Online Resource Allocation via Static Bundle Pricing

Online Resource Allocation addresses the problem of efficiently allocating limited resources to buyers with incomplete knowledge of future requests. In our setting, buyers arrive sequentially requesting a set of items, each with a value drawn from a known distribution. We study the efficiency of static and anonymous bundle pricing in environments where the buyers’ valuations exhibit strong complementarities. In such settings, standard item pricing fails to leverage item multiplicities, while static bundle pricing mechanisms are only known for very restricted domains and their analysis relies on domain-specific arguments. We develop a unified bundle pricing framework for online resource allocation in three well-studied domains with complementarities: (i) single-minded combinatorial auctions with maximum bundle size $d$; (ii) general single-minded combinatorial auctions; and (iii) network routing, where each buyer aims to route a unit of flow from a source node $s$ to a target node $t$ in a capacitated network. Our approach yields static and anonymous bundle pricing mechanisms whose performance improves exponentially with item multiplicity. For the $d$-single-minded setting with minimum item multiplicity $B$, we obtain an $O(d^{1/B})$-competitive mechanism. For general single-minded combinatorial auctions and online network routing, we obtain $O(m^{1/(B+1)})$-competitive mechanisms, where $m$ is the number of items. We complement these results with information-theoretic lower bounds. We show that no online algorithm can achieve a competitive ratio better than $ \widetildeΩ(m^{1/(B+2)})$ for single-minded combinatorial auctions and $ \widetildeΩ(d^{1/(B+1)})$ for the $d$-single-minded setting. Our constructions exploit a deep connection to the extremal combinatorics problem of determining the maximum number of qualitatively independent partitions of a ground set.

💡 Research Summary

The paper tackles the classic online resource allocation problem in a Bayesian setting, where a seller must allocate limited copies of items to a sequence of buyers arriving online. Each buyer requests a specific bundle of items and draws a value for that bundle from a known distribution. The central challenge is the presence of strong complementarities: a buyer derives value only if the entire requested bundle is allocated. Traditional item‑pricing mechanisms fail dramatically in such environments, achieving only Ω(m) competitive ratios, where m is the number of distinct items.

To overcome this limitation, the authors develop a unified framework for static, anonymous bundle‑pricing mechanisms that works across three well‑studied domains: (i) d‑single‑mindful combinatorial auctions (each requested bundle has size at most d), (ii) general single‑mindful combinatorial auctions (no bound on bundle size), and (iii) online network routing where each buyer wishes to route a unit of flow from a source s to a target t in a capacitated graph. The mechanisms are “static” because the entire menu of bundle‑price pairs is fixed before any buyer arrives, and “anonymous” because the menu does not depend on the identity of the buyer.

The design starts from the ex‑ante linear program (EA‑LP), a relaxation that upper‑bounds the expected social welfare of an optimal offline “prophet”. The authors scale down all item capacities by a factor γ, creating slack to absorb stochastic demand fluctuations. The optimal fractional solution of the γ‑scaled EA‑LP has a remarkably simple structure: for each buyer‑value pair, allocation probability is zero below a cutoff, full (tight) above it, and at most one intermediate “non‑tight” value. By aggregating over all buyers who request the same bundle S, the authors define an “important value” w_S for each bundle.

The static menu then offers:

• Unlimited copies of S at price w_S + 1 (the smallest price strictly larger than w_S).
• ⌊x_{S,w_S}⌋ copies of S at price w_S, where x_{S,w_S} is the total EA‑LP allocation to buyers of bundle S at its important value. If x_{S,w_S}<1, a randomized decision is made to possibly add one extra copy at price w_S.

When a buyer arrives, the lowest price at which their bundle S_b appears in the menu is w_b(S_b). If the realized value v_b ≥ w_b(S_b) and all items in S_b still have residual capacity, the buyer receives the bundle, pays the price, and capacities are decremented.

To analyze the mechanism, the authors introduce a “capacity‑unconstrained” version that ignores item capacities but respects the menu. This virtual process is easy to analyze: its expected welfare is within a constant factor of the optimal value of the γ‑scaled EA‑LP, denoted FRACOPT_γ, which itself is within O(γ) of the prophet’s welfare. The real mechanism’s welfare equals the unconstrained welfare minus the expected loss due to “blocked” buyers (those who find a menu entry but cannot receive the bundle because some item is exhausted). By carefully choosing γ and applying concentration bounds, the authors ensure that for each item the expected number of allocated copies is at most B/γ, and with high probability at most B. Consequently, the loss from blocked buyers is bounded by a constant fraction of FRACOPT_γ, yielding an overall competitive ratio O(γ).

Specific results:

• d‑single‑mindful setting: Choose γ = e^{10d/B}. With this choice, any allocated bundle is blocked with probability ≤0.1, and the loss from blocked buyers is ≤FRACOPT_γ/10. Hence the mechanism achieves a competitive ratio O(d^{1/B}).
• General single‑mindful auctions and online network routing: By setting γ = (m)^{1/(B+1)} (or the analogous expression for the routing graph), the same analysis yields a competitive ratio O(m^{1/(B+1)}), where m is the total number of distinct items (or edges in the network).

The paper also establishes information‑theoretic lower bounds. By embedding the construction of Babaioff et al. (2007) into the single‑mindful setting and exploiting a deep connection to the extremal combinatorial problem of the maximum number of qualitatively independent partitions of a ground set (originally posed by Rényi), the authors prove:

• No online algorithm can achieve a competitive ratio better than Ω̃(d^{1/(B+1)}) for the d‑single‑mindful case.
• No online algorithm can achieve a competitive ratio better than Ω̃(m^{1/(B+2)}) for the general single‑mindful case.

Thus, the proposed static bundle‑pricing mechanisms are essentially optimal up to polylogarithmic factors, and their performance improves exponentially as item multiplicity B grows. The work provides a clean, unified design principle for static pricing in the presence of strong complementarities, bridging gaps between prophet‑inequality theory, mechanism design, and extremal combinatorics, and offering practical pricing schemes for settings such as inventory with multiple copies, cloud resource bundles, and multicommodity network routing.