Secretary, Prophet, and Stochastic Probing via Big-Decisions-First

Abstract: We revisit three fundamental problems in algorithms under uncertainty: the Secretary Problem, Prophet Inequality, and Stochastic Probing, each subject to general downward-closed constraints. When elements have binary values, all three problems admit a tight $\tildeΘ(\log n)$-factor approximation guarantee. For general (non-binary) values, however, the best known algorithms lose an additional $\log n$ factor when discretizing to binary values, leaving a quadratic gap of $\tildeΘ(\log n)$ vs. $\tildeΘ(\log² n)$. We resolve this quadratic gap for all three problems, showing $\tildeΩ(\log² n)$-hardness for two of them and an $O(\log n)$-approximation algorithm for the third. While the technical details differ across settings, and between algorithmic and hardness proofs, all our results stem from a single core observation, which we call the Big-Decisions-First Principle: Under uncertainty, it is better to resolve high-stakes (large-value) decisions early.

• The paper introduces the Big-Decisions-First principle, prioritizing high-value decisions to achieve near-tight approximations in uncertain online selection settings.
• It presents an O(log n)-competitive algorithm for the secretary problem and establishes quadratic lower bounds for the prophet inequality under adversarial orders.
• By analyzing adaptive versus non-adaptive strategies in stochastic probing, the study highlights structural gaps that shape algorithm performance under uncertainty.

Big-Decisions-First Principle in Secretary, Prophet, and Stochastic Probing Problems

Introduction and Problem Landscape

This paper investigates approximation guarantees for three canonical online selection problems under uncertainty, all involving downward-closed feasibility constraints: the Secretary Problem, Prophet Inequality, and Stochastic Probing. In each setting, the objective is to construct a subset of $n$ elements (with uncertain values) maximizing a set function (additive, XOS, or submodular), where selection decisions are irrevocable and constrained.

The primary focus is elucidating the collection of existing lower and upper bounds in terms of their relationship to the value distribution of the elements—specifically, the difference in tightness between binary-valued and general-valued (weighted) instances. For binary values (i.e., $\{0,1\}$), all three problems are settled up to logarithmic factors with $\tilde{\Theta}(\log n)$ approximation. However, for general values, previous discretization-based reductions amplified this to $\tilde{\Theta}(\log^2 n)$, incurring a quadratic gap. The paper resolves this gap, establishing tight (up to $\operatorname{poly}(\log\log n)$) bounds and revealing an instructive separation between random-arrival and adversarial-arrival models.

The Big-Decisions-First Principle

The paper identifies a unifying structural phenomenon—the Big-Decisions-First Principle—driving both algorithmic design and lower bound constructions across all three problems. The core observation is that in the presence of uncertainty, resolving larger-value (higher-stakes) decisions as early as possible leads to tighter approximations. The extent to which each problem allows reordering (or influences the order) of big versus small decisions is pivotal to both achievable approximations and inapproximability gaps.

This principle formally dictates that if an online algorithm can process value buckets in the order of decreasing value (i.e., resolve high-value elements first), union bounds over low-value element selection do not propagate loss to high-value approximations. Conversely, when forced to process low-value elements first, as by an adversarial order or combinatorial constraint structure, the union-bound effect is cumulative, yielding a multiplicative gap.

Main Results

Secretary Problem

For the Secretary Problem (uniformly random order), the random permutation of elements is sufficient to allow the "big decisions first" strategy to yield an $O(\log n)$-approximation for general values, matching the existing lower bound up to $\operatorname{poly}(\log\log n)$, and closing the aforementioned quadratic gap. This result holds even for XOS functions, extending the result's applicability beyond additivity. The construction leverages random partitioning, careful value bucketing, and interleaving selection rounds, ensuring large-value elements are addressed in near-optimal order—a scenario feasible only under random arrival.

Prophet Inequality

In the Prophet Inequality (adversarial order), the analysis introduces a "double error-correcting code" structure in the feasibility constraint construction, forcing the algorithm to resolve small-value decisions first and tightly linking the selection of low-value and high-value elements. The resulting lower bound instance achieves $\tilde{\Omega}(\log^2 n)$ inapproximability for general values, demonstrating that no online algorithm can break the quadratic barrier imposed by worst-case arrivals and downward-closed constraints. The construction leverages combinatorial coding theory to engineer instance families for which any deviation from optimal path selection leads to irrevocable loss from big-value buckets.

Stochastic Probing

For Stochastic Probing, the adaptivity gap (ratio of value of adaptive to non-adaptive policy) is shown to also be $\tilde{\Omega}(\log^2 n)$, via a blockwise concatenated-tree construction (akin to meta-trees and blocks). The outer (probing) and inner (selection) constraints together enforce constraints echoing the "worst-case first" dynamics of the Prophet Inequality, despite allowing more adaptive flexibility. The analysis establishes that probing adaptivity is necessary: any non-adaptive policy cannot achieve better than a multiplicative $\tilde{\Theta}(\log^2 n)$ gap, precluding the existence of a universal non-adaptive strategy with better performance in such downward-closed environments.

The following table summarizes the main tight bounds established in the paper:

Problem	Order of Arrival	Upper Bound	Lower Bound	Tight up to
Secretary	Random	$\{0,1\}$0	$\{0,1\}$1	$\{0,1\}$2
Prophet Inequality	Adversarial	$\{0,1\}$3	$\{0,1\}$4	$\{0,1\}$5
Stochastic Probing	Adaptive	$\{0,1\}$6	$\{0,1\}$7	$\{0,1\}$8

Technical Contributions

Tight Algorithm for Secretary with General Values/XOS

The $\{0,1\}$9-competitive algorithm for general-valued secretary instances partitions the element stream into $\tilde{\Theta}(\log n)$0 buckets by value (powers of two). Within each round, optimal subsets are selected among elements of uniform scale, compatible with prior higher-value selections. The analysis critically addresses overfitting risks between sample, current, and real elements, ensuring concentration bounds are sufficient to union bound over all feasible selections. The analysis extends to XOS objectives using a clause-based bucketing and dynamic programming across phases.

Hard Instance for Prophet Inequality: Double Error-Correcting Codes

The prophet lower-bound construction generates a layered forest instance with combinatorially-structured intersection properties, ensuring that selection of a small fraction of low-value elements commits any algorithm to a specific feasible path, with all remaining large-value opportunities forfeited. The use of double error-correcting codes guarantees high distance both among siblings and among distinct feasible sets, precluding any significant overlap unless the selection aligns exactly with the offline optimum.

Large Adaptivity Gap for Stochastic Probing

The stochastic probing instance couples meta-trees (as in prophet construction) with blockwise tree codes, embedding local and global separation constraints. The outer (probing) constraint is defined over caterpillar graphs, while the inner constraint restricts to paths. The analysis shows any deviation from the path structure via non-adaptive probing leads to only $\tilde{\Theta}(\log n)$1 proportional value, yielding a tight adaptivity gap.

Broader Implications and Future Directions

The results establish that, contrary to prior expectations arising from similarities in the secretary and prophet settings for binary instances, the order of arrival (random vs worst-case) can have a dramatic effect—logarithmic vs quadratic gaps—in the presence of general value distributions and rich combinatorial (downward-closed) constraints. This separation demonstrates fundamental limits on reductions and sample-based approaches for complex constraint systems.

From a theoretical standpoint, the "big-decisions-first" paradigm offers a lens for analyzing selection under uncertainty in other online and combinatorial settings, especially those where feasible set structure interacts adversarially with the value distribution. Practically, the findings suggest that in mechanism design, online auctions, and adaptive probing models, architectural decisions about the order of access or information revelation have an outsized impact on approximability guarantees.

Potential future research may explore analogous gaps for submodular or more general subadditive objectives, tighten the remaining $\tilde{\Theta}(\log n)$2 slack, or extend the analysis to multi-dimensional or dynamic constraint systems. Additionally, coding-theoretic reductions embedded in feasibility structures could yield further applications in complexity-theoretic lower bounds for online algorithms.

Conclusion

This paper provides a nearly complete resolution of the tight approximation factors for secretary, prophet, and stochastic probing problems with general downward-closed feasibility constraints and general value distributions. By revealing a structural principle governing the interplay between uncertainty, element value, and order of decision-making, it identifies both the power and limits of online selection algorithms across adaptive, adversarial, and random-arrival environments. The results advance both algorithmic understanding and lower-bound methodologies in constrained online selection problems, with lasting implications for both theory and application.