{"id":"9ddda366-468c-4d2c-b1e4-1ca5def509d6","arxiv_id":"2508.07567","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An extended Arimoto-Blahut algorithm computes the rate-distortion function for bistatic ISAC systems by converting non-convex distortion constraints into linear constraints via auxiliary variables.","lead":"This paper develops an extended Arimoto-Blahut algorithm to compute rate-distortion tradeoffs in bistatic integrated sensing and communication (ISAC) systems, where distortion constraints are non-convex. The method introduces auxiliary variables to turn those constraints into linear ones and gives closed-form updates for squared-error and log-loss distortion metrics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global optimality of Algorithm 1 is unproven: monotone surrogate convergence (Thm 2) is insufficient for the non-concave objective F(p), so the claim to compute C(D) is not established.","rationale":"I read the manuscript as claiming a computational method for the non-convex rate-distortion optimization in bistatic ISAC. The main technical steps are: (i) Theorem 1's equivalence between (1) and the linearly constrained (3); (ii) construction of the surrogate \\tilde{F}; (iii) the alternating closed-form updates; (iv) Theorem 2's convergence. The weakest point is (iii)/(iv): the algorithm is a block-coordinate ascent on a function that is concave in each block separately but not jointly concave. For such problems, standard results guarantee convergence to a stationary point only, and there is no reason to expect the global maximum unless the problem is convex or the objective has special structure. The paper does not provide such structure; it even states the objective is non-convex. The numerical section cannot substitute for a proof: Fig. 2 shows agreement with a theoretical value in one case and superiority to a lower bound in another, but neither demonstrates global optimality. I also note minor issues: the KKT step in Theorem 1 assumes an active constraint in a way that could fail when λ=0 (though the theorem's conclusion is salvageable), and the log-loss algorithm is not fully specified. These are secondary; the global-convergence gap is the decisive concern. The reader's verdict of CONDITIONAL is appropriate: the gap is addressable by adding a global-convergence proof (e.g., by showing F is concave under the given Markov-chain and Gaussian assumptions, or by proving the fixed point is a global maximizer via duality), but without it the central claim is not rigorously established.","tokens_in":8831,"tokens_out":21029,"duration_ms":225619,"concrete_test":"Define a small finite-alphabet instance (e.g., binary S,X,Y,Z and U with |U|≤2) with transition probabilities p(y,z,s|x) and a distortion D strictly between the minimum and maximum achievable distortions. Compute the exact global optimum of (1) by exhaustive enumeration over all p(u,x) on a sufficiently fine grid (or via a global optimization oracle). Run Algorithm 1 from at least 20 random initializations of p0(u,x) and c0(u,z). If any run terminates with F(p_k) strictly below the exhaustive global optimum (within a tolerance accounting for the grid), the central claim is refuted. Conversely, if all runs converge to the exhaustive optimum, the practical concern is weakened, but a formal global-convergence proof would still be required for the theoretical claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that Algorithm 1 computes the capacity-distortion function C(D) of problem (1). This requires the alternating maximization of the surrogate \\tilde{F} in (4) to converge to a global maximum of F(p)=I(X;Y|U,S)+I(U;Z) over the distortion-feasible set. Theorem 2 only proves monotone non-decrease and boundedness of \\tilde{F} along the iterates; it does not identify the limit as the global maximizer. This gap is load-bearing because the optimization is non-convex: with U non-degenerate, I(U;Z)=H(Z)-H(Z|U) is the difference of concave entropies and is generally not concave in p(u,x), and I(X;Y|U,S) is concave but the sum need not be. The exact p-update (9) maximizes \\tilde{F}(·,q_k) for a fixed q_k, but \\tilde{F} is merely a lower bound of F that is tight only at the conditional q; a monotone bounded sequence on a non-concave objective can converge to a stationary point (local maximum or saddle) with value strictly below C(D). No convexity, strong-duality, or global-convergence argument is supplied. The log-loss variant (Sec. III-B) is even less supported, since its updates and convergence proof are omitted. Thus the abstract's assertion that the algorithm 'calculate(s) the rate-distortion trade-off' is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops Arimoto-Blahut-type algorithms for computing the capacity-distortion function C(D) of a bistatic ISAC system under squared-error and log-loss distortions. The original non-convex optimization problems (1) and (2) are reformulated by introducing auxiliary estimator variables c(u,z) and f(u,s,z), which turn the distortion constraints into linear constraints. The authors prove a formal equivalence (Theorem 1), derive closed-form alternating updates for p(u,x), q, and c/f, prove monotone non-decrease of the surrogate function (Theorem 2), and report numerical comparisons for Gaussian channels. The central claim is that Algorithm 1 calculates the rate-distortion trade-off.","tokens_in":9231,"tokens_out":11786,"duration_ms":129936,"significance":"If the global-convergence gap can be closed, this is a useful extension of the AB method to non-convex rate-distortion problems and would make the information-theoretic limits of bistatic ISAC computable in practice. The reformulation is natural, the updates are closed-form, and the monotonicity theorem is a useful partial result. The numerical agreement with known theoretical values in the log-loss case is encouraging. However, the central claim is currently only partially supported: monotone convergence of the surrogate does not by itself establish that the limit equals C(D), and the continuous-alphabet implementation is unspecified. These issues are load-bearing and require revision.","major_comments":[{"comment":"The proof of Theorem 2 establishes only \\tilde{F}(p_{n-1},q_{n-1}) ≤ \\tilde{F}(p_{n-1},q_n) ≤ \\tilde{F}(p_n,q_n), i.e., monotone non-decrease and boundedness of the surrogate. Since \\tilde{F} is biconcave but not jointly concave, and since F(p) in (1) is not concave (I(U;Z) contains -H(Z|U)), coordinate ascent can converge to a stationary point or saddle with value strictly below max_p F(p). The p-update (9) maximizes the surrogate for fixed q, not F itself; the q-update (6) makes the surrogate tight only at the current p. No global-convergence, strong-duality, or KKT-sufficiency argument is supplied. Thus the abstract's claim that Algorithm 1 'calculate[s] the rate-distortion trade-off' is not established. Please add a global optimality proof under stated conditions, or revise the claims to convergence to a local/stationary point and support the computation claim with additional argumen","section":"Section III-C, Theorem 2"},{"comment":"In the proof of Theorem 1, the KKT stationarity condition L'_c(p*,c*)=0 yields c*=\\hat{s} only when the multiplier λ is strictly positive, i.e., when the distortion constraint is active. If λ=0, the stationarity condition is vacuous and the argument fails. The monotonicity assertion later in Section III-A implies activity only for D below the unconstrained-achievable distortion, a condition not stated in Theorem 1. Moreover, the algorithm's Step 3, G(λ)=0, assumes equality in the constraint; for D larger than the distortion achieved by the unconstrained maximizer, no λ≥0 solves G(λ)=0. The equivalence can be repaired by observing that c enters only through the constraint and can be replaced by the conditional-mean estimator \\hat{s}(p) without changing feasibility or objective, but the manuscript should state and prove this and handle the inactive-constraint case explicitly.","section":"Section III-A, Theorem 1"},{"comment":"Algorithm 1 is formulated for finite alphabets with sums over p(u,x), p(y,s,z|x), etc. The numerical section applies it to continuous Gaussian channels with a power constraint, but no discretization, grid size, initialization, stopping criterion, or numerical method for solving the multiplier equations (λ, μ) is provided. Consequently, the validation in Figs. 1 and 2 cannot be reproduced or fully assessed, and the extension of the discrete-alphabet AB updates to continuous alphabets is not justified. Please provide full implementation details or a separate derivation for the Gaussian case.","section":"Section IV and Remark 1"},{"comment":"The log-loss variant is presented only as a set of update formulas; the pseudocode, the derivation of the p-update, and any convergence analysis are omitted ('details are omitted due to space limitations'). Since log-loss is one of the two distortion metrics advertised in the abstract and is used in Fig. 2, this is a substantive omission. Please include the full algorithm and at least a monotonicity/convergence statement analogous to Theorem 2, or clearly state that no convergence guarantee is claimed.","section":"Section III-B"}],"minor_comments":[{"comment":"The displayed expression for F(p) - \\tilde{F}(p,q) writes the first KL term as D(p(x,y|u,s) || q(x,y|u,s)); since q has no y component, this should be D(p(x|u,y,s) || q(x|u,y,s)), and the second term should be D(p(u|z) || q(u|z)).","section":"Section III-A, text after Eq. (4)"},{"comment":"No stopping criterion is given; please specify one (e.g., tolerance on function values or on p, or a maximum iteration count).","section":"Algorithm 1"},{"comment":"The labels '1/σ2=1' and '1/σ2=3' are ambiguous; specify whether these refer to σ_1^2, σ_2^2, or their inverses, and state the other parameters used for each curve.","section":"Section IV, Fig. 2"},{"comment":"The statement that Step 3 is replaced by solving a set of equations for λ and μ is too terse: no uniqueness, monotonicity, or solution method is given for the two-multiplier case.","section":"Section III-A, Remark 1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a real computational gap and the reformulation idea is promising. The main risk is the unproven global convergence: if the authors cannot provide a global-optimality argument or a clear convergence-to-local-optimum statement, the central claim should be downgraded. The numerical section also needs full implementation details. I believe major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper extends Arimoto-Blahut algorithms to compute the rate-distortion trade-off in bistatic ISAC systems with squared-error and log-loss distortion. The new idea is introducing an auxiliary variable to convert the non-convex distortion constraint into a linear constraint, then alternating maximization of an AB surrogate. The reformulation in Theorem 1 is algebraically correct (once you assume the state is independent of the input), and the closed-form updates are a genuine extension of AB to a problem class it didn't previously handle. The numerical results match known theory in the cases shown, and the comparison against [12] is appropriate.\n\nThe main soft spot is the gap between what the algorithm provably does and what the abstract claims. Theorem 2 proves only monotone non-decrease of the surrogate; it does not establish global convergence of the alternating scheme. Since the objective I(X;Y|U,S)+I(U;Z) is generally not concave in p(u,x), the iterates could converge to a stationary point below C(D). So the claim that the algorithm 'calculates the rate-distortion trade-off' is not proven. This is the load-bearing issue. It's fixable: a duality/concavity argument, a global-convergence proof, or a carefully worded statement that the algorithm provides a lower bound would close the gap.\n\nTwo smaller issues. First, the proof of Theorem 1 assumes the distortion constraint is active (λ > 0) without stating that condition; the argument can be patched (the monotonicity of C(D) helps), but as written it's incomplete. Second, the log-loss section omits the actual update details and convergence analysis (\"omitted due to space limitations\"), and the Gaussian numerical section never explains how the continuous alphabet is discretized. These omissions make the log-loss results hard to reproduce.\n\nThe core idea is sound, the extension is real, and the citation pattern is honest. The paper deserves a serious referee, but not in its current form. I'd recommend sending it to peer review with a request for a global-convergence analysis (or a sharpened claim), plus the missing details for log-loss and the Gaussian discretization.","headline":"A useful AB-style extension for bistatic ISAC rate-distortion, but the headline claim of computing C(D) goes beyond the proved convergence; worth a revision, not a desk reject.","tokens_in":9647,"tokens_out":11029,"would_cite":true,"duration_ms":110333,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A17","94A34","90C26"],"pacs":[],"model":"deepseek-v4-flash","headline":"A classical alternating-maximization method, extended by replacing each estimator with a free auxiliary variable, computes the capacity-distortion tradeoff of bistatic integrated sensing and communication systems for both squared-error and","keywords":["integrated sensing and communication","bistatic ISAC","rate-distortion function","capacity-distortion function","AB algorithm","non-convex constraints","squared-error distortion","logarithmic loss"],"falsifier":"Take a small finite-alphabet instance, enumerate or globally optimize over all input distributions $p(u,x)$, and compare the true optimal value with the value Algorithm 1 returns from many random initializations. Any run that stops at a strictly smaller value while satisfying the distortion constraint, or two initializations that settle at different values, would falsify the claim that the algorithm calculates the rate-distortion tradeoff.","tokens_in":8739,"feed_emoji":"📡","tokens_out":7315,"duration_ms":72276,"temperature":0.7,"pith_summary":"Bistatic integrated sensing and communication (ISAC) systems face a fundamental tradeoff: the more a codeword tells the communication receiver, the less accurately a separated sensing receiver can estimate the environment from the same transmission. The optimal tradeoff is the capacity-distortion function, but the optimization problems defining it carry distortion constraints that are non-convex in the input distribution, which blocks the standard AB algorithm. The paper's central move is to replace the estimator in the distortion term with a free auxiliary variable ($c(u,z)$ for squared error, $f(s,u,z)$ for log-loss), which makes the constraint linear while provably preserving the optimum. On that reformulation, the paper constructs extended AB algorithms with closed-form updates, proves the objective is monotonically non-decreasing along iterations, and shows numerical rate-distortion curves that match known values or exceed known lower bounds. A sympathetic reader takes away a practical method for drawing the full rate-distortion curve of bistatic ISAC channels instead of relying on theoretical bounds alone.","feed_headline":"Bistatic ISAC rate-distortion curves now computable by AB algorithm","feed_subtitle":"A linear constraint trick, plus closed-form updates, handles squared-error and log-loss sensing distortions.","key_machinery":"The auxiliary-variable linearization of the distortion constraint. For squared error, the estimator $\\hat{S}(U,Z)$ is pulled out of the constraint and treated as an independent function $c(u,z)$, so the constraint becomes a sum of terms $p(u,x)(s-c(u,z))^2$; for log-loss, the conditional estimator $p(s|u,z)$ is replaced by a free soft function $f(s,u,z)$. The AB surrogate $\\tilde{F}(p,q)$ then separates the variables, and the Lagrange multiplier $\\lambda$, found by solving the one-dimensional equation $G(\\lambda)=0$, pins the iterate to the distortion boundary. Each block update ($q$, then $p$, then $c$ or $f$) has a closed form.","core_discovery":"The paper claims that the capacity-distortion optimization of a bistatic ISAC system can be solved by an extended AB algorithm. Its pivotal result (Theorem 1) equates the optimum of the original non-convex problem (1), with squared-error distortion and objective $I(X;Y|U,S)+I(U;Z)$, to the optimum of a reformulated problem in which the distortion constraint is linear because the estimator $\\hat{S}(U,Z)$ is replaced by a free variable $c(u,z)$. At the optimum, $c(u,z)$ recovers exactly the conditional mean $\\hat{S}(u,z)$, so the reformulation's feasible set collapses back to the original one and the two optima coincide. The same equivalence is developed for log-loss distortion with a free sof","pith_inferences":["If the auxiliary-variable linearization is as general as the proof suggests, it should also apply to other bounded distortion measures whose optimal estimator has a closed form, such as absolute error or Hamming loss under discrete alphabets; the paper does not test this.","The global-convergence question is left open by Theorem 2; a natural extension is a formal proof that every limit point of the iterates satisfies the KKT conditions and that the surrogate has no spurious local maxima, or a counterexample if it does.","For the second log-loss Gaussian regime the computed curve sits above the known lower bound; if that bound is not tight, the algorithm may be providing the first numerical evidence of a larger rate-distortion region than theory currently certifies.","The same reformulation appears portable to monostatic ISAC or multi-antenna settings where the distortion constraint is non-convex, although the paper restricts itself to the bistatic model."],"forward_implications":["The non-convex distortion constraint can be replaced by a linear one without changing the optimum, so the original problem becomes tractable by alternating maximization.","Both the squared-error and log-loss versions admit closed-form updates for $p$, $q$, and the auxiliary estimator, avoiding the multiplier traversal of classical AB rate-distortion computation.","The same framework extends directly to Gaussian channels with a power constraint by adding a second Lagrange multiplier, yielding rate-distortion curves for practical bistatic channels.","The monotone convergence proof guarantees that any stopping point of the iteration is at least a local optimum of the surrogate, making the algorithm a stable numerical tool for capacity-distortion evaluation."],"supporting_citations":[{"why":"States Lemma 1: the squared-error capacity-distortion problem (1) that the algorithm solves.","marker":"[11]"},{"why":"States Lemma 2: the log-loss capacity-distortion problem (2) and supplies the theoretical value or lower bound used in the numerical comparisons.","marker":"[12]"},{"why":"Provides the original alternating-maximization AB algorithm for channel capacity whose update structure the extended algorithm generalizes.","marker":"[13]"},{"why":"Provides the original AB framework for rate-distortion functions, including the surrogate and multiplier ideas adapted here.","marker":"[14]"},{"why":"Demonstrates that an AB-type algorithm can be extended to a non-convex broadcast-channel capacity problem, motivating the extension to non-convex constraints.","marker":"[15]"},{"why":"Supplies AB-type algorithms for capacity bounds of broadcast channels, the algorithmic pattern the paper follows for the $p$ and $q$ updates.","marker":"[16]"}],"fun_headline_variants":["Extended AB solves bistatic ISAC rate-distortion trade-off","Linearization trick enables AB for non-convex ISAC constraints","Bistatic ISAC: AB algorithm now handles non-convex distortion","Extended AB algorithm computes ISAC rate-distortion curves","AB algorithm takes on non-convex ISAC rate-distortion"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The algorithm's alternating updates are assumed to reach the global maximum of the auxiliary objective, so that the value at the stopping point really is the capacity-distortion function; the paper proves only that the objective never goes down.","fun_headline_variants_meta":{"raw":{"variants":["Extended AB solves bistatic ISAC rate-distortion trade-off","Linearization trick enables AB for non-convex ISAC constraints","Bistatic ISAC: AB algorithm now handles non-convex distortion","Extended AB algorithm computes ISAC rate-distortion curves","AB algorithm takes on non-convex ISAC rate-distortion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000802,"raw_usage":{"total_tokens":3321,"prompt_tokens":663,"completion_tokens":2658,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":2572}},"tokens_in":407,"tokens_out":2658,"duration_ms":18251,"temperature":1.0,"reasoning_tokens":2572,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:02:39.387453+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small finite-alphabet instance, enumerate or globally optimize over all input distributions $p(u,x)$, and compare the true optimal value with the value Algorithm 1 returns from many random initializations. Any run that stops at a strictly smaller value while satisfying the distortion constraint, or two initializations that settle at different values, would falsify the claim that the algorithm calculates the rate-distortion tradeoff.","supporting_citations":[],"review_version":1}