{"id":"3ada6966-c25d-4a14-93a9-4f9c1e987c1b","arxiv_id":"2502.04895","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A compilation of deep learning methods for channel modeling, neural decoding, mutual information estimation, and capacity learning, applied to power line communications.","lead":"This doctoral thesis applies deep learning to physical layer communications, presenting neural decoders, autoencoders, and capacity estimators for channels without analytic models. The work compiles previously published results by the author into a single framework, with applications to power line communications.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Capacity-learning claim rests on an unproven assumption that f-DIME is an unbiased MI objective and that the cooperative generator/discriminator game converges to the true capacity-achieving distribution; the visible non-Shannon validations do not eliminate this risk.","rationale":"The reader's weakest assumption and my analysis coincide on the MI-estimator bias and training-convergence risk. The available text benchmarks f-DIME on known synthetic distributions and compares against AWGN capacity, but none of these tests verifies that maximizing the neural MI estimate over a generator family recovers the true capacity-achieving distribution for a non-Shannon channel. Since the thesis's headline contribution is exactly this claim, the missing control experiment is load-bearing. MIND and other results are credible but do not mitigate this gap. The verdict remains CONDITIONAL, as the reader already stated.","tokens_in":56110,"tokens_out":3664,"duration_ms":44713,"concrete_test":"Run CORTICAL on a scalar peak-power-constrained AWGN channel with several amplitude constraints, compute the exact capacity and optimal discrete input PMF via the Blahut-Arimoto or Smith algorithm, and compare against CORTICAL's learned MI and input distribution using an independent high-accuracy MI evaluation, with multiple seeds and discriminator capacities.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Ch. 8, summarized in the abstract and Sec. 1.2) is that CORTICAL estimates channel capacity and builds capacity-achieving input distributions for any discrete-time continuous memoryless vector channel. The training objective in Sec. 8.2 maximizes a variational lower bound on I(X;Y) using f-DIME (Ch. 7) with respect to the generator's parameters. For this to recover the true capacity C = sup_p I(X;Y), three conditions must hold: (i) the f-DIME variational estimate is tight and unbiased at the maximizing distribution, or at least has a bias whose argmax coincides with the true argmax; (ii) the parametric generator family contains (or can approximate) the capacity-achieving input distribution; and (iii) the stochastic cooperative training converges to a global optimum rather than a spurious local one. None of these is established in the visible text. Sections 7.3-7.6 benchmark f-DIME on Gaussian, cubic, and other synthetic distributions, but estimation accuracy on a fixed distribution does not imply that maximizing that estimate over a parametric family recovers the true capacity-achieving distribution; even a small estimator bias can shift the argmax. Moreover, peak-power AWGN channels (Fig. 8.2) have discrete capacity-achieving distributions, and the text does not specify how CORTICAL's continuous generator represents a discrete PMF. The only quantitative capacity comparisons shown are against known AWGN capacity; no non-Shannon channel with an independently known optimum is used to verify the learned distribution. Thus the thesis's headline contribution is plausible but unsupported by the provided evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This doctoral thesis develops deep-learning methods for physical-layer communications, with contributions in copula-based density estimation (SGN, CODINE), GAN-based channel and noise synthesis, a mutual-information neural decoder (MIND), mutual-information-regularized and capacity-driven autoencoders, a family of f-divergence mutual-information estimators (f-DIME), a cooperative capacity-learning framework (CORTICAL), and applications to power-line communications and trajectory interpolation. The central claims are that MIND achieves the performance of a genie MAP decoder in unknown channels and that CORTICAL estimates channel capacity and learns capacity-achieving input distributions for arbitrary discrete-time continuous memoryless vector channels. The numerical validation visible in the thesis relies mainly on toy channels, comparisons against MAP genie decoders, and closed-form AWGN capacity curves.","tokens_in":56407,"tokens_out":4962,"duration_ms":53590,"significance":"If the capacity-learning claims hold, the thesis would represent a meaningful step toward data-driven physical-layer design for channels without tractable models. The variational derivations for f-DIME and the cooperative CORTICAL training are internally consistent, and the thesis provides useful sanity checks by comparing against known MAP decoding and closed-form AWGN capacity. The presence of machine-checkable code for MIND and CORTICAL and the detailed proofs in Sec. 7.6.2 are additional strengths. However, the significance of the central capacity-learning claim is currently limited because the visible experiments do not validate the method on any non-Shannon channel with an independently known capacity or capacity-achieving distribution.","major_comments":[{"comment":"The claim that CORTICAL estimates the channel capacity C = sup_p I(X;Y) rests on three unproven conditions: (i) the f-DIME variational estimate is tight and unbiased, or at least has an argmax that coincides with the true argmax, over the family of generator-induced input distributions; (ii) the parametric generator family can represent the capacity-achieving distribution; and (iii) the cooperative training converges to a global optimum. Sections 7.3-7.6 benchmark f-DIME on fixed distributions, which does not establish that maximizing the estimate over a parametric family preserves the true capacity-achieving distribution. The only quantitative capacity comparisons shown in Sec. 8.3 are against closed-form AWGN capacity. I ask the authors to add at least one non-Shannon validation with an independently known capacity or capacity-achieving distribution, and to report the gap between the f-DIME objective and an independent MI estimate at the learned distribution.","section":"Sec. 8.2"},{"comment":"The capacity-achieving input distribution for a peak-power-limited AWGN channel is discrete, with finitely many mass points, but the CORTICAL generator is described as a continuous neural sampler. Figure 8.2 shows markers whose radius is proportional to the PMF, but the text does not specify how a continuous generator output realizes a discrete PMF, for example through output quantization, a Gumbel-softmax relaxation, or convergence to collapsed mixture components. Without this specification, the learned 'distribution' and the reported capacity for this scenario are not well defined. The authors should state the exact generative family and show, analytically or empirically, that it can represent discrete atoms.","section":"Sec. 8.3.1, Fig. 8.2"},{"comment":"The capacity-driven autoencoder is claimed to construct capacity-approaching codes, but the supporting numerical results in Figs. 6.6-6.10 plot the same neural MI estimator that appears in the training loss. These curves are therefore not independent evidence that the achieved rate approaches the true channel capacity. The authors should compare the learned codes against known capacity or against an independently computed MI for at least one channel, or explicitly frame the results as 'estimated MI achieved by the trained code' rather than as capacity-approaching performance.","section":"Sec. 6.4"},{"comment":"MIND is presented as a general decoding principle, but the decoding rule in Eq. (5.21) requires evaluating the a-posteriori information over all x in the support T_x. For continuous input alphabets this is computationally intractable, and all numerical experiments in Sec. 5.3 use discrete alphabets. The thesis's abstract and Sec. 1.2 claim optimal coding-decoding for arbitrary communication media, but the visible MIND results support only the discrete-input case. The authors should either state this restriction explicitly or provide a tractable decoding procedure for continuous alphabets.","section":"Sec. 5.2.2, Lemma 5.2.2.1"}],"minor_comments":[{"comment":"The opening sentence of this section says 'we provide extra details on the derivation of the supervised loss function from the supervised one'; this should read 'from the unsupervised one.'","section":"Sec. 5.5.1"},{"comment":"The theorem assumes f is a convex lower semicontinuous function with f(1)=0 and then uses its derivative f'; please state the additional regularity assumptions needed for the Fenchel-conjugate result and for the pointwise identity c_U(u) = (f^*)'(\\hat{T}(u)).","section":"Sec. 3.2.1, Theorem 3.2.1.1"},{"comment":"The phrases 'any arbitrary communication medium' and 'for any arbitrary communication medium' are broader than what the visible analysis establishes; the thesis should qualify these statements with the discrete-time, memoryless, stationary assumptions used in Chs. 5-8.","section":"Abstract and Sec. 1.2"},{"comment":"The abbreviation list and Chapter 9 use 'Nakagami-m' noise, while the table of contents spells it 'Nagakami-m'; please make the spelling consistent.","section":"Chapter 9"}],"recommendation":"major_revision","confidential_remarks":"The core derivations in Chapters 5, 7, and 8 are internally consistent, and the thesis has the useful property of shipping code for MIND and CORTICAL. The main risk is the breadth of the capacity-learning claim: the visible validation is confined to AWGN and to the authors' own MI estimator, so the central claim needs either a stronger independent benchmark or a careful narrowing of scope. I see this as fixable within the manuscript's scope, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe one-sentence take: treat this as a polished doctoral thesis that compiles previously published results; the individual pieces are mostly sound and sometimes clever, but the headline claim—capacity learning for arbitrary channels—is not supported by the experiments shown, and I don't think the thesis closes that gap.\n\nWhat's actually good: MIND is a clean reformulation of MAP decoding as a discriminator that estimates the a-posteriori ratio, and the toy-channel results match the genie MAP decoder; that is a real, useful contribution. The f-DIME family and the derangement sampling trick address a genuine variance problem in neural MI estimation, and the bias/variance tables are a solid piece of empirical work. CODINE's copula-density formulation is also neat and self-consistent. CORTICAL—generator/discriminator cooperative capacity learning—is a plausible idea, and the AWGN peak-power benchmark is the right kind of external check: it recovers a known discrete capacity-achieving distribution, which is non-trivial.\n\nWhere it gets soft: the thesis claims capacity estimation and capacity-achieving input distributions for any discrete-time continuous memoryless vector channel. The visible evidence doesn't get there. The f-DIME benchmarks show accurate MI estimation on fixed distributions; that does not imply that maximizing the estimate over a parametric generator family recovers the true capacity-achieving distribution. A small estimator bias can shift the argmax, and the text doesn't analyze that. The CORTICAL sections show learned distributions for AICN and fading channels, but there is no independent optimum to compare against, so the reported capacity is the maximum of the estimator itself. There is also an unresolved representational question: for peak-power AWGN the capacity-achieving input is a discrete PMF, and the text doesn't specify how CORTICAL's continuous generator produces point masses; Fig. 8.2 shows marker radii proportional to PMF, but the mechanism is not explained. These are genuine gaps, not manufactured ones. The reader's report says Ch. 7–8 are missing; they are actually in the text, but they are thin, and the gap analysis above stands.\n\nThe thesis is honest about being a compilation; Sec. 1.4 lists the papers, which is good practice. My main caveat is that the abstract and Sec. 1.2 overclaim 'any arbitrary medium.'\n\nWho benefits: grad students and researchers new to DL-for-comms get a readable map of one group's work; MI-estimation people will want the f-DIME chapter. The capacity-learning claim needs more work before it convinces me. I'd send this to serious peer review, but with a clear request to either prove or qualify the argmax/bias issue and to validate on at least one non-Shannon channel with an independent capacity upper bound.","headline":"A coherent, honest compilation of solid but incremental DL-for-comms work; the headline capacity-learning claim is plausible but not established beyond the AWGN benchmark.","tokens_in":56946,"tokens_out":3296,"would_cite":false,"duration_ms":35386,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A17","94A40","68T07","62H05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Optimal decoding and capacity-achieving coding can be reformulated as neural estimation problems and learned directly from paired input–output samples, with no tractable channel model required.","keywords":["deep learning","physical layer communications","channel capacity","mutual information estimation","neural decoding","autoencoders","generative adversarial networks","copulas"],"falsifier":"Train CORTICAL on a scalar AWGN channel under a strict peak-power constraint, whose capacity-achieving input distribution is known analytically to be discrete with finite alphabet, and compare the learned input distribution and the estimated capacity to the exact values. A learned distribution that does not converge to the known discrete optimum, or an estimated capacity that exceeds the true channel capacity — impossible for an unbiased estimator at the optimum — would show that the training objective is chasing a biased estimate of mutual information.","tokens_in":55883,"feed_emoji":"📡","tokens_out":10395,"duration_ms":88044,"temperature":0.7,"pith_summary":"This thesis tries to restate the two foundational problems of physical-layer communications — optimal decoding and channel capacity — in the language of machine learning, and to solve them with neural networks trained on paired input–output samples. It claims that no tractable mathematical model of the channel is needed: a discriminative network can learn the a-posteriori probabilities that an optimal decoder needs, and a cooperative generator–discriminator pair can learn the input distribution that achieves capacity. The payoff, if these claims hold, is a data-driven design path for channels where no optimal strategy is known, including power-line and other non-Gaussian media. The argument throughout rests on the same neural machinery: an information-estimation principle that extracts densities and mutual information from the equilibrium of an adversarial value function.","feed_headline":"Optimal decoding and channel capacity, learned from data alone","feed_subtitle":"Neural estimates of mutual information replace the channel model in decoding and in capacity-achieving code design.","key_machinery":"The central object is the variational representation of the f-divergence, stated as Theorem 3.2.1.1: for convex lower-semicontinuous $f$ with $f(1)=0$ and Fenchel conjugate $f^*$, the divergence between a copula density and the uniform density equals a supremum over functions $T$ of $\\mathbb{E}_{c}[T(u)] - \\mathbb{E}_{\\pi}[f^*(T(u))]$, and the maximizer satisfies $(f^*)'(T(u)) = c(u)$. This identity is the machine that carries the whole argument: it turns density and mutual-information estimation into a trainable adversarial game, and the optimum of that game supplies the density ratio from which the a-posteriori distribution (MIND), the mutual information (f-DIME), and the capacity objective (CORTICAL) are all read off. The copula step — projecting data into the unit cube via their marginal cumulative distributions — is what makes the density ratio well behaved in high dimensions.","core_discovery":"The central discovery claimed is that mutual information, the quantity whose maximization defines channel capacity and whose minimization defines an optimal decoder, can be turned into a trainable objective. The thesis derives a family of estimators (referred to collectively as f-DIME) from the variational representation of the f-divergence: maximizing a value function built from paired and unpaired samples yields, at equilibrium, the density ratio of paired to product distributions, from which the a-posteriori distribution and the mutual information can be read off. This single device is then used three times: the MIND decoder learns the a-posteriori probability of each codeword given the received signal and decodes by minimizing the resulting a-posteriori information, matching a genie MAP decoder in non-uniform-source, non-linear-channel, and non-Gaussian-noise experiments; the capacity-driven autoencoder adds the estimated mutual information as a regularizer to cross-entropy training to design capacity-approaching codes; and the CORTICAL framework couples a generator that shapes the input distribution with a discriminator that estimates the achieved mutual information, claiming to estimate channel capacity and recover capacity-achieving input distributions in non-Shannon scenarios such as peak-power-limited, non-Gaussian, and fading channels.","pith_inferences":["The density-ratio machinery could be inverted into a model-validation tool: for a proposed parametric channel model, the learned ratio between observed and model-generated outputs should be flat, and any structure would localize exactly where the model fails — a use the thesis does not discuss.","Because the copula step separates dependence from marginals, a copula learned on one medium could plausibly be recombined with the marginals of another, allowing cross-medium channel and noise synthesis without retraining — a transfer possibility the thesis leaves implicit.","A stress test the thesis does not run is distribution-level verification: on a channel whose capacity-achieving input is known analytically and is non-Gaussian, CORTICAL's learned input distribution should match it exactly, not merely produce a capacity number that lands near the true value."],"forward_implications":["If the MIND claim holds, receivers for channels with no analytic model — non-linear amplifiers, non-Gaussian noise, unknown source statistics — can be trained from samples, and the same network also returns the achieved information rate and the decoding error probability.","If the capacity-driven autoencoder claim holds, constellation shaping follows the channel's mutual information instead of a hand-picked distance metric, and the reported results show lower block error rates at fixed rate for the tested channels.","If the CORTICAL claim holds, channel capacity becomes a computable number for any discrete-time continuous memoryless vector channel: two cooperating networks return both the capacity estimate and the input distribution that achieves it, matching known results for AWGN and producing new distributions for peak-power-limited, non-Gaussian, and fading channels.","If the f-DIME claim holds, the thesis supplies a family of mutual information estimators with controlled bias and variance, and every decoder, autoencoder, and capacity learner built on them inherits their reliability."],"supporting_citations":[{"why":"The classical block formulation of the communication chain and channel capacity, the problem the thesis restates in machine-learning terms.","marker":"[1]"},{"why":"Introduces the autoencoder-based end-to-end communication model that the thesis extends with mutual-information regularization.","marker":"[11]"},{"why":"Supplies the variational f-divergence lower bound via Fenchel conjugacy used in Theorem 3.2.1.1 and throughout the estimators.","marker":"[86]"},{"why":"The neural mutual information estimator (MINE) whose bias and gradient behavior motivate the thesis's own f-DIME family.","marker":"[101]"},{"why":"The MIND decoder paper: documents the value function whose equilibrium yields the a-posteriori probabilities for decoding and information-rate estimation.","marker":"[20]"},{"why":"The capacity-driven autoencoder paper: introduces the mutual-information regularizer that shapes capacity-approaching codes.","marker":"[21]"},{"why":"The f-DIME paper: the family of discriminative mutual information estimators from the f-divergence variational representation.","marker":"[22]"},{"why":"The CORTICAL paper: the cooperative generator-discriminator framework for capacity estimation and capacity-achieving input distributions.","marker":"[23]"},{"why":"The copula density neural estimation paper (CODINE): the variational copula-density estimator underlying the mutual-information estimates.","marker":"[17]"}],"fun_headline_variants":["Neural estimates of mutual information enable data-driven decoding and capacity","Mutual information as trainable objective for physical layer communication","Data alone yields optimal decoding and channel capacity via deep learning","f-DIME estimator turns mutual information into a learnable decoding objective","Deep learning decodes and designs capacity-approaching codes without channel models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the premise that the neural mutual-information estimators are unbiased enough to train against, and that the cooperative generator–discriminator training converges to the true capacity-achieving input distribution, so that optimizing the estimated quantity really optimizes the true one; the thesis validates this against analytically known capacity in only a few figures, not across the claimed generality.","fun_headline_variants_meta":{"raw":{"variants":["Neural estimates of mutual information enable data-driven decoding and capacity","Mutual information as trainable objective for physical layer communication","Data alone yields optimal decoding and channel capacity via deep learning","f-DIME estimator turns mutual information into a learnable decoding objective","Deep learning decodes and designs capacity-approaching codes without channel models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000751,"raw_usage":{"total_tokens":3356,"prompt_tokens":971,"completion_tokens":2385,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":2298}},"tokens_in":587,"tokens_out":2385,"duration_ms":17006,"temperature":1.0,"reasoning_tokens":2298,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T21:03:16.062849+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CORTICAL on a scalar AWGN channel under a strict peak-power constraint, whose capacity-achieving input distribution is known analytically to be discrete with finite alphabet, and compare the learned input distribution and the estimated capacity to the exact values. A learned distribution that does not converge to the known discrete optimum, or an estimated capacity that exceeds the true channel capacity — impossible for an unbiased estimator at the optimum — would show that the training objective is chasing a biased estimate of mutual information.","supporting_citations":[{"cited_title":"Copula Density Neural Estimation","cited_arxiv_id":"2211.15353","evidence_quote":"The copula density neural estimation paper (CODINE): the variational copula-density estimator underlying the mutual-information estimates."}],"review_version":1}