Windowed Decoding of Protograph-based LDPC Convolutional Codes over Erasure Channels
Aravind Iyengar, Marco Papaleo, Paul Siegel, Jack Wolf, Alessandro Vanelli-Coralli, Giovanni Corazza
11 Windowed Decoding of Protograph-basedLDPC Convolutional Codes over Erasure Channels
Aravind R. Iyengar, Marco Papaleo, Paul H. Siegel,
Fellow, IEEE , Jack K. Wolf,
Life Fellow, IEEE ,Alessandro Vanelli-Coralli and Giovanni E. Corazza
Abstract —We consider a windowed decoding scheme for LDPCconvolutional codes that is based on the belief-propagation (BP)algorithm. We discuss the advantages of this decoding schemeand identify certain characteristics of LDPC convolutional codeensembles that exhibit good performance with the windoweddecoder. We will consider the performance of these ensemblesand codes over erasure channels with and without memory. Weshow that the structure of LDPC convolutional code ensemblesis suitable to obtain performance close to the theoretical limitsover the memoryless erasure channel, both for the BP decoderand windowed decoding. However, the same structure imposeslimitations on the performance over erasure channels withmemory.
Index Terms —Low-density parity-check codes, Convolutionalcodes, Iterative decoding, Windowed decoding, Belief propaga-tion, Erasure channels, Decoding thresholds, Stopping sets.
I. I
NTRODUCTION L OW-DENSITY parity-check (LDPC) codes, although in-troduced in the early 1960’s [4], were established as state-of-the-art codes only in the late 1990’s with the applicationof statistical inference techniques [5] to graphical modelsrepresenting these codes [6], [7]. The promising results fromLDPC block codes encouraged the development of convolu-tional codes defined by sparse parity-check matrices.LDPC convolutional codes (LDPC-CC) were first intro-duced in [8]. Ensembles of LDPC-CC have several attractivecharacteristics, such as thresholds approaching capacity withbelief-propagation (BP) decoding [9], and BP thresholds closeto the maximum a-posteriori (MAP) thresholds of randomensembles with the same degree distribution [10]. Whereasirregular LDPC block codes have also been shown to have BPthresholds close to capacity [11], the advantage with convo-lutional counterparts is that good performance is achieved byrelatively simple regular ensembles. Also, the construction offinite-length codes from LDPC-CC ensembles can be readily
A. R. Iyengar and P. H. Siegel are with the Department of Electrical andComputer Engineering and the Center for Magnetic Recording Research,University of California, San Diego, La Jolla, CA 92093 USA (e-mail:[email protected], [email protected]). J. K. Wolf (deceased) was with theDepartment of Electrical and Computer Engineering and the Center forMagnetic Recording Research, University of California, San Diego, La Jolla,CA 92093 USA. The work of A. R. Iyengar is supported by the NationalScience Foundation under the Grant CCF- .M. Papaleo is with Qualcomm Inc., San Diego, CA USA (e-mail: [email protected]). A. Vanelli-Coralli and G. E. Corazza are with the Uni-versity of Bologna, DEIS-ARCES, Viale Risorgimento, 2 - 40136 Bologna,Italy (e-mail: [email protected], [email protected])Parts of this work were presented at the 2010 Information Theory Workshop(ITW), Cairo, Egypt [1]; the 2010 International Communications Conference(ICC), Cape Town, South Africa [2]; and as an invited paper at the 2010 Int’lSymp. on Turbo Codes & Iterative Information Processing, Brest, France [3]. optimized to ensure desirable properties, e.g. large girthsand fewer cycles, using well-known techniques of LDPCcode design. Most of these attractive features of LDPC-CCare pronounced when the blocklengths are large. However,BP decoding for these long codes might be computationallyimpractical. By implementing a windowed decoder , one canget around this problem.In this paper, a windowed decoding scheme brought tothe attention of the authors by Liva [12] is considered. Thisscheme exploits the convolutional structure of the parity-checkmatrix of the LDPC-CC to decode non-terminated codes,while maintaining many of the key advantages of iterativedecoding schemes like the BP decoder, especially the lowcomplexity and superior performance. Note that althoughsimilar decoding schemes were proposed in [13], [14], theaim in these papers was not to reduce the decoding latency orcomplexity. When used to decode terminated (block) LDPC-CC, the windowed decoder provides a simple, yet efficientway to trade-off decoding performance for reduced latency.Moreover, the proposed scheme provides the flexibility to setand change the decoding latency on the fly. This proves tobe an extremely useful feature when the scheme is used todecode codes over upper layers of the internet protocol.Our contributions in this paper are to study the requirementsof LDPC-CC ensembles for good performance over erasurechannels with windowed decoding (WD). We are interestedin identifying characteristics of ensembles that present a goodperformance-latency trade-off. Further we seek to find suchensembles that are able to withstand not just random erasuresbut also long bursts of erasures. We reiterate that we will beinterested in designing ensembles that have the aforementionedproperties, rather than designing codes themselves. Althoughthe channels considered here are erasure channels, we note thatthe WD scheme can be used when the transmission happensover any channel.This paper is organized as follows. Section II introducesLDPC convolutional codes and the notation and terminologythat will be used throughout the paper. In Section III we de-scribe the decoding algorithms that will be considered. Alongwith a brief description of the belief-propagation algorithm, wewill introduce the windowed decoding scheme that is basedon BP. Possible variants of the scheme will also be discussed.Section IV deals with the performance of LDPC-CC on thebinary erasure channel. Starting with a short recapitulation ofknown results for BP decoding, we will discuss the asymptoticanalysis of the WD scheme in detail. Finite-length analysiswill include performance evaluation using simulations thatreinforce the observations made in the analysis. For erasure a r X i v : . [ c s . I T ] O c t channels with memory, we analyse LDPC-CC ensembles bothin the asymptotic setting and for finite lengths in Section V. Wealso include simulations illustrating the good performance ofcodes derived from the designed protographs over the Gilbert-Elliott channel. Finally, we summarize our findings in SectionVI. II. LDPC C ONVOLUTIONAL C ODES
In the following, we will define LDPC-CC, give a construc-tion starting from protographs , and discuss various ways ofspecifying ensembles of these codes.
A. Definition
A rate R = b/c binary, time-varying LDPC-CC is definedas the set of semi-infinite binary row vectors v [ ∞ ] , satisfying H [ ∞ ] v T[ ∞ ] = T[ ∞ ] , where H [ ∞ ] is the parity-check matrix H [ ∞ ] = H (1) H (1) H (2) ... H (2) . . . H m s (1) ... . . . H ( t ) H m s (2) . . . H ( t ) . . .. . . ... . . . H m s ( t ) . . .. . . (1)and [ ∞ ] is the semi-infinite all-zero row vector. The elements H i ( t ) , i = 0 , , · · · , m s in (1) are binary matrices of size ( c − b ) × c that satisfy [15] • H i ( t ) = , for i < and i > m s , ∀ t ≥ • ∃ t > such that H m s ( t ) (cid:54) = • H ( t ) has full rank ∀ t ≥ .The parameter m s is called the memory of the code and ν s =( m s + 1) c is referred to as the constraint length . The firsttwo conditions above guarantee that the code has memory m s and the third condition ensures that the parity-check matrixis full-rank. In order to get sparse graph codes, the Hammingweight of each column h of H [ ∞ ] must be very low, i.e., w H ( h ) (cid:28) ν s . Based on the matrices H i ( t ) , LDPC-CC can beclassified as follows [8]. An LDPC-CC is said to be periodic if H i ( t ) = H i ( t + τ ) ∀ i = 0 , , · · · , m s , ∀ t and for some τ > . When τ = 1 , the LDPC-CC is said to be time-invariant ,in which case the time dependence can be dropped from thenotation, i.e. H i ( t ) = H i ∀ i = 0 , , · · · , m s , ∀ t . If neitherof these conditions holds, it is said to be time-variant .Terminated LDPC-CC have a finite parity-check matrix H [ L ] = H (1) H (1) H (2) ... H (2) . . . H m s (1) ... . . . H ( L ) H m s (2) . . . H ( L ) . . . ... H m s ( L ) where we say that the convolutional code has been terminatedafter L instants . Such a code is said to be ( J, K ) regular if H [ L ] has exactly J ’s in every column and K ’s in every rowexcluding the first and the last m s ( c − b ) rows, i.e. ignoringthe terminated portion of the code. It follows that for a given J , the parity-check matrix can be made sparse by increasing c or m s or both, leading to different code constructions [16]. Inthis paper, we will consider LDPC-CC characterized by large c and small m s . As in [9], we will focus on regular LDPC-CCwhich can be constructed from a protograph. B. Protograph-based LDPC-CC
A protograph [17] is a relatively small bipartite graph fromwhich a larger graph can be obtained by a copy-and-permuteprocedure—the protograph is copied M times, and then theedges of the individual replicas are permuted among the M replicas to obtain a single, large bipartite graph referred to asthe derived graph . We will refer to M as the expansion factor . M is also referred to as the lifting factor in literature [11].Suppose the protograph possesses N P variable nodes (VNs)and M P check nodes (CNs), with degrees J j , j = 1 , · · · , N P ,and K i , i = 1 , · · · , M P , respectively. Then the derived graphwill consist of n = N P M VNs and m = M P M CNs. Thenodes of the protograph are labeled so that if the VN V j isconnected to the CN C i in the protograph, then V j in a replicacan only connect to one of the M replicated C i ’s.Protographs can be represented by means of an M P × N P bi-adjacency matrix B , called the base matrix of the protographwhere the entry B i,j represents the number of edges betweenCN C i and VN V j (a non-negative integer, since parallel edgesare permitted). The degrees of the VNs (CNs respectively) ofthe protograph are then equal to the sum of the correspondingcolumn (row, respectively) of B . A ( J, K ) regular protograph-based code is then one with a base matrix where all VNshave degree J and all CNs, excluding those in the terminatedportion of the code, have degree K .In terms of the base matrix, the copy-and-permute operationis equivalent to replacing each entry B i,j in the base matrixwith the sum of B i,j distinct size- M permutation matrices.This replacement is done ensuring that the degrees are main-tained, e.g., a in the matrix B is replaced by a matrix H ( M )2 = P ( M )1 ⊕ P ( M )2 where P ( M )1 and P ( M )2 are twopermutation matrices of size M chosen to ensure that eachrow and column of H ( M )2 has two ones. The resulting matrixafter the above transformation for each element of B , whichis the biadjacency matrix of the derived graph, correspondsto the parity-check matrix H of the code. The derived graphtherefore is nothing but the Tanner graph corresponding to theparity-check matrix H of the code.For different values of the expansion factor M , differentblocklengths of the derived Tanner graph can be achieved,keeping the original graph structure imposed by the pro-tograph. We can hence think of protographs as definingcode ensembles that are themselves subsets of random LDPCcode ensembles. We will henceforth refer to a protograph B and the ensemble C it represents interchangeably. Thismeans that the density evolution analysis for the ensemble of codes represented by the protograph can be performedwithin the protograph. Furthermore, the structure imposedby a protograph on the derived graph can be exploited todesign fast decoders and efficient encoders. Protographs givethe code designer a refined control on the derived graph edgeconnections, facilitating good code design.Analogous to LDPC block codes, LDPC-CC can also bederived by a protograph expansion. As for block codes,the parity-check matrices of these convolutional codes arecomposed of blocks of size- M square matrices. We now givetwo constructions of ( J, K ) regular LDPC-CC ensembles. Classical construction : We briefly describe the con-struction introduced in [18]. For convenience, we will refer tothis construction as the classical construction of ( J, K ) regularLDPC-CC ensembles. Let a be the greatest common divisor(gcd) of J and K . Then there exist positive integers J (cid:48) and K (cid:48) such that J = aJ (cid:48) , K = aK (cid:48) , and gcd ( J (cid:48) , K (cid:48) ) = 1 . Assumingwe terminate the convolutional code after L instants, we obtaina block code, described by the base matrix B [ L ] = L (cid:122) (cid:125)(cid:124) (cid:123) B B B ... B . . . B m s ... . . . B B m s . . . B . . . ... B m s where m s = a − is the memory of the LDPC-CC and B i , i =0 , · · · , m s are J (cid:48) × K (cid:48) submatrices that are all identical andhave all entries equal to . Note that an LDPC-CC constructedfrom the protograph with base matrix B [ L ] could be time-varying or not depending on the expansion of the protographinto the parity-check matrix.The protograph of the terminated code has N P = LK (cid:48) VNs and M P = ( L + m s ) J (cid:48) CNs. The rate of the LDPC-CCis therefore R L = 1 − (cid:18) L + m s L (cid:19) J (cid:48) K (cid:48) = 1 − (cid:16) m s L (cid:17) (1 − R ) (2)where R = 1 − J (cid:48) K (cid:48) is the rate of the non-terminated code.Note that R L → R and the LDPC-CC has a regular degreedistribution [9] when L → ∞ . We will assume that theparameters satisfy K (cid:48) > J (cid:48) and L ≥ − RR m s so that therates R and R L of the non-terminated and terminated codes,respectively, are in the proper range.The classical construction was proposed in [18] and itproduces protographs for some ( J, K ) regular LDPC-CCensembles. However, not all ( J, K ) regular LDPC-CC can beconstructed, e.g. m s becomes zero if J and K are relativelyprime and consequently the resulting code has no memory. In[9], the authors addressed this problem by proposing a con-struction rule based on edge spreading . We denote an ensembleof ( J, K ) regular LDPC-CC constructed as described here as C c ( J, K ) with the subscript c for “classical” construction. Modified construction : We propose a modified con-struction that is similar to the classical construction exceptthat we do not require that m s = a − , i.e. the memoryof the LDPC-CC is independent of its degree distribution.We further disregard the requirement that the B i matricesare identical and have only ones, i.e. parallel edges in theprotograph are allowed. However, the sizes of the submatrices B i , i = 0 , , · · · , m s will still be J (cid:48) × K (cid:48) . We will denote a ( J, K ) regular LDPC-CC ensemble constructed in this manneras C m ( J, K ) , with subscript m for “modified” construction.Note that the rate of the C m ( J, K ) ensemble is still given byEquation (2). Further, the independence of the code memoryand the degree distribution allows us to construct LDPC-CCeven when J and K are co-primes. This is illustrated in thefollowing example. Example 1:
Let J = 3 and K = 4 . Clearly, a classicalconstruction of this ensemble is not possible. However, withthe modified construction, we can set m s = 1 and define theensemble C m ( J, K ) given by B = , B = with design rate R L = 1 − (cid:0) L +1 L (cid:1) for a termination length L .Note that these submatrices are by no means the only possibleones. Another set of submatrices satisfying the constraints is ˆ B = , ˆ B = . (cid:3) The above example brings out the similarity between theproposed modified construction and the technique of edgespreading employed in [9], wherein the edges of the proto-graph defined by the matrix B (cid:48) = are “spread” between the matrices B and B (or between ˆ B and ˆ B ) to obtain a (3 , regular LDPC-CC ensemble withmemory m s = 1 . The advantage of the modified constructionis thus clear—it gives us more degrees of freedom to design theprotographs in comparison with the classical construction. Inparticular, the ensemble specified by the classical constructionis contained in the set of ensembles allowed by the modifiedconstruction, meaning that the best performing C m ( J, K ) ensemble (with memory the same as that of the C c ( J, K ) ensemble) is at least as good as the C c ( J, K ) ensemble. Notethat in [9], there was no indication as to how edges are to bespread between matrices. With windowed decoding, we willshortly show that different protographs (edge spreadings) havedifferent performances. We will also identify certain designcriteria for efficient modified constructions that suit windoweddecoding. C. Polynomial representation of LDPC-CC ensembles
We have thus far specified LDPC-CC ensembles by givingthe parameter L and the matrices B i , i = 0 , , · · · , m s . An al-ternative specification of terminated protograph-based LDPC-CC ensembles using polynomials is useful in establishingcertain properties of ( J, K ) regular ensembles and is describedbelow.Instead of specifying ( m s +1) matrices B i of size J (cid:48) × K (cid:48) ,we can specify the K (cid:48) columns of the ( m s + 1) J (cid:48) × K (cid:48) matrix B [1] = B B ... B m s using a polynomial of degree no more than d = ( m s +1) J (cid:48) − for each column. The polynomial of the j th column p j ( x ) = p (0) j + p (1) j x + p (2) j x + · · · + p ( d ) j x d (3)is defined so that the coefficient of x i , p ( i ) j , is the ( i +1 , j ) entryof B [1] for all i = 0 , , · · · , d and j = 1 , , · · · , K (cid:48) . Therefore,an equivalent way of specifying the LDPC-CC ensemble is bygiving L and the set of polynomials { p j ( x ) , j = 1 , , · · · , K (cid:48) } .With this notation, the l th column of B [ L ] is specified bythe polynomial x J (cid:48) i p j ( x ) where l = iK (cid:48) + j for unique ≤ i ≤ L − and ≤ j ≤ K (cid:48) . We can hence use “thecolumn index” and “the column polynomial” interchangeably.Further, to define ( J, K ) regular ensembles, we will need theconstraints p j (1) = J ∀ ≤ j ≤ K (cid:48) and K (cid:48) (cid:88) j =1 p [ m ] j (1) = K ∀ ≤ m ≤ J (cid:48) − , where p [ m ] j ( x ) is the polynomial of degree no larger than m s obtained from p j ( x ) by collecting the coefficients of termswith degrees l where l = hJ (cid:48) + m for some ≤ h ≤ m s , i.e. l = m (mod J (cid:48) ) : p [ m ] j ( x ) = p ( m ) j + p ( J (cid:48) + m ) j x + · · · + p ( m s J (cid:48) + m ) j x m s = m s (cid:88) h =0 p ( hJ (cid:48) + m ) j x h . (4)We will refer to these polynomials as the modulo polyno-mials . Let us denote the set of polynomials defining anLDPC-CC ensemble as P = { p j ( x ) , j ∈ [ K (cid:48) ] } , where [ K (cid:48) ] = { , , · · · , K (cid:48) } , and the modulo polynomials as P l = { p [ l ] j ( x ) , j ∈ [ K (cid:48) ] } , l = 0 , , · · · , J (cid:48) − . Later in thepaper, we will say “the summation of polynomials p i ( x ) and p j ( x ) ” to mean the collection of the i th and the j th columnsof B [1] . The following example illustrates the notation. Example 2:
For ( J, J ) codes, we have J (cid:48) = 1 and K (cid:48) = 2 ,the component base matrices B i , i = 0 , ..., m s are × matrices. With the first column of the protograph B [1] , weassociate a polynomial p ( x ) = p (0)1 + p (1)1 x + · · · + p ( m s )1 x m s of degree at most m s . Similarly, with the second column we associate a polynomial p ( x ) = p (0)2 + p (1)2 x + · · · + p ( m s )2 x m s ,also of degree at most m s . Then, the (2 i + 1) th column of B [ L ] can be associated with the polynomial x i p ( x ) , and the (2 i + 2) th column with the polynomial x i p ( x ) . As notedearlier, we will use the polynomial of a column and its indexinterchangeably, e.g. when we say “choosing the polynomial x i p ( x ) ,” we mean that we choose the (2 i + 1) th column of B [ L ] . Similarly, by “summations of polynomials p ( x ) and p ( x ) ,” we mean the collection of the corresponding columnsof B [ L ] . In order to define ( J, J ) regular ensembles, we willfurther have the constraint p (1) = p (1) = J . In this case,since J (cid:48) = 1 , p [0]1 (1)+ p [0]2 (1) = 2 J is the same as the previousconstraint, because p [0]1 (1) + p [0]2 (1) = p (1) + p (1) . (cid:3) We define the minimum degree of a polynomial a ( x ) as theleast exponent of x with a positive coefficient and denote it as min deg( a ( x )) . Clearly, ≤ min deg( a ( x )) ≤ deg( a ( x )) . Letus define a partial ordering of polynomials with non-negativeinteger coefficients as follows. We write a ( x ) (cid:22) b ( x ) if min deg( a ( x )) = min deg( b ( x )) , deg( a ( x )) = deg( b ( x )) andthe coefficients of a ( x ) are no larger than the correspondingones of b ( x ) . The ordering (cid:22) satisfies the following propertiesover polynomials with non-negative integer coefficients: if a ( x ) (cid:22) b ( x ) and c ( x ) (cid:22) d ( x ) , then a ( x ) + c ( x ) (cid:22) b ( x ) + d ( x ) a ( x ) c ( x ) (cid:22) b ( x ) d ( x ) . We define the boundary polynomial β ( a ( x )) of a polynomial a ( x ) to be β ( a ( x )) = x i + x j where i = min deg( a ( x )) and j = deg( a ( x )) . Note that when i = j , we define β ( a ( x )) = x i . We have for any polynomial a ( x ) , β ( a ( x )) (cid:22) a ( x ) .III. D ECODING A LGORITHMS
LDPC-CC are characterized by a very large constraintlength ν s = ( m s + 1) K (cid:48) M . Since the Viterbi decoder has acomplexity that scales exponentially in the constraint length,it is impractical for this kind of code. However, the sparsityof the parity-check matrix can be exploited and an iterativemessage passing algorithm can be adopted for decoding. Weconsider two specific iterative decoders here—a conventionalbelief-propagation decoder [6], [19] and a variant called awindowed decoder. A. Belief-Propagation (BP)
For terminated LDPC-CC, decoding can be performed asin the case of an LDPC block code, meaning that each framecarrying a codeword obtained through the termination can bedecoded with the sum-product algorithm (SPA) [19].Note that since the BP decoder can start decoding only afterthe entire codeword is received, the total decoding latency Λ BP is given by Λ BP = T cw + T dec , where T cw is the timetaken to receive the entire codeword and T dec is the timeneeded to decode the codeword. In many practical applicationsthis latency is large and undesirable. Moreover, for non-terminated LDPC-CC, a BP decoder cannot be employed. W K M W J Mm s K M ( L + m s ) J MLK M Fig. 1. Illustration of windowed decoding (WD) with window of size W = 4 for a C m ( J, J ) LDPC-CC with m s = 2 and L = 16 at the fourth decodinginstant. This window configuration consists of J W = W J (cid:48) M = 4 M rows of the parity-check matrix and all the ( W + m s ) K (cid:48) M = 12 M columns involvedin these equations: this comprises the red (vertically hatched) and the blue (hatched) edges shown within the matrix. Note that the symbols shown in green(backhatched) above the parity-check matrix have all been processed. The targeted symbols are shown in blue (hatched) above the parity-check matrix andthe symbols that are yet to be decoded are shown in gray above the parity-check matrix. B. Windowed Decoding (WD)
The convolutional structure of the code imposes a constrainton the VNs connected to the same parity-check equations—two VNs of the protograph that are at least ( m s + 1) K (cid:48) columns apart cannot be involved in the same parity-checkequation. This characteristic can be exploited in order toperform continuous decoding of the received stream througha “window” that slides along the bit sequence. Moreover,this structure allows for the possiblity of parallelizing theiterations of the message passing decoder through severalprocessors working in different regions of the Tanner graph.A pipeline decoder based on this idea was proposed in [8].In this paper we consider a windowed decoder to decodeterminated codes with reduced latency. Note that whereasa similar sliding window decoder was used to bound theperformance of BP decoding in [14], we are interested inevaluating the performance of the windowed decoder from aperspective of reducing the decoding complexity and latency.Consider a terminated ( J, K ) regular parity-check matrix H built from a base matrix B . The windowed decoder workson sub-protographs of the code and the window size W isdefined as the number of sets of J (cid:48) CNs of the protograph B considered within each window. In the parity-check matrix H ,the window thus consists of J W = W J (cid:48) M = W ( c − b ) rowsof H and all columns that are involved in the check equationscorresponding to these rows. We will henceforth refer to thesize of the window only in terms of the protograph with thecorresponding size in the parity-check matrix implied. Thewindow size W ranges between ( m s + 1) and ( L − becauseeach VN in the protograph is involved in at most J (cid:48) ( m s + 1) check equations; and, although there are a total of M P = J (cid:48) ( L + m s ) CNs in B , the decoder can perform BP when allthe VN symbols are received, i.e. when L ≤ W ≤ L + m s . Apart from the window size, the decoder also has a (typicallysmall) target erasure probability δ ≥ as a parameter . Theaim of the WD is to reduce the erasure probability of everysymbol in the codeword to a value no larger than δ .At the first decoding instant, the decoder performs belief-propagation over the edges within the window with the aimof decoding all of the first K (cid:48) symbols in the window, calledthe targeted symbols . The window slides down J (cid:48) rows andright K (cid:48) columns in B after at least a fraction (1 − δ ) ofthe targeted symbols are recovered (or, in general, after amaximum number of belief-propagation iterations have beenperformed), and continues decoding at the new position at thenext decoding time instant.We refer to the set of edges included in the window atany particular decoding time instant as the window config-uration . In the terminated portion of the code, the windowconfiguration will have fewer edges than other configurationswithin the code. Since the WD aims to recover only thetargeted symbols within each window configuration, the entirecodeword is recovered in L decoding time instants. Fig. 1shows a schematic representation of the WD for W = 4 .The decoding latency of the K (cid:48) targeted symbols with WDis therefore given by Λ W D = T W + T dec ( W ) , where T W isthe time taken to receive all the symbols required to decodethe K (cid:48) targeted symbols, and T dec ( W ) is the time taken todecode the targeted symbols. The parameters T cw and T W arerelated as T W = ( W + m s ) K (cid:48) LK (cid:48) T cw = W + m s L T cw , since at most ( W + m s ) K (cid:48) symbols are to be received toprocess the targeted symbols. The relation between T dec and We will see shortly that setting δ = 0 is not necessarily the most efficientuse of the WD scheme. T dec ( W ) is given by T dec ( W ) = WL T dec , since the complexity of BP decoding scales linearly in block-length and the WD uses BP decoding over W K (cid:48) symbolsin each window configuration. We assume that the numberof iterations of message passing performed is fixed to be thesame for the BP decoder and the WD. Thus, in latency-limitedscenarios, we can use the WD to obtain a latency reductionof Λ W D ≤ W + m s L Λ BP ∆ = w Λ BP . The smallest latency supported by the code-decoder systemis therefore at most a fraction w min = m s +1 L that of theBP decoder. As pointed out earlier, the only choice for non-terminated codes is to use some sort of a windowed decoder.For the sequence of ensembles indexed by L , with the choiceof the proposed WD with a fixed finite window size W ,the decoding latency vanishes as O ( L ) . We will typically beinterested in small values of W where large gains in decodinglatencies are achievable. Since the decoding latency increasesas W increases, the trade-off between decoding performanceand latency can be studied by analyzing the performance ofthe WD for the entire range of window sizes. Latency Flexibility:
Although reduced latency is an impor-tant characteristic of WD, what is perhaps more useful practi-cally is the flexibility to alter the latency with suitable changesin the code performance. The latency can be controlled byvarying the parameter W as required. If a large latency can behandled, W can be kept large ensuring good code performanceand if a small latency is required, W can be made small whilepaying a price with the code performance (We will see shortlythat the performance of WD is monotonic in the window size).One possible variant of WD is a decoding scheme whichstarts with the smallest possible window size and the sizeis increased whenever targeted symbols cannot be decoded,i.e., the target erasure probability cannot be met within thefixed maximum number of iterations. Other schemes wherethe window size is either increased or decreased based on theperformance of the last few window configurations are alsopossible. IV. M EMORYLESS E RASURE C HANNELS
In this section, we confine our attention to the performanceof the LDPC-CC when the transmission occurs over a mem-oryless erasure channel, i.e. a binary erasure channel (BEC)parameterized by the channel erasure rate ε . A. Asymptotic analysis
We consider the performance of the LDPC-CC in terms ofthe average performance of the codes belonging to ensemblesdefined by protographs in the limit of infinite blocklengthsand in the limit of infinite iterations of the decoder. As in thecase of LDPC block codes, the ensemble average performanceis a good estimate of the performance of a code in the ensemble with high probability. We will therefore concentrateon the erasure rate thresholds [11] of the code ensemblesas a performance metric in our search for good LDPC-CCensembles. BP : The asymptotic analysis of LDPC block codes withthe BP decoder over the BEC has been well studied [20]–[23].For LDPC-CC based on protographs, the BP decoding thresh-olds can be numerically estimated using the Protograph-EXIT(P-EXIT) analysis [24]. This method is similar to the standardEXIT analysis in that it tracks the mutual information betweenthe message on an edge and the bit value corresponding to theVN on which the edge is incident, while maintaining the graphstructure dictated by the protograph .The processing at a CN of degree d C results in an updatingof the mutual information on the d th C edge as I out ,d C = C ( I in , , · · · , I in ,d C − ) = d C − (cid:89) i =1 I in ,i (5)and the corresponding update at a VN of degree d V gives I out ,d V = V ( I ch , I in , , · · · , I in ,d V − ) = 1 − ε d V − (cid:89) i =1 (1 − I in ,i ) (6)where I ch = 1 − ε is the mutual information obtained from thechannel. Note that the edge multiplicities are included in theabove check and variable node computations. The a-posteriorimutual information I at a VN is found using I = 1 − ε d V (cid:89) i =1 (1 − I in ,i ) = 1 − (1 − I out ,d V )(1 − I in ,d V ) where the second equality follows from (6). The decoder issaid to be successful when the a-posteriori mutual informationat all the VNs of the protograph converges to as thenumber of iterations of message passing goes to infinity.The BP threshold ε ∗ BP ( B ) of the ensemble described by theprotograph with base matrix B is defined as the supremum ofall erasure rates for which the decoder is successful. Example 3:
The protograph B , = (3 3) has a BPthreshold of ε ∗ BP ≈ . . Note that all the CNs in theprotograph are of degree while all the VNs are of degree . This BP threshold is expected because B , corresponds tothe (3 , regular LDPC block code ensemble. The followingprotograph B (cid:48) , has a BP threshold ε ∗ BP ≈ . for L = 40 .Note that, as before, all VNs are of degree and all the CNsexcept the ones in the terminated portion of the code are of We will use the phrase “mutual information on an edge” to mean the mu-tual information between the message on the edge and the bit correspondingto the adjacent VN. degree . B (cid:48) , = · · · · · · ... ... · · · · · · · · · · · · ... ... ... ... ... ... . . . · · · This is the C c (3 , ensemble constructed in [18]. In terms ofthe notation introduced, this is given as B = B = B =[1 1] ; or equivalently as p ( x ) = p ( x ) = 1 + x + x . (cid:3) The above example illustrates the strength of protographs—they allow us to choose structures within an ensemble definedby a pair of degree distributions that may perform better thanthe ensemble average. In fact, the BP performance of regularLDPC-CC ensembles has been related to the maximum-a-posteriori (MAP) decoder performance of the correspondingunstructured ensemble [10]. WD : We now analyze the performance of the WDdescribed in Section III-B in the limit of infinite blocklengthsand the limit of infinite iterations of belief-propagation withineach window.
Remark:
In the limit of infinite blocklength, each term inthe base protograph B is replaced by a permutation matrix ofinfinite size to obtain the parity-check matrix, and therefore thelatency of any window size is infinite, apparently defeating thepurpose of WD. Our interest in the asymptotic performance,however, is justified as it allows us to establish lower boundson the probability of failure of the windowed decoder torecover the symbols of the finite length code. In practice, itis to be expected that the gap between the performance ofa finite length code with WD and the asymptotic ensembleperformance of the ensemble to which the code belongsincreases as the window size reduces due to the reductionin the blocklength of the subcode defined by the window.The asymptotic analysis for WD is very similar to that of theBP decoder owing to the fact that the part of the code within awindow is itself a protograph-based code. However, the maindistinction in this case is the definition of decoding success. Inthe case of BP decoding, the decoding is considered a successonly when, for any symbol in the codeword, the probabilityof failing to recover the symbol goes to (or equivalently,the a-posteriori mutual information goes to ) as the numberof rounds of message-passing goes to infinity. On the otherhand, the decoding within a window is successful as longas the probability of failing to recover the targeted symbolsbecomes smaller than a predecided small value δ . The decoderperformance therefore depends on two parameters: the windowsize W and the target erasure probability δ .We define the threshold ε ∗ ( i ) ( B , W, δ ) of the i th windowconfiguration to be the supremum of the channel erasure ratesfor which the WD succeeds in retrieving the targeted symbols of the i th window with a probability at least (1 − δ ) , giventhat each of the targeted symbols corresponding to the first ( i − window configurations is known with probability − δ .Fig. 2 illustrates the threshold ε ∗ ( i ) ( B , W, δ ) of the i th windowconfiguration. The windowed threshold ε ∗ ( B , W, δ ) is then W K M W J M − εm s K M − δ Fig. 2. Illustration of the threshold of the i th window configuration ε ∗ ( i ) ( B , W, δ ) . The targeted symbols of the previous window configurationsare known with probability − δ . The targeted symbols within the window arehighlighted with a solid blue bar on top of the window. The symbols withinthe blue (hatched) region in the window are initially known with probability − ε . The task of the decoder is to perform BP within this window until theerasure probability of the targeted symbols is smaller than δ . The window isthen slid to the next configuration. defined as the supremum of channel erasure rates for which thewindowed decoder can decode each symbol in the codewordwith probability at least − δ .We assume that between decoding time instants, no infor-mation apart from the targeted symbols is carried forward, i.e.when a particular window configuration has been decoded,all the present processing information apart from the decodedtargeted symbols themselves is discarded. With this assump-tion, it is clear that the windowed threshold of a protograph-based LDPC-CC ensemble is given by the minimum of thethresholds of its window configurations. For the classicaland modified constructions of LDPC-CC described in SectionII-B, all window configurations are similar except the onesat the terminated portion of the code. Since the windowconfigurations at the terminated portions can only performbetter, the windowed threshold is determined by the thresholdof a window configuration not in the terminated portion of thecode. Note that the performance of WD when the informationfrom processing the previous window configurations is madeuse of in successive window configurations, e.g. when symbolsother than the targeted symbols that were decoded previouslyare also retained, can only be better than what we obtain here.We now state a monotonicity property of the WD the proofof which is relegated to Appendix I. Proposition 1 (Monotonicity of WD performance in W ): For any C m ( J, K ) ensemble B , ε ∗ ( B , W, δ ) ≤ ε ∗ ( B , W + 1 , δ ) . (cid:4) It follows immediately from the definition of the windowedthreshold that ε ∗ ( B , W, δ ) ≤ ε ∗ ( B , W, δ (cid:48) ) ∀ δ ≤ δ (cid:48) . Furthermore, from the continuity of the density evolutionequations (6) and (5), we have that when we set δ = 0 , we decode not only the targeted symbols within the window butall the remaining symbols also. Since the symbols in the rightend of the window are the “worst protected” ones within thewindow (in the sense that these are the symbols for which theleast number of constraints are used to decode), we expect thewindowed thresholds ε ∗ ( B , W, δ = 0) to be dictated mostly bythe behavior of the submatrix B under BP. In the following,when the base matrix B of the protograph corresponding toan ensemble C is unambiguous, we will write ε ∗ ( B , W, δ ) and ε ∗ ( C , W, δ ) interchangeably.We next turn to giving some properties of LDPC-CCensembles with good performance under WD. We start withan example that illustrates the stark change in performance asmall difference in the structure of the protograph can produce. Example 4:
Consider WD with the ensemble C c (3 , inExample 3 with a window of size W = 3 . The correspondingprotograph defining the first window configuration is and we have ε ∗ ( C c (3 , , W = 3 , δ = 0) = 0 . This is seenreadily by observing that there are VNs of degree that areconnected to the same CNs. In fact, from this reasoning, wesee that ε ∗ ( C c ( J, K (cid:48) J ) , W, δ = 0) = 0 ∀ J ≤ W ≤ L .As an alternative, we consider the modified construction ofSection II-B2 to obtain the C m ( J, K ) ensemble B (cid:48) given by B = [2 2] , B = [1 1] . This ensemble has a BP threshold ε ∗ BP ( B (cid:48) ) ≈ . for L = 40 which is quite close to thatof the ensemble C c (3 , , ε ∗ BP ( C c (3 , ≈ . . WD witha window of size for this ensemble has the first windowconfiguration which has a threshold ε ∗ ( B (cid:48) , W = 3 , δ = 0) ≈ . , i.e.we can theoretically get close to . of the BP thresholdwith < of the latency of the BP decoder. Note thatthis improvement in threshold has been obtained while alsoincreasing the rate of the ensemble, since m s = 1 for the B (cid:48) ensemble in comparison with m s = 2 for C c (3 , . (cid:3) The above example illustrates the tremendous advantageobtained by using C m ( J, K ) ensembles for WD even underthe severe requirement of δ = 0 . The following is a goodrule of thumb for constructing LDPC-CC ensembles that havegood performance with WD. Design Rule 1:
For C m ( J, K (cid:48) J ) ensembles, set p ( d j ) j ≥ for all j ∈ [ K (cid:48) ] where d j = min deg( p j ( x )) .The above design rule says that for ( J, K (cid:48) J ) ensembles, itis better to avoid degree- VNs within a window. Note thatnone of the C c ( J, K (cid:48) J ) ensembles satisfy this design rule. Wenow illustrate the performance of LDPC-CC ensembles withWD when we allow δ > . Example 5:
We compare three LDPC-CC ensembles. Thefirst is the classical LDPC-CC ensemble C = C c (3 , . Thesecond and the third are LDPC-CC ensembles constructed asdescribed in Section II-B2. The ensemble C is defined by thepolynomials p ( x ) = 2 + x , p ( x ) = 2 + x and C is defined by q ( x ) = q ( x ) = 2 + x. We first observe that all three ensembles have the sameasymptotic degree distribution, i.e. all are (3 , regular LDPC-CC ensembles when L → ∞ . While C and C have a memory m s = 2 , C has a memory m s = 1 . Therefore, for a fixed L , while C and C have the same rate, C has a higherrate. Another consequence of a smaller m s is that C can bedecoded with a window of size W min ( C ) = 2 . Further notethat whereas C and C satisfy Design Rule 1, C does not.For a window of size , the subprotographs for ensembles C and C are as shown in Example 4, and that for ensemble C is as shown below In Fig. 3, we show the windowed thresholds plotted againstthe window size for the three ensembles C , C and C byfixing L = 100 for δ ∈ { − , − } . W ε ∗ ( B , W , δ ) ε ∗ ( C , W, e − ε ∗ ( C , W, e − ε ∗ ( C , W, e − ε ∗ ( C , W, e − ε ∗ ( C , W, e − ε ∗ ( C , W, e − Fig. 3. Windowed threshold as a function of the window size for theensembles C i , i = 1 , , with δ ∈ { − , − } . The rates of theensembles C and C are . whereas that of C is . . The correspondingShannon limits are therefore . for C and C , and . for C . A few observations are in order. The monotonicity of ε ∗ ( B , W, δ ) in W as proven in Proposition 1 is evident. Thewindowed thresholds ε ∗ ( B , W, δ ) for C and C are fairly closeto the maximum windowed threshold even when W = W min .The windowed thresholds for ensembles C and C are robustto changes in δ , i.e. the thresholds are almost the same (thepoints overlap in the figure) for δ = 10 − and δ = 10 − .Further, the windowed thresholds ε ∗ ( C i , W, δ ) are fairly close to the BP thresholds ε ∗ BP ( C i ) , i = 1 , , for W ≥ . We willsee next that this last observation is not always true. (cid:3) Effect of termination:
The better BP performance of the C c ( J, K ) ensemble in comparison with that of the ( J, K ) -regular block code ensemble (cf. Example 3) is because ofthe termination of the parity-check matrix of C c ( J, K ) codes.More precisely, the low-degree CNs at the terminated portionof the protograph are more robust to erasures and their erasure-correcting power is cascaded through the rest of the protographto give a better threshold for the convolutional ensemblein comparison with that for the corresponding unstructuredensemble [16]. From the definition of the WD, we can see thatthe sub-protograph within a window does not have the lower-degree checks if previous targeted symbols are not decoded.Therefore, we would expect a deterioration in the performance.Furthermore, the Design Rule 1 increases the degrees of theCNs in the terminated portion. Therefore, the effect of differenttermination on the WD performance is of interest. Example 6:
Tables I and II illustrate the WD thresholds for C m ( J, J ) ensembles that satisfy Design Rule 1 except when J = m s + 1 . These ensembles are defined by the polynomials p ( x ) = p ( x ) = ( J − m s ) + x + x + · · · + x m s . Note that J ≥ m s + 1 . The ensembles are terminated sothat the rate is R L = 0 . . The worst threshold with WD(corresponding to the least window size W min = m s + 1 )is denoted ε ∗ m s +1 . The largest threshold with WD is denoted ε ∗ L − m s and the BP threshold as ε ∗ BP . The increase in the TABLE I m s = 1 , R L = 0 . , C m ( J, J ) , δ = 10 − J ε ∗ m s +1 ε ∗ L − m s ε ∗ BP . . . . . . . . . . . . . . . . . . . . . . . . TABLE II m s = 2 , R L = 0 . , C m ( J, J ) , δ = 10 − J ε ∗ m s +1 ε ∗ L − m s ε ∗ BP . . . . . . . . . . . . . . . . . . . . . . . . gap between ε ∗ L − m s and ε ∗ BP with increasing J illustrates theloss due to edge multiplicities (“weaker” termination). This isbecause the terminations at the beginning and at the end ofthe code are different, i.e. the CN degrees in the terminatedportion at the beginning of the code are J − m s ) whichincreases with J ; whereas those at the end of the code are , a constant. Thus, much of the code performance is determinedby the “stronger” (smaller check-degree) termination, the oneat the end of the code for J > m s + 1 . This is also seen bythe fact that the gap between ε ∗ m s +1 and ε ∗ L − m s decreases as J increases, meaning that the termination at the beginning ofthe code is weak and increasing the window size helps little.Note that the C m (3 , ensemble in Table II is in fact the C c (3 , classical ensemble, and that ε ∗ L − m s is larger than thecorresponding BP threshold. This is possible since WD onlydemands that the erasure probability of the targeted symbolsis reduced to δ . In contrast, BP demands that the erasureprobability of all the symbols is reduced to . (cid:3) From the above discussion, we can add the following asanother design rule.
Design Rule 2:
For C m ( J, K ) ensembles, keep the ter-mination at the beginning of the code strong, preferablystronger than the one at the end of the code. That is, usepolynomials P = { p j ( x ) , j ∈ [ K (cid:48) ] } such that each of thesums K (cid:48) (cid:88) j =1 p (0) j , · · · , K (cid:48) (cid:88) j =1 p ( J (cid:48) − j is kept as small as possible. Targeted symbols:
We have thus far considered only thefirst K (cid:48) VNs in the sub-protograph contained within thewindow to be the targeted symbols. However, as an alternativeway to trade-off performance for reduced latency, it is possibleto consider other VNs also as targeted symbols. In thiscase, the window would be shifted beyond all the targetedsymbols after processing each window configuration. For awindow of size W , let us denote by ε ∗ i ( B , W, δ ) the windowedthreshold when the targeted symbols are the first iK (cid:48) VNs, ≤ i ≤ W . Hence, ε ∗ ( B , W, δ ) = ε ∗ ( B , W, δ ) . By definition, ε ∗ i ( B , W, δ ) ≤ ε ∗ ( B , W, δ ) . Example 7:
Consider the C m (6 , ensemble with m s = 1 defined by p ( x ) = p ( x ) = 3 + 3 x , denoted C ; and the C m (4 , ensemble with m s = 1 defined by q ( x ) = q ( x ) =2 + 2 x , denoted C . Also consider ensembles C and C givenby r ( x ) = r ( x ) = 2 + 4 x and s ( x ) = s ( x ) = 2 + 2 x +2 x respectively. Both C and C are C m (6 , ensembles,but with memory m s = 1 and respectively. Table III givesthe windowed thresholds ε ∗ i ( C j , W = 4 , δ ) with iK (cid:48) targetedsymbols for a window of size for j = 4 , , , . TABLE IIIW
INDOWED THRESHOLDS ε ∗ i ( C j , W = 4 , δ = 10 − ) , j = 4 , , , i C C C C . . . . . . . . . . . . . . . . One might expect the windowed threshold ε ∗ ( B , W, δ ) to be higher for an ensemble for which ε ∗ W ( B , W, δ ) is higher. This is not quite right: ε ∗ ( C , , − ) ≈ . > . ≈ ε ∗ ( C , , − ) whereas ε ∗ i ( C , , − ) >ε ∗ i ( C , , − ) ∀ i < . This can again be explained as theeffect of stronger termination in C in comparison with C .This is also evident in the larger thresholds for the (6 , ensemble C with same memory as C , but stronger termina-tion. Also, keeping the same termination and increasing thememory improves the performance, as is exemplified by thelarger thresholds of C in comparison with those of C . (cid:3) The windowed thresholds ε ∗ i ( B , W, δ ) quantify the unequalerasure protection of different VNs in the sub-protographwithin the window. Furthermore, it is clear that for good per-formance, it is advantageous to keep fewer targeted symbolswithin a window. B. Finite length performance evaluation
The finite length performance of LDPC codes under iterativemessage-passing decoding over the BEC is dependent on thenumber and the size of stopping sets present in the parity-check matrix of the code [23], [25]. Thus, the performanceof the codes varies based on the parity-check matrix usedto represent the code and, consequently, the performanceof iterative decoding can be made to approach that of MLdecoding by adding redundant rows to the parity-check matrix(See e.g. [26]). However, since we are exploiting the structureof the parity-check matrix of the convolutional code, wewill not be interested in changing the parity-check matrixby adding redundant rows as this destroys the convolutionalstructure. The ensemble stopping set size distribution for someprotograph-based LDPC codes was evaluated in [27] whereit was shown that a minimum stopping set size that growslinearly in blocklength is important for the good performanceof codes with short blocklengths. This analysis is similar to theanalysis of the minimum distance growth rate of LDPC-CCensembles—see [28] and references therein. It is worthwhileto note that although the minimum stopping set size growslinearly for protograph codes expanded using random permu-tation matrices, the same is not true for codes expanded usingcirculant permutation matrices [29]. In the following we willevaluate the finite length performance of codes constructedfrom C m ( J, K ) ensembles with BP and WD through MonteCarlo simulations. WD was considered with only the first K (cid:48) M symbols as the targeted symbols.In Figs. 4 and 5, the symbol error rate (SER) and thecodeword error rate (CER) performance are depicted for codes C ∈ C and C ∈ C , where the ensembles C and C weredefined in Example 5. The codes used were those constructedby Liva [12] by expanding the protographs using circulantmatrices (and sums of circulant matrices) and techniques ofprogressive edge growth (PEG) [30] and approximate cycleextrinsic message degree (ACE) [31] to avoid small cycles inthe Tanner graphs of the codes. The girth of both the codes C and C was . The parameters used for the constructionwere L = 20 and M = 512 so that the blocklength n = LK (cid:48) M = 20480 and R L = 0 . . The BP thresholdsfor ensembles C and C with L = 20 were . and . respectively. As is clear from Figs. 4 and 5, code C −4 −3 −2 −1 ε S E R C - W = 3 C - W = 3 C - W = 5 C - W = 5 C - W = 10 C - W = 10 C - BP C - BP Fig. 4. SER performance for BP and Windowed Decoding over BEC. −4 −3 −2 −1 " CER C − W = 3 C − W = 3 C − W = 5 C − W = 5 C − W = 10 C − W = 10 C −BP C −BPSB Fig. 5. CER performance for BP and Windowed Decoding over BEC. Alsoshown is the (Singleton) lower bound P SB as SB. outperforms code C for small window sizes ( W = 3 , ),confirming the effectiveness of the proposed design rules forwindowed decoding. For larger window sizes ( W = 10 ),there is no marked difference in the performance of the twocodes. It was also observed that for small M values ( < ),the performance of codes constructed through circulant per-mutation matrices was better than those constructed throughrandom permutation matrices. This difference in performancediminished for larger M values.We include in Fig. 5, for comparison, a lower bound onthe CER P cw . The Singleton bound, P SB , represents theperformance achievable by an idealized ( n, k ) binary MDScode. This bound for the BEC can be expressed as P cw ≥ n (cid:88) j = n − k +1 (cid:18) nj (cid:19) ε j (1 − ε ) n − j = P SB . Note that by the idealized ( n, k ) binary MDS code, we meana binary linear code that achieves the Singleton bound d min ≤ n − k + 1 with equality. This code does not exist for all valuesof k and n .V. E RASURE C HANNELS WITH M EMORY
We now consider the performance of LDPC-CC ensemblesand codes over erasure channels with memory. We considerthe familiar two-state Gilbert-Elliott channel (GEC) [32], [33]as a model of an erasure channel with memory. In this model,the channel is either in a “good” state G , where we assume theerasure probability is , or in an “erasure” state E , in whichthe erasure probability is . The state process of the channel isa first-order Markov process with the transition probabilities P { E → G } = g and P { G → E } = b . With these parameters,we can easily deduce [34] that the average erasure rate ε andthe average burst length ∆ are given by ε = P { E } = bb + g , ∆ = 1 g . We will consider the GEC to be parameterized by the pair ( ε, ∆) . Note that there is a one-to-one correspondence betweenthe two pairs ( b, g ) and ( ε, ∆) . Discussion:
The channel capacity of a correlated binaryerasure channel with an average erasure rate of ε is given as (1 − ε ) , which is the same as that of the memoryless channel,provided the channel is ergodic. Therefore, one can obtaingood performance on a correlated erasure channel through theuse of a capacity-achieving code for the memoryless channelwith an interleaver to randomize the erasures [27], [35]. This isequivalent to permuting the columns of the parity-check matrixof the original code. We are not interested in this approachsince such permutations destroy the convolutional structure ofthe code and as a result, we are unable to use the WD forsuch a scheme.Construction of LDPC block codes for bursty erasure chan-nels has been well studied. The performance metric of acode over a bursty erasure channel is related to the maximumresolvable erasure burst length (MBL) denoted ∆ max [35],which, as the name suggests, is the maximal length of a singlesolid erasure burst that can be decoded by a BP decoder.Methods of optimizing codes for such channels therefore focuson permuting columns of parity-check matrices to maximize ∆ max , e.g. [36]–[41]. Instead of permuting columns of theparity-check matrix, in order to maintain the convolutionalstructure of the code, we will consider designing C m ( J, K ) ensembles that maximize ∆ max . A. Asymptotic Analysis1) BP : As noted earlier, the performance of LDPC-CC en-sembles depends on stopping sets. The structure of protographsimposes constraints on the code that limit the stopping set sizesand locations, as will be shown shortly.Let us define a protograph stopping set to be a subset S ( B ) of the VNs of the protograph B whose neighboringCNs are connected at least twice to S ( B ) . These are alsodenoted as S ( P ) , in terms of the set of polynomials definingthe protograph. We define the size of the stopping set as the cardinality of S ( B ) , denoted | S ( B ) | . We call the leastnumber of consecutive columns of B that contain the stoppingset S ( B ) the span of the stopping set, denoted (cid:104) S ( B ) (cid:105) . Letus denote the size of the smallest protograph stopping setof the protograph B by | S ( B ) | ∗ , and the minimum numberof consecutive columns of the protograph B that contain aprotograph stopping set by (cid:104) S ( B ) (cid:105) ∗ . When the protographunder consideration is clear from the context, we will drop itfrom the notation and use | S | ∗ and (cid:104) S (cid:105) ∗ . The minimum spanof a stopping set is of interest because we can give simplebounds for ∆ max based on (cid:104) S ( B ) (cid:105) ∗ . Note that the stoppingset of minimal size and the stopping set of minimal span arenot necessarily the same set of VNs. However, we always have | S ( B ) | ∗ ≤ (cid:104) S ( B ) (cid:105) ∗ . The following example clarifies the notation.
Example 8:
Let us denote the base matrix corresponding tothe protograph of the ensembles C i of Example 5 as B ( i ) , i =1 , , . For ensembles C and C , the first two columns of B ( i ) , i = 1 , form a protograph stopping set, i.e. S ( B ( i ) ) = { V , V } , i = 1 , is a stopping set. This is clear from thehighlighted columns below B (1) = · · · · · · · · · · · · · · · ... ... ... ... ... ... . . . , B (3) = · · · · · · · · · · · · ... ... ... ... ... ... . . . . Therefore, | S ( B ( i ) ) | ∗ ≤ and (cid:104) S ( B ( i ) ) (cid:105) ∗ ≤ . Since nosingle column forms a protograph stopping set, | S ( B ( i ) ) | ∗ ≥ and (cid:104) S ( B ( i ) ) (cid:105) ∗ ≥ , implying | S ( B ( i ) ) | ∗ = (cid:104) S ( B ( i ) ) (cid:105) ∗ =2 , i = 1 , .For ensemble C , the highlighted columns of B (2) inthe following matrix form a protograph stopping set, i.e. S ( B (2) ) = { V , V } is a stopping set. B (2) = · · · · · · · · · · · · · · · ... ... ... ... ... ... . . . . Thus, | S ( B (2) ) | ∗ ≤ and (cid:104) S ( B (2) ) (cid:105) ∗ ≤ . As no singlecolumn of B (2) is a protograph stopping set and no threeconsecutive columns of B (2) contain a protograph stoppingset, it is clear that | S ( B (2) ) | ∗ ≥ and (cid:104) S ( B (2) ) (cid:105) ∗ ≥ , sothat | S ( B (2) ) | ∗ ≤ (cid:104) S ( B (2) ) (cid:105) ∗ = 4 . (cid:104) S l ,l (cid:105) = K (cid:48) ( m s ( l , l ) −
1) + ( l − l + 1) , i ( l , l ) = i l ≤ i l , j l ≤ j l = j ( l , l ) K (cid:48) ( m s ( l , l ) −
1) + 1 , i ( l , l ) = i l ≤ i l , j l < j l = j ( l , l ) K (cid:48) ( m s ( l , l ) −
1) + 1 , i ( l , l ) = i l < i l , j l ≤ j l = j ( l , l ) K (cid:48) ( m s ( l , l ) − − ( l − l − , i ( l , l ) = i l < i l , j l < j l = j ( l , l ) . (7)In these cases, it so happened that the stopping set with theminimal size and the stopping set with the minimal span werethe same. (cid:3) Our aim in the following will be to obtain bounds for themaximal (cid:104) S ( B ) (cid:105) ∗ over C m ( J, K ) ensembles with memory m s ,which we denote (cid:104) S ( J, K, m s ) (cid:105) ∗ , and design protographs thatachieve minimal spans close to this optimal value.The analysis of the minimal span of stopping sets for un-structured LDPC ensembles was performed in [42]. However,the structure of the protograph-based LDPC-CC allows us toobtain (cid:104) S ( J, K, m s ) (cid:105) ∗ much more easily for some C m ( J, K ) ensembles.We start by observing that if one of the VNs in theprotograph is connected multiple times to all its neighboringCNs, then it forms a protograph stopping set by itself. In orderto obtain a larger minimum span of stopping sets, it is desirableto avoid this case, and we include this as one of our designcriteria. Design Rule 3:
For a C m ( J, K ) ensemble, choose thepolynomials p j ( x ) such that for every j ∈ [ K (cid:48) ] , thereexists ≤ i j ≤ ( m s + 1) J (cid:48) − such that p ( i j ) j = 1 .Using the polynomial representation of LDPC-CC ensem-bles is helpful in this case since we can easily track stoppingsets as those subsets that have polynomials whose coefficientsare all larger than . From this fact, we can prove thefollowing. Proposition 2 ( (cid:104) S (cid:105) ∗ for C m ( J, J ) protographs): For C m ( J, J ) protographs of memory m s defined by polynomials p ( x ) and p ( x ) , (cid:104) S (cid:105) ∗ can be upper bounded as (cid:104) S (cid:105) ∗ ≤ m s , i ≤ i , j ≤ j = m s m s − , i ≤ i , j < j = m s m s − , i < i , j ≤ j = m s m s − , i < i , j < j = m s where i l = min deg( p l ( x )) and j l = deg( p l ( x )) , l = 1 , . (cid:4) We give the proof in Appendix II. We see from the abovethat (cid:104) S ( J, J, m s ) (cid:105) ∗ ≤ m s and a necessary condition forachieving this span is the first of four possible cases listedabove, which we include as another design criterion. Design Rule 4:
For C m ( J, J ) ensembles with mem-ory m s , set min deg( p ( x )) = 0 and deg( p ( x )) = m s . Corollary 3 (Optimal C m ( J, J ) protographs): For C m ( J, J ) protographs with memory m s and J > , (cid:104) S ( J, J, m s ) (cid:105) ∗ = 2 m s . (cid:4) The proof is given in Appendix III. Note that ensemble C in Example 5 achieves (cid:104) S ( J, J, m s ) (cid:105) ∗ , as was observed inExample 8. It also satisfies design rules 1, 3 and 4. We bringto the reader’s attention here that constructions other than theone given in the proof of the above corollary that achieve (cid:104) S (cid:105) ∗ = 2 m s are also possible. These constructions allow usto design C m ( J, J ) ensembles for a wide range of required (cid:104) S (cid:105) ∗ . We quickly see that a drawback of the convolutionalstructure is that if m s is increased to obtain a larger (cid:104) S (cid:105) ∗ , thecode rate R L decreases linearly for a fixed L .We give without proof the following upper bound for (cid:104) S (cid:105) ∗ for C m ( J, K (cid:48) J ) ensembles, as it follows from Proposition 2. Proposition 4 ( (cid:104) S (cid:105) ∗ for C m ( J, K (cid:48) J ) protographs): For C m ( J, K (cid:48) J ) protographs defined by polynomials P = { p j ( x ) , j ∈ [ K (cid:48) ] } , we have (cid:104) S (cid:105) ∗ ≤ min ( l ,l ) ∈ [ K (cid:48) ] ,l By looking at the stopping sets confined withincolumns corresponding to two polynomials only, we can useProposition 2 to upper bound the span of these stopping sets.The minimal such span over all possible choices of the twocolumns therefore gives an upper bound on the minimal spanof the ( J, K (cid:48) J ) protograph. Since (cid:104) S l ,l (cid:105) ≤ K (cid:48) m s ∀ l , l from Equation (7), we have (cid:104) S ( J, K (cid:48) J, m s ) (cid:105) ∗ ≤ K (cid:48) m s , whichis similar to the result in Proposition 2. This bound is, however,loose in general.For terminated codes, we can give an upper bound for (cid:104) S (cid:105) ∗ that is tighter in some cases. Corollary 5 ( (cid:104) S (cid:105) ∗ for C m ( J, K ) protographs): For C m ( J, K ) protographs terminated after L instants, (cid:104) S (cid:105) ∗ ≤ K (cid:48) L . Proof: From the Singleton bound for the protograph, wehave (cid:104) S (cid:105) ∗ ≤ J (cid:48) ( L + m s ) . Since we need m s ≤ R − R L for apositive code rate in (2), (cid:104) S (cid:105) ∗ ≤ J (cid:48) L − R = K (cid:48) L .Note that for C m ( J, K (cid:48) J ) protographs, this is tighter thanthe bound (cid:104) S (cid:105) ∗ ≤ K (cid:48) m s ≤ K (cid:48) ( K (cid:48) − L , which, in the worstcase, is a factor ( K (cid:48) − times larger. However, since we areinterested mainly in ensembles for which m s (cid:28) L , this bound might be looser than the one in Proposition 4 for C m ( J, K (cid:48) J ) ensembles. Example 9: Consider the C m ( J, K (cid:48) J ) ensemble with mem-ory m s = u ( K (cid:48) − 1) + 1 , m s ≤ ( K (cid:48) − L defined by thepolynomials p l ( x ) = ( J − 1) + x j l , l ∈ [ K (cid:48) ] ,j l = m s − u ( l − . It can be shown by an argument similarto the one used to prove Corollary 3 that for the protographof this ensemble, (cid:104) S (cid:105) ∗ = K (cid:48) u + 2 . This is exactly the boundin Proposition 4 since min l 1) + 2 = (cid:24) K (cid:48) K (cid:48) − m s (cid:25) which is roughly only a fraction of the (loose) upper bound for (cid:104) S ( J, K (cid:48) J, m s ) (cid:105) ∗ suggested in the discussion of Proposition4. The constructed C m ( J, K (cid:48) J ) protographs are thus optimalin the sense of maximizing the minimal span of stopping sets,i.e. (cid:104) S ( J, K (cid:48) J, u ( K (cid:48) − 1) + 1) (cid:105) ∗ = K (cid:48) u + 2 ∀ u ∈ [ L − . They also satisfy Design Rules 1 and 3 for J > . AlthoughProposition 4 gave a tight bound for (cid:104) S (cid:105) ∗ in this case, it isloose in general. (cid:3) We can show that the C m ( J, K ) protographs have minimalspans at least as large as the corresponding spans of C m ( a, K ) protographs. Proposition 6: (cid:104) S ( J, K, m s ) (cid:105) ∗ ≥ (cid:104) S ( a, K, m s ) (cid:105) ∗ where a = gcd( J, K ) ≥ . Proof: The equality is trivial when a = J . When ≤ a < J = aJ (cid:48) , one way of constructing the C m ( J, K ) ensembles with memory m s is to let each set of modulopolynomials P l themselves define C m ( a, K ) ensembles withmemory m s . The result then follows by noting that a stoppingset for the polynomials P has to be a stopping set for everyset of polynomials P l , l = 0 , , · · · , J (cid:48) − .The construction proposed above often allows us to strictlyincrease the minimal span of the C m ( J, K ) ensemble incomparison with the C m ( a, K ) ensemble, as illustrated by thefollowing example. Example 10: Consider the construction of a C m (4 , en-semble with memory . Let us call it C . The differentparameters in this case are J = 4 , K = 6 , a = 2 , J (cid:48) = 2 , K (cid:48) = 3 and m s = 3 . Since m s = u ( K (cid:48) − 1) + 1 with u = 1 , we have for C m (2 , protographs, (cid:104) S (2 , , (cid:105) ∗ = 5 from Example 9 and we will define the modulo polynomials P to be the optimal construction that achieves this minimalspan, i.e. P = { x , x , x } . Then, by defining P = { x , x , x } , we can show that (cid:104) S ( P ) (cid:105) ∗ = 6 and hence (cid:104) S (4 , , (cid:105) ∗ ≥ > (cid:104) S (2 , , (cid:105) ∗ . Note that the protograph defined by P has no degree- VNs associated with the component matrix B . In fact, theconstructed C m (4 , ensemble has ε ∗ ( C , m s + 1 , − ) ≈ . , fairly close to the Shannon limit of ε Sh = , even withthe smallest possible window size. Table IV lists the windowed TABLE IV ε ∗ i ( C , m s + 1 , − ) i ε ∗ i . . . . thresholds of this ensemble with different numbers of targetedsymbols within the smallest window for δ = 10 − . (cid:3) WD : The asymptotic analysis for WD is essentiallythe same as that for BP. We will consider WD with onlythe first K (cid:48) symbols within each window as the targetedsymbols. We are now interested in the sub-protograph stoppingsets, denoted S ( B , W ) , that include one or more of thetargeted symbols within a window. Let us denote the minimalspan of such stopping sets as (cid:104) S ( B , W ) (cid:105) ∗ . Since stoppingsets of the protograph of the LDPC-CC are also stoppingsets of the sub-protograph within a window, and since suchstopping sets can be chosen to include some targeted symbolswithin the window, we have (cid:104) S ( B , W ) (cid:105) ∗ ≤ (cid:104) S ( B ) (cid:105) ∗ . In fact, (cid:104) S ( B , W ) (cid:105) ∗ = (cid:104) S ( B ) (cid:105) ∗ when W ≥ (cid:24) (cid:104) S (cid:105) ∗ K (cid:48) (cid:25) + m s since in this case the first K (cid:48) (cid:108) (cid:104) S (cid:105) ∗ K (cid:48) (cid:109) columns are completelycontained in the window. Further, we have (cid:104) S ( B , W ) (cid:105) ∗ ≤ (cid:104) S ( B , W + 1) (cid:105) ∗ . This is true because a stopping set for window size W involving targeted symbols is not necessarily a stopping setfor window size W + 1 , whereas a stopping set for windowsize W + 1 is definitely a stopping set for window size W . Remark: When the first iK (cid:48) symbols within a window arethe targeted symbols, we have for i ≤ W − m s (cid:104) S i ( B , W ) (cid:105) ∗ = (cid:104) S ( B , W − i + 1) (cid:105) ∗ where (cid:104) S i ( B , W ) (cid:105) ∗ denotes the minimal span of stopping setsof the sub-protograph within the window of size W involvingat least one of the iK (cid:48) targeted symbols, and (cid:104) S ( B , W ) (cid:105) ∗ = (cid:104) S ( B , W ) (cid:105) ∗ . Consequently, we have (cid:104) S i ( B , W ) (cid:105) ∗ ≤ (cid:104) S ( B , W ) (cid:105) ∗ . The definition of (cid:104) S i ( B , W ) (cid:105) ∗ can be extended to accommo-date W − m s +1 ≤ i ≤ W , as in the case of windowed thresh-olds. In particular, we have (cid:104) S W ( B , W ) (cid:105) ∗ = (cid:104) S ( B ) (cid:105) ∗ ≤ J (cid:48) ,where the last inequality is from the Singleton bound. Example 11: Consider the ensemble C defined in Example5. With a window of size W = m s + 1 = 3 , we have (cid:104) S ( C , (cid:105) ∗ = 2 with the corresponding stopping set S = { V , V } highlighted below and with a window size W = 4 , we have (cid:104) S ( C , (cid:105) ∗ = (cid:104) S ( C ) (cid:105) ∗ = 4 , and the corresponding stopping sets S = { V , V } and S (cid:48) = { V , V , V } are as follows , . Note that for window size , whereas the minimal span ofa stopping set involving VN V is , that of a stopping setinvolving V is . However, for window size , the stoppingset involving V with minimal span, denoted S , and thatinvolving V , S (cid:48) , each have a span of , although theircardinalities are and respectively. We have in this case, S ⊂ S (cid:48) . Notice that (cid:104) S ( C , (cid:105) ∗ = (cid:104) S ( C , (cid:105) ∗ = 2 . (cid:3) B. Finite length analysis1) BP : We now show the relation between the parameters ∆ max and (cid:104) S ( B ) (cid:105) ∗ . We shall assume in the following that (cid:104) S (cid:105) ∗ ≥ , i.e. every column of the protograph has at least oneof the entries equal to . We will consider the expansion ofthe protographs by a factor M to obtain codes. Proposition 7: For any ( J, K ) regular LDPC-CC, ∆ max ≤ M (cid:104) S (cid:105) ∗ − . Proof: Clearly, the set of the M (cid:104) S (cid:105) ∗ columns of theparity-check matrix corresponding to the (cid:104) S (cid:105) ∗ consecutivecolumns of B that contain the protograph stopping set withminimal span must contain a stopping set of the parity-check matrix. Therefore, if all symbols corresponding to thesecolumns are erased, they cannot be retrieved. Corollary 8: A terminated C m ( J, K (cid:48) J ) LDPC-CC with m s = u ( K (cid:48) − 1) + 1 , u ∈ [ L − can never achieve theMBL of an MDS code. Proof: From the Singleton bound, we have ∆ max ≤ n − k = ( L + m s ) M , assuming that the parity-check matrix isfull-rank. From Proposition 7 we have, ∆ max ≤ M (cid:104) S ( J, K (cid:48) J, u ( K (cid:48) − 1) + 1) (cid:105) ∗ − (cid:24) K (cid:48) K (cid:48) − m s (cid:25) M − where the second equality follows from the discussion inExample 9. Since we require m s ≤ R − R L = ( K (cid:48) − L for anon-negative code rate in (2), ∆ max ≤ (cid:24) ( K (cid:48) − m s + ( K (cid:48) − LK (cid:48) − (cid:25) M − < ( L + m s ) M which shows that the MBL of an MDS code can never beachieved. Remark: Although the idealized binary ( n, k ) MDS codedoes not exist, there are codes that achieve MDS performance when used over a channel that introduces a single burst oferasures in a codeword. For example, the (2 n, n ) code with aparity-check matrix H = [ I n I n ] has an MBL of ∆ max = n .Despite the discouraging result from Corollary 8, we canguarantee an MBL that linearly increases with (cid:104) S (cid:105) ∗ as follows. Proposition 9: For any ( J, K ) regular LDPC-CC, ∆ max ≥ M ( (cid:104) S (cid:105) ∗ − 2) + 1 . Proof: From the definition of (cid:104) S (cid:105) ∗ , it is clear that if one ofthe two extreme columns is completely known, all other sym-bols can be recovered, for otherwise the remaining columnswithin the span of the stopping set S will have to containanother protograph stopping set, violating the minimality ofthe stopping set span (cid:104) S (cid:105) ∗ (The two extreme columns are pivots of the stopping set [41].) The largest solid burst thatis guaranteed to have at least one of the extreme columnscompletely known is of length M ( (cid:104) S (cid:105) ∗ − 2) + 1 . Therefore, ∆ max ≥ M ( (cid:104) S (cid:105) ∗ − 2) + 1 . Example 12: For the C m ( J, K (cid:48) J ) ensemble with memory m s = u ( K (cid:48) − 1) + 1 , u ∈ [ L − in Example 9, we have M K (cid:48) (cid:18) m s − K (cid:48) − (cid:19) + 1 ≤ ∆ max ≤ M K (cid:48) (cid:18) m s − K (cid:48) − (cid:19) + 2 M − from Propositions 7 and 9. Thus, we can construct codes withMBL proportional to m s . (cid:3) WD : The MBL for WD ∆ max ( W ) can be bounded asin the case of BP based on (cid:104) S ( B , W ) (cid:105) ∗ . Assuming that thewindow size is W ≥ m s + 1 , the targeted symbols are the first K (cid:48) symbols within the window, and the polynomials definingthe ensemble are chosen to satisfy Design Rule 3, we have (cid:104) S ( B , W ) (cid:105) ∗ ≥ . Propositions 7 and 9 in this case imply that M ( (cid:104) S ( B , W ) (cid:105) ∗ − 2) + 1 ≤ ∆ max ( W ) ≤ M (cid:104) S ( B , W ) (cid:105) ∗ − . C. Numerical results The MBL for codes C and C (the same codes usedin Section IV-B) was computed using an exhaustive searchalgorithm, by feeding the decoder with a solid burst of erasuresand testing all the possible locations of the burst. The MBLfor the codes we considered was and for codes C and C , respectively. Note that for code C , the MBL ∆ max =1023 = 2 M − , i.e., code C achieves the upper boundfrom Proposition 7. More importantly, the maximum possible ∆ max was achievable while maintaining good performanceover the BEC with the BP decoder. However, the MBL forcode C , ∆ max = 1751 < M − , is much smallerthan the corresponding bound from Proposition 7. In this case,although other code constructions with ∆ max up to werepossible, a trade-off between the BEC performance and MBLwas observed, i.e. the code that achieved ∆ max = 2045 wasfound to be much worse over the BEC than both codes C and C considered here. Such a trade-off has also been observedby others, e.g. [40]. This could be because the codes thatachieve large ∆ max are often those that have a very regularstructure in their parity-check matrices. Nevertheless, our codedesign does give a large increase in MBL ( > ) whencompared with the corresponding codes constructed from C c ensembles, without any decrease in code rate (same m s ). TheMBL achieved as a fraction of the maximum possible MBL ∆ max / ( n − k ) was roughly . and . for codes C and C , respectively.In Figs 6, 7 and 8, we show the CER performance obtainedfor codes C and C over GEC channels with ∆ = 10 , and respectively, and ε ∈ [0 . , . . As can be seen from −4 −3 −2 −1 " CER C − W = 3 C − W = 3 C − W = 5 C − W = 5 C − W = 10 C − W = 10 C −BP C −BPSB Fig. 6. CER Performance on GEC with ∆ = 10 with Singleton bound (SB). −4 −3 −2 −1 " CER C − W = 3 C − W = 3 C − W = 5 C − W = 5 C − W = 10 C − W = 10 C −BP C −BPSB Fig. 7. CER Performance on GEC with ∆ = 50 with Singleton bound (SB). the figures, for W = 3 , code C always outperforms code C , while for W = 5 there is no such gain when ∆ = 100 .However, for W = 10 and for BP decoding, code C slightlyoutperforms C .Note that the code C outperforms C for small ε when theaverage burst length ∆ = 100 for large window sizes and forBP decoding. This can be explained because in this regime,the probability of a burst is small but the average burst lengthis large. Therefore, when a burst occurs, it is likely to resemblea single burst in a codeword, and in this case we know thatthe code C is stronger than C . Also note the significantgap between the BP decoder performance and the Singleton −4 −3 −2 −1 " CER C − W = 3 C − W = 3 C − W = 5 C − W = 5 C − W = 10 C − W = 10 C −BP C −BPSB Fig. 8. CER Performance on GEC with ∆ = 100 with Singleton bound(SB). bound, suggesting that unlike some moderate length LDPCblock codes with ML decoding [38], LDPC-CC are far fromachieving MDS performance with BP or windowed decoding.VI. C ONCLUSIONS We studied the performance of a windowed decodingscheme for LDPC convolutional codes over erasure channels.We showed that this scheme, when used to decode terminatedLDPC-CC, provides an efficient way to trade-off decodingperformance for reduced latency. Through asymptotic perfor-mance analysis, several design rules were suggested to avoidbad structures within protographs and, in turn, to ensure goodthresholds. For erasure channels with memory, the asymptoticperformance analysis led to design rules for protographs thatensure large stopping set spans. Examples of LDPC-CC en-sembles that satisfy design rules for the BEC as well as erasurechannels with memory were provided. Finite length codesbelonging to the constructed ensembles were simulated and thevalidity of the design rules as markers of good performancewas verified. The windowed decoding scheme can be used todecode LDPC-CC over other channels that introduce errorsand erasures, although in this case error propagation due towrong decoding within a window will have to be carefullydealt with.For erasure channels, while close-to-optimal performance(in the sense of approaching capacity) was achievable for theBEC, we showed that the structure of LDPC-CC imposed con-straints that bounded the performance over erasure channelswith memory strictly away from the optimal performance (inthe sense of approaching MDS performance). Nevertheless, thesimple structure and good performance of these codes, as wellas the latency flexibility and low complexity of the decodingalgorithm, are attractive characteristics for practical systems.A CKNOWLEDGMENT The authors are very grateful to the anonymous reviewersfor their comments and suggestions for improving the presen-tation of the paper. They would also like to thank Gianluigi Liva who suggested the trade-off of decoding performance forreduced latency through the use of a windowed decoder, andR¨udiger Urbanke for pointing out an error in an earlier versionof the paper. A PPENDIX IP ROOF OF P ROPOSITION i th window configuration for window sizes W and W + 1 shown in Fig. 9. We are interested in a windowconfiguration that is not at the terminated portion of the code.Call the Tanner graphs of these windows A = ( V A , C A , E A ) W W + 1 Fig. 9. Sub-protographs of window sizes W and W +1 . The edges connectedto targeted symbols from previous window configurations are shown in darkershade of gray. and B = ( V B , C B , E B ) respectively, where V A , V B and C A , C B are the sets of VNs and CNs respectively and E A , E B are the sets of edges. Clearly, V A ⊂ V B , C A ⊂ C B , and E A ⊂ E B . Any VN in V A that is connected to some variablein V B \ V A has to be connected via some CN in C B \ C A . Theedges between these CNs and VNs in V A are shown hatchedin Fig. 9. Consider the computation trees for the a-posteriorimessage at a targeted symbol in V A and that for the samesymbol in V B . Call them T A and T B respectively. Then wehave T A ⊂ T B .We now state two lemmas which will be made use ofsubsequently. The proofs of these lemmas are straightforwardand have been omitted. Lemma 10 (Monotonicity of C ): The CN operation in (5) ismonotonic in its arguments, i.e., ≤ x ≤ x (cid:48) ≤ ⇒ C ( x, y ) ≤ C ( x (cid:48) , y ) ∀ y ∈ [0 , , where the two-argument function C ( x, y ) = xy . (cid:4) Lemma 11 (Monotonicity of V ): The VN operation in (6) ismonotonic in its arguments, i.e., ≤ x ≤ x (cid:48) ≤ ⇒ V ( x, y ) ≤ V ( x (cid:48) , y ) ∀ y ∈ [0 , , where V ( x, y ) = 1 − (1 − x )(1 − y ) . (cid:4) The operational significance of the above lemmas is the fol-lowing: if we can upper (lower) bound the mutual informationon some incoming edge of a CN or a VN, and use the boundto compute the outgoing mutual information from that node,we get an upper (lower) bound on the actual outgoing mutualinformation. Thus, by bounding the mutual information onsome edges of a computation tree and repetitively applying Lemmas 10 and 11, one can obtain bounds for the a-posteriorimutual information at the root of the tree.We start by augmenting T A , creating another computationtree T + A that has the same structure as T B . In particular, T + A includes the additional edges corresponding to the hatchedregion. In T + A and T B , we denote the set of these edgesby E u ( T + A ) and E u ( T B ) respectively. In T + A , we assign zeromutual information to each edge in E u ( T + A ) .Now, let I T A , I T + A and I T B be the a-posteriori mutual infor-mation at the roots of the trees T A , T + A and T B respectively.Then it is clear that I T A = I T + A , since the messages onedges in E u ( T + A ) are effectively erasures and zero out thecontributions from the checks in C + A \ C A = C B \ C A .On the other hand, if we denote by I e ( T B ) the mutual infor-mation associated with an edge e ∈ E u ( T B ) , and by I e ( T + A ) the mutual information associated with the corresponding edgein T + A , we know that I e ( T + A ) = 0 so that I e ( T + A ) ≤ I e ( T B ) .Hence, we have from Lemmas 10 and 11 that I T + A ≤ I T B .Since I T A = I T + A , it follows that I T A ≤ I T B , as desired.A PPENDIX IIP ROOF OF P ROPOSITION ≤ i l < j l ≤ m s , l = 1 , . We assume i l < j l inorder to satisfy Design Rule 3. Since the code has memory m s ,we have i = min { i , i } = 0 and j = max { j , j } = m s .Consider the subset of columns of B corresponding to thepolynomial r ( x ) = p ( x ) b ( x ) + p ( x ) b ( x ) where b ( x ) = (cid:40) x i , i = j − x i + x i +1 + · · · + x j − , i < j − and b ( x ) = (cid:40) x i , i = j − x i + x i +1 + · · · + x j − , i < j − . We claim that this is a protograph stopping set. To see this,consider the columns corresponding to the above subset with β ( p ( x )) and β ( p ( x )) as the column polynomials defining B . We have ˆ r ( x ) = β ( p ( x )) b ( x ) + β ( p ( x )) b ( x )= ( x i + x j )( x i + x i +1 + · · · + x j − )+ ( x i + x j )( x i + x i +1 + · · · + x j − )= x i + i + x i + i +1 + · · · + x i + j − + x j + i + x j + i +1 + · · · + x j + j − + x i + i + x i + i +1 + · · · + x j + i − + x i + j + x i + j +1 + · · · + x j + j − = 2 x i + i + · · · + 2 x i + j − + 2 x i + j + · · · + 2 x j + j − when j l > i l + 1 , l = 1 , . Similarly, it can be verified that ˆ r ( x ) has all coefficients equal to in all other cases also.Clearly, ˆ r ( x ) (cid:22) r ( x ) and thus r ( x ) can only differ from ˆ r ( x ) in having larger coefficients. Therefore, r ( x ) also hasall coefficients greater than . This shows that the chosen subset of columns form a protograph stopping set. Based onthe parameters i l , j l , l = 1 , , we can count the number ofcolumns included in the span of this stopping set and thereforegive upper bounds on (cid:104) S (cid:105) ∗ as claimed : (cid:104) S (cid:105) ∗ ≤ j − i ) , i ≤ i , j ≤ j = m s j − i ) − , i ≤ i , j < j = m s j − i ) − , i < i , j ≤ j = m s j − i − , i < i , j < j = m s . A PPENDIX IIIP ROOF OF C OROLLARY p ( x ) =( J − 1) + x m s and p ( x ) = ( J − 1) + x . Let the polynomial r ( x ) = p ( x ) a ( x ) + p ( x ) a ( x ) represent an arbitrary subset(chosen from the m s − − non-empty subsets) of the first (2 m s − columns of B , for any choice of polynomials a ( x ) and a ( x ) with coefficients in { , } and maximal degrees ( m s − and ( m s − respectively: a i ( x ) = d i (cid:88) j =0 a ( j ) i x j , i = 1 , , d = m s − , d = m s − where a ( j ) i ∈ { , } and not all a ( j ) i s are zeros. When a ( x ) (cid:54) =0 , let i = deg( a ( x )) . Clearly, r ( x ) is a monic polynomialof degree ( m s + i ) . When a ( x ) = 0 and a ( x ) (cid:54) = 0 , let i = deg( a ( x )) . Then, r ( x ) is a monic polynomial of degree (1+ i ) . Since in both these cases r ( x ) is a monic polynomial,there is at least one coefficient equaling . Thus, (cid:104) S (cid:105) ∗ > m s − . Finally, notice that p ( x ) + x m s − p ( x ) = ( J − 1) + x m s +( J − x m s − + x m s = ( J − 1) + ( J − x m s − + 2 x m s , with all coefficients strictly larger than . Note that p ( x ) cor-responds to the first column of the protograph and x m s − p ( x ) to the m th s column. Thus, we have (cid:104) S (cid:105) ∗ = 2 m s . Since wehave (cid:104) S ( J, J, m s ) (cid:105) ∗ ≤ m s from Proposition 2, we concludethat (cid:104) S ( J, J, m s ) (cid:105) ∗ = 2 m s .R EFERENCES[1] M. Papaleo, A. R. Iyengar, P. H. Siegel, J. K. Wolf, and G. Corazza,“Windowed erasure decoding of LDPC convolutional codes,” in , Cairo, Egypt, Jan. 2010, pp. 78–82.[2] A. R. Iyengar, M. Papaleo, G. Liva, P. H. Siegel, J. K. Wolf, and G. E.Corazza, “Protograph-based LDPC convolutional codes for correlatederasure channels,” in Proc. IEEE Int. Conf. Comm. , Cape Town, SouthAfrica, May 2010, pp. 1–6.[3] G. E. Corazza, A. R. Iyengar, M. Papaleo, P. H. Siegel, A. Vanelli-Coralli, and J. K. Wolf, “Latency constrained protograph-based LDPC-CC,” in , Brest, France, Sep. 2010, pp. 6–10.[4] R. G. Gallager, Low Density Parity Check Codes . Cambridge, Mas-sachusetts: MIT Press, 1963.[5] J. Pearl, Probabilistic Reasoning in Intelligent Systems: Networks ofPlausible Inference . Morgan Kaufmann, San Francisco, 1988.[6] N. Wiberg, “Codes and decoding on general graphs,” Ph.D. dissertation,Linkoping University, Linkoping, Sweden, 1996.[7] D. MacKay and R. Neal, “Near Shannon limit performance of lowdensity parity check codes,” Electronics Letters , vol. 33, no. 6, pp. 457–458, Mar 1997. [8] A. J. Felstrom and K. Zigangirov, “Time-varying periodic convolutionalcodes with low-density parity-check matrix,” IEEE Trans. Inf. Theory ,vol. 45, no. 6, pp. 2181–2191, Sep. 1999.[9] M. Lentmaier, G. P. Fettweis, K. S. Zigangirov, and D. J. Costello,“Approaching capacity with asymptotically regular LDPC codes,” in Proc. Inf. Theory and Applications , San Diego, California, 2009.[10] S. Kudekar, T. Richardson, and R. L. Urbanke, “Threshold saturationvia spatial coupling: Why convolutional LDPC ensembles perform sowell over the BEC,” CoRR , vol. abs/1001.1826, 2010.[11] T. Richardson and R. Urbanke, Modern Coding Theory . CambridgeUniversity Press, New York, 2008.[12] G. Liva, private communication, 2008.[13] M. Lentmaier, A. Sridharan, K. S. Zigangirov, and D. J. Costello, “Ter-minated LDPC convolutional codes with thresholds close to capacity,”in Proc. IEEE Int. Symp. Inf. Theory , Sep. 4-9, 2005, pp. 1372–1376.[14] M. Lentmaier, A. Sridharan, D. J. Costello, and K. S. Zigangirov,“Iterative decoding threshold analysis for LDPC convolutional codes,” IEEE Trans. Inf. Theory , vol. 56, no. 10, pp. 5274–5289, Oct. 2010.[15] A. Pusane, A. J. Feltstrom, A. Sridharan, M. Lentmaier, K. Zigangirov,and D. J. Costello, “Implementation aspects of LDPC convolutionalcodes,” IEEE Trans. Commun. , vol. 56, no. 7, pp. 1060–1069, Jul. 2008.[16] A. Sridharan, “Design and analysis of LDPC convolutional codes,” Ph.D.dissertation, University of Notre Dame, Notre Dame, Indiana, 2005.[17] J. Thorpe, “Low-density parity-check (LDPC) codes constructed fromprotographs,” JPL INP, Tech. Rep., Tech. Rep., Aug. 2003.[18] A. Sridharan, D. V. Truhachev, M. Lentmaier, D. J. Costello, and K. S.Zigangirov, “Distance bounds for an ensemble of LDPC convolutionalcodes,” IEEE Trans. Inf. Theory , vol. 53, no. 12, pp. 4537–4555, Dec.2007.[19] F. R. Kschischang, B. J. Frey, and H.-A. Loeliger, “Factor graphs andthe sum-product algorithm,” IEEE Trans. Inf. Theory , vol. 47, no. 2, pp.498–519, Feb. 2001.[20] M. Luby, M. Mitzenmacher, M. Shokrollahi, and D. Spielman, “Efficienterasure correcting codes,” IEEE Trans. Inf. Theory , vol. 47, no. 2, pp.569–584, Feb 2001.[21] ——, “Improved low-density parity-check codes using irregular graphs,” IEEE Trans. Inf. Theory , vol. 47, no. 2, pp. 585 –598, Feb 2001.[22] T. Richardson and R. Urbanke, “The capacity of low-density parity-check codes under message-passing decoding,” IEEE Trans. Inf. Theory ,vol. 47, no. 2, pp. 599–618, Feb 2001.[23] ——, “Efficient encoding of low-density parity-check codes,” IEEETrans. Inf. Theory , vol. 47, no. 2, pp. 638 –656, Feb 2001.[24] G. Liva and M. Chiani, “Protograph LDPC codes design based on EXITanalysis,” in Proc. IEEE Globecom , Washington, D.C., Nov. 27-30,2007, pp. 3250–3254.[25] C. Di, D. Proietti, I. Telatar, T. Richardson, and R. Urbanke, “Finite-length analysis of low-density parity-check codes on the binary erasurechannel,” IEEE Trans. Inf. Theory , vol. 48, no. 6, pp. 1570 –1579, Jun.2002.[26] J. Han and P. H. Siegel, “Improved upper bounds on stopping redun-dancy,” IEEE Trans. Inf. Theory , vol. 53, no. 1, pp. 90 –104, Jan. 2007.[27] D. Divsalar, S. Dolinar, and C. Jones, “Protograph LDPC codes overburst erasure channels,” JPL INP, Tech. Rep., Tech. Rep., Oct. 2006.[28] M. Lentmaier, D. G. M. Mitchell, G. P. Fettweis, and D. J. Costello,“Asymptotically regular LDPC codes with linear distance growth andthresholds close to capacity,” in Proc. Inf. Theory and Applications 2010 ,San Diego, California, 2010.[29] B. K. Butler and P. H. Siegel, “On distance properties of quasi-cyclicprotograph-based LDPC codes,” in Proc. IEEE Int. Symp. Inf. Theory ,Austin, TX, USA, Jun. 13-18, 2010, pp. 809–813.[30] X.-Y. Hu, E. Eleftheriou, and D. Arnold, “Regular and irregular pro-gressive edge-growth Tanner graphs,” IEEE Trans. Inf. Theory , vol. 51,no. 1, pp. 386–398, Jan. 2005.[31] T. Tian, C. Jones, J. Villasenor, and R. Wesel, “Selective avoidance ofcycles in irregular LDPC code construction,” IEEE Trans. Commun. ,vol. 52, no. 8, pp. 1242–1247, Aug. 2004.[32] E. Gilbert, “Capacity of a burst-noise channel,” Bell Syst. Tech. J , vol. 39,pp. 1253–1265, Sep. 1960.[33] E. Elliott, “Estimates of error rates for codes on burst-noise channels,” Bell Syst. Tech. J , vol. 42, pp. 1977–1997, Sep. 1963.[34] L. Wilhelmsson and L. B. Milstein, “On the effect of imperfect inter-leaving for the Gilbert-Elliott channel,” IEEE Trans. Commun. , vol. 47,no. 5, pp. 681–688, May 1999.[35] M. Yang and W. E. Ryan, “Design of LDPC codes for two-statefading channel models,” in The 5th International Symposium on WirelessPersonal Multimedia Communications , vol. 3, Oct. 2002, pp. 986–990. [36] S. J. Johnson and T. Pollock, “LDPC codes for the classic burstychannel,” in Proc. IEEE Int. Symp. Inf. Theory , Aug. 2004, pp. 184–189.[37] E. Paolini and M. Chiani, “Improved low-density parity-check codes forburst erasure channels,” in Proc. IEEE Int. Conf. Comm. , Istanbul, Jun.2006, pp. 1183–1188.[38] G. Liva, B. Matuz, Z. Katona, E. Paolini, and M. Chiani, “On construc-tion of moderate-length LDPC codes over correlated erasure channels,”in Proc. IEEE Int. Conf. Comm. , Dresden, Jun. 2009, pp. 1–5.[39] G. Sridharan, A. Kumarasubramanian, A. Thangaraj, and S. Bhashyam,“Optimizing burst erasure correction of LDPC codes by interleaving,”in Proc. IEEE Int. Symp. Inf. Theory , Jul. 2008, pp. 1143–1147.[40] S. J. Johnson, “Burst erasure correcting LDPC codes,” IEEE Trans.Commun. , vol. 57, no. 3, pp. 641–652, Mar. 2009.[41] E. Paolini and M. Chiani, “Construction of near-optimum burst erasurecorrecting low-density parity-check codes,” IEEE Trans. Commun. ,vol. 57, no. 5, pp. 1320–1328, May 2009.[42] T. Wadayama, “Ensemble analysis on minimum span of stopping sets,”in