Correlated Pandora’s Problem

• Correlated Pandora’s Problem is a generalization of Pandora’s Box where box values come from an arbitrarily correlated joint distribution, making sequential decision-making complex.
• The model requires strategies that optimally balance opening costs and revealed values, using techniques like reservation values and scenario-based updates.
• Recent studies leverage LP relaxation and Poisson rounding to achieve near-optimal approximations, aligning with NP-hardness benchmarks inherited from related set cover problems.

Correlated Pandora’s Problem is a generalization of Pandora’s Box in which the vector of box values is drawn from a known joint distribution and may be arbitrarily correlated across boxes. In the minimization convention used by much of the recent literature, a strategy sequentially chooses boxes to open, pays their opening costs, and aims to minimize the expected sum of opening costs plus the minimum revealed value among the opened boxes (Gergatsouli et al., 2023). The model strictly generalizes Min Sum Set Cover and is related to Uniform Decision Tree, so it sits at the intersection of stochastic search, adaptive information acquisition, and approximation algorithms (Bansal et al., 21 Sep 2025).

1. Formal model and information structure

A standard formulation has a finite set of boxes $B$, nonnegative opening costs $c_b$, and a random value vector

$v=(v_1,\ldots,v_n)\in \mathbb R_+^n$

drawn from a distribution $D$. A strategy sequentially decides which box to open next and when to stop and select one of the opened boxes. If $\mathcal O\subseteq B$ is the set of opened boxes, the objective is

$$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$

In the correlated setting, the coordinates of $v$ may be arbitrarily correlated, so revealing one value changes the posterior over all unopened boxes (Gergatsouli et al., 2023).

Much of the analysis uses an explicit finite-support scenario representation. There are $m$ scenarios $s\in S$, scenario $s$ occurs with probability $c_b$0, and under scenario $c_b$1, box $c_b$2 has value $c_b$3. In this view, opening a box reveals information about the underlying scenario, not merely the value of that single coordinate. This scenario-based representation is central in reductions to decision-tree and set-cover formulations (Chawla et al., 2021).

The correlated model should be distinguished from the classical independent-box setting. Under independence, opening one box reveals only its own value and does not affect beliefs about the others. Under correlation, the posterior is state-dependent, so the next action is inherently an information-acquisition decision as well as a search decision (Gergatsouli et al., 2023).

2. Adaptivity classes, benchmarks, and hardness

Three policy classes recur throughout the literature. In a Fully-Adaptive (FA) strategy, both the next box and the stopping time may depend on previously observed values. In a Partially-Adaptive (PA) strategy, the probing order is fixed in advance, but the stopping rule may depend on observations. In a Non-Adaptive (NA) strategy, both the probed set and the stopping rule are fixed before any values are revealed (Chawla et al., 2019).

This distinction is substantive rather than cosmetic. With arbitrary correlations, observations can encode the identity of future promising boxes, so FA strategies can exploit information in ways unavailable to PA or NA policies. In the sample-access model of "Pandora's Box with Correlations: Learning and Approximation" (Chawla et al., 2019), fully adaptive strategies cannot be efficiently approximated to within any sub-linear factor, which motivates the use of the best partially adaptive strategy as the benchmark in that line of work.

For explicit-support distributions, "Approximating Pandora's Box with Correlations" (Chawla et al., 2021) shows that correlated Pandora’s Box is approximation-preserving equivalent, up to small losses, to Uniform Decision Tree and a feedback version of Min-Sum Set Cover. Consequently, for support size $c_b$4, the problem admits a polynomial-time $c_b$5-approximation, while constant-factor approximations are achievable in subexponential time; at that stage, a polynomial-time constant-factor approximation remained open.

The hardness landscape is now sharp for the partially adaptive benchmark. Min Sum Set Cover is a special case of correlated Pandora, which implies that it is NP-hard to approximate the partially-adaptive optimum better than $c_b$6 (Gergatsouli et al., 2023). "Optimal 4-Approximation for the Correlated Pandora's Problem" (Bansal et al., 21 Sep 2025) strengthens the interpretation: factor $c_b$7 is optimal because the problem strictly generalizes MSSC, and factor $c_b$8 is NP-hard even with the power of fully adaptive algorithms.

A further distinction concerns minimization versus maximization. The correlated literature on approximation focuses on the cost-minimization form,

$c_b$9

because the maximization version with correlations has no bounded approximation (Bansal et al., 21 Sep 2025).

3. Reservation values under arbitrary correlations

The classical independent-box solution is Weitzman’s rule: for each box $v=(v_1,\ldots,v_n)\in \mathbb R_+^n$0, compute a reservation value $v=(v_1,\ldots,v_n)\in \mathbb R_+^n$1 from

$v=(v_1,\ldots,v_n)\in \mathbb R_+^n$2

open boxes in increasing order of $v=(v_1,\ldots,v_n)\in \mathbb R_+^n$3, and stop when the best observed value is at most the next reservation value. Under correlation, these indices are no longer static marginals, because opening one box changes the posterior distribution of all unopened boxes (Gergatsouli et al., 2023).

"Weitzman's Rule for Pandora's Box with Correlations" (Gergatsouli et al., 2023) shows that a direct posterior-based generalization remains effective. At each step, under the current distribution $v=(v_1,\ldots,v_n)\in \mathbb R_+^n$4, compute $v=(v_1,\ldots,v_n)\in \mathbb R_+^n$5 from

$v=(v_1,\ldots,v_n)\in \mathbb R_+^n$6

open the box with minimum $v=(v_1,\ldots,v_n)\in \mathbb R_+^n$7, and stop if the observed value $v=(v_1,\ldots,v_n)\in \mathbb R_+^n$8. The paper analyzes two posterior-update rules: $v=(v_1,\ldots,v_n)\in \mathbb R_+^n$9 for Variant 1: partial updates, and

$D$0

for Variant 2: full updates.

The guarantees are approximation results against the optimal partially-adaptive policy, not exact optimality under arbitrary correlations. The main theorem states that Algorithm 1 is a $D$1-approximation for Variant 1 and a $D$2-approximation for Variant 2 (Gergatsouli et al., 2023). The same paper also proves an NP-hardness lower bound of $D$3, so the $D$4 bound is close to the known hardness barrier.

A key analytical ingredient is an equivalent scenario-based characterization of reservation values. If the support is a set of scenarios $D$5, then

$D$6

Under uniform support this becomes

$D$7

This formula enables the histogram and tree-histogram analyses that underlie the $D$8 and $D$9 bounds (Gergatsouli et al., 2023).

The same work also gives a learning theorem. For unit opening costs and any $\mathcal O\subseteq B$0, polynomially many samples from the correlated distribution suffice so that Variant 1 achieves a $\mathcal O\subseteq B$1 approximation to the partially-adaptive optimum with probability at least $\mathcal O\subseteq B$2 (Gergatsouli et al., 2023).

4. Optimal approximation for the partially adaptive problem

The approximation status of the partially adaptive Correlated Pandora’s Problem is settled by "Optimal 4-Approximation for the Correlated Pandora's Problem" (Bansal et al., 21 Sep 2025). The paper studies the minimization model with arbitrary deterministic opening costs $\mathcal O\subseteq B$3, arbitrary joint distribution $\mathcal O\subseteq B$4 over scenarios $\mathcal O\subseteq B$5, and arbitrary correlations among box volumes. The benchmark is the optimal partially adaptive policy: the algorithm fixes an ordering of the boxes up front, opens boxes in that fixed order, and stops adaptively.

The main theorem states that Balanced Stopping with Poisson Rounding is a $\mathcal O\subseteq B$6-approximation to the optimal partially adaptive algorithm (Bansal et al., 21 Sep 2025). The factor is optimal because Correlated Pandora contains Min Sum Set Cover as a special case. The same paper explicitly identifies the factor-$\mathcal O\subseteq B$7 lower bound through this reduction.

The proof is LP-based rather than purely combinatorial. In continuous time, the relaxation uses cumulative start-time variables $\mathcal O\subseteq B$8 and scenario-dependent variables $\mathcal O\subseteq B$9, with per-scenario LP objective

$$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$0

The algorithm converts the fractional solution into randomized arrivals through non-homogeneous Poisson processes and then applies a stopping rule based on the balanced time

$$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$1

If $$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$2 is the first-arrival time of box $$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$3, define

$$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$4

Balanced Stopping stops at $$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$5 and takes the corresponding box (Bansal et al., 21 Sep 2025).

The core scenario-wise guarantee is

$$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$6

Taking expectation over $$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$7 yields the global $$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$8-approximation theorem (Bansal et al., 21 Sep 2025). Methodologically, the paper replaces earlier combinatorial amortization arguments with an LP/convex relaxation, Poisson rounding, and an algebraic linearization that compares each scenario directly to the relaxation.

This result improves on the earlier $$\mathbb E_{v\sim D}\!\left[\sum_{b\in \mathcal O} c_b + \min_{b\in \mathcal O} v_b\right].$$9-approximation of the posterior-Weitzman approach (Gergatsouli et al., 2023) and on the earlier $v$0-approximation from the first sample-access work on correlated Pandora (Chawla et al., 2019), while matching the hardness inherited from MSSC (Bansal et al., 21 Sep 2025).

5. Structured correlated variants and adjacent models

A major structured subclass is "Markovian Pandora's box" (Yang et al., 31 Jan 2025). In that model, rewards are correlated through a directed acyclic graph $v$1 that specifies both order constraints and Markovian dependence. Along a directed line $v$2,

$v$3

The objective returns to the classical maximization form,

$v$4

This structured correlation admits stronger positive results than the arbitrary-joint-distribution model: for a forest-structured precedence graph, there exists a fully adaptive algorithm that achieves the optimal expected payoff in polynomial time and space; under static transition, the paper gives near-optimal truncation results for multi-line graphs and a $v$5-approximation in forests (Yang et al., 31 Jan 2025).

This structured-correlated line should be contrasted with models that generalize Pandora in other directions but retain independence across boxes. "Pandora's Box Problem with Order Constraints" (Boodaghians et al., 2020) proves optimal threshold strategies for tree-like precedence constraints under independent rewards, but not for correlated rewards. A plausible implication is that precedence constraints and statistical dependence are orthogonal sources of difficulty: tree-structured feasibility can preserve Weitzman-style greedy optimality under independence, whereas arbitrary correlation already destroys static indexability in the unconstrained setting.

A separate distinction concerns delegation. "Delegated Pandora's box" (Bechtel et al., 2022) allows correlation within each element between the principal’s value $v$6 and the agent’s value $v$7, but assumes independence across elements: $v$8 Its positive approximation results rely on that product structure, so it is not a correlated-Pandora result in the sense of arbitrary joint distributions over boxes (Bechtel et al., 2022).

The terminology can also be misleading in broader costly-information frameworks. "Commitment Gap via Correlation Gap" (Chawla et al., 27 Aug 2025) studies Costly Information Combinatorial Selection under product-distributed item values. Its use of correlation gap refers to the correlation gap of a feasibility system and ex ante versus ex post relaxations, not to arbitrary cross-item statistical dependence. Thus it is directly relevant to generalized Pandora under combinatorial constraints, but not a solution to fully correlated Pandora’s Box (Chawla et al., 27 Aug 2025).

6. Learning, applications, and scope

The first systematic treatment of correlated Pandora with sample access is "Pandora's Box with Correlations: Learning and Approximation" (Chawla et al., 2019). In that framework, the algorithm does not know the joint distribution explicitly and receives i.i.d. samples from it. The paper proves polynomial sample complexity for learning near-optimal partially adaptive strategies and gives a $v$9-competitive PA algorithm against the optimal PA benchmark for the single-item minimization problem. It also gives constant-factor and $m$0-type results for richer feasibility constraints, while showing that fully adaptive strategies are not efficiently approximable to within any sub-linear factor in the sample-access model (Chawla et al., 2019).

As a baseline, "Generalizing Complex Hypotheses on Product Distributions: Auctions, Prophet Inequalities, and Pandora's Problem" (Guo et al., 2019) provides near-tight sample-complexity results for Pandora’s problem on product distributions, including an upper bound of

$m$1

samples and a lower bound of

$m$2

The paper is explicitly an independence benchmark rather than a correlated result, but it clarifies how much of the classical learning theory depends on product structure (Guo et al., 2019).

Pandora-style reasoning has also been exported to correlated Bayesian optimization. "Cost-aware Bayesian Optimization via the Pandora's Box Gittins Index" (Xie et al., 2024) interprets classical Pandora as a cost-aware black-box optimization problem and defines the Pandora’s Box Gittins Index acquisition function by

$m$3

The method plugs a correlated Gaussian-process posterior into the independent-box Gittins/Pandora formula. The paper is explicit that this is not an exact solution to the fully general correlated Pandora problem, but rather a posterior-plug-in heuristic that performs well empirically, particularly in medium-high dimensions (Xie et al., 2024).

The present scope of the field is therefore stratified. For arbitrary joint distributions over boxes, the minimization problem against the partially adaptive benchmark now has an optimal polynomial-time factor $m$4 approximation (Bansal et al., 21 Sep 2025). Weitzman-style posterior reservation values give a simpler and near-optimal $m$5 approximation, together with polynomial-sample learnability for the partial-update variant (Gergatsouli et al., 2023). Fully adaptive approximation remains tied to decision-tree complexity in the general explicit-support model (Chawla et al., 2021). Structured subclasses, such as Markovian DAG models, can recover exact fully adaptive tractability on forests (Yang et al., 31 Jan 2025). This suggests that the decisive issue is not merely the presence of correlation, but the interaction between correlation, benchmark class, and representational structure.