Pith. sign in

REVIEW 2 major objections 4 minor 17 references

Relaxed protographs with Bernoulli edge probabilities can be optimized by gradient descent on density-evolution BER, and the resulting codes beat 5G LDPC baselines of the same size under min-sum decoding.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 06:42 UTC pith:NM6BRRBL

load-bearing objection Clean Bernoulli-ensemble DE for min-sum protographs; modest but real 5G gains under matching constraints, with scope limits the authors already flag. the 2 major comments →

arxiv 2607.08484 v1 pith:NM6BRRBL submitted 2026-07-09 cs.IT math.IT

Learning LDPC codes with quantized density evolution over relaxed protographs

Gennady Shutkov , Dmitry Artemasov , Alexey Frolov , Pavel Rybin , Kirill Andreev This is my paper
classification cs.IT math.IT MSC 94B0594B3568T05
keywords LDPC codesprotographsdensity evolutiongradient descentcode optimizationmin-sum decodingrelaxed ensembles
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Designing a good LDPC parity-check matrix for a concrete iterative decoder is still a hard combinatorial search. This paper replaces random flips and genetic search with a continuous, fully differentiable objective. It treats each entry of a small base matrix (the protograph) as a probability that an edge exists, defines a density-evolution recursion that is valid on those fractional entries, and proves that the recursion returns exactly the average performance over the ensemble of binary matrices those probabilities induce. Projected gradient descent on the predicted bit-error rate therefore moves many edge probabilities at once with a deterministic, low-variance signal. After a few non-integer entries remain, they can be enumerated and the best binary matrix lifted; the resulting codes outperform the corresponding 5G base graphs under normalized min-sum decoding while using the same dimensions, rate and puncturing budget.

Core claim

For the normalized min-sum decoder, a density-evolution recursion written on a matrix of Bernoulli edge probabilities produces, at every iteration and for every edge, the exact ensemble-averaged message distributions of the binary protograph ensemble defined by those probabilities; the resulting differentiable DE bit-error rate can therefore be minimized by ordinary projected gradient descent, yielding protographs that, after enumeration of residual fractional entries, outperform 5G LDPC codes of identical size.

What carries the argument

Relaxed density evolution (Theorem 1): each absent edge is replaced by a neutral message distribution so that the multilinear check-node and variable-node updates collapse the 2^{D-1}-term average over Bernoulli configurations into a single closed-form mixture; the continuous matrix is thereby both an optimization variable and a faithful ensemble representation.

Load-bearing premise

The collapse that equates continuous density evolution to the ensemble average works only for min-sum-style check nodes; the same algebraic trick fails for sum-product because grid interpolation breaks the needed linearity.

What would settle it

Optimize a protograph under the claimed relaxed DE loss, lift the best binary matrix, and measure whether its block-error rate under normalized min-sum decoding is strictly better than the matching 5G base graph of the same dimensions at BLER 10^{-4}; a null or reverse result falsifies the practical claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes a deterministic gradient-descent method for designing long protograph-based LDPC codes under a fixed iterative decoder. It represents a protograph by a relaxed matrix Ω whose entries are Bernoulli edge-retention probabilities, defines a quantized density-evolution (DE) recursion that operates directly on these continuous entries, and proves (Theorem 1) that the resulting DE bit-error rate equals the ensemble-averaged DE performance of the induced binary-protograph ensemble. Projected gradient descent is then applied to log of the DE BER; after training, the few remaining non-integer entries are enumerated and the best binary protograph is lifted (ACE) and simulated. For normalized min-sum decoding the optimized protographs of the same dimensions as 5G BG1 outperform the standardized codes by roughly 0.03 dB (rate 0.88) and 0.18 dB (rate 1/2) at BLER 10^{-4} under three decoder regimes.

Significance. If the claims hold, the work supplies a practical, low-variance alternative to Monte-Carlo decoder-in-the-loop and genetic search for large protograph spaces. The central technical contribution—Theorem 1 together with the multilinear collapse that lets continuous DE equal the ensemble average—is cleanly proved under standard tree-like assumptions and is specific to the min-sum family. The numerical gains, while modest, are obtained against a strong industrial baseline under identical rate, lifting factor and puncturing budget, and the training analysis shows that gradient steps explore the discrete space far more efficiently than random flips. The method is fully autonomous once the DE parameters are fixed and therefore constitutes a useful addition to the LDPC design toolbox within its stated decoder scope.

major comments (2)
  1. Section IV-A and the high-rate experiment: the raw optimization produces a weight-one row/column degeneracy that is removed manually before reporting results. Because this step is outside the autonomous GD procedure claimed in the abstract and Section III-D, the paper should either (i) incorporate an explicit rate-preserving regularizer that prevents the degeneracy or (ii) document the frequency of such degeneracies and the precise reduction rule so that the pipeline remains reproducible.
  2. Section IV-A, BLER evaluation protocol: block-error rate is measured over information bits for the 5G references but over the whole codeword for the optimized codes. While the authors note that this follows the protograph-level objective, the asymmetry can inflate the reported gains; a side-by-side comparison under a common information-bit BLER definition (or an explicit statement that the difference is negligible) is needed to keep the 0.03 dB / 0.18 dB claims fully comparable.
minor comments (4)
  1. Remark 1 and Section V correctly flag that the multilinear collapse fails for sum-product because grid interpolation breaks linearity; a short sentence in the abstract clarifying that all reported gains are for normalized min-sum would prevent over-generalization.
  2. Table II lists a fixed learning rate 5e-3 and a hard [0,1] projection; a brief ablation or sensitivity remark would strengthen the claim of reliable convergence.
  3. Figures 5–6 show the final non-integer entries; adding the corresponding DE-BER values of the enumerated binary members would make the “small ensemble” claim fully quantitative.
  4. Notation: the same symbol Ω is used for both the relaxed matrix and, occasionally, for local neighborhoods; a local-row/column notation (already introduced later) could be used consistently from Definition 1.

Circularity Check

0 steps flagged

No significant circularity: relaxed DE equals ensemble-averaged DE by multilinear construction, and reported gains are external finite-length BLER vs 5G baselines.

full rationale

The paper's central derivation (Definition 1, Eqs. 6–15, Theorem 1) shows that the continuous DE recursion on Bernoulli edge probabilities Ω produces the marginal message PMFs averaged over the induced binary-protograph ensemble. This is an algebraic identity under the stated multilinear/min-sum assumptions (neutral elements p∞/pδ, affine mixtures, total-probability collapse), not a fit of a free parameter later re-labeled as a prediction. The optimization objective is log of that DE BER; after projected GD the few remaining non-integer entries are enumerated and the resulting binary protographs are ACE-lifted and evaluated by independent Monte-Carlo BLER against the external 5G BG1 standards of identical dimensions (Figs. 3–4, Table I). No uniqueness theorem is imported from the authors, no ansatz is smuggled via self-citation, and the only self-references are to standard DE literature or the authors' prior quantized-DE implementation details that are fully restated in Appendix A. The min-sum specificity (Remark 1) is an explicit scope limit, not a circular reduction. Consequently the derivation chain is self-contained and the empirical claim is externally falsifiable; circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central claim rests on standard density-evolution independence assumptions, the multilinear property of the min-sum check-node update on a quantized grid, and a handful of hand-chosen numerical hyperparameters. No new physical entities are postulated; the only invented mathematical object is the relaxed Bernoulli protograph ensemble itself, which is given a clear ensemble interpretation and is falsifiable by comparing continuous DE to explicit enumeration of the binary members.

free parameters (4)
  • NMS scaling coefficient α = 0.75
    Fixed at 0.75 for both DE training and finite-length simulation; chosen by convention rather than optimized.
  • LLR grid parameters (Lc, N) = Lc=50, N=1000
    Clipping level Lc=50 and 2001-point grid (N=1000) are set by hand; they control quantization error of the DE.
  • Learning rate νt = 5e-3
    Fixed projected-GD step size 5×10^{-3}; not adapted or cross-validated.
  • DE iteration count and training SNR schedule = 60 iters; threshold-triggered SNR decrease
    60 flooding iterations and the rule that SNR is lowered when DE BER falls below a threshold are design choices that affect the loss landscape.
axioms (4)
  • domain assumption Incoming messages at a protograph node are independent (tree-like neighborhood of the lifted Tanner graph).
    Standard DE assumption stated in §III-B; required for the total-probability expansions (9) and (23).
  • standard math The min-sum check-node update is multilinear in its input PMFs when absent edges are replaced by the neutral distribution p∞.
    Used to collapse the 2^{D-1}-term average into a single evaluation of fc on mixed inputs (Eqs. 8, 12); holds for the sign-split tail product on a regular grid.
  • domain assumption Zero-codeword assumption and AWGN-BPSK channel LLR distribution for initialization of transmitted variable nodes.
    Standard for DE of binary linear codes; stated in §III-B.
  • ad hoc to paper After training, the few remaining non-integer entries can be exhaustively enumerated and the best binary protograph selected by the same DE metric.
    Empirical observation reported in §IV; not proved that the continuous optimum always crystallizes to a small ensemble.
invented entities (1)
  • Relaxed Bernoulli-edge protograph ensemble H(Z)(Ω) no independent evidence
    purpose: Provides a continuous, differentiable representation of an ensemble of binary protographs so that DE BER can be evaluated and differentiated without Monte-Carlo sampling.
    Definition 1; each entry ωij is the retention probability of the corresponding permutation-block edge. Independent evidence is partial: the continuous DE matches the average of the enumerated binary members (Remark 2–3), but the entity is introduced for this paper’s optimization.

pith-pipeline@v1.1.0-grok45 · 18447 in / 3425 out tokens · 29337 ms · 2026-07-10T06:42:48.118304+00:00 · methodology

0 comments
read the original abstract

We consider the design of low-density parity-check (LDPC) codes for a given iterative decoder. Despite tools such as direct simulation, density evolution (DE), and EXIT-chart analysis, selecting a parity-check matrix remains a difficult combinatorial optimization problem. Existing approaches often rely on population-based search, random mutations, genetic algorithms, or related heuristics, which require careful parameter tuning and may be computationally expensive. Recent gradient descent (GD)-based methods optimize relaxed parity-check matrices by differentiating through decoder simulations. However, such decoder-in-the-loop strategies rely on noisy Monte Carlo estimates, require line search over soft matrix representations, and remain costly for long LDPC codes. Moreover, although optimization is performed in a relaxed domain, the loss is typically evaluated only at integer-valued parity-check matrices. In this work, we focus on the design of long protograph-based LDPC codes and propose a deterministic GD-based framework that operates directly on a relaxed protograph representation. Each protograph entry is interpreted as the probability that the corresponding element is equal to one. The loss function is based on DE bit error rate (BER) performance and can be evaluated directly for relaxed protographs. To justify this relaxation, we associate the relaxed representation with an ensemble of binary protographs and show that the proposed relaxed DE gives the ensemble-averaged DE performance. The resulting optimization procedure is fully autonomous and uses standard GD methods. Owing to deterministic DE evaluation and informative gradients, the proposed approach provides fast and reliable convergence. Numerical experiments for the min-sum decoder show that the optimized protographs outperform 5G LDPC codes with the same protograph dimensions.

Figures

Figures reproduced from arXiv: 2607.08484 by Alexey Frolov, Dmitry Artemasov, Gennady Shutkov, Kirill Andreev, Pavel Rybin.

Figure 1
Figure 1. Figure 1: 5G base graph 1 for the longest codes with rates [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Parity-check matrix and corresponding Tanner graph. Squares denote [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Block error rate performance for the BG1-medium case with rate [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Block error rate performance for the BG1-high case with rate [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Optimized relaxed protograph for the BG1-medium case with rate [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Optimized relaxed protograph for the BG1-high case with rate [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · 1 internal anchor

  1. [1]

    Low-density parity-check codes,

    R. G. Gallager, “Low-density parity-check codes,”IRE Transactions on Information Theory, vol. 8, no. 1, pp. 21–28, 1962

    work page 1962

  2. [2]

    A recursive approach to low complexity codes,

    R. M. Tanner, “A recursive approach to low complexity codes,”IEEE Transactions on Information Theory, vol. 27, no. 5, pp. 533–547, 1981

    work page 1981

  3. [3]

    The capacity of low-density parity- check codes under message-passing decoding,

    T. J. Richardson and R. L. Urbanke, “The capacity of low-density parity- check codes under message-passing decoding,”IEEE Transactions on Information Theory, vol. 47, no. 2, pp. 599–618, 2001

    work page 2001

  4. [4]

    Design of capacity-approaching irregular low-density parity-check codes,

    T. J. Richardson, M. A. Shokrollahi, and R. L. Urbanke, “Design of capacity-approaching irregular low-density parity-check codes,”IEEE Transactions on Information Theory, vol. 47, no. 2, pp. 619–637, 2001

    work page 2001

  5. [5]

    Convergence behavior of iteratively decoded parallel concatenated codes,

    S. ten Brink, “Convergence behavior of iteratively decoded parallel concatenated codes,”IEEE Transactions on Communications, vol. 49, no. 10, pp. 1727–1737, 2001

    work page 2001

  6. [6]

    Protograph LDPC codes design based on EXIT analysis,

    G. Liva and M. Chiani, “Protograph LDPC codes design based on EXIT analysis,” inProc. IEEE Global Telecommunications Conference, 2007, pp. 3250–3254

    work page 2007

  7. [7]

    Regular and irregular pro- gressive edge-growth Tanner graphs,

    X.-Y . Hu, E. Eleftheriou, and D.-M. Arnold, “Regular and irregular pro- gressive edge-growth Tanner graphs,”IEEE Transactions on Information Theory, vol. 51, no. 1, pp. 386–398, 2005

    work page 2005

  8. [8]

    Selective avoid- ance of cycles in irregular LDPC code construction,

    T. Tian, C. R. Jones, J. D. Villasenor, and R. D. Wesel, “Selective avoid- ance of cycles in irregular LDPC code construction,”IEEE Transactions on Communications, vol. 52, no. 8, pp. 1242–1247, 2004

    work page 2004

  9. [9]

    On the design of low-density parity-check codes within 0.0045 db of the Shannon limit,

    S.-Y . Chung, G. D. Forney, T. J. Richardson, and R. Urbanke, “On the design of low-density parity-check codes within 0.0045 db of the Shannon limit,”IEEE Communications Letters, vol. 5, no. 2, pp. 58–60, 2001

    work page 2001

  10. [10]

    Low-density parity-check (LDPC) codes constructed from protographs,

    J. Thorpe, “Low-density parity-check (LDPC) codes constructed from protographs,” Jet Propulsion Laboratory, INP Progress Report, Tech. Rep. 42-154, 2003

    work page 2003

  11. [11]

    Protograph LDPC codes with node degrees at least 3,

    D. Divsalar, S. Dolinar, C. Jones, and K. Andrews, “Protograph LDPC codes with node degrees at least 3,” inProc. IEEE Global Telecommu- nications Conference, 2006, pp. 1–6

    work page 2006

  12. [12]

    Decoder-in-the-loop: Genetic optimization-based LDPC code design,

    A. Elkelesh, M. Ebada, S. Cammerer, L. Schmalen, and S. ten Brink, “Decoder-in-the-loop: Genetic optimization-based LDPC code design,” IEEE Access, vol. 7, pp. 141 161–141 170, 2019

    work page 2019

  13. [13]

    Factor Graph Optimization of Error-Correcting Codes for Belief Propagation Decoding

    Y . Choukroun and L. Wolf, “Factor graph optimization of error- correcting codes for belief propagation decoding,”arXiv preprint arXiv:2406.12900, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  14. [14]

    Learning linear block codes with gradient quantization,

    L.-A. Dufrene, Q. Lampin, and G. Larue, “Learning linear block codes with gradient quantization,”IEEE Transactions on Communi- cations, 2025, doi:10.1109/TCOMM.2025.3615681; also available as arXiv:2503.16169

    work page doi:10.1109/tcomm.2025.3615681 2025

  15. [15]

    NR; multiplexing and channel coding,

    3GPP, “NR; multiplexing and channel coding,” 3rd Generation Partner- ship Project, Tech. Rep. TS 38.212, 2024

    work page 2024

  16. [16]

    Learning to decode linear codes using deep learning,

    E. Nachmani, Y . Be’ery, and D. Burshtein, “Learning to decode linear codes using deep learning,” inProc. 54th Annual Allerton Conference on Communication, Control, and Computing, 2016, pp. 341–346

    work page 2016

  17. [17]

    Gener- alized ldpc codes with low-complexity decoding and fast convergence,

    D. Simegn, D. Artemasov, K. Andreev, P. Rybin, and A. Frolov, “Gener- alized ldpc codes with low-complexity decoding and fast convergence,” IEEE Wireless Communications Letters, vol. 14, no. 11, pp. 3700–3704, 2025

    work page 2025