Computer Science · Artificial Intelligence
Axioms and Algorithms for Inferences Involving Probabilistic Independence - eScholarship
Dan Geiger, Azaria Paz, Judea Pearl
See on the mapAbstract
TECHNICAL REPORT R-119 March 1991 Reprinted from INFORMATION AND COMPUTATION AH Rights Reserved by Academic Press, New York and London Vol. 91, No. 1, March 1991 Printed in Belgium Axioms and Algorithms for Inferences Involving Probabilistic Independence* DAN GEIGER, A Z A R I A P A Z , AND JUDEA PEARL Cognitive Systems Laboratory, Computer Science Department, Los Angeles, California 90024-1596 This paper offers an axiomatic characterization of the probabilistic relation X is independent of Y (written (X, Y)) where X and Y are two disjoint sets of variables. Four axioms for {X, Y) are presented and shown to be complete. Based on these axioms, a polynomial membership algorithm is developed to decide whether any given independence statement (X, Y) logically follows from a set E of such statements, i.e., whether (X, Y) holds in every probability distribution that satisfies E. The complexity of the algorithm is 0(|E| -k + |E| - n), where | E | is the number of given statements, n is the number of variables in E u {(X, Y)}, and k is the number of variables in (X, Y). © 1991 Academic Press, Inc. 1. I N T R O D U C T I O N Consider a collection of information sources, each reflecting a different aspect of some underlying probabilistic phenomenon. These sources can be regarded as a set of random variables, governed by an unknown proba- bility distribution, some of which are dependent and some independent. In this paper, we are concerned with the following problem: Assume we know that some groups of variables are mutually independent, either by statistical analysis or by conceptual understanding of the underlying phenomenon; we need to infer new independencies without resorting to additional measurements or expensive numerical analysis. We formalize this question as follows: Let an {independence) statement be a sentence of the form (X, Y), where X= { x , x } and Y= {y y ) are disjoint finite sets of variables. The meaning of (X, Y) is that X and Y are probabilistically independent, namely, n u n P(X, Y) = P(X) -P{Y) * This work was supported in part by the National Science Foundation Grant IRI-8821444. Permanent affiliation: Computer Science Department, Technion, Haifa, Israel 32000. f 0890-5401/91 S3.00 Copyright © 1991 by Academic Press, Inc. All rights of reproduction in any form reserved.