Sample-Based Matroid Prophet Inequality

• The paper introduces a framework for prophet inequalities under matroid constraints using limited sample information to achieve competitive ratios against an offline benchmark.
• Key methodologies include reductions from order-oblivious secretary algorithms and direct threshold-based policies that maintain feasibility while optimizing online selections.
• The study offers mechanism-design insights and outlines open questions on the equivalence of SSPI approaches and the inherent sample complexity needed for general matroid settings.

Searching arXiv for relevant papers on sample-based matroid prophet inequalities and single-sample matroid prophet inequalities. Sample-based matroid prophet inequality is the study of prophet inequalities under matroid feasibility when the value distributions are unknown and the algorithm has access only to samples from those distributions. In the standard model, a ground set $E$ is endowed with a matroid $M=(E,\mathcal{I})$; each element $e\in E$ has an independent nonnegative random value drawn from a distribution $D_e$; elements arrive online in an adversarial order; and the algorithm must irrevocably accept or reject each arrival while maintaining independence in $M$. The benchmark is the offline prophet’s maximum-weight independent set, and the objective is to obtain a constant fraction of $\mathbb{E}[\mathrm{OPT}]$ using either a single sample per element or a larger but still limited sample budget. The literature has evolved from single-sample guarantees for special matroid classes to $(\tfrac14-\varepsilon)$-competitive algorithms for general matroids with polylogarithmic samples, and then to nearly tight sample complexity bounds for that regime (Azar et al., 2013, Fu et al., 2024, Feldman et al., 13 Jul 2025).

1. Formal model and benchmark

A matroid prophet inequality instance consists of a finite ground set $E$, a matroid $M=(E,\mathcal{I})$, and independent nonnegative random values on the elements. In the single-sample model, the algorithm receives one independent sample $s_e$ from each $M=(E,\mathcal{I})$0 before the online phase; in the multi-sample model, it receives $M=(E,\mathcal{I})$1 i.i.d. samples from each $M=(E,\mathcal{I})$2. During the online phase, realized values are revealed one by one in an adversarial order, and the algorithm must preserve feasibility by accepting only sets in $M=(E,\mathcal{I})$3 (Caramanis et al., 2021, Fu et al., 2024).

The offline benchmark is

$M=(E,\mathcal{I})$4

and a policy is $M=(E,\mathcal{I})$5-competitive if

$M=(E,\mathcal{I})$6

equivalently $M=(E,\mathcal{I})$7-competitive in the reciprocal convention

$M=(E,\mathcal{I})$8

Several papers analyze the strongest adversarial model, an almighty fully adaptive adversary that knows all samples and realizations and may choose the arrival order adaptively (Caramanis et al., 2021, Caramanis et al., 2021).

This formulation differs from the classical full-information matroid prophet inequality by removing access to the distributions themselves. The resulting technical problem is not merely threshold estimation: in the general matroid setting, limited-information algorithms must simultaneously infer useful thresholds, preserve independence online, and cope with adversarial order without the distributional primitives used in full-information proofs (Fu et al., 2024).

2. Secretary reductions and the first single-sample guarantees

The first systematic limited-information results for matroid prophet inequalities were obtained through a black-box reduction from order-oblivious secretary algorithms to single-sample prophet inequalities. In this framework, one sample vector is used to simulate the observation phase of an order-oblivious secretary policy, and the actual online values are then fed through the secretary algorithm’s acceptance rule. If the secretary algorithm is $M=(E,\mathcal{I})$9-competitive and order-oblivious, the induced single-sample prophet inequality inherits the same $e\in E$0 guarantee (Azar et al., 2013).

This reduction yielded the first single-sample prophet inequalities for several matroid classes under adversarial arrival. The stated guarantees include graphic matroids with factor $e\in E$1, transversal matroids with factor $e\in E$2, laminar matroids with factor $e\in E$3, and general matroids under the i.i.d. random assignment model with factor $e\in E$4 (Azar et al., 2013). The same work also introduced a novel single-sample algorithm for the $e\in E$5-uniform matroid, achieving

$e\in E$6

which asymptotically matches the best possible ratio even with full distributional knowledge (Azar et al., 2013).

The secretary-reduction paradigm established the first broad bridge between sample-based prophet inequalities and matroid secretary algorithms. At the same time, later work identified a structural limitation of this route: it does not fully exploit the available samples, because some rewards are never observed online and some samples are never used in setting acceptance thresholds. This loss became a central motivation for direct single-sample analyses that do not pass through secretary reductions (Caramanis et al., 2021).

3. Direct single-sample matroid methods and the $e\in E$7-partition framework

Direct single-sample analyses replaced the secretary reduction by threshold policies derived directly from the samples. A central technique is the greedy sample-path or greedy-ordered selection framework, based on a coupling in which each element receives two i.i.d. draws and a fair coin decides which draw becomes the sample and which becomes the realized reward. This symmetry permits pointwise comparison between a greedy benchmark built from samples and a greedy-like prophet solution built from rewards (Caramanis et al., 2021, Caramanis et al., 2021).

For matroids, the key abstraction is the $e\in E$8-partition property. In one formulation, after observing a sample-dependent subset $e\in E$9, one constructs a simple partition matroid $D_e$0 on $D_e$1, partitioned into groups $D_e$2, such that

$D_e$3

For each group $D_e$4, the threshold is

$D_e$5

and the online rule accepts $D_e$6 iff no earlier element from $D_e$7 was accepted and $D_e$8 (Caramanis et al., 2021).

The resulting meta-theorem states that any matroid satisfying an $D_e$9-partition property admits a $M$0-competitive single-sample prophet inequality: $M$1 If the partition transformation is polynomial-time computable, the policy is efficient (Caramanis et al., 2021). In the equivalent “weak $M$2-partition property” formulation, the same $M$3-competitive conclusion holds, again by reducing the matroid instance to disjoint rank-$M$4 groups and applying the optimal $M$5-competitive single-choice single-sample policy inside each group (Caramanis et al., 2021).

Representative improvements over the earlier order-oblivious-secretary route are summarized below.

Matroid family	Previous single-sample factor	Direct single-sample factor
Graphic matroid	$M$6	$M$7
Co-graphic matroid	$M$8	$M$9
Low-density matroid	$\mathbb{E}[\mathrm{OPT}]$0	$\mathbb{E}[\mathrm{OPT}]$1
Column $\mathbb{E}[\mathrm{OPT}]$2-sparse linear matroid	$\mathbb{E}[\mathrm{OPT}]$3	$\mathbb{E}[\mathrm{OPT}]$4
Transversal matroid	$\mathbb{E}[\mathrm{OPT}]$5	$\mathbb{E}[\mathrm{OPT}]$6

Here $\mathbb{E}[\mathrm{OPT}]$7 is the density parameter (Caramanis et al., 2021). The factor-of-$\mathbb{E}[\mathrm{OPT}]$8 improvement arises because prior secretary-based reductions used a $\mathbb{E}[\mathrm{OPT}]$9-competitive single-choice core, whereas the direct matroid constructions use the optimal $(\tfrac14-\varepsilon)$0-competitive single-sample single-choice policy of Rubinstein–Wang–Weinberg inside each rank-$(\tfrac14-\varepsilon)$1 component (Caramanis et al., 2021, Caramanis et al., 2021).

These direct methods also produced specialized single-sample results that are matroidal or matroid-adjacent but do not fit solely through the partition abstraction. The same line of work gave an $(\tfrac14-\varepsilon)$2-competitive policy for transversal matroids via bipartite vertex-arrival analysis and an $(\tfrac14-\varepsilon)$3-competitive policy for truncated partition matroids, both under an almighty adversary (Caramanis et al., 2021).

4. Uniform matroids: asymptotics and exact rank-$(\tfrac14-\varepsilon)$4 optimality

Uniform matroids have been a testing ground for how far single-sample information can go under matroid feasibility. For the $(\tfrac14-\varepsilon)$5-uniform matroid, the “Rehearsal” algorithm sets thresholds from a single sample vector and accepts online values into thresholded slots. Its competitive ratio is

$(\tfrac14-\varepsilon)$6

which was presented as asymptotically matching the best possible even with full distributional knowledge (Azar et al., 2013).

A later sharp result resolved the rank-$(\tfrac14-\varepsilon)$7 case exactly for deterministic mechanisms. For the uniform matroid $(\tfrac14-\varepsilon)$8, the algorithm sets a single threshold $(\tfrac14-\varepsilon)$9 equal to the second largest sample value and accepts any arriving realized value $E$0 with $E$1 until two items have been taken. The theorem states that

$E$2

even under adversarial arrival order; equivalently, the mechanism is $E$3-competitive (Pashkovich et al., 2023).

The same paper proves that no deterministic single-sample mechanism can achieve a competitive ratio strictly greater than $E$4 for $E$5. Thus, in rank $E$6, the simple rule “threshold at the second largest sample, accept up to two realizations above it” is optimal among deterministic mechanisms (Pashkovich et al., 2023). The analysis proceeds through a two-point paired reduction, explicit formulas for the prophet’s selection probabilities $E$7, lower bounds $E$8 for the gambler’s acceptance probabilities, and prefix inequalities of the form

$E$9

which imply the global $M=(E,\mathcal{I})$0 bound (Pashkovich et al., 2023).

This rank-$M=(E,\mathcal{I})$1 theorem also motivates a broader conjecture: for $M=(E,\mathcal{I})$2, setting the threshold to the $M=(E,\mathcal{I})$3-th largest sample and accepting up to $M=(E,\mathcal{I})$4 realizations above that threshold may remain $M=(E,\mathcal{I})$5-competitive for all $M=(E,\mathcal{I})$6 (Pashkovich et al., 2023). That conjecture remains open.

5. General matroids with polylogarithmic samples: quantiles, OCRS, and near-optimal sample complexity

For general matroids, single-sample constant-factor competitiveness is closely entangled with the matroid secretary problem, and no constant-factor algorithm with even a sublinear number of samples was known until the development of sample-based OCRS methods. The 2024 breakthrough gives a $M=(E,\mathcal{I})$7-competitive matroid prophet inequality using $M=(E,\mathcal{I})$8 samples per element (Fu et al., 2024).

The construction has two parts. First, it uses a quantile-based reduction from matroid prophet inequalities to online contention resolution schemes. For each element $M=(E,\mathcal{I})$9, one defines the exchange threshold

$s_e$0

From $s_e$1 sampled instances, the algorithm forms empirical threshold levels $s_e$2, and on a fresh online draw activates $s_e$3 with probability $s_e$4 whenever $s_e$5. The resulting activation vector $s_e$6 satisfies $s_e$7, the matroid polytope, and the active set preserves almost all prophet value: $s_e$8 when the learned quantiles are accurate (Fu et al., 2024).

Second, the active set is filtered by a sample-based matroid OCRS. In the OCRS language, one seeks $s_e$9-selectability: $M=(E,\mathcal{I})$00 The 2024 construction obtains a $M=(E,\mathcal{I})$01-selectable sample-based matroid OCRS using $M=(E,\mathcal{I})$02 samples, and hence a $M=(E,\mathcal{I})$03-competitive matroid prophet inequality against an almighty adversary (Fu et al., 2024).

A 2025 refinement improves the OCRS sample complexity to near optimal. It gives a polynomial-time sample-based randomized OCRS for any matroid of rank $M=(E,\mathcal{I})$04 with selectability $M=(E,\mathcal{I})$05, using

$M=(E,\mathcal{I})$06

samples per element to construct the scheme. Via the reduction of Fu–Lu–Tang–Wu–Wu–Zhang, this yields a sample-based matroid prophet inequality with

$M=(E,\mathcal{I})$07

and total per-variable sample complexity

$M=(E,\mathcal{I})$08

This matches the best known competitiveness for the almighty-adversary setting even when the distributions are fully known (Feldman et al., 13 Jul 2025).

The OCRS construction is chain-based. It samples a distribution over spanning chains, measures “chain-freeness” for each element, and uses a chain-to-OCRS reduction: a $M=(E,\mathcal{I})$09-balanced spanning chain distribution for $M=(E,\mathcal{I})$10 yields a $M=(E,\mathcal{I})$11-selectable OCRS. The balance theorem states that for $M=(E,\mathcal{I})$12, $M=(E,\mathcal{I})$13, and $M=(E,\mathcal{I})$14, the sampled chain is $M=(E,\mathcal{I})$15-balanced; setting $M=(E,\mathcal{I})$16 gives selectability approximately $M=(E,\mathcal{I})$17 (Feldman et al., 13 Jul 2025).

The sample complexity is nearly tight. A lower bound from Fu–Lu–Tang–Wu–Wu–Zhang shows that even offline contention resolution requires $M=(E,\mathcal{I})$18 samples to obtain any constant selectability; for matroids with $M=(E,\mathcal{I})$19, this implies $M=(E,\mathcal{I})$20. Hence the upper bound

$M=(E,\mathcal{I})$21

is tight up to a $M=(E,\mathcal{I})$22 factor (Feldman et al., 13 Jul 2025).

6. Mechanism-design consequences, structural limits, and open questions

Sample-based matroid prophet inequalities were originally motivated in part by prior-independent posted-price mechanisms. The secretary-reduction framework already implied limited-information posted-price and multidimensional auction mechanisms for matroidal environments, including welfare guarantees that track the underlying prophet-inequality factor and revenue guarantees obtained by combining comparison-based selection with sample reserves under regular or MHR assumptions (Azar et al., 2013).

Direct single-sample analyses improved these mechanism-design consequences by improving the underlying prophet factors. In particular, improved SSPIs can be translated into truthful posted-price mechanisms for welfare and revenue maximization, and for revenue an extra sample for “lazy reserves” may be required. The threshold-based nature of the direct matroid policies makes them natural candidates for posted-price style mechanisms under matroid feasibility (Caramanis et al., 2021). In the more explicit mechanism-design translation, an $M=(E,\mathcal{I})$23-competitive SSPI that is monotone in values yields order-oblivious posted-price guarantees such as welfare $M=(E,\mathcal{I})$24 and revenue $M=(E,\mathcal{I})$25 for MHR distributions, and welfare and revenue $M=(E,\mathcal{I})$26 for identical regular distributions in the comparison-based setting (Caramanis et al., 2021).

The main structural limitation of current single-sample techniques is their relation to order-oblivious secretary algorithms. A partial converse shows that for any downward-closed feasibility system, including matroids, an $M=(E,\mathcal{I})$27-competitive pointwise single-sample prophet inequality implies a $M=(E,\mathcal{I})$28-competitive order-oblivious secretary policy. Observation 6.1 states that every known SSPI policy, including those in the direct 2021 framework, is a P-SSPI; Theorem 6.2 then places constant-factor single-sample matroid prophet inequalities squarely inside the hardness landscape of the matroid secretary problem (Caramanis et al., 2021).

This yields several concrete open directions. One is whether SSPI and P-SSPI are equivalent up to constant factors; if they are, then a constant-factor SSPI for general matroids would imply a constant-factor OOS algorithm and would resolve a longstanding open problem (Caramanis et al., 2021). A second is whether general matroids admit constant-factor single-sample prophet inequalities by techniques that avoid this converse altogether (Caramanis et al., 2021). A third concerns sample complexity: the general-matroid $M=(E,\mathcal{I})$29 bound now has nearly optimal dependence on the rank, but tight lower bounds for the prophet-inequality problem itself remain open, as does the question of whether more than one sample is inherently necessary in the general matroid setting (Feldman et al., 13 Jul 2025). Finally, uniform matroids still present an internal frontier: the rank-$M=(E,\mathcal{I})$30 deterministic threshold rule is optimal, but the conjectured $M=(E,\mathcal{I})$31 guarantee for the $M=(E,\mathcal{I})$32-th-largest-sample threshold in $M=(E,\mathcal{I})$33 has not been proved for $M=(E,\mathcal{I})$34 (Pashkovich et al., 2023).