{"id":"50a87860-d617-4fb3-95cb-37766f30f725","arxiv_id":"1908.06329","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"For multiclass many-server queues with alternating service interruptions and renewal arrivals in the Halfin-Whitt regime, the paper proves that the optimal diffusion-scale costs converge to the optimal costs of a limiting compound-Poisson jump-diffusion control problem.","lead":"This paper proves that optimal scheduling for large multiclass service systems with random service interruptions and renewal arrivals is well approximated by a jump-diffusion control problem. It derives the limit value functions for discounted and long-run average costs, giving a principled way to design near-optimal server allocation policies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 5.2 omits the proof of the I0 nonempty case, and that case is required for the ergodic lower bound (5.10) in Theorem 3.3.","rationale":"The reader's weakest-assumption field names Assumption 3.2(ii), the bounded mean residual life condition. That assumption is indeed needed for the Foster-Lyapunov machinery, but it is an explicit, stated hypothesis of Theorems 3.2 and 3.3; if it fails, the theorems simply do not apply. By contrast, Lemma 5.2 is asserted without an essential part of its proof. The omitted I0 nonempty case is not an assumption on the model, but a case that the model permits and that the lower-bound argument must cover. The reader's rationale already notes this omission, so the agreement is partial: we agree the omission is load-bearing, but we would rank it above Assumption 3.2(ii) as the single most important gap. The discounted claim (Theorem 3.2) is not affected by this gap. Since the proof gap concerns only the ergodic lower bound and the result may well be true, the appropriate verdict is unchanged: the paper should remain CONDITIONAL, with the condition being that Lemma 5.2's proof for I0 nonempty must be supplied before the ergodic theorem is considered proven.","tokens_in":36435,"tokens_out":4890,"duration_ms":45245,"concrete_test":"Complete the proof of Lemma 5.2 for I0 nonempty by carrying out the calculation indicated in the final paragraph: apply Lemma B.2 with the full generator (4.12) and combine the estimates (B.19)-(B.22) to derive the analog of (B.23) uniformly for all admissible Markov policies with sup_n J_hat(X_hat_n(0), z_n) < infinity, and then verify that the resulting inequality implies (5.11). If the calculation requires an additional structural assumption (e.g., a positive lower bound on the service allocation to I0 classes under the policy, or an epsilon-term that cannot be absorbed), then Lemma 5.2 is not established and the lower bound (5.10) for Theorem 3.3 does not follow from the material in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central ergodic claim (Theorem 3.3) is proved via the lower bound (5.10) in Section 5.3.1. That proof applies Lemma 5.2 to obtain the uniform moment bound (5.11) for any sequence of admissible Markov policies with bounded cost. The proof of Lemma 5.2, however, only sketches the case I0 = empty; the last paragraph asserts that the case I0 nonempty follows by repeating the above argument and applying Lemma B.2, without giving the calculation. This is not a cosmetic omission. The whole point of the modified priority policy in Definition 4.1 and the Lyapunov function in (4.17) is to control classes with zero abandonment (I0), and the estimates (B.19)-(B.22) in the sketch are tailored to the case where every class has a positive drift toward zero. For i in I0, the term gamma_i q_i vanishes, so the drift of |x_i| can be non-positive only through the service allocation; the asserted repetition would need to show that (B.23) still holds uniformly for all policies in the lower-bound argument, which is not immediate. Because the assumptions allow I0 nonempty (only gamma_d > 0 is assumed), the stated proof of Theorem 3.3 is incomplete for a legitimate case.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a sequence of multiclass GI/M/n+M queues with renewal arrivals, exponential services and abandonments, and an alternating renewal (up-down) environment, in the Halfin-Whitt regime. The authors prove a functional central limit theorem showing that diffusion-scaled state processes under non-anticipative work-conserving policies converge to a controlled compound-Poisson jump diffusion. They then study the infinite-horizon discounted and long-run average (ergodic) scheduling problems and claim asymptotic optimality: the optimal values of the queueing models converge to the optimal values of the limiting jump-diffusion control problem. The main technical tool is an augmented Markovian model that includes the age processes of the renewal arrivals and the residual/age process of the down state; generator convergence for this augmented model, together with Foster-Lyapunov moment bounds, is used in place of the usual martingale arguments for mean empirical measures. The discounted result is proved in the style of Atar-Mandelbaum-Reiman, while the ergodic result is proved through lower and upper bounds, with the lower bound relying on a uniform moment estimate for arbitrary admissible Markov policies.","tokens_in":36665,"tokens_out":9242,"duration_ms":97339,"significance":"If the results are correct, this is a substantial contribution: it appears to be the first asymptotic-optimality result for ergodic control of multiclass many-server queues with general renewal arrivals, and the first treatment of optimal scheduling control in an alternating-renewal service-interruption environment at this scaling. The augmented-generator method, the construction of the approximate test functions in (5.12), and the Foster-Lyapunov estimates are nontrivial and likely to be reusable. The discounted theorem and the FCLT are developed in detail and look sound. However, the ergodic lower bound depends on Lemma 5.2, whose proof explicitly treats only the case I0=empty and asserts the nonempty case without calculation. Since the assumptions allow zero-abandonment classes, this is a load-bearing gap, not a cosmetic omission. The central ergodic theorem is therefore not fully established as written.","major_comments":[{"comment":"The proof of Lemma 5.2 explicitly treats only the case I0=empty; after equation (B.23) it says \"We may show that the result also holds when I0 is nonempty by repeating the above argument and applying Lemma B.2,\" with no accompanying calculation. This is a load-bearing gap. Lemma 5.2 supplies the uniform moment bound (5.11) that is used in Section 5.3.1 to pass from mean empirical measures to ergodic occupation measures of the limiting diffusion, and hence to prove the lower bound (5.10) of Theorem 3.3. Since Assumption 2.1 only requires gamma_d>0, the case I0 nonempty is allowed by the theorem's hypotheses. The omitted case is exactly where the modified priority policy of Definition 4.1 and the special Lyapunov terms for I0 in (4.17) are needed; for i in I0 the abandonment term gamma_i q_i vanishes, so the estimate (B.20) for the polynomial part over arbitrary policies does not follow from the given argument. The sentence \"repeating the above argument\" does not constitute a proof, and no analogue of (B.23) is displayed for I0 nonempty.","section":"Appendix B, Proof of Lemma 5.2, final paragraph"},{"comment":"Theorem 5.3 is stated for any sequence of policies satisfying (5.11), but the only mechanism in the paper for verifying (5.11) in the lower-bound argument is Lemma 5.2. Consequently the gap described above also invalidates Theorem 5.3 for sequences of policies when zero-abandonment classes are present. The proof of (5.10) takes an arbitrary sequence with sup_n J_hat(X_hat^n(0), z_n) < infinity, applies Lemma 5.2 and Theorem 5.3, and concludes that the limit of the mean empirical measures is in G. Without a complete Lemma 5.2, the lower bound, and hence the equality in Theorem 3.3, is not established in the parameter region I0 nonempty. The authors should either supply the missing calculation or state Theorem 3.3 under the stronger assumption gamma_i>0 for all i and discuss what fails in the general case.","section":"Section 5.3.1, Theorem 5.3 and proof of (5.10)"}],"minor_comments":[{"comment":"The bounded mean residual life condition is written as a displayed quotient whose denominator 1-F(t) can be zero when the distribution has bounded support; please state explicitly that the inequality is required only where the denominator is positive, or interpret the condition through the usual limiting convention.","section":"Assumption 3.2(ii), equation (3.10)"},{"comment":"The formula for the modified priority policy uses the denominator sum_{i in I0} rho_i; when I0 is empty this expression is undefined. Please add a convention, for example that the first block of the definition is vacuous when I0 is empty.","section":"Definition 4.1"},{"comment":"There is a duplicated phrase \"by by Assumptions 2.1 and 2.2\" in the sentence following (B.16); please correct the typographical error.","section":"Appendix B, around (B.16)"},{"comment":"The proof of Lemma 5.5 is delegated to [5, Lemma 7.2] with the sentence \"the proof of this lemma is the same as that of Lemma 7.2 in [5].\" Given that the present model has renewal age processes and compound-Poisson jumps, the transfer is not completely immediate; please spell out how the age-process terms and the jump operator are handled, or state which estimates in Sections 5.3.1 and 4 make the proof identical.","section":"Lemma 5.5"},{"comment":"The notation O(g) is used both for a function space and for a generic member of that space; this can confuse an inequality such as \"= O(1/sqrt(n))(||x||+||q||)\". It would be clearer to write explicit constants or use a distinct symbol for the space.","section":"Section 1.2, Notation"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically strong and the main ideas are interesting, but the stated proof of the ergodic lower bound is incomplete because Lemma 5.2 is not proved for the allowed case I0 nonempty. This is a localized but load-bearing gap. If the authors can supply the missing calculation, I would support acceptance. The heavy dependence on the companion preprint [17] is appropriate as a dependency rather than a circularity, but the editor may wish to verify that [17] is available and accepted before final decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe headline: this is a serious paper that extends heavy-traffic scheduling theory from Markovian arrivals without interruptions to renewal arrivals in an alternating up-down environment. The FCLT in Theorem 3.1 is new and appears correctly proved, and the discounted asymptotic optimality in Theorem 3.2 follows a sound route. The ergodic theorem (3.3) is the main claim, and I think it is likely true, but the proof as written has a hole in a load-bearing lemma.\n\nWhat is genuinely good: the construction of an augmented Markov process with renewal ages and residual downtime, the generator convergence in Lemmas 5.3 and 5.4, and the modified priority policy that explicitly handles classes with zero abandonment. The moment bounds in Theorem 4.1 are nontrivial. The paper also states its own limitations clearly—the omitted case in Lemma 5.2 is flagged in the text, which is honest but does not remove the gap.\n\nWhere it is soft: the proof of Lemma 5.2 in Appendix B only treats the case I0 empty. The authors write that the nonempty case follows 'by repeating the above argument and applying Lemma B.2.' That is not automatic. When I0 is nonempty, the drift term gamma_i q_i vanishes for those classes, and the Lyapunov function in (4.17) has a special construction precisely there. The estimates (B.19)–(B.22) are tailored to the case where every class has positive drift toward the origin. Repetition is not a proof unless the service allocation controls the zero-abandonment classes uniformly. This is load-bearing for the lower bound (5.10). If a referee cannot fill the gap, the ergodic theorem is not established for a legitimate parameter regime. This is likely a fixable omission rather than a fatal flaw, but as it stands the paper promises more than it proves.\n\nThe dependency on [17] is a real dependency; the limiting-diffusion optimality theory is imported from that companion. That is not circular, because [17] is a general structural result, but it means the present paper's ergodic result is only as solid as [17].\n\nWho should read it: researchers in queueing control, heavy-traffic theory, and stochastic operations. It deserves a serious peer review—I would send it to referees, with instructions to verify Lemma 5.2 or require the authors to supply the full argument. My recommendation: conditional accept, contingent on a complete proof of that lemma.","headline":"A genuinely new FCLT and asymptotic optimality for multiclass queues with renewal arrivals and service interruptions, but the ergodic proof has a flagged gap in Lemma 5.2 that needs closing.","tokens_in":37226,"tokens_out":2792,"would_cite":true,"duration_ms":27132,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90B22","90B36","60K37","60K25","60J75","60F17"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that, in the Halfin-Whitt regime, the optimal scheduling value of a multiclass many-server queue with renewal arrivals and service interruptions converges to the optimal value of a limiting compound-Poisson…","keywords":["multiclass many-server queues","Halfin-Whitt regime","service interruptions","renewal arrivals","alternating renewal process","jump diffusions","ergodic control","asymptotic optimality"],"falsifier":"Take interarrival times with finite moments of all orders but unbounded mean residual life, for instance a Weibull distribution with shape $1/2$ scaled to mean $1$, run the diffusion-scaled system under the modified priority policy, and check whether the long-run average moment bound (4.19) still holds; if $\\sup_n \\limsup_{T\\to\\infty} \\frac{1}{T}E\\int_0^T |\\hat X^n(s)|^\\kappa\\,ds$ diverges, then Theorem 4.1, and with it Theorem 3.3, fails in that case.","tokens_in":36207,"feed_emoji":"🔄","tokens_out":11375,"duration_ms":101993,"temperature":0.7,"pith_summary":"This paper studies scheduling control for large multiclass service systems with renewal arrivals and random service interruptions that alternate between up and down states, in the Halfin-Whitt critical-loading regime. It tries to establish that the optimal discounted cost and the optimal long-run average cost of the diffusion-scaled queue converge, as the system grows, to the corresponding optimal costs of a tractable limiting controlled jump diffusion driven by Brownian noise and a compound Poisson process. If true, operators can base scheduling decisions on this simpler limiting control problem and be assured that the real queue's performance approaches the optimum. The main obstacle is that the queue-length counting process is not Markov; the paper overcomes it by augmenting the state with the ages of the renewal arrivals and of the down times and proving convergence of the augmented generators.","feed_headline":"Interrupted queues: optimal scheduling matches a jump-diffusion limit","feed_subtitle":"Discounted and long-run average costs both converge to the optimum of a controlled compound-Poisson jump diffusion.","key_machinery":"Three mechanisms carry the argument. First, the non-Markovian queueing process is augmented with the age processes of the renewal interarrival times and of the down state, which makes the state Markov; convergence of the generators to the generator of the limiting jump diffusion is proved with test functions $\\varphi_n[f]$ built from residual-life functions $\\eta_i^n$ and $\\kappa_i^n$. Second, long-run average moment bounds for the diffusion-scaled processes are obtained from Foster-Lyapunov inequalities for this augmented process under a modified priority policy, using Lyapunov functions with renewal residual-life corrections. Third, asymptotic optimality is assembled from a lower bound via convergence of mean empirical measures to ergodic occupation measures and an upper bound via a spatial-truncation, concatenated scheduling policy that implements an $\\epsilon$-optimal control of the limit inside a compact set and the modified priority policy outside.","core_discovery":"The central claim is Theorem 3.2 and Theorem 3.3. Under the Halfin-Whitt assumptions and the bounded mean residual life condition, if the initial scaled state $\\hat X^n(0)$ converges to $x$, then the optimal discounted value $\\hat V_\\alpha^n(\\hat X^n(0))$ converges to $V_\\alpha(x)$, and the optimal ergodic cost $\\hat\\rho_n(\\hat X^n(0))$ converges to $\\rho^*$, where $V_\\alpha$ and $\\rho^*$ are the optimal values of the controlled jump diffusion $$dX_t = b(X_t,U_t)dt + \\Sigma dW_t + \\$\\lambda$ dL_t$$ with drift $$b(x,u)=\\ell - M(x-\\langle e,x\\rangle^+ u)-\\langle e,x\\rangle^+ \\Gamma u,$$ diffusion coefficient $\\Sigma=\\operatorname{diag}(\\sqrt{\\lambda_i(1+c_{a,i}^2)})$, and $L$ a compound Poisson process representing accumulated downtime. The paper establishes that scheduling policies for the original queue can be chosen so that their performance approaches the optimum of this limiting control problem, for both discounted and ergodic criteria.","pith_inferences":["My inference: the same augmented-generator scheme should extend to other regenerative environments, such as Markov-modulated up-down cycles whose rates depend on $n$, with the jump-diffusion limit's Lévy measure changed accordingly, provided a bounded residual-life condition holds.","My inference: the bounded mean residual life condition could plausibly be relaxed to polynomial growth of residual-life functions at the cost of higher-degree Lyapunov functions, suggesting a threshold effect where heavy-tailed interarrival or downtime distributions eventually destroy uniform moment bounds under fixed priority policies.","My inference: one could numerically test convergence rates by comparing finite-$n$ costs under the policy induced by the limit HJB against the predicted value, although the paper does not quantify the rate of convergence."],"forward_implications":["The optimal cost of the original scheduling problem is asymptotically computed by solving the HJB equation of the limiting jump diffusion; stationary Markov optimal controls of the limit induce asymptotically optimal scheduling policies for the queue.","Non-exponential renewal arrivals enter the limit only through their squared coefficients of variation, so higher-order interarrival distribution details are washed out in the Halfin-Whitt scaling.","Asymptotically negligible service interruptions add an independent compound-Poisson noise term to the limiting dynamics, so their effect on optimal cost can be quantified and priced into the limiting control problem.","A concrete modified priority policy yields uniform long-run average moment bounds; outside a compact set, running this policy and inside it following the limit's $\\epsilon$-optimal control achieves cost arbitrarily close to $\\rho^*$ for large $n$."],"supporting_citations":[{"why":"supplies the template for the discounted asymptotic-optimality proof, including the moment estimate in Lemma 3 that becomes Lemma 5.1.","marker":"[1]"},{"why":"supplies the spatial-truncation technique, the modified priority policy, and the Foster-Lyapunov argument reused for the ergodic upper bound.","marker":"[4]"},{"why":"supplies the alternating-renewal interruption model and the weak limit of the scaled downtime process as a compound Poisson process.","marker":"[13]"},{"why":"supplies the HJB characterization and epsilon-optimal controls for the limiting controlled jump diffusion used in Theorems 5.1 and 5.2.","marker":"[17]"},{"why":"supplies the Lyapunov-function construction with residual-life corrections for renewal age processes.","marker":"[19]"},{"why":"supplies the Foster-Lyapunov criterion used to prove positive Harris recurrence of the augmented Markov process.","marker":"[20]"},{"why":"supplies exponential ergodicity of the limiting jump diffusion under constant control, used in Proposition 4.1.","marker":"[24]"},{"why":"supplies the mean empirical measure convergence lemma invoked for concatenated policies in the ergodic upper bound.","marker":"[5]"}],"fun_headline_variants":["Optimal scheduling in random down-times matches a jump-diffusion limit","Queue scheduling under intermittent service: a jump-diffusion limit","Halfin-Whitt queues with random downtimes: optimal control convergence","Convergent optimal policies for multiclass queues in alternating environments","When service is interrupted: optimal scheduling converges to jump diffusions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 3.2(ii), the bounded mean residual life condition (3.10): the mean residual life of each interarrival time and of the downtime variable is bounded by a constant over all horizons, and without it the paper provides no alternative bound, so the ergodic asymptotic optimality theorem would not be established.","fun_headline_variants_meta":{"raw":{"variants":["Optimal scheduling in random down-times matches a jump-diffusion limit","Queue scheduling under intermittent service: a jump-diffusion limit","Halfin-Whitt queues with random downtimes: optimal control convergence","Convergent optimal policies for multiclass queues in alternating environments","When service is interrupted: optimal scheduling converges to jump diffusions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000749,"raw_usage":{"total_tokens":3362,"prompt_tokens":996,"completion_tokens":2366,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":2276}},"tokens_in":612,"tokens_out":2366,"duration_ms":16547,"temperature":1.0,"reasoning_tokens":2276,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:49:27.199321+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take interarrival times with finite moments of all orders but unbounded mean residual life, for instance a Weibull distribution with shape $1/2$ scaled to mean $1$, run the diffusion-scaled system under the modified priority policy, and check whether the long-run average moment bound (4.19) still holds; if $\\sup_n \\limsup_{T\\to\\infty} \\frac{1}{T}E\\int_0^T |\\hat X^n(s)|^\\kappa\\,ds$ diverges, then Theorem 4.1, and with it Theorem 3.3, fails in that case.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the template for the discounted asymptotic-optimality proof, including the moment estimate in Lemma 3 that becomes Lemma 5.1."},{"cited_title":"Arapostathis, A","cited_arxiv_id":null,"evidence_quote":"supplies the spatial-truncation technique, the modified priority policy, and the Foster-Lyapunov argument reused for the ergodic upper bound."},{"cited_title":"Pang and W","cited_arxiv_id":null,"evidence_quote":"supplies the alternating-renewal interruption model and the weak limit of the scaled downtime process as a compound Poisson process."},{"cited_title":"Ergodic control of diffusions with compound Poisson jumps under a general structural hypothesis","cited_arxiv_id":"1908.01068","evidence_quote":"supplies the HJB characterization and epsilon-optimal controls for the limiting controlled jump diffusion used in Theorems 5.1 and 5.2."},{"cited_title":"Konstantopoulos and G","cited_arxiv_id":null,"evidence_quote":"supplies the Lyapunov-function construction with residual-life corrections for renewal age processes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Foster-Lyapunov criterion used to prove positive Harris recurrence of the augmented Markov process."},{"cited_title":"Arapostathis, G","cited_arxiv_id":null,"evidence_quote":"supplies exponential ergodicity of the limiting jump diffusion under constant control, used in Proposition 4.1."},{"cited_title":"Arapostathis and G","cited_arxiv_id":null,"evidence_quote":"supplies the mean empirical measure convergence lemma invoked for concatenated policies in the ergodic upper bound."}],"review_version":1}