{"id":"1e0c1b2a-26d7-449d-aa2c-fa1bd16d3f4c","arxiv_id":"2506.09197","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An iterative semi-static bandwidth sharing and scheduling policy is proposed that provably approaches the quality of the optimal static sharing policy while updating operator coordination only once per hyperperiod.","lead":"This paper proposes a bandwidth-sharing algorithm for wireless operators that updates how much spectrum they share once every few periods instead of continuously. If it works, operators can deliver real-time video with near-optimal quality while coordinating with each other far less often.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The O(1/V) optimality gap in Theorem 1 depends on an unproved V-independent bound on the stochastic gradient g, but Lagrange multipliers from tight per-period constraints grow with V and virtual-queue backlog, likely making the bound O(V^2) and the gap O(V).","rationale":"The reader's weakest assumption, strict static feasibility (20), is explicit in Theorem 1 and is standard in drift-plus-penalty proofs, so it is less load-bearing than the proof's handling of the stochastic gradient. My concern targets the core O(1/V) optimality bound: the proof requires a V-independent second-moment bound on g, but the definition of g through Lagrange multipliers makes it depend on V and P. A direct KKT inspection of RA(S_t,A(k)) shows λ is at least proportional to V whenever the resource constraint is active, which is exactly the regime the policy is designed for. Thus the constant C in Lemma 5 is not shown to be independent of V, and the subsequent division by V in Lemma 7 could turn a V^2 drift bound into an O(V) gap. This is a technical gap in the central theorem, not just a missing feasibility verification. Nevertheless, the paper presents a plausible algorithm and the gap may be repairable by choosing η to scale with V or by showing a sharper bound on P·Q' terms; the reader's CONDITIONAL verdict is still appropriate, so I do not change it. I disagree with the reader's identification because the strict-feasibility condition is at least honestly stated as an assumption, whereas the gradient bound is silently assumed without proof despite being essential.","tokens_in":1307,"tokens_out":1043,"duration_ms":245362,"concrete_test":"Run the per-period subproblem RA(S_t,A(k)) with the paper's two-region, two-operator simulation settings: set arrival rates γ = 0.1 and 0.9, T = 20, Q_min = 0.3, α = 0.008, and compute λ_i^r from KKT conditions for a region with high load over several random arrival vectors. Then compute ||g(S_t,A(k))||^2 per (18) for V = 1, 10, 100, 1000 while holding P and S fixed. If ||g||^2 grows linearly or quadratically in V (rather than remaining bounded or growing sublinearly), the constant C in Lemma 5 is not V-independent. Additionally, simulate ABS for V = 10, 100, 1000 with otherwise identical parameters and fixed η=0.01; if the empirical gap between ABS's average objective and the static benchmark does not shrink roughly as 1/V, Theorem 1's stated bound is not reflected in the regime the simulations actually study.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 1 in Appendix B defines constants B, C, D in Lemma 5, with C an upper bound on ||g(S_t,P_t,A_t)||^2 for the stochastic gradient in (18)—the sum of negative Lagrange multipliers over a hyperperiod. This bound must be uniform in V and in the virtual-queue state P. But for each period k, the per-region resource constraint (16) couples all clients in operator i/region r, and the KKT conditions for the convex program RA(S_t,A(k)) with penalty term in F_k show λ_i^r = O(V + max_n P_i,r,n) Q'(τ*) in any region where the constraint is tight, because the objective F_k contains -V Q(τ) and P_i,r,n (Qmin - Q + α)+. The virtual queue P evolves with bounded increments and in standard drift-plus-penalty analyses only its time-average is controlled, not its supremum, so P can transiently grow with V. Consequently λ, and hence ||g||^2, can plausibly scale as V^2, making C = Θ(V^2) in Lemma 5. If so, dividing (41) by V in Lemma 7 yields an objective gap of order V, not O(1/V), so Theorem 1's stated optimality guarantee is not established. The proof's assertion that the gradient is bounded because Q is twice-differentiable on a compact set only bounds Q' w.r.t. τ, not the Lagrange multiplier w.r.t. S; this is a non-sequitur. Even informally, the claimed asymptotic performance 'within O(1/V) of optimum' relies on ηC/(2V) being small, which the proof does not ensure.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ABS, an adaptive semi-static bandwidth-sharing policy for multi-operator wireless networks serving real-time video. Sharing variables are updated once per hyperperiod by a projected stochastic-gradient step, while per-period client scheduling is obtained by solving a convex resource-allocation problem with a virtual-queue penalty for the long-run minimum-quality constraint. The main theoretical claim is Theorem 1: under a strict static-feasibility condition, ABS satisfies the sharing and minimum-quality constraints and attains an objective within O(1/V) of the optimal stationary randomized policy with static sharing. The paper also reports simulations comparing ABS with optimal static and dynamic sharing policies, studying effects of step size, hyperperiod length, and arrival/channel variability.","tokens_in":20316,"tokens_out":11026,"duration_ms":148417,"significance":"If the O(1/V) guarantee were fully established, the paper would make a useful contribution: it proposes a concrete, low-coordination mechanism for inter-operator bandwidth sharing with a QoE-aware scheduling component, and it gives a convex relaxation plus a Lyapunov/SGD-style proof architecture that is credible in outline. The authors also provide a clear algorithm, parameter studies, and comparisons against baselines. However, the current proof has a load-bearing gap concerning the boundedness of the stochastic gradient, one supporting convexity lemma is asserted without proof, and the optimality benchmark is internally weaker than the abstract's claim about the optimal static-sharing policy. These issues prevent the central theoretical claim from being accepted as stated.","major_comments":[{"comment":"The O(1/V) claim rests on a uniform upper bound C on E[||g(St, At)||^2 | Theta] in (24), (28), and (32), used in the final gap (43). The proof after (29) asserts that the gradient is bounded because Q is twice differentiable on a compact set. That only bounds Q' with respect to tau; it does not bound the Lagrange multipliers in (18). For RA(St, A(k)), KKT stationarity for a tight per-region constraint (16) gives lambda_i^r = (V + P_i^r,n(k)) Q'(tau*_i^r,n) when the penalty term is active, so lambda, and hence g, scale with V and with the virtual-queue state P. Lemma 6 controls only time averages of P, not a uniform bound independent of V. If C = Theta(V^2), then the term eta C/(2V) in (43) is Theta(V), and the stated O(1/V) optimality gap in Theorem 1 is not established. This needs to be repaired either by proving a true V-uniform bound on the second moment of g or by restating the guarantee with the correct dependence on V and the queue backlog.","section":"Appendix B, Lemma 5 and Lemma 7"},{"comment":"Convexity of F_k(tau*(S_t, k), A_t) as a function of S_t is load-bearing: it converts (30) into (31), and hence produces the -2 eta [f(S_t) - f(S**)] term in the drift bound. The proof is omitted and replaced by a citation to Corollary 2.2 of [47]. Given that the entire drift-plus-penalty argument depends on this inequality, the authors should either supply the argument or state exactly which hypotheses of the cited theorem are verified for the parametric program RA(S_t, A(k)).","section":"Appendix B, Lemma 4"},{"comment":"The benchmark in Theorem 1 is the asymptotically optimal stationary randomized policy with static sharing defined by OPT-SS*, i.e., the convex relaxation with constraint (9), not the original optimal static-sharing problem OPT-SS. Lemma 1 shows only that enforcing (8) implies constraint (1); it does not establish equivalence, and OPT-SS* replaces the expectation by a time average. Therefore the abstract's statement that ABS 'converges to the optimal static sharing policy' and the analogous wording in the conclusion are stronger than what is proved. The claims should be restated relative to OPT-SS*, or an equivalence between the relaxation and OPT-SS must be established.","section":"Section V-A and abstract"},{"comment":"The strict static feasibility condition (20) is the hypothesis under which Theorem 1 applies, yet the simulations never verify that this condition holds in any of the reported scenarios. The paper itself notes that the no-sharing baseline is infeasible for Qmin = 0.3, which shows the condition is doing real work. The authors should check (20) for the simulation parameters, or otherwise limit the empirical claims to the feasible regime; otherwise the numerical experiments do not demonstrate the theorem's regime.","section":"Section VI-A and Theorem 1 hypothesis"}],"minor_comments":[{"comment":"There is a typo in the first sentence: 'thoerem' should be 'theorem'.","section":"Appendix B"},{"comment":"The symbol P is used both for the set of periods in a hyperperiod, P_t, and for the virtual-queue array P(k); this is confusing in Section V and in the proof and should be disambiguated.","section":"Notation"},{"comment":"Lemma 2's convexity argument is only sketched; a one-sentence justification that (Qmin - Q(tau) + alpha)_+ is convex in tau because Q is concave and increasing would remove any doubt.","section":"Section IV-C"},{"comment":"The adaptive/variable step-size curve ('green curve') is described in the text, but no adaptive step-size rule is defined in Algorithm 1 or in the parameter table; this makes the simulation result hard to reproduce.","section":"Section VI-B, Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The skeptic's concern about the V-dependence of C appears legitimate and is the main technical obstacle. If the authors can provide a genuine uniform bound, or change the statement to an honest O(V + eta C(V)/V) guarantee, the paper may become publishable. The overclaim connecting Theorem 1 to the optimal static-sharing policy should be corrected regardless. The manuscript's scope and system model are appropriate for a networking venue, so a major revision is preferable to rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the ABS policy is a reasonable engineering contribution, and the simulations suggest it works, but the headline optimality guarantee in Theorem 1 is not established by the proof. The stress-test note identifies the right culprit.\n\nWhat's actually new: the packaged combination of a hyperperiod-level stochastic gradient update (driven by dual multipliers from the per-period allocation) with virtual queues for the minimum-quality constraint is a sensible semi-static policy, and I don't think it appears verbatim in the prior work. The intuition that Lagrange multipliers measure the marginal value of sharing is sound, and the O(1/V) target is the right one. The simulation study covers step size, hyperperiod length, and nonstationary arrivals/channels, which is useful; the no-sharing baseline is handled honestly, including the infeasibility caveat.\n\nWhere it breaks down: the proof of Theorem 1 has three load-bearing gaps. First, Lemma 4 (convexity of the optimal-value function of RA(S) in S) is asserted, not proved; the Fiacco-Kyparisis result may supply it, but the referee would need to check the conditions. Second, the constants B, C, D in Lemma 5 are not shown to be V-independent. C bounds ||g||^2, where g is a sum of Lagrange multipliers over a hyperperiod. Those multipliers are sensitivities of an objective containing -V Q(τ) plus virtual-queue penalty terms, so they plausibly grow with V and with queue backlog; standard drift-plus-penalty does not uniformly bound the latter. The paper's sentence that the gradient is bounded because Q is twice-differentiable on a compact set is a non-sequitur: it bounds Q', not λ. If C = Θ(V^2), Lemma 7 gives a gap O(V), not O(1/V). The stress-test note holds up. Third, the benchmark in Theorem 1 is the optimum of the relaxed problem OPT-SS*, not the original static-sharing problem, and the abstract omits that caveat when claiming convergence to the optimal static sharing policy. Related work also skips the authors' own [33], which is a close predecessor.\n\nBottom line: the algorithm is plausible and the simulations are encouraging, but the central theorem needs real work. This deserves peer review, not desk rejection; a serious referee will push the proof to be honest.","headline":"The ABS policy is a plausible engineering contribution with encouraging simulations, but Theorem 1's O(1/V) guarantee is not actually established; the proof gaps are real but fixable, so it deserves a serious referee.","tokens_in":20846,"tokens_out":3189,"would_cite":false,"duration_ms":31585,"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":"This paper claims that a semi-static bandwidth-sharing policy, coordinating only once per hyperperiod, reaches a total QoE within $O(1/V)$ of the optimal static-sharing policy whenever a fixed-sharing policy can meet the 95% quality…","keywords":["adaptive bandwidth sharing","Quality of Experience (QoE)","real-time video","semi-static policy","Lyapunov drift","stochastic gradient descent","static vs dynamic sharing","quality percentile constraint"],"falsifier":"Using the paper's simulation parameters (two operators, two regions, 30 clients each, Bernoulli arrivals with rates 0.1 and 0.9, $Q_{\\min}=0.3$, $\\alpha=0.008$, $T=20$, channel capacity $10^7$ bits per timeslot), compute the minimum of $E[(Q_{\\min}-Q(\\tau)+\\alpha)^+]$ over stationary randomized static-sharing policies. If that minimum is not strictly below $0.0004=0.05\\alpha$, strict static feasibility fails and Theorem 1's premise is absent in exactly the regime where the paper demonstrates convergence.","tokens_in":19705,"feed_emoji":"📶","tokens_out":15892,"duration_ms":164909,"temperature":0.7,"pith_summary":"Real-time video packets are only useful if delivered before their deadline, so an operator that has spare timeslots in one region and a shortage in another gains from sharing, but renegotiating the share every period is expensive. This paper claims that a semi-static policy, ABS, can get most of the benefit with coordination only once per hyperperiod: it freezes the sharing vector for $H$ periods, schedules clients period by period, and then updates the sharing vector using Lagrange multipliers from those scheduling problems. The headline theoretical result is Theorem 1, which states that for any strictly static feasible minimum quality $Q_{\\min}$ (meaning some fixed-sharing policy already meets the 95% quality constraint with a strict margin), the policy satisfies the sharing and minimum-quality constraints and reaches a total QoE within $O(1/V)$ of the optimal stationary randomized policy with static sharing; increasing $V$ shrinks the gap at the cost of longer virtual queues. Simulations with two operators and two regions show ABS converging close to the optimal static sharing in about 150 hyperperiods regardless of initial sharing or arrival rates, and they show static sharing nearly matching dynamic sharing in this setting. The practical claim is that near-optimal real-time video QoE does not require per-period inter-operator coordination.","feed_headline":"Rarely updated sharing provably nears optimal static video QoE","feed_subtitle":"Updating once per hyperperiod still lands near the optimal static-sharing benchmark.","key_machinery":"ABS, the Adaptive Bandwidth-sharing and Scheduling policy, is the central object. Its two mechanisms are the virtual queue $P_{r,n}^i(k)$, which tracks a client's accumulated shortfall against the smoothed 95% quality constraint through the update $(P+(Q_{\\min}-Q+\\alpha)^+-0.05\\alpha)^+$, and the sharing update $S_{t+1}=\\Pi_\\Omega(S_t-\\eta g(S_t,A_t))$, where $g$ is assembled from the Lagrange multipliers of the per-region constraints in the period scheduling problem. That update is a projected stochastic gradient ascent on the hyperperiod value function, and the proof's Lyapunov function $L=\\frac12\\sum (P_{r,n}^i)^2+\\frac{1}{2\\eta}\\|S_t-S^{**}\\|^2$ couples the two: the negative queue squared terms show the quality constraints are met, while the objective terms yield the $O(1/V)$ optimality gap. The hyperperiod length $H$ is the tunable knob: it sets how often operators coordinate.","core_discovery":"The central claim is that the difficult joint problem of dynamic sharing and scheduling can be replaced, at negligible QoE cost, by a semi-static policy. ABS keeps the sharing vector $S_t$ fixed during hyperperiod $P_t$, solves per period a convex scheduling problem $RA(S_t,A(k))$ that maximizes total video quality minus virtual-queue penalties, and then updates $S_{t+1}=\\Pi_\\Omega(S_t-\\eta g(S_t,A_t))$ using the optimal Lagrange multipliers of the per-region resource constraints. The proof of Theorem 1 shows that this update is a projected gradient step on the expected per-hyperperiod optimal value, and a Lyapunov-drift argument with virtual queues drains the quality deficits while the $V$-scaled objective term yields the $O(1/V)$ optimality gap. In the paper's own summary, the policy converges to the optimal static sharing policy irrespective of initial conditions or fluctuations in arrival rates. The simulation section additionally claims that optimal static sharing achieves performance nearly equivalent to optimal dynamic sharing, so the coordination savings are not bought with much QoE.","pith_inferences":["A finite-horizon statement is left implicit: the theorem is asymptotic in the number of hyperperiods, so a natural next step is a regret bound showing how $V$, the hyperperiod length $H$, and the step size $\\eta$ trade off over a fixed number of hyperperiods.","The manuscript says the proofs of Lemma 2 (convexity of OPT-SS*) and Lemma 4 (convexity of the value function in $S_t$) are omitted for brevity; both supply the convexity on which the projected-gradient update rests, so filling in those proofs would complete the chain backing Theorem 1.","Because the Lagrange multipliers already measure the marginal value of an extra shared timeslot, the same update could be reinterpreted as a price-based coordination market among operators, moving toward the incentive mechanisms the paper lists as future work.","Since the optimality benchmark is a static-sharing policy, ABS inherits the static-versus-dynamic gap; a testable prediction is that this gap widens under strong arrival imbalance or channel variability, in which case the hyperperiod would have to shrink to keep the semi-static policy near-optimal."],"forward_implications":["For any strictly static feasible target, ABS reaches a long-run average total QoE within $O(1/V)$ of the optimal stationary randomized static-sharing policy, and $V$ can be chosen large enough to make the gap arbitrarily small.","ABS satisfies the smoothed 95% minimum-quality constraint in the long run, so clients with hard deadlines get the quality-percentile guarantee without per-period inter-operator negotiation.","The convergence statement holds irrespective of the initial sharing configuration and of fluctuations in arrival rates, which the simulations confirm for different starting points and different arrival processes.","In the simulated two-operator, two-region setting, static sharing captures most of the QoE gain of dynamic sharing, so the coordination savings of ABS come at little performance cost.","Step size controls the trade-off: too large a step converges fast but misses the optimum, too small a step is accurate but slow, and an annealed step size combines fast initial progress with accurate final convergence."],"supporting_citations":[{"why":"Supplies the projection operator for the sharing update and the sensitivity result that identifies the per-period Lagrange multipliers as the gradient of the optimal value used in the update.","marker":"[34]"},{"why":"Provides Lemma 8, the bounded-variation lemma that turns the bounded limsup of the virtual-queue average into the vanishing queue term used in Lemma 9.","marker":"[21]"},{"why":"Provides Lemma 4.3 bounding the change in squared backlog, used to derive the drift inequality for the virtual queues, and the queue-stability background for infrequent updates.","marker":"[17]"},{"why":"Supplies the convexity-of-optimal-value result used in Lemma 4 to show the per-hyperperiod value function is convex in the sharing vector, giving the gradient inequality in the drift proof.","marker":"[47]"},{"why":"Supplies the dominated convergence theorem used to interchange expectation and gradient when identifying $g(S_t,A_t)$ as the gradient of the expected per-hyperperiod value.","marker":"[48]"},{"why":"The per-period dynamic sharing approach that serves as the dynamic benchmark; ABS is designed to approach its performance without per-period coordination.","marker":"[33]"},{"why":"The earlier bandwidth-sharing solution for real-time traffic that this paper extends by adding perceived-video-quality objectives and a general resource allocation model.","marker":"[26]"},{"why":"Supplies the frame-based timely-throughput model for real-time traffic that underlies the paper's arrival and deadline structure.","marker":"[19]"}],"fun_headline_variants":["Semi-static sharing gets near-optimal QoE with far less coordination","Less frequent updates still hit near-optimal video QoE","Hyperperiod updates preserve near-optimal QoE for live video","Coordination slashed without sacrificing video QoE","Infrequent sharing updates still near-optimal for real-time video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole guarantee depends on strict static feasibility: some stationary randomized policy with fixed sharing must already satisfy $E[(Q_{\\min}-Q+\\alpha)^+]\\le \\kappa\\,0.05\\,\\alpha$ with $0<\\kappa<1$ for every client, so the quality target is met with a strict margin by a policy that does not see the future.","fun_headline_variants_meta":{"raw":{"variants":["Semi-static sharing gets near-optimal QoE with far less coordination","Less frequent updates still hit near-optimal video QoE","Hyperperiod updates preserve near-optimal QoE for live video","Coordination slashed without sacrificing video QoE","Infrequent sharing updates still near-optimal for real-time video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000283,"raw_usage":{"total_tokens":1684,"prompt_tokens":970,"completion_tokens":714,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":626}},"tokens_in":586,"tokens_out":714,"duration_ms":7872,"temperature":1.0,"reasoning_tokens":626,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:55:39.791419+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Using the paper's simulation parameters (two operators, two regions, 30 clients each, Bernoulli arrivals with rates 0.1 and 0.9, $Q_{\\min}=0.3$, $\\alpha=0.008$, $T=20$, channel capacity $10^7$ bits per timeslot), compute the minimum of $E[(Q_{\\min}-Q(\\tau)+\\alpha)^+]$ over stationary randomized static-sharing policies. If that minimum is not strictly below $0.0004=0.05\\alpha$, strict static feasibility fails and Theorem 1's premise is absent in exactly the regime where the paper demonstrates convergence.","supporting_citations":[{"cited_title":"Scheduling heterogeneous real-time traffic over fading wireless channels,","cited_arxiv_id":null,"evidence_quote":"Provides Lemma 8, the bounded-variation lemma that turns the bounded limsup of the virtual-queue average into the vanishing queue term used in Lemma 9."},{"cited_title":"Resource allocation and cross-layer control in wireless networks,","cited_arxiv_id":null,"evidence_quote":"Provides Lemma 4.3 bounding the change in squared backlog, used to derive the drift inequality for the virtual queues, and the queue-stability background for infrequent updates."},{"cited_title":"Convexity and concavity properties of the optimal value function in parametric nonlinear programming,","cited_arxiv_id":null,"evidence_quote":"Supplies the convexity-of-optimal-value result used in Lemma 4 to show the per-hyperperiod value function is convex in the sharing vector, giving the gradient inequality in the drift proof."},{"cited_title":"Hajek, Random processes for engineers","cited_arxiv_id":null,"evidence_quote":"Supplies the dominated convergence theorem used to interchange expectation and gradient when identifying $g(S_t,A_t)$ as the gradient of the expected per-hyperperiod value."},{"cited_title":"Maximizing Real-Time Video QoE via Bandwidth Sharing under Markovian setting","cited_arxiv_id":"2401.10681","evidence_quote":"The per-period dynamic sharing approach that serves as the dynamic benchmark; ABS is designed to approach its performance without per-period coordination."},{"cited_title":"Optimizing bandwidth sharing for real- time traffic in wireless networks,","cited_arxiv_id":null,"evidence_quote":"The earlier bandwidth-sharing solution for real-time traffic that this paper extends by adding perceived-video-quality objectives and a general resource allocation model."},{"cited_title":"A theory of QoS for wireless,","cited_arxiv_id":null,"evidence_quote":"Supplies the frame-based timely-throughput model for real-time traffic that underlies the paper's arrival and deadline structure."}],"review_version":1}