On the Cost of Non-Adaptivity in Matroid Prophet Inequalities

Abstract: Matroid prophet inequalities admit an optimal 2-competitive algorithm, which relies on adaptively updating thresholds based on previous outcomes. Motivated by applications to posted-price mechanisms and the structural simplicity of fixed-threshold policies, recent work initiated the study of non-adaptive matroid prophet inequalities. The central question is to understand how much the performance deteriorates when thresholds must be fixed in advance. We explore the fundamental limits of non-adaptive algorithms and show new structural barriers and algorithmic insights. We first identify a simple case where non-adaptive algorithms admit a lower bound strictly above $2$: for truncated partition matroids where every local partition has rank $1$, there is an instance giving a lower bound of $\approx 2.179$, and we give a non-adaptive OCRS-style algorithm that exactly matches this ratio. We then show that richer matroid structures can amplify this barrier: we obtain stronger lower bounds of $2.217$ for laminar matroids and $3$ for graphic matroids, and complement the hardness results with improved upper bounds for these matroids.

• The paper quantifies the cost of non-adaptivity in matroid prophet inequalities by proving competitive ratio lower bounds (e.g., 2.179 for truncated R1-partition matroids) using Poisson binomial analysis.
• It amplifies these bounds in complex matroid classes, establishing ratios of 2.217 for laminar and up to 3 for graphic matroids, highlighting the impact of structure on algorithm performance.
• Novel non-adaptive algorithms based on OCRS-style processes and probabilistic partitioning yield improved upper bounds, bridging theoretical limits with practical mechanism design applications.

Cost of Non-Adaptivity in Matroid Prophet Inequalities

Introduction and Motivation

This paper systematically analyzes the competitive loss—quantified as the “cost of non-adaptivity”—for online selection algorithms under matroid constraints when the algorithm is restricted to non-adaptivity, i.e., thresholds must be set in advance, independent of online realizations. Such settings are of central relevance in Bayesian mechanism design and posted-price mechanisms for multidimensional buyers, where adaptivity (thresholds or prices that depend on prior choices) can compromise truthfulness or practical implementation. While adaptive matroid prophet inequalities admit a tight 2-competitive algorithm, the deterioration under the non-adaptive constraint was, prior to this work, poorly understood.

Main Contributions and Technical Insights

The core contribution is a sharp characterization of the structural and algorithmic limits of non-adaptive algorithms for matroid prophet inequalities in several natural classes of matroids. The following are the principal findings, along with their algorithmic and theoretical interpretations:

Truncated $R_1$-Partition Matroids: Exact Separation

For truncated $R_1$-partition matroids (where every local partition has rank 1 in a global cardinality constraint), the paper constructs a specific hard instance showing any non-adaptive algorithm can do no better than a competitive ratio of approximately $2.179$—strictly larger than the adaptive lower bound of $2$. This is achieved via a sophisticated instance that embeds the classic two-variable prophet inequality and utilizes multiple low-value Bernoulli items to force tradeoff decisions for a non-adaptive threshold policy. Notably, a non-adaptive OCRS-style algorithm is analyzed and proven to match this lower bound, yielding an exact characterization of non-adaptivity's cost for this class.

A key technical aspect is a Poisson binomial variable analysis: the tension arises from balancing the probability of consuming options on lower-value items versus preserving global feasibility for rare but high-value late arrivals. The upper bound analysis leverages this structure to deliver precise guarantees.

Amplification in Richer Matroid Families

For more structured matroids, the paper demonstrates amplification of non-adaptivity's cost:

• Laminar Matroids: A recursive construction extends the lower bound to at least $2.217$.
• Graphic Matroids: New constructions exploiting parallel edge structure yield a lower bound of $3$.

Crucially, for simple graphic matroids (i.e., simple graphs), the best-known lower bound remains $2$, and extending the approach to achieve a strict separation remains an open problem.

Improved Upper Bounds via Direct and Relaxed Analysis

Complementing the lower bounds, new non-adaptive algorithms for these matroid classes are proposed:

• The general non-adaptive OCRS-style process from the $R_1$-partition setting is extended and analyzed for truncated partition and laminar matroids, yielding competitive ratios of $3$ and $6.311$, respectively.
• For graphic matroids, an innovative relaxation based on spanning forests, combined with random partitioning (rather than the ex-ante matroid polytope), allows for an $R_1$0-competitive non-adaptive algorithm. With further optimization via probabilistic partitioning and a subsampling prophet inequality bound, the competitive ratio on simple graphs is improved to $R_1$1.

These upper bounds employ a mixture of Markov inequalities, convolution techniques, and coupling arguments, as well as leveraging refined prophet inequality analyses under random subinstance selection.

Complexity of Static vs. Non-Adaptive Thresholds

The paper also provides a thorough examination of when non-adaptive policies offer gains over static thresholds. For $R_1$2-uniform matroids, it is shown that non-adaptive (item-specific, offline) thresholds give no worst-case improvement over static (unified) threshold approaches. For richer matroid structures (partition, laminar, graphic), static thresholds become super-constant in competitive ratio, while non-adaptive item-specific thresholds still yield constant-competitive algorithms in these cases.

Numerical Results and Contradictory Claims

Key numerical findings:

• Exact lower/upper bound for truncated $R_1$3-partition: $R_1$4.
• Amplified lower bounds: $R_1$5 (laminar), $R_1$6 (graphic).
• Matching upper bounds: $R_1$7 (truncated $R_1$8-partition), $R_1$9 (truncated partition), $2.179$0 (laminar), $2.179$1 (graphic), $2.179$2 (simple graphic).

A notable contradictory claim—definitive separation from adaptive policies for these classes—contrasts with previous conjectures that the gap might be purely an artifact of previous analyses or be universally tight at $2.179$3.

Theoretical and Practical Implications

These results elucidate a fundamental barrier—the requirement of fixed thresholds introduces unavoidable loss due to inability to condition on future feasibility, especially as matroid structure becomes more complex (laminar nesting, parallelism in graphics). For mechanism design, this quantifies the price paid to achieve truthfulness and simplicity (e.g., fixed posted prices), which is essential in multi-dimensional and constrained environments.

Algorithmically, knowing the precise (or nearly precise) cost of non-adaptivity informs the design of posted-price mechanisms, online selection algorithms, and guides the interface between prophet inequalities and OCRS in broader Bayesian optimization settings. Open directions include closing the lower/upper gap for laminar and graphic matroids, understanding the simple graphic matroid gap, and further algorithmic development to match lower bounds in the most general settings.

Future Directions

• Closing tightness gaps: For general truncated partition and laminar matroids, can the exact competitive ratio (especially the conjectured $2.179$4) be achieved by extending tight Poisson binomial analysis?
• Simple graphic matroids: Is it possible to construct instances separating adaptive and non-adaptive algorithms for simple graphs?
• Broader matroid classes: Extending these analyses to further exotic/combinatorial matroid types or matroid unions.

Conclusion

The paper provides a comprehensive quantitative and structural theory for the cost of non-adaptivity in matroid prophet inequalities, resolving several open problems and mapping out new ones. The results provide an explicit algorithmic and lower-bound framework for optimal posted-price mechanisms under matroid constraints, revealing nuanced limitations and opportunities in online stochastic optimization and Bayesian mechanism design.

Reference: "On the Cost of Non-Adaptivity in Matroid Prophet Inequalities" (2607.02766)