{"id":"fb57d5ef-b725-42d0-b49c-18610fb78c0e","arxiv_id":"2506.03508","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper introduces MDDRA, a model-data dual-driven scheme combining Lyapunov queues and a GRAF network for energy-efficient 6G resource allocation, with claimed IREE gains of 10.2-20.7% in simulations.","lead":"The paper proposes a hybrid resource allocation algorithm that pairs a Lyapunov queue with a graph neural network to predict traffic and allocate bandwidth and power in 6G networks, aiming to cut energy waste from capacity-demand mismatch. It reports 10 to 20 percent gains in a new energy efficiency metric, but the mathematical guarantees rely on unproven convergence steps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's convergence proof is unsupported: Appendix B assumes, rather than proves, that correction factors ν∈[0,2] can force every virtual queue residual to zero, and its drift argument (ΔQ≤0 ⇒ Q→0) is logically invalid. The claimed lim ηT_IREE=ηT,*_IREE therefore lacks a sound basis.","rationale":"The reader's weakest assumption and my independent reading converge on the same point: the proof of Lemma 1 and Theorem 2 depends on an unproven feasibility of correction factors ν∈[0,2] and on an invalid inference from non-positive Lyapunov drift to queue convergence. This is load-bearing because the entire convergence-to-optimal claim is built on those queues being forced to zero. Appendix B does not show that (19)–(21) can be satisfied simultaneously; it merely postulates the ν that would satisfy them and then draws a convergence conclusion that does not follow from ΔQ≤0. The recurrence Q_{τ+1}=(1−ν)Qτ can be non-contracting (ν=0) or merely sign-flipping (ν=2), so an additional contraction argument is required and absent. The draft's simulations and the real-map extension in Section V-C are useful empirical evidence, and the GRAF universal approximation result has independent plausibility, but they do not repair a mathematical proof whose central mechanism is assumed rather than established. Since the reader already recommended REJECT with high correctness risk, my stress-test confirms that verdict rather than moving it. A single feasibility counterexample or the trivial Q=1, ν=0 drift counterexample would settle the specific concern; if neither lands, the rejection would need to rest on the other issues (reproducibility, circularity of Theorem 1) that I did not need to weigh.","tokens_in":36078,"tokens_out":3480,"duration_ms":40297,"concrete_test":"Run a one-time-slot feasibility check of Problem 4. Fix Bn,τ, P a_{n,τ}, queues Qζ, Qκ, Qη, and ηT_IREE to representative values (including small, large, positive and negative queue realizations), and solve the feasibility problem over κτ, hτ, νζ, νκ, νη subject to (19)–(22) with ν∈[0,2] and hτ∈[0,1]. If even a single tuple with nonzero queues is infeasible, the correction-factor mechanism fails. Separately, verify the drift inference directly: set Q0=1 and choose νκ_τ=0 for all τ; then ΔQ=0 for every step yet Qτ≡1, demonstrating that the implication in Appendix B ('ΔQ≤0 and Q²≥0 ⇒ lim Q=0') is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 2: lim_{k,τ→∞} ηT_IREE = ηT,*_IREE. The proof chain passes through Lemma 1 and Appendix B, which is the load-bearing step.\n\nAppendix B defines Qκ_τ = Σ_{t=0}^{τ-1}(ξ(t)−κt/T) and asserts existence of νκ_τ∈[0,2] satisfying ξ(τ)−κτ/T = −νκ_τ Qκ_τ (Eq. 30). This is an assumption of exact feasibility, not a derived property. For given Qκ_τ, the required value is νκ_τ = (κτ/T − ξ(τ))/Qκ_τ; nothing in the constraints or traffic model guarantees this lies in [0,2], especially when Qκ_τ is small or has the wrong sign. The same issue appears in (19) and (21): νζ and νη must simultaneously satisfy three nonlinear equalities with variables κτ, hτ and resource allocations, and no existence proof is given.\n\nEven if such ν existed, the Lyapunov argument is invalid. From Q_{τ+1} = Qτ + ξ(τ)−κτ/T = (1−νκ_τ)Qτ, the drift is ΔQ = [(1−ν)²−1](Qτ)² ≤ 0. The paper concludes 'Since (Qκ_τ)² ≥ 0, lim_{τ→∞} Qκ_τ = 0.' Non-positive drift does not imply convergence to zero: with ν=0 every step, ΔQ=0 but Q never converges; with ν=2, Q merely alternates sign. A contraction rate bounded away from 1, or an independent residual-convergence argument, would be needed and is absent.\n\nAppendix D inherits this gap. Its Part A assumes approximate stationarity and Adam convergence but never proves that constraints (19)–(21) are feasible with ν∈[0,2] at each time step. Part B then uses the unproven queue convergence to claim optimality of ηT_IREE. Moreover, (21) contains the target ηT_IREE itself, making the constraint circular unless a fixed-point argument is supplied.\n\nThus the proof's weakest link is not a minor technicality: the mechanism claimed to enforce long-term constraints is an existence assumption in disguise. Without a feasibility theorem for the correction factors and a valid queue-contraction argument, Theorem 2 does not follow from the material presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a model-data dual-driven resource allocation (MDDRA) scheme for maximizing integrated relative energy efficiency (IREE) in 6G wireless networks under dynamic, partially observable traffic. The method combines a model-driven Lyapunov queue for long-term correction of cumulative mismatch errors with a data-driven Graph Radial Basis Fourier (GRAF) network for short-term traffic prediction under incomplete spatial data. The central theoretical claims are Theorem 1 (universal approximation property of GRAF) and Theorem 2 (convergence of MDDRA to the optimal IREE, i.e., lim_{k,τ→∞} ηT_IREE = ηT,*_IREE). Numerical experiments compare MDDRA against several baselines and report IREE gains of 10.2%–20.7%, as well as robustness to traffic incompleteness and real-world map data.","tokens_in":36667,"tokens_out":3937,"duration_ms":44372,"significance":"The problem addressed—energy-efficient resource allocation under incomplete and dynamically mismatched traffic—is timely and practically relevant. If the convergence guarantee (Theorem 2) were rigorously established, the MDDRA framework would be a useful contribution to 6G network optimization, particularly for scenarios where complete spatial-temporal traffic data are unavailable. The paper also includes a substantial simulation study, comparison with multiple baselines, and an extension to real-world maps, which are commendable. However, the central theoretical guarantee is not sound: the proof of Lemma 1 (Appendix B) relies on an invalid implication from non-positive Lyapunov drift to queue convergence, and the proof of Theorem 2 (Appendix D) assumes the very constraints it must prove. These are load-bearing gaps that the numerical experiments cannot compensate for. The paper does not provide reproducible code or machine-checked proofs, so the claimed convergence cannot be independently verified.","major_comments":[{"comment":"The proof of Lemma 1 assumes that for each time slot τ there exists a correction factor νκ_τ ∈ [0,2] such that ξ(τ) − κτ/T = −νκ_τ Qκ_τ (Eq. 30). This is an additional feasibility assertion, not a consequence of the traffic model or constraints (19)-(22). For a given Qκ_τ, the required value is νκ_τ = (κτ/T − ξ(τ))/Qκ_τ, and nothing in the paper guarantees this lies in [0,2], especially when Qκ_τ is near zero or has the opposite sign. Even granting that such ν exists, the Lyapunov drift argument is invalid: from Qκ_{τ+1} = (1−νκ_τ)Qκ_τ, the drift ΔQκ_τ = [(1−νκ_τ)²−1](Qκ_τ)² is non-positive, but non-positive drift does not imply convergence to zero. For example, νκ_τ = 0 gives zero drift while Qκ_τ is constant, and νκ_τ = 2 gives Qκ_{τ+1} = −Qκ_τ, which never converges. A contraction argument with the scaling factor bounded away from 1, or an independent residual-convergence argument, is needed and is absent. Thus Lemma 1, which underpins the queue-stability part of Theorem 2, is unproven.","section":"Appendix B, Eqs. (30)-(31)"},{"comment":"The proof of Theorem 2 is circular. In Appendix D, after listing the first-order conditions, the authors state: 'At the same time, we should satisfy the constraints (19), (20) and (21).' They then claim that the ADMM update rules 'easily obtain (19) according to (39), obtain (21) according to (41) and obtain (20) according to (38).' However, these derivations assume that the constraint equalities hold in the first place. For instance, obtaining (20) from the stationarity condition in (38) requires substituting the very equality ξ(τ)+νκ_τ Qκ_τ = κτ/T that the proof is supposed to establish. Similarly, the claim that Adam converges to a bounded region (Ref. [54]) does not imply that the constraints (19)-(21) become feasible; it only gives stationarity of the unconstrained augmented Lagrangian. The existence of ν∈[0,2] satisfying all three nonlinear constraints simultaneously is never proven, and the argument does not establish the queue convergence that the optimality statement depends on.","section":"Appendix D, Part A"},{"comment":"The final step of the proof, Eq. (44), constructs an increasing sequence {ηT,τ_IREE} and argues that 'according to Appendix B, we can guarantee the optimality of the IREE when τ → ∞.' Since Appendix B's Lemma 1 is invalid, this optimality guarantee is unsupported. Moreover, the theorem's hypothesis 'lim_{k→∞} Λ(ητ,(k)_IREE) = 0' is itself a convergence assumption that is never established; it is exactly the kind of statement the proof should deliver. The footnote (Footnote 8) weakens the claim by saying the convergence 'still hold[s] in many practical scenarios,' but no sufficient conditions for practical convergence are provided. Therefore, the central claim of the paper, lim_{k,τ→∞} ηT_IREE = ηT,*_IREE, lacks a valid proof.","section":"Theorem 2 and Appendix D, Part B"},{"comment":"The universal approximation result in Theorem 1 is a composition of known results: the RBF approximation property from Ref. [11] and the universal approximation of recurrent networks from Ref. [53], with an error bound that additionally depends on the unproven assumption that the FGO block can learn the autoregressive dynamics of the RBF parameters with bounded errors ε2, ε3 (Eq. 34). The paper does not provide a new approximation theorem, only a re-statement of existing results in a specific architecture. Since this theorem is presented as a contribution, the authors should either make precise the new element (e.g., the architectural constraint of FGO layers) or explicitly label the result as a corollary of prior work.","section":"Appendix C, Theorem 1"}],"minor_comments":[{"comment":"The acronym IREE is used in the abstract before the metric is formally introduced in Section II-B. Please define the acronym and give a one-sentence description in the introduction to help readers not familiar with Ref. [24].","section":"Abstract and Section I"},{"comment":"The GRAF network is described as predicting traffic demands D(L,τ), but Eq. (27) defines CT(L,τ), the network capacity. The mechanism by which the network output is converted to a traffic prediction (Algorithm 1, line 3) is not specified. Please clarify the relationship between CT(L,τ) and D(L,τ) in the forward pass.","section":"Section IV-C, Eq. (27) and Algorithm 1"},{"comment":"The sentence 'in (24) we make some equivalent transformations on (21) to ensure the quadratic nature of κτ' is unclear. The quadratic nature of κτ in the penalty term is already apparent from the expression; please rephrase to explain why the particular form of the transformed constraint is chosen.","section":"Section IV-B, Eq. (24)"},{"comment":"In the row for Baseline 4, 'Aaveraged historical traffic' contains a typo; it should read 'Averaged historical traffic.'","section":"Table IV"},{"comment":"The performance gap between MDDRA and Baseline 2 is attributed to 'prediction error of the GRAF network,' but no direct measurement of this prediction error in terms of IREE is provided. A quantitative decomposition of the gap (e.g., prediction error vs. queue tracking error) would strengthen the analysis.","section":"Section V-A and Figure 5"},{"comment":"The footnote 'Although the above convergence analysis relies on idealized traffic assumptions... the proposed MDDRA scheme can still converge in many practical scenarios as discussed in Section V' overstates the evidence: Section V shows numerical convergence for specific instances, which does not substitute for the missing theoretical guarantees. The authors should note that the practical convergence is empirical rather than proven.","section":"Section IV-D, Footnote 8"}],"recommendation":"reject","confidential_remarks":"The manuscript file contains an appended authors' response letter to a different journal's review (IoT-46919-2025), which suggests that the same paper may have been submitted elsewhere. This dual-submission concern, combined with the fundamental proof gaps in Lemma 1 and Theorem 2, makes the paper unsuitable for publication in its current form. Even if the presentation issues were fully resolved, the central convergence claim would require a substantially new proof strategy, well beyond a minor revision. I recommend rejection, not because the topic is unimportant, but because the load-bearing theoretical results are not mathematically sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here is my read of arXiv:2506.03508. The MDDRA scheme and GRAF network are a real engineering contribution worth discussing, but the paper's main theoretical guarantee is unsupported.\n\nWhat is actually new: the combination of a Lyapunov queue with a GRAF predictor (FourierGNN plus RBF layer) for IREE maximization under incomplete traffic data is a legitimate extension of prior work. The pre-training and fine-tuning strategy is sensible and the comparisons against FEDformer, GCN, and FourierGNN show consistent prediction gains. The IREE improvements of 10–20% over baselines and the robustness experiments under varying sampling rates and a real-world map give the paper practical value. The writing is clear and the system model is spelled out.\n\nThe soft spot is the proof of Theorem 2, and the stress-test note is correct: Appendix B assumes the existence of correction factors ν∈[0,2] that exactly force the queue residual to zero at each step. That is an assumption, not a derived property. No argument shows that the required ν lies in [0,2] for arbitrary queue states, especially when the queue is small or has the wrong sign. And the Lyapunov drift argument is invalid: ΔQ≤0 does not imply Q→0. With ν=0 the drift is zero but the queue never shrinks; with ν=2 it merely alternates sign. A contraction rate bounded away from 1, or an independent residual-convergence argument, is missing.\n\nAppendix D inherits this gap. It assumes constraints (19)–(21) are satisfied and then uses the unproven queue convergence to claim optimality. Equation (21) is also circular because it contains the target ηT_IREE being bounded. Theorem 1 is a composition of known RBF and universal approximation results, so it is not a new theoretical result. No code is released, so the numerical claims cannot be independently checked. The appended response letter shows an editor already recommended rejection with major concerns; the proof gaps in the current version are consistent with that.\n\nThe algorithmic idea is worth a serious referee, and I would send it to peer review rather than desk reject, but only with a clear request: either prove the feasibility of the correction factors and give a genuine contraction argument, or remove the optimality claim and present the convergence as empirical. If they can do that, this could become a solid method paper. As it stands, I would not cite the convergence theorem, though I might cite the GRAF architecture in related work.","headline":"A useful algorithm and solid simulations, but the central convergence proof does not hold as written.","tokens_in":37216,"tokens_out":2105,"would_cite":false,"duration_ms":26397,"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":"A hybrid model-data scheduler is claimed to converge to the optimal integrated relative energy efficiency in 6G networks even when traffic data is incomplete.","keywords":["integrated relative energy efficiency","model-data dual-driven","Lyapunov queue","graph neural network traffic prediction","6G resource allocation","traffic-capacity mismatch","radial basis function","energy efficiency"],"falsifier":"Run MDDRA on a constructed traffic trace with a sudden demand surge after a long quiet period and check whether constraints (19)–(21) admit $\\nu\\in[0,2]$ at every slot; if a slot forces $\\nu$ outside that interval and IREE stops tracking the full-information optimum, the claim fails. A cheaper check is to record the required $\\nu$ values from the paper's own simulator and test whether they ever leave $[0,2]$.","tokens_in":35881,"feed_emoji":"⚡","tokens_out":11181,"duration_ms":104489,"temperature":0.7,"pith_summary":"The paper tries to establish that a hybrid model-data scheduler can maximize integrated relative energy efficiency (IREE) — a 6G energy metric that penalizes spatial-temporal mismatch between capacity and traffic — without ever holding a complete traffic map. The algorithmic claim is Theorem 2: under smoothness of the fine-tuning loss, the IREE value converges to the optimum as both iteration and time grow. The practical stake is that operators could allocate bandwidth and power from volunteered user reports and short traffic histories, yet track the full-information optimum closely; the reported gap is about 5.1–7.6%. If true, the result turns traffic prediction from a precondition into an optional accuracy booster, because a Lyapunov queue corrects prediction drift over the long run.","feed_headline":"Hybrid scheduler claims optimal 6G energy efficiency from traffic gaps","feed_subtitle":"Lyapunov queue plus graph predictor stays within 7.6% of full-data allocation while using only volunteer reports.","key_machinery":"The machinery is the pair of a Lyapunov virtual queue and a correction factor. Three virtual queues $Q^\\zeta_\\tau$, $Q^\\kappa_\\tau$, and $Q^\\eta_\\tau$ store the accumulated errors in the utility constraint, the JS divergence, and the transient IREE; correction factors $\\nu^\\zeta_\\tau,\\nu^\\kappa_\\tau,\\nu^\\eta_\\tau\\in[0,2]$ are meant to zero each queue residual at every slot, so transient optimization tracks the long-term optimum. Around them sits a recursive ADMM on the augmented Lagrangian. The data-driven partner is the GRAF network, which uses Fourier Graph Operator layers over the base-station adjacency graph to forecast bandwidth and power configurations and a radial-basis (RBF) layer to interpolate missing traffic while preserving spatial continuity.","core_discovery":"The central claim is that the MDDRA algorithm drives $\\eta_T^{\\mathrm{IREE}}$ to the optimal value, i.e., $\\lim_{k,\\tau\\to\\infty}\\eta_T^{\\mathrm{IREE}}=\\eta_T^{*,\\mathrm{IREE}}$ (Theorem 2), under dynamic traffic and incomplete spatial-temporal data. The argument couples a data-driven GRAF network, which predicts current traffic from incomplete reports using radial-basis interpolation and Fourier graph operators, with a model-driven Lyapunov queue that accumulates historical errors while correction factors force the transient allocation to satisfy long-term constraints. The paper also claims a universal approximation property for GRAF on spatially continuous, temporally autoregressive traffic, and reports IREE gains of 10.2–20.7% over purely data-driven or model-driven baselines in simulations.","pith_inferences":["A finite-time version of Theorem 2 would be the natural next step: the asymptotic statement does not say how many slots an operator needs before the queue correction overtakes prediction error, and the numerical study uses $T=30$.","The feasibility of the correction factors inside $[0,2]$ can be tested adversarially; a constructed demand sequence that forces a factor outside that interval would reduce the guarantee from exact optimality to approximate tracking.","The same Lyapunov-plus-graph template transfers to other long-term network objectives such as latency, fairness, or carbon intensity whenever a transient metric and a mismatch term like the JS divergence can be defined.","Because GRAF trains on complete historical resource configurations rather than on sparse traffic readings, its learned topology abstraction may transfer to deployments with different traffic but similar base-station layouts."],"forward_implications":["Operators could run bandwidth and power allocation for IREE maximization using only volunteered user location reports, without collecting complete spatial-temporal traffic data.","The 10.2% gain over greedy scheduling with predicted traffic and the 20.7% gain over scheduling with averaged historical traffic give a concrete split of what the long-term queue correction and the graph predictor each contribute.","The gap to complete-information benchmarks stays near 5.1% when current traffic is known and 7.6% for full ADMM, so the price of data incompleteness is bounded in the tested scenarios.","The design-principle result steers extra power budget toward roads with large speed limits and higher driving visibility, because those conditions raise the JS divergence faster than they raise network utility.","GRAF's spatial interpolation makes the gains persist as the sampling rate drops; at half the user reports the reported IREE advantage over graph or frequency baselines reaches 37.6%."],"supporting_citations":[{"why":"defines the IREE metric that the whole allocation problem maximizes.","marker":"[24]"},{"why":"supplies the radial-basis capacity approximation and the RBF-based IREE optimization that MDDRA extends to incomplete traffic.","marker":"[11]"},{"why":"provides the Lyapunov rate-stability conditions that generate the virtual queues and correction constraints.","marker":"[38]"},{"why":"provides the Fourier Graph Operator layers used in GRAF and the pre-training complexity estimate.","marker":"[39]"},{"why":"supplies the distributed ADMM updates used to solve the recursive augmented Lagrangian.","marker":"[37]"},{"why":"provides the convection-diffusion traffic model used to generate dynamic traffic in simulations and design-principle analysis.","marker":"[25]"},{"why":"gives the universal approximation result used to justify the FGO block learning the autoregressive resource-configuration map.","marker":"[53]"},{"why":"gives the adaptive-gradient convergence property that Theorem 2 relies on for the fine-tuning loss.","marker":"[54]"}],"fun_headline_variants":["Dual-driven 6G scheduler hits optimal energy efficiency","Hybrid algorithm closes 6G traffic-capacity mismatch","10-20% energy gains from hybrid 6G resource allocation","Lyapunov and graph net tweak 6G for peak IREE","Model-data mix reaches optimal 6G energy efficiency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that at every time step the correction factors $\\nu^\\zeta_\\tau,\\nu^\\kappa_\\tau,\\nu^\\eta_\\tau$ can be chosen inside $[0,2]$ to drive the virtual-queue residuals exactly to zero; if any traffic sequence makes that infeasible, the claimed convergence to the optimal IREE has no proof.","fun_headline_variants_meta":{"raw":{"variants":["Dual-driven 6G scheduler hits optimal energy efficiency","Hybrid algorithm closes 6G traffic-capacity mismatch","10-20% energy gains from hybrid 6G resource allocation","Lyapunov and graph net tweak 6G for peak IREE","Model-data mix reaches optimal 6G energy efficiency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000677,"raw_usage":{"total_tokens":3068,"prompt_tokens":924,"completion_tokens":2144,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":2058}},"tokens_in":540,"tokens_out":2144,"duration_ms":16671,"temperature":1.0,"reasoning_tokens":2058,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:01:36.447728+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MDDRA on a constructed traffic trace with a sudden demand surge after a long quiet period and check whether constraints (19)–(21) admit $\\nu\\in[0,2]$ at every slot; if a slot forces $\\nu$ outside that interval and IREE stops tracking the full-information optimum, the claim fails. A cheaper check is to record the required $\\nu$ values from the paper's own simulator and test whether they ever leave $[0,2]$.","supporting_citations":[{"cited_title":"IREE Oriented Gree n 6G Networks: A Radial Basis Function Based Approach,","cited_arxiv_id":null,"evidence_quote":"defines the IREE metric that the whole allocation problem maximizes."},{"cited_title":"Location-and-Preference Joint P rediction for Task Assign- ment in Spatial Crowdsourcing,","cited_arxiv_id":null,"evidence_quote":"supplies the radial-basis capacity approximation and the RBF-based IREE optimization that MDDRA extends to incomplete traffic."},{"cited_title":"A multi-scale analysis of 27,000 urban street networks: Eve ry US city, town, urban- ized area, and Zillow neighborhood,","cited_arxiv_id":null,"evidence_quote":"provides the convection-diffusion traffic model used to generate dynamic traffic in simulations and design-principle analysis."}],"review_version":1}