Bayesian epistemology II: Arguments for Probabilism

Similar documents
Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space.

APPENDIX A Some Linear Algebra

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Affine transformations and convexity

CS286r Assign One. Answer Key

FACTORIZATION IN KRULL MONOIDS WITH INFINITE CLASS GROUP

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Week 2. This week, we covered operations on sets and cardinality.

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017

THE WEIGHTED WEAK TYPE INEQUALITY FOR THE STRONG MAXIMAL FUNCTION

a b a In case b 0, a being divisible by b is the same as to say that

REAL ANALYSIS I HOMEWORK 1

Complete subgraphs in multipartite graphs

The Second Anti-Mathima on Game Theory

Google PageRank with Stochastic Matrix

More metrics on cartesian products

2.3 Nilpotent endomorphisms

and problem sheet 2

12 MATH 101A: ALGEBRA I, PART C: MULTILINEAR ALGEBRA. 4. Tensor product

(1 ) (1 ) 0 (1 ) (1 ) 0

Lecture 17 : Stochastic Processes II

Lecture 3: Probability Distributions

Fully simple singularities of plane and space curves

BOUNDEDNESS OF THE RIESZ TRANSFORM WITH MATRIX A 2 WEIGHTS

Stochastic Structural Dynamics

Volume 18 Figure 1. Notation 1. Notation 2. Observation 1. Remark 1. Remark 2. Remark 3. Remark 4. Remark 5. Remark 6. Theorem A [2]. Theorem B [2].

DIFFERENTIAL FORMS BRIAN OSSERMAN

MAT 578 Functional Analysis

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0

Exercise Solutions to Real Analysis

DIFFERENTIAL SCHEMES

Appendix for Causal Interaction in Factorial Experiments: Application to Conjoint Analysis

k t+1 + c t A t k t, t=0

Assortment Optimization under MNL

Polynomials. 1 More properties of polynomials

= z 20 z n. (k 20) + 4 z k = 4

Representation theory and quantum mechanics tutorial Representation theory and quantum conservation laws

arxiv: v1 [math.co] 1 Mar 2014

j) = 1 (note sigma notation) ii. Continuous random variable (e.g. Normal distribution) 1. density function: f ( x) 0 and f ( x) dx = 1

ON THE EXTENDED HAAGERUP TENSOR PRODUCT IN OPERATOR SPACES. 1. Introduction

MATH 5707 HOMEWORK 4 SOLUTIONS 2. 2 i 2p i E(X i ) + E(Xi 2 ) ä i=1. i=1

A 2D Bounded Linear Program (H,c) 2D Linear Programming

Affine and Riemannian Connections

Department of Computer Science Artificial Intelligence Research Laboratory. Iowa State University MACHINE LEARNING

On the Operation A in Analysis Situs. by Kazimierz Kuratowski

Perfect Competition and the Nash Bargaining Solution

FINITELY-GENERATED MODULES OVER A PRINCIPAL IDEAL DOMAIN

Engineering Risk Benefit Analysis

6. Stochastic processes (2)

MATH 241B FUNCTIONAL ANALYSIS - NOTES EXAMPLES OF C ALGEBRAS

COMPLEX NUMBERS AND QUADRATIC EQUATIONS

6. Stochastic processes (2)

GELFAND-TSETLIN BASIS FOR THE REPRESENTATIONS OF gl n

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4)

SL n (F ) Equals its Own Derived Group

Online Appendix. t=1 (p t w)q t. Then the first order condition shows that

Economics 101. Lecture 4 - Equilibrium and Efficiency

Lecture 21: Numerical methods for pricing American type derivatives

Games of Threats. Elon Kohlberg Abraham Neyman. Working Paper

On the set of natural numbers

Genericity of Critical Types

p 1 c 2 + p 2 c 2 + p 3 c p m c 2

Advanced Introduction to Machine Learning

Appendix B. Criterion of Riemann-Stieltjes Integrability

Formulas for the Determinant

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction

Subset Topological Spaces and Kakutani s Theorem

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009

10-801: Advanced Optimization and Randomized Methods Lecture 2: Convex functions (Jan 15, 2014)

Math 101 Fall 2013 Homework #7 Due Friday, November 15, 2013

Expected Value and Variance

P exp(tx) = 1 + t 2k M 2k. k N

Differential Polynomials

20. Mon, Oct. 13 What we have done so far corresponds roughly to Chapters 2 & 3 of Lee. Now we turn to Chapter 4. The first idea is connectedness.

SUCCESSIVE MINIMA AND LATTICE POINTS (AFTER HENK, GILLET AND SOULÉ) M(B) := # ( B Z N)

Discussion 11 Summary 11/20/2018

Equilibrium with Complete Markets. Instructor: Dmytro Hryshko

Lecture 3. Ax x i a i. i i

Another converse of Jensen s inequality

1 Matrix representations of canonical matrices

R n α. . The funny symbol indicates DISJOINT union. Define an equivalence relation on this disjoint union by declaring v α R n α, and v β R n β

C/CS/Phy191 Problem Set 3 Solutions Out: Oct 1, 2008., where ( 00. ), so the overall state of the system is ) ( ( ( ( 00 ± 11 ), Φ ± = 1

Fixed points of IA-endomorphisms of a free metabelian Lie algebra

ECE 534: Elements of Information Theory. Solutions to Midterm Exam (Spring 2006)

Singular Value Decomposition: Theory and Applications

On the Correlation between Boolean Functions of Sequences of Random Variables

Rapid growth in finite simple groups

Weighted Voting Systems

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

Calculation of time complexity (3%)

ELASTIC WAVE PROPAGATION IN A CONTINUOUS MEDIUM

GENERAL EQUILIBRIUM IN INFINITE SECURITY MARKETS

Math 2534 Final Exam Review Answer Key 1. Know deþnitions for the various terms: a. Modus Ponens b. Modus Tollens c. divides d. rational e.

χ x B E (c) Figure 2.1.1: (a) a material particle in a body, (b) a place in space, (c) a configuration of the body

CHALMERS, GÖTEBORGS UNIVERSITET. SOLUTIONS to RE-EXAM for ARTIFICIAL NEURAL NETWORKS. COURSE CODES: FFR 135, FIM 720 GU, PhD

The Geometry of Logit and Probit

THE CLASS NUMBER THEOREM

Matrix Approximation via Sampling, Subspace Embedding. 1 Solving Linear Systems Using SVD

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016

Transcription:

Bayesan epstemology II: Arguments for Probablsm Rchard Pettgrew May 9, 2012 1 The model Represent an agent s credal state at a gven tme t by a credence functon c t : F [0, 1]. where F s the algebra of propostons about whch the agent has an opnon. 1 If A F, then c t (A) = 0 ff the agent has mnmal credence n A at t. If A F, then c t (A) = 1 ff the agent has maxmal credence n A at t. Note: It s an emprcal assumpton that agents are capable of maxmal and mnmal credences; t s not a normatve clam. 2 The norms At any tme t n her epstemc lfe, an agent ought to have a credence functon c t such that Probablsm c t s a probablty functon. That s, c t ( ) = 0 and c t ( ) = 1. c t (A B) = c t (A) + c t (B) f A and B are mutually exclusve. Countable addtvty c t s countably addtve. That s, f F s nfnte, c t ( n A n ) = n c t (A n ) f A 1, A 2,... are parwse mutually exclusve. 1 Snce F s an algebra, t s closed under conjunctons, dsjunctons, and negatons. 1

3 Dutch Book arguments 3.1 Probablsm We wll assume that F s fnte. Ths assumpton s not necessary, but t smplfes proofs. We begn by gvng an alternatve formulaton of Probablsm. Defnton 1 An assgnment of truth values to the propostons n F s a functon v : F 0, 1} such that 0 f v(a) = 1 v( A) = 1 f v(a) = 0 and 0 f v(a) = 0 and v(b) = 0 v(a B) = 1 otherwse Defnton 2 Let V be the set of all assgnments of truth values to propostons n F. We mght thnk of each assgnment of truth values as a possble world. Thus, V s the set of all possble worlds. Note that snce F s fnte, V s fnte. Defnton 3 Let V + be the convex hull of V. That s, V + := λ v v : λ v > 0 and λ = 1 Another characterzaton of V + : t s the smallest convex set that contans all elements of V. Lemma 1 An agent satsfes Probablsm ff her credence functon c t s n V +. Proof. Suppose c t V +. To see that c t s a probablty functon, t suffces to note that: () Each v V s a probablty functon. () If p and p are probablty functons, then λp + (1 λ)p s a probablty functon. For the converse, suppose that c t s a probablty functon. Then, for each v V, let A v := A A v(a)=0 Thus, A v s the unque proposton n F such that v 0 f v = v (A v ) = 1 f v = v That s, A v s made true by v but by no other truth assgnment. Thus: A = A v } 2

And the A v s are dsjont. Now let Then, snce c t s a probablty functon, c t (A) = c t ( A v ) = λ v := c t (A v ) c t (A v ) = v(a)c t (A v ) = λ v v(a) as requred. So far, we have been treatng credence functons and truth value assgnments as functons from F nto [0, 1]. But, f F = A 1,..., A n }, then we mght just as well consder them as vectors n an n-dmensonal vector space. Thus, f c s a credence functon, we represent t as c = (c 1,..., c n ) where c = c(a ). Smlarly, f v s a truth value assgnment, we represent t as v = (v 1,..., v n ) where v = v(a ). Usng ths notaton, we can better state the Dutch Book argument. A Dutch book s a book of bets on the propostons n F and a prce for that book such that the prce s greater than the payoff of the book of bets n every possble world. We represent ths mathematcally as follows: Then a book of bets on F s represented by a vector (s 1,..., s n ). Ths s a set of n bets B 1,..., B n, where: B wll pay s f A s true; B wll pay 0 f A s false. Suppose p s the prce for bet B. Then the prce of the book (s 1,..., s n ) s p. The payoff of the book B = (s 1,..., s n ) at v V s v s Thus, a book (s 1,..., s n ) wth prces p for bet B s a Dutch Book ff for all v V. p > v s 3

3.1.1 The argument (1) Credences as bettng odds An agent wth credence r n proposton A should consder rs as a far prce for a bet that pays S f A s true and 0 f A s false. (2) Package prncple If an agent consders r as a far prce for bet 1 and r as a far prce for bet 2, then she should consder (r + r ) as a far prce for the book of bets that conssts of bets 1 and 2. (3) Undutchbookable An agent should not have credences that lead her to acceptng a Dutch book as far. (3) Theorem 1 (Dutch book theorem) (I) If c V +, then there s a book of bets (s 1,..., s n ) such that, for all v V, c s > v s (II) If c V +, then there s no book of bets (s 1,..., s n ) such that, for all v V, Therefore, (4) An agent ought to obey Probablsm. 3.1.2 Proof of Theorem 1 c s > v s (I) Suppose c V +. Then let p V + be the pont n V + that s closest to c. Then let s = c p. Then, by a classcal geometrcal result, we have that the angle between s and v p s not acute, for any v V. Thus s (v p) 0. Ths gves s v s p. But we also have s 2 > 0. But s 2 = s s = s (c p) = s c s p Thus, s p < s c. So s v < s c. That s, as requred. c s > v s (II) Suppose c V +. And let s be a vector. Then, ether the angle between v c and s s rght for all v V or for at least one v V, the angle between s and v c s acute. If the angle between s and v c s rght for all v V, then s (v c) = 0, so c s = v s for all v V. If the angle between s and v c s obtuse, then s (v c) > 0, so Ths completes the proof. c s < v s 4

4 Accuracy domnaton arguments In accuracy domnaton arguments, we treat epstemc states as epstemc acts, we ntroduce measures of epstemc utlty for those acts, and we employ the machnery of decson theory to derve norms that govern epstemc states. An epstemc utlty argument requres: An epstemc utlty functon For each credence functon c and possble world w, EU(c, w) measures the epstemc goodness of havng c at w. An example: The Brer score s the followng measure: B(c, w) = 1 (c(a) v w (A)) 2 A F where v w (A) = 0 f A s false 1 f A s true A norm of decson theory Ths tells us how an agent should choose from a range of dfferent acts on the bass of the epstemc utlty of those acts at dfferent worlds. Example: Act-Type Domnance Suppose there are two sorts of act: D 1 and D 2. Say that an act D s domnated by another act D f D has hgher utlty than D n every world. Now suppose: Every D n D 1 s domnated by some D n D 2. No D n D 2 s domnated by any D n D 1 or n D 2. Then the agent should choose an act from D 2. 4.1 The argument (1) The Brer score measures epstemc utlty (2) Act-Type Domnance (3) Theorem 2 (de Fnett) (I) If c V +, then there s p V + such that for all worlds w. B(c, w) < B(p, w) (II) If c V +, then there s no credence functon p such that for all worlds w. Therefore, (4) An agent ought to obey Probablsm. B(c, w) < B(p, w) 5