Ex Ante Free-Order Prophet Inequalities

• Ex Ante Free-Order Prophet Inequalities are online decision problems where the gambler selects the order based on known distributions before observing outcomes.
• They use ex ante relaxations over matroid polytopes and Bernoulli reductions to develop algorithms with competitive ratios up to 0.7258 and 1-1/e.
• Practical applications include optimal stopping and online contention resolution, with insights into threshold design and free-order strategy improvements.

Searching arXiv for the cited and related prophet-inequality papers to ground the article and retrieve arXiv identifiers. Ex ante free-order prophet inequalities occupy two closely related positions in the prophet-inequality literature. In the classical single-choice setting, “free order” or “order selection” means that the gambler may choose the arrival order of independent distributions before observing any realizations; the objective is a competitive ratio against the prophet’s expected offline maximum (Giambartolomei et al., 2023). In the matroid setting, an “ex-ante prophet inequality” compares an online policy not to the offline optimum directly, but to an ex-ante relaxation defined over the matroid polytope; this viewpoint yields optimal online contention resolution schemes (OCRSs) and random-order contention resolution schemes (RCRSs) for matroids (Lee et al., 2018). The two usages share the same ex-ante informational structure—distributions are known before values are realized—but differ in benchmark, feasibility constraints, and algorithmic consequences.

1. Conceptual scope and model distinctions

The classical prophet-inequality problem starts from independent nonnegative random variables $V_1,\ldots,V_n$ drawn from known distributions $D_1,\ldots,D_n$. A prophet observes all realizations and achieves

$E[\max_i V_i].$

An online gambler observes the values sequentially in some order $\pi$, accepts one value irrevocably or continues, and aims to maximize

$E[V_{\pi_\tau}],$

where $\tau$ is the stopping time. The central objective is the largest constant $c$ such that

$E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$

for every instance; $c$ is the competitive ratio (Giambartolomei et al., 2023).

Within this framework, three order models are distinguished. In adversarial order, the order is fixed by an adversary, and the optimal ratio is $1/2$. In random order, the permutation is drawn uniformly from $D_1,\ldots,D_n$0. In order-selection (OS), also described as “ex-ante free-order,” the gambler chooses $D_1,\ldots,D_n$1 after seeing all distributions but before observing realizations; the data states that clearly OS $D_1,\ldots,D_n$2 random-order $D_1,\ldots,D_n$3 adversarial (Giambartolomei et al., 2023).

A separate but connected line of work concerns matroid prophets. There, the feasible set is an independent set of a matroid $D_1,\ldots,D_n$4, and the online algorithm may accept multiple elements subject to independence. The distinguishing feature is the benchmark: instead of competing directly with the offline optimum, the algorithm is analyzed against an ex-ante relaxation over the matroid polytope $D_1,\ldots,D_n$5 (Lee et al., 2018). This distinction is essential. A common misconception is to identify ex-ante prophet inequalities with free-order prophet inequalities in the OS sense; the literature represented here uses “ex ante” in both senses, but the associated benchmarks and algorithmic structures are different.

Setting	Order/benchmark structure	Best guarantee in the data
Classical adversarial-order	Offline maximum benchmark	$D_1,\ldots,D_n$6
Classical random-order	Offline maximum benchmark	hardness $D_1,\ldots,D_n$7; best algorithms $D_1,\ldots,D_n$8
Classical order-selection	Offline maximum benchmark, order chosen ex ante	$D_1,\ldots,D_n$9
Matroid adversarial ex-ante	Ex-ante LP benchmark $E[\max_i V_i].$0	$E[\max_i V_i].$1
Matroid random-order ex-ante	Ex-ante LP benchmark $E[\max_i V_i].$2	$E[\max_i V_i].$3

2. Ex-ante relaxation for matroid prophet inequalities

For matroid prophets, let $E[\max_i V_i].$4 be a matroid with rank function $E[\max_i V_i].$5, and let $E[\max_i V_i].$6 be the known distribution of $E[\max_i V_i].$7 for each $E[\max_i V_i].$8. The matroid polytope is

$E[\max_i V_i].$9

If $\pi$0 denotes the inverse CDF of $\pi$1, define

$\pi$2

so that

$\pi$3

The ex-ante relaxation is the convex program

$\pi$4

or equivalently $\pi$5. Its optimum is denoted

$\pi$6

This objective is an upper bound on the expected offline optimum (Lee et al., 2018).

This benchmark isolates the ex-ante structure of the instance before any realizations occur. The resulting optimization problem is deterministic and convex, and it compresses the stochastic instance into quantile masses $\pi$7 and conditional top-quantile values $\pi$8. A plausible implication is that this formulation is especially well suited to reductions to rounding problems, because the decision variables already live in the matroid polytope.

3. Adversarial-order ex-ante prophet inequality and the optimal $\pi$9-OCRS

The adversarial-order theorem states: for any matroid $E[V_{\pi_\tau}],$0 and independent distributions $E[V_{\pi_\tau}],$1, there is a deterministic online algorithm in adversarial arrival order whose expected reward is at least

$E[V_{\pi_\tau}],$2

The proof first reduces to the case where each $E[V_{\pi_\tau}],$3 is Bernoulli taking value $E[V_{\pi_\tau}],$4 with probability $E[V_{\pi_\tau}],$5. It then considers a correlated ex-ante value vector obtained by sampling an independent set $E[V_{\pi_\tau}],$6 from an $E[V_{\pi_\tau}],$7-decomposition of $E[V_{\pi_\tau}],$8 and setting $E[V_{\pi_\tau}],$9 if $\tau$0, and $\tau$1 otherwise (Lee et al., 2018).

The algorithm maintains the accepted set $\tau$2. If $\tau$3 denotes the set of the first $\tau$4 arrivals accepted, it defines the residual function

$\tau$5

where $\tau$6 is the value of the maximum $\tau$7-weight independent set in the contraction $\tau$8. With the uniform threshold parameter $\tau$9, the algorithm accepts element $c$0 at state $c$1 if and only if $c$2 and

$c$3

where

$c$4

is the base-price of $c$5 at state $c$6 (Lee et al., 2018).

The analysis decomposes the reward into

$c$7

Two telescoping arguments show

$c$8

and hence $c$9 (Lee et al., 2018).

This ex-ante prophet inequality is then converted into an OCRS by a standard dual-fitting argument. For online rounding of $E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$0, the preprocessing computes an optimal $E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$1 and the corresponding $E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$2, and then uses the same residual base-prices $E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$3. Online, element $E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$4 becomes active with probability $E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$5; if active and $E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$6 and $E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$7, it is accepted, otherwise rejected. The resulting randomized contention resolution scheme selects each $E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$8 with probability at least $E[V_{\pi_\tau}] / E[\max_i V_i] \ge c$9 while always maintaining $c$0, hence it is a $c$1-OCRS (Lee et al., 2018).

4. Random-order ex-ante prophet inequalities and optimal $c$2-RCRSs

In the random-order prophet model for matroids, values are drawn independently at time $c$3, the adversary may see $c$4, but the permutation of arrivals is uniformly random. On observing $c$5 and its time $c$6, the algorithm must decide irrevocably. Under the same ex-ante benchmark $c$7, there is an online policy in random arrival order with expected reward at least

$c$8

This is the random-order ex-ante prophet inequality for matroids (Lee et al., 2018).

As in the adversarial case, the proof reduces to Bernoulli $c$9-values. Let $1/2$0 denote the set of accepted elements arriving before time $1/2$1, and define

$1/2$2

The algorithm uses the time-varying threshold

$1/2$3

When element $1/2$4 arrives at time $1/2$5 with $1/2$6, it is accepted if $1/2$7 and

$1/2$8

The reward is again split into revenue and utility, with

$1/2$9

Because $D_1,\ldots,D_n$00 satisfies $D_1,\ldots,D_n$01, the total reward obeys

$D_1,\ldots,D_n$02

which equals $D_1,\ldots,D_n$03 (Lee et al., 2018).

The induced rounding scheme is an RCRS, and the guarantee is tight. The data states that tightness for rank $D_1,\ldots,D_n$04 shows that no $D_1,\ldots,D_n$05-OCRS in adversarial order can do better than $D_1,\ldots,D_n$06, and no $D_1,\ldots,D_n$07-RCRS in random order can do better than $D_1,\ldots,D_n$08. The exchange lemma,

$D_1,\ldots,D_n$09

holds because the $D_1,\ldots,D_n$10 are the critical values of the usual greedy algorithm in the contracted matroid, and this lemma underpins both the adversarial and random-order analyses (Lee et al., 2018).

5. Classical free-order prophet inequalities: order selection

In the classical single-choice model, order-selection is the setting in which the gambler chooses the arrival order using the distributions alone, then observes the realized values in that order. The data explicitly identifies this as the “Order-Selection (OS, also ‘ex-ante free-order’)” model, and notes that one only needs to consider nonrandomized stopping times $D_1,\ldots,D_n$11 by standard arguments (Giambartolomei et al., 2023).

Historically, Hill (1983) showed that by choosing the order one can break the $D_1,\ldots,D_n$12 barrier of adversarial order. Chawla–Hartline et al. (2010) gave a simple posted-price style single-threshold algorithm attaining $D_1,\ldots,D_n$13. Later refinements increased the ratio to $D_1,\ldots,D_n$14, $D_1,\ldots,D_n$15, and $D_1,\ldots,D_n$16, with the latter associated in the data with blind strategies of Correa–Saona–Ziliotto (2021). Peng–Tang (FOCS 2022) introduced a continuous-time arrival-design with a one-parameter threshold function and proved a $D_1,\ldots,D_n$17-competitive ratio for OS. Bubna–Chiplunkar (EC 2023) refined the threshold-optimization analysis to obtain $D_1,\ldots,D_n$18 (Giambartolomei et al., 2023).

The Peng–Tang scheme embeds the $D_1,\ldots,D_n$19 variables on $D_1,\ldots,D_n$20 by selecting arrival times $D_1,\ldots,D_n$21, one per distribution, so that the conditional distribution of the maximum is spread out in time; equivalently, one chooses an absolutely continuous order-measure on $D_1,\ldots,D_n$22. A threshold function $D_1,\ldots,D_n$23 is then associated to each time $D_1,\ldots,D_n$24, and the gambler accepts the first arriving value $D_1,\ldots,D_n$25 with $D_1,\ldots,D_n$26. The function $D_1,\ldots,D_n$27 is chosen by solving an integral equation or differential equation arising from continuous-time backward induction (Giambartolomei et al., 2023).

The refinement by Bubna–Chiplunkar splits $D_1,\ldots,D_n$28 into more segments and solves a small convex program for the worst-case instance. The data states that the improvement uses analytic lower bounds on the worst-case ratio and explicitly bounds the integrals arising in the backward-induction analysis, thereby avoiding brute-force simulation. The outcome is an explicit family of thresholds, given in closed form up to a few parameters, guaranteeing ratio at least $D_1,\ldots,D_n$29 on every instance (Giambartolomei et al., 2023).

6. Separation from random order, methodological themes, and open questions

A central recent development is the rigorous separation between order selection and random order. For random order, the data reports a new hardness result of $D_1,\ldots,D_n$30: no algorithm can achieve a larger ratio in the random-order model. The proof uses an explicit “three-point-plus-constant” instance consisting of $D_1,\ldots,D_n$31 i.i.d. random variables taking $D_1,\ldots,D_n$32 with probability $D_1,\ldots,D_n$33, $D_1,\ldots,D_n$34 with $D_1,\ldots,D_n$35, $D_1,\ldots,D_n$36 else}) and one extra constant $D_1,\ldots,D_n$37, with parameters $D_1,\ldots,D_n$38, $D_1,\ldots,D_n$39, and $D_1,\ldots,D_n$40. By asymptotically analyzing exact backward-induction thresholds as $D_1,\ldots,D_n$41, the gambler-to-prophet ratio is upper-bounded by $D_1,\ldots,D_n$42 (Giambartolomei et al., 2023).

This establishes that order selection strictly improves over simply taking a uniformly random order, because OS admits a $D_1,\ldots,D_n$43-competitive algorithm while the random-order model has hardness at most $D_1,\ldots,D_n$44. The separation is significant precisely because random order and ex-ante order choice are sometimes conflated. The data makes the distinction explicit: carefully choosing the order, instead of simply taking it at random, benefits the gambler (Giambartolomei et al., 2023).

Across both the matroid and single-choice lines of work, a common methodological pattern is visible. In the matroid setting, ex-ante relaxations over $D_1,\ldots,D_n$45, Bernoulli reductions, residual-value processes, and duality connect prophet inequalities to contention resolution (Lee et al., 2018). In the single-choice free-order setting, continuous-time relaxations, backward induction, and threshold functions $D_1,\ldots,D_n$46 organize both algorithm design and lower-bound analysis (Giambartolomei et al., 2023). This suggests a broader unifying principle: ex-ante structure is exploited either through relaxations over feasible polytopes or through ex-ante control of temporal placement.

The open questions stated in the data remain substantial. For classical OS, the optimal competitive ratio is unknown; the IID special case has tight ratio $D_1,\ldots,D_n$47, leaving a gap of roughly $D_1,\ldots,D_n$48. The data also asks whether full OS reaches $D_1,\ldots,D_n$49, whether there are extremal instances forcing OS ratio below $D_1,\ldots,D_n$50, and whether richer free-order choices can lower or characterize the optimum more sharply. For combinatorial prophet inequalities, the open directions include extending the $D_1,\ldots,D_n$51-OCRS guarantee to the fully online adversary of [FSZ16] and determining whether similar optimal ex-ante free-order prophet inequalities exist for richer constraints such as intersections of matroids (Lee et al., 2018). Together, these questions place ex ante free-order prophet inequalities at the intersection of optimal stopping, stochastic combinatorial optimization, and online rounding.