WorksComputer Science
The Data Debt of Stockouts: Partial Identification and Minimax-Regret Inventory Control
Abstract
Problem definition: Inventory decisions select which demand enters sales logs. Unrecorded forecasts can leave demand frequencies unidentified even when stocking levels are recorded. We study which multistage ordering decisions these logs can justify and what finite samples can guarantee. Methodology/results: Nested stockout thresholds yield exact sets of compatible demand laws and finite representations for policy evaluation. We choose one deterministic inventory rule to minimize its largest expected cost gap from the informed optimum under the same law. This population value is data debt for the specified information and policy class. Under independent demand, carryover can create positive debt despite common optimal actions in every isolated period. Finite samples add learning error above debt. The required sample count grows linearly with the selection bound, with upper and lower orders matching up to logarithmic factors. This rate requires fixed demand and cost bounds, sufficient revelation, and a uniformly positive probability of clearing all feasible stock across the confidence sets. Controlled and semi-synthetic studies separate policy improvements from tighter guarantees. Managerial implications: More records can tighten guarantees while debt persists. Deeper revelation can establish common optimality without fully identifying demand. Using these guarantees requires validated logging, a credible selection range, and a regret tolerance in cost units. Information acquisition must distinguish tighter assurance from operating savings.
Introduction
Demand censoring is common in supply chains and logistics. Firms observe sales or deliveries but not all demand because stock and capacity limit what can be served [11]. Sales below the available stock reveal demand exactly. A sellout reveals that demand reached the recorded boundary but not how far it exceeded it.
Our log records sales and stock available before demand. A manager may stock more when an unrecorded forecast predicts a busy day, making high-stock days unrepresentative. The sensitivity bound limits the difference between revealed and hidden demand compositions. Each period’s demand distribution stays stable between learning and deployment.
To see how unrecorded demand can change an inventory decision, consider an item with demand zero or one. Its long-run log contains 20% revealed zeros, 20% revealed ones, and 60% censored at zero inventory. The revealed zero-to-one odds are 1:1. A selection bound allows hidden odds from 1:11 to 11:1, giving overall zero-demand probabilities from 0.25 to 0.75 in Figure 1(a). With setup, holding, and shortage costs of 0.30, 0.20, and 0.80 and no unit ordering cost, the endpoint explanations recommend opposite actions. Figure 1(c) shows that the wrong action costs an extra 0.25 in expectation. Both values appear in uncensored records, but their population frequencies remain unresolved.
Figure 1. The same stockout log supports different demand frequencies and opposite optimal actions. (a) Two compatible completions at . Percentages refer to the full population. (b) The selection bound varies while the log stays fixed. (c) Ordering and no-order costs. The actions tie at , and the other action at either endpoint adds 0.25 in expected cost.
A deterministic rule specifies one order for each period and inventory level. Its regret is the expected cost above the best such rule knowing the demand law. We minimize the largest regret across compatible laws. The resulting population data debt is 0.25 in the opening example, even with exact log frequencies. Carryover links decisions through leftover stock: isolated periods can share optimal actions while the full horizon has no common optimal policy.
Our analysis proceeds from demand identification to common ordering rules and then to finite-sample guarantees. Here is the horizon, bounds integer demand and inventory, and counts independent and identically distributed (iid) historical trajectories, each recording an -period cycle.
Exact demand identification from nested thresholds. Theorem 1 characterizes the demand laws that reproduce the log under hidden selection. Higher stocking levels reveal demand beyond lower ones, giving the records a nested structure. Proposition 2 uses this structure to eliminate unobserved allocations of censored demand, leaving demand probabilities and linear constraints. Proposition 3 gives at most distributions from lower or upper stopping probabilities; their weighted averages fill the demand set. Corollary 1 establishes equivalence with the full periodwise sales-and-stocking record. Additional historical stock-flow restrictions can narrow this set.
A carryover mechanism for multistage data debt. Using each period’s log distribution, Proposition 5 identifies the smallest worst-case regret attainable by one deterministic rule. Proposition 6 shows that this debt vanishes exactly when compatible models share an optimal rule. Proposition 7 then isolates the carryover mechanism: with identical costs across periods, both isolated periods can have zero debt while the full horizon has positive debt.
Explicit sensitivity dependence in learning above debt. Finite logs add uncertainty about the demand set. Enlarging it for sampling error gives an upper regret bound, or certificate (Theorems 2–3). We implement this step with Clopper–Pearson (CP) binomial confidence intervals. A second construction uses demand hazards: the probabilities of stopping at a level conditional on reaching it. Theorems 4–5 and Corollary 2 give the matching trajectory order , up to logarithms, in the stated fixed-confidence and small-error regime. Here bounds excess worst-case regret above debt, and bounds the chance a record reaches a threshold and reveals whether demand stops there. The upper bound holds with fixed demand and cost bounds, uniformly positive stock-clearing probability across confidence sets, and exact hazard-set optimization. The lower family has positive debt.
The numerical studies compare hazard and CP, trace cost conditions that change orders, and examine common optimality under deeper revelation. Section 3.2 describes a compatible logging architecture, and Section 7 explains how a cost-unit tolerance can guide deployment and information collection.
Related Literature
Learning with observed and censored demand. Levi et al. [23] bound empirical newsvendor performance, and Besbes and Mouchtaki [3] characterize worst-case performance across sample sizes. Related learning guarantees include Xie et al. [39], Zhang et al. [42], Fan et al. [10], Lyu et al. [25]. Censored-demand learning uses threshold hazards, stochastic approximation, and adaptive ordering [11, 17, 16, 4, 1, 41, 24, 9]. In the single-period censored newsvendor, Hssaine and Sinclair [15] derive irreducible regret and statistical guarantees for distributions agreeing below the largest historical stock. Kumar and Mouchtaki [22] study worst-case performance and targeted exploration.
In offline pricing, Bu et al. [5] characterize when censored sales permit near-optimal decisions and attain the best-achievable minimax loss otherwise. We characterize the exact demand set under bounded informative stocking and its consequences for multistage ordering with carryover.
Our closest multistage comparisons, Qin et al. [31] and Han et al. [13], also observe sales and stocking boundaries. Qin et al. (2026, Sections 2.2 and 5) fix period demand laws conditional on the complete boundary array; coverage for ordering and carryover enables policy learning even when absolute costs remain unidentified. Han et al. [13] allow history-dependent stocking with current demand laws fixed conditional on past observations. Their biased sample-average approximation (SAA) learns from offline coverage. and generates it online. Here stocking may use omitted current-demand predictors, leaving frequencies within observed support unresolved. Planning demand remains independent; policies are precommitted and deterministic. § 5.3 distinguishes our explicit dependence above positive debt from the cubic scalar-record horizon order in Qin et al. [31].
Sensitivity to hidden selection. Kaciroti et al. [19] compare event odds across observed and missing-data patterns in a Bayesian product-limit model. Here threshold odds bounds between revealed and hidden groups, combined with nested stockout records, yield linear demand constraints and an endpoint recursion. Odds-ratio sensitivity, confounding-robust policy learning, and sequential evaluation provide established tools [35, 43, 40, 20, 29]. Hess et al. [14] derive sharp value bounds, efficient estimators, and learning guarantees, including baseline improvement. Jin et al. [18] minimize sensitivity-set regret against the same-law class optimum, allowing deployment randomization. We build on this criterion. The inventory contribution links nested censored records to an exact marginal set, a common-law carryover problem, and explicit sensitivity-dependent finite-sample limits. Table EC.Table 1 separates the information restrictions and targets.
| Method | Excess risk | True-law regret | Width | ||
| 250 | Hazard | 0.2547 | 0.001849 | 0.002936 | 0.7559 |
| 250 | CP (binary) | 0.2370 | 0.001850 | 0.002998 | 0.7038 |
| 1,000 | Hazard | 0.2299 | 0.001809 | 0.007410 | 0.5962 |
| 1,000 | CP (binary) | 0.2102 | 0.001828 | 0.005479 | 0.5493 |
| 4,000 | Hazard | 0.2068 | 0.001798 | 0.010757 | 0.4808 |
| 4,000 | CP (binary) | 0.1904 | 0.001784 | 0.007320 | 0.4484 |
Table 1. Notes. Each entry equally weights 12 fixed instances and 200 paired datasets per instance. CP uses the specialized five-interval construction. Excess risk is ; width refers to the first-period zero-demand probability. Mean debt is 0.165034. These design averages do not pool the instances into an iid confidence calculation.
Inventory decisions under partial information. Minimax regret is established in inventory [30] and statistical decisions under partial identification [32, 33, 27]. Unlike moment- or distance-based uncertainty sets [8, 26], our demand laws must have a stocking mechanism that reproduces the retained record. A sharp demand set retains every law compatible with the historical model and excludes every incompatible law; additional joint observations or stock-flow constraints change that benchmark. Dynamic and learning inventory models examine procurement commitments, alternative feedback, and demand dependence [37, 2, 6, 21, 28, 12, 36]; revenue-management unconstraining supplies demand completions [34].
Model Setup
We first define future inventory decisions and their cost, then describe the historical logs used to learn a rule. Historical stocking may use demand information absent from those logs.
The planning problem
There are periods, and inventory is counted in integer units up to capacity . Before period ’s order, the stock is . Initial stock is fixed, with in the numerical studies. The decision is the stock level after ordering. The order quantity is . Demand then arrives, sales are the smaller of demand and available stock, and leftover stock carries to the next period. Unmet demand is lost rather than backlogged.
The four nonnegative coefficients charge for placing an order (), each unit ordered (), leftover stock (), and unmet demand (). and is one when event holds and zero otherwise. A policy is a lookup table with . This deterministic Markov policy uses only the period and current inventory. These policies form the finite class . Write for the demand distribution in period and for its probability vector, where . Independence gives the joint law , called a product law. denotes expectation under that law.
Thus is the expected cost of a proposed rule, is the informed optimum, or oracle value, in the same class, and is their nonnegative difference in cost units per modeled horizon. For a fixed action cap , grouping larger physical demands at (top coding) preserves inventory paths and regret after a common cost subtraction. Section EC.§ 1 proves this reduction. For the perturbation proofs, we also subtract a policy-invariant component from the stage cost:
Let and . Because , the transformed and original regrets coincide. The subtraction removes a demand-dependent cost that no policy can affect, so later probability perturbations need only track the policy-sensitive part of cost.
The historical observation
The baseline historical file records both sales and the effective censoring level . The historical observation model imposes no interperiod stock-flow restriction on . We call the reveal depth. Demand beyond that level is not distinguished by the record, while depth reveals the full modeled range. The revelation indicator is determined by . It equals one when and zero at a sell-through tie. The probabilities and summarize revealed and censored records, respectively:
Write for the recorded triple and for its population law. The coarsened record has law , summarized by and . Sales of four reveal demand of four when stock remains, but imply only at a sellout. We use to construct demand sets. Corollary 1 shows why these summaries suffice for periodwise identification under the selection restriction below. Files retaining only sales and a validated revelation indicator use the same construction. A physical capacity or stock proxy must be validated as the effective boundary before it can be treated as (Section EC.1.6).
A pre-demand signal can predict current demand despite period independence (Section EC.1.6). One illustrative use learns a fixed rule from historical cycles allowing leftover stock to be returned, reallocated, or removed, then deploys with carryover. Demand marginals stay stable across these protocols. Without removals, adds constraints; combining periodwise demand sets then gives a conservative benchmark, as § 4.3 explains.
At each threshold , consider records with . Those with reveal whether demand stops at or exceeds it. Those with do not. In either group, the stopping probability is its demand hazard, and its odds are that probability divided by the probability of exceeding . The hidden-selection bound limits the unrevealed-to-revealed odds ratio:
Cross-multiplication covers zero probabilities and leaves empty-group comparisons unrestricted. Thus equates the hazards when both groups have positive mass. Full demand–stock independence is sufficient but not necessary. Figure 1(b) shows how relaxing this bound widens the compatible demand interval.
The censoring proportion measures how often demand is hidden. The bound limits how different its composition may be. Independent stocking can cause frequent censoring at . Varying at a fixed log changes the permitted selection, while its censoring frequencies and depths remain fixed.
Population and deployment assumptions
The assumptions below collect the observation, demand, sensitivity, and deployment requirements. Independent historical trajectories are additionally required for the finite-sample guarantees in § 5.1.
Assumption A1 (Observation architecture). The baseline historical record contains and follows (4), with sell-through ties treated as censored. Its derived indicator is .
Identification and certification use a validated effective boundary and revelation indicator.
Assumption A2 (Planning demand). Demand has support and is independent across periods, with potentially period-specific marginals. There is no terminal cost.
Independence makes periodwise marginals sufficient for multistage cost evaluation. It does not require . With serial dependence, the same marginals can give different policy costs.
Assumption A3 (Sensitivity). The historical law satisfies the odds bound defined in § 3.2 at the stated on the full prespecified universe .
Calibrating requires substantive knowledge, external information, or validated deeper revelation (Section EC.1.6). Understating it can exclude the true law.
Assumption A4 (Deployment class). The learner commits before deployment to one policy in .
The informed comparator uses the same class. A deployment mixture draws one of several rules before use. Such mixtures and rules using additional sales history define different benchmarks.
Assumption A5 (Population stability). For each period, the deployment demand marginal equals the latent demand marginal associated with historical observation law .
A5 allows stocking changes while holding demand marginals fixed. Population shifts require new data or a shift model. § 6.2 checks alternative post-cutoff evaluation laws.
From Demand Identification to Inventory Decisions
We first treat log probabilities as known and identify the demand laws that explain them. We then simplify this set for policy evaluation, combine periods, and ask whether one rule is optimal throughout. Finite-sample uncertainty enters in § 5.
Demand laws supported by the log
A record censored at stock tells us only that demand was at least . We distribute that record’s probability among possible demands , writing for the amount assigned to each. These assignments must reproduce the recorded censored frequencies. Adding them to revealed demand gives a candidate demand distribution:
We now express the hidden-selection bound from Section 3.2 through these allocations. The revealed stopping and continuing masses are and . For records hidden at threshold , is the probability mass stopping at , and is the mass continuing beyond it:
The revealed and unrevealed odds are and . Cross-multiplying the bound gives
The observed masses are constants, while are linear in the allocations, so the sensitivity restrictions are linear.
Let collect all allocations satisfying (5)–(8). Keeping only the resulting demand probabilities gives , the demand-marginal projection. A projection here means discarding the allocation variables after checking that at least one valid allocation exists. Because the restrictions are linear, the demand set is a polytope, a bounded region described by linear constraints.
Theorem 1 (Periodwise-sharp identification). is exactly the set of demand marginals generated by joint laws of that reproduce and satisfy the threshold-specific restrictions.
Each feasible allocation supplies a compatible historical law, and conversely. Having identified the set, we next ask whether recording each censoring level narrows it.
Corollary 1 (Recorded-boundary equivalence). Let be a law of consistent with the observation architecture in (4), and let be its coarsened probabilities. Under the threshold-specific restrictions, with no additional restrictions on the joint law of , the demand set reproducing the full periodwise record satisfies
The full record fixes each revealed cell for , with row sum . Appending a feasible hidden allocation reproduces while preserving the demand marginal and threshold odds restrictions. Section EC.1.1 proves the result, including zero-mass cases. This is a population-identification result; finite-sample confidence sets are constructed in Section 5.1. Propositions 2 and 3 therefore represent the same set for both the full periodwise record and its summary.
The case connects the identified set to product-limit estimation, which combines successive threshold hazards into a demand distribution.
Proposition 1 (Product-limit benchmark). If for every , then at the demand hazard equals , and is the singleton discrete product-limit law.
The hazards determine the discrete product-limit, or Kaplan–Meier, law used in censored inventory learning [16, 24]. A threshold with no revealed risk cannot be learned merely by setting . The proof is in Section EC.1.1.
From latent allocations to endpoint distributions
The latent allocation proves which demand laws are possible, but its unobserved cells need not all appear in the optimization. Because a censored depth can contribute only to larger demands, cumulative tail constraints suffice to decide whether a proposed marginal admits an allocation. This yields a direct representation in demand probabilities. Suppress time subscripts and write .
Proposition 2 (Direct demand-marginal representation). consists exactly of probability vectors satisfying
together with . The projected set has variables and constraints.
The first two constraint groups preserve revealed demand and assign censored mass only at or above its reveal depth. The last two impose the selection bound. The hazard construction below establishes sufficiency and the endpoint representation (Section EC.1.2).
The same nesting gives a constructive description. At threshold , let be the unresolved probability mass whose revelation has ended by and whose demand reaches . Its fraction stopping at is the unrevealed hazard . The remaining fraction continues to the next threshold, where mass newly censored there is added. When , this gives
The interval is when both revealed masses are zero, when only , and when only .
The convex hull of endpoint distributions consists of all their weighted averages. The next result shows that these averages recover the entire demand set.
Proposition 3 (Endpoint-hazard representation). is the convex hull of distributions generated by choosing each at or . There are at most endpoint images before duplicate and convex-hull reduction.
Holding the other hazards fixed, the demand vector in (13) is affine, or linear plus a constant, in the remaining hazard. Thus every interior distribution is a weighted average of endpoint distributions (Section EC.1.2). Allocations establish compatibility, direct constraints simplify the variables, and endpoints enable finite evaluation.
Combining the identified marginals
Carryover requires demand laws for the whole horizon. A rectangular family allows any admissible marginal in one period to be paired with any in another. Write for the joint law of and for the retained marginals. The family is
Each point of (14) specifies one product demand law for the entire horizon. The law is not selected again after each inventory decision. By Corollary 1, retaining the full periodwise laws gives the same family. Joint observation information can impose additional restrictions. Let contain the product demand laws generated by joint historical mechanisms reproducing the entire and satisfying the same periodwise restrictions.
An outer set contains every compatible law and may include additional laws.
Proposition 4 (Rectangular outer set). Under product demand, , and inclusion can be strict. If admissible historical mechanisms factor across periods and , equality holds.
Thus the rectangle is exact for retained periodwise information and conservative when joint or stock-flow restrictions narrow the historical class (Section EC.1.3).
One common policy and rectangular data debt
Several demand laws remain possible after identification, so a multistage policy must be evaluated across them. We compare its cost with the informed optimum under the same law and take the largest gap across the rectangle. Rectangular data debt is the smallest such gap achievable by one common policy.
Definition 1 (Rectangular data debt). The deterministic-Markov rectangular data debt is
A minimizer is a rectangular -robust policy.
To interpret debt as an information limit, we must show that the rectangle can reproduce the retained logs.
Proposition 5 (Deterministic reduced-information minimax value). Within the historical observation model of Section 3.2, fix and . A deterministic rule observes only these marginals and returns a policy in . Allow every joint historical mechanism with product demand marginal, observation marginals , and valid periodwise restrictions. The smallest attainable worst-case regret is exactly . The same value obtains if the rule instead observes all periodwise full-record laws and mechanisms must reproduce them.
Compatible periodwise completions can be multiplied within this observation model to attain every point of (14). Marginalization proves the reverse inclusion. Corollary 1 extends the construction to fixed . This proves both minimax characterizations (Section EC.1.3).
A vertex is a demand law that is not a mixture of others in its set; a product extreme model combines periodwise vertices. These models determine when debt vanishes.
Proposition 6 (Monotonicity and decision sufficiency). With the observation law, costs, and policy class fixed, is nondecreasing in . It is zero if and only if all product extreme models share an optimal policy in .
Thus demand can remain unidentified while the optimal decision is identified. A rule that stays selected as varies can still incur positive regret. Section EC.1.4 proves the zero-debt characterization.
From isolated actions to a multistage policy
A natural shortcut would be to decide whether each period’s information identifies its optimal order and then combine those answers. Independent demand does not justify this shortcut, because leftover stock couples the costs of successive decisions. Let denote the minimax regret of period solved alone, from zero inventory and without continuation value. The following two-period family isolates this carryover effect under identical costs.
Proposition 7 (Carryover under identical period costs). Let , , , and independent demand probabilities , where . In both periods take , , and . Then , while
is positive whenever , a nonempty interval.
In isolation, every compatible model prefers no order because . The same dominance removes second-period ordering from zero in the dynamic problem. The remaining choices are never to order, costing , or buy one unit initially and never replenish, costing . Early stock is worth , the probability of demand occurring at least once. Its range can straddle even though neither single-period probability does. Section EC.1.5 proves the reduction.
Figure 2 gives a periodwise censoring instance. Here , , , and yield in both periods. At , two zero isolated debts yield deterministic dynamic debt , attained by initial stocking without replenishment.
Figure 2. Rectangular data debt in the two-period instance of Proposition 7. (a) Each point is a product demand law with demand-one probabilities . Colors show the informed optimal policy, with indifference on . L and H denote the endpoints 0.35 and 0.65, in period order. (b) Both undominated deterministic policies are evaluated at the four corners, where their worst regrets occur. The smaller row maximum is .
Within the stated observation model, the endpoints share identical periodwise sales-and-boundary laws, so deterministic debt persists with exact knowledge of those laws. A deployment mixture that initially stocks with probability has minimax regret , which remains positive (Remark EC.1.3). With exact log laws, the mixture meets a two-period regret tolerance of 0.10; every deterministic rule has worst-case regret at least 0.1225. The deterministic separation also persists for sufficiently small positive common holding costs (Section EC.1.5).
Remark 1 (Dynamic debt need not add across periods). Dynamic debt can also fall below the sum of isolated debts. With time-varying costs, free early stock can remove a later replenishment choice and its positive isolated debt (Section EC.1.5). Together with Proposition 7, this establishes that there is no general ordering.
Policy Guarantees and Computation
Population debt assumes known log probabilities. With finite records, a confidence set enlarges the compatible demand set to allow for sampling error. Optimizing over it gives a regret certificate: an upper bound on the selected rule’s worst-case regret. We first establish its validity, then bound its gap above debt, derive the sample requirements, and explain computation.
From finite records to a policy guarantee
We observe iid historical trajectories of periods. Informative dependence between demand and reveal depth within each trajectory remains allowed. The statistics use , with derived from the recorded boundary. Corollary 1 justifies the population target; finite-sample protection comes from the confidence construction below.
Hidden selection prevents treating the revealed demand as a representative population sample. It does not prevent estimating the hazard within the revealed group. For record at period and threshold , let and . Their counts are and . The first counts records distinguishing stopping from continuation. The second counts stops. Conditioning on that observable risk count gives the following inference target.
Lemma 1 (Conditional binomial representation). For positive population revealed-risk mass, under arbitrary within-period dependence of ,
The result follows by counting three mutually exclusive cases: revealed demand stopping at , revealed demand exceeding , and all other records (Section EC.2.1). It supplies inference for the revealed group’s hazard. The sensitivity constraints then relate that hazard to the unrevealed group. A zero observed risk count receives the full interval . When population risk is zero, the cross-multiplied restrictions handle the boundary directly.
Fix a desired failure probability , distinct from the isolated debts . The CP intervals bound observable probabilities and revealed hazards [7]. Dividing across intervals (Bonferroni allocation) protects their simultaneous coverage. Write for the interval for . Varying the observable probabilities over their intervals and relaxing each sensitivity constraint outward gives the demand projection . In particular, the unknown true revealed hazard is replaced by the appropriate confidence endpoint in each inequality:
Let be the true demand law shared by history and deployment under A5. Section EC.2.1 gives the full CP system. CP and the hazard construction in § 5.3 are alternative confidence enlargements of the same population set. CP supports the linear implementation, while the hazard recursion exposes the explicit selection-dependent rate.
Theorem 2 (Containment and a policy certificate). Under Assumptions A1–A5 and the independent sampling scheme above, with probability at least ,
If minimizes worst-case regret over this outer set, with value , then on the same event,
Simultaneous containment precedes policy selection, so it protects the rule chosen from the same data. For any policy define its population worst-case regret by
For a selected rule , let be its checked upper certificate. On the containment event, the benchmark debt, selected-rule risk, certificate, and true-law regret satisfy
Figure 3(b) separates the gap (policy suboptimality) and (assessment conservatism). True-law regret has no general ordering with population debt. Theorem 2 requires no uniform clearance, and a rule can be certified before optimization converges (Section EC.2.4).
Figure 3. Controlled CP diagnostics at . (a) Means over 20 datasets. Shading is the interquartile range of . (b) Population debt, selected-policy suboptimality, and assessment conservatism at . (c) Dispersion and the count meeting tolerance 0.15.
Separating sampling uncertainty from persistent debt
Containment establishes validity. To assess precision, we measure how far the enlarged set extends beyond the population set, then translate this probability error into cost error. Let denote a fixed confidence construction and its optimized worst-case regret. The largest distance of an outer demand vector from its nearest population-compatible vector, measured by the sum of absolute probability differences, is
To translate this error into cost, let be the range of over feasible states, actions, and demands, and its largest demand range at a fixed state and action. The clearance floor is taken over the relevant periodwise family. Demand clears every feasible stock position. Section Section EC.2.2 bounds the remaining transformed cost range across starting inventories by and constructs the conversion coefficients as
In Lemma 2, are the marginals of two product demand laws ; here denotes a demand law.
Lemma 2 (Policy-value perturbation). *For product laws in the family defining and any ,
The same bound holds for the model-specific optimal transformed values.*
The lemma bounds both a fixed rule’s cost and the informed optimum. Regret is their difference, giving the factor two in the certificate gap below. Sections Sections EC.2.2–EC.2.3 give the proofs.
Theorem 3 (Certificate contraction). *On the simultaneous containment event for the fixed construction , with coefficients uniform over its outer family,
The sample cost of hidden selection
Theorem 3 reduces learning to demand-set error. We next bound this error using sample size and , then test whether any learner can improve its order. Making the earlier population risk’s dependence on explicit, write
Thus is the selected rule’s excess over the best common rule with known log laws. Learning can remove this excess while population debt remains positive. The deployed policy belongs to . The learning algorithm may randomize which such policy it returns.
A confidence set built from the hazards. For a revealed stopping probability , define
These hidden-hazard endpoints in (13) let us track how transforms estimation error. Give each observable cell probability a Hoeffding interval and each revealed hazard a conditional binomial Kullback–Leibler (KL) interval . Let . The cell intervals have radius , and the hazard interval consists of satisfying , where for and denotes binary Kullback–Leibler divergence. At a zero count it is . Vary the cell vector over its intervals and the probability simplex (nonnegative entries summing to one), vary each hidden hazard over , and generate demand vectors using (13). Their convex hull is , which contains the population set on simultaneous coverage.
Let minimize worst-case regret over the product of these sets, with optimized value . A known valid clearance restriction may be imposed on the outer sets if it holds for every population-compatible marginal. This ensures that sufficiently large demand can erase inventory differences between policies. Section EC.§ 3 gives the complete construction and proofs.
Here bounds the joint revealed-risk probability , which includes demand survival as well as stocking depth.
A Hoffman constant converts violations of linear constraints into distance from their feasible set; the hazard construction bounds demand-set error directly.
Theorem 4 (An explicit hidden-selection upper bound). Under Assumptions A1–A5, suppose the historical trajectories are iid and every positive revealed-risk probability is at least . Suppose , , , , and the outer families on the simultaneous coverage event have clearance at least . Set and . If , exact optimization gives, with probability at least ,
The constants , , , are held fixed when comparing rates in , , , . No Hoffman constant is required.
The odds transforms amplify error measured through square roots of probabilities by at most , yielding the linear sample dependence. Following unresolved mass through successive thresholds converts hazard error into demand-set error. Under shared demands and a common policy, an initial inventory difference then has expected duration at most , keeping fixed when the cost and clearance bounds are fixed. Section EC.§ 3.1 derives outer clearance from population clearance after sufficiently many samples. Without a uniform floor, the present argument yields , up to logarithms. The rate describes exact optimization over the hazard sets; CP containment and directional approximation are analyzed separately.
The next result tests necessity using similar logs that require different robust decisions.
Theorem 5 (Hidden-selection lower bound above positive debt). For every , , , and integer , there is a family of valid binary-demand logging experiments with positive deterministic population debt, revealed risk at least , uniformly bounded known stage costs, and population clearance at least . For every learning rule returning a policy in , allowing randomization in that choice,
The same family satisfies
Costs are common to the alternatives at fixed , and may vary across these parameter-indexed families. The benchmark is deterministic population debt, not the value for ex ante deployment mixtures.
Near revealed hazard , statistically close logs at fixed revealed risk require opposite robust actions, each wrong action incurring excess above its own positive debt. Reset periods with certain stock-clearing demand remove carryover between decisions. Adding their losses supplies the factor in total excess regret. Comparing both bounds gives sample requirements for the same excess-over-debt target.
Corollary 2 (Sample complexity with explicit sensitivity). Under Theorem 4, a sufficient trajectory count for with probability at least is
For , , fixed failure probability , and , Theorem 5 implies the necessary order
Thus the , , , exponents match up to logarithms for the excess-over-debt target, with support, cost and clearance constants fixed. The lower family satisfies these conditions even though its debt is positive.
Here bounds total excess regret over periods. For an average tolerance , the upper-bound assumptions and Corollary 2 give the sufficient count
The necessary order is in the corollary’s lower-bound regime, with . Thus the polynomial factor concerns fixed total tolerance; average tolerance retains only logarithmic horizon dependence in the upper bound. The revealed-risk gate still applies. Each of the trajectories contains period records.
Relation to noninformative logging. Counting each period observation as one scalar record gives records; the matched order is up to logarithms. Qin et al. [31] already obtain the cubic horizon order under their logging and coverage conditions. Our addition is the explicit selection dependence for learning above positive population debt. Their usable fraction and our joint revealed risk need not coincide. A tolerance below cannot be met by more records of the same type within .
Policy evaluation and numerical certification
The statistical results assume exact optimization. Implementation must distinguish selecting a policy from certifying its regret. Each period has feasible lookup tables, so . Proposition EC.1.2 reduces the search for worst-case demand to at most endpoint products, because a fixed-policy cost difference is affine in each marginal. Small problems permit enumeration.
Column-and-constraint generation (CCG) solves a smaller policy problem over selected demand scenarios, called the master. It then searches the full set for the chosen policy’s worst-case scenario, the separation step, and adds that scenario. A globally solved master gives a lower bound, and complete separation gives a policy upper bound; their gap measures optimization error. Each positive gap adds a new endpoint scenario, ensuring finite termination under exact global solves, but no polynomial-time guarantee. Section EC.4.4 compares these calculations at the reported small scales.
A finite-sample CP set also varies the observable probabilities. Substituting estimates into the fixed-observation endpoint formula would omit this uncertainty. For each chosen weight vector, or support direction, we bound the weighted sum of demand probabilities. These inequalities give an enclosing polytope with directions, distinct from setup cost . On coverage, . More data address sampling uncertainty. More directions refine the computational enclosure. Section EC.2.3 bounds both value errors by the same perturbation argument. Fixed directions can leave a positive gap even with unlimited data.
For , the checked implementation encloses the CP set in a verified polytope. At products of its vertices, dynamic programming (DP) computes upper policy-cost bounds and lower informed-cost bounds, accounting for rounding and vertex-storage error (Section EC.2.4). The largest resulting cost gap gives the selected rule’s upper bound . Numerical optimization error, sampling uncertainty, and directional approximation are reported separately.
Numerical Experiments
Three studies follow the theoretical sequence: compare finite-log guarantees with population debt, vary costs to locate order changes, and test common optimality after deeper revelation. All policies are deterministic and use the same-law informed comparator.
Policy learning and guarantees from finite logs
Comparing the confidence constructions. We first use binary demand, taking values zero or one, to compare policy risk and bound tightness for the hazard construction of Theorem 4 and the specialized five-interval CP construction on identical logs. We use , , , , and both signs of a fixed hazard perturbation around the lower-bound decision threshold. Each of these 12 instances keeps its demand law and costs fixed across , 1,000, 4,000, with 200 paired datasets per cell and . Four policies and all endpoint products are enumerated. Section EC.4.1 gives construction and coverage details.
Table 1 averages the fixed instances. Specialized CP has lower mean interval width and certificate in all 36 design cells. Each paired 95% interval, computed separately for a fixed design cell, excludes zero. Mean excess population risk favors hazard in eight cells, CP in nine, and ties in 19. The paired intervals exclude zero in three hazard-favoring cells and four CP-favoring cells. In the negative-perturbation instances, the true-law oracle differs from the population minimax policy. Learning the latter can therefore increase true-law regret. The constructions differ in intervals, parameterization and coverage allocation; this fixed-design comparison does not isolate interval efficiency or establish an asymptotic rate.
Policy error and assessment conservatism. The next study adds a third demand value and separates policy error above debt from conservatism in the bound. It uses , true and candidate , eight support directions, and 20 datasets per . From to 1,000, population risk remains 0.0773 while the mean checked bound falls from 0.2167 to 0.1500. More records chiefly tighten the guarantee over this range. At
, true-law regret is 0.0122 and debt is 0.0565. Policy suboptimality is 0.0208 and assessment conservatism is 0.0727, as Figure 3(b) shows. Both sampling and projection can affect selection. Each selected policy is evaluated on all 64 population endpoint products.
The eight-direction population optimum is 0.0678, exceeding debt by 0.0113. On one separate sample, 64 directions change the bound from 0.151021 to 0.150956. These are not replication means. Figure 3(c) reports dispersion and shows that 13 of 20 bounds meet 0.15 at . In a separate binary check with true sensitivity 3, using lowers demand containment from 98.5% to 25.4% as rises from 100 to 1,000. More data cannot repair a misspecified class (Section EC.§ 4.1).
When costs change the decision
We vary costs on fixed logs to identify when demand uncertainty changes replenishment. Naturally uncensored two-day FreshRetailNet paths (Wang et al. [38]) are a nonrandom sample, scaled and clipped to demand proxies in and artificially censored. Training through November 24, 2024 alone determines selection, scales, and capacity calibration. Of 2,000 selected series, 1,177 supply qualifying held-out paths. The formal target pairs independent draws from the periodwise training distributions. Held-out targets stress stability and dependence. These semi-synthetic results do not estimate retailer demand, hidden selection, or realized savings (Section EC.§ 4.2).
Plug-in inference treats empirical log frequencies as exact; CP allows them to vary over confidence intervals. We pair each method with minimax regret or worst expected original cost, naming the objectives regret and cost. Original cost retains the demand term that cancels in same-law regret. Adverse censoring lowers the demand-dependent reveal depth as demand rises; aligned censoring raises it. The grid fixes 30 adverse logs and varies , , and shortage cost , with . Both periods share these costs. The generator’s maximum selection odds ratio is 6.9872. Thus covers it, whereas is a misspecified benchmark. Every rule is assessed on the same population family and log-specific CP outer set.
At , CP-regret and plug-in-regret behave identically in all runs in 18 of 27 cost cells. In the other nine, behavior changes in 3.3%–100% of runs. Table 2 shows mixed risk effects despite lower common certificates, which are themselves the CP selection objective. Table EC.Table 2 retains all 27 cells.
| Metric | Lower | Equal | Higher | Range of mean differences |
| CP-regret minus plug-in-regret, both at | ||||
| Common certificate | 9 (7) | 18 | 0 (0) | [−0.0522, 0.0000] |
| Excess population risk | 6 (4) | 18 | 3 (3) | [−0.0089, +0.0778] |
| True-law regret | 6 (4) | 18 | 3 (3) | [−0.0314, +0.0523] |
| CP-regret: minus | ||||
| Common certificate | 14 (13) | 13 | 0 (0) | [−0.4223, 0.0000] |
| Excess population risk | 12 (11) | 13 | 2 (2) | [−0.0560, +0.0712] |
| True-law regret | 11 (11) | 13 | 3 (2) | [−0.2154, +0.0344] |
Table 2. Notes. Counts are among 27 cost cells, each using the same 30 logs. Parentheses count cells whose pointwise paired 95% interval excludes zero in that direction; there is no multiplicity adjustment. All policies use the same population and CP assessment sets. Lower is the CP-regret selection objective, not by itself an operating gain.
At , all 30 CP policies replenish second-period inventory to two units. Plug-in keeps one. The extra unit costs conditional on this state. It raises cost throughout the population family because the tail probability is at most 0.3875, below the break-even level 0.4375. All CP assessment sets admit tails above that threshold. The extra order therefore protects against tails outside the population family. Mean certificate falls by 0.0100 while excess population risk rises by 0.0712 and true-law regret by 0.0344.
At , CP instead raises the second-period order from one to two when inventory is empty in 9 of 30 runs. Both actions already pay the fixed fee, so the extra unit costs . Its break-even tail is 0.2308, below the true value 0.2883. Mean true-law regret falls by 0.0121 and excess population risk by 0.0031. Whether another fixed fee is incurred distinguishes these two responses to uncertainty.
Replacing by 10 in CP-regret lowers mean true-law regret in 11 cost cells and raises it in three. Ignoring hidden selection can therefore be costly, but widening the set need not improve true-law performance everywhere. Changing the CP objective to worst original cost lowers that bound in nine cells while raising regret certificates and excess risk in those cells. The cost parameters determine which uncertainty matters for action and which criterion the policy protects.
At baseline costs, four-method performance agrees at . For adverse logs, the CP-regret population-risk difference is , with paired interval [, ]. The mean improvement remains inconclusive. Under aligned censoring, CP-cost raises true-law regret by , interval [, ], while its mean worst-cost-bound change has interval [, ]. All intervals are pointwise and conditional on fixed calibration and path pools.
Evaluation targets and computation. Holding high-service-cost policies fixed, we compare plug-in candidates 1 and 10, the two values. Evaluation can pool series before drawing independently by period, draw a series first and then its independent period demands, or retain empirical two-day pairs. The corresponding mean cost changes are −0.0433 for the pooled marginal product, +0.0031 for a mixture of series products, and +0.0521 for empirical joint paths. Their single-period marginals coincide, yet rankings reverse. Changing cannot repair pooling or temporal dependence. Section EC.§ 4.2 gives paired intervals. Computation covers 36 or 216 complete policies for , or 3. Independent enumeration confirms the stored-set optima. Section EC.§ 4.4 reports timings for these enumerated problems.
A controlled revelation study
Deeper revelation supplies information about previously hidden demand. We test whether it establishes common optimality while demand remains partly identified. Two binary periods share a true observation marginal, estimated from one sample of 1,000 marginal observations. Both CP methods reuse this sample and growing prefixes of one independent revelation stream. CP/direct retains an interval for the hidden group’s stopping probability; CP/ compresses it into a symmetric bound. At 1,600 new observations, CP/direct has zero bounds in all 500 runs per scenario, while CP/ retains means 0.3808 and 0.4040. Both methods initially select true-law oracle policies. The zero bounds certify common optimality over demand sets containing multiple distributions, as in Proposition 6, without a positive operating gain. Smaller samples produce 18 cost increases after CP/direct policy updates. Collection costs are not deducted, and the guarantee is for fixed sample size, not choosing when to stop after inspecting the bound (Section EC.§ 4.3).
Discussion and Conclusion
The analyst validates sales, reveal depths, pooling, independence, population stability, and logging architecture before choosing a selection range (Section EC.1.6). An audit records intended stock before independently assigning full revelation. Under demand-preserving revelation and stable intended stocking, audited demand identifies threshold risk groups. Hazard intervals can bound odds or constrain demand directly. Calibrated ranges require calibration-error control.
A rule qualifies when its checked bound meets a cost-unit regret tolerance per horizon or period across the credible selection range. More records address sampling error; refined computation addresses approximation; deeper revelation or richer rules can change debt. Collection uses a prespecified budget with fixed sample sizes or planned checkpoints with simultaneous protection.
Revelation may certify common optimality without identifying demand. Acquisition compares guarantee changes, old and new policy costs under a common law, and collection costs over future uses.
Electronic Companion
Partial Identification and Minimax-Regret Inventory Control
Sections EC.1–EC.3 contain proofs; Section EC.4 gives experimental details.
| Study | Information restriction | Decision target |
| Qin et al. [31] | Demand law fixed conditional on the full boundary array. | Inventory policy gaps and value estimation from scalar records. |
| Han et al. [13] | History-based stocking; current demand law fixed given history. | Offline policy gaps; online cumulative regret. |
| Jin et al. [18] | Marginal sensitivity model for unobserved treatment confounding. | Same-law class-relative minimax regret; randomized deployment allowed. |
| Hess et al. [14] | Marginal sensitivity model for unobserved treatment confounding. | Sharp value bounds, policy learning, and baseline improvement. |
| This paper | Nested sales/stocking records with bounded hidden-selection odds. | Common deterministic rule; learning above debt from iid trajectories. |
Table EC.1. Information restrictions and targets in adjacent learning frameworks. Notes. The three inventory settings share sales and stocking boundaries as observables. Treatment-policy settings have different records and policy classes. Our revealed-risk probability includes demand survival and is not a stocking-only coverage fraction.
Identification and Data Debt
Notation follows the manuscript; periodwise arguments suppress . Norms are unless specified, and . The transformed cost in (3) subtracts , independently of the policy, so
For physical demand with finite first moments and a fixed action cap , top coding preserves sales and inventory paths. Only shortage cost changes, by , whose expected sum is common to every policy and its class oracle. Thus regret is unchanged. Changing the action cap changes the policy class and is a separate comparison.
One completion for both observation records
Proof of Theorem 1 and Corollary 1. Compatibility here refers to the historical observation model of § 3.2. Any compatible joint law supplies the latent masses in (5), the marginal projection, and the cross-multiplied restrictions (7)–(8); necessity follows. Conversely, choose a feasible hidden array and revealed cells for with . Define
For a coarsened record, take and the other revealed cells zero. For a given full periodwise record , take its observed revealed cells. The array has total mass and reproduces every required record: revealed cells are fixed by , and censored cell has mass . Its demand marginal is . The revealed stopping and continuing masses remain and , and the hidden masses depend only on . Thus the same restrictions hold for both records. This construction uses no division and includes zero-mass cases; if the latent set is empty, both demand sets are empty.
Proof of Proposition 1. At , . Since , this implies , so the demand hazard is . These hazards uniquely determine the product-limit distribution, including boundary values zero and one.
A common construction for the two demand representations
Proof of Propositions 2 and 3. Every feasible allocation satisfies the direct-demand constraints: is its row sum, columns can use only rows , and the hidden event and tail masses are
Conversely, take a probability vector satisfying those constraints. They give . For , put ; normalization and probability accounting give and , with terminal . If , set ; otherwise choose any point in the hazard interval of (13). The inequalities give exactly that interval, including the stated zero-mass conventions. Consequently the recursion in (13) recovers .
For any hazards in that box, a compatible allocation is
Each column sums to . Its unresolved mass at is , giving row sum and terminal row sum . The hazard bounds ensure the sensitivity restrictions. This proves sufficiency of the direct constraints and that the hazard box maps onto the demand set. The direct description has variables and constraints. Finally, the recursion is affine in each hazard separately, so multilinear interpolation expresses every image as a convex combination of endpoint images. All endpoint images are feasible and the demand set is convex, proving the endpoint representation.
Periodwise information and the common policy
Proof of Propositions 4 and 5. Marginalizing any compatible historical mechanism gives a member of . Conversely, multiplying compatible periodwise completions attains any point of that rectangle and reproduces the specified , or by Corollary 1. This product construction is admissible because the historical observation model imposes no interperiod stock-flow constraint. It proves demand-image equality for marginal-only information, even when allowed mechanisms also include dependent records. At a fixed information input a deterministic rule returns one ; every such output is attainable by a constant rule. Minimizing its worst regret over the demand image therefore gives exactly (15).
If the full observation law is fixed, marginalization gives inclusion in the rectangle. For , the product completions also reproduce and prove equality under factorized mechanisms. The following example establishes that inclusion can otherwise be strict.
Example EC.1 (A strict inclusion). In the unrestricted reveal-depth model, let be independent Bernoulli. Reveal both at depth two when they agree; otherwise censor both at zero. The observation law gives mass to each of the jointly revealed zeros and ones, and to joint censoring. Each period has , , , and the generating mechanism satisfies . At candidate , the periodwise interval is . Its lower-endpoint product violates , which is necessary to reproduce the jointly revealed-zero event. Thus the full observation law excludes a point of the rectangle.
Endpoint evaluation and zero debt
Proposition EC.1.2 (Product-vertex reduction). If with finite , then every satisfies
Proof. Write regret as . For a fixed pair, its expectation is the sum of pathwise cost differences weighted by , hence is affine in each marginal separately. Successive maximization therefore attains a product of generating points. The outer supremum commutes with the finite policy maximum, proving the claim. □
Proof of Proposition 6. Increasing enlarges the demand sets and cannot reduce minimax regret. Finiteness of guarantees a minimizer. Its nonnegative worst regret is zero exactly when it is optimal at every product extreme model, by Proposition EC.1.2. □
Carryover and isolated-period debt
Proof of Proposition 7. Since , no order is optimal in each isolated period and dominates replenishment from zero in period two. At inventory one no positive order is feasible. Replacing any period-two rule by no replenishment weakly lowers its cost under every demand law; the same reduction applies to the informed oracle. Only two undominated choices remain: never order, costing , or order initially and never replenish, costing . The difference is governed by the increasing function . Thus their worst regrets are and , giving (16). The positive-debt interval is nonempty because exceeds both and . □
For the main figure’s instance, and give worst regrets and . The isolated and final-period dominance margin is and remains strict with nonnegative holding cost. For each policy, costs are affine in the common holding cost ; the policy class is finite and the demand rectangle compact. Worst regrets and their minimum are therefore continuous in . The strict positive-debt separation and initial-order preference persist for sufficiently small .
Remark EC.1.3 (Deployment randomization in Figure 2). Allow an ex ante mixture, retaining the deterministic oracle. Let be the probability of initial stocking; otherwise never order. The demand law is fixed before the independent policy draw. The preceding dominance argument removes second-period ordering from every mixture component. Since increases over the rectangle, worst expected regret is
Balancing the terms gives and . The mixed regret remains positive despite zero isolated debts.
Reverse example for Remark 1. Let , , , and be fully revealed. Set all first-period costs to zero. At , second-period records and give by (13). Take . In isolation, not ordering costs and ordering costs , with worst regrets and . Thus . Buying one unit for free in period one and carrying it forward incurs zero cost under every compatible law. Nonnegative costs imply a zero oracle value as well, hence . □
Information scope and sensitivity calibration
For any periodwise law of , draw with ’s marginal, set , then draw conditionally on . This imposes no stock-flow equations. Inventory without removals and with nonnegative replenishment also requires . These joint restrictions shrink the rectangle, even with periodwise records; its minimax value bounds the narrower class from above. Recorded-boundary equivalence uses the stated model. The retail proxy supplies calibration; treating it as an effective boundary requires validation.
Certificates retain every row . Externally justified revealed- and hidden-hazard intervals bound both odds ratios; their largest bound across gives under the extended boundary convention. These intervals need external support, such as validated deeper revelation; stocking or recorded forecasts alone leave hidden hazards unidentified. Zero observed risk receives .
Confidence and Numerical Guarantees
Hats here denote the experimental CP construction. The explicit rate in Section EC.3 concerns a separate hazard family.
Simultaneous coverage and containment
Suppress the period index. Form two-sided CP intervals for the observable cells , , the risk probabilities , and the revealed hazards, allocating error to each. At use for the hazard. The latent confidence polytope varies over their intervals and the probability simplex, imposes
and requires to lie in its interval. For defined in (6), impose the two outer inequalities (18). Its demand projection is .
Proof of Lemma 1 and Theorem 2. The counts in , , and are multinomial. For , conditioning on the first two counts totaling makes the first count binomial with probability . Conditional CP coverage, averaged over , therefore gives marginal coverage for the hazard; cell and risk counts are ordinary binomials. A zero population risk has almost surely. The union bound yields an event of probability at least covering all intervals.
On this event, every population-feasible allocation satisfies, on positive-risk rows,
Zero-risk rows have , , making both outer inequalities vacuous. True cells satisfy all confidence constraints. Hence the entire population latent set, its demand projection, and their periodwise products are contained in the outer sets. Since belongs to the product set, its regret is bounded by the optimized supremum. Monotonicity under set inclusion gives the population-debt inequality.
Lemma EC.2.1 (Binary coefficient extrema). For , put , . Given simultaneous bounds , and an outer region for , set
Use if and if . Then lies between and .
Proof. Writing , the restrictions are and . Their solution coefficients increase in and decrease in . Endpoint extrema and prove the bounds, including the empty-group conventions. □
Value spans and perturbation
Write over the indicated finite domain.
Proof of Lemma 2. Use , , , from (23), uniform over the rectangular family under consideration. Let be the remaining transformed cost, with terminal value zero. Couple two starting inventories using the same demands and policy. Current costs differ by at most ; demand sends both next states to zero. Backward induction therefore gives
Replace by one period at a time. Adjacent hybrids have the same starting-state distribution at that period and the same future model. Conditional on state , the range over of is at most . Total variation bounds the expectation difference by . Summing gives the policy bound; taking minima gives the oracle bound because the coefficients are uniform over policies. □
A common set-distance bound
Proof of Theorem 3. For nonempty periodwise sets , let , with uniform over . The preceding lemma bounds the policy and oracle changes separately. Choosing a nearby for each and taking suprema gives, for every ,
Approximate minimizers suffice by taking their error to zero. Minimize over , choose the confidence and population sets, and use containment for the lower inequality. □
For a directional enclosure , let be its directed distance to the CP set and its optimized regret. Applying (4) and the triangle inequality gives
using coefficients uniform over the enlarged family. Fixed need not make vanish as increases; this approximation does not inherit the hazard-family rate.
Verified projection and numerical evaluation
The numerical evaluation considered here is restricted to . Confidence levels and are represented exactly as rational numbers. Guarantees use costs stored in binary64 (double precision).
Proposition EC.2.2 (One-sided numerical certificate). Let be an outward encoding of the latent CP polytope, with demand projection . Verified dual support bounds define an outer polytope . Suppose its vertices have stored simplex representatives with . For a selected policy, let and be interval-DP upper and lower bounds on its cost and the same-law oracle cost at each stored product . With uniform coefficients over the outer and stored families,
On this also bounds true-law regret, irrespective of policy optimization accuracy.
Proof. For a count and tail error , nonboundary CP endpoints are certified by exact binomial-tail comparisons:
Tail monotonicity gives outward enclosure; boundary endpoints are zero or one, and cases beyond the exact-tail verification range return . A sensitivity row , , remains valid after componentwise downward rounding of . Thus contains the exact CP set.
For each chosen direction , rationally verified dual feasibility and an upward-rounded support value imply by weak duality. Exact intersection of these halfspaces with the probability simplex gives . Vertices are stored on a dyadic simplex grid with exact unit sum; their maximum error is computed rationally. Applying the same convex weights to stored vertices shows that every point of is within of their convex hull, which need not contain . Equation (4) supplies the storage allowance. Proposition EC.1.2 reduces the stored-hull supremum to endpoint products, where directed-rounding DP gives . The allowance uses rationally computed cost spans and valid clearance floors; its final sum is rounded upward. Failed projection or numerical checks prevent issuance of a bound. □
Sample Complexity with Explicit Hidden Selection
We prove the bounds for excess over deterministic population debt in (26). A learning rule may randomize which policy in it returns; the adversarial law evaluates that returned policy.
Hazard sensitivity and the upper bound
Use the hazard construction of § 5.3, with . Here , with and infinite divergence when assigns zero mass to an event of positive mass under . Denote the cell confidence region by , the revealed-hazard interval by , and its transformed interval by . Applying (13) to these parameters and taking the convex hull gives . The untrimmed family is compact and nonempty: the cell region contains the empirical vector and the hazard intervals are nonempty. A known valid clearance restriction may be imposed. If that intersection is empty, return an arbitrary policy and certificate ; this cannot occur on coverage.
Lemma EC.3.1 (Hazard confidence-set containment and distance). For and iid historical trajectories, suppose every positive revealed risk is at least and . Then, with probability at least , simultaneously for all ,
A valid clearance intersection preserves these conclusions.
Proof. Endpoint sensitivity. Put and . For or , , with or . Since , the inequality bounds the transformed root distance by . Cauchy–Schwarz bounds total variation by root distance, and Jensen gives . Thus
The second inequality uses the triangle inequality through and . The formulas extend to zero probabilities by continuity.
Demand coupling. The recursion is a mixture of revealed point masses with weights and censored stopping chains with weights . A chain starts at , stops at threshold with probability , and terminates at . Changing mixture weights has effect at most . With weights fixed, couple the chains using the same uniforms: their first separation probability is at most . Terminal distance is at most twice that probability. Averaging over censored weights, whose sum is at most one, gives
Coverage and projection. Each cell interval has failure probability at most by Hoeffding; each nonempty hazard interval has the same bound conditional on its risk count by binomial Chernoff tails. A union bound over intervals gives coverage with probability at least . For positive population risks, the counts satisfy except on an event of probability at most .
On both events, monotonicity of and Proposition 3 give containment. For a generated outer point, replace its cells by the true cells, changing their total mass by at most . Clip each hidden hazard to its true interval; (7) bounds its change by . A zero population risk has zero count almost surely and both hidden-hazard intervals equal , requiring no change. The resulting vector is exact-feasible, and (8) gives . For convex combinations, use the same weights on their exact-feasible projections; convexity preserves the bound. A valid clearance intersection removes no exact point and cannot increase the directed distance. □
Proof of Theorem 4. On the preceding event, containment and Theorem 3 give
Here by the value-span recursion. Since , (6) yields (27). If only the exact sets have clearance , their distance bound gives outer clearance ; and the risk-count condition ensure . Without uniform clearance, gives the stated upper order. □
A binary experiment and the lower bound
Fix , , and . For , draw with and set . The full recorded triple has law
Its threshold-zero revealed risk is . Given , any zero-demand probability in gives a compatible completion without changing the log. Hence the exact demand-zero interval is
Let and . Start empty, with unit shortage cost, setup cost , and zero holding and per-unit ordering costs. This known cost is common to both alternatives below at fixed , lies in , and bounds realized stage costs by one.
Lemma EC.3.2 (A hard pair with positive debt). Set
Both experiments have positive debt. Their optimal common actions are no order at and order one at . A wrong action has excess worst regret at least , and every compatible law has . In either direction the -record KL divergence is at most , the corresponding total variation is at most , and .
Proof. The action costs are and , giving worst regrets
At these are equal, and , imply . Since and , the midpoint remains strictly inside both perturbed intervals. Both regrets are therefore positive. Also , proving clearance.
For , gives . Hence , and has opposite signs and magnitude at least at .
For either , and , so . Thus gives, in either direction,
Pinsker gives total variation at most . Finally, and give .
Proof of Theorem 5. Put . For , decision period has log . Its following reset period has demand one, setup cost , and all other costs zero. If is odd, append a cost-free period with demand one. Reset and final periods have records , identifying their demand exactly with revealed risk one. Independent periodwise completions and iid trajectories satisfy A1–A5; all compatible laws have clearance at least .
Every reset empties inventory regardless of the policy. Ordering there adds cost without affecting future inventory, so no reset order can improve the learner or oracle. Each decision starts empty. Without reset orders, both policy cost and informed minimum split across decision periods. With the worst regret for action , rectangularity allows every marginal to maximize its own same-law regret independently. Therefore
where counts wrong robust actions. Equality in the separable regret calculation is attained by policies making no reset orders.
Under the uniform prior on , pair alternatives differing only in one coordinate. All other samples have identical laws, so the full-sample total variation remains at most by Lemma EC.3.2. Adjoining an independent random seed preserves this bound. Any learning rule therefore has average testing error at least per coordinate, giving some with . As ,
For this same , implies . On that event, (12) gives excess at least , proving the weaker constants and in (29). Expectations and probabilities include the learning rule’s random selection of a deterministic deployed policy; the benchmark remains its own experiment’s population debt. □
Proof of Corollary 2. For , . Squaring Theorem 4 and including its risk-count condition gives (30). For and , Theorem 5 precludes a uniform guarantee whenever , yielding (31). The binary lower family has uniformly bounded costs and obeys the known clearance restriction throughout its exact sets. Imposing that restriction on the hazard outer family fixes its clearance constant independently of . Thus the stated exponents match up to logarithms. □
Experimental Designs and Decision Diagnostics
Controlled studies
All controlled studies start empty with . The binary design uses , , revelation probabilities for demands , costs , true sensitivity 3, candidates 1.5, 3, 5, and 1,000 datasets per . Per period, Lemma EC.2.1 bounds four cells and the revealed tail with five CP intervals, each at 0.05/10. Regret coverage is 8,000/8,000 for valid candidates (cellwise Wilson intervals [99.6, 100.0]%); at , misspecified candidate 1.5 covers regret in 77.0% ([74.3, 79.5]%) and demand in 25.4%.
The ternary study uses , , , , and true sensitivity 5.5. Costs are , , , and . For each , 20 independent panels sample observations per period, shared across candidates 3, 5.5, 8. Thirty CP intervals each receive 1/600. Eight binary64 directions , , in give eight stored vertices per period and 512 products. Population risks in Figure 3 use all 216 policies and 64 population endpoint products; directional comparisons are fixed-sample, not replication means.
Hazard–CP comparison. The comparison uses , , , , both signs, and , with 200 paired datasets per cell. Define
The first-period log follows (9), with true hidden hazard and intervals from (10). Set , , , and other costs zero. Period two has . Law and costs remain fixed across ; methods share iff two-period records.
Hazard uses four-cell Hoeffding intervals on the simplex, conditional KL intervals with , and the odds transforms. Failure budgets are 0.025 for intervals and 0.025 for the rate proof’s risk-count event; containment uses only the first. CP divides 0.05 across ten two-sided intervals. These budgets bound error probabilities. Both enumerate four policies and endpoint products with outward-rounded endpoints and exact rational evaluation of represented costs. No clearance restriction is added; six , cells fail the rate theorem’s risk-count gate. Population-set coverage is 7,200/7,200 for hazard and 7,195/7,200 for CP; CP true-law coverage is 7,197/7,200, with matching certificate counts. All five CP misses involve observable-cell noncoverage. Counts pool heterogeneous instances; paired and Wilson intervals are pointwise by fixed cell.
Retail construction and matched comparisons
The strict window requires initial stock, at most one persistent stockout transition, and no positive out-of-stock (OOS) hour sales. Eligibility requires 30 naturally uncensored days through November 24, 2024 and a positive training 90th-percentile scale . Fixed-seed selection chooses 2,000 of 17,681 eligible series; the proxy is , with ties to even and frozen scales. No path crosses the cutoff.
With , training paths are sampled with replacement. Capacity calibration uses all 484,628 strict-clean training-day records in the selected 2,000 series, including censored days and days outside two-day paths. Capacities equal 3 on naturally uncensored days and the clipped quantity otherwise, with counts (28,678, 88,914, 42,339, 324,697) at 0, 1, 2, 3. Independently per cell, draw accordingly, and . When , replace by (adverse) or (aligned). Sensitivities are 6.987 and 24.983. Baseline costs are ; , 1,000, 30 paired panels, and candidates 1, 3, 10, 30. Plug-in uses unsmoothed cells; general CP allocates 0.05/20 per interval. Four methods cross plug-in/CP inference with regret/original-cost optimization over all 36 policies; each candidate uses a shared eight-direction CP assessment set.
The grid reuses 30 adverse panels for , , , , four methods and candidates 1, 10. All 27 cells use common population and log-specific CP assessment sets (Table EC.Table 2). For policies these common- sets differ from the selection sets. Pointwise paired 95% intervals condition on fixed calibration and pools; reusing panels adds no independent samples.
For high-service costs , 20 adverse pairs select plug-in regret policies at candidates 1 and 10. With path shares , marginals and joint laws , targets are
Holding policies fixed, candidate-10 minus candidate-1 cost changes are respectively [, ], 0.0031 [0.0027, 0.0034], and 0.0521 [0.0464, 0.0579] (paired 95% intervals). The 823 series without held-out paths have zero weight. These reversals concern distinct targets, without causal attribution or population-resampling uncertainty.
A controlled revelation study
Here , and both periods share with , revealed hazard 0.20, and hidden hazard 3/7 or 3/5 (sensitivities 3 or 6), under prior . Costs are
| Changed | ||||||
| 0.1 | 0 | 0.5 | -0.0009 | -0.0009 | -0.0012 | 3.3% |
| 0.1 | 0 | 1 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.1 | 0 | 2 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.5 | 0 | 0.5 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.5 | 0 | 1 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.5 | 0 | 2 | -0.0314 | +0.0302 | -0.0522 | 70.0% |
| 1 | 0 | 0.5 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 1 | 0 | 1 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 1 | 0 | 2 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.1 | 0.2 | 0.5 | -0.0034 | -0.0024 | -0.0054 | 30.0% |
| 0.1 | 0.2 | 1 | +0.0025 | -0.0043 | -0.0123 | 20.0% |
| 0.1 | 0.2 | 2 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.5 | 0.2 | 0.5 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.5 | 0.2 | 1 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.5 | 0.2 | 2 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 1 | 0.2 | 0.5 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 1 | 0.2 | 1 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 1 | 0.2 | 2 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.1 | 0.6 | 0.5 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.1 | 0.6 | 1 | +0.0344 | +0.0712 | -0.0100 | 100.0% |
| 0.1 | 0.6 | 2 | -0.0171 | +0.0778 | -0.0425 | 73.3% |
| 0.5 | 0.6 | 0.5 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 0.5 | 0.6 | 1 | -0.0176 | -0.0044 | -0.0120 | 16.7% |
| 0.5 | 0.6 | 2 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 1 | 0.6 | 0.5 | 0.0000 | 0.0000 | 0.0000 | 0.0% |
| 1 | 0.6 | 1 | +0.0523 | -0.0089 | -0.0157 | 33.3% |
| 1 | 0.6 | 2 | -0.0121 | -0.0031 | -0.0192 | 30.0% |
Table EC.2. Complete cost grid: CP-regret minus plug-in-regret at . Notes. Means over 30 common logs at each cost, with . Here before a metric denotes the paired difference. Changed is the fraction with different actions on at least one of the nine demand paths from empty inventory. All cells, including ties and adverse changes, are retained.
Each of 500 replications per scenario reuses one multinomial vector of 1,000 iid marginal -observations across periods, without paired trajectories. Future demand is a product; confidence marginals vary separately. An independent stream of 1,600 Bernoulli observations from the baseline-unrevealed group supplies prefixes , shared by both methods. Randomized deeper inventory is assumed to reveal unchanged demand. CP/direct intersects the prior set with
CP/ instead takes the largest bidirectional odds ratio over historical and pilot hazard intervals, capped at 10. Six historical intervals receive 0.025/6 each and the pilot interval 0.025. Shared true parameters give both-period coverage with failure probability 0.05 at each fixed , without optional stopping.
For , , rationally minimize and maximize over the historical cell bounds and , $1-o_{1,U}-s_{1,U} \le a+c \le 1-o_{1,L}-s_{1,L}$, where is revealed risk. Four policies and four endpoint products suffice. Initial policies are true-law oracles; direct zero bounds at retain positive-width demand intervals. All 18 negative gains have sensitivity 3: 14 at , four at $n = 50.
Computational comparisons
Table EC.3 reports three fresh-process runs per instance for column-and-constraint generation (CCG), all-endpoint optimization, and independent policy-by-policy dynamic programming (DP) (§ 5.4). Independent enumeration checks 480 CP instances over 36 policies and 49–64 endpoint products; the maximum optimal-value discrepancy is .
Finite-set agreement checks optimization. Certifying also requires one-sided endpoint and projection checks, vertex-representation error control, and interval policy/oracle costs (Proposition EC.2.2).
| Policies | Products | CCG (ms) | Full set (ms) | Enumerate (ms) | Peak (MiB) | |
| 2 | 36 | 16 | 34.015 | 2.578 | 10.163 | 80.65 |
| 3 | 216 | 64 | 147.764 | 18.265 | 491.107 | 82.84 |
Table EC.3. Computation on two fixed population instances with . Notes. Medians of three fresh-process runs at , four endpoints per period; all three optimal values agree. Full set uses vectorized policy enumeration; CCG uses occupancy masters with enumeration checks. Peak memory covers the process. Timings compare alternative optimization methods, excluding end-to-end certification. Platform: Intel Core Ultra 5 225H, Windows 11, Python 3.14.6, SciPy 1.18.0. Thread environment variables were one; actual HiGHS threads were unavailable.
References
- [1]Ban GY (2020) Confidence intervals for data-driven inventory policies with demand censoring. Oper. Res. 68(2):309–326, URL http://dx.doi.org/10.1287/opre.2019.1883.
- [2]Bekci RY, Gumus M, Miao S (2023) Inventory control and learning for one-warehouse multistore system with censored demand. Oper. Res. 71(6):2092–2110, URL http://dx.doi.org/10.1287/opre.2021.0694.
- [3]Besbes O, Mouchtaki O (2023) How big should your data really be? Data-Driven newsvendor: Learning one sample at a time. Management Sci. 69(10):5848–5865, URL http://dx.doi.org/10.1287/mnsc.2023.4725.
- [4]Besbes O, Muharremoglu A (2013) On implications of demand censoring in the newsvendor problem. Management Sci. 59(6):1407–1424, URL http://dx.doi.org/10.1287/mnsc.1120.1654.
- [5]Bu J, Simchi-Levi D, Wang L (2023) Offline pricing and demand learning with censored data. Management Sci. 69(2):885–903, URL http://dx.doi.org/10.1287/mnsc.2022.4382.
- [6]Chen B, Wang Y, Zhou Y (2024) Optimal policies for dynamic pricing and inventory control with non-parametric censored demands. Management Sci. 70(5):3362–3380, URL http://dx.doi.org/10.1287/mnsc.2023.4859.
- [7]Clopper CJ, Pearson ES (1934) The use of confidence or fiducial limits illustrated in the case of the binomial. Biometrika 26(4):404–413, URL http://dx.doi.org/10.1093/biomet/26.4.404.
- [8]Delage E, Ye Y (2010) Distributionally robust optimization under moment uncertainty with application to data-driven problems. Oper. Res. 58(3):595–612, URL http://dx.doi.org/10.1287/opre.1090.0741.
- [9]Ding J, Huh WT, Rong Y (2024) Feature-based inventory control with censored demand. Manufacturing Service Oper. Management 26(3):1157–1172, URL http://dx.doi.org/10.1287/msom.2021.0135.
- [10]Fan X, Chen B, Olsen T, Qin H, Zhou Z (2026) Don’t follow reinforcement learning blindly: Lower sample complexity of learning optimal inventory control policies with fixed ordering cost. Production Oper. Management 35(4):1449–1468, URL http://dx.doi.org/10.1177/10591478251378851.
- [11]Godfrey GA, Powell WB (2001) An adaptive, distribution-free algorithm for the newsvendor problem with censored demands, with applications to inventory and distribution. Management Sci. 47(8):1101–1112, URL http://dx.doi.org/10.1287/mnsc.47.8.1101.10231.
- [12]Gundem K, Qi Z (2025) Offline dynamic inventory and pricing strategy: Addressing censored and dependent demand, arXiv:2504.09831.arxiv.org/abs/2504.09831
- [13]Han Y, Fan X, Zhang J, Zhou Z (2026) Towards optimal inventory control under censored demand: A biased sample-average approximation approach. arXiv preprint arXiv:2609.39397, URL http://dx.doi.org/10.48550/arXiv.2609.39397, version 1, September 30, 2026.
- [14]Hess K, Frauen D, Melnychuk V, Feurriegel S (2026) Efficient and sharp off-policy learning under unobserved confounding, arXiv:2502.13022v3.arxiv.org/abs/2502.13022
- [15]Hssaine C, Sinclair SR (2026) The data-driven censored newsvendor problem, arXiv:2412.01763, version 3, revised April 2026.arxiv.org/abs/2412.01763
- [16]Huh WT, Levi R, Rusmevichientong P, Orlin JB (2011) Adaptive data-driven inventory control with censored demand based on Kaplan–Meier estimator. Oper. Res. 59(4):929–941, URL http://dx.doi.org/10.1287/opre.1100.0906.
- [17]Huh WT, Rusmevichientong P (2009) A nonparametric asymptotic analysis of inventory planning with censored demand. Math. Oper. Res. 34(1):103–123, URL http://dx.doi.org/10.1287/moor.1080.0355.
- [18]Jin Z, Xu G, Zheng X, Zhou Y (2026) Policy learning under unobserved confounding: A robust and efficient approach, arXiv:2507.20550v2.arxiv.org/abs/2507.20550
- [19]Kaciroti NA, Raghunathan TE, Taylor JMG, Julius S (2012) A Bayesian model for time-to-event data with informative censoring. Biostatistics 13(2):341–354, URL http://dx.doi.org/10.1093/biostatistics/kxr048.
- [20]Kallus N, Zhou A (2021) Minimax-optimal policy learning under unobserved confounding. Management Sci. 67(5):2870–2890, URL http://dx.doi.org/10.1287/mnsc.2020.3699.
- [21]Katehakis MN, Teymourian E, Yang J (2024) Data-driven inventory control involving fixed setup costs and discrete censored demand. Naval Res. Logist. 71(8):1220–1236, URL http://dx.doi.org/10.1002/nav.22211.
- [22]Kumar R, Mouchtaki O (2026) What is the value of censored data? An exact analysis for the data-driven newsvendor, URL https://arxiv.org/abs/2602.16842, arXiv:2602.16842.
- [23]Levi R, Perakis G, Uichanco J (2015) The data-driven newsvendor problem: New bounds and insights. Oper. Res. 63(6):1294–1306, URL http://dx.doi.org/10.1287/opre.2015.1422.
- [24]Lyu C, Zhang H, Xin L (2024) UCB-type learning algorithms with Kaplan–Meier estimator for lost-sales inventory models with lead times. Oper. Res. 72(4):1317–1332, URL http://dx.doi.org/10.1287/opre.2022.0273.
- [25]Lyu J, Xie J, Yuan S, Zhou Y (2025) A minibatch stochastic gradient descent-based learning metapolicy for inventory systems with myopic optimal policy. Management Sci. 71(7):5572–5588, URL http://dx.doi.org/10.1287/mnsc.2023.00920.
- [26]Mohajerin Esfahani P, Kuhn D (2018) Data-driven distributionally robust optimization using the Wasserstein metric: Performance guarantees and tractable reformulations. Math. Programming 171(1–2):115–166, URL http://dx.doi.org/10.1007/s10107-017-1172-1.
- [27]Montiel Olea JL, Qiu C, Stoye J (2026) Decision theory for treatment choice problems with partial identification. Rev. Econom. Stud. rdag015, URL http://dx.doi.org/10.1093/restud/rdag015.
- [28]Mukherjee S, Pham AD, Pibernik R, Xu Y (2026) In-context learning for data-driven censored inventory control, arXiv:2605.14840.arxiv.org/abs/2605.14840
- [29]Namkoong H, Keramati R, Yadlowsky S, Brunskill E (2020) Off-policy policy evaluation for sequential decisions under unobserved confounding. Advances in Neural Information Processing Systems, volume 33, 18819–18831.arxiv.org/abs/2003.05623
- [30]Perakis G, Roels G (2008) Regret in the newsvendor model with partial information. Oper. Res. 56(1):188–203, URL http://dx.doi.org/10.1287/opre.1070.0486.
- [31]Qin H, Simchi-Levi D, Zhu R (2026) Information limits of multistage inventory control: Learning, valuation, and censoring. arXiv preprint arXiv:2609.37380, URL http://dx.doi.org/10.48550/arXiv.2609.37380, version 1, September 29, 2026.
- [32]Stoye J (2007) Minimax regret treatment choice with incomplete data and many treatments. Econometric Theory 23(1):190–199, URL http://dx.doi.org/10.1017/S0266466607070089.
- [33]Stoye J (2012) Minimax regret treatment choice with covariates or with limited validity of experiments. J. Econometrics 166(1):138–156, URL http://dx.doi.org/10.1016/j.jeconom.2011.06.012.
- [34]Talluri KT, van Ryzin GJ (2004) The Theory and Practice of Revenue Management (Boston, MA: Kluwer Academic Publishers).DOI
- [35]Tan Z (2006) A distributional approach for causal inference using propensity scores. J. Amer. Statist. Assoc. 101(476):1619–1637, URL http://dx.doi.org/10.1198/016214506000000023.
- [36]Tang J, Qi Z, Fang E, Shi C (2025) Offline feature-based pricing under censored demand: A causal inference approach. Manufacturing Service Oper. Management 27(2):535–553, URL http://dx.doi.org/10.1287/msom.2024.1061.
- [37]Wang T, Gong X, Zhou SX (2017) Dynamic inventory management with total minimum order commitments and two supply options. Oper. Res. 65(5):1285–1302, URL http://dx.doi.org/10.1287/opre.2017.1622.
- [38]Wang Y, Gu J, Long L, Li X, Shen L, Fu Z, Zhou X, Jiang X (2026) FreshRetailNet-LT: A stockout-annotated censored demand dataset for latent demand recovery and forecasting in fresh retail, URL https://arxiv.org/abs/2505.16319v3, arXiv:2505.16319v3, revised 9 June 2026.
- [39]Xie Y, Ma W, Xin L (2026) VC theory for inventory policies, URL https://arxiv.org/abs/2404.11509, arXiv:2404.11509v3, revised February 2026.
- [40]Yadlowsky S, Namkoong H, Basu S, Duchi J, Tian L (2022) Bounds on the conditional and average treatment effect with unobserved confounding factors. Ann. Statist. 50(5):2587–2615, URL http://dx.doi.org/10.1214/22-AOS2195.
- [41]Yuan H, Luo Q, Shi C (2021) Marrying stochastic gradient descent with bandits: Learning algorithms for inventory systems with fixed costs. Management Sci. 67(10):6089–6115, URL http://dx.doi.org/10.1287/mnsc.2020.3799.
- [42]Zhang X, Ye ZS, Haskell WB (2024) Error propagation in asymptotic analysis of the data-driven (s, S) inventory policy. Oper. Res. 73(1):1–21, URL http://dx.doi.org/10.1287/opre.2020.0568.
- [43]Zhao Q, Small DS, Bhattacharya BB (2019) Sensitivity analysis for inverse probability weighting estimators via the percentile bootstrap. J. Royal Statist. Soc. Ser. B 81(4):735–761, URL http://dx.doi.org/10.1111/rssb.12327.


