Combinatorial Auctions with Budgets

We consider budget constrained combinatorial auctions where bidder $i$ has a private value $v_i$, a budget $b_i$, and is interested in all the items in $S_i$. The value to agent $i$ of a set of items $R$ is $|R \cap S_i| \cdot v_i$. Such auctions capture adword auctions, where advertisers offer a bid for ads in response to an advertiser-dependent set of adwords, and advertisers have budgets. It is known that even of all items are identical and all budgets are public it is not possible to be truthful and efficient. Our main result is a novel auction that runs in polynomial time, is incentive compatible, and ensures Pareto-optimality for such auctions when the valuations are private and the budgets are public knowledge. This extends the result of Dobzinski et al. (FOCS 2008) for auctions of multiple {\sl identical} items and public budgets to single-valued {\sl combinatorial} auctions with public budgets.

💡 Research Summary

The paper studies a combinatorial auction model that captures many real‑world settings such as keyword advertising, where each bidder $i$ is characterized by three parameters: a private single‑unit value $v_i$, a publicly known budget $b_i$, and a set of items $S_i$ that the bidder is interested in. The bidder’s valuation for any bundle $R$ is linear in the number of desired items received, i.e., $u_i(R)=|R\cap S_i|\cdot v_i$. The central question is whether one can design a mechanism that is simultaneously (i) incentive compatible (truthful) with respect to the private values, (ii) Pareto‑optimal (efficient) given the budget constraints, and (iii) computationally tractable.

Previous work by Dobzinski et al. (FOCS 2008) showed that even when all items are identical and budgets are public, no mechanism can be both truthful and efficient if bidders’ values are private. Their impossibility hinges on the symmetry of items and the fact that each bidder’s valuation is a simple multiple of a single price. The present paper relaxes the “identical items” assumption while retaining a strong structural restriction: each bidder values every item in $S_i$ at the same private unit price $v_i$. This single‑valued combinatorial setting enables a novel auction design that overcomes the earlier impossibility.

Mechanism Overview
The authors propose a price‑raising auction that proceeds in discrete rounds. Initially the price $p$ is set to zero. In each round the mechanism identifies the smallest remaining budget among active bidders, $b_{\min}$, and raises the price by the smallest increment $\Delta$ that would either exhaust $b_{\min}$ or allocate at least one more item. All bidders whose declared value $v_i$ is at least the current price are eligible to receive an item from their still‑unallocated set $S_i$. The mechanism picks any such item (e.g., arbitrarily or via a fixed ordering), assigns it to the bidder, and deducts $p$ from that bidder’s remaining budget. If a bidder’s residual budget falls below the current price, she becomes inactive and receives no further items. The process repeats until either all items are allocated or no active bidder can afford the current price.

Incentive Compatibility (DSIC)
The proof of truthfulness follows the classic “critical price” argument. If a bidder reports a value $v_i’$ higher than her true $v_i$, she may be forced to purchase items at a price exceeding $v_i$, which would reduce her utility because each additional item yields only $v_i$ value but costs $p>v_i$. Conversely, under‑reporting $v_i$ can only cause the bidder to miss allocations that would have been profitable at the true price, again lowering utility. Hence reporting the true $v_i$ is a dominant strategy for each bidder, regardless of others’ reports. The public nature of budgets is crucial: the price increments are computed from known $b_i$, preventing a bidder from manipulating the price trajectory by hiding budget information.

Pareto Optimality
When the auction terminates, one of two conditions holds: (1) every item has been allocated, or (2) the current price $p$ exceeds the remaining budget of every active bidder, so no further allocation is affordable. In either case there is no feasible reallocation that can make some bidder strictly better off without hurting another, establishing Pareto optimality. The mechanism therefore achieves the strongest possible efficiency notion compatible with incentive constraints in this setting.

Computational Complexity
Each round requires identifying the minimum remaining budget and selecting an eligible item, operations that can be performed in $O(n+m)$ time. The number of price increments is bounded by $O(\log U)$ where $U$ is the maximum budget, because each increment either eliminates at least one bidder or allocates at least one item. Consequently the overall running time is $O(nm\log U)$, i.e., polynomial in the input size, making the mechanism practical for large‑scale advertising platforms.

Necessity of Public Budgets
The authors also present a simple counterexample showing that if budgets were private, no mechanism could retain both DSIC and Pareto optimality. A bidder could misreport her budget to influence the price trajectory, thereby obtaining a more favorable allocation, breaking truthfulness. Thus the public‑budget assumption is not merely a technical convenience but a fundamental requirement for the positive result.

Relation to Prior Work
When every $S_i$ equals the whole item set $