Quasi-Monte Carlo Methods Fall 2013 By Yaohang Li, Ph.D. Review • Last Class – Monte Carlo Solver for PDE • This Class – Quasi-Monte Carlo • Next Class – Markov Chain Monte Carlo Random Numbers • Random Numbers – Pseudorandom Numbers • Monte Carlo Methods – Quasirandom Numbers • Uniformity • Low-discrepancy • Quasi-Monte Carlo Methods – Mixed-random Numbers • Hybrid-Monte Carlo Methods Discrepancy •Discrepancy – For one dimension DN* DN* ( x1 ,..., xn ) sup | 0u 1 1 N N n 1 [ 0 ,u ) ( xn ) u | • is the number of points in interval [0,u) – For d dimensions DN* DN* ( x1 ,..., xn ) sup | E # ofxi E m( E ) | N • E: a sub-rectangle • m(E): the volume of E A Picture is Worth a Thousand Words Quasi-Monte Carlo •Motivation – Convergence • Monte Carlo methods: O(N-1/2) • quasi-Monte Carlo methods: O(N-1) – Integration error bound • Koksma-Hlwaka Inequality Theorem 1 1 N | f ( xn ) f ( x)dx | V ( f ) DN* N n1 0 – V(f): bounded variation • Criterion [ f ] V [ f ]DN* V [ f ]c(log N ) k N 1 – k is a dimension dependent constant Quasi-Monte Carlo Integration • Quasi-Monte Carlo Integration – If x1, …, xn are from a quasirandom number sequence 1 0 1 n f ( x)dx f ( xi ) n i 1 – Compared with Crude Monte Carlo • Only difference is the underlying random numbers – Crude Monte Carlo » pseudorandom numbers – Quasi-Monte Carlo » quasirandom numbers Discrepancy of Pseudorandom Numbers and Quasirandom Numbers • Discrepancy of Pseudorandom Numbers – O(N-1/2) • Discrepancy of Quasirandom Numbers – O(N-1) Analysis of Quasi-Monte Carlo • Convergence Rate – O(N-1) • Actual Convergence Rate – O((logN)kN-1) • k is a constant related to dimension – when dimension is large (>48) • the (logN)k factor becomes large • the advantage of quasi-Monte Carlo disappears Quasi-random Numbers •van der Corput sequence – digit expansion n a j (n)b j j 0 – radical-inverse function b (n) a j (n)b ( j 1) j 0 • for an integer b>1, the van der Corput sequence in base b is {x0, x1, …} with xn=b(n) for all n>=0 Halton Sequence • Halton Sequence – s dimensional van der Corput sequence • xn=(b1(n), b2(n),…, bs(n)) – b1, b2, … bs are relatively prime bases • Scrambled Halton Sequence – Use permutations of digits in the digit expansion of each van der Corput sequence – Improve the randomness of the Halton sequence Discussion • In low diemensions (s<30 or 40), quasi-Monte Carlo methods in numerical integrations are better than usual Monte Carlo methods • Quasi-Monte Carlo method is deterministic method – Monte Carlo methods are statistic methods • There are serially efficient implementation of quasirandom number sequences – – – – Halton Sobol Faure Niederreiter • quasi-Monte Carlo can now efficiently used in integration – Still in research in other areas Summary • Quasirandom Numbers – Discrepancy – Implementation • van der Corput • Halton • Quasi-Monte Carlo – Integration – Convergence rate – Comparison with Crude Monte Carlo What I want you to do? • Review Slides • Review basic probability/statistics concepts • Select your presentation topic