Lecture Note Informational Friction

February 7, 2014

1

Since

Global Games
Agent

i [0, 1] u (a, A, ) = a (1A> c)1

Utility function

BR (A, ) =

1 A> , 0 A 0 1 0<<1 unique unique 2

eqm eqm

ai = 1i ai = 0i ai = 0i or ai = 1i. 0< <1

does not depend on fundamendals

self fullling eqa

The key observation is that equilibria in the most interesting area

(, c)

at all.

Morris Shin (1998), unintended consequence of complete information

Agent

i [0, 1] u (a, A, ) = a (1A> c)

2

Utility function Private signal

1 xi = + i , i N 0 , x
,

Public signal

z=+

1 N 0, z

Denition

A symmetric Bayesian Equilibrium is a strategy function such that

a (x, z )

and a aggregate attack function

A (, z )

a (x, z ) arg max E [u (a, A (, z ) , ) |x, z ] A (, z ) = Ex [a (x, z ) |z ] = a (x, z ) x ( x (x )) dx

Guess & verify

Fix arbitrary

z.

Suppose stategy function is a form of monotone threshold function

, pdf shifs to the left and

a (x, z ) 0, so A

. Hence, there exists a unique

a (x, z ) = 1(,x (z)] (x). Then, (z ) such that A ( (z ) , z ) = (z ).

This justies the form of the strategy function, as will be shown.

1 u (1, A, ) u (0, A, ) = 2 CDF

of x is

1c A> is increasing in others' decision A, c A ( x (x )) and pdf of x is f (x) = x ( x (x )).

showing strategic complementarity.

First, given individual threshold

x (z ),

get global threshold

(z ).

( x (x (z ) )) A ( (z ) , z ) = (z ) ( x (x (z ) (z ))) = (z ) 1 x (z ) = (z ) + 1 ( (z )) x A (, z ) =
Second, given global threshold

(1)

(z ),

get individual threshold

x (z ).3

Since the payo

E [u (a, A (, z ) , ) |x, z ]

= a P|x,z ( (z )) c x z = a (z ) x+ z
the optimal strategy is

c
such that

is strictly decreasing in

from

a (1 c)

to

ac,

a (x, z ) = 1(,x (z)] (x) 0 1 1 (c)

(z )

x z x (z ) + z c = x z = x (z ) + z (z )

(2)

Existence

To show the existence of

( (z ) , x (z ))

satisfying (1) and (2), substitute (1) to (2).

(z )

x 1 (z ) + 1 ( (z )) x z (z ) 1 ( (z )) x

= =

1 z z + 1 (c) z z + 1 (c) . x x G () =
to

(3)

LHS

z 1 () is continuous G () := x existence of ( (z ) , x (z ))is proved.

Recall

and ranging fromlim 0

lim1 G () = ,
z x

so the

Uniqueness
=

1 1 ( x) = f ( f 1 (x)) . Since x f 1 and respectively, 2

G () =

z x

1 (1 ( )) takes values in

z 2, x

at

z x

2 x

2 2z z, ! ( (z ) , x (z ))

Conclusion (Morris Shin Limit)

When

x ( x 0)

for xed

or

idiosyncratic signal, the unique equilibrium

z 0 ( z ) for xed x ,4 i.e., when there (z ) converges to 1 c for any realization z .5

is continuous at

is relatively small

Proof

Note LHS minus RHS of (3) is the form of and

H (1 c, z, z , 0) = 0

(z, z , x )

H ( (z, z , x ) , z, z , x ) = 0. It suces to prove H (1 c, z, 0, x ) = (0, x ) and (z , 0). For the former,

1 () = 1 (c) = 1 c.
For the latter, by implicit function theorem, it suces to show for all

H1 (1 c, z, 0, x ) = H1 (1 c, z, z , 0) = 0

z x + z

z.

Indeed, if

c < 1, (11 (1c)) = 0

for all

z.
1 z, + z . x

3 The

x key technical step is to get the conditional distribution |x, z N x+ x + z

With improper prior,

p (|x, z )

x (x )2 z (z )2 p (x, z |) p () p (x, z |) 2 2 = p (x|z, ) p (z |) = p (x|) p (z |) e e p (x, z ) p (x, z |) d 1 2

4 This

e 2 ((x +z )

2(x x+z z ) )

=e

1 2

1 1 x +z

x x+z z 2 2 + x z

1 2

1 1 x + z

x x+z z + x z

z x is subtle, and not , because there are and that require dierence speed of convergence. z x x 5 Observe equilibrium now depends on fundmendals c and . < 1 c i no regime change occurs.

Grossman Stiglitz (1980), a paradox on information cost

Time structure

t = 0, 1. t=0
are riskless bond with interest rate

Bonds availabe at xed supply Agents

and risky asset

1 d N d, d

for price

p,

with

s.

i [0, 1]. p.
Only the

All observe

fraction of informed observes

z = d+

1 N 0, z

Information sets are

U = {p} , I = {z, p}.

CARA utility function

U (c| ) = E [ec | ].

Initial wealth

w0 . p (z ) and demand functions xI (z, p) , xU (p)

Denition

A rational expectation equilibrium (REE) is a price function

such that

x ( ) = arg max U (c| ) s.t. c = (w0 px) (1 + r) + dx

c,x

xI (z, p (z )) + (1 ) xU (p (z )) = sz.
Demand function in CARA normal specication is known to be

x ( ) =

E[d| ](1+r )p 6 . V [d| ]

Guess & Verify

Price function is linear observable.

p (z ) = 0 + 1 z , 1 = 0.7
+z z d d 1 d +z , d +z

For informed agent, knowing

is redundant because

is

d|p, z = d|z N

8 For uninformed agent, since the functional form is known and .

1 = 0, z

can be backed up by

p.

So the informedness is the same as informed agents. Hence,

xI (z, p)

+z z d d d +z

(1 + r) p
1 d +z
p0 1

+z d d

xU (p)

d +z

(1 + r) p

By solving

1 d +z
and

(= xI (z, p) = xI (z, p) + (1 ) xU (p)) .

xI (z, p) + (1 ) xU (p) = s, 0 p (z ) = =

1 = 0

can be found.

1 + z z s d d (1 + r) (d + z ) s d d z + z (1 + r) (d + z ) (1 + r) (d + z )

0 =

s d d z , 1 = (1 + r) (d + z ) (1 + r) (d + z ) 1 , xU (p) =
z d +z 1 1 (1+r ) 1 d +z

Sensitivity Welfare
6 When 7 Exact

Demand is not sensitive to price. Plugging

= 0.

Since in equilibrium,

xI (z, p) = xU (p) = s,

both type's equilibrium utilities are the same. This cannot

justify costly information-collection activity (Grossman-Stiglitz paradox).

consumption is normal U (c|) E [c|] + V [c| ] = (w0 px) (1 + r) + E [d| ] x + V [d | ] x 2 . 2 2 values of 0 and 1 are known to agents, although derivation follows
z (z d)2 2

8 p (d|z ) p (z |d) p (d) e

)2 d (dd 2

e 2 [(d +z )d

)d] 2(z z +d d

1 2

1 1 d +z

z z +d d d d +z

3.1 Solving paradox by introducing noise

Both cannot see

u,

but know

s. s + u

is not observable.

2 u N 0, u

Denition

A noisy REE is a price function

p (z, u)

and demand functions

xI (z, p) , xU (p)

such that

x ( ) = arg max U (c| ) s.t. c = (w0 p (z, u) x) (1 + r) + dx

c,x

xI (z, p (z, u)) + (1 ) xU (p (z, u)) = s + u z, u

Guess & Verifty

Price function is linear

is known, and

](1+r )p p (z ) = 0 + z (z + u u), z = 0, u = 0. x ( ) = E[d|V . [d| ] u does not contain information on d, demand function is as before +z z d d d +z

For informed agent, since

xI (z, p) =
p0 z

(1 + r) p
1 d +z

Uninformed agent can only deduce Hence,

p (p) :=

= z + u u = d + + u u
+p p d d (p) d +p

from

1 p. p N d, p

1 p

2 2 = z + 2 u u .

(1 + r) p .
gives

xU (p) =
Market clearing condition,

1 d +p

xI (z, p (z, u)) + (1 ) xU (p (z, u)) = s + u,

+ z z (d + z ) (1 + r) p + 1 d d + p p d d (d + p ) (1 + r) p = s + u
Since

p has

two representations,

p =

p0 z

= z + u u,

both of the following equations are valid identity

(1 ) p 0 + z z + (1 ) p 1 (1 + r) (d + z + (1 ) p ) p = d d z z d d + (z + (1 ) p ) z (1 + r) (d + z + (1 ) p ) p + ((1 ) p u ) u =
Hence,

(s + u) s.

(0 , z , u ) is 0
z

a solution of one of the following nonlinear systems

= = = =

u z p u =
z (and therefore

1 (1+r )(d +z +(1)p )(1)p z z 1 (1+r )(d +z +(1)p )(1)p z 1 (1+r )(d +z +(1)p )(1)p z 1 2 +2 2 z u u

(1)p 0 s d d z

or

0 z u z
p

= = = =

s d d (1+r )(d +z +(1)p ) z +(1)p (1+r )(d +z +(1)p ) (1)p u (1+r )(d +z +(1)p ) 1 2 +2 2 z u u

p )

is easy to get, but

is solution to a quadratic polynomial.

Sensitivity
agents

is. Larger

|u | is an index of how sensitive equilibrium price is to supply noise, or how noisy the equilibrium price 2 and z make informed agents want to trade less, while larger allows price to relect informed 9 activity more. Demand is sensive to price xU (p) < 0.
1 2 +2 2 z u u

9 To

see this, recall p =

p

< z

and plug z equation in the latter system into xU (p) =

= = = p

d +p

p 1 d +p z

(1 + r)

. Then,

1 (1 + r) (d + p ) z

(1 + r) (d + z + (1 ) p ) (1 + r) (d + p ) z + (1 ) p p (d + z + (1 ) p ) (z + (1 ) p ) (d + p ) z + (1 ) p p d (z + (1 ) p ) d p d p d < (1 + r) =0 z + (1 ) p z + (1 ) p

(1 + r) (1 + r)

Welfare

If all agents have the same initial wealth,

EecI d + p = <1 EecU d + z

consumptions are linear function of price, so the utility should be

form ?

Angelotos Werning (2006), Morris Shine meet Grossman Stiglitz

Angelotos Werning (2006) connects public

Morris Shin (1998) does not connect private signal to public signal. continum of agent, changing not available.

signal to private signal by introducing price and 2-stage game. Since price reects private signal, and there are a

aects

z ,

and therefore Morris Shine type limit, such as xing

and

x 0,

is

True return of risky asset traded in stage 1, At stage 1, agnt

i,

is drawn from improper uniform.

i [0, 1] p

observes

xi = +

2 N 0, x

and trade risky and no interest riskless asset to

solve CARA normal problem. Euilibrium price equates demand and noisy supply

s= .

N (0, 1)

of the risky asset.

At stage 2, agents observe

p,

which gives information on

i solves, given p and w0 , maxk,a E [V (w0 pk + k ) + u (a, A, ) |xi , p] where V (c) = ec u (a, A, ) = a (1A> c).10
In summary, agent

and

Denition
action

An equilibrium is a price function

p (, ),

inidividual strategies

k (x, p)

and

a (x, p),

and collective

A (, p)

such that

N REE stage 1 :

k (x, p) arg max E [V (w0 pk + k ) |xi , p] = k (x, p (, )) f (x) dx a (x, p) arg max E [u (a, A (, p) , ) |xi , p] . A (, p) = a (x, p) f (x) dx

P BE stage 2 :

Guess & Verify

Suppose stage 1 price is

p (, ) = + p

. Demand function is symmetric among agents:

k (xi , p) =

E[ |xi ,p]p V [ |x,p] . By

solving the signal extraction problem,

|xi ,p N =

p p+x xi 1 p +x , x +p

. Hence,

k (xi , p) =

p p+x xi p +x 1 x + p

x (xi p)

K (, p) = p (, ) = k (x, p) f (x) dx =

x ( p)

x ( p (, ))
and

2 , p = x x

This veries the guess. In stage 2, since applies. Hence, multiplicity occurs i

i are independent, exactly the same argument as Morris Shine (1998)

x >

1 2 3 2 2 2p x < . 2
is because which has

Observations

Recall in Morris Shin (1998),

better private signal with rate of convergence

x 0 gives uniqueness. Here, it gives multiplicity. This x is reected in much better public signal, price

rate of convergencex , and therefore the second stage environment is nearer to pure public signal case, where multiple equilibria occur. This results depend on the good information aggregation, so if the aggregation is not as good as

x ,

the result could be reversed.

that the additive utility implies the rst stage choice does not aect the second stage payo. Hence, optimization does not have to be backward. Indeed, solving the rst stage rst gives a linear p = + p , so makes the second stage easier.
5

10 Observe

Beauty Contest with full information

Agent

i [0, 1] U (ai , a , ) = (1 r) (ai ) r (ai a ). a =

2

Utility function:

[0,1]\{i}

aj dj =

aj dj .11

Best response is

Uai = 2 [ (1 r) (ai ) r (ai a )] = 0 ai ( a, ) = (1 r) + ra .

Since

ai = aj i, j ,

unless

r = 1, ai = a =

is the only equilibrium.

12

Equilibrium with incomplete information

Agent

i [0, 1] U (ai , a , ) = (1 r) (ai ) r (ai a )

i, i 1 N 0, x
. Public signal

Utility function: Private

a =

[0,1]\{i}
.

aj dj =

aj dj .

xi = +

y=+

1 N , y

Denition

An equilibrium is

a (x, y )

and

a (, y )

such that

a (xi , y ) max E (1 r) (a ) r (a a (, y )) |xi , y

a

a (xi , y ) dj =

a (xi , y ) f (x) dx

a (, y ) =
Best response is

ai = (1 r) E [|xi , y ] + rE [ a|xi , y ]. r

controles degree of strategic complementarity.

Guess & Verify

Suppose

ai = x xi + y y .

In equilibrium, Since

a =

(1 r + rx ) E [|xi , y ] + ry y . ai = (1 r + rx )

|xi ,y N

a di [0,1]\{i} i x xi +y y 1 x +y , x +y

=
,

ai di = x + y y 13

and therefore

ai =

x xi + y y (1 r + rx ) x + ry y = xi + x + y x + y (x , y )
such that

(1 r + rx ) y + ry y. x + y =
(1r +rx )y x +y

The guess is veried i there exists a

x =

(1r +rx )x , y x +y

+ ry .

Indeed,

14

x =

(1 r) x y , y = . (1 r) x + y (1 r) x + y r
is big.

Aggregate action is sensitive to noise shock when complementarity parameter

a = x + y y = (x + y ) + y = +

y . (1 r) x + y

Uniqueness

Dene

(0) [] := and Ei [] = E [|xi , y ], E (1) [] := E Ej [] dj (k) [] E E (k1) [] = := E

(k1) [] dj. Ej E

11 Just

want to make sure a is given in optimization.

ai = ai ( a, ) ,which gives intuition on why r > 1 a = ai is a subtle technicality here. xi di = xf (x) dx. there is no complementarity, r = 0, ai is just a bayesian weighted average.

12 Graphical 13 There 14 When

way to solve is to plot

makes equilibrium unstable.

By repeatedly substituting

a ,15
K

ai
In this model,

(1 r)
k=0

(k) + rK (1 r) Ei E (K ) [ rk Ei E a] =
x 16 x +y ,

(k) [] Ei E

can be explicitly calculated. By induction with

(k) [] = E (k) [] = Ei E

1 k y + k , 1 k+1 y + k+1 xi (K ) [ limK rK (1 r) Ei E a] = 0

holds is

Hence, the unique equilibrium within the class where

ai

= =

(1 r)
k=0

(k) = (1 r) rk Ei E
k=0

rk

1 k+1 y + k+1 xi y= (1 r) x y xi + y. (1 r) x + y (1 r) x + y

(1 r)

xi y y + 1r 1 r

(1 r) (1 r) = xi + 1 1 r 1 r

Morris Shin (2002), Social value of public information

Agent

i [0, 1]
2 U (ai , a , ) = (1 r) (ai ) r Li L
. No private signal. .

Utility function: Public signal

Li =

[0,1]\{i}

2 = (ai aj ) dj , L

Lj dj .17

1 y N 0, y

15 The

calculation describes explicitly how higher order beliefs aect decision.

ai = = = = = = (1 r) Ei + rEi a = (1 r) Ei + rEi (1 r) Ei + rEi (1 r) Ej + rEj (1) + r2 Ei E (1) (1 r) Ei + r (1 r) Ei E (1) + r2 Ei E (1) (1 r) Ei + r (1 r) Ei E (1 r) Ej [] + rEj aj dj ak dk (1) + r2 (1 r) Ei E (2) + r2 (1 r) Ei E (2) (1 r) Ei + r (1 r) Ei E
K

aj dj ak dk dj

dj

aj dj

(1 r)
k=0

(k) + rK (1 r) Ei E (K ) a rk Ei E .

16 Recall | xi ,y N

x xi +y y 1 , + x +y x y

. Ei = xi + (1 ) y. For k = 1,
(1) E (1) Ei E = = = Ej [] dj = + (1 ) y (xi + (1 ) y ) + (1 ) y 2 xi + 1 2 y (k) [] dj = Ej E Ej 1 k y + k dj = 1 k y + k ( + (1 ) y )

When k holds,
(k+1) E = = (k+1) Ei E = =
17 L

E (k) = E

k+1 + 1 k+1 y k+1 (xi + (1 ) y ) + 1 k+1 y. 1 k+2 y + k+2 xi

is not intuitive here, but is necessary to get a particular welfare function.

7

Social welfare function

W (a, ) =

1 1r

2 u (ai , a , ) di = (ai ) di. ai = (1 r) E [|y ] + rE [ a|y ].

No gain from closeness of agents. In addition, since there is no private

FOC and therefore BR is the same as before. information,

ai = aj = a = E [|y ] = y .