{"id":"7cb5b760-53b3-41f1-829f-34c6eb92579d","arxiv_id":"2509.22194","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An integrated MSP-MDP model is shown to admit a nested dynamic reformulation, and its optimal value and solution sets are proven stable under Kantorovich and Fortet-Mourier distribution perturbations.","lead":"This paper builds a single mathematical model that combines multistage stochastic programming and Markov decision processes, separating external random inputs from internal state-transition randomness, and proves error bounds for how much near-optimal costs and policies change when those random distributions are perturbed. The results give theorists a common framework and quantitative stability guarantees for a broad class of dynamic decision problems under uncertainty.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'without loss of generality' independence of ζ_t in §2 is false: a correlated process with identical stagewise marginals can change the value, so Theorem 4.1's marginal-only bound cannot hold for correlated ζ.","rationale":"The reader's weakest-assumption identification is correct and, in my view, the most load-bearing concern. Theorem 4.1's proof is inseparable from stagewise independence: (4.17) telescopes the perturbation by changing one marginal at a time, and the Kantorovich dual bound (4.16) applies only to expectations of functions of the individual ζ_t. If ζ_t are correlated, the intermediate objects in (4.8) are not valid joint laws, and the RHS of (4.5) cannot see changes in the dependence structure. The Bernoulli construction is a crisp falsification of the WLOG statement: the marginals are unchanged while the optimal value changes. This does not invalidate the stability theory within the independent-ζ model, but it narrows the claimed scope and undermines the paper's wording that the framework covers endogenous uncertainties generally. A secondary gap is that Theorem 4.1's proof invokes convexity of g_t in x_t from Theorem 3.2 without restating that condition in its hypotheses; this is reparable but should also be clarified. The reader's CONDITIONAL verdict remains appropriate: the paper needs an explicit independence assumption (or a genuine reduction argument) and a toning down of the subsumption claims.","tokens_in":69462,"tokens_out":18827,"duration_ms":170706,"concrete_test":"Run the two-stage Bernoulli test. Let T=2, ζ0,ζ1∈{0,1}, ζ2≡0, s0=0, s1=ζ0, s2=s1·ζ1, C0=C1=0, C2(s2)=-s2, and all decisions fixed in a compact set. Compute ϑ under (i) ζ0,ζ1 independent Bernoulli(1/2) and (ii) ζ1=ζ0. The stagewise marginals are identical in both cases, so all d_K(P_t,P̃_t)=0 and the RHS of (4.5) is 0, while the optimal values differ by 1/4. This verifies that the additive stagewise Kantorovich bound fails for correlated ζ and that the §2 'without loss of generality' independence claim must be replaced by an explicit independence assumption or a state-augmentation argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2, immediately after Eq. (2.1d), asserts that ζ0,...,ζT are mutually independent 'without loss of generality.' This is load-bearing in Theorem 4.1: the proof (4.8)–(4.17) telescopes the perturbation one stage at a time, bounding each increment by d_K(P_t,P̃_t) via the Kantorovich duality in (4.16). If ζ_t are correlated, a stagewise marginal perturbation does not determine a joint perturbation, and cross-stage dependence is invisible to the RHS of (4.5). Concretely, take T=2, s0=0, s1=ζ0, s2=s1·ζ1, C0=C1=0, C2(s2)=-s2, with ζ0,ζ1 Bernoulli(1/2). Under independence the optimal value is -1/4; under perfect correlation ζ1=ζ0 the marginals are identical, so all d_K(P_t,P̃_t) are zero, yet the value is -1/2. This contradicts the bound (4.5) if it were meant to cover correlated ζ. Thus the WLOG claim collapses: the theorem holds only for models where independence of the endogenous uncertainties is an explicit assumption, not a consequence of the MSP-MDP setup.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a finite-horizon integrated MSP-MDP model in which endogenous noises ζ_t drive state transitions and an exogenous process ξ_t generates history-dependent costs and constraints. It derives a nested dynamic reformulation (Theorem 3.1), establishes continuity, convexity and Lipschitz properties of the stagewise value functions (Proposition 3.2, Theorem 3.2), and then develops quantitative stability bounds for the optimal value and optimal solution sets under perturbations of the endogenous distributions in Kantorovich metric (Theorem 4.1) and of the exogenous process in Fortet-Mourier and conditional Kantorovich metrics (Theorems 4.2–4.6). The paper also includes several worked examples intended to compare the new bounds with the filtration-distance and nested-distance bounds of Heitsch–Römisch and Pflug–Pichler.","tokens_in":69757,"tokens_out":16125,"duration_ms":139915,"significance":"If the stability theorems were correct, they would supply the first explicit distributional-perturbation bounds for this integrated MSP-MDP class, with constants built from Lipschitz moduli, feasible-set diameters and the Slater constant, and they would complement the existing filtration/nested-distance results. The paper contains detailed proofs, explicit constants, and worked examples including a nonlinear example for which the new bound is computable. However, two load-bearing technical gaps—the unjustified 'without loss of generality' independence claim for the endogenous noise and the incorrect inequality direction in the Fortet-Mourier estimates—mean that the central stability claims are not yet established as stated.","major_comments":[{"comment":"The assertion that ζ0,...,ζT can be assumed mutually independent 'without loss of generality' is false for model (2.1). Since ζ_t enters the state transition S_t^M and the cost C_t, cross-stage dependence of (ζ0,...,ζT) affects the value even when all stagewise marginals are fixed. Concretely, take T=2, s0=0, s1=ζ0, s2=s1·ζ1, C0=C1=0, C2(s2)=−s2, with ζ0 and ζ1 Bernoulli(1/2). Under independence the optimal value is −1/4; under perfect correlation ζ1=ζ0 the value is −1/2, while all stagewise marginal distributions are identical. Applying Theorem 4.1 to this pair would give d_K(P_t,P̃_t)=0 for all t but |ϑ(ζ)−ϑ(ζ̃)|=1/4, contradicting (4.5). The telescoping proof at (4.17) relies precisely on inter-stage independence. The manuscript must either add an explicit assumption that ζ0,...,ζT are independent and treat this as a modeling restriction, or generalize Theorem 4.1 to a metric on the joint distribution of ζ; the current WLOG claim and the theorem as stated are not valid.","section":"§2, after Eq. (2.1d); Theorem 4.1, proof around (4.17)"},{"comment":"The passage from E_{ξ,ξ̃}[max{1,‖ξ‖^r,‖ξ̃‖^r}‖ξ−ξ̃‖] to the Fortet-Mourier metric d_FM,r+1 has the wrong inequality direction. With the definition (4.1)–(4.2), d_FM,p is a supremum over test functions and, by Kantorovich-Rubinstein duality, the infimum over couplings of the weighted L1 cost max{1,‖ξ‖^{p−1},‖ξ̃‖^{p−1}}‖ξ−ξ̃‖. Therefore, for any fixed coupling—in particular the product coupling implicitly used by writing E_{ξ,ξ̃}—the expectation of the cost is ≥ d_FM,p, not ≤. For instance, in (4.52) the chain 'E[...] ≤ Lϑ d_FM,3T+1' is only valid if (ξ,ξ̃) is chosen as an optimal coupling for the Fortet-Mourier cost; otherwise the inequality goes the wrong way. The same incorrect direction appears in (4.30), (4.61), and in (4.85), where E‖ξ_T−ξ̃_T‖ ≤ d_K is asserted despite d_K being the infimum over couplings. The proofs can likely be repaired by fixing, for each pair of measures, a coupling attaining the infimum in the relevant Kantorovich/Fortet-Mourier metric and performing all estimates under that coupling, or by using the dual representation directly on the value functions, but as written the Fortet-Mourier bounds are not established.","section":"§4.2, Proposition 4.1 Eq. (4.30); Theorem 4.2 Eqs. (4.45), (4.52); Theorem 4.4 Eq. (4.61); Theorem 4.6 Eq. (4.85)"},{"comment":"The theorem statement concludes E_{ξ,ξ̃}[H(X*(ξ),X*(ξ̃))] ≤ ϵ, but the proof ends with '<3ϵ' after accumulating the distance through the orthogonal projection and the ϵ-neighborhood argument. The constants in the definition of δ(ξ,ξ̃) should be rescaled (e.g., by replacing ϵ with ϵ/3 throughout) to match the stated conclusion, or the statement should be changed to a 3ϵ bound. This is a quantitative mismatch in one of the main stability theorems and needs to be corrected in revision.","section":"Theorem 4.3, proof leading to (4.53)"}],"minor_comments":[{"comment":"The claim that the obtained results 'subsume the main conclusions of [18], [29] and [36]' is stronger than what the paper itself demonstrates; Example 4.2 shows regimes where the new bound is looser than the nested-distance bound, and the authors later write that 'theoretical evidence is yet to be established'. The wording should be softened to 'complement'.","section":"§1 and the paragraph after Theorem 4.2"},{"comment":"The displayed identity d_K(Q1,Q̃1)=∫‖ξ1‖dQ1−∫‖ξ̃1‖dQ̃1 uses only the particular 1-Lipschitz test function ‖·‖; this gives a lower bound on the Kantorovich metric, not its value. The tightness comparison in the example should be reworked by computing or bounding the actual supremum over all 1-Lipschitz functions.","section":"Example 4.2, around Eqs. (4.95)–(4.97)"},{"comment":"The constants L_{X,t} appearing in the definition of L̂_t are not defined in the theorem statement; they should be defined before the statement. In the proof they are introduced as L_{X,t}=A L_g L_S/ρ, so this is only a presentation issue, but it should be fixed for readability.","section":"Theorem 4.1 statement, Eq. (4.5)"},{"comment":"Reference [55], 'David Wozabal. Stability of Markovian stochastic programming', is incomplete: no publication venue or year is given. Please complete the citation or replace it with a published version.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a strong claim to subsume prior stability results in multistage stochastic programming, but its own comparison example shows that the new bounds can be either tighter or looser depending on the problem data. I would ask the authors to temper that claim and to state clearly which assumptions are modeling restrictions (especially the independence of ζ_t) rather than consequences of the framework. The Kantorovich/Fortet-Mourier direction errors are likely fixable with a standard optimal-coupling argument, but they affect several central theorems and should be addressed carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Key thing to know: the integrated MSP-MDP model is genuinely new and the stability analysis is mostly careful, but the claim that the endogenous uncertainties can be taken as mutually independent 'without loss of generality' is false and load-bearing for Theorem 4.1. The model as stated in (2.1) does not assume independence, and §2's WLOG claim after (2.1d) doesn't survive contact with correlated ζt. Concrete failure: take T=2, s0=0, s1=ζ0, s2=s1·ζ1, C0=C1=0, C2(s2)=-s2, ζ0,ζ1 Bernoulli(1/2). Under independence the value is -1/4; under perfect correlation ζ1=ζ0 the stagewise marginals are unchanged, so every d_K on the right-hand side of (4.5) is zero, yet the value is -1/2. Thus Theorem 4.1 cannot hold as stated for the general model; the proof's telescoping at (4.17) leans on exactly the independence that was asserted rather than proven. The fix is easy conceptually: make independence an explicit modeling assumption (natural for many MDP transition noises) or replace the right-hand side with a metric that captures the joint distribution/copula. Either way, the 'without loss of generality' has to go.\n\nWhat's actually good: the unified formulation is a fair first of its kind, and the dynamic nested reformulation (Theorem 3.1) and Lipschitz/convexity properties (Prop 3.2, Theorem 3.2) are worked out in serious detail. The stability theorems give explicit constants in terms of Lipschitz moduli, the Slater constant ρ, and feasible-set diameter A — no fitted parameters. The worked examples (4.2, 4.3) honestly compare their bounds with the nested-distance and filtration-distance bounds, and the comparison shows complementarity rather than superiority, which is fine; it is not a defect that sometimes their constant is worse.\n\nThe weaker spots are proportionate. The claim in the introduction that the results 'subsume' Heitsch–Römisch and Pflug–Pichler is overstated: those papers quantify dependence of the whole data process, which these Kantorovich bounds do not, and Example 4.2 itself shows the bounds are not uniformly tighter. The right word is 'complement.' Theorem 4.3's qualitative continuity is more of a template result and adds little beyond Theorem 4.4's quantitative version. None of these are dealbreakers.\n\nVerdict: the paper deserves a serious referee, but only after the independence issue is fixed and the subsumption claim is toned down. If the authors restrict Theorem 4.1 to independent ζ's, the paper still has plenty of value — the integrated model and the exogenous-uncertainty stagewise bounds are new. I'd send it back for major revision rather than desk-reject. Not something I'd personally cite in its current form.","headline":"New integrated MSP-MDP model with careful stability bounds, but the 'WLOG' independence of endogenous noise is false and breaks Theorem 4.1 as stated.","tokens_in":70239,"tokens_out":3626,"would_cite":false,"duration_ms":31839,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C15","90C40"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that perturbing either the endogenous or exogenous randomness in an integrated MSP-MDP model changes the optimal value and optimal solution set only by weighted sums of stagewise probability-metric distances.","keywords":["multistage stochastic programming","Markov decision process","stability analysis","Kantorovich metric","Fortet-Mourier metric","endogenous uncertainty","exogenous uncertainty","nested reformulation"],"falsifier":"Build a two-stage instance satisfying all Lipschitz and Slater assumptions, set $\\zeta_1=\\zeta_0+\\varepsilon$ with a coupling that preserves each marginal, and perturb only the joint distribution. If the resulting change in optimal value exceeds the sum of stagewise Kantorovich terms in (4.5), the paper's independence claim fails; if it stays within the bound, the independence assumption is not the limiting premise.","tokens_in":1487,"feed_emoji":"🎲","tokens_out":1820,"duration_ms":58151,"temperature":0.7,"pith_summary":"The paper tries to establish that an integrated model combining Markov decision processes and multistage stochastic programming is quantitatively stable under distributional perturbations of both types of randomness. It derives a nested reformulation, then proves that the optimal value and optimal solution set change by amounts controlled by stagewise Kantorovich (or Fortet-Mourier) distances between original and perturbed distributions. This matters because when the environment shifts or the data contain estimation error, a model user wants a guarantee that small distribution errors cannot cause large changes in the recommended policy. The results offer explicit error bounds without relying on filtration distance or nested distance, making them easier to compute in practice.","feed_headline":"MSP-MDP value errors obey a stagewise probability-metric bound","feed_subtitle":"Endogenous and exogenous distribution perturbations are controlled by sums of Kantorovich metrics, stage by stage.","key_machinery":"The central machinery is the dynamic nested reformulation (3.2)-(3.4), combined with a Lipschitz-continuity result for the stagewise value functions (Theorem 3.2) obtained under Slater's condition and Lipschitz assumptions on costs, transitions, and constraints. The Kantorovich metric, the supremum expectation gap over 1-Lipschitz test functions, converts these Lipschitz moduli into metric bounds through its dual representation. For exogenous perturbations, an additional conditional-Lipschitz condition on the kernels $Q_t(\\xi_t\\mid\\xi_{[t-1]})$ lets the effect of a stagewise perturbation propagate recursively through the value functions.","core_discovery":"For the integrated model, with endogenous uncertainties $\\zeta_0,\\dots,\\zeta_T$ independent across stages, Theorem 4.1 establishes $|\\vartheta(\\zeta)-\\vartheta(\\tilde\\zeta)|\\le \\sum_{t=0}^{T-1}\\hat L_{t+1}d_K(P_t,\\tilde P_t)+L_C d_K(P_T,\\tilde P_T)$, together with a corresponding Hausdorff bound for the optimal solution sets. For exogenous uncertainty, Theorems 4.5 and 4.6 give analogous bounds through conditional Kantorovich metrics, while Theorems 4.2 and 4.4 provide whole-process Fortet-Mourier bounds under Lipschitz conditions on conditional distributions. The stability theory is built on the nested dynamic reformulation and does not require the relatively complete recourse condition used in earlier filtration-distance results.","pith_inferences":["The paper treats mutual independence of $\\zeta_0,\\dots,\\zeta_T$ as 'without loss of generality' but supplies no reduction argument; if the endogenous shocks are correlated across stages, the additive stagewise sum in Theorem 4.1 is not guaranteed, and a nested or joint-distance formulation would likely be needed.","The same dual-representation argument could be adapted to distributionally robust versions of the integrated model: the Lipschitz moduli computed here directly supply worst-case gaps for ambiguity sets measured in Kantorovich or Fortet-Mourier metrics.","For infinite-horizon problems, the recursive constants $\\hat L_t$ would need to be shown contractive; the paper stops at a finite horizon, so whether the stability bounds survive as $T\\to\\infty$ remains open.","A testable extension would be numerical validation of bound tightness on inventory or energy-dispatch instances with misspecified demand and loss-rate distributions."],"forward_implications":["Small estimation error in a single stage's endogenous distribution yields value error bounded by that stage's Kantorovich distance times a constant, without modeling the whole process's filtration distance.","The bounds decompose perturbation effects stage by stage, so a decision maker can identify which stage's distribution error contributes most to the total value error.","For exogenous uncertainty, the conditional Lipschitz condition implies that perturbations at early stages propagate to later stages at a rate controlled by $L_{Q}$ products.","When a growth condition holds, similar quantitative Hausdorff bounds apply to the optimal solution sets, not only to optimal values.","In the one-stage case the solution-set bound yields a linear dependence on the Kantorovich distance, strengthening earlier square-root-type estimates."],"supporting_citations":[{"why":"Supplies the filtration-distance stability bound that the new results compare against and aim to complement.","marker":"[18]"},{"why":"Supplies the nested-distance stability baseline for convex multistage stochastic programs.","marker":"[36]"},{"why":"Introduces filtration distance for scenario tree modeling, an existing perturbation measure.","marker":"[17]"},{"why":"Provides prior quantitative stability for linear multistage stochastic programs extended by Theorem 4.2.","marker":"[29]"},{"why":"Supplies first-order sensitivity of MDP value functions under transition-kernel perturbation, contrasted with the global bounds here.","marker":"[25]"},{"why":"Provides the Kantorovich metric and stability theory used for the dual representation.","marker":"[42]"},{"why":"Supplies the probability-metric method underlying Kantorovich and Fortet-Mourier metrics.","marker":"[41]"},{"why":"Provides a one-stage solution-set stability bound that the new solution-set results strengthen in the $T=1$ case.","marker":"[43]"},{"why":"Supplies a conditional-distribution Lipschitz assumption comparable to condition (4.71).","marker":"[55]"}],"fun_headline_variants":["MSP-MDP value errors follow Kantorovich sum bounds","Stagewise metric bounds for MSP-MDP stability","New MSP-MDP stability results via Kantorovich/Fortet-Mourier","Perturbation bounds for MSP-MDP without recourse condition"],"cache_read_input_tokens":72448,"weakest_assumption_plain":"The claim that $\\zeta_0,\\dots,\\zeta_T$ can be taken mutually independent without loss of generality is unproved; if the endogenous uncertainties are correlated across stages, the additive stagewise decomposition behind the main endogenous-stability bound collapses.","fun_headline_variants_meta":{"raw":{"variants":["MSP-MDP value errors follow Kantorovich sum bounds","Stagewise metric bounds for MSP-MDP stability","New MSP-MDP stability results via Kantorovich/Fortet-Mourier","Perturbation bounds for MSP-MDP without recourse condition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000429,"raw_usage":{"total_tokens":2218,"prompt_tokens":992,"completion_tokens":1226,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":1154}},"tokens_in":608,"tokens_out":1226,"duration_ms":8882,"temperature":1.0,"reasoning_tokens":1154,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:44:11.827596+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a two-stage instance satisfying all Lipschitz and Slater assumptions, set $\\zeta_1=\\zeta_0+\\varepsilon$ with a coupling that preserves each marginal, and perturb only the joint distribution. If the resulting change in optimal value exceeds the sum of stagewise Kantorovich terms in (4.5), the paper's independence claim fails; if it stays within the bound, the independence assumption is not the limiting premise.","supporting_citations":[{"cited_title":"Stability of multistage stochastic programs.SIAM Journal on Optimization, 17(2):511–525, 2006","cited_arxiv_id":null,"evidence_quote":"Supplies the filtration-distance stability bound that the new results compare against and aim to complement."},{"cited_title":"A distance for multistage stochastic optimization models","cited_arxiv_id":null,"evidence_quote":"Supplies the nested-distance stability baseline for convex multistage stochastic programs."},{"cited_title":"Scenario tree modeling for multistage stochastic programs.Mathematical Programming, 118:371–406, 2009","cited_arxiv_id":null,"evidence_quote":"Introduces filtration distance for scenario tree modeling, an existing perturbation measure."},{"cited_title":"On stability of multistage stochastic programs.SIAM Journal on Optimization, 19(2):952–968, 2008","cited_arxiv_id":null,"evidence_quote":"Provides prior quantitative stability for linear multistage stochastic programs extended by Theorem 4.2."},{"cited_title":"First-order sensitivity of the optimal value in a markov decision model with respect to deviations in the transition probability function","cited_arxiv_id":null,"evidence_quote":"Supplies first-order sensitivity of MDP value functions under transition-kernel perturbation, contrasted with the global bounds here."},{"cited_title":"Stability of stochastic programming problems.Handbooks in operations research and management science, 10:483–554, 2003","cited_arxiv_id":null,"evidence_quote":"Provides the Kantorovich metric and stability theory used for the dual representation."},{"cited_title":"Quantitative stability in stochastic programming: The method of probability metrics.Mathematics of Operations Research, 27(4):792–818, 2002","cited_arxiv_id":null,"evidence_quote":"Supplies the probability-metric method underlying Kantorovich and Fortet-Mourier metrics."},{"cited_title":"Stability of solutions for stochastic programs with complete recourse.Mathematics of Operations Research, 18(3):590–609, 1993","cited_arxiv_id":null,"evidence_quote":"Provides a one-stage solution-set stability bound that the new solution-set results strengthen in the $T=1$ case."},{"cited_title":"Stability of markovian stochastic programming","cited_arxiv_id":null,"evidence_quote":"Supplies a conditional-distribution Lipschitz assumption comparable to condition (4.71)."}],"review_version":2}