2007 IEEE International Symposium

on Signal Processing and Information Technology

Constructing Quasi-Cyclic LDPC codes Using a

Search Algorithm
Gabofetswe Malema
Department of Computer Science
University of Botswana
P/Bag 0022 Gaborone
Botswana
malemag @mopipi.ub.bw

Abstract- This article presents a search algorithm for con- proposed algorithm with some obtained codes sizes, girths
structing regular and irregular quasi-cyclic LDPC codes. Code and rates. Section III evaluates bit-error rate performances
rows and columns are divided into j (column weight) and k
(row-weight) or more groups respectively. Rows or columns in
a group are connected to the same group and in order of Section IV.
appearance. Grouping of rows and columns forms sub-matrices.
The sequential row-column connections order creates a cyclic
II. PROPOSED QC-LDPC SEARCH ALGORITHM
structure in sub-matrices. The row-column constraint is observed In [8] quasi-cyclic column-weight two codes are constructed
to avoid four-cycles. The proposed algorithm is flexible compared based on a distance graph representation. A desired girth g, is
to other methods. It obtains codes over a wide range of girths, obtained by searching and connecting two rows that are at a
rates and lengths. Bit-error rate simulations show that obtained
codes have good performance with randomly searched codes distance of atleastg We extend the search algorithmiproposed
performing better that sequentially searched codes. in [8] to construct QC-LDPC with column-weights higher than
two. LDPC code rows and columns are divided into groups
I. INTRODUCTION of equal sizes, p. The row and column groups are then paired
Low-density parity-check (LDPC) codes have been shown such that each row-group and each column-group appears k
to have good error correcting capabilities coming closer to and j times respectively for an (N, j, k) regular code, where
the Shannon limit[l]. However, constructing LDPC codes N is code length, k is row weight and j is column weight.
that meet hardware and decoding performance requirements A row, ri and column, cx that are at a desired distance, g,
is not easy. Construction of LDPC codes is not unique are then searched sequentially or randomly from each row-
and several parameters have to be considered. Flexible con- column group pair. The distance g is the shortest path (number
structions are important especially at short lengths to meet of edges) between the column and row. The grouping or rows
practical constraints. Random search algorithms such as bit- and columns divides the code into sub-matrices. The remaining
filling (BF)[2] and progressive-edge growth (PEG)[3] have rows and columns in the row-column group pair are connected
been developed to construct short LDPC codes with opti- with respect to ri and cx which produces a cyclic connection
mized girth, rate and performance. Though these methods in the sub-matrices. The algorithm only guarantees a girth of
are good at finding short length performing codes, obtained six. Despite this, girths higher than six were easily obtained.
codes are not easy to implement because of the random row- The algorithm is more successful in constructing high girths
column interconnect[4]. Quasi-cyclic LDPC codes are one codes (girths of ten and twelve) compared to when a distance
structured codes that are easy to implement because of their graph is constructed. The algorithm is described below with
block and cyclic interconnections[5][6]. Although, there are connections performed from the row side. It is to be noted that
several methods for constructing QC-LDPC codes, they do rows and columns are often refereed to as check and variable
not allow arbitrary sizes and rates[7]. They also do not allow nodes in a Tanner graph. In our description we choose to use
arbitrary sub-matrix configurations. In this paper we present the terms rows and columns.
a search algorithm that is more flexibility (code size, matrix Tanner Graph Search Algorithm
configuration etc) in code construction compared to existing 1) Divide rows into j' equal groups of size p,
methods. The algorithm which is based or similar to BF (RG, ... RG j' is at least equal to j (column-weight).
and PEG algorithms obtains both regular and irregular quasi- Columns are also divided into k' groups of size p,
cyclic codes. The proposed algorithm was first introduced in (CG ... CGk'), k' is at least equal to k(row-weight).
[8] for constructing column-weight two quasi-cyclic LDPC If the number of rows or columns is unknown or not
codes based on distance graph representation. We extend the given, start with a theoretical minimum number if known
algorithm to construct quasi-cyclic codes with column-weights otherwise start with a group size of k (row-weight).
higher than two by directly constructing a Tanner graph. rx is row x, and cx is column x.
This paper is organized as follows. Section II presents the Urx is a set of columns within a distance (number of

978-1 -4244-1 835-0/07/$25.00 ©C2007 IEEE 956

 rows
columns
CX rows O
columns
rows
columns
at least a distance of g from ri is sequentially or randomly
o || FI F0 2 /°searched from CG of RCGt. A sequential search is where
0| 0// the group is searched in order of position whereas in random
0 Dt | |D 200 / search the rows or columns of a group are picked arbitrarily.
0 D D 0
l ////0 If c, is not found the algorithm fails and exits. In the second
0 0 0 part of step 3, RG rows are connected to CG columns relative
F r\ /
0
°
05 0 Q° to the positions of ri and c, provided the girth condition
2 \ 0 0 is not violated. If one of the connections violates the girth
D 01 0 0I condition the algorithm fails and exits. The complexity of the
F-I
\F,1
1 \\ o / algorithm is O(M), where M is the number of rows[8]. If
I o
0
-I \\ 0
o ri L// ° edges to the Tanner graph are added from the column side,
F- ol06 0 0 the complexity is O(N), where N is the number of columns.
3
F-I
0
F6 \0 0 0
F \ I0 D 0 0
F- \°0 0 0 Figure 1 shows a construction of a (20,3,4) girth-six
0 0°
0~~~~~~~
0
code (p =5, j
0
0
=3,k1
4, M =15, N
There are 3 (j) row groups labelled (1,2,3) and 4 (k)
6). =20,g9
C 0 0 column groups labelled (4,5,6,7) of size 5 (p). To have a
0 Rowgroup2 Rowgroup3 regular (3,4) code, row and column groups are paired as
connections connechons connehons [ 1,4],[1,5],[1,6],[1,7],[2,4],[2,5],[2,6],[2,7],[3,4],[3,5],[3,6]
.1. Tanner graph representation of a (20,3,4) code.
and [3,7]. Figure 1 shows the general construction procedure.
Fig. For each row-column group pair ri and c, are first found
== = _ == == _ = === _ T == = or chosen. The remaining rows and columns are connected
= l l = = = l l == = l l == = l l = = relative to ri and c,. The figure shows 3 sets of connections
= 1 = l = = l l= = l l= = llfor the 3 row groups. In each case the reference row, ri,
_ ___ _ _ 1_ _1_ _ _ _-_ ___ _ is chosen as the first row in the group. A column,cx, at a
= l _= = l _= == _ = = = l === l _ distance of at least six from ri is then found in each column
__. l _ ___ _l ___
l_ _group. Only the first group and (ri, c,) connections are shown
l 1 l l l ll l for clarity. Girths of 4 are avoided by observing the row-
= =-l = _ l = == = = =-l =_ l = = column constraint. With girths higher than six, connecting the
= =-==-_ = =_== _ = = == ===_ = rest of row-column group rows and columns may result in
cycles less than the desired distance of g. Figure 2 shows the
resulting code matrix. The 3 row sub-matrices correspond to
the 3 row groups and the 4 column sub-matrices correspond
edges) of g from r,. to the 4 column groups. Figure 3 shows a general structure
Un, is a set of rows within a distance of g from c. for a (N, 2, 4) LDPC code when ri is always the first row of
2) Pair row and column groups such that each row-group a row group, where I is a p x p identity matrix, Ix is shifted
appears k times and each column-group j times. We p x p identity matrix and 0 is a p x p sub-matrix. If the
denote these row-column group pairs as RCG. number of row and column groups is equal to j and k the
RCG={(RG1, CG) ... (RGj, CGk')}. There are kj' or resulting matrix does not have zero sub-matrices as in part
jk' row-column group pairs. (a). If the number of groups is larger, the resulting code has
3) For t = 1 to kj' { zero sub-matrices as in part (b). If ri is selected randomly
RG E RCGj, CG CG C RCG, and c, is also searched randomly all the identity sub-matrices
select ri RCGt,
select e RG
o
RCGtare shifted.
The proposed algorithm is versatile in that sub-matrices (I
sequchnthatl ranelysealorithm failsome
or c, c CG
and 0) configurations can be arbitrary. This flexibility is
For zhatc1top,, e
useful in accommodating techniques such as check-variable
node overlapping in LDPC decoder designs. The arrangement
ri+z is connected to c±+z if c±+z V U,,+z else of sub-matrices is determined by row-column group pairings.
algorithm fails.

4) Use obtained Tanner graph to form a LDPC parity- A. Girth-Six and Eight Codes
check matrix. Tables I and II show obtained girth six and eight code
minimum group sizes using a sequential search. Obtained
The algorithm constructs the Tanner graph in step 3. A codes are about the same size as those found with algebraic
reference row, ri, is arbitrarily selected from a row-group,RG, methods in [9][10]. For girth-six codes, obtained codes have a
from the row-column group pair RCGt. A column,cx, that is minimum group size of k and k+1 when the row-weight is odd

957
 '~~~~~~~~
I '~~~~~~~~~
'~ ~ ~ ~ ~ ~ ~~BER vs SNR

(a)~~~~~~~~~~~~~~~~~~~~~~~~~~~1
I 0 0 0I
I 0 I~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~------- --------------
o I I I0
0 I 0~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~-------

I I0 0 I0 0 1 0 I o2

0000 ~~~~~~~~~~~~~~~~~~~10
ran------(1200 ---)g -6
Fig. 3. Quasi-cyclic code sub-matrices arrangement (a) with non-zero sub-ran (120036)------------------- g
(b) with zero sub-matices.
matrices " "

TABLE
I~~~~~~~~~~~~~~~~~~~~~~~~~Y--- T--- ---

GIRTH-SIX MINIMUM GROUP SIZES WITH A SEQUENTIAL SEARCH.~~~~~~~~~~~~LU---- m seq-(1200,3,6) SNRg=6

oP G.(ra n112 003068g=,-- --3-- -,6 -

Fig. 3. Quasi-cyclic (a) with
code sub-matrices arrangement
non-zero sub- H
matrices (b) with zero sub-matrices.
ran-(1200,3,6),g=8---------------3,6
ran-(1200,3,6)g=10 10------C(008,3,6

and even respectively. This minimum was found algebraically

in [9]. Smaller codescould be obtained by using random
searches for girth-eight codes as shown in Table III. Larger~~~~~~~~~~~~~~--- 10~~~~~~~~~~~~~~~~---- I--- 1

----
------------

si an eih coe fo ove a wid rang of rate and legts 10.4

E efrac uvsfr(6 eua oe ih2trtos
In[9] a 7sftar cale MAGMA1 1] was use to serh o
minimum 1grou sies ive thealebai githcodiios AsR10SN

Mackay and~~
PEG~ codesaofte0same3siz

rrgarQC-LDPC(1008, 3,6)
searches were not exhaustive, the code sizes obtained are~~~~~~~~~~~~~~~~~~-------------F------

Codevnstructedcodvesy oTaiinedmusin the proposd algobraithm

for9(1200,3,6 regula coudes Coesobtained byusing sequntial 103---- ------------

girth-ih codes.a
peartcularl poronucdio honi Girth-en1 codges SR----(-----------------)

co e r aiyobtainedb
ra doasachs.T
using erc rode s .hv abo uter the---------------------
58-
 TABLE II TABLE IV
(N,3,K) AND (N,4,K) GIRTH-EIGHT CODES MINIMUM GROUP SIZES USING (N,3,K) LDPC CODES SIZES WITH GIRTH TEN AND TWELVE.
SEQUENTIAL SEARCH.

k group size code size girth

k 4 5 6 7 8 9 10 11 12 4 47 141 x 188 10
4 110 330x440 12
3 9 19 21 25 27 55 57 61 63 5 134 402x670 10
4 31 38 44 56 73 84 108 121 5 341 1023 x 1705 12
6 161 483x966 10
TABLE III 6 513 1539 x 3072 12
OBTAINED (N, 3, k) GIRTH-EIGHT LDPC CODES USING RANDOM 7 231 693x 1617 10
SEARCHES 7 851 2553 x 5957 12
8 356 1068 x2848 10
k group size code size 9 515 1545x4635 10
4 9 27x36
5 13 39x65

shows BER performances of the codes with 80 iterations. Our

6 18 54x 108
7 22 66x 154 code performs as well as the other codes at this size despite
8 27 81 x216 its constraints in structure. The PEG code is better by about
9 38 114x342 0.1dB at 10-6 BER.
10 41 123x410 The irregular PEG code outperforms our irregular quasi-cyclic
11 56 168x616 code by about 0.2dB at 10-5 BER as shown in Figure 6. Our
12 58 174x 696 codes were obtained using the structures shown in Figure 7.
In the Figure each '1' entry represents a p x p shifted identity
sub-matrix and each '0' entry is a p x p zero sub-matrix. The
number and connections of groups is arbitrary. The irregular
same average girth (average of minimum cycles from all code is constructed by moving some connections from parity
nodes). Randomly searched codes show good performance as groups to information groups such that information groups
their BERs approach 10-7 at about 3dB. have four connections. The shift value in the identity sub-
Short block codes obtained by MacKay[12] and by Hu et matrices is determined by the proposed algorithm.
al[3] using the PEG algorithm were used for BER compar-
ison. These codes are some of the best known short length IV. CONCLUSION
performing codes and could be found in [13]. The codes used In this paper a search algorithm based on bit-filling and
are regular and irregular (1008,3,6) codes. For regular codes
PEG algorithms for constructing quasi-cyclic LDPC codes was
MacKay's algorithm obtains a girth of 8 and average of 9.
WithctheyPEG algorithmoba a girth of 8 and averagegh 9.66 presented. The algorithm could be used to construct a distance
or Tanner graph with a complexity linear to the code size.
are obtained. Using our proposed algorithm we obtain QC-
Both regular and irregular codes could be obtained over a wide
LDPC codes with a girth and average girth of 10. Figure 5 .g.i
range of rates, girths and lengths. Although girth conditions
are not sufficient beyond six, higher girths were easily ob-
1 01 1 ll o o0l01 100 .....
l tained with a Tanner graph construction. BER performance
0 0 00 _ .011
o 1
0 010
0 simulations show than randomly searched codes are better
0 0 1110
0 1
I010 01 oo
001oo ° 0
010
0 11

0 1 0I0 001
0
0
0 than °
those obtained with a sequential search. Obtained regular
01 0 __ 0_ 1T 0 0 1 quasi-cyclic codes perform as well as random codes. However,
001o o _oo_o o °o 0 0 1
o
random irregular codes using PEG algorithm outperform our
(a) codes. Use of performance optimization techniques could be
Ilo 0 0 0 0 0l 1 employed in the future to improve performance of the codes.
0 1 0 1 l l 1 l 0l l l l
1° 1° 1' 1' 1° 1' 1° 1' 1° 1° 1° 1° 1' 1° 1° 1' 1° 1°
1000lo 1 l l o0001000
1 0 00l 10 REFERENCES
0101 0 l 0ol0
o0 0 0 1

|10 l1 101 o o l l111 ° °° IX ° Io° l

o 01 1101 o
ol oo1I ' 1 ° [1] s. Chung,
.sign of low-density parity-check codes within 0.0045dB of the Shannon
G. D. Forney, J. J. Richardson, and R. Urbanke, "On the de-

1° l 1°1° l (b)
1°ll l 1°1° ° ll1° l 1°1° ° 1°I ....limit,"
[~~~~~~~~~2]
IEEE Communication Letters, vol. 5, pp. 58-60, February 2001.
J. Campello, D. s. Dodha, and s. Rajagopalan, "4Designing LDPC codes
using bit-filling," in Proceedings of the International Conference on
Fig. 7. QC-LDPC code structures showing the number and arrangement . ...Communications, vol. 1, June 2001.
Of row and column groups or sub-matrices (a) regular structure (b) irregular [3] X. Hu, E. Eleftheriou, and D. Arnold, "Progressive edge-growth Tanner
structure. graphs," Proc. IEEE GLOBECOM,San Antonio,TX, November 2001.

959
 [4] A.Blanksby and C.Howland, "A 690-mW 1-Gb/s 1024-b, Rate-1/2
Low-Density Parity-Check Code Decoder," IEEE Journal of Solid-Sate
Circuits, vol. 37, pp. 402-412, March 2002.
[5] H. Fujita and K. Sakaniwa, "Some Classes of Quasi-Cyclic LDPC
Codes: Properties and Efficient Encoding Method," IEICE Fundamen-
tals, vol. E88-A, no. 12, pp. 3627-3635, 2005.
[6] S. Olcer, "Decoding architecture for array-code-based LPDC codes,"
Proc. IEEE GLOBECOM, pp. 2046-2050, December 2003.
[7] S. Johnson and S. Weller, "Regular low-density parity-check codes from
combinatorial designs," Information Theory Workshop, Cairns, Australia,
September 2001.
[8] G. Malema and M. Liebelt, "Quasi-Cyclic LDPC Codes of Column-
Weight Two Using a Search Algorithm," EURASIP Journal on Advances
in Signal Processing, vol. 2007, pp. Article ID 45768, 8 pages, 2007.
[9] M. P. Fossorier, "Quasi-Cyclic Low-Density Parity-Check Codes From
Circulant Permutation Matrices," IEEE Transactions on Inforamtion
Theory, vol. 50, pp. 1788-1793, August 2004.
[10] M. O'Sullivan, "Algebraic Construction of Sparse Matrices with Large
Girth," IEEE Transactions on Information Theory, vol. 52, pp. 719-727,
February 2006.
[11] U. of Sydney, "MAGMA Algebraic Software," Department of Mathe-
matics, http://www.magma.maths. usyd.edu.au/magma.
[12] D. MacKay, S. Wilson, and M. Davey, "Comparison of constructions
of Gallager codes," IEEE Transactions on Communications, vol. 47,
pp. 1449-1454, October 1999.
[13] D. MacKay, "Encyclopedia of sparse graph codes.
Database of codes.," Department of Physics,
http://www.inference.phy.cam.ac. uk/mackay/codes/data.html.

960