Pandora’s Box Gittins Index in Cost-Aware BO

• PBGI is a cost-aware Bayesian optimization acquisition rule that applies Pandora’s Box reservation-value logic via an expected-improvement equation.
• It integrates elements of Gittins indices with cost-sensitive decision making, distinguishing itself from classic ratio-based methods.
• Empirical studies indicate that PBGI enhances search efficiency in medium-to-high dimensions and reliably couples evaluation decisions with stopping rules.

Searching arXiv for recent and foundational papers on Pandora’s Box Gittins Index and closely related Pandora/Gittins formulations. Pandora’s Box Gittins Index (PBGI) is a cost-aware acquisition rule for Bayesian optimization that imports the reservation-value logic of Weitzman’s Pandora’s box problem into a posterior-updating setting. In the modern Bayesian-optimization usage, PBGI assigns to each candidate $x$ a scalar $g$ defined implicitly by an expected-improvement indifference equation, and then ranks candidates by that scalar (Xie et al., 2024). The term is not standard across the broader Pandora literature: in classical Pandora’s box and in bandit-learning work on Pandora’s box, the relevant index is usually Weitzman’s reservation value rather than a discounted-bandit Gittins index (Gatmiry et al., 2022). The most precise contemporary usage therefore distinguishes three layers: the classical Weitzman reservation index for independent costly search, the generalized Gittins-based reinterpretation available in certain structured extensions (Olszewski et al., 2015), and the PBGI acquisition used in cost-aware Bayesian optimization as a posterior-substitution heuristic or approximation beyond the independent setting (Xie et al., 2024, Xie et al., 16 Jul 2025).

1. Conceptual origin and terminological scope

The classical Pandora’s box problem considers sequential inspection of boxes with independent hidden prizes and opening costs, with the objective of maximizing expected best-found prize minus inspection costs. In that setting, Weitzman’s rule assigns each box a reservation value $\sigma_i$ determined by

$\mathbb E[(X_i-\sigma_i)^+]=c_i,$

opens boxes in decreasing order of $\sigma_i$, and stops when the best observed value exceeds the next reservation value (Gatmiry et al., 2022, Beyhaghi et al., 2023). This is the foundational Pandora index.

The expression “Pandora’s Box Gittins Index” is newer and more specific. In cost-aware Bayesian optimization, PBGI denotes the scalar $g$ solving a posterior expected-improvement equation of the form

$\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$

or, in cost-per-sample form,

$\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$

depending on the formulation (Xie et al., 2024, Xie et al., 16 Jul 2025). This object is motivated by Pandora’s box, and some tutorial treatments explicitly call the Bayesian-optimization extension “PBGI,” while also noting that exact Gittins optimality applies only in the independent Pandora case, not in correlated Gaussian-process settings (Scully et al., 12 Jun 2025).

The term should not be conflated with the classical discounted Gittins index of multi-armed bandits. The discounted Gittins index is defined through a retirement-option dynamic program or stopping-time ratio for repeated sampling of a Markov arm (Edwards, 2019, Baas et al., 2023). By contrast, the index used in classical Pandora’s box is a one-shot search reservation value, not a discounted retirement index (Gatmiry et al., 2022). A careful reading of the literature therefore supports a qualified usage: PBGI is a Pandora-inspired, Gittins-framed index in Bayesian optimization, whereas the exact index in the original Pandora problem remains Weitzman’s reservation value (Gatmiry et al., 2022, Scully et al., 12 Jun 2025).

2. Classical Pandora structure and the Weitzman index

In the standard independent-box model, box $i$ has known opening cost $c_i$ and prize distribution $g$0. If $g$1 denotes the opened boxes and $g$2 the discovered prizes, the classical objective is

$g$3

The optimal policy is characterized by reservation values $g$4 or $g$5, defined equivalently by

$g$6

at the stopping boundary, or in equality form

$g$7

for the optimal reservation value (Olszewski et al., 2015, Beyhaghi et al., 2023).

A closely related gain-function formulation appears in the bandit-learning study of Pandora’s box: $g$8 with optimal threshold $g$9 satisfying

$\sigma_i$0

The interpretation given there is that $\sigma_i$1 is the expected additional utility from opening box $\sigma_i$2 when current value $\sigma_i$3 is already in hand (Gatmiry et al., 2022).

This reservation value fully determines both ordering and stopping. In the known-distribution setting, the optimal policy is: compute $\sigma_i$4 for every box, inspect in decreasing order of $\sigma_i$5, and stop when the current best observed value exceeds the next box’s reservation value (Gatmiry et al., 2022). That index role is exact and classical; it is the main structural ancestor of PBGI.

A useful interpretive refinement comes from the tutorial treatment that embeds classical Pandora’s box into a general Markov-chain-selection framework. There, the Pandora reservation value is presented as a Gittins index in the exact independent case, with expected improvement

$\sigma_i$6

and index $\sigma_i$7 determined by

$\sigma_i$8

Under that interpretation, an unopened box has index $\sigma_i$9, an already opened box has index equal to its realized value, and the optimal rule is “always take the action of greatest Gittins index” (Scully et al., 12 Jun 2025). This identification is exact for independent required-inspection Pandora, but does not extend unchanged to correlated settings.

3. Generalized Pandora rules and the Gittins connection

The strongest formal bridge between Pandora-style reservation values and genuine Gittins-index reasoning is the characterization of generalized Pandora problems in which a Gittins reduction exists (Olszewski et al., 2015). There, the objective is generalized from dependence on only the best prize to dependence on all discovered prizes: $\mathbb E[(X_i-\sigma_i)^+]=c_i,$0 A generalized reservation value is then defined by

$\mathbb E[(X_i-\sigma_i)^+]=c_i,$1

The paper proves a sharp characterization: if a generalized Pandora rule is to be optimal for all costs and prize distributions, then the utility must be “Strongly Pandora’s Rule compatible” (SPR), meaning

$\mathbb E[(X_i-\sigma_i)^+]=c_i,$2

Equivalently, with ordered prizes $\mathbb E[(X_i-\sigma_i)^+]=c_i,$3,

$\mathbb E[(X_i-\sigma_i)^+]=c_i,$4

Under SPR, the generalized Pandora problem can be transformed into a multi-armed bandit target process, and the Gittins index theorem yields an optimal policy (Olszewski et al., 2015).

In that reduction, the Gittins calibration equation is

$\mathbb E[(X_i-\sigma_i)^+]=c_i,$5

while the generalized Pandora reservation equation is

$\mathbb E[(X_i-\sigma_i)^+]=c_i,$6

The identification

$\mathbb E[(X_i-\sigma_i)^+]=c_i,$7

shows that the Gittins index and the generalized reservation value are monotone-equivalent, hence induce the same policy (Olszewski et al., 2015).

This result is important for PBGI because it clarifies what is and is not being claimed. It shows that a rigorous Pandora-to-Gittins equivalence exists, but only in a restricted structural class. Outside SPR, generalized Pandora rules need not be true indices in the Gittins sense, even if reservation-value heuristics remain useful (Olszewski et al., 2015). Consequently, later PBGI terminology in Bayesian optimization should be read as conceptually Gittins-like, not as a universal extension theorem.

4. PBGI in cost-aware Bayesian optimization

The modern technical meaning of PBGI comes from cost-aware Bayesian optimization. In this setting, an unknown black-box function $\mathbb E[(X_i-\sigma_i)^+]=c_i,$8 is queried sequentially, with each evaluation at $\mathbb E[(X_i-\sigma_i)^+]=c_i,$9 incurring cost $\sigma_i$0. One of the main formulations uses simple regret plus cumulative evaluation cost; another uses expected budget constraints (Xie et al., 2024, Xie et al., 16 Jul 2025).

The defining PBGI acquisition in the budget-constrained formulation is

$\sigma_i$1

with $\sigma_i$2 a cost-scaling hyperparameter (Xie et al., 2024). For the cost-per-sample formulation, the corresponding equation is

$\sigma_i$3

for unevaluated $\sigma_i$4, while already evaluated points are assigned their observed value (Xie et al., 16 Jul 2025). In both cases the acquisition value is the baseline level $\sigma_i$5 at which expected one-step improvement exactly matches evaluation cost.

This differs from expected improvement per cost,

$\sigma_i$6

in a decisive way. EIPC is ratio-based and invariant to multiplying all costs by a constant, whereas PBGI is threshold-based and becomes more conservative when the overall cost scale rises (Xie et al., 2024). That non-multiplicative cost sensitivity is one of its main conceptual motivations.

The Bayesian-optimization literature explicitly interprets PBGI as a posterior-substitution of the exact Pandora rule. In the independent finite Pandora setting,

$\sigma_i$7

is Bayes-optimal (Xie et al., 2024). General Gaussian-process Bayesian optimization violates the independence assumption because evaluating one point changes the posterior at other points. The PBGI heuristic proceeds by updating the posterior correctly and then acting as if each candidate’s marginal posterior could be scored by the Pandora reservation equation (Xie et al., 2024, Scully et al., 12 Jun 2025). A tutorial treatment states this explicitly: classical exactness holds for independent Pandora, while the Bayesian-optimization PBGI is a correlated-Pandora extension that “accounts for correlations from the past, but disregard correlations in the future” (Scully et al., 12 Jun 2025).

Under a Gaussian posterior $\sigma_i$8, expected improvement has the closed form

$\sigma_i$9

so PBGI amounts to solving a one-dimensional root equation in $g$0 for each $g$1 (Xie et al., 2024). The same work derives the acquisition gradient

$g$2

which permits standard continuous optimization over $g$3 (Xie et al., 2024).

The same paper introduces PBGI-D, a dynamic-decay variant in which $g$4 is replaced by a time-varying $g$5, motivated by robustness to hyperparameter misspecification (Xie et al., 2024). Unknown-cost variants replace $g$6 by a posterior moment from a log-GP cost model: $g$7 in the defining equation (Xie et al., 2024).

5. Stopping rules and cost guarantees

A central consequence of importing Pandora logic into Bayesian optimization is that “where to sample next” and “when to stop” become coupled. The cost-aware stopping paper makes this explicit and treats PBGI mainly as the acquisition half of an acquisition–stopping pair (Xie et al., 16 Jul 2025).

In the classical independent Pandora setting, the exact adaptive stopping rule is to stop when the best remaining unopened index no longer beats the best observed reward. In the Bayesian-optimization notation used there, the stopping rule is

$g$8

where $g$9 is the best observed value so far (Xie et al., 16 Jul 2025). By the defining equation of PBGI, this is equivalent to

$\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$0

Thus the operational rule is: stop when no candidate’s expected improvement is worth its evaluation cost (Xie et al., 16 Jul 2025).

The paper also defines a related acquisition

$\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$1

whose stopping condition $\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$2 is equivalent to PBGI stopping in the known-cost case (Xie et al., 16 Jul 2025).

Its main theorem provides a stopping-cost guarantee. If $\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$3 is compact, $\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$4 is a random function with constant mean $\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$5, and

$\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$6

then for a Bayesian-optimization algorithm using either PBGI or LogEIPC acquisition together with the PBGI/LogEIPC stopping rule,

$\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$7

where $\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$8 is the initial evaluation cost and $\mathrm{EI}_{f\mid y_{1:t}}(x;g)=\lambda c(x)$9 is the stopping time (Xie et al., 16 Jul 2025). The proof rests on the fact that before stopping, the selected point always satisfies

$\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$0

so each evaluation pays for itself in expected improvement (Xie et al., 16 Jul 2025).

This result does not prove Bayes-optimality for correlated Bayesian optimization. It is instead a guarantee on cumulative cost at stopping, and the paper is explicit that exact Bayes-optimality remains confined to the independent Pandora case (Xie et al., 16 Jul 2025).

6. Variants, generalizations, and limits of indexability

The broader Pandora literature makes clear that exact index structure is delicate. A survey emphasizes that the term PBGI itself is absent from most Pandora work, but also notes that Weitzman’s reservation value is “a special case of indices proposed by Gittins,” and points to staged inspection, order-constrained search, and branching-bandit connections as natural generalizations of reservation-value policies (Beyhaghi et al., 2023).

One such generalization is CVaR-aware Pandora’s box. There, the objective is

$\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$1

and the paper proves an exact Weitzman-style solution after a one-dimensional variational reduction (Ji, 18 May 2026). For fixed $\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$2, define transformed prizes

$\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$3

then define $\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$4 by

$\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$5

After optimizing the scalar $\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$6, opening boxes in decreasing $\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$7 is optimal (Ji, 18 May 2026). The paper describes this as retaining “the index structure of the classical problem” and even calls the resulting reservation value “a Gittins-type index” (Ji, 18 May 2026). This is not PBGI in the Bayesian-optimization sense, but it shows that Pandora-style indexability can survive non-expected-value objectives.

By contrast, the bandit-learning work on Pandora’s box under utility-only feedback is explicit that its “indices” are still just Weitzman reservation values, not Gittins indices: $\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$8 That work learns confidence intervals for $\mathrm{EI}_{f\mid x_{1:t},y_{1:t}}(x;g)=c(x),$9 and their ordering from repeated rounds, but does not introduce a separate Gittins-style dynamic-allocation index (Gatmiry et al., 2022). This is relevant because PBGI terminology can otherwise suggest a stronger formal connection than the underlying mathematics supports.

A further limitation emerges in generalized Pandora problems. Outside the SPR-compatible class, a generalized reservation rule need not be a true state-independent index and need not be optimal for all costs and prize distributions (Olszewski et al., 2015). Likewise, in correlated Bayesian optimization, PBGI is a principled acquisition derived from Pandora logic, but not one supported by the exact Gittins theorem (Xie et al., 2024, Scully et al., 12 Jun 2025).

7. Computation, empirical behavior, and robustness issues

The computational burden of PBGI in Bayesian optimization is moderate relative to nonmyopic lookahead methods because each acquisition evaluation requires only a one-dimensional root solve. The original cost-aware BO paper reports using 100 iterations of bisection without early stopping (Xie et al., 2024), and the stopping-rule paper follows that implementation choice directly (Xie et al., 16 Jul 2025). Acquisition optimization over $i$0 is then handled by multi-start continuous optimization; one reported setup uses $i$1 Sobol points, $i$2 starts, and multi-start L-BFGS-B (Xie et al., 2024).

Empirically, PBGI performs especially well in medium- to high-dimensional settings and often outperforms cost-aware EI-per-cost baselines in multimodal problems, while sometimes lagging on unimodal tasks such as Rosenbrock or certain robotics benchmarks (Xie et al., 2024). The proposed stopping rule paired with PBGI is reported to “consistently match or outperform” other acquisition–stopping combinations on cost-adjusted simple regret and to stop reliably when other methods frequently fail to stop within a budget cap (Xie et al., 16 Jul 2025).

Several computational lessons from the broader Gittins literature carry over structurally, even though not all formulas apply verbatim. Classical Gittins computation relies on a one-armed retirement option, dynamic programming, and root-finding in the outside-option reward $i$3 (Gui et al., 2019). Sampling-based approximation work further shows that if an index can be represented as the unique root of a charge-based optimal stopping value, then one can combine horizon truncation, bounded-state stopping-value approximation, nested simulation, stochastic approximation, and explicit confidence intervals for the resulting index estimate (Baas et al., 2023). This suggests a computational pathway for more complex PBGI variants, though such a transfer requires a corresponding stopping-value representation for the Pandora-derived index itself.

Robustness is another concern. The robust Gittins scheduling paper shows that naive use of exact indices computed from slightly misspecified distributions can be highly non-robust: even arbitrarily small multiplicative tail perturbations can make the classical Gittins policy incur unbounded approximation loss in scheduling (Moseley et al., 14 Apr 2025). Its remedy is to preserve the predicted index order while extending continuation quanta by a multiplicative slack factor (Moseley et al., 14 Apr 2025). This is not a Pandora or PBGI result, but it highlights a general phenomenon relevant to index-based search: when an index depends sensitively on conditional continuation distributions, mild model error can distort policy quality sharply. A plausible implication is that PBGI may require analogous robustness analysis under posterior or cost-model misspecification.

A final practical distinction concerns learning. In classical Gittins problems with unknown transitions, reinforcement-learning and sampling-based methods now exist that learn the retirement threshold or index directly from experience, rather than first estimating the full model (Dhankhar et al., 2024, Baas et al., 2023). By contrast, current PBGI implementations in Bayesian optimization compute the index from a maintained posterior model rather than learning the index itself end-to-end (Xie et al., 2024, Xie et al., 16 Jul 2025). This is a methodological difference between PBGI as currently used and classical unknown-model Gittins methods.

8. Interpretation and contemporary significance

PBGI sits at the intersection of three traditions. From Pandora’s box it inherits the reservation-value viewpoint: compare each uncertain option to an outside alternative and assign a scalar threshold at indifference (Gatmiry et al., 2022, Beyhaghi et al., 2023). From Gittins theory it inherits the broader design principle that optimal choice can sometimes be reduced to ranking options by a scalar index derived from a single-option calibration problem (Olszewski et al., 2015, Scully et al., 12 Jun 2025). From Bayesian optimization it inherits posterior adaptation, continuous action spaces, and cost-aware acquisition design (Xie et al., 2024, Xie et al., 16 Jul 2025).

Its contemporary significance lies less in establishing a new universal theorem than in clarifying a useful structural approximation. In the exact independent Pandora problem, the relevant index is the classical Weitzman reservation value, which some modern treatments legitimately regard as a Gittins index under an appropriate Markov-chain-selection embedding (Scully et al., 12 Jun 2025). In generalized SPR-compatible Pandora problems, a true Gittins reduction is available (Olszewski et al., 2015). In cost-aware Bayesian optimization, the same scalar-indifference logic becomes the PBGI acquisition

$i$4

which is not exact Gittins optimality but remains a principled cost-aware rule (Xie et al., 2024).

A persistent source of confusion is therefore terminological. “PBGI” is best reserved for the Bayesian-optimization acquisition and closely related correlated-search adaptations (Xie et al., 2024, Xie et al., 16 Jul 2025), while “reservation value” remains the correct term for the exact index in classical Pandora’s box (Gatmiry et al., 2022). The connection to Gittins is real but qualified: sometimes exact through a reduction, often conceptual through indexability, and in the BO setting explicitly heuristic beyond the independent model (Olszewski et al., 2015, Scully et al., 12 Jun 2025).

On that reading, Pandora’s Box Gittins Index is not a single monolithic object but a family resemblance class of scalar search scores. Its canonical exact member is Weitzman’s reservation value; its most prominent modern extension is the cost-aware Bayesian-optimization acquisition defined by a posterior EI–cost indifference equation; and its broader research agenda concerns when such scalar rankings remain exact, approximately valid, learnable, robust, and computationally tractable under richer objectives and informational structures (Xie et al., 2024, Xie et al., 16 Jul 2025, Ji, 18 May 2026).