Sampling according to the multivariate normal density

This paper deals with the normal density of n dependent random variables. This is a function of the form: ce(-x/sup T/Ax) where A is an n/spl times/n positive definite matrix, a: is the n-vector of the random variables and c is a suitable constant. The first problem we consider is the (approximate) evaluation of the integral of this function over the positive orthant /spl int/(x/sub 1/=0)/sup /spl infin///spl int/(x/sub 2/=0)/sup /spl infin///spl middot//spl middot//spl middot//spl int/(x/sub n/=0)/sup /spl infin//ce(-x/sup T/Ax). This problem has a long history and a substantial literature. Related to it is the problem of drawing a sample from the positive orthant with probability density (approximately) equal to ce(-x/sup T/Ax). We solve both these problems here in polynomial time using rapidly mixing Markov Chains. For proving rapid convergence of the chains to their stationary distribution, we use a geometric property called the isoperimetric inequality. Such an inequality has been the subject of recent papers for general log-concave functions. We use these techniques, but the main thrust of the paper is to exploit the special property of the normal density to prove a stronger inequality than for general log-concave functions. We actually consider first the problem of drawing a sample according to the normal density with A equal to the identity matrix from a convex set K in R/sup n/ which contains the unit ball. This problem is motivated by the problem of computing the volume of a convex set in a way we explain later. Also, the methods used in the solution of this and the orthant problem are similar.

[1]  G. Martynov.  Evaluation of the normal distribution function , 1981 .

[2]  D. Aldous.  Random walks on finite groups and rapidly mixing markov chains , 1983 .

[3]  P. Moran.  CALCULATION OF MULTIVARIATE NORMAL PROBABILITIES—ANOTHER SPECIAL CASE , 1985 .

[4]  Andrei Z. Broder,et al.  How hard is it to marry at random? (On the approximation of the permanent) , 1986, STOC '86.

[5]  D. Aldous.  On the Markov Chain Simulation Method for Uniform Combinatorial Distributions and Simulated Annealing , 1987, Probability in the Engineering and Informational Sciences.

[6]  S. Iyengar.  Evaluation of Normal Probabilities of Symmetric Regions , 1988 .

[7]  M. Evans,et al.  Monte carlo computation of some multivariate normal probabilities , 1988 .

[8]  Mark Jerrum,et al.  Approximating the Permanent , 1989, SIAM J. Comput..

[9]  Martin E. Dyer,et al.  A random polynomial-time algorithm for approximating the volume of convex bodies , 1991, JACM.

[10]  Martin E. Dyer,et al.  A Random Polynomial Time Algorithm for Approximating the Volume of Convex Bodies , 1989, STOC.

[11]  M. Dyer.  Computing the volume of convex bodies : a case where randomness provably helps , 1991 .

[12]  David Applegate,et al.  Sampling and integration of near log-concave functions , 1991, STOC '91.

[13]  Miklós Simonovits,et al.  Random Walks in a Convex Body and an Improved Volume Algorithm , 1993, Random Struct. Algorithms.

[14]  Ravi Kannan,et al.  Markov chains and polynomial time algorithms , 1994, Proceedings 35th Annual Symposium on Foundations of Computer Science.

[15]  Nicholas G. Polson,et al.  Sampling from log-concave distributions , 1994 .

[16]  Persi Diaconis,et al.  What do we know about the Metropolis algorithm? , 1995, STOC '95.