Transcript g(x`,t`)
Econ 805
Advanced Micro Theory 1
Dan Quint
Fall 2008
Lecture 4 – Sept 11 2008
Today: Necessary and Sufficient Conditions
For Equilibrium
Today: necessary and sufficient conditions for a particular
bidding function to be a symmetric equilibrium
1
Tuesday’s big result was the Envelope
Theorem
Theorem. Suppose that
For all t, x*(t) is nonempty
For all (x,t), gt(x,t) exists
For all x, g(x,-) is absolutely continuous
gt has an integrable bound: supx X | gt(x,t) | B(t) for almost all t,
with B(t) some integrable function
Then for any selection x(s) from x*(s),
V(t) = V(0) + 0t gt(x(s),s) ds
Then we applied this to auctions with symmetric
independent private values
2
Another fairly general necessary condition:
monotonicity
In symmetric IPV auctions, equilibrium bid strategies will
generally be increasing in values; how to prove?
Equilibrium strategies are solutions to the maximization
problem maxx g(x,t)
What conditions on g makes every selection x(t) from x*(t)
nondecreasing?
Recall supermodularity and Topkis
If g(x,t) has increasing differences in (x,t), then the set x*(t) is
increasing in t (in the strong set order)
For g differentiable, this is when 2g / x t 0
But let t’ > t; if x* is not single-valued, this still allows some points
in x*(t) to be above some points in x*(t’), so it wouldn’t rule out
equilibrium strategies which are decreasing at some points
3
Single crossing and single crossing
differences properties (Milgrom/Shannon)
A function h : T R satisfies the strict single crossing
property if for every t’ > t,
h(t) 0 h(t’) > 0
(Also known as, “h crosses 0 only once, from below”)
A function g : X x T R satisfies the strict single crossing
differences property if for every x’ > x, the function
h(t) = g(x’,t) – g(x,t) satisfies strict single crossing
That is, g satisfies strict single crossing differences if
g(x’,t) – g(x,t) 0 g(x’,t’) – g(x,t’) > 0
for every x’ > x, t’ > t
(When gt exists everywhere, a sufficient (but not necessary)
condition is for gt to be strictly increasing in x)
4
What single-crossing differences gives us
Theorem.* Suppose g(x,t) satisfies strict single crossing
differences. Let S X be any subset. Let
x*(t) = arg maxx S g(x,t), and let x(t) be any (pointwise)
selection from x*(t). Then x(t) is nondecreasing in t.
Proof. Let t’ > t, x’ = x(t’) and x = x(t).
By optimality, g(x,t) g(x’,t) and g(x’,t’) g(x,t’)
So g(x,t) – g(x’,t) 0 and g(x,t’) – g(x’,t’) 0
If x > x’, this violates strict single crossing differences
* Milgrom (PATW) theorem 4.1, or a special case of theorem 4’ in Milgrom/Shannon 1994
5
So now, given a symmetric equilibrium with
the bid function b…
Define g(x,t) as the expected payoff, given bid x and value t,
when everyone else uses the equilibrium bid function
If g satisfies strict single crossing differences, then b must be
(weakly) increasing
And (from Tuesday), if g is absolutely continuous and
differentiable in t with an integrable bound, then
V(t) = V(0) + 0t gt(b(s),s) ds
6
All these conditions are almost always
satisfied by symmetric IPV auctions
Suppose b : T R+ is a symmetric equilibrium of some auction
game in our general setup
Assume that the other N-1 bidders bid according to b;
g(x,t)
=
t Pr(win | bid x) – E(pay | bid x)
=
t W(x) – P(x)
So g(x,t) is absolutely continuous and differentiable in t, with
derivative bounded by 1
What about strict single crossing differences?
For x’ > x,
g(x’,t) – g(x,t) =
[ W(x’) – W(x) ] t – [ P(x’) – P(x) ]
When does this satisfy strict single-crossing?
7
When is strict single crossing satisfied by
g(x’,t) – g(x,t) = [ W(x’) – W(x) ] t – [ P(x’) – P(x) ] ?
Assume W(x’) W(x) (probability of winning nondecreasing in bid)
g(x’,t) – g(x,t) is weakly increasing in t, so if it’s strictly positive at t, it’s
strictly positive at t’ > t
Need to check that if g(x’,t) – g(x,t) = 0, then g(x’,t’) – g(x,t’) > 0
This can only fail if W(x’) = W(x)
If b has convex range, W(x’) > W(x), so strict single crossing differences
holds and b must be nondecreasing (e.g.: T convex, b continuous)
If W(x’) = W(x) and P(x’) P(x) (e.g., first-price auction, since P(x) = x),
then g(x’,t) – g(x,t) 0, so there’s nothing to check
But, if W(x’) = W(x) and P(x’) = P(x), then bidding x’ and x give the same
expected payoff, so b(t) = x’ and b(t’) = x could happen in equilibrium
Example. A second-price auction, with values uniformly distributed
over [0,1] [2,3]. The bid function b(2) = 1, b(1) = 2, b(vi) = vi
otherwise is a symmetric equilibrium.
But other than in a few weird situations, g will satisfy strict single
8
crossing differences, so we know b will be nondecreasing
In fact, b will almost always be strictly
increasing
Suppose b(-) were constant over some range of types [t’,t’’]
Then there is positive probability
(N – 1) [ F(t’’) – F(t’) ] FN – 2(t’)
of tying with one other bidder by bidding b* (plus the additional
possibility of tying with multiple bidders)
Suppose you only pay if you win; let B be the expected payment,
conditional on bidding b* and winning
Since t’’ > t’, either t’’ > B or B > t’, so either you strictly prefer to win at
t’’ or you strictly prefer to lose at t’
Assume that when you tie, you win with probability greater than 0 but
less than 1
Then you can strictly gain in expectation either by reducing b(t’) by a
sufficiently small amount, or by raising b(t’’) by a sufficiently small
amount
9
b will almost always be strictly increasing
In first-price auctions, equilibrium bid distributions don’t have point
masses even when type distributions do
When there is a positive probability of each bidder having a particular
value, they play a mixed strategy at that value
Otherwise, there’d be a positive probability of ties, and the same logic
would hold – either prefer to increase by epsilon to win those ties, or
decrease by epsilon if you’re indifferent about winning them
In second-price auctions, if the type distribution has a point mass,
bidders still bid their valuations
Still a dominant strategy
So in that case, there are positive-probability ties
10
So to sum up, in “well-behaved” symmetric
IPV auctions, except in very weird situations,
any symmetric equilibrium bid function will be strictly
increasing,
and the envelope formula will hold
Next: when are these sufficient conditions for a bid
function b to be a symmetric equilibrium?
11
Sufficient Conditions
12
What are generally sufficient conditions for
optimality in this type of problem?
A function g(x,t) satisfies the smooth single crossing
differences condition if for any x’ > x and t’ > t,
g(x’,t) – g(x,t) > 0
g(x’,t’) – g(x,t’) > 0
g(x’,t) – g(x,t) 0
g(x’,t’) – g(x,t’) 0
gx(x,t) = 0 gx(x,t+d) 0 gx(x,t – d)
for all d > 0
Theorem. (PATW th 4.2) Suppose g(x,t) is continuously
differentiable and has the smooth single crossing differences
property. Let x : [0,1] R have range X’, and suppose x is the
sum of a jump function and an absolutely continuous function. If
x is nondecreasing, and
the envelope formula holds: for every t,
g(x(t),t) – g(x(0),0) = 0t gt(x(s),s) ds
then x(t) arg maxx X’ g(x,t)
(Note that x only guaranteed optimal over X’, not over all X)
13
Why do we need smooth single crossing
differences, not just strict s.c.d.?
Consider g(x,t) = (x – t)3, x(t) = x, X = [0,1]
Clearly, x(t) is nondecreasing
And it satisfies the envelope theorem: since
gt(x(t),t) = – (x(t) – t)2 = 0 and g(x(t),t) = 0
But g(x,t) is maximized at x = 1, so x(t) = t is not a
solution
14
Let’s prove the sufficiency theorem
Lots of extra terms due to the possibility of
discontinuities – we’ll just do the case where x(t)
is continuous (and therefore absolutely contin)
x(t) absolutely continuous means it’s
differentiable almost everywhere
So almost everywhere, by the chain rule,
d/dt g(x(t),t) = gx(x(t),t) x’(t) + gt(x(t),t)
But we know the envelope condition holds, so
V’(t) = d/dt g(x(t),t) = gt(x(t),t)
So almost everywhere, gx(x(t),t) x’(t) = 0
15
Let’s prove the sufficiency theorem
Suppose at type t, instead of x(t), I played x(t’), with t’ > t
My gain from the change is
g(x(t’),t) – g(x(t),t) = tt’ gx(x(s),t) x’(s) ds
Now, we know from before that at almost every s,
gx(x(s),s) x’(s) = 0, so either gx(x(s),s) = 0 or x’(s) = 0
If x’(s) = 0, then gx(x(s),t) x’(s) = 0 as well
If gx(x(s),s) = 0, then since x(s) x(t), smooth single crossing
differences says gx(x(s),t) 0 (recall t < s)
And we know x’(s) 0
So for almost every s > t, gx(x(s),t) x’(s) ds 0
So integrating up, the switch from x(t) to x(t’) can’t be good
The symmetric argument rules out t’ < t
16
So there you have it…
If g is differentiable and satisfies smooth single crossing
differences, then x weakly increasing and satisfying the
Envelope condition implies x(t) is optimal AMONG THE
RANGE OF x
We haven’t said anything about other possible x outside the
range of x
Now, when will smooth single crossing differences be satisfied
in auctions?
Well, g(x,t) = t W(x) – P(x), so gx = t W’(x) – P’(x) is weakly
increasing in t as long as W is nondecreasing in x
So as long as higher bids win more often, the only condition we
have to worry about is g being differentiable wrt x
17
Applying this to auctions,
Let b be a bid function, g the implied expected payoff function
If
g is differentiable with respect to x
b is weakly increasing
b and g satisfy the envelope condition
then b(t) is a best-response among the range of b
Still need to check separately for deviations outside the
range of b
If the range of b is convex, that is, T is convex and b is
continuous, only really have to worry about highest type deviating
to higher bids, lowest type deviating to lower bids
18
Pulling it all together,
Theorem (Constraint Simplification). Let g(x,t) be a
parameterized optimization problem. Suppose that
g is differentiable in both its arguments
gt has an integrable bound
g satisfies strict and smooth single crossing differences
Let x : [0,1] X be the sum of an absolutely continuous
function and a jump function, and let X* be the range of x
Then x(t) arg max x X* g(x,t) for every t if and only if
x is nondecreasing
the envelope condition holds
19
And in well-behaved symmetric IPV
auctions,
b : T R+ is a symmetric equilibrium if and only if
b is increasing, and
b (and the g derived from it) satisfy the envelope formula
20
Up next…
Recasting auctions as direct revelation mechanisms
Optimal (revenue-maximizing) auctions
Might want to take a look at the Myerson paper, or
the treatment in one of the textbooks
If you don’t know mechanism design, don’t worry, we’ll go
over it
21
(I probably won’t get to)
First-Order Stochastic Dominance
22
When is one probability distribution “better”
than another?
Two probability distributions, F and G
F first-order stochastically dominates G if
- u(s) dF(s) - u(s) dG(s)
for every nondecreasing function u
So anyone who’s maximizing any increasing function
prefers the distribution of outcomes F to G
(Very strong condition.)
Theorem. F first-order stochastically dominates G if
and only if F(x) G(x) for every x.
23
Proving FOSD “F(x) G(x) everywhere”
Proof for differentiable u. Rewrite it using a basis consisting of
step functions
dq(s) = 0 if s < q, 1 if s q
Up to an additive constant,
u(s) = - u’(q) dq(s) dq
To see this, calculate
u(s’) – u(s) = - u’(q) (dq(s’) – dq(s)) dq = ss’ u’(q) dq
So F FOSD G if and only if - dq(s) dF(s) - dq(s) dG(s) for
every q
24
Proving FOSD “F(x) G(x) everywhere”
But
- dq(s) dF(s) = Pr(s q) = 1 – F(q)
and similarly
- dq(s) dG(s) = 1 – G(q)
So if F(x) G(x) for all x,
Es~F u(s) Es~G u(s)
for any increasing u
“Only if” is because dq(x) is a valid increasing function of x
25