{"id":"18d1376b-36f3-42dd-b5ba-6bf1b5e2744d","arxiv_id":"2412.19025","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Without common randomness, a hybrid coding scheme can beat both separation-based and uncoded architectures for channel-aware optimal transport on binary and Gaussian channels.","lead":"This paper analyzes how to move a random sequence through a communication channel so that the output has a desired distribution while the distortion is minimized. It shows that separating source coding from channel coding is optimal when sender and receiver share randomness, but not otherwise, and it constructs a hybrid scheme that beats both standard approaches.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equality case of Theorem 2 depends on an unproved strong-data-processing lemma for the augmented distributions (157)-(158); since the binary and Gaussian hybrid schemes operate at equality, the ≤ version of (41) needs a self-contained proof.","rationale":"The central claim has three components: optimality of separation with common randomness, strict suboptimality of separation without common randomness, and an achievable hybrid scheme that can outperform both separation-based and uncoded schemes. Theorem 1 is well supported by a standard converse, and the strict-inequality part of Theorem 2 is a standard random-coding/likelihood-encoding argument. The single load-bearing weak point is the equality-case augmentation in Appendix B. The paper's own flagship demonstrations—the binary hybrid scheme with condition (64) potentially at equality and the Gaussian hybrid scheme with equality in (104)/(108)—fall precisely in the case where the augmentation is used. The assertion that (157)-(158) force I(X^(k);Z) < I(X;Z) unless the mutual information is zero is a strong data-processing inequality; it is not proved in the text, and the cited textbook problem is not quoted or checked. This matches the reader's weakest_assumption, so I agree with that identification. There is no evidence of a fatal flaw; the theorem is likely repairable. A direct proof of the contraction property, or replacement of the external citation, would settle the issue. Minor issues such as the 'R > I(Z;V)' phrase in the decoding step of Appendix B (which should be 'R < I(Z;V)' to match the random-coding union bound) and the Λ/Σ notation slip in Theorem 3 do not change the assessment. The proper verdict remains CONDITIONAL, so no change to the reader's verdict is needed.","tokens_in":23849,"tokens_out":23496,"duration_ms":220718,"concrete_test":"Give a self-contained proof of the following claim: for the kernel p_{X^(k)|X} defined in (157), with all off-diagonal entries positive, there exists η_k<1 such that for every Z with the Markov chain X^(k)-X-Z and I(X;Z)>0, we have I(X^(k);Z) ≤ η_k I(X;Z), then apply the same argument to Y^(k). One way is to identify the exact statement of [36, p.402, Problem 25] and verify its hypotheses; another is to prove the strong data-processing inequality directly via the Dobrushin contraction coefficient of the kernel. If the claim is proved, insert the argument before the sentence 'Hence, property 3) is also satisfied' in Appendix B.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Appendix B, the strict-inequality achievability is careful and standard. The fragile step is the equality case max{I(X;Z), I(Y;Z)} = I(Z;V), used to state condition (41) with ≤. The paper augments p_X and p_Y via (157)-(158) and asserts that because p_{XX^(k)} is 'indecomposable', [36, p.402, Problem 25] gives I(X^(k);Z) < I(X;Z) or I(X^(k);Z)=0, and similarly for Y; hence max{I(X^(k);Z), I(Y^(k);Z)} < I(Z;V). This is a strong data-processing-type statement about the kernel from X to X^(k): for every Z with X^(k)-X-Z, the mutual information must strictly decrease unless it is zero. The paper does not state the definition of indecomposability used in [36], nor does it connect the hypotheses of Problem 25 to the specific coupling (157)-(158). This is not a purely cosmetic gap: the binary hybrid scheme enforces (64) as an equality, and the Gaussian hybrid scheme operates at equality through (104) and (108). If the cited lemma fails, the achievability proofs of the headline hybrid-coding comparisons in Sections IV and V are incomplete, and Theorem 2 would be formally established only for the strict-inequality version of (41). The argument is probably repairable, since the all-positive transition in (157) should admit a contraction coefficient, but the paper as written does not supply the proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies channel-aware optimal transport (CAOT), where a block of i.i.d. source variables is transmitted over a memoryless channel to generate a block of i.i.d. output variables with a prescribed marginal distribution, minimizing end-to-end distortion. The main results are: (i) with unlimited common randomness, the source-channel separation architecture is asymptotically optimal (Theorem 1); (ii) without common randomness, separation is generally suboptimal, as shown by binary and Gaussian toy examples; and (iii) a hybrid coding scheme is proposed (Theorem 2) that combines uncoded transmission with digital coding, and it is shown to outperform both separation-based and uncoded schemes in binary and Gaussian settings. The paper provides explicit single-letter expressions and analytic comparisons in Sections IV and V.","tokens_in":24188,"tokens_out":22045,"duration_ms":189905,"significance":"If the main claims are fully established, the paper makes a valuable contribution to generative communication: it shows that the classical point-to-point source-channel separation theorem fails when a perception constraint (exact output marginal) is imposed and common randomness is absent, and it provides a constructive hybrid coding scheme that exploits channel stochasticity. The clean converse in Theorem 1, the explicit toy examples, and the detailed binary/Gaussian analyses are strengths. The proof of the strict-inequality version of Theorem 2 is a careful random-coding argument using the likelihood encoder and soft covering, and the paper gives concrete, checkable formulas for the schemes. However, two load-bearing technical points currently prevent the results from being fully supported: the equality-case argument in Appendix B relies on an unproved strong-data-processing assertion, and an algebraic threshold in the binary comparison appears to be incorrect.","major_comments":[{"comment":"The proof that the augmented distributions (157)-(158) satisfy max{I(X^(k);Z), I(Y^(k);Z)} < I(Z;V) rests entirely on the assertion that 'pXX^(k) is indecomposable' and the citation to [36, p. 402, Problem 25]. The paper neither states the strong-data-processing lemma being invoked nor verifies its hypotheses for the specific kernel (157)-(158), and it does not address the fact that (157)-(158) require pmin_X and pmin_Y to be positive, which is not assumed in Theorem 2. This gap is load-bearing because the binary hybrid scheme in Section IV enforces (64) as an equality and the Gaussian hybrid scheme in Section V operates at equality through (106)-(108); thus the ≤ version of (41) is essential for those comparisons. As written, only the strict-inequality version of Theorem 2 is fully proved.","section":"Appendix B, equality case of Theorem 2"},{"comment":"The algebraic step claiming that (-ρ^2 + (2ρ^2 - 2ρ + 1)θ)θ ≥ 0 for θ ≥ ρ^2/(2ρ^2 - ρ + 1) is incorrect: substituting the stated threshold gives -ρ^3/(2ρ^2 - ρ + 1)θ, not 0. The correct threshold appears to be ρ^2/(2ρ^2 - 2ρ + 1) (note the changed denominator). Consequently, the proof that D'_H < DU over the stated interval [ρ^2/(2ρ^2 - ρ + 1), 1/2) is invalid, and the analytical claim of hybrid superiority over the uncoded scheme in this portion of the binary case is not established. The numerical plots suggest the qualitative conclusion may still hold, but the interval and the proof need to be corrected.","section":"Section IV, equation (82)"}],"minor_comments":[{"comment":"Both definitions denote the minimum achievable distortion by the same symbol DJ(Γ), which is confusing because the two scenarios have different values; please use distinct notations (e.g., D_J^CR(Γ) and D_J^{no-CR}(Γ)) or introduce a convention after Theorem 1.","section":"Section II, Definitions 1 and 2"},{"comment":"The conditional expectation should be E[X|g^T X + N] = s \\tilde X (a column vector times a scalar); the superscript T in s^T \\tilde X is a typo and makes the expression dimensionally inconsistent. The same notation issue appears in (162)-(163) and (171)-(174).","section":"Appendix C, equations (161)-(163)"},{"comment":"The statement says Y ~ N(0, Λ), but the proof in (171) uses W_2^2(N(0,Σ), N(0,γ s_1 s_1^T)), which presumes Y has covariance Σ. Either set Λ = Σ in the theorem or adjust the proof to use Λ explicitly.","section":"Appendix C, Theorem 3 statement"},{"comment":"There are several typos: 'Combing' should be 'Combining' before (57); 'when when θ is close to' has a duplicated 'when'; 'speration-based scheme' in Section V should be 'separation-based scheme'; 'intially remains at zero' should be 'initially remains at zero'; 'it follows it follows the trajectory' in Section IV has a duplicated phrase; 'supercript' should be 'superscript' in Section V.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a genuine contribution to generative communication, and the core ideas—especially the common-randomness separation theorem and the hybrid coding construction—are valuable. The two major issues I identified are technical and appear repairable within the manuscript's scope: the equality case of Theorem 2 needs a self-contained strong-data-processing argument (and a treatment of zero-probability source symbols), and the binary threshold in Section IV needs a corrected calculation. The self-citations to [5], [17], [21], and [22] are appropriate because those works provide the source-coding building blocks; the paper's own results do not reduce to them. I would support publication after these load-bearing points are fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is worth a serious read. Its central finding is that for optimal transport through a channel with a hard marginal constraint on the reconstruction, source-channel separation is asymptotically optimal when unlimited common randomness is available, but generally suboptimal without it, even in the point-to-point setting. The second part is the conceptually important one, and it gives theoretical cover to the common empirical observation that deep joint source-channel coding can beat separation under perception constraints.\n\nWhat is genuinely new: the problem formulation itself (channel-aware optimal transport with p_{Y^n}=p_Y^n exactly), and the hybrid coding scheme of Theorem 2, which is a real construction and not a tweak. The strict-inequality achievability proof is a careful random-coding argument using the likelihood encoder and soft covering; I read it through and the steps hold. The binary and Gaussian sections are explicit and analytically checked, with the uncoded-vs-separation comparisons worked out in closed form. Theorem 1's converse is standard but correct, and the Gaussian linear-scheme optimality in Theorem 3 is a nice extra.\n\nThe soft spot is real but local. In Appendix B, the equality case of condition (41), where max{I(X;Z), I(Y;Z)} = I(Z;V), is dispatched by augmenting p_X and p_Y via (157)-(158) and asserting that the coupling is 'indecomposable', citing [36, p. 402, Problem 25] for the needed strict drop in mutual information. The paper never states the relevant definition or checks the hypotheses. That is a genuine gap, because the binary and Gaussian hybrid schemes operate exactly at equality (see (64) and (104)-(108)); without the strict-decrease lemma, the ≤ version of (41) is not fully proved. I think the claim is likely true: the transition in (157) is all-positive when p_X has full support, which should give a contraction coefficient. But as written it is a cite-out, and a referee should ask for the proof. There is also a silent full-support assumption: if pmin = 0, the augmentation degenerates and the strict decrease is simply false.\n\nMinor stuff: the vector Gaussian benchmark relies on unreviewed preprints from the same group ([21], [22]). Self-citation is not the issue; the numerical comparisons inherit the unreviewed status. Theorem 3 has notational inconsistencies (Λ vs Σ, s vs s1). Typos throughout ('supercript', 'intially').\n\nWho should read it: information theorists working on rate-distortion-perception and generative JSCC; it gives a concrete theoretical target. It deserves a serious referee. I would send it out and ask for a repaired equality-case proof, a full-support caveat, and a notation cleanup. The theorems are solid enough that the fixes are routine rather than creative.","headline":"A new and conceptually important result for generative communication, with a repairable proof gap in the equality case of the main achievability theorem.","tokens_in":24664,"tokens_out":5117,"would_cite":true,"duration_ms":48390,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A17","94A34","49Q22"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that source-channel separation is asymptotically optimal for channel-aware optimal transport when unlimited common randomness is available, but generally suboptimal without it, and gives a hybrid coding scheme that…","keywords":["channel-aware optimal transport","generative communication","common randomness","source-channel separation","hybrid coding","rate-distortion-perception","soft covering","likelihood encoder"],"falsifier":"Take a concrete instance of the equality case $\\max\\{I(X;Z),I(Y;Z)\\}=I(Z;V)$, e.g. $X\\sim\\mathrm{Bernoulli}(1/2)$ and $Z$ the output of a binary symmetric channel from $X$, and compute $I(X^{(k)};Z)$ for the perturbation (157); if any $k$ gives $I(X^{(k)};Z)\\ge I(X;Z)$ while $p_{X^{(k)}}=p_X$, the Appendix B argument for the boundary case collapses. A broader refutation would be a finite-alphabet example satisfying the hypotheses of Theorem 2 in which no sequence of augmented distributions with properties 1)-3) exists, showing the $\\le$ version requires an extra condition.","tokens_in":23629,"feed_emoji":"📡","tokens_out":8490,"duration_ms":81352,"temperature":0.7,"pith_summary":"This paper studies channel-aware optimal transport: a block of i.i.d. random variables is sent through a memoryless channel so that the receiver can generate another block with a prescribed marginal distribution while minimizing end-to-end distortion. The central claim is that whether the classical source-channel separation architecture is asymptotically optimal hinges on common randomness. With unlimited shared randomness, separation achieves the fundamental limit, $D_J(\\Gamma)=D_S(\\Gamma)$. Without common randomness, separation is generally strictly suboptimal, and the paper constructs a hybrid coding scheme whose achieved distortion obeys $D_J(\\Gamma)\\le \\mathbb{E}[d(X,Y)]$ whenever a rate-matching condition $\\max\\{I(X;Z),I(Y;Z)\\}\\le I(Z;V)$ holds. The binary and Gaussian analyses show this scheme can dominate both separation-based and uncoded schemes, giving an information-theoretic rationale for generative joint source-channel coding in point-to-point links.","feed_headline":"Hybrid coding beats separation when no common randomness exists","feed_subtitle":"Why it matters: separation is suboptimal when the decoder must generate a prescribed distribution.","key_machinery":"The load-bearing object is the auxiliary random variable $Z$ together with the condition $\\max\\{I(X;Z),I(Y;Z)\\}\\le I(Z;V)$, which balances how much information about the source and the reconstruction must be carried through $Z$ against how much of $Z$ the channel output $V$ can reveal. The hybrid scheme splits the channel input into an uncoded part, whose noise is deliberately used as a generative resource, and a coded part that transmits a digital message derived from $Z$; the decoder combines the recovered $Z$ with the raw channel output and applies a maximal coupling to enforce the prescribed $p_Y$. The proof machinery consists of the likelihood encoder, which stochastically maps $X^n$ to a codeword $Z^n(m)$, the soft-covering lemma, which guarantees the generated output approximately follows $p_Y^n$, and joint typicality decoding, which ensures the digital message survives the channel when the rate $R$ lies between $\\max\\{I(X;Z),I(Y;Z)\\}$ and $I(Z;V)$.","core_discovery":"On the paper's own terms, the discovery is a pair of results about the minimum distortion $D_J(\\Gamma)$ achievable when transporting $p_X$ to $p_Y$ through a memoryless channel under an input cost constraint. Theorem 1 states that with unlimited common randomness, $D_J(\\Gamma)=D_S(\\Gamma)$, where $D_S(\\Gamma)$ is the distortion achieved by converting the channel into a rate-limited bit pipe at capacity and then performing rate-limited optimal transport; the converse is proved by a time-sharing argument that extracts a single-letter pair $(X_T,Y_T)$ and bounds its mutual information by the channel capacity. Theorem 2 states that without common randomness, $D_J(\\Gamma)$ is no larger than $\\mathbb{E}[d(X,Y)]$ for any auxiliary $Z$ with $X\\leftrightarrow Z\\leftrightarrow V$, $Y\\leftrightarrow Z\\leftrightarrow V$ structure satisfying $\\max\\{I(X;Z),I(Y;Z)\\}\\le I(Z;V)$ and $\\mathbb{E}[c(U)]\\le\\Gamma$; the associated hybrid codebook, likelihood encoder, joint typicality decoder, and soft-covering decoder achieve the bound. The binary BSC analysis and the Gaussian AWGN analysis then display parameter regimes where the optimized hybrid scheme beats both the separation benchmark $D_S(\\Gamma)$ and the best uncoded scheme, with the optimizer switching modes as the channel degrades.","pith_inferences":["Beyond the paper's claims, the boundary gap suggests a testable refinement: one could compute $I(X^{(k)};Z)$ explicitly for small-alphabet examples to check whether the asserted strict decrease in Appendix B always holds, and if not, state Theorem 2 with a strict-inequality hypothesis.","The hybrid principle points to a practical architecture for generative image transmission: reserve a fraction of channel uses for uncoded analog transmission so the channel noise acts as a generative prior, and use the remaining bandwidth for a digital representation, with the fraction tuned by the mutual-information condition.","A natural extension, not pursued here, is bandwidth mismatch: replacing $I(Z;V)$ by a per-symbol capacity and rebalancing the analog/digital split should yield a generalized curve interpolating between uncoded and separation schemes.","The authors' one-shot direction could be made quantitative by replacing soft covering with finite-blocklength covering bounds, which would turn the asymptotic theorem into a bound on achievable distortion for generative codecs at practical blocklengths."],"forward_implications":["In point-to-point generative communication with a prescribed reconstruction distribution and no shared randomness, separation is provably suboptimal even in the infinite-blocklength limit, unlike classical source-channel communication.","The achievability bound gives a quantitative design rule: allocate channel input between an analog component (left for the channel to randomize) and a digital component (protected by coding), choosing the split so that the digital rate matches the auxiliary mutual information condition (41).","In the binary BSC case, the optimized hybrid scheme operates in distinct modes—separation for small crossover probability, uncoded for intermediate values, and a hybrid form for large values—with explicit threshold structure.","In the vector Gaussian case, the hybrid scheme dominates both separation and uncoded schemes at every power level, and it reduces to the uncoded scheme below an explicit power threshold $\\Gamma^*$.","If Theorem 2 is correct, deep joint source-channel coding systems that let the channel's randomness shape the reconstruction are not merely finite-blocklength heuristics; they realize a genuine asymptotic advantage."],"supporting_citations":[{"why":"Supplies the rate-limited optimal transport formulas for the with- and without-common-randomness cases that define the separation benchmarks $D_S(\\Gamma)$ and $\\bar D_S(\\Gamma)$.","marker":"[5]"},{"why":"Gives the binary rate-distortion-perception functions used to evaluate the separation benchmark and the binary hybrid scheme.","marker":"[17]"},{"why":"Gives the vector Gaussian rate-distortion-perception function used for the common-randomness benchmark in the Gaussian analysis.","marker":"[21]"},{"why":"Gives the scalar Gaussian rate-distortion-perception functions used in the Gaussian toy example.","marker":"[22]"},{"why":"Supplies the indecomposability argument (Problem 25, p. 402) on which the equality-case proof of Theorem 2 depends.","marker":"[36]"},{"why":"Provides the reverse waterfilling formula for the Gaussian distortion-rate function used by the separation-based scheme.","marker":"[42]"},{"why":"Supplies the likelihood encoder whose stochastic mapping is the key to the hybrid scheme's soft covering and distortion analysis.","marker":"[43]"},{"why":"Supplies the soft-covering lemma used to guarantee the decoder's output distribution approximates $p_Y^n$ and to control total variation distances in the proof.","marker":"[44]"}],"fun_headline_variants":["Hybrid coding beats separation without shared randomness","Channel-aware transport: hybrid coding wins without common randomness","No common randomness? Hybrid coding outperforms separation","Hybrid scheme bests separation in generative communication"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorem's boundary case rests on an unproved assertion that the perturbed source and sink distributions defined in (157)-(158) are indecomposable, so their mutual information with $Z$ strictly decreases; if that assertion fails, only the strict-inequality version of the achievability bound is fully established.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid coding beats separation without shared randomness","Channel-aware transport: hybrid coding wins without common randomness","No common randomness? Hybrid coding outperforms separation","Hybrid scheme bests separation in generative communication"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1447,"prompt_tokens":1031,"completion_tokens":416,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":357}},"tokens_in":647,"tokens_out":416,"duration_ms":4383,"temperature":1.0,"reasoning_tokens":357,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:59:27.172777+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a concrete instance of the equality case $\\max\\{I(X;Z),I(Y;Z)\\}=I(Z;V)$, e.g. $X\\sim\\mathrm{Bernoulli}(1/2)$ and $Z$ the output of a binary symmetric channel from $X$, and compute $I(X^{(k)};Z)$ for the perturbation (157); if any $k$ gives $I(X^{(k)};Z)\\ge I(X;Z)$ while $p_{X^{(k)}}=p_X$, the Appendix B argument for the boundary case collapses. A broader refutation would be a finite-alphabet example satisfying the hypotheses of Theorem 2 in which no sequence of augmented distributions with properties 1)-3) exists, showing the $\\le$ version requires an extra condition.","supporting_citations":[{"cited_title":"Output constrained lossy source coding with limited common randomness,","cited_arxiv_id":null,"evidence_quote":"Supplies the rate-limited optimal transport formulas for the with- and without-common-randomness cases that define the separation benchmarks $D_S(\\Gamma)$ and $\\bar D_S(\\Gamma)$."},{"cited_title":"On the r ate-distortion-perception function,","cited_arxiv_id":null,"evidence_quote":"Gives the binary rate-distortion-perception functions used to evaluate the separation benchmark and the binary hybrid scheme."},{"cited_title":"Rate-Distortion-Perception Tradeoff for Gaussian Vector Sources","cited_arxiv_id":"2406.18008","evidence_quote":"Gives the vector Gaussian rate-distortion-perception function used for the common-randomness benchmark in the Gaussian analysis."},{"cited_title":"Csisz´ ar and J","cited_arxiv_id":null,"evidence_quote":"Supplies the indecomposability argument (Problem 25, p. 402) on which the equality-case proof of Theorem 2 depends."},{"cited_title":"The likelihood encode r for lossy compression,","cited_arxiv_id":null,"evidence_quote":"Supplies the likelihood encoder whose stochastic mapping is the key to the hybrid scheme's soft covering and distortion analysis."},{"cited_title":"Soft covering with high probability,","cited_arxiv_id":null,"evidence_quote":"Supplies the soft-covering lemma used to guarantee the decoder's output distribution approximates $p_Y^n$ and to control total variation distances in the proof."}],"review_version":1}