{"id":"3ceed96f-7ac7-437c-9c7e-adf964482d49","arxiv_id":"2501.12148","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A deep-unfolded primal-dual power control algorithm reaches about 101 percent of the FPLinQ benchmark in under 10 iterations, but its convergence theorem relies on a false monotonicity lemma.","lead":"This paper trains a small neural network folded into a classic wireless power-control loop, aiming to maximize total data rate in dense device-to-device networks. The trained algorithm matches a standard benchmark with fewer iterations, but the paper's convergence proof contains a false mathematical step.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 relies on a false monotonicity lemma: log-concavity only yields an inner-product inequality, not the componentwise monotonicity needed to prove that the update in (12) is a Yates standard interference function.","rationale":"The reader's weakest assumption is exactly the load-bearing defect I find: Lemma 1 is false, and the proof of Theorem 1 depends on it to establish monotonicity of the fixed-point map. Without monotonicity, the Yates standard-interference-function argument collapses, and with it the paper's central convergence guarantee. I agree with the REJECT verdict, so no further adjustment is needed. I am not asserting that the LPDA numerical method is useless; the experiments are suggestive but incomplete, lacking error bars and code. However, the manuscript's advertised contribution is a theoretical guarantee, and that guarantee is not supported by the submitted proof. The concrete counterexample to Lemma 1 is decisive for the proof as written, while the proposed numerical search would determine whether the theorem itself might still be true under the full set of assumptions and merely needs a corrected proof.","tokens_in":9393,"tokens_out":7477,"duration_ms":79391,"concrete_test":"Analytical check: evaluate the claimed componentwise inequality in Lemma 1 for f(x,y)=e^{-(x-y)^2} at x=(3,0) and x'=(2,0). Since (3,0) \\ge (2,0) componentwise but \\nabla log f(3,0)=(-6,6) and \\nabla log f(2,0)=(-4,4), the second component rises and the lemma fails. This directly removes the only support for monotonicity of \\tilde I in (21). To distinguish 'theorem false' from 'theorem unproved', additionally run a numerical search over standard, feasible, log-concave interference functions with scale-invariant gradients, such as positive homogeneous concave examples, and test whether the map in (12) is componentwise monotone; any violation would refute Theorem 1 as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central theoretical claim is that (12) converges because its right-hand side is a standard and feasible interference function. The proof of Theorem 1 establishes monotonicity of \\tilde I in (21) through Lemma 1, which asserts that log-concavity of f implies componentwise monotonicity of \\nabla f / f. This lemma is false. Negative semidefiniteness of the Hessian of log f gives only (\\nabla log f(x) - \\nabla log f(x'))^T (x - x') \\le 0, which does not imply the componentwise inequality used in (22). A concrete counterexample is f(x,y)=e^{-(x-y)^2}, which is log-concave, with x=(3,0) \\ge x'=(2,0), yet \\nabla log f(3,0)=(-6,6) and \\nabla log f(2,0)=(-4,4); the second component violates the claimed inequality. Since monotonicity is one of the three defining properties of a Yates standard interference function, the convergence guarantee for (12) is unsupported. The feasibility part of the proof also invokes convergence of the CCCP update to establish existence of the fixed point, which is close to circular because that update is essentially the algorithm whose convergence is being proven. The numerical LPDA may still be practical, but the advertised theoretical guarantee is not established as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep-unfolding-based primal-dual algorithm (LPDA) for weighted sum rate (WSR) maximization in Gaussian interference channels. The theoretical part derives a closed-form fixed-point update from a difference-of-convex (DCA/CCCP) reformulation and claims that, under log-concavity and gradient scale-invariance of the interference function, the update is a Yates standard and feasible interference function, hence the fixed-point iteration converges. The algorithmic part introduces a learned primal-dual method with a neural-network q-update and reports numerical experiments on a D2D network model showing competitive performance with FPLinQ in about 8 iterations.","tokens_in":9693,"tokens_out":7021,"duration_ms":66495,"significance":"If the theoretical claim were valid, the paper would extend the standard interference function framework to a broader class of log-concave interference functions and provide a principled justification for a practical deep-unfolding algorithm. The numerical results are a genuine strength: the LPDA achieves roughly 101% of the FPLinQ objective on unseen networks with only 8 iterations when trained on random weights, and the experimental setup is described in enough detail to be reproduced. However, the central theoretical guarantee rests on a false lemma, so the advertised convergence result is not established. Since the theory is a major advertised contribution, the paper cannot be accepted in its current form.","major_comments":[{"comment":"Lemma 1 is false as stated. The negative semidefiniteness of the Hessian of log f implies only the inner-product inequality (∇ log f(x) − ∇ log f(x′))T(x − x′) ≤ 0 for all x, x′, which does not imply the componentwise inequality claimed in the lemma for x ≥ x′. A concrete counterexample is f(x,y) = e^{-(x-y)^2}, which is log-concave. For x = (3,0) and x′ = (2,0), ∇ log f(x) = (−6,6) and ∇ log f(x′) = (−4,4), so the second component violates the claimed inequality (6 ≤ 4 is false). This lemma is the basis for the monotonicity proof of the candidate interference function in (21) via equations (22)–(25). Without monotonicity, the function in (21) is not a standard interference function, and the convergence guarantee from Theorem 2 of [28] cannot be invoked.","section":"Appendix, Lemma 1"},{"comment":"The feasibility part of the proof is circular. It uses convergence of the CCCP/DCA update (11) to assert the existence of a fixed point p̃ ≥ Ĩ(p̃), but (12) is exactly the closed-form solution of (11), so this argument assumes the convergence that Theorem 1 is meant to establish. The manuscript states in Section III-A that iteration (11) 'is guaranteed to converge to a stable point of (10)' if I_i is log-concave, but no proof or appropriate reference is supplied; the cited Theorem 2 of [30] concerns CCCP convergence under conditions that are not verified here. Thus the feasibility of the interference function in (21) is not independently established.","section":"Appendix, proof of Theorem 1, Feasibility"},{"comment":"The paper applies Theorem 1 to the primal-dual update (19), but the proof in the appendix only treats the function (21). Equation (19) contains an additional term −q_i γ_i(q)^{-1} and a denominator perturbation λ_i, and the statement that 'the right hand side is again a standard and feasible interference function' if λ and q are constant is asserted without proof. Furthermore, the right-hand side of (19) is evaluated at the outer iterate p^{(k)} and q, not as a function of the current variable p in the manner of (21), so the fixed-point interpretation used in Algorithm 1 is not the same as the one analyzed in Theorem 1. A separate, precise proof of the fixed-point property of (19) is needed.","section":"Section III-B, Eq. (19) and Algorithm 1"}],"minor_comments":[{"comment":"There is a typo: 'opitmization' should be 'optimization'.","section":"Section I"},{"comment":"There is a typo: 'follwing' should be 'following'. The displayed equivalence arrows in (22)–(25) are not true bi-implications as typeset and should be replaced by directional implications or inequalities.","section":"Appendix, proof of Lemma 1"},{"comment":"The scale-invariance condition is imprecise: the phrase 'scale invariant as a function in q' is ambiguous. The proof uses the explicit condition ∇p I_i(αq) = ∇p I_i(q) for all α > 0 and q, which should be stated in the theorem.","section":"Theorem 1 statement"},{"comment":"Line 3 of Algorithm 1 is difficult to parse; the formula for p^{(k+1)}_i, including the term −q_i^{(k)} γ(q^{(k)})^{-1}_i and the min with Pmax, should be written more clearly and connected to equation (19).","section":"Algorithm 1"},{"comment":"The metric (20) is relative to FPLinQ after 100 iterations; values above 100% indicate the LPDA's output is on average slightly better than the benchmark on the test set, but the text says 'same performance.' This should be clarified to avoid overstatement.","section":"Section V, Table II"},{"comment":"The training section does not specify the number of training epochs, batch size, or the schedule of the decreasing learning rate; adding these details would improve reproducibility.","section":"Section IV-B"}],"recommendation":"reject","confidential_remarks":"The central theoretical claim is invalid because Lemma 1 is false, and the feasibility proof is circular. The numerical results are promising but are likely insufficient to carry the paper on their own, especially because the paper's stated contribution is the theoretical guarantee under log-concavity. The authors would need to either replace the false lemma with a correct proof under additional assumptions or substantially reframe the paper as an empirical study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the practical core: the deep-unfolded primal-dual algorithm (LPDA) is a legitimate new method. The idea of replacing the q-update in the inner loop with a small FCNN and training end-to-end on the WSR objective is clean, and the experiments, though lightweight, show it reaches FPLinQ's performance in 8 iterations on 500 unseen networks. That part is worth someone's attention.\n\nThe problem is the theory. Theorem 1 claims convergence because the right-hand side of (12) is a standard interference function. Monotonicity is established via Lemma 1 in the appendix, and that lemma is false. Log-concavity gives you the inner-product inequality only; it does not give componentwise monotonicity of the log-gradient. Your counterexample is right: f(x,y)=e^{-(x-y)^2} is log-concave, but at (3,0) and (2,0) the second component of the log-gradient goes from 4 to 6, which violates the claimed direction. So the monotonicity proof for \\tilde I does not hold, and with it the Yates-based convergence argument falls.\n\nThe feasibility step is also circular. It invokes the convergence of the DCA/CCCP update (11) to assert that a fixed point exists, but that update is exactly the one in (12). Using the algorithm's own convergence to prove the existence of a fixed point for the same iteration, before establishing convergence, is reasoning in a circle.\n\nWhat is not circular is the affine case. There the algorithm reduces to the known Tan-Chiang-Srikant fixed-point iteration, and the paper says so. That part is fine.\n\nThe numerical section is thin: no error bars, no code, and only one system model. The performance metric in (20) is reasonable, but 101% and 97.9% without variance tells you little about robustness.\n\nNet: the method might be repairable. If someone fixes the monotonicity argument, maybe by proving the needed componentwise property under the specific structure of the interference functions used in (9) and (13), the paper could become a solid applied contribution. As written, the central guarantee is unsupported. I'd send it to review because the method is plausible and the flaw is isolated rather than the whole enterprise being incoherent, but I'd expect a major revision or rejection. I would not cite the current version.","headline":"False monotonicity lemma sinks the convergence proof, but the deep-unfolded method and experiments may be worth salvaging.","tokens_in":10206,"tokens_out":2662,"would_cite":false,"duration_ms":25533,"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":"Under log-concavity, a closed-form power update provably converges, and a deep-unfolded version matches FPLinQ in eight iterations.","keywords":["weighted sum rate","interference function","log-concavity","fixed-point algorithm","deep unfolding","primal-dual algorithm","power control","difference-of-convex programming"],"falsifier":"Check Lemma 1 on $f(x,y)=e^{-(x-y)^2}$, which is log-concave. At $p=(3,2)$ and $p'=(1,1)$, $p \\geq p'$ componentwise, but the second entry of $\\nabla \\log f$ is $2$ at $p$ and $0$ at $p'$, so the componentwise monotonicity asserted in Lemma 1 fails; this would invalidate the monotonicity proof of update (12).","tokens_in":9178,"feed_emoji":"📡","tokens_out":5061,"duration_ms":48744,"temperature":0.7,"pith_summary":"The paper attacks the non-convex weighted sum rate (WSR) maximization problem in Gaussian interference channels. It claims that when the interference function is log-concave and its gradient is scale invariant, a closed-form fixed-point update rule derived from difference-of-convex programming is a standard feasible interference function in Yates' sense, so the iteration provably converges. Building on this, the authors propose a primal-dual algorithm and then a deep-unfolded version (LPDA) that runs only eight iterations. Numerical experiments on device-to-device networks suggest LPDA matches the FPLinQ benchmark while converging faster. If the theoretical claim holds, it gives a broader class of interference models, beyond the usual affine linear one, for which WSR maximization has a provably convergent fixed-point algorithm.","feed_headline":"Power update provably converges for log-concave interference","feed_subtitle":"A deep-unfolded primal-dual algorithm matches the FPLinQ benchmark with only eight iterations.","key_machinery":"The load-bearing object is the closed-form update (12), obtained by solving the first-order optimality conditions of the difference-of-convex subproblem. The proof checks that this update satisfies the three defining properties of a Yates standard interference function (positivity, scalability, monotonicity) under log-concavity and scale invariance of $\\nabla I$. Monotonicity is established via the paper's Lemma 1, which asserts that log-concavity forces the log-gradient map to be componentwise monotone; scalability uses scale invariance of the gradient and the standard property $\\alpha I(p) > I(\\alpha p)$. The deep-unfolded LPDA replaces the inner primal-dual $q$-update with a fully connected neural network, leaving the $p$-update as the provably convergent interference-function iteration.","core_discovery":"The central claim is Theorem 1: for any standard, feasible interference function whose components are log-concave and whose gradients are scale invariant, the update rule (12), $p_i = \\min\\left\\{ w_i \\left[ \\sum_j w_j \\frac{\\nabla_p I_j(p)_i}{I_j(p)} \\right]^{-1}, P_{\\max} \\right\\}$, is itself a standard and feasible interference function. Hence the fixed-point iteration $p^{(k+1)} = \\text{update}(p^{(k)})$ converges to a fixed point that corresponds to a stationary point of the approximating WSR problem. This extends a known fixed-point algorithm for affine linear interference to arbitrary log-concave interference functions.","pith_inferences":["The fixed-point framework could extend to other nonconvex problems expressible as a difference of log-concave functions, not just WSR.","End-to-end training with the WSR objective as the loss suggests a general recipe: unfold any provably convergent interference-function iteration and train only the auxiliary updates.","The reported performance above 100% on random weights hints that LPDA exploits the training distribution, so testing on truly unseen weight distributions is a natural next step."],"forward_implications":["For affine linear interference, the algorithm reduces to a known fixed-point power control method, now subsumed as a special case.","For a wider class of log-concave interference models, including the concave example in (13), WSR maximization gets a provably convergent fixed-point algorithm.","The deep-unfolded LPDA achieves benchmark-level WSR with only 8 iterations, making it suitable for fast, scalable resource allocation.","Because the $p$-update is an interference function, standard convergence and feasibility results for power control apply directly to the learned algorithm's architecture."],"supporting_citations":[{"why":"Supplies the definition of standard interference function and the fixed-point convergence theorems that Theorem 1 relies on.","marker":"[28]"},{"why":"Provides the CCCP convergence result used to argue the DCA iteration reaches a stationary point of the approximating problem, supporting feasibility.","marker":"[30]"},{"why":"Introduces the FPLinQ benchmark algorithm used as the numerical comparison target.","marker":"[17]"},{"why":"Develops the fractional programming framework underlying FPLinQ, the state-of-the-art benchmark.","marker":"[18]"},{"why":"Derives the same fixed-point update for the affine linear interference special case and proves its convergence via interference functions.","marker":"[3]"},{"why":"Introduces the deep unfolding technique that motivates the LPDA architecture.","marker":"[6]"}],"fun_headline_variants":["Deep-unfolded PDA matches FPLinQ in 8 iterations","Log-concave interference yields fixed-point convergence","Deep unfolding reduces WSR algorithm complexity","Non-convex WSR solved via deep-unfolded fixed-point","Provable convergence for log-concave interference"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the claim that log-concavity forces each ratio $\\nabla_p I_j(p)_i / I_j(p)$ to decrease when any entry of $p$ increases; if that specific step fails, the standard-interference-function guarantee is not established.","fun_headline_variants_meta":{"raw":{"variants":["Deep-unfolded PDA matches FPLinQ in 8 iterations","Log-concave interference yields fixed-point convergence","Deep unfolding reduces WSR algorithm complexity","Non-convex WSR solved via deep-unfolded fixed-point","Provable convergence for log-concave interference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001134,"raw_usage":{"total_tokens":4643,"prompt_tokens":810,"completion_tokens":3833,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":426,"completion_tokens_details":{"reasoning_tokens":3755}},"tokens_in":426,"tokens_out":3833,"duration_ms":29355,"temperature":1.0,"reasoning_tokens":3755,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:28:02.015080+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check Lemma 1 on $f(x,y)=e^{-(x-y)^2}$, which is log-concave. At $p=(3,2)$ and $p'=(1,1)$, $p \\geq p'$ componentwise, but the second entry of $\\nabla \\log f$ is $2$ at $p$ and $0$ at $p'$, so the componentwise monotonicity asserted in Lemma 1 fails; this would invalidate the monotonicity proof of update (12).","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the definition of standard interference function and the fixed-point convergence theorems that Theorem 1 relies on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CCCP convergence result used to argue the DCA iteration reaches a stationary point of the approximating problem, supporting feasibility."},{"cited_title":"Shen and W","cited_arxiv_id":null,"evidence_quote":"Introduces the FPLinQ benchmark algorithm used as the numerical comparison target."},{"cited_title":"Shen and W","cited_arxiv_id":null,"evidence_quote":"Develops the fractional programming framework underlying FPLinQ, the state-of-the-art benchmark."},{"cited_title":"Dahrouj, W","cited_arxiv_id":null,"evidence_quote":"Derives the same fixed-point update for the affine linear interference special case and proves its convergence via interference functions."},{"cited_title":"Gregor and Y","cited_arxiv_id":null,"evidence_quote":"Introduces the deep unfolding technique that motivates the LPDA architecture."}],"review_version":1}