Expert Advice for Amateurs

Similar documents
Hamilton- J acobi Equation: Weak S olution We continue the study of the Hamilton-Jacobi equation:

Supplementary Material

EXERCISES FOR SECTION 1.5

Online Appendix for "Customer Recognition in. Experience versus Inspection Good Markets"

INCENTIVE COMPATIBILITY AND MECHANISM DESIGN

MODULE 3 FUNCTION OF A RANDOM VARIABLE AND ITS DISTRIBUTION LECTURES PROBABILITY DISTRIBUTION OF A FUNCTION OF A RANDOM VARIABLE

T L. t=1. Proof of Lemma 1. Using the marginal cost accounting in Equation(4) and standard arguments. t )+Π RB. t )+K 1(Q RB

d 1 = c 1 b 2 - b 1 c 2 d 2 = c 1 b 3 - b 1 c 3

Essential Microeconomics : OPTIMAL CONTROL 1. Consider the following class of optimization problems

Inventory Analysis and Management. Multi-Period Stochastic Models: Optimality of (s, S) Policy for K-Convex Objective Functions

Example on p. 157

Hamilton Jacobi equations

Optimality Conditions for Unconstrained Problems

Lecture 4 Notes (Little s Theorem)

Chapter 7: Solving Trig Equations

Section 3.5 Nonhomogeneous Equations; Method of Undetermined Coefficients

On Customized Goods, Standard Goods, and Competition

Robust estimation based on the first- and third-moment restrictions of the power transformation model

The Asymptotic Behavior of Nonoscillatory Solutions of Some Nonlinear Dynamic Equations on Time Scales

A Note on Superlinear Ambrosetti-Prodi Type Problem in a Ball

Chapter 2. First Order Scalar Equations

t is a basis for the solution space to this system, then the matrix having these solutions as columns, t x 1 t, x 2 t,... x n t x 2 t...

Designing Information Devices and Systems I Spring 2019 Lecture Notes Note 17

1 Consumption and Risky Assets

Vehicle Arrival Models : Headway

Morning Time: 1 hour 30 minutes Additional materials (enclosed):

This document was generated at 1:04 PM, 09/10/13 Copyright 2013 Richard T. Woodward. 4. End points and transversality conditions AGEC

Two Coupled Oscillators / Normal Modes

11!Hí MATHEMATICS : ERDŐS AND ULAM PROC. N. A. S. of decomposiion, properly speaking) conradics he possibiliy of defining a counably addiive real-valu

15. Vector Valued Functions

1 Solutions to selected problems

CHAPTER 2 Signals And Spectra

THE WAVE EQUATION. part hand-in for week 9 b. Any dilation v(x, t) = u(λx, λt) of u(x, t) is also a solution (where λ is constant).

CHAPTER 12 DIRECT CURRENT CIRCUITS

SOLUTIONS TO ECE 3084

2. Nonlinear Conservation Law Equations

Y 0.4Y 0.45Y Y to a proper ARMA specification.

Final Spring 2007

Lecture 2 April 04, 2018

Chapter 6. Systems of First Order Linear Differential Equations

8. Basic RL and RC Circuits

Math Week 14 April 16-20: sections first order systems of linear differential equations; 7.4 mass-spring systems.

Solutions from Chapter 9.1 and 9.2

10. State Space Methods

Chapter 7 Response of First-order RL and RC Circuits

Econ107 Applied Econometrics Topic 7: Multicollinearity (Studenmund, Chapter 8)

Two Popular Bayesian Estimators: Particle and Kalman Filters. McGill COMP 765 Sept 14 th, 2017

Guest Lectures for Dr. MacFarlane s EE3350 Part Deux

O Q L N. Discrete-Time Stochastic Dynamic Programming. I. Notation and basic assumptions. ε t : a px1 random vector of disturbances at time t.

WEEK-3 Recitation PHYS 131. of the projectile s velocity remains constant throughout the motion, since the acceleration a x

Linear Response Theory: The connection between QFT and experiments

Online Appendix for Optimal Taxation and Human Capital Policies over the Life Cycle

T. J. HOLMES AND T. J. KEHOE INTERNATIONAL TRADE AND PAYMENTS THEORY FALL 2011 EXAMINATION

t 2 B F x,t n dsdt t u x,t dxdt

Challenge Problems. DIS 203 and 210. March 6, (e 2) k. k(k + 2). k=1. f(x) = k(k + 2) = 1 x k

Finish reading Chapter 2 of Spivak, rereading earlier sections as necessary. handout and fill in some missing details!

Hamilton- J acobi Equation: Explicit Formulas In this lecture we try to apply the method of characteristics to the Hamilton-Jacobi equation: u t

Overview. COMP14112: Artificial Intelligence Fundamentals. Lecture 0 Very Brief Overview. Structure of this course

On Measuring Pro-Poor Growth. 1. On Various Ways of Measuring Pro-Poor Growth: A Short Review of the Literature

3.1.3 INTRODUCTION TO DYNAMIC OPTIMIZATION: DISCRETE TIME PROBLEMS. A. The Hamiltonian and First-Order Conditions in a Finite Time Horizon

Lecture Notes 2. The Hilbert Space Approach to Time Series

Online Appendix to Solution Methods for Models with Rare Disasters

Approximation Algorithms for Unique Games via Orthogonal Separators

Let us start with a two dimensional case. We consider a vector ( x,

V L. DT s D T s t. Figure 1: Buck-boost converter: inductor current i(t) in the continuous conduction mode.

Explaining Total Factor Productivity. Ulrich Kohli University of Geneva December 2015

A Primal-Dual Type Algorithm with the O(1/t) Convergence Rate for Large Scale Constrained Convex Programs

5. Stochastic processes (1)

HOMEWORK # 2: MATH 211, SPRING Note: This is the last solution set where I will describe the MATLAB I used to make my pictures.

Sections 2.2 & 2.3 Limit of a Function and Limit Laws

18 Biological models with discrete time

Predator - Prey Model Trajectories and the nonlinear conservation law

Chapter Floating Point Representation

1 Review of Zero-Sum Games

Math 2142 Exam 1 Review Problems. x 2 + f (0) 3! for the 3rd Taylor polynomial at x = 0. To calculate the various quantities:

The Contradiction within Equations of Motion with Constant Acceleration

Logic in computer science

TO our knowledge, most exciting results on the existence

Reading from Young & Freedman: For this topic, read sections 25.4 & 25.5, the introduction to chapter 26 and sections 26.1 to 26.2 & 26.4.

Lecture 20: Riccati Equations and Least Squares Feedback Control

Physics 235 Chapter 2. Chapter 2 Newtonian Mechanics Single Particle

Problem Set #3: AK models

Math 315: Linear Algebra Solutions to Assignment 6

Math 10B: Mock Mid II. April 13, 2016

E β t log (C t ) + M t M t 1. = Y t + B t 1 P t. B t 0 (3) v t = P tc t M t Question 1. Find the FOC s for an optimum in the agent s problem.

Stochastic models and their distributions

( ) ( ) ( ) ( u) ( u) = are shown in Figure =, it is reasonable to speculate that. = cos u ) and the inside function ( ( t) du

Stability and Bifurcation in a Neural Network Model with Two Delays

ANSWERS TO EVEN NUMBERED EXERCISES IN CHAPTER 6 SECTION 6.1: LIFE CYCLE CONSUMPTION AND WEALTH T 1. . Let ct. ) is a strictly concave function of c

Notes for Lecture 17-18

Some Basic Information about M-S-D Systems

Comparing Means: t-tests for One Sample & Two Related Samples

Math 334 Fall 2011 Homework 11 Solutions

ON THE DEGREES OF RATIONAL KNOTS

MA 214 Calculus IV (Spring 2016) Section 2. Homework Assignment 1 Solutions

Economic Growth & Development: Part 4 Vertical Innovation Models. By Kiminori Matsuyama. Updated on , 11:01:54 AM

6. Stochastic calculus with jump processes

RANDOM LAGRANGE MULTIPLIERS AND TRANSVERSALITY

KEY. Math 334 Midterm I Fall 2008 sections 001 and 003 Instructor: Scott Glasgow

In this chapter the model of free motion under gravity is extended to objects projected at an angle. When you have completed it, you should

Transcription:

Exper Advice for Amaeurs Ernes K. Lai Online Appendix - Exisence of Equilibria The analysis in his secion is performed under more general payoff funcions. Wihou aking an explici form, he payoffs of he exper and he decision maker, U e (a,, b) and U d (a, ), are assumed o saisfy he following condiions: ) U e ( ) and U d ( ) are wice coninuously differeniable, 2) U( ) i <, 3) U(a, i ) = for some a R, 4) U2( ) i >, i = e, d, 5) U d (a, ) = U e (a,, ) for all (a, ), and 6) U3( ) e > everywhere. Also define s argmax a(r, s) = U d (a, )d, if r < s, a r a d (r), if r = s. The emergence of off-equilibrium beliefs in he amaeur model, which canno be eliminaed by having all messages used, poses a challenge in he equilibrium characerizaion. A full characerizaion requires consideraion of all possible off-equilibrium beliefs, which represen a very large se. In he following, I illusrae he issues and provide a specificaion of beliefs ha, coupled wih a mild condiion on he exper s payoff, ensures he exisence of pariional equilibria. I firs show ha, for whaever beliefs afer false advice, he exper never fully reveals his informaion. Inuiively, if he exper reveals, he amaeur s own informaion will become useless, implying ha he amaeur will respond in exacly he same way as does he novice. Since a biased exper does no fully reveal o he novice, he also will no do so o he amaeur. Proposiion 5. There exiss no separaing equilibrium in he amaeur model. Proof. If he exper in he amaeur model fully reveals his informaion, for all Θ, induces a d () on all inerval ypes. Accordingly, a ype- exper s payoff is U e (a d (),, b). Suppose here exiss a fully separaing equilibrium. For any η >, we can find a < η ha induces a d ( ) in he equilibrium. Suppose deviaes by sending m ha in he equilibrium is reserved for = ɛ, where ɛ > is such ha a d ( ) < a d ( ɛ) < a e (, b). The coninuiy of he payoff funcions and ha U d (a, ) U e (a,, ) guaranee ha such an ɛ exiss. Upon receiving m, all inerval ypes wih / (, ɛ) will ake acion a d ( ɛ), which, given he choice of ɛ and ha U( ) e <, is sricly preferred over a d ( ) by. To he remaining ypes wih (, ɛ), all being low-inerval ypes, m is a false advice, and hey ake acion under off-equilibrium beliefs ψ( l ). For any such beliefs, given ha U2( ) d >, he se of induced acions aken by hese inerval ypes is bounded by a d () and a d ( ɛ). Since U( ) e <, we

have U e (a d (),, b) U e (a,, b) for all a [a d (), a d ( ɛ)]. Thus, he payoff for o send m is, regardless of he specificaion of ψ( l ), bounded below by (A.) U e (a d ( ɛ),, b) ɛ U e (a d (),, b) ɛ U e (a d ( ɛ),, b). Subracing s equilibrium payoff from (A.) gives: (A.2) [U e (a d ( ɛ),, b) U e (a d ( ),, b)] ɛ ɛ [U e (a d ( ɛ),, b) U e (a d ( ),, b)]. [U e (a d (),, b) U e (a d ( ),, b)] We can choose η sufficienly small such ha, for any ɛ ha saisfies he above crierion of choosing he deviaing message, he firs and he hird posiive erms in (A.2) dominae he second negaive erms for some < η. The sric incenive for some o deviae poses a conradicion o he exisence of fully separaing equilibrium. In he CS model, when he indifference condiion is saisfied under a pariional sraegy of he exper (p.7-8) where N 2, i follows ha he following incenive-compaibiliy condiions are saisfied: ) every boundary ype i will (weakly) prefer sending messages in M i over any oher messages; and 2) in he inerior of every I i he inerior ypes will (sricly) prefer he same. The exper s sraegy hus consiues an informaive pariional equilibrium. In he amaeur model, even if he indifference condiion holds under such sraegy, in which (A.3) V e (M i, i, b) = V e (M i, i, b), i =,..., N, = and N =, since he off-equilibrium beliefs may generae a benefi of lying under false advice he incenivecompaibiliy condiions is no necessarily saisfied. Before illusraing how benefi of lying may arise, I pause o characerize he induced acions of he amaeur under a pariional sraegy. I disinguish beween wo ypes of induced acions, effecively induced and ineffecively induced. An acion a is effecively induced by m if he amaeur updaes her beliefs µ( m, s ) using Bayes rule and here exiss Θ such ha, in her maximizaion problem of which a is he soluion, µ( m, s ) φ( s ). Lemma 4. A high-inerval ype h akes effecively induced acions if and only if she receives. subsiuing advice: her hreshold I i, i =,..., N, and he exper reveals ha I j, i < j N; or 2. complemenary advice: her hreshold I i, i =,..., N, and he exper reveals ha I i, and she akes ineffecively induced acions if and only if she receives Wih a sligh abuse of noaions, I use V e (M i, i, b) o sand for V e (m, i, b) for all m M i. 2

3. redundan advice: her hreshold I N, and he exper reveals ha I N ; or 4. false advice: her hreshold I i, i = 2,..., N, and exper reveals ha I k, k < i. A low-inerval ype l akes effecively induced acions if and only if she receives 5. subsiuing advice: her hreshold I i, i = 2,..., N, and he exper reveals ha I k, k < i; or 6. complemenary advice: her hreshold I i, i = 2,..., N, and he exper reveals ha I i, and she akes ineffecively induced acions if and only if she receives 7. redundan advice: her hreshold I, and he exper reveals ha I ; 8. false advice: her hreshold I i, i =,..., N, and he exper reveals ha I j, i < j N. Proof. I prove he cases for high-inerval ypes; he cases for low-inerval ypes are similar. Consider firs Condiions and 2. Suppose h wih I i, i =,..., N, receives m indicaing ha I j. For j > i, h s updaed beliefs under Bayes rule are µ( m, h ) = /( j j ) for ( j, j ] and zero elsewhere. For j = i, h s updaed beliefs are µ( m, h ) = /( i ) for [, i ] and zero elsewhere. In boh cases, here exiss [, ] such ha µ( m, h ) φ( h ), and he resuling acions are effecively induced. This proves he sufficiencies. The necessiies is proved by conraposiive. Suppose h wih I i, i =,..., N, receives m indicaing ha I j, j < i. Noe ha hen Θ σ (m) h = ; Bayes rule canno be applied, and he resuling acion canno be effecively induced. Finally, suppose h wih I N receives m indicaing ha I N. Her updaed beliefs are µ( m, h ) = /( ) for [, ] and zero elsewhere. This is equivalen o φ( h ) for all [, ], and he resuling acion canno be effecively induced. Since here are only wo ypes of acions, effecively induced and ineffecively induced, and hey are muually exclusive, he sufficiencies (necessiies) in Condiions 3 and 4 for ineffecively induced acions follow from he above necessiies (sufficiencies) for effecively induced acions. Using Lemma 4, he profile of acions effecively induced on h, I i, i =,..., N, is { a(, i ), for m M i (complemenary advice), ρ(m, h ) = a( j, j ), for m M j, i < j N (subsiuing advice); and ha induced on l, I i, i = 2,..., N is { a( i, ), for m M i (complemenary advice), ρ(m, l ) = a( k, k ), for m M k, k < i (subsiuing advice). The profile of acions ineffecively induced by redundan advice is ρ(m, h ) = a(, ) and ρ(m, l ) = a(, ). The profile of acions induced by false advice depends on he off-equilibrium beliefs. To 3

illusrae he benefi of lying, suppose he indifference condiion holds wih N = 3 in a proposed equilibrium. Consider boundary ype (Figure 3). 2 If is he rue sae, all high-inerval ypes will have and all low-inerval ypes >. Suppose sends m M, indicaing ha [, ]. Then, all high-inerval ypes will ake acion a(, ), [, ], and all lowinerval ypes will ake acion a(, ) (he second line in Figure 3). Given ha saisfies he indifference condiion, he will be indifferen beween inducing hese acions and hose induced by m M 2, which are a(, 2 ) and a(, ), (, 2 ] (no shown in he figure). Deviaion high-inerval ypes hresholds low-inerval ypes hresholds acions induced: a( 2, ) a? a( 2, ) Proposed Equilibrium high-inerval ypes hresholds low-inerval ypes hresholds acions induced: a(, ) a(, ) 2 M M 3 Figure : Incenives for Deviaions Now, suppose lies by sending m M 3, indicaing ha ( 2, ]. All high-inerval ypes will ake a( 2, ) and all low-inerval ypes wih ( 2, ] will ake a( 2, ) (he firs line in Figure 3). These inerval ypes canno deec he lie and are effecively induced o ake hese acions. And o hese acions are less favorable han hose induced by M (or M 2 ). To he low-inerval ypes wih (, 2 ], he advice can be deeced as false. Wihou any resricion, one can come up wih off-equilibrium beliefs so ha he ineffecively induced acions aken by hese low-inerval ypes will be closer o s ideal han is a(, ), he effecively induced acion aken by hese ypes in he proposed equilibrium. This creaes a benefi of lying, which is absen in he CS model. I is conceivable, especially in equilibria wih more seps, ha such benefi of lying may ouweigh he cos of inducing less favorable acions. Wha equilibria may emerge in siuaions of his sor require ad hoc and deailed specificaions of beliefs. The following proposiion saes, however, ha here is a se of off-equilibrium beliefs ha, ogeher wih a mild resricion on he exper s payoff, ensures he sufficiency of he indifference condiion for he exisence of pariional equilibria. Denoe ψ o be he se of off-equilibrium beliefs of all inerval ypes: ψ = {,s} T {l,h} ψ( s). Proposiion 6. There exiss a se of off-equilibrium beliefs ψ such ha, provided U2( ) e is sufficienly large, he boundary ypes { i } N i= ha saisfy (A.3) always consiue an equilibrium. 2 There could be a benefi of lying for boundary ypes whenever N 3. However, for he inerior ypes, such benefi also arises for N = 2. Thus, he indifference condiion is no always sufficien even for wo-sep equilibria. Indeed, in he CS model, incenive compaibiliy for he inerior ypes is a consequence of ha for he boundary ypes. As will be discussed below, his is also no rue in he amaeur model. 4

Proof. I firs consruc ψ and sae he cases of he exper s payoff V e (m,, b) under ψ. I hen show ha, if ψ = ψ and U e 2( ) is sufficienly large, hen (A.3) is sufficien for he following o always hold: for all [ i, i ], (A.4) V e (M i,, b) = max V e (M j,, b), i, j =,..., N. j The se of off-equilibrium beliefs ψ is consruced as follows. Suppose here exiss a monoone soluion, {,..., N } (, ), o (A.3). If a high-inerval ype h wih ( i, i ], i =,..., N, receives a false advice, her beliefs are ha is disribued on [, i ] wih densiy /( i ) and zero elsewhere; if a low-inerval ype l wih ( i, i ] receives a false advice, her beliefs are ha is disribued on ( i, ) wih densiy /( i ) and zero elsewhere. Then, when [ i, i ] sends m M i under he pariional sraegy and deviaes from i by sending m M g, g i, he profile of his expeced payoff will be (A.5) V e (m,, b) = k i U e (a( k, k ),, b) i r=k i r r U e (a(, r ),, b) U e (a(, i ),, b) U e (a( k, k ),, b), U e (a( i, i ),, b) i U e (a(, i ),, b) i i U e (a( i, ),, b) U e (a( i, i ),, b), U e (a( j, j ),, b) i U e (a( i, ),, b) j r r=i j r U e (a( r, ),, b) U e (a( j, j ),, b), if m M k, k < i, if m M i, if m M j, i < j N. Using he second and he hird cases in (A.5), he expeced payoff for i o send m M i and 5

m M i are, respecively, (A.6) (A.7) i i U e (a( i, i ), i, b) U e (a( i, i ), i, b) i i i U e (a(, i ), i, b) i U e (a( i, i ), i, b) U e (a( i, ), i, b) U e (a( i, i ), i, b). i i Thus, he indifference condiion (A.3) becomes he following second-order difference equaion: (A.8) V ( i, i, i, b) = i [U e (a( i, i ), i, b) U e (a( i, i ), i, b)] i i [U e (a( i, i ), i, b) U e (a(, i ), i, b)] i i [U e (a( i, ), i, b) U e (a( i, i ), i, b)] i [U e (a( i, i ), i, b) U e (a( i, i ), i, b)] =, i =,..., N, =, N =. Suppose here is a sricly increasing pariion,,..., i, ha saisfies (A.8). Tha U e ( ) <, a(, ) is sricly increasing in is argumens, and he coninuiy of V ( i, i,, b) in ensure ha here exiss a unique i > i ha saisfies (A.8). Turning o incenive compaibiliy, I begin by showing ha (A.4) holds for i, i =,..., N, ha saisfy (A.3). If N = 2, here exiss no oher se of messages ha i can send, and (A.4) is saisfied vacuously. So, consider N 3. Suppose i sends message m M in, 2 n N i. Then, from he hird case in (A.5) his expeced payoff is (A.9) i U e (a( in, in ), i, b) in in 2 r=i in U e (a( in, ), i, b) r r U e (a( r, ), i, b) in U e (a( in, in ), i, b). Subracing (A.9) from (A.7), we have D 3 = i [U e (a( i, i ), i, b) U e (a( in, in ), i, b)] i i in 2 r r=i r in [U e (a( i, ), i, b) U e (a( i, ), i, b)] [U e (a( i, i ), i, b) U e (a( r, ), i, b)] in [U e (a( i, i ), i, b) U e (a( in, ), i, b)] [U e (a( i, i ), i, b) U e (a( in, in ), i, b)]. in 6

Noe ha (A.8) implies ha he exper s ideal acion a e ( i, b) (a( i, i ), a( i, i )). Since a( in, in ) > a( i, i ), a( in, ) > a( i, i ) for ( in, in ), and a( j, ) > a( i, i ) for ( j, j ), j = i,..., i n 2, given U e ( ) < and he maximum of U e (a, i, b) is achieved for a (a( i, i ), a( i, i )), he firs, hird, fourh and fifh erms are posiive. Also, he second erm vanishes. Thus, D 3 >. Nex, suppose i sends m M i η, η i. From he firs case in (A.5), his expeced payoff is (A.) i η U e (a( i η, i η ), i, b) i r=i η r i η r U e (a(, r ), i, b) i η U e (a(, i η ), i, b) i U e (a( i η, i η ), i, b). Subracing (A.) from (A.6), we have D 4 = i η [U e (a( i, i ), i, b) U e (a( i η, i η ), i, b)] i η i η [U e (a( i, i ), i, b) U e (a(, i η ), i, b)] i 2 r=i η i r r [U e (a( i, i ), i, b) U e (a(, r ), i, b)] i [U e (a(, i ), i, b) U e (a(, i ), i, b)] i [U e (a( i, i ), i, b) U e (a( i η, i η ), i, b)]. Similar o he above, since a( i, i ) > a( i η, i η ), a( i, i ) > a(, i η ) for ( i η, i η ), and a( i, i ) > a(, j ), for ( j, j ), j = i η,..., i 2, he firs, second, hird and fifh erms are posiive, while he fourh erm vanishes. Thus, D 4 >. Tha D 3 > and D 4 > imply ha (A.4) holds for i, i =,..., N. I show nex ha given (A.8) and for sufficienly large U e 2( ), all ( i, i ) prefer sending m M i over m M i, and all ( i, i ) prefer sending m M i over m M i, i =,..., N, so ha (A.4) holds for all inerior. Consider an arbirary ( i, i ). From he hird case in (A.5), his expeced payoff from sending m M i is (A.) U e (a( i, i ), i, b) i i U e (a( i, ), i, b) U e (a( i, ), i, b) U e (a( i, i ), i, b). i i 7

Subracing his expeced payoff from sending m M i in (A.5) from (A.), we have (A.2) D 5 = i [U e (a( i, i ),, b) U e (a( i, i ),, b)] i [U e (a( i, i ),, b) U e (a(, i ),, b)] i i i [U e (a( i, ),, b) U e (a( i, ),, b)] [U e (a( i, ),, b) U e (a( i, i ),, b)] i [U e (a( i, i ),, b) U e (a( i, i ),, b)]. Differeniaing D 5 wih respec o gives D 5 = i [U e (a( i, i ),, b) U e (a( i, i ),, b)] [U e (a( i, i),, b) U e (a(, i ),, b)] i i i [U e (a( i, ),, b) U e (a( i, i ),, b)] [U e (a( i, i),, b) U e (a( i, i ),, b)] i [U e (a( i, i ),, b) U e (a(, i ),, b)]. Since a( i, i ) > a( i, i ), a( i, i ) > a(, i ) for ( i, ), and a( i, ) > a( i, i ) for ( i, i ), U e 2( ) > implies ha he firs four erms are posiive and he las erm is negaive; when decreases from i, here are negaive effecs on D 5 from he firs four erm and a posiive effec from he las erm. However, for a sufficienly large U e 2( ) a, he negaive effecs ouweigh he posiive. A sufficienly large U e 2( ) hen ensures, given (A.8), D 5 for. Consider nex an arbirary ( i, i ). From he firs case in (A.5), his expeced payoff from sending m M i is (A.3) i U e (a( i, i ),, b) i i U e (a(, i ),, b) i U e (a(, i ),, b) U e (a( i, i ),, b). 8

Subracing (A.3) from he expeced payoff from sending m M i in (A.5), we have (A.4) D 6 = i [U e (a( i, i ),, b) U e (a( i, i ),, b)] i i [U e (a( i, i ),, b) U e (a(, i ),, b)] i [U e (a(, i ),, b) U e (a(, i ),, b)] i [U e (a( i, ),, b) U e (a( i, i ),, b)] i [U e (a( i, i ),, b) U e (a( i, i ),, b)]. Differeniaing D 6 wih respec o gives D 6 = i [U e (a( i, i ),, b) U e (a( i, i ),, b)] i [U e (a( i, i),, b) U e (a(, i ),, b)] i i [U e (a( i, ),, b) U e (a( i, i ),, b)] [U e (a( i, i),, b) U e (a( i, i ),, b)] i [U e (a( i, ),, b) U e (a( i, i ),, b)]. Since a( i, i ) > a( i, i ), a( i, i ) > a(, i ) for ( i, ), and a( i, ) > a( i, i ) for ( i, i ), U e 2( ) > implies ha he firs four erms are posiive. While he las erm is negaive, similar o he above, for a sufficienly large U e 2( ) a, he posiive effecs on D 6 from he firs four erms ouweigh he negaive effec from he las erm; (A.8) hen implies D 6 a. When U e 2( ) is sufficienly large for all inerior ypes, (A.4) holds for all of hem. The off-equilibrium beliefs specified in he proof ha for ( i, i ], h and l believe ha is uniformly disribued on, respecively, [, i ] and ( i, ) are sufficien for he incenive compaibiliy condiion o hold for he boundary ypes. However, since he inerior ypes induce a se of acions no induced by he boundary ypes, a sufficienly large U e 2( ) is called ino he picure o ensure ha incenive compaibiliy also holds for hem. Figure 4 illusraes he raionale wih an example of wo-sep equilibrium. Given ha he indifference condiion holds, he boundary ype s expeced payoff from he profile of acions a(, ) and a(, ), [, ], is he same from ha from a(, ) and a(, ), (, ] (he wo upper lines). Consider he acions induced when he inerior ype sends messages in M and M 2. If we compare he profile of acions in he lower pair of lines wih hose in he upper pair, we can see ha hey are he same excep for (, ]. While gives no 9

boundary ype M M 2 inerior ype M a(, ) a(, ) a(, ) a(, ) M 2 a(, ) a(, ) a(, ) a(, ) a(, ) a(, ) Figure 2: Acions Induced by Boundary and Inerior Types false advice when he sends messages in eiher M and M 2, here is one when sends m M 2. The specificaion of ψ, which allows incenive compaibiliy o hold for, (ineffecively) induce he acion a(, ) (wih aserisk) for he inerior ype if he sends messages in M 2, which is he same as he acion effecively induced by m M. If we could fix he profile of acions, ha U e 2( ) > would have guaraneed ha < sricly prefers o send messages in M over M 2. However, when he exper s ype changes, he profile of acions also changes, and, insofar as he acions aken by (, ] are concerned, is indifferen beween M and M 2. Thus, we have o ensure ha, overall, prefers M enough for / (, ] so ha even wih he indifference for (, ] he incenive compaibiliy sill holds. For his, a sufficienly large U e 2( ) is required. A large U e 2( ) means ha he ideal acion of a higher is sufficienly higher han ha of a lower. This addiional resricion is nohing bu a srenghening of he already exising soring condiion.