{"id":"3948dae5-b617-4115-9954-4ee2b88101db","arxiv_id":"2506.16098","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Importance-sampling weighted loss functions enable autoencoder-based probabilistic constellation shaping with exact automatic-differentiation gradients, matching prior methods in AWGN and IM/DD simulations.","lead":"This paper presents a new loss function that lets machine-learning autoencoders optimize the probabilities of constellation points in digital communication without manually correcting gradients. The method uses importance sampling and automatic differentiation, and its simulations match existing techniques on two channel models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exact-gradient proof hinges on equating the fixed proposal q_m with the batch's empirical frequency |B_m|/N; under the algorithm's actual per-epoch q_m, the gradient of (2) is biased by ~(1-q_m)/(2N q_m), so the central claim is not established for finite batches.","rationale":"The central claim of the paper is that the proposed loss functions (2) and (5), combined with the per-epoch sampler QM, provide exact gradients of MI/BMI with respect to constellation point probabilities using only automatic differentiation. The proof of this claim is Eq. (4), which is derived from Eq. (3) under the substitution q_m = |B_m|/N. That substitution is exactly the weak point: in Algorithm 1, q_m is the variational-distance approximation computed once per epoch, not the empirical frequency of the current batch. Once this distinction is made, the actual gradient of (2) with respect to p_m contains a bias term of order (1-q_m)/(2N q_m) relative to the gradient of (1). For rare constellation points, where the variational constraint q_m N >= 1 is active, this bias is order one rather than order 1/N, so it is not negligible at the stated batch size of 2^15. The reader's weakest assumption identified the same conflation, and the finite-sample bias of the estimator is precisely what makes the analytical claim unproven. The concern is load-bearing because it targets the paper's theoretical novelty: the method's practical success in matching [1] is plausible, but the claim of exact gradients is what distinguishes it from a heuristic importance-sampling scheme. The concrete test isolates the issue by computing the empirical gradient and comparing it with the exact gradient of the MI bound; if the test shows small bias at N = 2^15, the concern would be resolved, and if it shows large bias, the paper would need to either revise the proof or restrict the claim to asymptotics. No change to the verdict is needed because the reader's CONDITIONAL decision already requires addressing this bias, and the test would settle whether the condition is actually violated.","tokens_in":6366,"tokens_out":18563,"duration_ms":205875,"concrete_test":"For a 4-PAM AWGN channel at Es/N0 = 5 dB, choose a strongly shaped P with at least one probability below 1/N, set N = 2^15, and run Alg. 1 exactly as specified. Across 10^4 independent batches, average the autodiff gradient of (2) with respect to p_m, using the epoch-level q_m and the batch's K_m, and compare elementwise with the exact gradient of (1) computed by Gauss-Hermite quadrature. Report the maximum absolute difference in bits/symbol; if it exceeds 0.01, the exactness claim fails. Repeat with N = 100 to test the stated short-sequence advantage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. II, after defining QM as the variational-distance approximation fixed for an epoch (Alg. 1), Eq. (2) is written with the empirical frequency |B_m|/N inside the first logarithm, and the text then states 'Then q_m = |B_m|/N'. This identification is used to convert Eq. (3) into Eq. (4), which is claimed to be the exact gradient of the MI bound (1). However, under the algorithm's actual definition w_m = p_m/q_m with q_m the epoch-level proposal, differentiating the first two terms of (2) with respect to p_m yields -(K_m/(N q_m))[log(K_m/N) + log(p_m/q_m) + log2(e)] plus channel-dependent terms, not -log2(p_m) - log2(e). Taking expectations, the difference from the gradient of (1) is (1-q_m)/(2N q_m) in nats due to the log-concavity bias of E[(K/N) log(K/N)]. For any constellation point whose q_m is at the variational lower bound q_m N = 1, this bias is O(1), not O(1/N). Thus Eq. (4) is not the gradient of the MI lower bound for the batch sizes used (2^15), and the claim that the loss (2) supplies exact gradients of MI/BMI with respect to the constellation probabilities is unproven. The method may still be consistent as N tends to infinity, but the analytical argument does not establish exactness, and the simulation plots lack error bars that would reveal the bias.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an autoencoder-based method for probabilistic constellation shaping in which the mutual information (MI) or bit-wise mutual information (BMI) is optimized with loss functions that use importance-sampling weights w_m = p_m/q_m and a proposal distribution Q_M refreshed once per epoch. The authors claim that, unlike prior work [1], their loss functions allow the constellation-point probabilities to be optimized using automatic differentiation alone, and that the resulting gradients are exact. The method is validated on a 256-QAM AWGN channel and an 8-PAM intensity-modulation/direct-detection channel, with results reported to match the gradient-addition baseline [1] within 0.01 bits/symbol. The paper also states that code and results are publicly available.","tokens_in":6728,"tokens_out":8432,"duration_ms":104447,"significance":"If the exact-gradient claim were correct, the paper would make a useful practical contribution: it would remove the error-prone manual gradient corrections of [1], allow shorter symbol sequences through importance sampling, and apply to arbitrary channel models where the demapper is learned. The paper is clearly written, the simulations cover two relevant channel models, and the release of code is a strength. However, the central theoretical claim is not established by the derivation as written, because the proof identifies the epoch-level proposal q_m with the random empirical frequency |B_m|/N of a finite batch. That identification is load-bearing for Eq. (4) and for the abstract's claim of exact gradients. The empirical agreement with baselines is encouraging but is reported without error bars, so it does not by itself resolve the finite-sample bias concern.","major_comments":[{"comment":"The derivation of Eq. (4) from Eq. (3) uses the identification q_m = |B_m|/N, but in Algorithm 1 q_m is the variational-distance approximation fixed for the whole epoch, whereas |B_m|/N is the random empirical frequency of a single batch. These two quantities are not equal. With q_m fixed and w_m = p_m/q_m, the expected value of the first term of Eq. (2) is -(1/N) \\sum_m w_m E[K_m log(K_m/N)] with K_m = |B_m|, which is not equal to -\\sum_m p_m log q_m. A Taylor expansion of the multinomial expectation gives E[K_m log(K_m/N)] = N q_m log q_m + (1-q_m)/2 + O(1/N), so the loss itself is a biased estimate of the bound (1). The gradient with respect to p_m divides by q_m, producing a bias of order (1-q_m)/(2N q_m); for a constellation point at the variational lower bound q_m = 1/N, this bias is O(1), not O(1/N). Therefore Eq. (4) is the gradient of the realized batch loss, not the exact gradient of the MI bound, and the abstract's claim that the method 'uses exact gradients' is unsupported. The authors should either prove unbiasedness of the gradient estimator under the actual sampling scheme or explicitly characterize the bias and reframe the claim as a consistent large-N approximation.","section":"Section III, Figs. 2-4"},{"comment":"The numerical comparisons report that differences between the proposed method and the baselines remain below 0.01 bits/symbol, but no confidence intervals, number of independent runs, or error bars are provided. Since the theoretical exactness of the gradients is the main claim of the paper, the simulations should include a direct finite-batch bias check, for example by comparing the batch gradient used in training with a high-accuracy estimate of the gradient of (1) at the same q_m, or by reporting the variability of the optimized probabilities across random restarts. Without such a check, the close agreement with [1] is suggestive but does not validate the exact-gradient claim at the batch size used in the paper.","section":"Section III, Figs. 2-4"}],"minor_comments":[{"comment":"Equation (4) contains the factor 1/|B_m|, which is undefined when a constellation point is not sampled in a batch. The corresponding automatic-differentiation derivative is zero because the sum over the empty set is empty. Please clarify how empty batches are handled and whether the displayed formula is meant to apply only when |B_m| > 0.","section":"Section II, Eq. (4)"},{"comment":"The algorithm updates W_M in the inner loop and then sets P_M = W_M Q_M at the end of the epoch, but the text does not state whether P_M is renormalized to sum to one. Since P_M is used for constellation normalization and as the target distribution, please clarify the normalization step explicitly.","section":"Algorithm 1"},{"comment":"The sentence 'Then q_m = |B_m|/N' is the crux of the derivation and is currently ambiguous. It should be replaced with an explicit statement of whether q_m denotes the epoch-level proposal or the empirical frequency, because the two choices lead to different and incompatible gradient derivations.","section":"Section II, paragraph after Eq. (2)"},{"comment":"The text mentions that the peak-power AWGN study agrees with Blahut-Arimoto but omits the results due to space constraints. Since the code is available, consider moving this result to an appendix or providing a brief quantitative summary in the text, as it is a strong validation of the method.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the exact-gradient claim in the abstract and Section II. The stress-test concern about the identification q_m = |B_m|/N lands: the derivation is internally inconsistent with Algorithm 1, and the finite-batch bias is O(1) for low-probability symbols at the variational lower bound. The paper can likely be repaired by reframing the contribution as a consistent stochastic-gradient method with a quantified bias, and by adding a bias analysis or experiments. The self-citation to [1] is appropriate because that work is used as a baseline, not as the source of the main claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this paper offers a genuinely simpler way to train autoencoder-based constellation shaping — replace the manual gradient correction of Aref–Chagnon with importance-weighted losses and refresh the proposal each epoch — and the simulations show it matches the baselines on AWGN and a simplified IM/DD channel. The code is public. That part is real and useful.\n\nThe soft spot is in the proof. The paper claims Eq. (4) is the exact gradient of the MI lower bound with respect to the symbol probabilities. The derivation hinges on the line \"Then q_m = |B_m|/N\" in Sec. II. That equates the per-epoch proposal distribution, which is fixed before sampling, with the batch's empirical frequency. These are different objects. Under the algorithm's actual definition w_m = p_m/q_m, differentiating (2) gives an extra bias term of order (1-q_m)/(2N q_m). For constellation points at the variational lower bound q_m N = 1 that bias is O(1), not O(1/N). So the exact-gradient claim is not established for finite batch sizes. The method may still be consistent as N→∞, and the experiments suggest it works, but the paper should say so honestly.\n\nTwo more minor issues: the simulation curves have no error bars, so we cannot tell whether the sub-0.01 bit/symbol differences are noise; and the advertised advantage for short symbol sequences is not experimentally tested — the paper only mentions the peak-power AWGN agreement, and those results are omitted.\n\nMy overall impression: the idea is worth taking seriously, and the empirical match to [1] suggests the biased gradient does no harm in these cases. But the central theoretical claim needs repair: either present (2) with the true proposal q_m in the first logarithm, or provide a finite-sample bias analysis. As written, the paper overstates its analytic guarantee.\n\nI'd send it to review — a good referee would catch the same issue and the authors can fix it. I'd bring it up in a reading group if you want a discussion about importance sampling in autoencoder training. I would not cite the exact-gradient claim in my own work until the bias is addressed.","headline":"Useful importance-sampling loss for autoencoder-based constellation shaping, but the exact-gradient proof conflates the proposal distribution with the empirical batch frequency; the claim needs a fix.","tokens_in":7209,"tokens_out":5933,"would_cite":false,"duration_ms":61337,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes weighted importance-sampling loss functions that let an autoencoder learn constellation point probabilities for probabilistic shaping, with automatic differentiation alone producing exact gradients of mutual information…","keywords":["probabilistic constellation shaping","importance sampling","automatic differentiation","mutual information","bit-wise mutual information","autoencoder","AWGN channel","IM/DD channel"],"falsifier":"Run Algorithm 1 on a strongly shaped 256-QAM distribution with a small batch size (for example $N = 2^8$) so that the rarest points violate $q_m N \\ge 1$, and compare the automatic-differentiation update of $w_m$ against the exact gradient of the finite-sample objective, or against the gradient-addition method's update. If the two move in different directions or diverge, the claimed exactness holds only in the regime where $|B_m|/N$ is a faithful surrogate for $q_m$.","tokens_in":6202,"feed_emoji":"📡","tokens_out":10788,"duration_ms":116146,"temperature":0.7,"pith_summary":"Probabilistic constellation shaping adjusts how often each constellation point is sent, and the previous autoencoder approach required manually adding correction terms to the gradients computed by automatic differentiation. This paper removes that step by writing the objective as an importance-sampling estimate: symbols are drawn from an auxiliary distribution $Q_M$ that approximates the target probabilities $P_M$, and every sample is re-weighted by $w_m = p_m/q_m$. The authors prove that gradient ascent on the two proposed losses, one for mutual information and one for bit-wise mutual information, yields exactly the same gradient in the constellation probabilities as the earlier gradient-correction method. They demonstrate on a 256-QAM AWGN channel and an 8-PAM intensity-modulation/direct-detection channel that the learned shaping gains match the baseline within numerical accuracy. If the claim holds, engineers can shape constellations for new channels with standard machine-learning tooling and no hand-derived gradient terms.","feed_headline":"Automatic gradients for probabilistic constellation shaping","feed_subtitle":"New losses with importance sampling match gradient-addition baselines on AWGN and IM/DD channels.","key_machinery":"The load-bearing object is the weighted importance-sampling loss, in which a batch of symbols is drawn from a sampling distribution $Q_M$ (a variational-distance approximation of the target $P_M$), each symbol's contribution is weighted by $w_m = p_m/q_m$, and the constellation normalization is evaluated with the induced distribution $\\tilde P_M = W_M \\cdot Q_M$. The identity that carries the exactness proof is $\\partial f/\\partial w_m = q_m\\,\\partial f/\\partial p_m$ combined with the fact that the statistics of the drawn batches do not depend on $w_m$, so automatic differentiation through the loss directly returns the true gradient in the constellation probabilities. The outer loop of Algorithm 1 refreshes $Q_M$ once per epoch from the current $P_M$, keeping the two distributions close and ensuring every point has $q_m N \\ge 1$; this is what lets the method use relatively short batches without rare-symbol dropout.","core_discovery":"The paper's central claim is that Eqs. (2) and (5) are valid autoencoder loss functions for learning constellation point probabilities, and that the gradients of these losses with respect to the weights $w_m = p_m/q_m$, computed by automatic differentiation, coincide with the exact gradients of the mutual information and bit-wise mutual information with respect to $p_m$. The analytical step is Eq. (4), which follows from $\\partial f/\\partial w_m = q_m\\,\\partial f/\\partial p_m$ and reproduces the gradient expression of the previous method. The paper further claims that this removes the need for the manual correction terms in the earlier approach, that the per-epoch refresh of the sampler $Q_M$ via variational-distance approximation keeps $q_m N \\ge 1$ and thereby avoids dropout of rare constellation points, and that simulations on AWGN and IM/DD channels match the earlier method to within about 0.01 bits/symbol.","pith_inferences":["The exact-gradient claim is best read as asymptotic in batch size: the proof identifies $|B_m|/N$ with $q_m$, so at small batches the estimated loss is biased and the guarantee should soften gradually rather than fail at a sharp threshold.","A natural extension the paper does not test is joint geometric and probabilistic shaping: with the constellation coordinates also trainable, the automatic-differentiation path through normalization should expose gradients with respect to coordinates as well, and the same importance-sampling loss could learn both.","On channels where noise depends on the transmitted symbol sequence, such as nonlinear optical fibers, the per-epoch refresh of $Q_M$ should beat fixed-distribution sampling; the paper motivates this scenario but leaves the experiment for future work."],"forward_implications":["The same optimization trajectory as the gradient-addition method is available from a simpler implementation, because the loss itself, rather than extra manual terms, carries the exact gradient.","Engineers can optimize for either symbol-metric decoding (MI) or bit-metric decoding (BMI) with the same weight-learning scheme, switching only the receiver output in the loss.","Training with relatively short symbol sequences remains viable for strongly shaped constellations, since the per-epoch sampler refresh keeps low-probability points represented ($q_m N \\ge 1$).","The reported MI gap matches the baseline within about 0.01 bits/symbol on 256-QAM AWGN, and the BMI gains match on 8-PAM IM/DD, so the method's practical behavior tracks the theoretically corrected baseline.","Since the demapper can be a neural network and the channel model is just a pass-through, the loss functions transfer to other channels without re-deriving gradients."],"supporting_citations":[{"why":"Baseline gradient-addition method; the paper shows that its Eq. (4) reproduces this method's gradient expression.","marker":"[1]"},{"why":"Related joint shaping algorithm that samples from a fixed distribution; the paper contrasts its per-epoch refreshed sampler against it.","marker":"[11]"},{"why":"Source for variational-distance distribution matching and the $q_m N \\ge 1$ criterion used to construct $Q_M$ each epoch.","marker":"[14]"},{"why":"Gives the achievable-information-rate lower bound (their Eq. 34) that the MI loss in Eq. (2) estimates via importance sampling.","marker":"[15]"},{"why":"Supplies the no-learning-rate optimizer used in the gradient-ascent simulations.","marker":"[18]"},{"why":"Provides the Gauss-Hermite quadrature used to validate the optimized MI and BMI values.","marker":"[19]"}],"fun_headline_variants":["Importance sampling makes constellation shaping gradients automatic","Auto-diff gradients for constellation shaping via importance sampling","Exact gradients for shaping without manual correction terms","Shaping gradients from importance sampling, no manual steps","Automatic gradients for probabilistic shaping via importance sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exact-gradient claim assumes that within each batch the empirical frequency $|B_m|/N$ of a constellation point is a faithful stand-in for its true sampling probability $q_m$, so the whole argument depends on batches being large enough and the variational-distance sampler giving every point $q_m N \\ge 1$.","fun_headline_variants_meta":{"raw":{"variants":["Importance sampling makes constellation shaping gradients automatic","Auto-diff gradients for constellation shaping via importance sampling","Exact gradients for shaping without manual correction terms","Shaping gradients from importance sampling, no manual steps","Automatic gradients for probabilistic shaping via importance sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000593,"raw_usage":{"total_tokens":2743,"prompt_tokens":871,"completion_tokens":1872,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":1800}},"tokens_in":487,"tokens_out":1872,"duration_ms":13325,"temperature":1.0,"reasoning_tokens":1800,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:45:32.623100+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a strongly shaped 256-QAM distribution with a small batch size (for example $N = 2^8$) so that the rarest points violate $q_m N \\ge 1$, and compare the automatic-differentiation update of $w_m$ against the exact gradient of the finite-sample objective, or against the gradient-addition method's update. If the two move in different directions or diverge, the claimed exactness holds only in the regime where $|B_m|/N$ is a faithful surrogate for $q_m$.","supporting_citations":[{"cited_title":"End-to-end learning of joint geometric and probabilistic constellation shaping,","cited_arxiv_id":null,"evidence_quote":"Baseline gradient-addition method; the paper shows that its Eq. (4) reproduces this method's gradient expression."},{"cited_title":"Joint learning of probabilistic and geo- metric shaping for coded modulation systems,","cited_arxiv_id":null,"evidence_quote":"Related joint shaping algorithm that samples from a fixed distribution; the paper contrasts its per-epoch refreshed sampler against it."},{"cited_title":"Probabilistic amplitude shaping,","cited_arxiv_id":null,"evidence_quote":"Source for variational-distance distribution matching and the $q_m N \\ge 1$ criterion used to construct $Q_M$ each epoch."},{"cited_title":"Simulation-based computation of information rates for channels with memory,","cited_arxiv_id":null,"evidence_quote":"Gives the achievable-information-rate lower bound (their Eq. 34) that the MI loss in Eq. (2) estimates via importance sampling."},{"cited_title":"Training deep networks without learning rates through coin betting,","cited_arxiv_id":null,"evidence_quote":"Supplies the no-learning-rate optimizer used in the gradient-ascent simulations."},{"cited_title":"A note on multivariate Gauss-Hermite quadrature,","cited_arxiv_id":null,"evidence_quote":"Provides the Gauss-Hermite quadrature used to validate the optimized MI and BMI values."}],"review_version":1}