Lecture 2: Equilibrium

Dominant Strategy Equilibrium, Bayes-Nash

Equilibrium, Characterization, Consequences,
Revelation Principle, BIC, DSIC
Reading: Chapter 2

Today:
• First-price Auction and Second-price Auction
• Bayes Nash equilibrium (BNE)
• BNE Characterization
• Revenue Equivalence
• Solving for BNE
• Revelation Principle
• Bayesian Incentive Compatibility (BIC)
• Dominant Strategy Incentive Compatibility (DSIC)

1
Review of probability
• Random variable, e.g., X ∼ U [0, 1]

• cumulative distribution function,

FX (z) = Pr[X < z], e.g., FX (z) = z.
dFX (z)
• density function fX (z) = dz , e.g., fX (z) = 1.
R∞
• expectation, E[X] = −∞ zfX (z)dz, e.g.,

Z 1 1
E[X] = z · 1dz = .
0 2
expectation is average weighted by distribution
expectation of uniform = midpoint of range

2
Auction Examples
Def: second-price auction
1. bidders bid
2. highest bid wins
3. winner pays second-highest bid
Def: first-price auction
1. bidders bid
2. highest bid wins
3. winner pays bid

Equilibrium
“given a game, what is outcome when players behave selfishly?”
Incomplete information games (i.e., auctions)
“players have private information that specifies their payoff”
Notation
• vectors v = (v1 , . . . , vn )

3
 • hiding coordinates:
v−i = (v1 , . . . , vi−1 , ?, vi+1 , . . . , vn ).

• filling in coordinates: (v−i , z) = (v1 , . . . , vi−1 , z, vi+1 , . . . , vn ).

Definition: A strategy is a function from private info to an
action.
Example: Strategy for second-price auction:
“bid your value”
Definition: An auction has a DSE if there is a profile of
strategies σ 1 , . . . , σ n (σ i maps value vi to bid bi ) such that for
all i and b−i , bidder i’s utility is maximized bidding bi = σ i (vi ).
Note: “truthtelling” is a DSE for second-price auction.

Bayes-Nash Equilibrium
“how do agents play, when no DSE?”
Note: First price auction has no DSE.
Definition: Players with a common prior know the distri-
bution of the private info, v ∼ F .
Definition: A strategy profile is a Bayes-Nash Equilibrium

4
(BNE) if for all i, σ i (vi ) is a best response when other agents
play σ −i (v−i ) with v−i ∼ F −i|vi (conditioned on vi ).
Example:
• First price auction

• Two bidders, values U [0, 1].

Question: What is BNE?
Answer:
z
• Guess σ i (z) = 2 and verify

• If player 2 bids b2 ∼ U [0, 12 ], how should player 1 bid?

• What is 1’s expected utility with bid b1 ?

5
 E[u1 ] = (v1 − b1 ) · Pr[1 wins]
= (v1 − b1 )Pr[b1 > b2 ]
= (v1 − b1 )Pr[b1 > v2 /2]
= (v1 − b1 )Pr[2b1 > v2 ]
= (v1 − b1 )F (2b1 )
= (v1 − b1 )2b1
= 2v1 b1 − 2b21

• To maximize, take derivative and set to zero, solve

v1
• b1 = 2.

Single-dimensional Games
“The agent’s private information is single-dimensional.”
Definition:
• Value: vi = value of agent i for “service.”

6
 • Outcome of game is x and p.

• Game outcome for i:


1

if i is served
xi = 
0 otherwise

pi = payment i makes.

• Utility: ui = vi xi − pi
a.k.a. linear utility
• Agents are risk neutral.
maximize expected utility
Game rules G map b to outcomes and payments.
• xG
i (b) = outcome to i when bids are b.

• pG
i (b) = payment for i when bids are b.

Compose G with σ to map v to outcomes and payments:

“ex post allocation/payment rules”

7
 • xi (v) = xG
i (σ(v)) = outcome when bidder values are v.

• pi (v) = pG
i (σ(v)) = payment for i when values are v.

For values v ∼ F :
“interim allocation/payment rules”
• xi (vi ) = E[xi (v) | vi ] = Ev−i [xi (v−i , vi )].

• pi (vi ) = E[pi (v) | vi ].

• ui (vi ) = vi xi (vi ) − pi (vi ).

Note: In notation xi (vi ), G, σ, F are implicit.
interim allocation/payment rules summarize G, σ, F suffi-
ciently to capture best response.
Prop: G, σ, and F are in BNE iff∗ for all i,

vi ∈ argmaxz vi xi (z) − pi (z)

∗
: for each agent and value, actions not in the range of σ are
dominated by actions in its range. (for ⇐)

8
Characterization of BNE
“Can we tell if an outcome can be a BNE?”
Theorem: G, σ, and product distribution F are in BNE iff∗
1. xi (vi ) is monotone non-decreasing,

2. pi (vi ) = vi xi (vi ) − 0vi xi (z)dz + pi (0),

R

and often pi (0) = 0.

∗
same.
Note: usually pi (0) = 0.
drop subscript i, and consider value v with interim alloca-
tion/payment rule x and p.
Proof: (BNE ⇐ characterization)
by picture
Show that agent prefers σ(v) over σ(z).
Case 1: z < v (opposite case analogous).
• u(v, z) = utility with value v playing σ(z).

• Calculate u(v, v) = vx(v) − p(v).

9
Illustrations:
See MDnA Figure 2.1
• Plot vx(v):
• Plot p(v):
• Subtract to get u(v):
• Calculate u(v, z) = vx(z) − p(z).
• Plot vx(z):
• Plot p(z):
• Subtract to get u(v, z):
• Agent loss is: u(v, v) − u(v, z).

Bidder’s Loss:
• Plot
Q.E.D.

Proof: (Characterization ⇒ BNE)

10
Monotonicity:
• Recall: u(v, z) = vx(z) − p(z)
u(v, z) is agent’s utility for strategy of z.
• BNE =⇒ u(v, v) ≥ u(v, z).
Take v = z1 and z = z2 and vice versa.

z2 x(z2 ) − p(z2 ) ≥ z2 x(z1 ) − p(z1 )

z1 x(z1 ) − p(z1 ) ≥ z1 x(z2 ) − p(z2 )
• Add and cancel:

z2 x(z2 ) + z1 x(z1 ) ≥ z2 x(z1 ) + z1 x(z2 )

• Regroup:

(z2 − z1 )x(z2 ) − (z2 − z1 )x(z1 ) ≥ 0

(z2 − z1 )(x(z2 ) − x(z1 )) ≥ 0

Then:
• If z2 − z1 > 0 =⇒ x(z2 ) ≥ x(z1 )

• =⇒ x(·) is monotone!

11
Payment identity (Proof 1, by picture):
See MDnA Figure 2.2
• Solve for ξ = p(z2 ) − p(z1 )

z2 (x(z2 ) − x(z1 )) ≥ ξ ≥ z1 (x(z2 ) − x(z1 ))

• Draw picture.

• Draw p(·) that satisfies bounds for all z1 < z2

• Plug in z2 = v and z1 = 0 for identity.

Payment identity (Proof 2, by calculus):

skip this example due to time limits?
• Recall: u(v, z) = vx(z) − p(z)

• BNE implies u(v, z) is maximized when z = v.

Derivative is zero at z = v.
• Differentiate with respect to z:

12
 d
u(v, z) = vx′ (z) − p′ (z)
dz

vx′ (v) − p′ (v) = 0

• Holds for all v, thus identity:

p′ (z) = zx′ (z)

• Integrate both sides from 0 to v:
Z v Z v
′
p (z)dz = zx′ (z)dz
0 0

Review “Integration by parts?” if necessay (next topic in

notes)

Z v
p(v) − p(0) = [zx(z)]v0 − x(z)dz
0
• Regroup:
Z v
p(v) = vx(v) − x(z)dz + p(0)
0

Q.E.D.

13
14
Integrate by parts?
Present in parallel with “Payment identity (Proof 2)” if
needed.
Recall product rule:

d
f (z)g(z) = f ′ (z)g(z) + f (z)g ′ (z)


dz
Integrate both sides:
Z Z
f (z)g(z) = f ′ (z)g(z)dz + f (z)g ′ (z)dz

Rearrange:
Z Z
f (z)g ′ (z)dz = f (z)g(z) − f ′ (z)g(z)dz

Use f (z) = z and g ′ (z) = x′i (z).

15
Consequences of BNE Characterization
Recall: Theorem: G, σ, and F are in BNE only if
1. xi (vi ) is monotone non-decreasing,
R vi
2. pi (vi ) = vi xi (vi ) − 0 xi (z) dz + pi (0),
and often pi (0) = 0.

Revenue Equivalence
“auctions with same BNE outcome have same profit”
notice that payment identity means “interim revenue equiva-
lence”
“revenue equivalence” = “payment identity”
Q: what is the outcome of Vickrey auction?
A: bidder with highest value.
Q: who wins in BNE of first-price auction?
A: bidder with highest value.
Result: Vickrey and first-price have same expected revenue.
Example: two agents, uniform values

16
 recall, uniform random variables evenly h divide
i interval in
expectation, i.e., E[valith[1]] = 2/3; E v(2) = 1/3

recall, BNE of FPA is σ(v) = v/2

h i h i
• FPA revenue: E b(1) = E v(1) /2 = (2/3)(1/2) = 1/3
h i
• FPA revenue: E v(2) = 1/3

Method of Revenue Equivalence

Equate two formulas for payments
1. auction rules
2. payment identity
Note: usually (2) is from revenue equivalent DSE auction.

Solving BNE
“given an auction, find the BNE”

17
BNE via Revenue Equivalence
e.g., FPA and SPA are revenue equnvalent so equate (1) from
FPA with (1) from SPA via the equivalence of (2). Solve for
bid strategy from equation.
Given non-truthful mechanism M :
1. guess allocation rule of M ,
2. calculate expected payment for agent i with value vi in
“truthful” revenue equivalent auction.
3. use this expected payment to derive bidding strategies in
M.
Example: first-price auction (two bidders, U [0, 1])
• guess: highest value bidder wins in BNE.
symmetric and monotone strategies
• so: revenue equivalent to second price auction.
• in second price auction, winner pays second highest bid.
• same expected payment in first price auction.
• first-price means: bid = expected payment conditioned
on winning.

18
 note: losers pay nothing in both auctions so can equate
“payment conditioned on winning”

σ 1 (v1 ) = E[p1 (vi )|1 wins]

= E[v2 |v1 > v2 ]
v1
= .
2

Example: all-pay auction

skip this example due to time limits? (see textbook)
• accept sealed bids.
• winner is highest bidder.
• all bidders pay their bids.
Solve for BNE: (two bidders, U [0, 1])
• guess: highest value always wins.
• in all-pay: bid = expected payment in second price.
• in second price auction:
v1
E[p1 (vi )|1 wins] = E[v2 |v1 > v2 ] = 2.

• p1 (v1 ) = E[p1 (v1 )|1 wins]Pr[i wins]+E[p1 (v1 )|1 loses]Pr[i loses].

19
 – Pr[i wins] = xi (vi ) = vi
– E[pi (vi )|1 wins] = vi /2
uniform v2 evenly divides [0, v1 ]
vi v2i
• = 2 · vi + 0 · (1 − vi ) = 2.

• but bid = payment!

v2i
• ⇒ σ i (vi ) = pi (vi ) = 2.

Revelation Principle
“if exists mechanism with good BNE outcome, exist mechanism
with truthtelling BNE and same outcome”
Definition: a mechanism is Bayesian Incentive Compatible
(BIC) if truth-telling is a BNE.
Theorem: can convert mechanism M with good BNE to BIC
mechanism M † with same BNE outcome.
Proof :
• let σ be BNE strategies in M .
• define M † :
1. accept bids, b,

20
 2. simulate σ(b) in M ,
3. output outcome of simulation.
• Claim: σ is BNE in M ⇒ truthtelling is BNE in M † .
Note: applies even to crazy multi-stage mechanisms.

Second-price Auction = (Ascending Auction)† .

Corollary: a mechanism is BIC iff
1. xi (vi ) is monotone non-decreasing,
2. pi (vi ) = vi xi (vi ) − 0vi xi (z)dz + pi (0).
R

Note: strategies σ(v) = v is “onto” (so iff has no assumption)

Note: when looking for good BNE mechanisms, assume BIC.

Dominant Strategy IC
“truth-telling is a dominant strategy”
(same BIC except for xi (vi , v−i ) and pi (vi , v−i ))
Corollary: a deterministic mechanism is DSIC iff
1. xi (vi , v−i ) is a step function at v̂i (v−i )
2.

21
 
v̂

i (v−i ) if vi ≥ v̂i (v−i )
pi (vi , v−i ) =  + pi (0, v−i ).
0 otherwise

Note: prefer DSIC to BIC (more robust), unless DSIC comes

with a loss.

22