{"id":"13df9cad-1b4f-4c88-a106-0f48c2cb4415","arxiv_id":"2608.12073","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Stock-based threshold policies, which set prices only by remaining inventory, achieve computable multi-unit prophet-inequality guarantees (e.g., 0.6269 for K=2 and 0.6816 for K=3) in continuous time with nonhomogeneous Poisson arrivals.","lead":"Using only the number of items left, not the time on the clock, a seller can still lock in a guaranteed fraction of what a clairvoyant benchmark would earn. The paper proves how to compute those fractions for inventory-based threshold policies and gives numbers such as 0.6269 for two units and 0.6816 for three units.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported ratios depend on the single-interval lemma for the first threshold level; its unverified Pontryagin proof is the critical spot, and a failure would overstate the true worst case.","rationale":"The reader's weakest assumption is the structural characterization in Theorem 4, especially the single-interval claim for D_2 and the island bounds for deeper levels. My read of the proof in Appendices C and D confirms that this is the most load-bearing link: the advertised numbers for both m=2 and m=3 are produced by programs obtained from exactly this structural reduction, and the reduction is supported solely by a long regularized optimal-control argument. I do not see an outright contradiction in the argument, and many parts are internally consistent, but the proof's complexity and lack of independent verification make it the right place to probe. The proposed R=2 versus R=1 continuous-time test directly isolates the single-interval claim without relying on discretization error, so it would settle whether the concern lands. Since this is the same concern that motivated the CONDITIONAL verdict, no change to the reader's verdict is needed; the paper should remain CONDITIONAL pending the check.","tokens_in":49398,"tokens_out":28105,"duration_ms":264520,"concrete_test":"For K=2, m=2, s2=1, formulate the exact continuous-time analogue of Theorem 3 with beta^2 allowed to be active on R=2 disjoint intervals instead of one (variables: S, Lambda_1, two interval endpoints, and the optional late-filled alpha masses). Solve this finite-dimensional non-convex program by multi-start local optimization plus a coarse global scan. If its minimum is strictly below min_{R=1} (the value behind 0.6269), Lemma 7 is false and the table entry is not a valid lower bound. Repeat with K=3, m=3 allowing D_2 on two intervals and/or D_3 on three islands to test Lemma 9 and the J_3<=2 bound. Equality within numerical tolerance supports the reduction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Tables 2 and 3 are valid lower bounds on gamma* requires PoisOPTRe_K(s) to be exactly represented by the finite-dimensional programs. The keystone is the single-interval characterization of the first nontrivial threshold level: Lemma 7 for m=2, and Lemma 9 for general m (with the m=3 case additionally using J_3<=2 from Lemma 8). These lemmas are proved by a regularized Pontryagin minimum principle: the switching gap Phi is shown to be strictly convex on mode-1 intervals, which forbids two separated active components of D_2, and Claims 6-7 transfer the costate representation into the island bound. The argument is long, not machine-checked, and must cover degenerate cases (Lambda_1=0, Lambda_1=S, zero-length islands, singular arcs where chi_i=0) plus normality of the PMP multipliers. If any step fails, for example if Phi is only convex rather than strictly convex on a singular arc, or if an abnormal multiplier appears at a boundary, then an optimal adversary could use two or more separated beta^2-active islands. In that case the 4-variable m=2 and 9-variable m=3 programs are strict restrictions of the adversary's problem: their minima exceed the true PoisOPTRe_K(s), so the reported ratios would not be valid lower bounds. No code, global-optimality certificate, or independent verification is provided; Appendices C-D are the sole support for the structural reduction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies a K-unit online resource allocation problem with nonhomogeneous Poisson arrivals and time-varying valuation distributions. It analyzes stock-based threshold policies whose thresholds depend only on remaining inventory, comparing them against the K-best hindsight prophet benchmark. The authors formulate a continuous-time type-covering dual, reduce the original minimax problem to a finite-constraint Poisson optimization problem PoisOPTRe_K(s) (Theorem 1), and then prove structural results: extremal/cutoff arrival structures (Theorem 2), a four-parameter exact formulation for m=2 (Theorem 3), and a finite nested-interval representation for general m (Theorem 4). Numerical lower bounds are reported in Tables 2–3, e.g., 0.6269 for K=2, m=2 and 0.6816 for K=3, m=3.","tokens_in":49689,"tokens_out":13445,"duration_ms":145966,"significance":"If the structural theorems and numerical claims are correct, the paper makes a substantial contribution: it converts an infinite-dimensional nonconvex adversarial problem into small finite-dimensional programs and provides quantitative evidence that simple inventory-based policies achieve strong guarantees. The derivations are self-contained and contain no fitted parameters; the structural claims are falsifiable and the ratios can be checked by simulation. The main caveat is that the final numbers depend on intricate optimal-control arguments and on unverified numerical global optimization, so the current evidence is not yet conclusive.","major_comments":[{"comment":"The reported ratios are asserted as lower bounds on γ*, computed by minimizing non-convex finite-dimensional programs via grid search (§5 and §6.3.1). For a minimization problem, the minimum over a finite grid is an upper bound on the true minimum, not a lower bound; hence the described computational method cannot certify that the reported values are valid lower bounds on PoisOPTRe_K(s) and therefore on γ*. The paper should either provide rigorous global-optimality certificates for each reported value (e.g., interval branch-and-bound, Lipschitz-based bounds, or exact analytical minimization), or supply verifiable code that certifies the lower bound, or explicitly weaken the claims to heuristic estimates.","section":"§6.3, Tables 2–3"},{"comment":"The proof that β-controls can be taken extremal replaces controls on subintervals by chattering between extreme points and then defines α̂ 'according to the conditions described in Lemma 4 with respect to û', asserting that the switching times ŵ_i equal the original w_i. Since Lemma 4's switching time is determined by the β-control and the target total mass, this equality is not automatic; the text does not establish that ∫ α̂^i = ∫ α^i, or even the approximation needed for the stated O(1/N) error bound. This step underpins Theorem 2 and therefore the whole structural reduction, so it needs to be made rigorous or replaced by a different argument.","section":"Lemma 5, Appendix B.4"},{"comment":"The finite-dimensional reduction for general m is conditional on a long regularized Pontryagin argument, in particular Lemmas 8–9 and Claims 6–9. The text handles singular arcs, abnormal multipliers, boundary cases, and zero-length islands in summarized form, and no machine-checked proof or independent verification is supplied. Since the m=3 ratios in Table 3 depend on the J_3≤2 island bound and on the exactness of the 9-variable program, a failure in any of these claims would undermine the reported numbers. I ask for a fully detailed proof with explicit treatment of all degenerate cases, or an independent verification of the island bounds (e.g., a formalized proof or a separate derivation).","section":"Theorem 4, Appendix D"}],"minor_comments":[{"comment":"The text says the reduced problem 'only contains 3 free variables', but (15) defines four variables S, a, b, Λ1; please correct this inconsistency.","section":"§5, before (15)"},{"comment":"The phrase 'as plotted in .' has a missing figure reference; Figure 1 should be cited explicitly.","section":"§3.1"},{"comment":"There are several typographical issues ('exprssion', 'almost entirely' instead of 'almost everywhere', inconsistent notation 'PoisOPTRe_K(s)' vs 'PoisOPTReK(s)'); a careful proofreading pass is needed.","section":"Appendix D"},{"comment":"In the displayed inequality before (29), the right-hand side contains a stray 'dt' after x^{k'}_1; the intended expression is the sum of terminal state probabilities.","section":"Lemma 3 proof, Appendix A.3"},{"comment":"A brief explanation of how the m=1 ratios are obtained from Chawla et al. (2024) and Jiang et al. (2025b) would help readers interpret the comparison.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The numerical claims in Tables 2–3 are central to the paper's impact, and at present they are not backed by reproducible code or global-optimality certificates. Requiring the authors to provide certified numerical methods (or to clearly reclassify the numbers as heuristic) seems essential before publication. The structural PMP proofs are long and currently difficult to verify; an independent proof or formalization of Lemmas 8–9 would substantially increase confidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a real advance, not incremental. It gives a principled reduction of continuous-time stock-based thresholds to a finite Poisson optimization problem, and the structural characterizations (Theorem 2, Lemmas 5, 7, 9) are exactly the kind of results that make this line of work useful. The m=2 reduction to four scalar variables is convincing and relatively transparent. The m=3 nested-island bound is more exotic but is actually argued in detail in the appendix rather than handwaved.\n\nThe advertised ratios (0.6269 for K=2, 0.6816 for K=3) are not proven optima of the reduced programs: they come from grid search on a non-convex problem, with no global-optimality certificate and no code. That matters because the finite-dimensional program is only exact if the structural theorem holds at every degenerate or singular arc. Lemma 7 (single interval for the first nontrivial level) and Lemma 9 carry the load; the proofs are long, use regularized Pontryagin arguments, and are not machine-checked. I did not find a clear error on a careful read, but this is exactly the kind of claim where small gaps in regularized PMP arguments are common. Independent verification, either by code or by a focused check of Claims 6 and 8, should be required before the numbers are quoted as established.\n\nA separate, smaller issue: Table 1 compares the m=2 and m=3 ratios against \"fully adaptive\" ratios from Jiang et al. (2025a), which are benchmarked through an LP relaxation rather than the prophet benchmark. The footnote flags this, but the table as printed invites the wrong reading. That should be fixed; it is not a fatal flaw.\n\nThe citation pattern looks honest. Prior work by Chawla et al. (2024) and Jiang et al. (2025b) is used as baseline and inspiration, not as an unjustified assumption of the target result. There is no circularity.\n\nWho is this for? Researchers in prophet inequalities, revenue management, and simple-policy analysis. The framework is significant even if the table values later need adjustment. It deserves a serious referee: the reduction and structural results are substantial enough to justify referee time, and the open questions about the numerical certification are addressable in revision.","headline":"A serious, unusually complete prophet-inequality paper whose computational banner numbers depend on a long, unverified optimal-control argument; worth refereeing, but the headline numbers need independent verification.","tokens_in":50214,"tokens_out":1932,"would_cite":true,"duration_ms":21947,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90B05","49K15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Stock-based thresholds with two price levels recover 0.6269 of the prophet's reward in the two-unit case, and three thresholds recover 0.6816 in the three-unit case, in continuous time with nonhomogeneous Poisson arrivals.","keywords":["prophet inequalities","stock-based thresholds","competitive ratio","Poisson arrivals","resource allocation","type-covering dual","bang-bang control","continuous time"],"falsifier":"Solve the nine-variable program for $m=3$, $K=3$ with a certified global optimizer: a value below 0.6816 would refute the claimed lower bound. Alternatively, search for a feasible $\\mathrm{PoisOPTRe}_3$ instance whose optimal $\\beta$ schedule has three $D_3$ islands inside the $D_2$ interval; the structural theorem forbids it.","tokens_in":1750,"feed_emoji":"📈","tokens_out":1937,"duration_ms":75465,"temperature":0.7,"pith_summary":"The paper asks how much performance is lost when an online seller ignores calendar time and posts thresholds that depend only on how many units remain. Its central claim is that such stock-based threshold policies with two thresholds achieve a competitive ratio of 0.6269 against the multi-unit prophet benchmark when $K=2$, and three thresholds achieve 0.6816 when $K=3$. The proof works by reducing the infinite-dimensional worst-case problem to a finite-dimensional Poisson optimization and then characterizing the worst-case arrival pattern as an extreme cutoff schedule with nested intervals. If correct, the result means simple booking-limit rules that are easy to communicate and robust to demand misestimation can capture most of the value of full hindsight knowledge.","feed_headline":"Two stock prices capture 62.7% of a hindsight prophet's reward","feed_subtitle":"A three-level stock-based menu pushes the guarantee to 0.6816 for K=3, without watching the clock.","key_machinery":"The key object is the reduced Poisson optimization $\\mathrm{PoisOPTRe}_K(s)$, a finite-constraint non-convex control problem whose value lower-bounds the competitive ratio of stock-based policies. Its structural characterization is the load-bearing mechanism: worst-case $\\beta$ controls are cutoff-valued, $\\alpha$ controls are late-filled, and, after fixing $\\alpha$ masses, the first nontrivial $\\beta$ level is active on one interval while level $i$ has at most $4(m-i+1)+2$ active islands inside the previous level's active time. This yields a finite-dimensional endpoint representation with at most $4m^2-8m-5$ timing variables, turning the original continuum adversary into a small grid-search program.","core_discovery":"The central discovery is that the adversarial problem of minimizing the competitive ratio of stock-based threshold policies has a finite-dimensional core. After a type-covering dual reformulation, the worst-case instance reduces to $\\mathrm{PoisOPTRe}_K(s)$, where the adversary picks arrival-rate functions over time; despite being non-convex, this reduced problem has an optimal solution with a bang-bang structure: at every moment the arrival rates across threshold levels form a cutoff extreme point, intermediate rates are late-filled, the first nontrivial level is active on a single interval, and each deeper level is active on a bounded number of nested islands. This shows the infinite-dimensional search over arrival patterns collapses to at most $O(m^2)$ scalar variables independent of $K$, and solving those finite programs gives the reported guarantees. For $m=2$ the reduction is an exact four-parameter problem; for $m=3$ it is a nine-variable program.","pith_inferences":["Because the structural theorem is independent of the horizon length and the valuation model, an immediate extension would be to reusable-resource or matching settings where inventory state still drives decisions.","The bang-bang characterization suggests the worst-case adversary concentrates its harm: valuations that fill thresholds late and abrupt cutoffs, meaning misspecified demand forecasts may be less damaging than time-varying value distributions.","A direct testable extension is to compute the $m=4$ guarantee for moderate $K$ using the same endpoint program; the island bounds predict the optimal controls remain sparse.","One could also iterate the $D_2$ single-interval argument to simplify the outer levels, possibly reducing the $O(m^2)$ variable count further."],"forward_implications":["If the claim is right, a seller who posts only two inventory-contingent prices is guaranteed at least 0.6269 of the best-$K$ hindsight reward for $K=2$, and larger $K$ levels push this above 0.76.","The same reduction certifies that ignoring calendar time entirely is nearly free: the loss from stock-based thresholds is bounded by the gap between the computed numbers and the prophet benchmark, not by the number of price levels.","The finite-parameter representation makes the guarantees computable for any $m$ and $K$ by deterministic grid search, with no discretization of the original continuous-time adversary.","The results also apply to posted pricing via virtual-valuation transfer, so a posted menu of $K$-inventory-contingent prices captures the same fraction of optimal revenue.","The structural bound is uniform in $K$, so adding thresholds improves guarantees predictably: each extra stock-level threshold adds a fixed number of timing variables."],"supporting_citations":[{"why":"Provides the static-threshold ($m=1$) ratios and the $1-\\Theta(\\sqrt{\\log K/K})$ scaling that this paper improves on with two and three thresholds.","marker":"Chawla et al. (2024)"},{"why":"Supplies the type-covering dual formulation and the proof that the $m=1$ guarantee is tight, which the current dual reduction builds on.","marker":"Jiang et al. (2025b)"},{"why":"Gives the best known fully adaptive multi-unit guarantees used as a benchmark in Table 1.","marker":"Jiang et al. (2025a)"},{"why":"Establishes the continuous-time Poisson revenue-management model whose arrival process this paper adopts.","marker":"Gallego and Van Ryzin (1994)"},{"why":"The KKM lemma is used to show every threshold interval contains a binding type, enabling the reduction to a finite-constraint Poisson program.","marker":"Knaster et al. (1929)"},{"why":"Provides the existence theorem for optimal control with compact convex control sets used in the structural analysis.","marker":"Fleming and Rishel (2012)"},{"why":"Supplies the Pontryagin minimum principle and endpoint constraint qualification used to derive the switching-gap structure.","marker":"Vinter (2000)"},{"why":"The variation-diminishing theorem for convolution kernels bounds the number of active islands in the nested beta schedule.","marker":"Schoenberg (1948)"}],"fun_headline_variants":["Stock-only thresholds hit 62.7% of hindsight prophet reward","Two stock thresholds capture 62.7% without timing knowledge","Prophet inequality: stock thresholds achieve 0.6269 for K=2","No clock needed: stock thresholds give 62.7% prophet guarantee","Stock thresholds: 62.7% (K=2), 68.2% (K=3) without clock"],"cache_read_input_tokens":52352,"weakest_assumption_plain":"The structural characterization of the reduced Poisson program—that worst-case beta controls are bang-bang, with $D_2$ on one interval and a bounded number of nested islands per level, established through a long Pontryagin argument with regularization—is the load-bearing premise; if any switching-gap or variation-diminishing claim fails on a singular arc, the finite-dimensional programs and the reported ratios are not justified.","fun_headline_variants_meta":{"raw":{"variants":["Stock-only thresholds hit 62.7% of hindsight prophet reward","Two stock thresholds capture 62.7% without timing knowledge","Prophet inequality: stock thresholds achieve 0.6269 for K=2","No clock needed: stock thresholds give 62.7% prophet guarantee","Stock thresholds: 62.7% (K=2), 68.2% (K=3) without clock"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001298,"raw_usage":{"total_tokens":5338,"prompt_tokens":1030,"completion_tokens":4308,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":4202}},"tokens_in":646,"tokens_out":4308,"duration_ms":30285,"temperature":1.0,"reasoning_tokens":4202,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:17:31.032307+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the nine-variable program for $m=3$, $K=3$ with a certified global optimizer: a value below 0.6816 would refute the claimed lower bound. Alternatively, search for a feasible $\\mathrm{PoisOPTRe}_3$ instance whose optimal $\\beta$ schedule has three $D_3$ islands inside the $D_2$ interval; the structural theorem forbids it.","supporting_citations":[],"review_version":1}