Pandora's Box: Optional Inspection Theory

• Pandora's Box with Optional Inspection is a sequential search problem where a decision maker may select a box without inspection, breaking classical index rules.
• The framework introduces threshold-based two-phase policies and committing strategies to manage costly information acquisition effectively.
• Research reveals NP-completeness, a PTAS for near-optimal policies, and links to prophet inequalities and stochastic submodular maximization.

Pandora’s Box with Optional Inspection, also called Pandora’s problem with nonobligatory inspection, is a sequential search problem in which a decision maker may select a box without first opening it. In contrast to Weitzman’s classical Pandora problem, where selection requires prior inspection, the optional-inspection variant lets the searcher either pay to inspect and reveal a realized value or stop and choose a box “closed.” This seemingly small change eliminates the classical index-rule solution, makes exact optimization structurally intricate, and has led to a literature centered on threshold structure, committing policies, approximation guarantees, and reductions to stochastic submodular maximization, prophet inequalities, and broader costly-information frameworks (Beyhaghi et al., 2019, Beyhaghi et al., 2022, Esfandiari et al., 2019).

1. Classical Pandora’s rule and the optional-inspection departure

In Weitzman’s original problem, there are $n$ boxes; box $i$ contains an independent random prize $v_i \sim F_i$; inspecting box $i$ costs $c_i$; and the searcher adaptively chooses which box to inspect next and when to stop. Utility is the value of the selected inspected box minus the sum of inspection costs. The optimal policy is an index policy: each box receives a reservation value $\sigma_i$ defined by

$\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$

boxes are inspected in decreasing $\sigma_i$, and search stops when some observed value exceeds all remaining reservation values (Beyhaghi et al., 2019).

Optional inspection changes exactly one requirement: a box may be selected without being inspected. The searcher can therefore inspect box $i$ first, or directly choose it closed. The survey literature treats this as one of the most important relaxations of Weitzman’s “open before select” rule and places it under “alternative inspection methods” (Beyhaghi et al., 2023). The central consequence is negative rather than simplifying: the simple reservation-value ranking is no longer exact, the optimal adaptive policy can depend intricately on realized values, and there is no known simple ranking rule or polynomial-time method for the exact optimum in the general model (Beyhaghi et al., 2019).

A recurring misconception is that optional inspection is merely an easier version of classical Pandora because the action set is larger. The literature shows the opposite. The relaxation makes the decision problem algorithmically harder, because the searcher must optimize both information acquisition and the outside possibility of taking an uninspected alternative, and these choices interact with future stopping decisions (Beyhaghi et al., 2022).

2. Formal model and tractable restricted policy classes

The canonical optional-inspection model keeps the standard stochastic primitives. Boxes $i=1,\dots,n$ have independent prizes $i$0 with known expectations $i$1, inspection costs $i$2, sequential search with recall, and a stopping rule that may select any box, inspected or not (Beyhaghi et al., 2019). A useful state representation is the generalized subproblem $i$3, where $i$4 is the set of remaining uninspected boxes and $i$5 is a free outside option, equivalently the current maximum observed value; this formulation is central to later structural results (Beyhaghi et al., 2022).

Because the unrestricted optimum is complex, early approximation work isolates a nearly nonadaptive family of committing policies. A committing policy pre-commits to a partition $i$6 of the boxes and to a total order on $i$7: boxes in $i$8 are never inspected, boxes in $i$9 are never selected before inspection, and boxes in $v_i \sim F_i$0 are inspected only in that fixed order (Beyhaghi et al., 2019). For a fixed reservation set $v_i \sim F_i$1, the policy $v_i \sim F_i$2 is defined by modifying each box in $v_i \sim F_i$3 to have inspection cost $v_i \sim F_i$4 and deterministic value $v_i \sim F_i$5, leaving $v_i \sim F_i$6 unchanged, running Weitzman’s rule on the modified instance, and replacing any hypothetical inspection of a box in $v_i \sim F_i$7 by direct uninspected selection (Beyhaghi et al., 2019).

The tractability of committing policies is stronger than it first appears. Although there are $v_i \sim F_i$8 candidate reservation sets, the optimal committing policy must be one of only $v_i \sim F_i$9 policies,

$i$0

where $i$1 is the original Weitzman policy and $i$2 reserves exactly one closed box. The expected value of each candidate can be computed in polynomial time in $i$3 and the maximum support size $i$4, so the best committing policy is efficiently computable (Beyhaghi et al., 2019).

3. Exact optimal structure, two-phase policies, and complexity

The strongest exact structural theorem currently described in the supplied literature is that optimal optional-inspection policies are far more organized than arbitrary adaptive trees. For every remaining set $i$5, there exists a threshold $i$6 such that if the outside option $i$7, then $i$8 and there exists an optimal policy using a backup box; if $i$9, no optimal policy uses a backup box (Beyhaghi et al., 2022). This threshold lemma induces a single-crossing property: once the outside option rises above the threshold of the current residual problem, it remains above the relevant thresholds later on (Beyhaghi et al., 2022).

From this, the paper proves a two-phase representation. There exists an optimal policy specified by an ordering $c_i$0, thresholds $c_i$1, and an index $c_i$2, such that while search continues, the policy opens the next box in $c_i$3 as long as the current maximum observed value is below threshold; at the cutoff $c_i$4, if still below threshold, it claims box $c_i$5 closed and stops; and if the observed maximum exceeds threshold, it switches to Weitzman’s optimal policy on the remaining boxes (Beyhaghi et al., 2022). Equivalently, the optimal policy follows only two different pre-determined orders of inspection and transitions from one to the other at most once (Beyhaghi et al., 2022).

This sharpens earlier structural understanding. The survey reports that recent work shows the general optimal policy is two-phased and can be fully specified through an initial inspection order and a threshold for each box, and that the problem belongs to NP (Beyhaghi et al., 2023). The 2022 structural paper makes this explicit, proving that Pandora’s problem with nonobligatory inspection belongs to class NP and, together with NP-hardness, is NP-complete (Beyhaghi et al., 2022).

The same paper also gives a PTAS. The approximation scheme relies on the two-phase structure, a unique backup-box viewpoint, threshold discretization, value discretization, and a reduction to the stochastic dynamic-program framework of Fu, Li, and Xu. The result is that, despite NP-completeness, policies with expected payoff at least a $c_i$6-fraction of optimum can be computed in polynomial time for arbitrarily small $c_i$7 (Beyhaghi et al., 2022).

4. Approximation guarantees and the evolution of policy design

Approximation work for optional inspection has progressed along several distinct axes: simple committing policies, structure-aware improvements, local randomization, and matroid-constrained costly-information frameworks. These guarantees are stated in different objective and feasibility settings, so they are complementary rather than directly comparable.

For the single-item maximization problem, the first non-trivial guarantee for committing policies is that the optimal committing policy is a $c_i$8-approximation to the optimal adaptive policy. The proof maps optional inspection to stochastic monotone submodular maximization with a partition matroid constraint by introducing $c_i$9 distributed as $\sigma_i$0 and $\sigma_i$1 deterministically equal to $\sigma_i$2, then combines an amortization lemma with the adaptivity gap of Asadpour and Nazerzadeh (Beyhaghi et al., 2019). In the special case of two boxes, at least one of $\sigma_i$3 achieves a $\sigma_i$4 fraction of the optimal utility, and this factor is tight for committing policies (Beyhaghi et al., 2019).

Later structural work strengthens the committing-policy guarantee itself. The two-phase/PTAS paper proves that at least one of the $\sigma_i$5 committing policies achieves at least $\sigma_i$6 of the optimal utility, and states this as a tight $\sigma_i$7 approximation with a novel proof for general classes of distributions, extending what had previously been shown only for discrete and finite distributions (Beyhaghi et al., 2022).

A different line studies the minimization analogue through local hedging. In this framework, each item is independently randomized into either obligatory-inspection mode or non-inspection mode. The loss of this local randomization is instance-dependent but at most $\sigma_i$8, yielding a $\sigma_i$9-approximation for single-item nonobligatory inspection and, when combined with obligatory-inspection combinatorial algorithms, the first approximation algorithms for optional-inspection versions of matroid basis, set cover, facility location, Steiner tree, and feedback vertex set (Scully et al., 2024).

For matroid-constrained maximization, the 2024 MDP-based framework introduces semilocal approximation because purely local approximation cannot beat the trivial $\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$0 barrier. It proves that every normalized box admits a semilocal $\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$1-approximation with $\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$2 and $\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$3, which composes into a $\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$4-approximation for max-matroid-PBOI via an efficient randomized committing policy (Chawla et al., 2024).

A subsequent line reframes the same matroid setting through the commitment gap in Costly Information Combinatorial Selection. The central bound is

$\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$5

which implies $\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$6 for matroids. Specialized back to Pandora’s Box with Optional Inspection under matroid feasibility, this yields an efficient $\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$7-approximation and matches the existential $\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$8 upper bound cited there (Chawla et al., 27 Aug 2025).

5. Online, prophet-inequality, and combinatorial generalizations

Optional inspection has also been developed in online models where boxes arrive sequentially in adversarial order. In the online Pandora model, box $\mathbb{E}\big[(v_i-\sigma_i)^+\big] = c_i,$9 arrives with a tuple $\sigma_i$0 drawn from a known joint distribution $\sigma_i$1; the type $\sigma_i$2 is observed upon arrival; the decision maker chooses whether to pay the inspection cost and open the box; and if it is opened, then $\sigma_i$3 is observed and the prize is either kept or rejected irrevocably (Esfandiari et al., 2019). The objective is

$\sigma_i$4

The key technical idea in the online literature is a reduction from costly inspection to a zero-cost prophet-inequality instance. Thresholds $\sigma_i$5 are chosen to satisfy

$\sigma_i$6

and transformed boxes

$\sigma_i$7

encode inspection costs into truncated values (Esfandiari et al., 2019). A cost-equivalence lemma preserves expected utility under this transformation, and threshold-based prophet-inequality algorithms then transfer directly back to the original optional-inspection problem. The resulting guarantees include a $\sigma_i$8-approximation for the single-item online setting, a $\sigma_i$9-approximation for online $i$0-Pandora’s box, a $i$1-approximation for online knapsack Pandora’s box, a $i$2-approximation for online matroid Pandora’s box, and a $i$3-approximation for the multi-arm Pandora’s box problem (Esfandiari et al., 2019).

This reductionist viewpoint has two broader consequences. First, it places optional inspection at the interface between Pandora-style costly search and prophet inequalities. Second, it shows that many of the useful algorithmic objects are not classical reservation values but transformed surrogate values, thresholds, and feasibility-aware selection rules. A later taxonomy of stopping problems explicitly describes optional inspection as a natural middle ground between Prophet Inequalities and Pandora’s Box, and in a minimization/no-commitment/cost regime identifies an exact connection to Decreasing Buying Cost Ski Rental with tight competitive ratio $i$4 (Correa et al., 29 Sep 2025).

6. Position in the broader Pandora literature

The survey literature describes optional inspection as the cleanest and most studied relaxation of the “inspect before select” requirement (Beyhaghi et al., 2023). Its importance comes from the fact that it preserves the core economics of costly information acquisition while breaking the exact indexability that made the classical problem special. The resulting theory therefore serves as a template for several neighboring generalizations.

One neighboring direction imposes order or precedence constraints on which boxes may be opened. There, generalized reservation values can still recover greedy optimality for rooted trees, but the moment the constraint graph becomes more general the problem becomes NP-hard to approximate well (Boodaghians et al., 2020). Another neighboring direction allows sequential inspections within a box, such as partial opening followed by full opening; that model introduces $i$5-thresholds, $i$6-thresholds, and $i$7-thresholds, proves NP-hardness, and again recovers constant-factor guarantees through committing policies and stochastic-optimization relaxations (Aouad et al., 10 Jul 2025). A planning-based ordered variant adds sunk scope costs and obtains scope-dependent reservation values indexed by paid scope, rather than a single static reservation index (Sonal et al., 9 Jun 2026).

These adjacent models do not coincide with optional inspection, but they clarify its role. A plausible implication is that optional inspection marks the point at which the Pandora literature moves away from one-dimensional reservation values and toward state-dependent threshold systems, surrogate-value amortization, and policy classes defined by partial commitment. That interpretation is consistent with the supplied survey, which emphasizes two-phase structure, threshold descriptions, PTAS results, and constant-factor approximations as the main algorithmic themes for nonobligatory inspection (Beyhaghi et al., 2023).

In contemporary research, Pandora’s Box with Optional Inspection is therefore best understood not as a minor variant of Weitzman’s rule, but as a canonical costly-information problem in which exact indexability fails, adaptivity matters, and tractable analysis emerges through structure theorems and approximation frameworks rather than a single universal index.