{"id":"a3e2d506-b78b-4aef-8a61-008018cf20f0","arxiv_id":"2504.21444","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A Lyapunov-optimized, adversarial-bandit slicing framework schedules 6G immersive traffic alongside eMBB and URLLC services, with a Kalman-filter refinement for non-stationary channels.","lead":"This paper proposes a wireless scheduling framework that lets data-heavy, low-latency immersive services share spectrum with legacy phone and sensor traffic without hard-slicing them apart. It combines queue-based Lyapunov optimization with an adaptive learning algorithm and a Kalman filter, reporting better throughput and latency in simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's sublinear regret bound relies on F*_k being linear in the context, but F*_k is the optimum value of a non-convex MINLP and is piecewise affine at best; Appendix C's linearity proof is circular.","rationale":"The reader's weakest_assumption correctly identifies the linear-reward hypothesis as the load-bearing point of Theorem 1. My stress-test confirms and sharpens that concern: Appendix C does not merely fail to prove linearity; the claimed linearity is actually false in general because F*_k is the value function of a maximization, hence convex and piecewise-affine rather than affine. Appendix D's affine approximation is not a rescue, since its coefficients depend on the allocation decisions that are themselves context-dependent. Without a linear reward model, the LinEXP3 machinery of [45] cannot be invoked, so the sublinear regret bounds (26) and (27) are not established. The numerical results are directionally favorable but do not test this structural assumption, and no code or confidence intervals are provided. The Lyapunov 'equivalence' is another overstatement (minimizing an upper bound is not equivalent to solving the original problem), but the linearity gap alone invalidates the main theorem's proof. Therefore the REJECT verdict is appropriate; I would not change it.","tokens_in":29988,"tokens_out":3922,"duration_ms":46287,"concrete_test":"Construct a small counterexample with N=1 user, |F|=2 sub-channels, fixed channel gains, fixed F_l, and fixed arrivals. For a sweep of virtual-queue values G in [1,100] (holding Q, Lambda, and noise fixed), compute the exact optimal value F*_k of Problem 3.1 by exhaustive search over the 4 binary allocations with optimal power allocation for each. Then check the affine identity F*_k(20) = (F*_k(10) + F*_k(30))/2 for representative G=10,20,30. If the identity fails (equivalently, if the optimal allocation changes between these points), F*_k is not linear in G, and the linear-reward hypothesis required by Theorem 1 is violated. A residual-based linear regression over the sweep would confirm the same conclusion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim, Theorem 1, applies the LinEXP3 regret bound of [45], which requires the per-super-frame reward (1/\\bar{K}) \\sum_k F*_k(F_l) to be linear in the context vector X_l. Appendix C attempts to prove this by asserting, in three bullet points, that F*_k(F_l) is linear with {G_n(k)} and {C_n(k)} \"according to the definition of F*_k(F_l)\". This is not a proof: Problem 3.1 defines F*_k(F_l) as the maximum over allocations {b_n(t,f), p(t,f)} of a linear objective. The maximum of affine functions is convex and piecewise-linear, not globally affine. For example, with two candidate allocations, F* = max(a1 G + b1, a2 G + b2), which has a kink where the optimal allocation switches. The context X_l also contains nonlinear features such as Q^2 and GQ, so even a fixed affine-in-rate objective would not be linear in X_l unless the corresponding coefficients of the quadratic/cross terms are exactly zero, which is not shown and is generically false. Appendix D does not repair this gap: it derives an affine approximation for r_n(k) with coefficients alpha_n(k), beta_n(k) that depend on the allocation variables b_n(t,f), p(t,f), which are themselves the outputs of the optimization and therefore depend on the context. Consequently, the \"unbiased estimation\" step in Appendix C presupposes a true linear parameter theta that need not exist; under misspecification the LinEXP3 regret bound of [45] does not apply. Separately, Lemma 1 provides only an upper bound on the drift-minus-utility, so the claimed \"equivalence\" in Problem 2 is an overstatement; however, the linearity failure is the more immediate obstruction to Theorem 1. Because the sublinear regret guarantee is the paper's main claimed contribution and it is unsupported, the REJECT verdict stands.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes a two-timescale radio access network slicing framework for multiplexing legacy eMBB/URLLC traffic with immersive MBBLL traffic. A Lyapunov drift transformation is used to convert the long-term throughput maximization into a per-frame weighted utility problem, a penalty-BCD algorithm (PBRA) is proposed for the frame-level mixed-integer resource allocation, and two contextual-bandit slicing algorithms (Ad2S and Ad2S-NR) are proposed for the super-frame-level slicing decisions, with an ME-KF channel tracker for non-stationary scenarios. The central theoretical claim is Theorem 1, which bounds the cumulative regret of both algorithms by O(L^{2/3}) under stationary and non-stationary channels. The paper also reports extensive numerical experiments showing throughput, latency, and convergence improvements over several baselines, including EXP3, contextual UCB, and a PPO-based scheme.","tokens_in":30400,"tokens_out":7059,"duration_ms":79146,"significance":"If the regret bounds were valid, the paper would provide a useful advance in non-stochastic network slicing with QoS awareness, and the numerical study is unusually thorough, covering multiple baselines, bursty traffic, varying traffic intensities, parameter sweeps, and an RL comparison. The ME-KF tracking component is a plausible engineering contribution, and the complexity analysis in Lemma 4 is useful. However, the theoretical guarantee is the paper's headline contribution, and it depends on a linear realizability condition that is neither proven nor generically true for the stated problem; the numerical results do not repair this gap because the regret plots are not accompanied by any alternative theoretical justification or by confidence intervals that would let the reader assess the variability of the reported gains.","major_comments":[{"comment":"The claimed sublinear regret relies on the assertion that F*_k(F_l) is linear in the context vector X_l. This is not established and is generically false: F*_k is defined in Problem 3.1 as the maximum over {b_n(t,f), p(t,f)} of an affine objective, and the maximum of affine functions is convex and piecewise linear, with kinks at points where the optimal allocation changes. The context vector contains nonlinear features such as Q_n^2 and G_n Q_n, but the max operator breaks global linearity regardless of feature choice. Appendix D does not repair the gap: Eq. (35) gives r_n(k) = alpha_n(k) E[hat R_n(l)] + beta_n(k), with alpha_n(k) and beta_n(k) explicitly depending on the allocation variables b_n(t,f) and p(t,f), which are outputs of the optimization and therefore context-dependent. Hence the linear model required by the LinEXP3 regret bound in [45] is misspecified, and Theorem 1's sublinear regret bound does not follow.","section":"Section VI (Theorem 1), Appendix C, Appendix D, Eq. (35)"},{"comment":"The proof drops the max operator in the queue dynamics (5) by asserting that r_n(k) is guaranteed to be less than Q_n(k). No constraint enforcing this inequality appears in Problem 1, Problem 2, or Problem 3.1, and the objective rewards large r_n(k), so the inequality is not guaranteed by the problem formulation. Without it, Eq. (30) and hence the drift bound (15) are not valid, and the claimed equivalence between Problem 1 and Problem 2 is not established. Even if the inequality did hold, minimizing the drift bound in (15) is a surrogate optimization, not an equivalent transformation, absent an optimality-gap analysis.","section":"Appendix A, proof of Lemma 1; Eq. (5), Eq. (15)"},{"comment":"The unbiasedness derivation assumes that there exists a true parameter theta_{l,f} such that the per-super-frame reward equals <X_l, theta_{l,f}>. Since the linearity assumption fails as described above, the estimator in Eq. (20) is the least-squares projection of the reward onto the span of X_l, not an unbiased estimator of a true parameter of the linear model. The importance-weighting factor I{...}/pi_l(f|X_l) does not restore the realizability condition required by the regret bound of [45], so the unbiased-estimation hypothesis of that theorem is not satisfied.","section":"Appendix C, unbiased estimation step; Eq. (20)"},{"comment":"The softmax denominator in the policy is written as sum over f' != f of exp(...), omitting the current action f. As written, pi_l(.|X_l) is not a normalized probability distribution over slicing actions, so Algorithm 2 is not the LinEXP3 policy whose regret bound is cited. If this is a typographical error, the denominator should be corrected to sum over all f' in the action set; as it stands, the policy is not implementable as stated and the exploration probability gamma/|F| does not combine with the exponential term to give a valid distribution.","section":"Algorithm 2, Eq. (19)"}],"minor_comments":[{"comment":"The word \"executively\" should be \"exclusively\": each time-frequency element can be allocated to only one user.","section":"Section II-A, after Eq. (2)"},{"comment":"The sentence \"we first formulate ... and than discuss\" contains a typo: \"than\" should be \"then\".","section":"Section III, first paragraph"},{"comment":"The statement \"approximately 1.3x additional regret\" in the paragraph after Theorem 1 is only a numerical illustration; the ratio of the two bounds depends on N and |F| and is not a constant 1.3 in general.","section":"Theorem 1, Eq. (26) and Eq. (27)"},{"comment":"The proof of Lemma 2 would be easier to follow if the notation for the penalty term and the surrogate function distinguished the iteration index of sigma from the block-coordinate inner iterations more clearly.","section":"Appendix B, Lemma 2 and Eq. (33)"},{"comment":"The labels \"1x\" and \"1.3x\" in Fig. 4 are not defined in the caption; the reader has to infer that they refer to the regret overhead ratio claimed in the text.","section":"Fig. 4"}],"recommendation":"reject","confidential_remarks":"The paper addresses a timely problem and has a substantial simulation study, but the central theoretical claim is not supported. The linearity of the reward in the context vector is asserted in Appendix C and implicitly contradicted by the allocation-dependent coefficients in Appendix D's own approximation. This is not a small missing-proof issue; the claimed linear model is generically false for the optimal value of the MINLP in Problem 3.1. I therefore recommend rejection, despite the engineering interest of the numerical results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the thing: this paper proposes a reasonable engineering framework for co-scheduling eMBB, URLLC, and the new immersive MBBLL service using Lyapunov drift, adversarial linear contextual bandits, a penalty-BCD resource allocator, and a Kalman filter for non-stationary channels. The combination is new and the simulations are promising. But the central theoretical claim—sublinear regret for the Ad2S and Ad2S-NR algorithms—depends on a linearity assumption that is asserted, not proven, and is false in general. The max over allocations of a linear objective is piecewise-linear, not linear, in the context vector. Appendix C's three-bullet 'proof' just says 'according to the definition of F*_k(F_l).' Appendix D's affine approximation leaves coefficients that depend on the allocation decisions, which themselves depend on the context. So the LinEXP3 regret bound from Neu and Olkhovskaya does not apply. This is load-bearing: without a linear reward, or a different bandit analysis, Theorem 1 is not established. The Lyapunov transformation is also an upper-bound minimization, not an equivalence; the paper overstates that.\n\nWhat I think is genuinely useful: the problem setup is well motivated, the dual-timescale decomposition is sensible, and the numerical study is more thorough than usual—bursty traffic, a PPO baseline, and parameter sweeps are all there. The ME-KF tracking idea is a nice touch. If the linearity issue were fixed—say by restricting the bandit to a finite set of slicing strategies or by analyzing the misspecified case—the framework could become a solid systems paper.\n\nThe lack of code and error bars on the simulations is a secondary concern; I wouldn't reject on that alone, but it limits how much I trust the reported numbers. The citation pattern looks fine, with external theorems used where appropriate.\n\nIn short: I read this as a revise-and-resubmit, not a desk reject. A serious referee should engage with it because the flaw is instructive and possibly fixable. If I were at an editor's desk, I'd send it out. But as it stands, I would not cite the regret guarantee as a proven result.","headline":"Useful slicing framework, but the main regret bound rests on an unproved and generally false linearity assumption—worth referee time, not acceptance as is.","tokens_in":30970,"tokens_out":4282,"would_cite":false,"duration_ms":41555,"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":"By learning slicing decisions from queue backlogs, two algorithms keep 6G immersive and legacy traffic on one RAN with sublinear regret and measured throughput and latency gains.","keywords":["network slicing","immersive communication","Lyapunov drift","adversarial contextual bandit","QoS-aware resource allocation","non-stationary channel tracking","Kalman filter","dual-timescale optimization"],"falsifier":"Take a two-user, two-subchannel version, compute the optimal $F_k^*(\\mathcal{F}_l)$ by exhaustive search over all binary allocations for a dense grid of backlog values, and fit $F_k^*$ against the context $X_l$: if the residual grows with $X_l$ or the slope depends on the allocation variables $b_n(t,f)$, the linearity assumption fails and the regret bound of Theorem 1 cannot be invoked. Alternatively, simulate Ad2S for large $L$ under i.i.d. channels and check whether the log-regret versus log-$L$ slope exceeds $2/3$ asymptotically.","tokens_in":29805,"feed_emoji":"📶","tokens_out":7049,"duration_ms":67707,"temperature":0.7,"pith_summary":"The paper tries to show that next-generation immersive traffic (AR, VR, holographic telepresence) can be carried in the same radio access network as legacy eMBB and URLLC traffic, without the usual practice of isolating them into separate slices. It argues that statistical isolation causes throughput loss and increased queue backlog, and proposes instead to make slicing decisions adapt online, using queue lengths as context. The central claim is that two algorithms—Ad2S for stationary channels and Ad2S-NR with a Kalman tracker for non-stationary ones—achieve sublinear regret and so learn the near-optimal slicing configuration over time. If the paper is right, operators can mix 6G and legacy services on the same spectrum with lower latency and higher throughput than fixed slicing, while preserving strict URLLC delay guarantees.","feed_headline":"Queue-aware slicing cuts latency 64% for mixed 6G and legacy traffic","feed_subtitle":"Online slicing learns from queue backlogs to keep immersive and legacy services on one RAN, proven sublinear regret.","key_machinery":"The load-bearing mechanism is a three-level control loop. (1) Virtual queues $G_n(k)$ are constructed from the real packet queues and the delay thresholds $\\delta_{e,M}$, so that the long-term average-backlog constraint becomes a queue-stability problem. (2) The Lyapunov drift-minus-utility bound converts the original problem into a per-frame utility maximization that is solved by the PBRA algorithm, a penalty successive upper-bound minimization with block coordinate descent. (3) A non-stochastic contextual-bandit learner, of the LinEXP3 type, chooses the frequency-domain slice partition $\\mathcal{F}_l$ at each super-frame, using the current virtual-queue and buffer-queue values (and, in Ad2S-NR, Kalman-extrapolated pathloss means $\\hat{\\mu}_n(l)$) as the context vector $X_l$.","core_discovery":"The paper's central claim is that the long-term throughput-maximization problem with mixed QoS constraints can be transformed, via the Lyapunov drift theorem, into a short-term utility maximization in which every slice choice is scored by a backlog-weighted rate sum. It then treats that utility as a non-stochastic reward and chooses slice configurations with a linear contextual-bandit learner whose context is the vector of virtual-queue and buffer-queue states, giving the Ad2S algorithm. The theorem states that with learning rate $\\eta = L^{-2/3}(|\\mathcal{F}|(1+4N))^{-1/3}(\\log|\\mathcal{F}|)^{2/3}$ and exploration parameter $\\gamma = L^{-1/3}(|\\mathcal{F}|(1+4N)\\log|\\mathcal{F}|)^{1/3}$, the cumulative regret after $L$ super-frames obeys $\\mathrm{Regret}_L^{\\mathrm{Ad2S}} \\le 5 L^{2/3} (|\\mathcal{F}|(1+4N)\\log|\\mathcal{F}|)^{1/3} F_{\\max}$, and the non-stationary refinement Ad2S-NR, which adds ME-KF pathloss estimates to the context, satisfies the analogous bound with $1+8N$ in place of $1+4N$. Both bounds are sublinear in $L$, which is what it means for the slicing rule to be asymptotically optimal.","pith_inferences":["Editors' extension: the same virtual-queue-as-context recipe could be applied to other shared infrastructures—edge clouds, fronthaul, or open RAN midhaul—where hard-QoS and elastic flows compete, and the regret analysis would carry over as long as a linear-utility approximation is plausible.","Editors' extension: a direct stress test is to run Ad2S on a scenario with a deliberately nonlinear utility (e.g., a fifth-order rate model); the regret is expected to degrade to linear, which would confirm that the linearity premise is the load-bearing part of Theorem 1.","Editors' extension: the ME-KF tracker is replaceable; feeding a learned channel predictor (say, a sequence model) into the context vector would test whether the sublinear bound survives when the extrapolation is no longer provably unbiased."],"forward_implications":["If the regret bounds hold, an online slicing rule can approach the optimal configuration without knowing the channel statistics in advance, because the sublinear regret guarantees that the time-averaged utility converges to the best fixed slice policy in hindsight.","Mixed 6G and legacy operation would no longer require hard statistical isolation: the simulations report a 3.86 Mbps throughput gain, a 63.96% latency reduction, and a 24.36% reduction in exploration time against the tested baselines.","The numerical study reports 92.8% URLLC QoS satisfaction, meaning the probability that the frame's delivered data clears the queue within the delay bound, which is higher than the EXP3, contextual-UCB, and non-adaptive baselines.","Operators can choose where to sit on the throughput-versus-latency trade-off by tuning the ratio $\\omega_Q/\\omega_T$ of the Lyapunov weights, with the suggested operating range $6\\times 10^{-5}$ to $8\\times 10^{-5}$."],"supporting_citations":[{"why":"Supplies the LinEXP3 adversarial linear contextual bandit whose regret bound Theorem 1 directly invokes.","marker":"[45]"},{"why":"Provides the Lyapunov drift theorem used to transform the long-term problem into the drift-minus-utility form.","marker":"[40]"},{"why":"The penalty successive upper-bound minimization (PSUM) framework on which the PBRA algorithm's convergence rests.","marker":"[44]"},{"why":"Defines the bursty traffic model and dual-timescale resource structure used in the formulation and experiments.","marker":"[13]"},{"why":"Contextual UCB baseline that the proposed scheme must beat and motivates the non-stochastic approach.","marker":"[26]"},{"why":"Baseline QoS-oriented resource allocation with slicing puncture (NAdS+DRAS) compared in simulations.","marker":"[9]"},{"why":"Larsen's delayed-measurement extrapolation method used by ME-KF to predict pathloss.","marker":"[49]"},{"why":"Establishes unbiasedness of the extrapolated Kalman measurement, needed for the regret analysis.","marker":"[51]"},{"why":"The EXP3 non-stochastic multi-armed bandit baseline and comparison.","marker":"[42]"}],"fun_headline_variants":["Latency down 64% with unified slicing for immersive and legacy","Adversarial slicing learns to mix 6G and legacy traffic","Queue-aware framework lifts throughput and cuts latency for 6G","Sublinear regret slicing for immersive and legacy coexistence","One framework, two traffic types: slicing learns on the fly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The sublinear regret bound rests on the assumption that the per-frame utility $F_k^*(\\mathcal{F}_l)$ produced by the resource-allocation solver is a linear function of the context vector fed to the bandit learner; the paper asserts this linearity rather than proving it, and the affine approximation it derives in the appendix has coefficients $\\alpha_n(k)$, $\\beta_n(k)$ that depend on the allocation decisions themselves.","fun_headline_variants_meta":{"raw":{"variants":["Latency down 64% with unified slicing for immersive and legacy","Adversarial slicing learns to mix 6G and legacy traffic","Queue-aware framework lifts throughput and cuts latency for 6G","Sublinear regret slicing for immersive and legacy coexistence","One framework, two traffic types: slicing learns on the fly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000533,"raw_usage":{"total_tokens":2649,"prompt_tokens":1115,"completion_tokens":1534,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":731,"completion_tokens_details":{"reasoning_tokens":1449}},"tokens_in":731,"tokens_out":1534,"duration_ms":11458,"temperature":1.0,"reasoning_tokens":1449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:03:10.301278+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-user, two-subchannel version, compute the optimal $F_k^*(\\mathcal{F}_l)$ by exhaustive search over all binary allocations for a dense grid of backlog values, and fit $F_k^*$ against the context $X_l$: if the residual grows with $X_l$ or the slope depends on the allocation variables $b_n(t,f)$, the linearity assumption fails and the regret bound of Theorem 1 cannot be invoked. Alternatively, simulate Ad2S for large $L$ under i.i.d. channels and check whether the log-regret versus log-$L$ slope exceeds $2/3$ asymptotically.","supporting_citations":[{"cited_title":"Efficient and robust algorithms for adversarial linear contextual bandits,","cited_arxiv_id":null,"evidence_quote":"Supplies the LinEXP3 adversarial linear contextual bandit whose regret bound Theorem 1 directly invokes."},{"cited_title":"Neely, Stochastic network optimization with application to com- munication and queueing systems , San Rafael, CA, USA: Morgan & Claypool, 2010","cited_arxiv_id":null,"evidence_quote":"Provides the Lyapunov drift theorem used to transform the long-term problem into the drift-minus-utility form."},{"cited_title":"Network slicing for service-oriented networks under resource constraints,","cited_arxiv_id":null,"evidence_quote":"The penalty successive upper-bound minimization (PSUM) framework on which the PBRA algorithm's convergence rests."},{"cited_title":"How should I orchestrate resources of my slices for bursty URLLC service provision?","cited_arxiv_id":null,"evidence_quote":"Defines the bursty traffic model and dual-timescale resource structure used in the formulation and experiments."},{"cited_title":"laco: a latency-driven network slicing orchestration in beyond-5G networks,","cited_arxiv_id":null,"evidence_quote":"Contextual UCB baseline that the proposed scheme must beat and motivates the non-stochastic approach."},{"cited_title":"Resource allocation and slicing puncture in cellular networks with eMBB and URLLC terminals coexistence,","cited_arxiv_id":null,"evidence_quote":"Baseline QoS-oriented resource allocation with slicing puncture (NAdS+DRAS) compared in simulations."},{"cited_title":"Incorporation of time delayed measurements in a discrete-time kalman filter,","cited_arxiv_id":null,"evidence_quote":"Larsen's delayed-measurement extrapolation method used by ME-KF to predict pathloss."},{"cited_title":"Extrapolation of delayed measurements for fusion in a distributed sensor network,","cited_arxiv_id":null,"evidence_quote":"Establishes unbiasedness of the extrapolated Kalman measurement, needed for the regret analysis."},{"cited_title":"The nonstochastic multiarmed bandit problem,","cited_arxiv_id":null,"evidence_quote":"The EXP3 non-stochastic multi-armed bandit baseline and comparison."}],"review_version":1}