On the Growth Rate of the Weight Distribution of Irregular Doubly-Generalized LDPC Codes

Recently, low-density parity-check (LDPC) codes have been intensively studied due to their near-Shannon-limit performance under iterative belief-propagation decoding. Binary regular LDPC codes were first proposed by Gallager in 1963 [1]. In the last decade the capability of irregular LDPC codes to outperform regular ones in the waterfall region of the performance curve and to asymptotically approach (or even achieve) the communication channel capacity has been recognized and deeply investigated (see for instance [2, 3, 4, 5, 6, 7]).

It is usual to represent an LDPC code as a bipartite graph, i.e., as a graph where the nodes are grouped into two disjoint sets, namely, the variable nodes (VNs) and the check nodes (CNs), such that each edge may only connect a VN to a CN. The bipartite graph is also known as a Tanner graph [8]. In the Tanner graph of an LDPC code, a generic degree-$q$ VN can be interpreted as a length-$q$ repetition code, as it repeats $q$ times its single information bit towards the CNs. Similarly, a degree-$s$ CN of an LDPC code can be interpreted as a length-$s$ single parity-check (SPC) code, as it checks the parity of the $s$ VNs connected to it.

The growth rate of the weight distribution of Gallager’s regular LDPC codes was investigated in [1]. The analysis demonstrated that, provided that the smallest VN degree is at least 3, for large enough codeword length $N$, the expected minimum distance of a randomly chosen code in the ensemble is a linear function of $N$.

More recently, the study of the weight distribution of binary LDPC codes has been extended to irregular ensembles. Important works in this area are [9, 10, 11]. In [11] a complete solution for the growth rate of the weight distribution of binary irregular LDPC codes was developed. One of the main results of [11] is a connection between the expected behavior of the weight distribution of a code randomly chosen from the ensemble and the parameter $\lambda^{\prime}(0)\rho^{\prime}(1)$, $\lambda(x)$ and $\rho(x)$ being the edge-perspective VN and CN degree distributions, respectively. More specifically, it was shown that for a code randomly chosen from the ensemble, one can expect an exponentially small number of small linear-weight codewords if $0\leq\lambda^{\prime}(0)\rho^{\prime}(1)<1$, and an exponentially large number of small linear-weight codewords if $\lambda^{\prime}(0)\rho^{\prime}(1)>1$.

This result establishes a connection between the statistical properties of the weight distribution of binary irregular LDPC codes and the stability condition of binary irregular LDPC codes over the binary erasure channel (BEC) [3, 4]. If $q^{*}$ denotes the LDPC asymptotic iterative decoding threshold over the BEC, the stability condition states that we always have

Prior to the rediscovery of LDPC codes, binary generalized LDPC (GLDPC) codes were introduced by Tanner in 1981 [8]. A GLDPC code generalizes the concept of an LDPC code in that a degree-$s$ CN may in principle be any $(s,h)$ linear block code, $s$ being the code length and $h$ the code dimension. Such a CN accounts for $s-h$ linearly independent parity-check equations. A CN associated with a linear block code which is not a SPC code is said to be a generalized CN. In [8] regular GLDPC codes (also known as Tanner codes) were investigated, these being GLDPC codes where the VNs are all repetition codes of the same length and the CNs are all linear block codes of the same type.

The growth rate of the weight distribution of binary GLDPC codes was investigated in [12, 13, 14, 15]. In [12] the growth rate is calculated for Tanner codes with BCH check component codes and length-2 repetition VNs, leading to an asymptotic lower bound on the minimum distance. The same lower bound is developed in [13] assuming Hamming CNs and length-2 repetition VNs. Both works extend the approach developed by Gallager in [1, Chapter 2] to show that, for sufficiently large $N$, the minimum distance is a linear function of $N$. The growth rate of the number of small weight codewords for GLDPC codes with a uniform CN set (all CN of the same type) and an irregular VN set (repetition VNs with different lengths) is investigated in [14]. It is shown that for sufficiently large $N$, a minimum distance increasing linearly with $N$ is expected when either the uniform CN set is composed of linear block codes with minimum distance at least $3$, or the minimum length of the repetition VNs is 3. On the other hand, if the minimum distance of the CNs and the minimum length of the repetition VNs are both equal to 2, then for a randomly selected GLDPC code in the ensemble we expect a minimum distance growing as a linear or sublinear function of $N$ (for large $N$) depending on the sign of the first order coefficient in the growth rate Taylor series expansion. The results developed in [14] were further extended in [15] to GLDPC ensembles with an irregular CN set (CNs of different types). It was there proved that, provided that there exist CNs with minimum distance $2$, a parameter $\lambda^{\prime}(0)C$, generalizing the parameter $\lambda^{\prime}(0)\rho^{\prime}(1)$ of LDPC code ensembles, plays in the context of the weight distribution of GLDPC codes the same role played by $\lambda^{\prime}(0)\rho^{\prime}(1)$ in the context of the weight distribution of LDPC codes. The parameter $C$ is defined in Section III.

Interestingly, this latter results extends to binary GLDPC codes the same connection between the statistical properties of the weight distribution of irregular codes and the stability condition over the BEC. In fact, it was shown in [16] that the stability condition of binary irregular GLDPC codes over the BEC is given by

Generalized LDPC codes represent a promising solution for low-rate channel coding schemes, due to an overall rate loss introduced by the generalized CNs [17]. Doubly-generalized LDPC (D-GLDPC) codes generalize the concept of GLDPC codes while facilitating much greater design flexibility in terms of code rate [18] (an analogous idea may be found in the previous work [19]). In a D-GLDPC code, the VNs as well as the CNs may be of any generic linear block code types. A degree-$q$ VN may in principle be any $(q,k)$ linear block code, $q$ being the code length and $k$ the code dimension. Such a VN is associated with $k$ D-GLDPC code bits. It interprets these bits as its local information bits and interfaces to the CN set through its $q$ local code bits. A VN which corresponds to a linear block code which is not a repetition code is said to be a generalized VN. A D-GLDPC code is said to be regular if all of its VNs are of the same type and all of its CNs are of the same type and is said to be irregular otherwise. The structure of a D-GLDPC code is depicted in Fig. 1.

In this paper the growth rate of the weight distribution of binary irregular D-GLDPC codes is analyzed for small weight codewords. It is shown that, provided there exist both VNs and CNs with minimum distance $2$, a parameter $1/P^{-1}(1/C)$ discriminates between an asymptotically small and an asymptotically large expected number of small linear-weight codewords in a D-GLDPC code randomly drawn from a given irregular ensemble (the function $P(x)$ is defined in Section III). The parameter $1/P^{-1}(1/C)$ generalizes the above mentioned parameters $\lambda^{\prime}(0)\rho^{\prime}(1)$ and $\lambda^{\prime}(0)C$ to the case where both generalized VNs and generalized CNs are present. The obtained result also represents the extension to the D-GLDPC case of the previously recalled connection with the stability condition over the BEC. In fact, it was proved in [16] that the stability condition of D-GLDPC codes over the BEC is given by

The paper is organized as follows. Section II defines the D-GLDPC ensemble of interest, and introduces some definitions and notation pertaining to this ensemble. Section III defines further terms regarding the VNs and CNs which compose the D-GLDPC codes in the ensemble. Finally, Section IV states and proves the main result of the paper regarding the growth rate of the weight distribution.

We define a D-GLDPC code ensemble ${\mathcal{M}}_{n}$ as follows, where $n$ denotes the number of VNs. There are $n_{c}$ different CN types $t\in I_{c}=\{1,2,\cdots,n_{c}\}$, and $n_{v}$ different VN types $t\in I_{v}=\{1,2,\cdots,n_{v}\}$. For each CN type $t\in I_{c}$, we denote by $h_{t}$, $s_{t}$ and $r_{t}$ the CN dimension, length and minimum distance, respectively. For each VN type $t\in I_{v}$, we denote by $k_{t}$, $q_{t}$ and $p_{t}$ the VN dimension, length and minimum distance, respectively. For $t\in I_{c}$, $\rho_{t}$ denotes the fraction of edges connected to CNs of type $t$. Similarly, for $t\in I_{v}$, $\lambda_{t}$ denotes the fraction of edges connected to VNs of type $t$. Note that all of these variables are independent of $n$.

The polynomials $\rho(x)$ and $\lambda(x)$ are defined by

and

If $E$ denotes the number of edges in the Tanner graph, the number of CNs of type $t\in I_{c}$ is then given by $E\rho_{t}/s_{t}$, and the number of VNs of type $t\in I_{v}$ is then given by $E\lambda_{t}/q_{t}$. Denoting as usual $\int_{0}^{1}\rho(x)\,{\rm d}x$ and $\int_{0}^{1}\lambda(x)\,{\rm d}x$ by $\int\rho$ and $\int\lambda$ respectively, we see that the number of edges in the Tanner graph is given by

and the number of CNs is given by $m=E\int\rho$. Therefore, the fraction of CNs of type $t\in I_{c}$ is given by

and the fraction of VNs of type $t\in I_{v}$ is given by

Also the length of any D-GLDPC codeword in the ensemble is given by

Note that this is a linear function of $n$. Similarly, the total number of parity-check equations for any D-GLDPC code in the ensemble is given by

A member of the ensemble then corresponds to a permutation on the $E$ edges connecting CNs to VNs.

The growth rate of the weight distribution of the irregular D-GLDPC ensemble sequence $\{{\mathcal{M}}_{n}\}$ is defined by

where $\mathbb{E}_{{\mathcal{M}}_{n}}$ denotes the expectation operator over the ensemble ${\mathcal{M}}_{n}$, and $N_{w}$ denotes the number of codewords of weight $w$ of a randomly chosen D-GLDPC code in the ensemble ${\mathcal{M}}_{n}$. The limit in (7) assumes the inclusion of only those positive integers $n$ for which $\alpha n\in\mathbb{Z}$ and $\mathbb{E}_{{\mathcal{M}}_{n}}[N_{\alpha n}]$ is positive (i.e., where the expression whose limit we seek is well defined). Note that the argument of the growth rate function $G(\alpha)$ is equal to the ratio of D-GLDPC codeword length to the number of VNs; by (6), this captures the behaviour of codewords linear in the block length, as in [11] for the LDPC case.

Note that for any D-GLDPC code, there is a bijective correspondence between the set of D-GLDPC codewords and the set of split assignments which are both variable-valid and check-valid.

The weight enumerating polynomial for CN type $t\in I_{c}$ is given by

Here $A_{u}^{(t)}\geq 0$ denotes the number of weight-$u$ codewords for CNs of type $t$. Note that $A_{r_{t}}^{(t)}>0$ for all $t\in I_{c}$. The bivariate weight enumerating polynomial for VN type $t\in I_{v}$ is given by

Here $B_{u,v}^{(t)}\geq 0$ denotes the number of weight-$v$ codewords generated by input words of weight $u$, for VNs of type $t$. Also, for each $t\in I_{v}$, corresponding to the polynomial $B^{(t)}(x,y)$ we denote the sets

and

We denote the smallest minimum distance over all CN types by

and the set of CN types with this minimum distance by

We also define

for each $t\in I_{c}$, and

We note that $C_{t}>0$ for all $t\in I_{c}$, so $C>0$.

Similarly, we denote the smallest minimum distance over all VN types by

and the set of VN types with this minimum distance by

In the specific case where $p=2$, we also introduce the following definitions. For each $t\in X_{v}$, define the set $L_{t}=\{i\in\mathbb{Z}\;:\;B^{(t)}_{i,2}>0\}$ – note that these sets are nonempty. Also define the polynomial $P(x)$ by

and denote its inverse by $P^{-1}(x)$. Since all the coefficients of $P(x)$ are positive, $P(x)$ is monotonically increasing and therefore its inverse is well-defined and unique. Note that in the case $r=p=2$, both $C$ and the polynomial $P(x)$ depend only on the CNs and VNs with minimum distance equal to $2$.

Finally, note that throughout this paper, the notation $e=\exp(1)$ denotes Napier’s number.

The following theorem constitutes the main result of the paper.

The theorem is proved next. For ease of presentation, the proof is broken into four parts.