{"id":"aef94936-9b77-4bb3-8816-d8bc55360781","arxiv_id":"2506.15836","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A neural polar decoder estimates channel mutual information and optimizes the input distribution, producing polar codes with lower error rates than uniform i.i.d. inputs on channels with memory.","lead":"Researchers trained a neural decoder for polar error-correcting codes and used it to estimate how much information an unknown noisy channel can carry. The same decoder then shapes the transmitted bit distribution, reducing error rates compared with uniform inputs on channels with memory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (26) replaces Q in Theorem 5 with the per-sample loss difference L(x,0;θ*_co)-L(x,y;θ*_ch), but Theorem 4 only proves consistency of averaged MI, not of this per-sample plug-in; if biased, gradient ascent maximizes the NPD's proxy, not capacity.","rationale":"Reader's weakest assumption is the same as the one I find load-bearing: the per-sample plug-in gradient in Eq. (26). I agree. The paper is good-faith and has real support: Theorem 4 gives an averaged consistency statement inherited from [19], the AWGN experiment shows convergence to the known optimum, and the HY/SCL construction is plausible. But the central contribution—simultaneous MI maximization and code design—could fail if the gradient step maximizes a biased NPD proxy. The gap is not merely 'no convergence proof in general': it is that the specific per-sample estimator used in Eq. (26) is not shown to be consistent or even unbiased, and Algorithm 1's stage-averaged loss is not the log-likelihood of U^N that Q requires. The proposed test would settle whether the gap is practical by comparing plug-in gradients against exact ones on a channel where the true posterior is computable. Since the concern is unresolved but not disproven, the appropriate disposition remains conditional: the method is plausible and empirically promising, but the optimization step needs either a proof of per-sample consistency/unbiasedness or an experimental validation against exact gradients before the capacity-optimization claim is accepted.","tokens_in":18374,"tokens_out":8689,"duration_ms":109118,"concrete_test":"Run Algorithm 3 on a small-N Ising channel (e.g., N=8 or 16) where the exact posterior P_{U|Y} can be computed by enumerating all 2^N input sequences and channel states. At each improvement step, compute (a) the exact REINFORCE gradient of true MI using the known channel model, (b) the plug-in gradient from Eq. (26), and (c) the true MI of the updated LSTM input distribution. If the cosine similarity between (a) and (b) is poor, or if estimated MI rises while true MI does not, the load-bearing substitution in Eq. (26) is optimizing a biased proxy and the central claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 3's improvement step (Section IV-C) is load-bearing: it asserts that the plug-in L(x^N,0^N;θ*_co)-L(x^N,y^N;θ*_ch) can stand in for Q(x^N,y^N)=log(P_{U|Y}/P_U) in Theorem 5's exact gradient expression. This substitution is not a corollary of Theorem 4. Theorem 4 (via Theorem 3 from [19]) gives asymptotic consistency for averaged conditional MI estimates over M→∞; it says nothing about per-sample values of Algorithm 1's stage-averaged loss, nor about the gradient of that plug-in objective with respect to ψ. Algorithm 1 averages cross-entropy losses over all n+1 stages, not just the final-stage log-likelihood of U^N, so L(x,0;θ_co)-L(x,y;θ_ch) is not obviously the log-ratio of the sequential conditional distributions needed for the score-function gradient. If the plug-in is biased—e.g., because the NPD fits conditional entropies in expectation but not individual log-ratios—the alternating procedure can converge to a distribution that maximizes the NPD's estimate rather than the true channel capacity. The experiments are encouraging but do not close this gap: Fig. 1 tracks only the estimated MI on AWGN, and the final Ising/Trapdoor numbers sit below published lower bounds (0.5415<0.5451, 0.571<0.572), which may be finite-N effects but is also what a biased proxy would produce.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an alternating training procedure for black-box channels: a neural polar decoder (NPD) is used to estimate the mutual information between channel inputs and outputs (Eq. (23)), and a parametric LSTM input distribution is updated by gradient ascent where the log-density ratio in the exact score-function gradient is replaced by a per-sample difference of NPD losses (Eq. (26)). In the inference phase, the trained NPD is used to construct polar codes for asymmetric input distributions via the Honda-Yamamoto scheme with adaptive frozen sets and SCL decoding. Experiments on binary-input AWGN, Ising, and Trapdoor channels report MI estimates approaching published capacity bounds and BER/FER improvements over uniform i.i.d. inputs for block lengths up to 1024.","tokens_in":18825,"tokens_out":12427,"duration_ms":138436,"significance":"If the method works as claimed, it would be a useful step toward joint capacity estimation and code design for channels with memory, where model-based methods are often intractable. The paper's strengths include a public implementation, comparison against published capacity bounds and an SCT benchmark, an exact score-function gradient identity (Theorem 5), and a consistency result for the averaged MI estimator conditional on the assumptions of [19]. However, the theoretical results as stated do not cover the actual stage-averaged loss used in training, the coupled minimization of the two NPD embeddings, or the per-sample plug-in in the gradient update; these gaps leave the central optimization claim supported mainly by the experiments.","major_comments":[{"comment":"The estimator in Eq. (23) is computed from the stage-averaged loss L(x^N,y^N;θ) = (1/(n+1)) Σ_{j=0}^n L_j(v_j,e_j;θ), where each L_j is an average cross-entropy over the stage-j bits v_j. Theorem 3, however, concerns the final-stage conditional log-loss -log σ(l_θ(y^N,u^i)) for H(U_i|U^{i-1},Y^N) and H(U_i|U^{i-1}). Minimizing a sum of stage-wise cross-entropies is a different variational problem, and the proof of Theorem 4 applies Theorem 3 directly to a quantity that is not the one being minimized. The claim in Section III-D that the guarantees 'remain intact' needs a proof; as written, consistency of bI^M_Θ is not established.","section":"III-D/E, Eq. (14), Theorems 3-4"},{"comment":"Eq. (23) defines the MI estimate as the difference of two independent minimizations, min over θ_co and min over θ_ch, yet the NPD parameters F_θF, G_θG, H_θH are shared between θ_co and θ_ch, and Algorithm 3 updates a single coupled θ via the sum L(x,0;θ_co)+L(x,y;θ_ch). The two conditional-entropy terms cannot generally be minimized independently under this coupling, so the estimator actually implemented is not the estimator in Eq. (23). Theorem 4, which relies on separate consistency of the two conditional entropy estimators, therefore does not apply to the implemented algorithm without an additional argument.","section":"IV-B and Algorithm 3, Eq. (23)"},{"comment":"The gradient step replaces Q(x^N,y^N)=log(P_{U|Y}/P_U) in the exact score-function identity of Theorem 5 with the per-sample plug-in L(x,0;θ*_co)-L(x,y;θ*_ch). Theorem 4 only proves consistency of the averaged conditional MI estimate; it does not prove that this per-sample difference is a consistent estimate of the log-density ratio, nor that the expectation of the plug-in times ∇_ψ log P^ψ_X converges to the true gradient. Without such a result, Algorithm 3 may converge to a stationary point of the NPD's proxy objective rather than of the true MI. The empirical gaps in Fig. 2 (0.5415 vs the lower bound 0.5451 for Ising; 0.571 vs 0.572 for Trapdoor) are consistent with this concern, though finite-N effects are also possible.","section":"IV-C, Eq. (26)"}],"minor_comments":[{"comment":"The two terms in the cross-entropy loss are identical; the second term should presumably be (1-v0,i) log(1-σ(l0,i)).","section":"Eq. (9)"},{"comment":"The proof differentiates the expectation of Q without accounting for the ψ-dependence of Q; the missing term vanishes in expectation, but the proof should state this explicitly.","section":"Theorem 5 proof"},{"comment":"The claim that the fast loss computation reduces the number of computational steps from N log N to log N is unclear, since each stage computes O(N) node LLRs; please clarify what quantity is being counted.","section":"III-D"},{"comment":"The theoretical statements are for indecomposable FSCs, while the AWGN experiment in Fig. 1 is memoryless and not covered by these statements; the paper should state the scope of the theory explicitly.","section":"Theorems 3-4 and Fig. 1"},{"comment":"The adaptive frozen set threshold t is heuristic, and no finite-block-length guarantee is given for the resulting code; this should be stated as a limitation.","section":"V-B3, Eq. (33)"},{"comment":"The notation l_θ(y^N,u^i) should be defined explicitly with the sign convention used in Eq. (9), since the relationship between the LLR and the posterior probability is essential to the loss.","section":"Eq. (16)"}],"recommendation":"major_revision","confidential_remarks":"The paper is an extended version of an ISIT 2024 contribution, and the incremental contribution beyond [19] is primarily the joint MI maximization and SCL integration. The main risk is theoretical: the implemented objectives do not match the proven ones. I would encourage the editor to require the authors to either align Algorithm 1 with Theorem 3 (e.g., use the final-stage loss for the MI estimator) or prove consistency for the stage-averaged loss, and to provide a consistency or stationarity statement for the gradient update in Eq. (26), or an ablation on a known channel showing that the learned gradient matches the true gradient."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nThe paper does something genuinely new: it closes the loop between neural capacity estimation and practical polar coding for black-box channels. The pieces—NPDs, HY polar codes, neural MI optimization—exist in the literature, but the specific integration, including SCL decoding of NPDs and adaptive frozen sets, is new and non-trivially engineered. They also ship code, and the fast loss computation (stage-wise instead of recursive) is a real improvement over the earlier NPD paper. The experiments on Ising and Trapdoor channels are encouraging: the optimized input distribution clearly beats uniform i.i.d. inputs in MI and BER, and the MI estimates track published capacity bounds, which is a decent external check.\n\nThe soft spot is exactly where the stress-test note points. Equation (26) replaces the true log-ratio Q(x^N,y^N) with the per-sample loss difference L(x,0;θ*_co)-L(x,y;θ*_ch), but Theorem 4 only proves consistency of the averaged MI estimator over M→∞. It says nothing about per-sample log-ratios, and Algorithm 1 averages losses across all stages, so the plug-in is not obviously the sequential conditional log-ratio that Theorem 5's score-function gradient requires. If that plug-in is biased, the alternating algorithm maximizes the NPD's proxy, not the true MI. The observed MI values sitting slightly below the lower bounds (0.5415 vs 0.5451 for Ising, 0.571 vs 0.572 for Trapdoor) is the kind of small shortfall a biased proxy could produce, though finite-N effects are an equally plausible explanation. The paper doesn't acknowledge this gap; it states the substitution as a practical consequence of consistency, which it isn't, formally.\n\nThat said, the flaw is in the theoretical justification, not in the empirical observation. The central idea is plausible, and the authors are honest in showing the polarization loss caused by optimized inputs. The lack of error bars on the BER/FER plots makes the order-of-magnitude claim hard to judge, but the MI plots show consistent behavior across runs.\n\nThis paper deserves a serious referee. A reviewer should ask for either a proof that the plug-in gradient is asymptotically unbiased, or an experiment comparing the optimized distribution against a brute-force alternative to check for bias. I would cite it if I were working on neural polar codes or black-box capacity optimization. Bring it to reading group if you want a lively discussion about what \"consistency\" does and doesn't buy you.","headline":"Useful integration of NPDs with input-distribution optimization and SCL decoding, but the gradient step rests on an unproven plug-in approximation.","tokens_in":19262,"tokens_out":2859,"would_cite":true,"duration_ms":30442,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A17","94B35"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single alternating algorithm uses neural polar decoders to estimate channel capacity and to build practical polar codes for black-box channels.","keywords":["neural polar decoders","mutual information estimation","capacity estimation","input distribution optimization","polar codes","channels with memory","list decoding","HY scheme"],"falsifier":"On a finite-state channel whose capacity-achieving input distribution is known exactly (for example a simple Markov channel), run Algorithm 3 and compute the true MI of the resulting distribution by brute-force dynamic programming. If the NPD's MI estimate continues to rise while the true MI falls below the uniform-i.i.d. value, the per-sample loss-difference surrogate is biased and the alternating loop can converge away from capacity.","tokens_in":18193,"feed_emoji":"📡","tokens_out":6270,"duration_ms":64459,"temperature":0.7,"pith_summary":"Channel capacity is usually computed separately from code design, and most estimators give no practical code. This paper claims a single training loop closes that gap: neural polar decoders (NPDs) estimate the mutual information between inputs and outputs of an unknown channel with memory, and a gradient step then reshapes the input distribution to push that estimate upward. The loop stops with both an optimized input distribution and a decoder, and the decoder is used to construct a polar code through the HY scheme with list decoding. On the Ising and Trapdoor channels the estimated mutual information approaches published capacity bounds while the code achieves lower bit error rates than uniform i.i.d. inputs at block lengths up to 1024. A sympathetic reader takes the paper's contribution to be a practical bridge between capacity estimation and finite-blocklength coding for channels without a model.","feed_headline":"Neural polar decoders estimate capacity and build codes in one pass","feed_subtitle":"Alternating MI estimation with input shaping lifts rates and cuts bit errors versus uniform inputs on channels with memory.","key_machinery":"The machinery is the neural polar decoder (NPD), which replaces the check-node, bit-node, embedding, and LLR functions of successive cancellation decoding with neural networks, giving a decoder whose complexity $O(LkdN\\log N)$ does not depend on the channel's state size. Carrying the argument are two derived objects: the MI estimator of Eq. (23), the difference between the NPD's cross-entropy losses on real channel outputs and on constant zero outputs, whose consistency is stated in Theorem 4; and the score-function gradient of Theorem 5, implemented in Eq. (26) by using the per-sample loss difference as the plug-in estimate of $\\log(P_{U|Y}/P_U)$. Code construction then uses the HY scheme (a polar-coding construction for non-uniform input distributions) with an adaptive frozen set to accommodate the optimized input, plus list decoding.","core_discovery":"The paper's central claim is that Algorithm 3 simultaneously estimates the mutual information $I(X_\\psi^N;Y^N)$ and optimizes the input distribution $P_\\psi$ by alternating two steps. In the estimation step, two NPDs conditioned on the channel output and on a constant input respectively provide consistent estimates of conditional entropies, so their difference estimates the MI. In the improvement step, with the NPD fixed, the algorithm applies the score-function gradient from Theorem 5, replacing the intractable log-ratio $\\log(P_{U|Y}/P_U)$ with the plug-in difference of NPD losses, and updates the LSTM parameters $\\psi$. The paper argues that the optimized distribution's MI approaches the channel capacity lower bounds for the Ising and Trapdoor channels and that polar codes built from it, using the HY scheme and SCL decoding, attain lower bit error rates than uniform i.i.d. inputs for block lengths up to 1024.","pith_inferences":["The plug-in log-ratio gradient in Eq. (26) is a surrogate for the true density ratio; if it is biased, the loop maximizes the NPD's estimate rather than capacity, so finite-sample validation on a known channel is the decisive test.","The same loss-difference trick could be reused as a density-ratio estimator for other neural coding or shaping schemes, since it only requires two NPD passes per sample.","Reduced polarization under optimized inputs suggests that longer blocks or list sizes may be needed to close the gap to capacity, and that combining this method with polarization-rate-preserving shaping could improve finite-length performance.","Treating the channel as a black box makes the procedure a candidate for adaptive link configuration in real systems, where the channel changes and the input distribution must be re-optimized from streaming samples."],"forward_implications":["For an unknown channel with memory, the same training run yields both an estimate of achievable rate and a working polar code, removing the usual separation between capacity estimation and code design.","On channels whose capacity-achieving input is non-uniform, the optimized distribution gives higher estimated MI and lower BER than uniform i.i.d. inputs at block lengths up to 1024, so non-uniform shaping can be realized within the polar-code framework.","Because NPD decoding complexity is $O(LkdN\\log N)$ and independent of channel state size, list decoding can scale to channels where trellis-based decoders are infeasible.","The adaptive frozen set trades a small rate loss for mitigation of the reduced polarization caused by optimized inputs, making the scheme practical for finite block lengths."],"supporting_citations":[{"why":"It supplies the neural polar decoder architecture and the consistency result (Theorem 3) that the MI estimator builds on.","marker":"[19]"},{"why":"It defines the polar transform and successive cancellation decoding that the NPD replaces and generalizes.","marker":"[16]"},{"why":"It provides successive cancellation list decoding, which the paper integrates with the NPD and uses in all BER experiments.","marker":"[17]"},{"why":"It gives the HY polar coding scheme that accommodates the optimized non-uniform input distribution.","marker":"[20]"},{"why":"It supplies the lower and upper capacity bounds for Ising and Trapdoor channels against which the estimated MI is compared.","marker":"[21]"},{"why":"It defines the trellis-based benchmark decoder for channels with memory and identifies the reduced-polarization problem addressed by the adaptive frozen set.","marker":"[24]"},{"why":"It defines the LSTM model that parameterizes the input distribution and is optimized by the improvement step.","marker":"[27]"},{"why":"It supplies the reparameterization trick used to compute gradients of the input-distribution parameters.","marker":"[28]"}],"fun_headline_variants":["Neural polar decoders optimize code rates via MI estimation","Polar codes get rate lift from neural MI estimators","Two-step NPD training yields better codes for memory channels","Joint MI estimation and distribution shaping for polar codes","Neural decoders estimate MI and shape inputs in one loop"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The optimization step assumes the per-sample difference of the two NPD losses is an accurate estimate of the log-ratio of posterior to prior on the transformed input, so that gradient ascent on that surrogate increases the true mutual information; only the averaged MI estimator, not this per-sample plug-in, is proved consistent in the paper.","fun_headline_variants_meta":{"raw":{"variants":["Neural polar decoders optimize code rates via MI estimation","Polar codes get rate lift from neural MI estimators","Two-step NPD training yields better codes for memory channels","Joint MI estimation and distribution shaping for polar codes","Neural decoders estimate MI and shape inputs in one loop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000592,"raw_usage":{"total_tokens":2822,"prompt_tokens":1038,"completion_tokens":1784,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":1705}},"tokens_in":654,"tokens_out":1784,"duration_ms":15689,"temperature":1.0,"reasoning_tokens":1705,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:51:12.618799+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a finite-state channel whose capacity-achieving input distribution is known exactly (for example a simple Markov channel), run Algorithm 3 and compute the true MI of the resulting distribution by brute-force dynamic programming. If the NPD's MI estimate continues to rise while the true MI falls below the uniform-i.i.d. value, the per-sample loss-difference surrogate is biased and the alternating loop can converge away from capacity.","supporting_citations":[{"cited_title":"Data-Driven Neural Polar Codes for Unknown Channels With and Without Memory","cited_arxiv_id":"2309.03148","evidence_quote":"It supplies the neural polar decoder architecture and the consistency result (Theorem 3) that the MI estimator builds on."},{"cited_title":"List decoding of polar codes,","cited_arxiv_id":null,"evidence_quote":"It provides successive cancellation list decoding, which the paper integrates with the NPD and uses in all BER experiments."},{"cited_title":"Computable upper bounds on the capacity of finite-state channels,","cited_arxiv_id":null,"evidence_quote":"It supplies the lower and upper capacity bounds for Ising and Trapdoor channels against which the estimated MI is compared."},{"cited_title":"Construction of polar codes for channels with memory,","cited_arxiv_id":null,"evidence_quote":"It defines the trellis-based benchmark decoder for channels with memory and identifies the reduced-polarization problem addressed by the adaptive frozen set."},{"cited_title":"Long short-term memory,","cited_arxiv_id":null,"evidence_quote":"It defines the LSTM model that parameterizes the input distribution and is optimized by the improvement step."},{"cited_title":"Gradient estimation using stochastic computation graphs,","cited_arxiv_id":null,"evidence_quote":"It supplies the reparameterization trick used to compute gradients of the input-distribution parameters."}],"review_version":1}