Random processes and probability distributions. Phys 420/580 Lecture 20

Transcription

1 Random processes and probability distributions Phys 420/580 Lecture 20

2 Random processes Many physical processes are random in character: e.g., nuclear decay (Poisson distributed event count) P (k, τ) = e λτ (λτ) k! motion of thermalized interacting particles (Maxwell-Boltzmann speed profile) f(v) v 2 exp mv2 2kT

3 Random processes Probabilistic descriptions often arise from the complex behaviour of many interacting degrees of freedom E.g., Temperature is an emergent phenomenon: it s a collective property of a large number of interacting particles particles exchange energy and establish a MB distribution, characterized by a single parameter T

4 Ergodicity For a large collection of noninteracting oscillators, the phase space trajectory of each individual p oscillator is a closed circle q

5 Ergodicity For a large collection of noninteracting oscillators, the phase space trajectory of each individual p oscillator is a closed circle As we turn on interactions, the trajectories are perturbed q

6 Ergodicity For a large collection of noninteracting oscillators, the phase space trajectory of each individual p oscillator is a closed circle As we turn on interactions, the trajectories are perturbed q The trajectory density averaged over all oscillators can be understood as a probability density

7 Ergodicity Consider an idealized billiards system... non-ergodic ergodic

8 Statistical physics Ergodicity is the basis for statistical physics We assume that all states in the phase space are accessible and have equal weight In a system at fixed temperature, states are visited according to the Boltzmann factor: Z = e E n/kt n Stochastic process with transition probabilities given by P n m e (E n E m )/kt

9 Random numbers What is a random number? Really no such thing Loose term of art referring to a sequence of independent numbers drawn randomly from some distribution Typically these are integer or real values uniformly distributed in some finite range

10 Random numbers An infinite sequence of digits: Is it random? (Humans are terrible at judging) Each of the digits 0 9occurs 10 of the time 1 Each pair of two successive digits occurs of the time

11 Random numbers Consider the first one million digits of the sequence: digit counts are distributed around the average every pattern is equally probable P ( )=P ( ) digits are completely uncorrelated given x, P (x = 0) = 1 10

12 Random numbers How do we generate sequences of random numbers? Strictly speaking, this isn t possible on a deterministic computer using finite arithmetic Nonetheless, it may be possible to construct long sequences with the appearance of randomness Probably okay if the relationship between numbers has no physical significance

13 Linear congruential generator Want a random sequence of real numbers (U n ) [0, 1) Popular strategy: use fractions U n = X n /m built from the sequence of integers (X n ) {0, 1, 2,...,m 1} linear congruence scheme (Lehmer 1948) X n+1 =(ax n + c) mod m

14 Linear congruential generator Want a random sequence of real numbers (U n ) [0, 1) Popular strategy: use fractions U n = X n /m built from the sequence of integers (X n ) {0, 1, 2,...,m 1} linear congruence scheme (Lehmer 1948) X n+1 =(ax n + c) mod m multiplier increment modulus

15 Linear congruential generator Recursion builds off an initial seed value, X 0 X 0 X 1 =(ax 0 + c) mod m X 2 =(a[(ax 0 + c) mod m]+c) mod m... Requires careful choice of parameters: (X n )=7, 6, 9, 0, 7, 6, 9, 0,... X 0 = a = c =7 m = 10

16 Linear congruential generator Recursion builds off an initial seed value, X 0 X 0 X 1 =(ax 0 + c) mod m X 2 =(a[(ax 0 + c) mod m]+c) mod m... Requires careful choice of parameters: (X n )=7, 6, 9, 0, 7, 6, 9, 0,... X 0 = a = c =7 m = 10 very non-random, period 4

17 Linear congruential generator Any generator of the form X n+1 = F (X n ) taking m distinct values must be periodic with period P m For a linear congruential generator, one can show that the period is maximum if c is relatively prime to m; b = a 1 is a multiple of p, for very prime p dividing m b is a multiple of 4, if m is a multiple of 4.

18 Linear congruential generator For the sake of efficiency, specialize to the case m =2 32 Advantages: each integer fits into exactly one computer word the modulus (an expensive division operation) is automatically handled in hardware by overflow A specialized version of the maximum period rule: c =1 a 5 mod 8

19 Linear congruential generator A long repeating cycle does not imply randomness: (X n )=0, 1, 2,...,m 1 (a = c = 1) We require weak correlation between elements, X j X k X j X k (j = k), especially when j k is small Judged empirically via statistical tests (Die Hard)

20 Linear congruential generator Some other considerations: a = 10,c= 23,m= 566 best if a/m is not too small least significant bits are more X 2n+1 highly correlated complete orbits lie in hyperplanes (Marsaglia 1968): X 2n (X 0,X 1,...,X q 1 ), (X q,x q+1,...,x 2q 1 ),...

21 Probability distributions Suppose there are N discrete events occur with probabilities p 1,p 2,...,p N Since something must happen, the total sum is p 1 + p p N =1 Given a randomly generated number ξ [0, 1], how can we select one of the N events?

22 Probability distributions Each even i occupies a width p i in the interval: 1 0 < ξ <p 1 2 p 1 < ξ <p 1 + p 2 3 p 1 + p 2 < ξ <p 1 + p 2 + p 3... N p p N 1 < ξ < 1 p 0 1 p 2 p N 1

23 Probability distributions The relevant quantity is the cumulative probability, P i = i j=1 p j Similarly, for continuous distributions, we construct a cumulative probability distribution P (x) = x from the probability density dy p(y) p(x)

24 Probability distributions Sampling via p(x) x P 1 ([0, 1]) P(x) ξ x x p(x) P(x) ξ x x

25 Inverse transform method Example: p(x) = (1/λ)e x/λ if 0 x< 0 if x<0 Back map gives ξ P (x) = x 0 dy p(y) =1 e x/λ By inversion, we see that x λ ln(1 ξ) with ξ drawn uniformly from [0, 1] is equivalent to x drawn from the nonuniform distribution p(x)

26 Inverse transform method What about cases where no analytic inverse exists? In some cases, related multivariate distributions are invertible: e.g. Gaussian distribution p(x) = e x2 /2σ 2 2πσ 2 Consider the product p(x, y) =p(x)p(y) = e (x2 +y 2 )/2σ 2 2πσ 2

27 Inverse transform method From radial coordinates r = x 2 + y 2 θ = tan 1 (y/x) we can further transform ρ = r 2 /2 so that p(ρ, θ)dρ dθ = 1 2π e ρ dρ dθ x = 2ρ cos θ y = 2ρ sin θ Sampling is now possible with two random variables: x 2 ln(1 ξ 1 ) cos(2πξ 2 ) (Box-Muller)

28 Rejection method When P (x) is not easily invertible and no other tricks can be applied, try the following rejection method: Generate a sequence (x 1,x 2,x 3, ) with the elements drawn uniformly from [x min,x max ] Generate a sequence (ξ 1, ξ 2, ξ 3, ) with the elements drawn uniformly from [0,p max ] Discard elements from the first sequence if p(x i ) < ξ i

29 Rejection method Works by throwing away p(x) reject results that are rare p max Some limitations: the probability distribution must be bounded and have a finite range x min x max x accept