code C 'o$er a finite field* can -e defined in terms of either a generator matri+ or parity3check matri+, 4enerator matri+ 4 'k5n*
C = 1mG/
C = 1c 7 cH2 = 0/
LDPC Codes
LDPC Codes 33 linear codes defined in terms of 6, 6 has a small a$erage n)m-er of non3zero elements per ro%
! ! ! ! ! ! ! ! ! ! ! H= ! ! ! ! ! ! ! ! ! ! ! ! ! ! !
H 'v0 c * !
& code%ord is an assignment of $2s s,t,7
...
...
x
v8c
= !c
Check nodes
Pro-a-ility Difference
x 0 y = p' x = 8 y * p' x = ! 8 y *
Properties7
B ' B' x** = x B ' x 0 y * = x 0 y B' x 0 y * = x 0 y
Message Domains
robability !i""erence
' x = ! 8 y* ' x = 8 y*
B '*
Likelihood Ratio
' y 8 x = !* 'y 8 x = *
B2 '*
Et is comp)ted like7
mv 0c
'i *
= xv 0 yv B 'mc 20v
c 28v c
'i *
v c ... ...
v 28c v
B'm
'i *
v 20 c
Decision R)le
&fter s)fficiently many iterations0 ret)rn the likelihood ratio7
!0 if x 0 y B 'mc 0v ' i * * < ! v v F= c 8v x other%ise 0
... ...
...
v 28c v
B2 'm
'i *
v 20c
*
'i *
I B2 ' !'m
v 20c
**
Camiliar (+ample7
Ef one die has density f)nction gi$en -y7
# ( $ ' & %
The density f)nction for the s)m of t%o dice is gi$en -y the con$ol)tion7
D,(, Threshold
Ci+ing the channel message densities0 the message densities %ill either ;con$erge; to min)s infinity0 or they %on2t, Cor the ga)ssian channel0 the smallest 'infim)m* SJR for %hich the densities con$erge is called the density e$ol)tion threshold,
M)lti3edge3type constr)ction
(dges of a partic)lar ;color; are connected thro)gh a perm)tation, (dges -ecome specialized, (ach edge type has a different message distri-)tion each iteration,
Re$ie%
Eterati$e Message Passing can -e )sed to decode LDPC codes, Density ($ol)tion can -e )sed to predict the performance of MP for infinitely large codes, More sophisticated classes of codes can -e designed to close the gap to the Shannon limit,