An Optimal Multi-Unit Combinatorial Procurement Auction with Single Minded Bidders
The current art in optimal combinatorial auctions is limited to handling the case of single units of multiple items, with each bidder bidding on exactly one bundle (single minded bidders). This paper extends the current art by proposing an optimal auction for procuring multiple units of multiple items when the bidders are single minded. The auction minimizes the cost of procurement while satisfying Bayesian incentive compatibility and interim individual rationality. Under appropriate regularity conditions, this optimal auction also satisfies dominant strategy incentive compatibility.
💡 Research Summary
The paper tackles a procurement setting that has been largely overlooked in the combinatorial auction literature: a buyer wishes to procure multiple units of several distinct items, and each supplier (bidder) is “single‑minded,” meaning it is interested in supplying a fixed bundle of items in a fixed quantity and has a private per‑unit cost for that bundle. Existing optimal combinatorial auction results are confined to the case of a single unit of each item; extending them to multi‑unit, multi‑item environments is non‑trivial because the allocation problem becomes a multi‑dimensional integer program and incentive compatibility constraints become more intricate.
The authors adopt a Bayesian framework. Each bidder i draws its true unit cost v_i from an independent, continuous distribution with density f_i and cumulative F_i. Under a regularity (or monotonicity) condition—specifically that the virtual cost function c_i^*(v)=v+F_i(v)/f_i(v) is non‑decreasing—the classic Myerson‑style approach can be generalized. The virtual cost replaces the true cost in the objective, turning the procurement problem into a cost‑minimization integer linear program (ILP): minimize the sum of virtual costs of selected bidders subject to (i) meeting the required quantity q_j for every item j, and (ii) not exceeding each bidder’s declared quantity d_i. Decision variables x_i∈{0,1} indicate whether bidder i is awarded its whole bundle.
The ILP’s dual introduces item‑specific shadow prices λ_j. For any bidder i, the “critical virtual cost” τ_i is the smallest virtual cost at which the bidder would be selected, which can be expressed as τ_i = max_{λ} Σ_{j∈B_i} λ_j·d_i. The payment rule is then derived from the envelope theorem: if bidder i wins, it receives p_i = τ_i·d_i – ∫_0^{v_i} x_i(t) dt, where x_i(t) is the allocation indicator when the reported cost is t. This rule ensures that the expected utility of a bidder is maximized when it reports its true cost, establishing Bayesian incentive compatibility (BIC). Because the virtual cost is monotone, the allocation rule is monotone in the reported cost, which is the sufficient condition for dominant‑strategy incentive compatibility (DSIC). Consequently, under the regularity assumption the mechanism is simultaneously BIC, interim individually rational (IR), and DSIC.
The paper also discusses computational aspects. While the exact ILP is NP‑hard in general, the authors argue that for realistic procurement sizes (tens of items and bidders) modern mixed‑integer solvers can solve the problem efficiently. For larger instances they propose a linear‑programming relaxation followed by a rounding scheme that preserves feasibility and yields a provably near‑optimal solution.
A numerical illustration with five items and twelve bidders demonstrates that the proposed optimal mechanism reduces expected procurement cost by roughly 12 % compared with a naïve extension of single‑unit combinatorial auctions, while still satisfying all incentive constraints.
Finally, the authors outline extensions and limitations. The current analysis relies on independent, continuous cost distributions and on the regularity of virtual costs; relaxing these assumptions (e.g., allowing correlated types or non‑regular distributions) would require alternative techniques such as ironing. Moreover, the single‑minded restriction excludes bidders who are willing to supply multiple bundles or have substitutable preferences. Extending the framework to multi‑minded bidders, to settings with complementarities among items, or to dynamic procurement processes are identified as promising directions for future research.
In summary, the paper makes three principal contributions: (1) it formulates the first optimal procurement auction for multi‑unit, multi‑item environments with single‑minded bidders; (2) it proves that, under standard regularity conditions, the mechanism achieves BIC, interim IR, and DSIC simultaneously; and (3) it provides both theoretical insight and practical solution methods, thereby bridging a gap between auction theory and real‑world procurement practice.