{"id":"0f0c1121-718e-427c-a579-d4ba5c68f909","arxiv_id":"1908.09295","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"In a two-class stock-rationing queue, the optimal dynamic rationing policy is claimed to be threshold after reordering states by a policy-dependent index, but the middle-region proof is not fully supplied.","lead":"Using performance difference equations, this paper derives an algebraic characterization of the optimal dynamic rationing policy for a two-class stock-rationing queue. The authors claim the optimal policy is always of 'transformational threshold type' and give three sufficient conditions for an ordinary threshold policy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 10 inherits from Theorem 9, whose proof assembles the optimal policy coordinate-wise from thresholds P_i^{(d)} of a fixed reference policy; this is invalid because π(d') and G^{(d')} change with every coordinate, so the wildcard sub-policies are not admissible policies.","rationale":"Good-faith reading: the paper develops a coherent sensitivity-based optimization framework, gives an explicit Poisson-equation solution, and derives two extreme-region results conditional on a class property. These are useful components. But the abstract's main claim, a 'complete algebraic solution,' and Theorem 10 are not established. The decisive gap is that Theorem 9's proof treats coordinates as independent even though the stationary distribution and perturbation factors are globally coupled. This is exactly the weakest assumption identified by the reader. My independent inspection shows the same coupling flaw appears earlier: the induction in Theorem 3 (Section 7) transfers a sign only at the flipped coordinate, then generalizes it to all elements of S(d,c). Under the per-policy threshold reading used in the proof, the sign-propagation claim can even fail in a small K=2, N=3 example. The proposed exhaustive enumeration settles whether the theorem itself is false or merely unproven: if the constructed policy wins for every (d,P) in the enumeration, the claim survives the test though a correct proof would still be required; if it loses for a single (d,P), the central claim is false. The numerical experiment in the paper does not repair this gap because it omits the all-zero threshold and does not enumerate the full policy set. Therefore the reader's rejection stands.","tokens_in":45142,"tokens_out":21839,"duration_ms":213108,"concrete_test":"Exhaustively enumerate all 2^K policies for small instances (e.g., K=4, N=8, and also K=15, N=100 using the parameters of Section 9, which is only 32768 policies) using the paper's own closed-form formulas (5)-(6) and (17). For each reference policy d in D and each P in the per-policy middle interval (PL(d), PH(d)), construct the Theorem 9 prediction: sort {1,...,K} by P_i^{(d)}, choose n0 with P_{i_{n0-1}} < P <= P_{i_{n0}}, set transformed coordinates i_1,...,i_{n0-1} to 0 and i_{n0},...,i_K to 1, and invert the permutation to obtain d_pred. Compare eta_{d_pred} to the exact arg max over D. Any (d,P) for which d_pred is not optimal falsifies Theorems 9-10. In addition, rerun Example 2 with the all-zero threshold theta=K+1 included and with full dynamic enumeration rather than comparing only two extreme policies.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 10) inherits entirely from Theorem 9, which asserts that in the middle penalty region, after sorting states by the P_i^{(d)} values of an arbitrary reference policy d, the optimal policy is a single threshold. The proof of Theorem 9 splits coordinates into two sets and claims that, because P exceeds the maximum P_i^{(d)} on one set and is below the minimum on the other, the optimal 'sub-policies' are respectively all-0 and all-1, which are then recombined. This is not a valid derivation. First, objects of the form (0; d_{i1},...,d_{in0-1}, ∗,...,∗; 1,...,1) are not policies in D: no stationary distribution, perturbation factor, or long-run average profit is defined for them. Second, and more fundamentally, P_i^{(d)} is the root of G^{(d)}(i)+b=0 for the fixed policy d. Once any coordinate is changed, π(d') and G^{(d')}(i) change, so the root for the modified policy is not P_i^{(d)}, and the sign of G^{(d')}(i)+b at the same P is not controlled by P_i^{(d)}. The performance difference equation (39) only gives local information for a single-coordinate change from the reference policy under which G is evaluated. The same coupling problem appears in Theorem 3: Eq. (50) transfers the sign of G+b only at the single coordinate where the two policies differ, but the proof then asserts the sign at every coordinate of S(d,c) ('for each j1 ...'), which is unjustified. Thus the 'complete algebraic solution' rests on an unproven coordinate-wise separability assumption, and no valid proof of Theorem 10 is supplied.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies a single-product, two-class stock-rationing queue with Poisson arrivals, exponential service times, lost sales, and a penalty cost P for serving Class 2 at low inventory. Using sensitivity-based optimization, the authors derive a performance difference equation (Lemma 2) and define P_i^(d) as the root of G^(d)(i)+b=0 for each policy d and state i. They divide the penalty range into P ≥ PH(d), 0 ≤ P ≤ PL(d), and PL(d) < P < PH(d). The paper claims that in the first two regions the optimal policy is the all-0 or all-1 threshold policy, and in the middle region the optimal policy is of 'transformational threshold type': after permuting states by the order of P_i^(d), the optimal policy is a single threshold. This is summarized as Theorem 10 and advertised as a complete algebraic solution to the optimal dynamic rationing problem.","tokens_in":45545,"tokens_out":9264,"duration_ms":85757,"significance":"If the main theorem were correct, it would give a structural characterization of optimal rationing policies that goes beyond the threshold results obtained by MDP/submodularity methods, and it would demonstrate the usefulness of sensitivity-based optimization for inventory rationing. The paper does contain some correct and useful building blocks: the policy-based birth-death description, the explicit stationary distribution, the linearity of the long-run average profit in P, and the single-coordinate performance difference equation are standard but carefully derived. The explicit formulas for the profit of static policies are also a useful computational reference. However, the central structural theorem is not established, and the numerical experiment intended to support it omits a key policy, so the claimed significance is not currently realized.","major_comments":[{"comment":"The proof of the middle-region result assumes that coordinates can be optimized independently. The objects ~d_a = (0; d1, ..., d_{n0-1}, *, ..., *; 1, ..., 1) and ~d_b = (0; *, ..., *, d_{n0}, ..., d_K; 1, ..., 1) are not policies in the policy space D: no stationary distribution, perturbation realization factor, or long-run average profit is defined for them. Moreover, P_i^(d) is defined as the root of G^(d)(i)+b=0 for the fixed reference policy d, and once any other coordinate is changed, both pi(d') and G^(d')(i) change; hence the sign of G^(d')(i)+b at the same P is not controlled by P_i^(d). The 'it is easy to see from the proof of Theorem 4' step is therefore not a valid derivation, and Theorems 9 and 10 are left without proof.","section":"Section 7.3, Theorem 9 (and Theorem 8)"},{"comment":"The induction in the proof of Theorem 3 is invalid because the sign-transfer step (the equation relating G^(d(1))(j1)+b to G^(d)(j1)+b) applies only at the single coordinate where the two adjacent policies differ. In Step one, the argument shows only that for the policy d(1) that differs from d at coordinate j1, the sign at that same coordinate j1 is nonpositive. In Step two, to conclude G^(d(2))(j2)+b <= 0, the proof needs G^(d(1))(j2)+b <= 0 at the next coordinate j2, which was never established because d(1) and d differ only at j1. The phrase 'for each j1' does not turn this into a property of one policy at all coordinates of S(d,c). Theorem 3(1) and (2), which are used in Theorems 4 and 6, therefore do not follow.","section":"Section 6.3, Theorem 3"},{"comment":"The numerical scan restricts theta to 1 <= theta <= 15, but the static policy set D_Delta defined in Section 8 includes theta = K+1 = 16. The omitted policy d_Delta,16 is exactly (0; 0, ..., 0; 1, ..., 1), which is the policy d* of Example 1 with eta = 22.3. Since theta = 16 is not included, the observation that eta_d_Delta,9 = 21.4 < eta_d* = 22.3 does not establish that the optimal static rationing policy is suboptimal in D, and it cannot support the conclusion that the optimal dynamic policy is not of threshold type. The numerical demonstration of the paper's central claim is therefore incomplete.","section":"Section 9, Example 2"},{"comment":"The 'transformational threshold' characterization is self-referential and non-constructive. The permutation used to define d*(Transfer) is determined by sorting the values P_i^(d), but P_i^(d) is a function of the stationary distribution pi(d) of the very policy being sorted; thus the theorem describes a property that an optimal policy would satisfy after being sorted by its own P-values, rather than providing an algebraic rule that determines the optimal policy or its permutation without enumerating the 2^K policies. The title's claim of a 'complete algebraic solution' is therefore stronger than what is actually shown.","section":"Section 7.3, Theorem 10 and Section 6.3, Eq. (45)"}],"minor_comments":[{"comment":"PL(d) is written with a max instead of a min in the proof; the displayed definition in Eq. (47) uses min, so the proof should be corrected.","section":"Section 6.3, proof of Lemma 4"},{"comment":"The vertical axis is labeled eta_d* in both figures, but the plotted quantity appears to be eta_d_Delta,theta, the long-run average profit of the static policy indexed by theta; the labels should distinguish the two quantities.","section":"Section 9, Example 2, Figures 3 and 4"},{"comment":"The reward function uses indicator notation 1_{i<N} and 1_{i=N}; this is fine, but the same symbol 1 is later used for both an indicator and the numeral one, which makes equations such as (17) harder to read.","section":"Section 4, Eq. (7)"},{"comment":"The claim that this is 'the first' general solution of Poisson equations with two free constants is not substantiated by a comparison with the cited references; please soften the claim or provide a precise novelty statement.","section":"Section 5, Theorem 1"}],"recommendation":"reject","confidential_remarks":"The main structural theorem (Theorem 10) rests on an invalid coordinate-wise separability argument, and the numerical experiment omits the very threshold policy that would be optimal among static policies. The performance-difference equation and the explicit formulas in Sections 4-6 are standard and appear correct, but they are not sufficient to support the claimed complete solution. A revision would need either a genuinely new proof of the middle-region assertion or a significant weakening of the claims, and the numerical study would need to include the full static policy set. As it stands, I do not see how the stated results can be repaired within the current framework."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the central structural result (Theorems 9–10) is not established. The proof takes a reference policy d, sorts states by the roots P_i^(d) of G^(d)(i)+b=0, and claims that because P lies above all roots on one block and below all roots on another, the optimal sub-policies are all-zero and all-one. But P_i and G change when you change the policy; the roots for a modified policy are not P_i^(d). The performance difference equation only compares policies that differ at one coordinate, and the induction in Theorem 3 tries to transfer sign information across coordinates after policy changes. Step one proves G^(d(1))(j1)+b <= 0 only at the one coordinate where d(1) differs from d; it does not give the sign at other coordinates for the same d(1), so Step two's use of the induction premise is unjustified. The wildcard sub-policies in Theorem 9 are not elements of D, so no stationary distribution or profit is defined for them. Once this goes, Theorem 10 goes with it.\n\nSecond, the numerics do not demonstrate what they claim. In Example 2, K=15 and they scan theta=1..15, omitting theta=16, which is the all-zero threshold policy. They report a static optimum theta*=9 with value 21.4, below the all-zero value 22.3, and conclude the dynamic optimum is non-threshold. But all-zero is itself in the static family; with theta=16 included it would be the best static policy. The example is internally inconsistent.\n\nThe paper is not without merit. The policy-based birth-death model and explicit stationary probabilities are clean, the general solution of the Poisson equation with two free constants is a nice observation, and the performance difference equation is standard. The idea of a 'transformational threshold' is a reasonable way to think about permuting state coordinates. But the central theorem is load-bearing and unproven, and the numerical evidence for the main claim does not hold up.\n\nWho is this for? Readers working on inventory rationing may find the modeling and computational framework useful, but they should not rely on the main theorem. This paper deserves a serious referee because the claims are formal and checkable, and a referee report could pinpoint the exact gap. My recommendation would be reject or major revision; the current version should not be accepted.","headline":"The 'complete algebraic solution' is not proven: Theorem 9 assembles the optimum from thresholds of a fixed reference policy, and the numerical experiment quietly omits the best threshold baseline.","tokens_in":46019,"tokens_out":4261,"would_cite":false,"duration_ms":46634,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90B22","90C40","60J27"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that in a two-class stock-rationing queue the optimal dynamic rationing policy is always a single-threshold policy after inventory levels are reordered by policy-dependent penalty breakpoints.","keywords":["stock-rationing queue","dynamic rationing policy","two demand classes","sensitivity-based optimization","performance difference equation","threshold policy","penalty cost","transformational threshold"],"falsifier":"Enumerate all policies for a small instance with K low-stock levels in the middle penalty region, for example K=5, N=100, $\\lambda$=3, mu_1=4, mu_2=2, C1=1, C2,1=4, C2,2=1, C3=5, C4=1, R=15, with P strictly between PL(d) and PH(d). Compute the long-run average profit of every policy from the birth-death stationary distribution, and also compute the best policy that becomes a block of zeros followed by ones after sorting levels by that policy's own breakpoints P_i^(d); if the unrestricted best profit exceeds the transformed-threshold best profit, the paper's central existence result fails.","tokens_in":44969,"feed_emoji":"📦","tokens_out":9875,"duration_ms":93828,"temperature":0.7,"pith_summary":"The paper studies a warehouse that stocks one product and faces two demand classes: high-priority demand is always served when inventory is positive, while low-priority demand can be served or rejected when inventory is low, with a penalty cost for serving it. It claims that the optimal dynamic rationing policy—the rule for when to reject low-priority demand—always has a threshold structure after a certain reordering of inventory levels. The reordering is by the penalty-cost values at which a marginal comparison for each inventory level changes sign. If true, this gives a complete and computable description of the optimal policy, including the difficult middle penalty range where earlier methods only suggested bang-bang control. The result matters for make-to-stock systems, supply chains, and other settings where inventory is rationed across customers with different priorities.","feed_headline":"One reordering turns optimal rationing into a threshold policy","feed_subtitle":"For a warehouse serving two demand classes, the best dynamic policy is a single cut after sorting stock levels by penalty breakpoints.","key_machinery":"The central object is the perturbation realization factor G^(d)(i)=g^(d)(i-1)-g^(d)(i), the difference between the performance potentials of adjacent inventory levels under policy d. Combined with the linear equation G^(d)(i)+b=0 in the penalty cost P, this produces the breakpoints P_i^(d) that order the low-stock states. The performance difference equation η_{d'}-η_d=μ_2 π^(d')(i)(d'_i-d_i)[G^(d)(i)+b], valid for policies differing at a single position, carries the argument: it reduces a profit comparison between policies to the sign of one factor per inventory level, and that sign is controlled by where the penalty cost P sits relative to P_i^(d).","core_discovery":"The paper proves that for any given policy d, the sign of G^(d)(i)+b, where G^(d)(i) is the perturbation realization factor between adjacent inventory levels and b=R+C_{2,2}-P, determines whether changing the service decision at level i improves the long-run average profit. The unique penalty value P_i^(d) where G^(d)(i)+b=0 orders the low-stock levels. Sorting those levels by the P_i^(d), the optimal dynamic rationing policy takes the transformed form d*(Transfer)=(0;0,...,0,1,...,1;1,...,1): reject Class 2 demand at the transformed levels with the smallest breakpoints and serve it at the rest. When the sorted order is the natural order 1,2,...,K, the original policy itself is a threshold policy; when it is not, the optimal policy is of what the paper calls transformational threshold type. The paper also gives three penalty regions—high, low, and middle—with explicit sufficient conditions under which the original policy is of threshold type.","pith_inferences":["The same sign-of-G^(d)(i)+b criterion suggests a natural extension to more than two demand classes: the single transformed threshold would likely become a staircase of cuts ordered by the analogous breakpoints, one cut per additional priority class.","Because the breakpoints P_i^(d) depend on the reference policy, the characterization suggests an iterative policy-improvement loop—sort states by current breakpoints, apply the transformed threshold update, recompute, and repeat—that could find the optimum without enumerating 2^K policies.","The algebra uses only the birth-death generator and the linearity of profit in P, so the threshold-after-reordering structure may persist under phase-type service times or Markovian arrival processes, although the paper does not claim that extension."],"forward_implications":["If the penalty cost P is at least PH(d), the optimal policy rejects every Class 2 demand while stock is low: d*=(0;0,...,0;1,...,1).","If 0<P<=PL(d) and PL(d)>0, the optimal policy serves Class 2 at every low-stock level: d*=(0;1,...,1;1,...,1).","In the middle range PL(d)<P<PH(d), the optimal policy becomes a threshold after sorting inventory levels by the breakpoints P_i^(d); when that sorted order is the natural order 1,2,...,K, the original policy itself is a threshold policy.","The long-run average profit is linear in the penalty cost for each fixed policy, and the paper derives closed-form profit expressions for the extreme optimal policies.","Restricting to static threshold policies is suboptimal in the middle penalty region, so a manager who insists on a fixed critical rationing level can lose profit relative to the optimal dynamic policy."],"supporting_citations":[{"why":"supplies the performance potential and performance difference equation used in Lemma 1 to compare any two policies.","marker":"[11]"},{"why":"introduces perturbation realization and potentials, the concepts behind G^(d)(i).","marker":"[12]"},{"why":"provides the Poisson-equation computation used to derive the policy-based Poisson equation and its general solution.","marker":"[57]"},{"why":"applies sensitivity-based optimization to queues and gives the bang-bang control that this paper refines in the middle penalty region.","marker":"[60]"},{"why":"establishes optimality of critical rationing levels for make-to-stock queues via MDP, the result this paper re-derives algebraically.","marker":"[39]"},{"why":"first proved dynamic rationing policies are optimal, the historical claim generalized by the transformed threshold result.","marker":"[81]"}],"fun_headline_variants":["Reordering inventory yields optimal threshold policy","Optimal rationing? Just sort by penalty breakpoints","Transformational threshold: key to optimal rationing","A single reorder makes rationing policy threshold-optimal","Algebraic solution reveals: reorder for threshold policy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The core assumption is that a policy can be assembled coordinate by coordinate: an improving change at one stock level stays improving after changes are made at other stock levels, even though those changes alter the system's steady-state probabilities and the marginal values that justify the first change.","fun_headline_variants_meta":{"raw":{"variants":["Reordering inventory yields optimal threshold policy","Optimal rationing? Just sort by penalty breakpoints","Transformational threshold: key to optimal rationing","A single reorder makes rationing policy threshold-optimal","Algebraic solution reveals: reorder for threshold policy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1269,"prompt_tokens":932,"completion_tokens":337,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":263}},"tokens_in":548,"tokens_out":337,"duration_ms":3909,"temperature":1.0,"reasoning_tokens":263,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:16:18.846340+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Enumerate all policies for a small instance with K low-stock levels in the middle penalty region, for example K=5, N=100, $\\lambda$=3, mu_1=4, mu_2=2, C1=1, C2,1=4, C2,2=1, C3=5, C4=1, R=15, with P strictly between PL(d) and PH(d). Compute the long-run average profit of every policy from the birth-death stationary distribution, and also compute the best policy that becomes a block of zeros followed by ones after sorting levels by that policy's own breakpoints P_i^(d); if the unrestricted best profit exceeds the transformed-threshold best profit, the paper's central existence result fails.","supporting_citations":[{"cited_title":"L., Cao, J","cited_arxiv_id":null,"evidence_quote":"provides the Poisson-equation computation used to derive the policy-based Poisson equation and its general solution."},{"cited_title":"Y., Xia, L., Li, Q","cited_arxiv_id":null,"evidence_quote":"applies sensitivity-based optimization to queues and gives the bang-bang control that this paper refines in the middle penalty region."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"establishes optimality of critical rationing levels for make-to-stock queues via MDP, the result this paper re-derives algebraically."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"first proved dynamic rationing policies are optimal, the historical claim generalized by the transformed threshold result."}],"review_version":1}