CS 700 Computational Mechanism Design Notes on Myerson Optimal Auction

Fall 2008 Prof. David Parkes Harvard University and EPFL October 15, 2008
Consider N = {1, . . . , n} agents with valuations V n [0, h]n and a seller with value v0 . There is a d distribution function Fi (z ) = P r[vi z ] and a density function fi (z ) = dz Fi (z ). Seek a direct-revelation mechanism (DRM) with allocation rule x : V n [0, 1]n and payment rule p : V n Rn , where xi (v ) is the probability that the item is allocated to agent i and pi (v ) is the expected payment. An optimal auction solves the following problem: max
x,p v

v0 (1
j

xj (v )) +
j

pj (v ) f (v )dv i, v

s.t.]
j

xj (v ) 1, xi (v ) 0, Ui (x, p, vi ) 0, Ui (x, p, vi )
Vi

i, vi [vi xi (wi , vi ) pi (wi , vi )] fi (vi )dvi , i, wi , vi ,

where Ui (x, p, vi ) =
Vi

(vi xi (v ) pi (v ))fi (vi )dvi

The road-map of Myersons proof is to show that the problem max

x,p

prot

s.t. feasibility IC IR is equivalent to the problem of max

x

expected virtual surplus

s.t. feasibility monotonicity, which can then be solved in term by an ecient auction designed with respect to virtual surplus under the technical condition of monotone non-decreasing hazard rate on the distribution function. Thus, Myerson converts the problem of optimal auction design into one of ecient auction design on a perturbed problem.
These

notes are based on discussions by Nisan and Hartline in Algorithmic Game Theory Cambridge University Press.

Lemma 1 Mechanism M = (x, p) is DSIC (in expectation) if and only if xi (v ) is monotone non-decreasing vi xi (z, vi )dz . and pi (v ) = vi xi (v ) z=0 Proof 1 (). Fix vi , and drop this from the notation. For DSIC, we need vx(v ) p(v ) vx(v ) p(v ) for every v . Substituting for p(v ), this requires
v v

vx(v ) vx(v ) +
0 v

x(z )dz vx(v ) v x(v ) +

0 v

x(z )dz

x(z )dz
0 v

x(z )dz (v v )x(v )

v

If v > v then we need (v v )x(v ) v x(z )dz . If v > v then we need (v v )x(v ) v x(z )dz . In both cases, the condition holds if x(v ) is monotone non-decreasing. (). For monotonicity, we require vx(v ) p(v ) vx(v ) p(v ) and v x(v ) p(v ) v x(v ) p(v ), and adding that (v v )x(v ) (v v )x(v ). If v v then we need x(v ) x(v ). For the payments, x vi and dene u(b) = vx(b) p(b). For truthfulness, we need u(b) b = 0,
b=v

so that whatever the value of agent i (and whatever vi ) the agent maximizes its utility by bidding b = v . Expanding, we require: v x(b) b
b=v

p(b) b

=0
b=v

To emphasize that this must hold for all vi , substitute z = vi , for 0 = zxi (z ) pi (z ), or pi (z ) = zxi (z ). Integrating both sides from 0 to some constant bi , we have
bi bi

pi (z )dz =
0 0

zxi (z )dz
b bi

pi (bi ) pi (0) = zxi (z )|0i = bi xi (b)

0

xi (z )dz
0 bi

xi (z )dz

Finally, by voluntary participation pi (v ) 0 and no positive transfers pi (v ) 0, and we have the pricing rule. A nice pictoral proof of suciency is available by drawing a plot of x(v ) vs. v (xing vi ) and noting that the payment is the area above the curve, prot the area below the curve. Remark. In the special case of a deterministic allocation rule, the payment is precisely that of the critical value (i.e., the minimal value at which the agent is allocated.) Denition 1 The virtual valuation of agent i with valuation vi is i (vi ) = vi Denition 2 The virtual surplus of allocation x(v ) is i (vi )xi (v ) + v0 (1
i j

1 Fi (vi ) fi (vi )

xj (v ))

Theorem 1 The expected prot of any truthful mechainsm is equal to its expected virtual surplus. Proof 2 Since the mechainsm is truthful, can adopt the lemma and have xi (v ) monotone and pi (v ) = vi vi xi (v ) z=0 xi (z, vi )dz . Now x vi and consider bidder i. Neglecting i from the notation, we now have:
h

Ev [p(v )] =
v =0 h

p(v )f (v )dv
h v

=
v =0 h

vx(v )f (v )dv
v =0 h z =0

x(z )f (v )dzdv
h

=
v =0 h

vx(v )f (v )dv
z =0 h

x( z )
v =z

f (v )dvdz

=
v =0 h

vx(v )f (v )dv
z =0

x(z )(1 F (z ))dz x(v )f (v )dv

=
v =0

1 F (v ) f (v )

= Ev [(v )x(v )], and thus the agents expected payment is exactly its expected virtual valuation. This completes the proof, since virtual surplus for an allocation is the sum of the virtual valuations of each agent and the valuation of the seller when the item is not sold. Thus, to maximize expected prot the task is to choose an allocation that maximizes expected virtual surplus across all monotonic and feasible allocation rules: maxEv
x i

i (vi )xi (v ) + v0 (1
j

xj (v ))

s.t

feasibility monotonicity

Consider the allocation rule x that maximizes virtual surplus; i.e., x(v ) sets xi (v ) = 1 for the bidder i that solves arg maxi i (vi ) when maxi i (vi ) v0 (breaking ties at random), or xi (v ) = 0 for all i otherwise. This is the Lemma 2 Virtual surplus maximization is monotone when the virtual valuations i (vi ) are monotone nondecreasing in vi for all i. A sucient condition for i (vi ) to be monotone non-decreasing is the monotone hazard rate condition. Denition 3 The hazard rate for distribution function Fi (v ) is f (v ) 1 F (v ) If
f (v ) 1F (v )

is monotonie non-decreasing then v

1F (v ) f (v )

is monotone non-decreasing.

Theorem 2 Allocating the item to maximize virtual surplus is optimal and truthful when the distribution has a monotone non-decreasing hazard rate. Collect payment equal to the critcal value, i.e. the smallest amount the winner could have bid and still been allocated.

Remark. In the symmetric case with Fi (v ) = F (v ) for all i, this is equivalent to Vickrey auction with reserve price 1 (v0 ), i.e. with a reserve price r that solves r 1 F (r ) = v0 f (r )

Remark. To gain some intuition for why this reserve price maximizes revenuve, consider selling an item to a single buyer. The seller xes a price r, and the buyer either accepts (and pays) the price or walks away. The optimal price solves max r(1 F (r)),
r

and by the rst-order conditions this is solved by r satisfying 1 (r f (r ) + F (r )) = 0 r 1 F (r ) =0 f (r )

Remark. Myerson also discusses the idea of input ironing in the case in which the allocation rule is non-monotonic. The idea is to construct the nearest monotonic transformation of the virtual valuations (in a well-dened sense) and then use those in place of virtual valuations. vi /100 Example. Consider vi U (0, 100). In this case, i (vi ) = vi 1 = vi (100 vi ) = 2vi 100. 1/100 1 Thus, the optimal auction is a Vickrey auction with reserve price (0) = 50. [The exponential distribution f (v F (v ) = 1 ev is another one with monotone virtual valuations, here the hazard rate is constant, 1 F (v ) = = .] Example. A two-step uniform distribution with a higher probability density in the initial section has a decreasing hazard rate because f (v ) falls from step 1 to step 2.
ev ev