Generalized second price auction

Formal model

Suppose that there are n bidders and k < n slots. Each slot has a probability of being clicked of α i . We can assume that top slots have a larger probability of being clicked, so:

We can think of n − k additional virtual slots with click-through-rate zero, so, α i = 0 for i > k . Now, each bidder has an intrinsic value for one slot v i submits a bid b i indicating the maximum he is willing to pay for a slot (which is his bid reported valuation – notice it doesn't need to be the same as his true valuation v i ). We order the bidders by their value, let's say:

and charge each bidder a price p i (this will be 0 if they didn't win a slot). Slots are sold in a pay-per-click model, so a bidder just pays for a slot if the user actually clicks in that slot. We say the utility of bidder i when allocated to slot j is u i = α j ( v i − p i ) . The total social welfare from owning or selling slots is given by: ∑ j α j v π ( j ) where π ( j ) is the bidder allocated to slot j . The total revenue is given by: ∑ i α i p i

GSP mechanism

To specify a mechanism we need to define the allocation rule (who gets which slot) and the prices paid by each bidder. In a generalized second-price auction we order the bidders by their bid and give the top slot to the highest bidder, the second top slot to the second highest bidder and so on. So, bidder i gets slot i . Each bidder pays the bid of the next highest bidder, so: p i = b i + 1 .

Non-truthfulness

There are cases where bidding the true valuation is not a Nash equilibrium. For example, consider two slots with α 1 = 1 and α 2 = 0.4 and three bidders with valuations v 1 = 7 , v 2 = 6 and v 3 = 1 . Bidding 7, 6 and 1 respectively is not a Nash equilibrium, since the first bidder could lower their bid to 5 and get the second slot for the price of 1, increasing their utility.

Equilibria of GSP

Edelman, Ostrovsky and Schwarz, working under complete information, show that GSP (in the model presented above) always has an efficient locally-envy free equilibrium, i.e., an equilibrium maximizing social welfare, which is measured as S W = ∑ i α i v π ( i ) where π ( i ) is the slot in which player i is allocated according to his bid (the permutation π is defined by the bid vector ( b 1 , … , b n ) ). Further, the revenue in any locally-envy free equilibrium is at least as high as in the (truthful) VCG outcome.

Bounds on the welfare at Nash equilibrium are given by Caragiannis et al., proving a Price of Anarchy bound of 1.282 . Dütting et al. and Lucier at al. prove that any Nash equilibrium extracts at least one half of the truthful VCG revenue from all slots but the first. Computational analysis of this game have been performed by Thompson and Leyton-Brown.

GSP and uncertainty

The classical results due to Edelman, Ostrovsky and Schwarz and Varian hold in the full information setting – when there is no uncertainty involved. Recent results as Gomes and Sweeney and Caragiannis et al. and also empirically by Athey and Nekipelov discuss the Bayesian version of the game - where players have beliefs about the other players, but do not necessarily know the other players' valuations.

Gomes and Sweeney prove that an efficient equilibrium might not exist in the partial information setting. Caragiannis et al. consider the welfare loss at Bayes-Nash equilibrium and prove a Price of Anarchy bound of 2.927. Bounds on the revenue in Bayes-Nash equilibrium are given by Lucier et al. and Caragiannis et al.

Budget Constraints

The impact of budget constraints in the sponsored search/position auction model is discussed in Ashlagi et al. and in the more general assignment problem by Aggarwal et al. and Dütting et al.