Goal: Maximize the profit.
Allocation Algorithm: Give the object to the player with the highest bid
.
Payment Scheme: The winner pays an amount equal to the second highest bid
.
Sometimes it makes sense to assume that some partial information is known. (部分数据已知)
Assume that the valuations come from a known probability distribution. The distribution is common knowledge! (估值来自一个概率分布, 这个分布是一个 knowledge)
Assume that you have a single player with v1∼U[0,1].
What sort of mechanism would we devise to maximize the expected revenue?
What is the optimal price p?
The optimal price is p=1/2 and gives us expected revenue 1/4.
Assume that there are two players with v1,v2∼U[0,1].
What is the expected revenue that we get from the second price auction?
` - Truthful auction: we consider b1(v1)=v1, b2(v2)=v2.
Assume that there are two players with v1,v2∼U[0,1].
First price auction is not truthful.
What is a Bayesian Nash equilibrium for the first price auction?
v1/2
, b2(v2) = v2/2
. (CANVAS)Notice that the social welfare is maximized in equilibrium, i.e. we assign the item to the highest true value!
All single-item auctions that allocate (
in equilibrium
) the item to the player withhighest value
and in whichlosers pay 0
, will haveidentical expected revenue
.
Assume that there are two players with v1,v2∼U[0,1].
Can we improve upon the 1/3 expected revenue in a truthful way?
Vickrey Auction with Reserve Price 1/2.
if v1 < 1/2 and v2 < 1/2 we do not allocate the item
if v1 < 1/2 and v2 ≥ 1/2 we allocate the item to bidder 2 for price of 1/2
if v1 ≥ 1/2 and v2 < 1/2 we allocate the item to bidder 1 for price of 1/2
if v1 ≥ 1/2 and v2 ≥ 1/2 we allocate the item to the highest bidder and charge him the second highest bid.
Is this auction truthful? Yes
Add a dummy bidder (auctioneer) with value 1⁄2. Run Vickrey auction.
This auction is an affine maximizer.
What is the expected revenue that we get from this auction?
What is the highest expected revenue that we get from this auction?
Definition
Assume that there are n bidders with vi ∈ [0,h].
Mechanism M(x, p1, … , pn)
Input: b=(b1,…,bn) the vector of declarations
Output: x(b) allocation function
xi(b)=1 if i gets the item
xi(b)=0 if i does not get the item
Output: p1(b),…,pn(b): payment functions
The utility of player i if his true value is vi but he reports bi is
ui(bi,b-i;vi)= vi⋅xi(bi,b-i)-pi(bi,b-i)
Notice that player i’s valuation is a single-parameter.
A mechanism is truthful if for every i, bi,vi,b-i
ui(vi,b-i;vi) ≥ ui(bi,b-i;vi)
vi⋅xi(vi,b-i)-pi(vi,b-i) ≥ vi⋅xi(bi,b-i)-pi(bi,b-i)
Theorem (9.39 AGT book). A mechanism is truthful if and only if, for any agent i and any fixed choice of bids by the other agents b−i,
Example:
Definition. The virtual valuation of agent i with valuation vi is
Definition. Given valuations, vi , and corresponding virtual valuations, φi(vi), the virtual surplus (social welfare) of allocation x is
Theorem. The expected profit of any truthful mechanism, M, is equal to its expected virtual surplus
Lemma. Consider any truthful mechanism and fix the bids v−i of all bidders except for bidder i. The expected payment of a bidder i satisfies:
proof:
Lemma. Virtual surplus maximization is truthful if and only if, for all i, φi(vi) is monotone nondecreasing in vi.
Definition: MyeF(b)
Given the bids b and F, compute “virtual bids”: b′i = φi(bi).
Run VCG on the virtual bids b′ to get x′ and p′
Output x = x′ and p with pi = φ−1i (p′i) (upon winning).
MyeF(v1,v2): Vickrey Auction with Reserve Price 1/2
Q: Can we do better?