LDPC Codes and Finite Geometries
Ian F. Blake (UBC)(joint with Qiuju Dao (LSI Logi ) and Shu Lin (UC Davis))May 20148JCC, Shanghai, May 2014
Purpose of talk
Comments on the uses of �nite geometries inthe onstru tion of LDPC ode to a hieve lowprobability of error �oors
8JCC, Shanghai, May 2014 1
Outline
I. The elements of LDPC oding for low error �oorsII. Classi al �nite geometries and odingIII. Partial geometries and generalized quadrangles
8JCC, Shanghai, May 2014 2
I. The elements of LDPC oding for low error �oors
8JCC, Shanghai, May 2014 3
Preliminaries
I using hard de isions at the re eiver in urs a ostof 2dB of Eb=N0 - annot a�ord this - ) soft de isionsI using soft de isions ) annot use algebrai de oding algsI must use soft de ision algorithms and try for ML performan eI ) message passing algorithms - an approximationI su h "real number" algorithms a dynami al system) onvergen e questionsI trapping sets
8JCC, Shanghai, May 2014 4
Model of a digital ommuni ation system
+n(t) psd N0=2fsi (t); i = 1; : : : ;Mg ^i
I Eb=N0 = SNR (power ratio) = (Eb=T )=(N0W )Eb = energy in transmitted signal per bit over (0; T )I Eb=N0dB = 10(log10(Eb=N0))I 3dB = doubling of power/energy per bitI many parameters of interest, power, bandwidth baud rate,modulation, signal shaping, : : :
8JCC, Shanghai, May 2014 5
Hard vs soft de isions
Re eived sequen e: r = (r1; r2; : : : ; rn)
` I Hard de ision: r 2 Fn2, Soft de ision: r 2 RnI If ri independent, hard de ision okayIf ri orrelated (as in oding), need soft de ision
8JCC, Shanghai, May 2014 6
Shannon oding theorem - informalTheorem (Shannon): For a dis rete memoryless hannel, there existsa quantity, apa ity, C, su h that ommuni ation with arbitrarily lowerror probability is possible for any rate R less than apa ity and thisis not possible at rates above apa ity. In parti ular, there exists afun tion Er(R) su h thatPe < exp(NEr(R); R < C; Er(R) � 0; 0 � R < C:Er(R)
RC1
8JCC, Shanghai, May 2014 7
Capa ity - AWGN hannel
AWGN: CAWGN = W log(1 + PN0W ) bps
BIAWGN: CBIAWGN = � ∫ ��(x) log2 ��(x)dx � 12 log2(2��2)where ��(x) � p8��2 = exp(�(x + 1)2=2�2+exp(�(x � 1)2=2�2:� a fun tion of Eb=N0.
8JCC, Shanghai, May 2014 8
Hard vs softEb=N0
C0
1soft hard2dB
PeEb=Nosoft hard2dB
ShannonI for a given rate ode - Shannon limit Eb=N0I oding to a hieve apa ity8JCC, Shanghai, May 2014 9
Error orre ting odes - binary
I (n; k) ode - k � n generator matrix, (n � k)� n parity he k matrixI 2k odewords, rate = k=n,I De oding: for a given re eived word r �nd losest odeword xI Hard de ision - algebrai de oding algorithms -in Fn2I Soft de ision - approximations to maximum likelihood (ML) - in Rn- minimizes Pe for equally likely inputs- ML algorithms too omputationally omplexI algebrai soft de ision algorithms known but not e�e tive?ML de oding: ^x = argmaxx2C P (x j r)
8JCC, Shanghai, May 2014 10
Maximum likelihood de oding - the problem
re eived r = (r1; r2; : : : ; rn) 2 Rn �nd argmaxx2C P (x j r)I Sear hing all 2k possibilities in Rn very omplexI We annot a�ord to give up 2-dB ) soft de isionI For good performan e we have to ome lose to MLI Real number algorithms - dynami systemsI Gallager understood this problem in the early 1960's !!!The answer: message passing algorithms on graphs
8JCC, Shanghai, May 2014 11
Gallager A de oding algorithm - hard de ision
I alphabet f�1; 1gI initially, re eived symbols VN's transferred to he k nodesI he k node j sends to variable node vi (if onne ted)the parity (produ t) of all re eived symbols ex ept from viI a VN sends to neighbor CN the original symbol unless allin oming symbols were the same - unanimous agreementI exa t analysis available - assuming independen e of in omingsymbols - indi ates gap to apa ity for biregular odesI Gallager B - variable thresholds
8JCC, Shanghai, May 2014 12
Belief propagation (BP)/message passing algorithms
I for soft de ision (BIAWGN) hannelsI an "approximation" to ML - although often far from itI omplexity of algorithm number of edges in graph= number of ones in parity he k matrixI onvergen e of algorithms, number of iterations, performan ewhen to stop?
8JCC, Shanghai, May 2014 13
Belief propagation algorithm equations
Important prin iple: message sent from VM v to CN must not take intoa ount message sent from to v in previous round - and for CN's to VN'sThe equations: log likelihood ratios:L(x j y) = P r(x = 0 j y)=P r(x = 1 j y); `i = lnL(xi j yi)`i =1) xi = 0m(`)v = { mv ` = 0mv + ∑ 02Cvn m`�1 0v ` � 1m` v = ln 1 + ∏v 0inV nv tanh(m`v 0 =2)1�∏v 0inV nv tanh(m`v 0 =2)
8JCC, Shanghai, May 2014 14
Belief propagation algorithm diagram
v ... ......
... m(`�1) vm(`)v I iterations build up belief - +1) x = 0
8JCC, Shanghai, May 2014 15
The error �oor problem
I a dynami al system over the realsI omplexity proportional to no. edges in graph
PeEb=N0
error �oor
8JCC, Shanghai, May 2014 16
Trapping sets
1. A (�; �) trapping set for the ode de�ned by the Tanner graph G is asubset � of VN's su h that the subgraph of G indu ed by the set �,denoted by G[�℄, has exa tly � odd degree CN's (and an arbitrary numberof even degree CN's).2. The trapping set is said to be elementary if all the asso iated he k nodesin G[�℄ have degree 1 or 2.3. The trapping set is alled small if � � pn (n is the ode length) and�=� � 4.8JCC, Shanghai, May 2014 17
Trapping sets
In essen e, if there is a small trapping set, there is a relativelyhigher probability the BP algorithm will fail to onverge asit iterates. The updates " ontradi t" or os illate as thealgorithm iterates.
8JCC, Shanghai, May 2014 18
LDPC odes - summing up
A good ode should have:I good minimum distan e for its rateI be e� ient to en odeI no small trapping sets for good error �oorI have sparse parity he k matrix for algorithm e� ien yI girth � 6?I performan e lose to the Shannon limit - about 1dB at 10�6?
8JCC, Shanghai, May 2014 19
II. Classi al �nite geometries and oding
8JCC, Shanghai, May 2014 20
The geometri odes
I a variety of related odes studied extensivelyI these in lude Reed-Muller (RM) odes, Generalized RM, proje tive RMEu lidean geometry odes, proje tive geometry odesI several ways of de�ning these odes:generator polynomial, geometri , multivariable (evaluation)I the lass of polynomial odes (Kasami, Lin and Peterson, 1968) in ludesall these lasses of odesI the book of Assmus and Key
8JCC, Shanghai, May 2014 21
Generalized Reed-Muller (GRM) odes - y li over FqI a natural generalization of the binary RM odesI Constru t (m + 1)� (qm � 1) array over Fqzeroth row all ones - rows i asso with xi ; i = 1; : : : ; m olumns = all nonzero m-tuples over Fq; u0; u1; : : : ; uqm�1"pun tured" GRM ode GRM� - drop the zero ol.I Pr(m; q) = all polys degree � r in m variables over FqI the "evaluation" ode: f = (f (u1) : : : ; f (uqn�1)); ui 2 FqGRM�q(r;m) = f f j f 2 Pr(m; q)g
8JCC, Shanghai, May 2014 22
Properties of GRM odes
For r < q n = qm � 1; k = (r +mm ); d = (q � r)q(m�1) � 1;where wq(j) = ∑i ji ; j = ∑i jiqi ; 0 � ji < q:Theorem: � 2 Fqm primitive. GRM�(r;m) y li with �j a root i�fj j 0 < j < qm � 1; 0 < wq(j) � m(q � 1)� r � 1g:(relates single variable to multivariable approa hes)
8JCC, Shanghai, May 2014 23
Proje tive geometry odes - PG(m; q)
De�nition 1: A q-ary r -th order proje tive geometry ode is the largest linear ode PG(q;m) with all r -�ats in its dual spa e.De�nition 2: A q-ary r -th order y li proje tive geometry ode of length�m;q = (qm+1 � 1)=(q � 1) has de�ning set of roots for the dual ode�h(q�1); � primitive 2 Fqm+10 < min0�`<mwq(p`h(q � 1)) � �(q � 1)
Theorem (Hamada): The two de�nitions are equivalent.
8JCC, Shanghai, May 2014 24
Parameters of the proje tive geometry odes - PG(m; q)
n = (m+1�1)=(q � 1); k = n � Rr(m; q); d � (qm�r�1 � 1)=(q � 1) + 1
where Rr(m; q) is the rank of r -�ats in PG(m:q). For the dual ode
d = (qr+1 � 1)=(q � 1)
8JCC, Shanghai, May 2014 25
Eu lidean geometry odes
De�nition 1: A q-ary r -th order Eu lidean geometry ode is the largest linear ode EG(q;m) with all r -�ats not going through the origin in its dual spa e.De�nition 2: A q-ary r -th order Eu lidean geometry ode of length qm is anextended y li ode with a de�ning set of roots for the dual ode�h; � 2 Fqm 0 < min0�`<mwq(p`h) < �(q � 1)Theorem (Hamada): The two de�nitions are equivalent.
8JCC, Shanghai, May 2014 26
Properties of EG odes
An r -th order q-ary EG ode of r -�ats in EG(m; q) in the dual ode is y li with parameters:n = qm � 1; k = n � fRr(m; q)� Rr(m � 1; q)g; d � qm�r + pqm�r�1
and the dual ode has minimum distan e d = qr � 1.I It is noted the GRM odes over prime �elds are also geometri odes.I expressions for Rr(m; q)? Hamada
8JCC, Shanghai, May 2014 27
Hamada's theorem
Theorem (Hamada): The q-rank of the in iden e matrix of pointsand r -�ats in PG(m; q) is equal to :
Rr(m; q) = ∑fs0;:::;smgm�1∏j=0 L(sj+1;sj)
∑i=0 (�1)i(m + 1i )(m + sj+1p � sj � ipm )
where q = ps and summation is taken over all ordered sets(s0; s1; : : : ; sm) denoted by Sm;r(pm) of (m + 1) integers s`; ` =0; 1; : : : ;, su h thatsm = s0; 0 � sj � m�r; and 0 � sj+1p�sj � (m+1)(p�1); j = 0; 1; : : : ; s�1:and L(sj+1; sj) = bsj+1p�sjp :8JCC, Shanghai, May 2014 28
Spe ial ases of Hamada's theorem
I Hamada's formula di� ult to gain insight fromI some parti ular ases of points and �ats wherethe formula is "simple" are summarizedI in many ases the formulae are of use in odingI let d(r;m; q) = Rr(m; q) be the p-rank, q = ph of points andr -�ats in PG(m; q) and d�(r;m; q) the rank of the dual ode8JCC, Shanghai, May 2014 29
Spe ial ases of Hamada's theorem - ontd.1d�(1; m; p) = (m+p�1p�1 )d�(m � 1; m; q) = qm+1�1q�1 � (m+p�1m )h � 1d�(2; m; p) = (m+2p�2m )� (p � 2)(m+p�2m�1 )d�(3; m; p) = (m+3p�3m )� (2p � 3)(m+2p�3m�1 ) + (p�22 )(m+p�3m�2 )
1Ce herini and Hirs hfeld, 1992, Hirs hfeld and Shaw, 19948JCC, Shanghai, May 2014 30
Spe ial ases of Hamada's theorem - ontd.2d�(r;m; 2) = ∑ri=1 (m+1i )d�(r;m; p) = pm+1�1p�1 �∑m�rs=0 ∑[s(p�1)=p℄i=0 (�1)i(m+1i )(m+s(p�1)�ipm )(S)d�(r;m; p) = ∑r�1j=0(�1)j((r�j)(p�1)�1j )(m+(r�j)p�rm�j ) (HS)
Note impli ations for d�(1; m; p)Hamada: rank of all r -�ats in EG(m; q) is = Rr(m; q)� Rr(m � 1; q)2Ce herini and Hirs hfeld, 1992, Hirs hfeld and Shaw, 19948JCC, Shanghai, May 2014 31
Spe ial ases of Hamada's theorem - Hirs hfeld, Shaw,van Lint
pm+1 � 1p � 1 �m�1
∑s=0 [s(p�1)=p℄∑i=0 (�1)i(m + 1i )(m + s(p � 1)� ipm ) = (m + p � 1p � 1 )
8JCC, Shanghai, May 2014 32
Two results in Eu lidean geometries
1) Key and Ma kenzie, 1991: The rank of m-�ats in EG(2m; p):m�1∑i=0 (�1)i(2mi )(m + (m � i)p2m )
2) Sin, DCC 2004: V a ve tor spa e, dimension n + 1 over Fpt In iden ematrix of e �ats and d-�ats in V where "in iden e" means nonzerointerse tion formula given for the p-rank of the in iden e matrix generalizesformula of Hamada Other "simple" formulae?8JCC, Shanghai, May 2014 33
III. Partial geometries and generalized quadrangles
8JCC, Shanghai, May 2014 34
Partial geometries
I in iden e stru tures of points and lines (in l. EG, PG)I a set N of n points and a set M of mI in iden e relations - parallel lines et I the geometri aspe t is helpful to analyze trapping sets
8JCC, Shanghai, May 2014 35
Partial geometries - ontd
A partial geometry PaG( ; �; Æ) if for � � 2; � 2 and Æ � 1:1. any two points are on at most one line;2. ea h point is on lines;3. Ea h line passes through � points;4. If a point v is not on a line L, then there areexa tly Æ lines, ea h passing through v and a point on L.5. When Æ = � 1, a net
8JCC, Shanghai, May 2014 36
Partial geometries - ontd
I n = (�(�� 1)( � 1) + Æ)=Æ pointsI m = ((�� 1)( � 1) + Æ)=Æ linesProfessor Lin will talk on these odes
8JCC, Shanghai, May 2014 37
Example! " # $ % & '( '' ') '* '!
'(!'"
'(!'!
'(!')
'(!'(
'(!%
'(!#
'(!!
'(!)
'((
+'#')&,'"*$)-./!012/34567869:;<5
=>?,@A
+'#')&,'"*$)-,AB?,CD6"(
+'#')&,'"*$)-,A0B?,CD6"(
+'#')&,'"*$)-,AB?,CD6'(
+'#')&,'"*$)-,A0B?,CD6'(
+'#')&,'"*$)-,AB?,CD6"
+'#')&,'"*$)-,A0B?,CD6"
=E:;;8;30C9CD
F;<8@5@3A2=G
Figure 1: Performan e urves for the (16129,15372)quasi- y li PaG8JCC, Shanghai, May 2014 38
Trapping sets for PaG odes
Theorem:3 If G is the Tanner graph of an LDPC ode with girth at least 6,with V N 0s of degree , and G ontains a (�; �) trapping set with � < then
� � ( + 1� �)�:
(Thus, if + 1� � > 4, not a small trapping set)3Diao et al, IT Trans, 20138JCC, Shanghai, May 2014 39
Trapping sets for PaG odes
I geometri stru ture allows interpretation of trapping setsI VN's are points, CN's linesI � � V N 0s a (�; �) trapping set )� lines passing through odd number points in �I mi = number of lines (CN's) passing through ipoints of � (number of CN 0s of degree i)
8JCC, Shanghai, May 2014 40
Trapping sets for PaG odes - ontd.
Theorem : Let GPaG be the Tanner graph of a PaG( ; �; Æ).If � is a (�; �) trapping set and � � then:
� � ( + 1� �)�+ ∑i odd(i � 1)2mi + ∑i even i(i � 2)mi
and equality holds if Æ = � and the sums go to 2b(�+1)=2 �1and 2b�=2 , respe tively.(Note that the �rst term on the right hand side of thisexpression is the general bound noted earlier.)
8JCC, Shanghai, May 2014 41
Trapping sets for PaG odes - ontd.
In the parti ular ase of a net the bound an be improved somewhat. LetL1; L2; : : : ; L� be a set of parallel lines in PaG( ; �; � 1) and for a set � ofV N 0s let �i =j � \ Li j. Then:Theorem: If the set � of V N 0s of a net is a (�; �) trapping set, then
� � ( � 1)�� �2 + �∑i=1 �2i + j f` : 1 � ` � �; �` oddg j :
(Agrees with previous bound if whenever �i � 2; 8i - better in other ases)
8JCC, Shanghai, May 2014 42
Codes from generalized quadrangles
I spe ial ase of partial geometries Æ = 1 i.e. for v =2 L thereis a unique line point v 0 2 L and line through v ; v 0I graph has girth 8- better performan e?I a deeply studied subje t - books by Thas (2), Payne and van MaldeghemI GQ odes - the dual spa e of the in iden e matrix of a GQI a subset of generalized polygonsI results on trapping sets for PaG's appli able hereI studied for odes for over 30 years
8JCC, Shanghai, May 2014 43
Codes from generalized quadrangles
I good survey on GQ odes by Liu and Pados (IT Trans, 2005)I stronger stru ture but more limited sets of parametersI girth 8 - bounds or exa t expressions fro dimension/ distan eI in iden e matrix an be made quasi- y li - blo ks of ir ulants- useful for en odingI di�erent size ir ulants for di�erent ordersI limited simulation - so far poor results - short odes
8JCC, Shanghai, May 2014 44
Codes from generalized quadrangles
I GQ of order (s; t) if ea h line is in ident with(s + 1) points and ea h ea h point is in ident with (t + 1) linesI only orders (q; q); (q; q2); (q2; q); (q2; q3); (q3; q2)and (q � 1; q + 1) and (q + 1; q � 1) (duals) knownI point set N and line set M then:j N j= (s + 1)(st)2 � 1st � 1 ; jM j= (t + 1)(st)2 � 1st � 1
8JCC, Shanghai, May 2014 45
Trapping sets in Codes from GQ's
Case i): An ovoid of the generalized quadrangle is a set of points O su h thatevery line in the quadrangle is in ident with a unique point of O. Su h ovoids,if they exist for a parti ular onstru tion, must have 1 + st points. Suppose aset of � � 1 + st VN's has the property that ea h line is in ident with atmost one point of S. The CN's orresponding to these VN's then is just theset of �(1 + t) lines that interse t the � VN's. Ea h su h CN has degree 1.Sin e it is assumed t will be quite large and, in parti ular t > 3, the ratio�=� = 1 + t > 4 and su h a set of VN's annot be a small trapping set.
8JCC, Shanghai, May 2014 46
Trapping sets in Codes from GQ's - ontd
Case ii) A spread S is a set of lines of the GQ su h that ea h point of thequadrangle is in ident with a unique line in S i.e. the lines of S are parallel.By duality if a spread exists it must have 1 + st lines. Suppose the � VN'sare ollinear. The orresponding CN's in the subgraph generated by su h a on�guration will have � = t� CN's and for large t does not orrespond to asmall trapping set. Again ea h CN has degree 1.
8JCC, Shanghai, May 2014 47
Trapping sets in Codes from GQ's
I the above suggests the geometri stru ture of GQ's usefuland an be exploited further for trapping set sizeI Question: For a given � of VN's, what is thesmallest � of CN's possible?I eviden e of usefulness of girth 8?
8JCC, Shanghai, May 2014 48
Summary of talk
I some understanding of stru ture of LDPC odes forlow error �oorI �nite geometries seem to give odes - good minimum distan evs rate, girth � 6, useful hara terization of trapping sets,e� ient en odings/quasi- y li I an analyti relationship between trapping set size anderror �oor would be very useful
8JCC, Shanghai, May 2014 49
Thank you for your attention8JCC, Shanghai, May 2014 50