We study the efficiency of Bayes-Nash equilibria in first-price procurement auctions. We measure efficiency by the Price of Anarchy (PoA): the ratio between the expected social cost in equilibrium and the minimum expected social cost resulting from a socially optimal allocation. We first show that the standard first-price procurement auction has no bounded PoA guarantee. We construct equilibria in which every bidder is indifferent between all bids in a common interval, so that bids can be arbitrarily large, and the PoA is unbounded. If costs lie in an interval $[\underline c,\bar c]$ but bids are not restricted, the PoA equals the trivial ratio $\bar c/\underline c$. We hence study ways to restore a finite guarantee, first by characterizing the PoA when bids lie in $[\underline c,\bar c]$, which leads to the tight bound $1 + \frac1{e^2}(\frac{\bar c}{\underline c}-1)$. Alternatively, we introduce the notion of a phantom bidder: the auctioneer accepts the lowest bid only if it does not exceed a phantom bid, drawn independently of the costs, and otherwise buys from an outside option at cost $\bar c$. We show that a simple fixed phantom bid, i.e., a deterministic cap, gives PoA $=\frac1e\sqrt {\frac{\bar c}{\underline c}}+O(1)$. We further show that no phantom mechanism whose construction depends only on the cost interval, under a mild regularity condition, whatever distribution it uses and whatever it reveals about the phantom bid before bidding, can improve on the $\sqrt{\frac{\bar c}{\underline c}}$ dependence.

翻译：暂无翻译