{"id":"a6241634-f8c2-453c-989a-d3e40ef569bb","arxiv_id":"2501.04408","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Distributed image semantic communication training can be scheduled by an alternating KKT and fractional-programming algorithm that minimizes weighted time and energy while keeping a fitted PSNR above a threshold.","lead":"The paper found a way to allocate bandwidth, power, computing speed, and image compression during the training of image-understanding wireless models so that phones and a base station use the least combination of time and battery. A smart generalist should read it because training such models on phones is slow and power-hungry, and this is a concrete step toward making it practical.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 3's 'optimal sol*' claim is unsupported: Theorem 1 states only a necessary condition for P4's global optimum, while the algorithm uses the converse; the cited sum-of-ratios theorem's hypotheses are not verified (denominator r_n is concave, not convex).","rationale":"The reader's strongest_claim is the global optimality of Algorithm 3. I examined the subproblem where that claim is least secure. P3 is convex and its KKT solution is fine modulo minor edge cases. P4 is a sum-of-pseudoconvex-ratios problem; global optimality there is the load-bearing step, because Algorithm 3's output and Section VI-B both depend on it. The paper's Theorem 1 gives only a necessary condition (global optimum implies existence of (gamma*, delta*)), not the sufficiency direction the algorithm employs. The attached proof 'given by Lemma 2.1 in [53]' is not reproduced, and the hypotheses of that lemma are not checked against P4. The mismatch in Lemma 1, namely denominator concave versus the cited convex-denominator condition, reinforces that the wrong theorem class may have been invoked. This is a proof gap, not an empirical observation, so it can be settled by reading [53] and by a grid-search counterexample. The PSNR surrogate concern raised by the reader is real and should remain a condition, but I judge the unproven global-optimality step to be more central to the strongest claim. Hence I keep the CONDITIONAL verdict: acceptance requires either a correct equivalence theorem with verified hypotheses or a softened claim such as 'near-optimal heuristic', plus out-of-sample validation of Eq. (58).","tokens_in":23863,"tokens_out":16725,"duration_ms":180962,"concrete_test":"Two-part check. (1) Read Lemma 2.1 of [53] and verify whether it supplies the missing converse for P4 with concave denominator r_n; if it does not, the proof is incomplete. (2) Numerically test global optimality: take a small feasible instance (e.g., N=2, discretize p_n and B_n on a fine grid satisfying (16a), (16c), (33a), (33b)), compute the true global minimum of P4 by exhaustive enumeration, and run Algorithm 2 from multiple initial points. If any run returns a point with objective above the grid minimum, the 'optimal sol*' claim for P4 is false; if all runs match over many random instances, the concern is weakened but still needs the missing proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"P4/P6 is the block whose global optimality is used to justify 'optimal sol*' in Algorithm 3 and the Section VI-B statement that 'global optimality of the subproblems can be secured.' The proof has two gaps. First, Theorem 1 as stated is one-way: a global optimum of P6 implies existence of (gamma*, delta*) satisfying (35) and solving P7. Algorithm 2 searches for (gamma, delta) with phi = 0 and then treats the P7 solution as the global optimum of P4; this converse is not asserted, let alone proved, in Theorem 1. It is imported from Lemma 2.1 of [53] without stating or checking that lemma's hypotheses. Second, Lemma 1 proves r_n is jointly concave and then invokes a cited condition for pseudoconvexity of a ratio that requires a positive convex denominator. The denominator r_n is concave, so the cited condition does not apply as written; whether the ratio is nevertheless of the class covered by [53] is simply left unverified. If the P4 global-optimality step fails, Algorithm 3 returns at best a locally optimized allocation, and the headline claim 'optimal sol* = (p*, B*, f*, h*, rho*)' is not established. This is independent of the empirical PSNR-surrogate concern: even granting Eq. (58), the optimization certificate is missing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper considers a distributed image semantic communication system in which local devices encode and transmit image features to a base station, which decodes, computes the loss, and updates the model. The authors formulate a joint resource allocation problem P1 that minimizes a weighted sum of the maximum per-device training completion time and the total energy consumption, subject to per-device constraints on transmit power, bandwidth, CPU frequencies, compression rate, and a minimum PSNR requirement. P1 is decomposed into P3 over (f,h,rho,T) with (p,B) fixed and P4 over (p,B) with the other variables fixed; P3 is solved by KKT conditions after convexity verification, while P4 is attacked as a sum-of-ratios problem using a parameter-refinement method attributed to Jong [53]. Algorithm 3 alternates between the two subproblems and is claimed to return the optimal solution sol* = (p*,B*,f*,h*,rho*). The experiments use a CIFAR-10 deep JSCC model, compare Algorithm 3 against four baselines under varying bandwidth, power, frequency, weights, and number of users, and also compare reconstructed image quality against JPEG.","tokens_in":24200,"tokens_out":9230,"duration_ms":91193,"significance":"If the results hold, the paper contributes a useful formulation that jointly controls latency, energy, and semantic quality during the training phase of an image semantic communication network, rather than only at inference. The decomposition into P3/P4 is natural, the Hessian calculation for r_n in Lemma 1 is explicit, and the experimental sweep over system parameters is extensive. However, the central claims are not fully established: the proof of Lemma 1 contradicts itself on the convexity of the denominator, Theorem 1 supplies only a necessary condition while Algorithm 2 uses the converse, and the PSNR feasibility constraint is enforced through an in-sample fitted curve with no uncertainty quantification. These are load-bearing gaps, so the significance of the contribution cannot be fully assessed until they are closed.","major_comments":[{"comment":"The proof of Lemma 1 first establishes that r_n(p_n,B_n) is jointly concave, but then invokes a pseudoconvexity criterion from [52] that requires the denominator to be positive, convex, and differentiable. A concave denominator does not satisfy that condition. Thus the pseudoconvexity of p_n rho_n d_n / r_n, which underpins the fractional-programming treatment of P4, is not proved as written.","section":"V-B, Lemma 1"},{"comment":"Theorem 1 states only that a global optimum of P6 implies the existence of (gamma*,delta*) satisfying (35) and solving P7. Algorithm 2 searches for (delta,gamma) with phi = 0 and then declares the resulting P7 solution to be the global optimum of P4. That converse is not proved in the manuscript; it is imported from Lemma 2.1 and Theorem 3.1 of [53] without stating or verifying their hypotheses. In particular, because r_n is concave rather than convex, it has not been shown that the sum-of-ratios framework of [53] applies. Consequently, Algorithm 3's output 'optimal sol*' and the Section VI-B statement that 'global optimality of the subproblems can be secured' are unsupported.","section":"V-B, Theorem 1 and Algorithm 2"},{"comment":"The PSNR constraint (16e) is enforced through the fitted model P(x) = 18.67 ln(3.35x + 5.11) with x = 1.52 rho + 0.03 S. This curve is fitted to the same CIFAR-10 deep JSCC model used in the subsequent simulations, and only a 5-fold average MAE of 0.4381 dB is reported, with no confidence intervals, no out-of-sample validation, and no sensitivity analysis. If the curve deviates from the true PSNR of the trained model, the allocations returned by Algorithm 3 do not actually guarantee P_min, so the feasibility claim is circular with respect to the model whose training is being optimized.","section":"III-C and VII-B, Eq. (58)"},{"comment":"The convergence proof for Algorithm 2 states that the function phi(v) is linear and therefore satisfies the hypotheses of Theorem 3.2 in [53]. However, phi in (38) is evaluated at (p^(i+1),B^(i+1)), which are themselves functions of (delta^(i),gamma^(i)) obtained by solving P7. The implicit dependence is not accounted for in the manuscript, so the linearity claim and the resulting invocation of the convergence theorem are not established.","section":"VI-C"}],"minor_comments":[{"comment":"The word 'Interploted' should be 'Interpolated', and the same typo appears in the caption of Fig. 2.","section":"VII-B, Fig. 2 caption"},{"comment":"The word 'Comparision' should be 'Comparison'.","section":"VII-G, Fig. 8 caption"},{"comment":"In the complexity analysis, the text says 'step 3 incurs a complexity of O(4N)' when referring to solving P3; this should be step 4, and the subsequent numbering of steps in Algorithm 2 should be aligned with the actual pseudocode.","section":"VI-A"},{"comment":"The definition 'rmin_n = T^* - T^{cmp}_n - T^{S}_n / rho_n d_n' is missing parentheses; it should be rmin_n = (T^* - T^{cmp}_n - T^{S}_n)/(rho_n d_n).","section":"V-B"},{"comment":"The sentence 'We defer a more detailed introduction of the model to Section III-A' should refer to Section VII-A, since the detailed model introduction appears there.","section":"III-A"},{"comment":"The baseline named 'Optimize f, h, rho only' is later referred to as 'Optimize f, rho only' in the discussion of Fig. 5; the names should be made consistent.","section":"VII-D"},{"comment":"The comment inside Algorithm 2 says the proof is given by Theorem 3.1 in [53], while Section VI-C cites Theorem 3.2 for the same convergence claim; this should be reconciled.","section":"VI-C and Algorithm 2"},{"comment":"Condition 1 writes P''(x) < 0 for a function of two variables; the notation should be clarified as a shorthand for concavity in each argument or expressed via the Hessian.","section":"III-C, Condition 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on [53] for the global-optimality and convergence certificates, but the hypotheses are never checked and one required condition (convex denominator) is contradicted by the paper's own Lemma 1. The PSNR surrogate is fitted in-sample on the same model used for evaluation, so the feasibility guarantee needs validation beyond a single average MAE. Both issues should be addressed before the claims of optimal resource allocation and guaranteed PSNR are accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has a genuinely new formulation — training-phase resource allocation for image semantic communication with a PSNR floor — and the decomposition is sensible. But the central optimality claim doesn't hold as written, and the PSNR constraint is validated on the same model it is then used to evaluate. Worth a serious referee, but it needs major revision.\n\nThe new thing is the problem itself. Prior work on SemCom resource allocation almost all assumes a trained model and optimizes inference. Here the authors minimize a weighted sum of latency and energy during collaborative training, over bandwidth, transmit power, local CPU frequency, BS frequency, and compression rate, subject to a PSNR floor. I checked the cited literature and the combination is not there. The solution strategy — split into a convex P3 and a sum-of-ratios P4, solve P3 via KKT and P4 via a fractional-programming refinement — is standard but applied cleanly. The KKT derivations in Section V-A are traceable and correct.\n\nThe soft spots are real. First, the claim that Algorithm 3 returns the global optimum of P1 is not supported. Theorem 1 only states a necessary condition: a global optimum of P6 implies existence of (γ*,δ*) satisfying (35) and solving P7. Algorithm 2 uses the converse — if φ=0 then the P7 solution is globally optimal — but that converse is neither stated nor proved, and the hypotheses of Lemma 2.1 in [53] are not checked. More concretely, Lemma 1 proves r_n is jointly concave and then cites a ratio-pseudoconvexity condition that requires the denominator to be positive and convex. That condition does not apply. So the global-optimality certificate for P4 is missing, and without it Algorithm 3 is at best a locally convergent heuristic. This is not a nitpick; the abstract and Section VI explicitly promise global optimality.\n\nSecond, the PSNR floor is enforced through a fitted function (58) trained on the same CIFAR-10 deep JSCC model used in the simulations. The fit reports only in-sample MAE, with no held-out validation. If the curve is off, allocations that satisfy the fitted constraint may violate the real PSNR floor. This is fixable — report cross-validated error, test on another dataset, or treat the fit as an uncertainty set.\n\nThird, the experiments compare against simple baselines and report averages over 100 runs but no error bars or actual training iterations. The 'performance' is the fitted PSNR, so the loop is partly circular.\n\nWho should read it: people working on wireless edge training or SemCom resource allocation. The formulation is a useful contribution and the proof gaps are addressable. I'd send it to review, but expect major revision: fix or retract the global optimality claim, validate the PSNR surrogate, add training-level experiments.","headline":"Genuinely new training-phase resource allocation formulation for image SemCom, but the global optimality proof has a load-bearing gap and the PSNR surrogate needs validation; worth a serious referee but not acceptable as is.","tokens_in":24696,"tokens_out":7684,"would_cite":true,"duration_ms":67841,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that jointly allocating transmit power, bandwidth, CPU frequencies, and semantic compression rate minimizes the weighted time-and-energy cost of distributed image-semantic-communication training while enforcing a PSNR…","keywords":["semantic communication","resource allocation","image transmission","deep JSCC","distributed training","convex optimization","fractional programming","PSNR constraint"],"falsifier":"Run the same deep JSCC model on held-out or higher-resolution images at a grid of (rho,SNR) values covering the ranges in Table II, measure the true PSNR, and check whether any point lies more than a small tolerance below the fitted surface; a single violation inside the feasible region would mean constraint (16e) is not certified. A sharper test: solve P1 with the true measured PSNR surface and compare the resulting weighted cost and allocation to Algorithm 3's output, since a material gap would refute the optimality claim.","tokens_in":23635,"feed_emoji":"⚡","tokens_out":8921,"duration_ms":80171,"temperature":0.7,"pith_summary":"This paper is trying to establish that the overhead of training an image semantic communication network can be minimized by jointly choosing transmission power, bandwidth, per-device CPU frequency, and the semantic compression rate, subject to a minimum reconstruction quality. The authors model a base station and mobile devices that collaboratively train a deep JSCC image codec, then formulate the joint problem as minimizing the weighted sum of the maximum per-device training time and the total energy consumption. They solve it by splitting the problem into a convex subproblem (frequencies, compression rate, and makespan) solved by KKT conditions, and a sum-of-ratios subproblem (power and bandwidth) solved by a parametric fractional-programming method, with the two alternating until convergence. The paper reports that this algorithm beats four baseline allocation strategies under varied bandwidth, power, frequency, and user counts. If the claim is right, a network operator could run near-optimal training resource allocation while still meeting the required image quality.","feed_headline":"Joint power-bandwidth plan cuts training cost for semantic image nets","feed_subtitle":"A 50-device system meets PSNR targets while minimizing weighted time and energy, beating four allocation baselines.","key_machinery":"The machinery that carries the argument is the decomposition of the joint problem into two alternately solved subproblems, together with a parametric identity that linearizes the sum-of-ratios cost. For the time/energy subproblem, the PSNR constraint is inverted into a lower bound on the compression rate, which leaves a convex problem solvable by KKT conditions. For the power/bandwidth subproblem, the paper introduces auxiliary variables gamma and delta so that each transmission-energy ratio p_n rho_n d_n D_n / r_n(p_n,B_n) is represented by delta_n, and the objective becomes gamma_n times (p_n rho_n d_n D_n - delta_n r_n(p_n,B_n)); the theorem borrowed from the fractional-programming literature states that solving this converted problem with the correct gamma and delta solves the original sum-of-ratios problem. The fitted PSNR function P(rho, S) = 18.67 ln(3.35(1.52 rho + 0.03 S) + 5.11) is what ties model quality to the resource variables and closes the loop.","core_discovery":"On the paper's own terms, the central discovery is that the joint problem P1 admits an efficient alternating solution with global optimality of each subproblem. After introducing the auxiliary variable T for the maximum training time, the authors decompose P1 into P3, which optimizes device frequency, BS frequency, compression rate, and T, and P4, which optimizes transmission power and bandwidth. P3 becomes the convex problem P5 once the compression rate is set to the smallest value satisfying the PSNR constraint, so the KKT conditions yield the global f*, h*, and T*. P4 is a sum of pseudoconvex ratios; the paper applies a fractional-programming theorem to convert it into P7, whose solution shares the global optimum with P4 when the auxiliary parameters satisfy gamma_n = omega2 / r_n and delta_n = p_n rho_n d_n D_n / r_n. Algorithm 3 alternates between these two solvers and returns sol* = (p*, B*, f*, h*, rho*), which the paper claims is the optimal solution of P1. The PSNR floor is enforced through the fitted concave surrogate P(x) = 18.67 ln(3.35x + 5.11) with x = 1.52 rho + 0.03S, obtained from CIFAR-10 evaluations of the deep JSCC model.","pith_inferences":["Editorial inference: since the PSNR surrogate is fitted once on CIFAR-10, any deployment on different data or channel statistics should re-fit the curve or replace it with a conservative lower bound, otherwise the feasibility guarantee in constraint (16e) is only as accurate as that fit.","Editorial inference: the same decomposition should carry over to other image quality metrics such as SSIM only if a concave, non-decreasing surrogate in (rho,S) can be validated; without that structure, the KKT step for the compression rate no longer holds directly.","Editorial inference: the centralized O(N) solver is a natural candidate to be re-run periodically as channel gains drift; a distributed or learning-based policy could trade a small optimality gap for better scalability in very dense networks, a direction the paper flags for future work.","Editorial inference: a prediction interval around the fitted PSNR curve, rather than a single mean curve, would turn the weakest assumption into a testable certificate that the system truly meets P_min."],"forward_implications":["If Algorithm 3 behaves as claimed, a fixed pair (p,B) yields globally optimal device frequencies, BS frequencies, compression rates, and makespan from the convex subproblem.","A fixed set (f,h,rho,T) yields globally optimal power and bandwidth from the sum-of-ratios solver, so neither subproblem leaves resources on the table.","The alternating procedure converges to an epsilon-accurate solution with complexity O(K(Ij+9I+4)N), making online re-allocation feasible as channels change.","Raising total bandwidth or maximum device frequency reduces training time, while raising the PSNR floor increases both time and energy and improves the achieved PSNR.","The trained semantic model reaches 26.95 dB at SNR=13 dB and rho=0.1, exceeding JPEG's 26.13 dB, so the quality constraint is satisfiable while resource costs are being minimized."],"supporting_citations":[{"why":"Supplies the deep JSCC encoder-decoder architecture whose PSNR is the performance metric being constrained.","marker":"[38]"},{"why":"Provides the fractional-programming equivalence and the modified Newton convergence proof that Algorithm 2 uses to solve the sum-of-ratios subproblem globally.","marker":"[53]"},{"why":"Supplies the KKT sufficiency/necessity and Slater regularity used to solve the convex subproblem P5.","marker":"[56]"},{"why":"Provides the FDMA computation/transmission models and the default simulation parameters (bandwidth, noise, capacitance) used in the experiments.","marker":"[42]"},{"why":"Establishes the pseudo-convexity of the concave-over-convex ratio p_n rho_n d_n / r_n used in Lemma 1.","marker":"[52]"},{"why":"Is the logarithmic-model source the fitted PSNR curve P(x)=a ln(c_rho rho + c_S S + b) adapts.","marker":"[60]"},{"why":"Grounds the shared-database iterative training procedure that defines the time and energy terms being optimized.","marker":"[16]"}],"fun_headline_variants":["Global-optimal subproblems slash training time and energy in semantic image nets","Alternating solver gives optimal power-bandwidth mix for fast semantic training","Joint frequency-power-bandwidth plan cuts time-energy for semantic nets","Global-optimal subproblem mix speeds semantic training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fitted curve P(x)=18.67 ln(3.35x+5.11), with x=1.52 rho+0.03S, accurately represents the true PSNR of the trained deep JSCC model; if the real model's PSNR dips below this curve anywhere in the operating range, the allocations returned by Algorithm 3 will not actually satisfy the quality requirement.","fun_headline_variants_meta":{"raw":{"variants":["Global-optimal subproblems slash training time and energy in semantic image nets","Alternating solver gives optimal power-bandwidth mix for fast semantic training","Joint frequency-power-bandwidth plan cuts time-energy for semantic nets","Global-optimal subproblem mix speeds semantic training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001023,"raw_usage":{"total_tokens":4331,"prompt_tokens":976,"completion_tokens":3355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":3284}},"tokens_in":592,"tokens_out":3355,"duration_ms":22768,"temperature":1.0,"reasoning_tokens":3284,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:34:06.813766+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same deep JSCC model on held-out or higher-resolution images at a grid of (rho,SNR) values covering the ranges in Table II, measure the true PSNR, and check whether any point lies more than a small tolerance below the fitted surface; a single violation inside the feasible region would mean constraint (16e) is not certified. A sharper test: solve P1 with the true measured PSNR surface and compare the resulting weighted cost and allocation to Algorithm 3's output, since a material gap would refute the optimality claim.","supporting_citations":[{"cited_title":"An efficient global optimization algorithm for nonlinear sum- of-ratios problem,","cited_arxiv_id":null,"evidence_quote":"Provides the fractional-programming equivalence and the modified Newton convergence proof that Algorithm 2 uses to solve the sum-of-ratios subproblem globally."},{"cited_title":"Cambini and L","cited_arxiv_id":null,"evidence_quote":"Establishes the pseudo-convexity of the concave-over-convex ratio p_n rho_n d_n / r_n used in Lemma 1."},{"cited_title":"Optimizing utility-energy efficiency for the metaverse over wireless networks under physical layer security,","cited_arxiv_id":null,"evidence_quote":"Is the logarithmic-model source the fitted PSNR curve P(x)=a ln(c_rho rho + c_S S + b) adapts."}],"review_version":1}