How does the computer generate observations from various distributions specified after input analysis?

Transcription

1 1

2 How does the computer generate observations from various distributions specified after input analysis? There are two main components to the generation of observations from probability distributions. 1. Random number generation. 2. Random variate generation. 2

3 Random number generation The generation of U(0,1) random variates (observations from Uniform (0,1) distribution). This serves as the foundation for the generation of observations from other distributions, which is called random variate generation. Random Number Generator is the term used to describe the procedure and parameters used to generate the U(0,1) observations. 3

4 Since the stream of random numbers generated is reproducible, random number generation procedures are also referred to as pseudo random number generators. The stream or sequence of numbers produced by a generator should pass statistical tests for randomness. An outside observer should not be able to tell the difference (statistically) between a stream of pseudo random numbers and an actual random number stream. 4

5 A pseudorandom process appears random, but isn t Pseudorandom sequences exhibit statistical randomness but generated by a deterministic process Pseudorandom sequences are easier to produce than a genuine random sequences Pseudorandom sequences can reproduce exactly the same numbers useful for testing and fixing software. 5

6 Random number generators typically compute the next number in the sequence from the previous number The first number in a sequence is called the seed to get a new sequence, supply a new seed (current machine time is useful) to repeat a sequence, repeat the seed 6

7 Desirable Attributes: Uniformity Independence Efficiency Replicability Long Cycle Length 7

8 Each random number R t is an independent sample drawn from a continuous uniform distribution between 0 and 1 1, 0 x 1 pdf: f(x) = 0, otherwise 8

9 9 1/12 4 1/ 3 1/ 2) (1/ 3] / [ )] ( [ ) ( 2 1/ 2] / [ ) ( x R E dx x R V x xdx R E x 0f(x) 1 PDF:

10 One early method the midsquare method (von Neumann and Metropolis 1940) Start with a four digit positive integer Z 0. Square Z 0 to get an integer with up to eight digits (append zeros if less than eight). Take the middle four digits as the next four digit integer Z 1. Place a decimal point to the left of Z 1 to form the first U(0,1) observation. Repeat 10

11 MidSquare Example: X 0 = 7182 (seed) 2 X 0 = ==> R 1 = X 0 = (5811) 2 = ==> R 2 = etc. 11

12 Note: Cannot choose a seed that guarantees that the sequence will not degenerate and will have a long period. Also, zeros, once they appear, are carried in subsequent numbers. Ex1: X 0 = 5197 (seed) = ==> R 1 = = ==> R 2 = Ex2: X 0 = 4500 (seed) = ==> R 1 = = ==> R 2 = X 0 2 X 1 2 X 0 X 1 12

13 The prior method does not work well. Degenerates to zero. What are good methods? Linear Congruential Generators (LCGs). Composite generators. Tausworthe generators. 13

14 Linear Congruential Generators (LCGs). A LCG generates a sequence of integers Z 1, Z 2, Z 3, using the following recursive formula, Z i ( az 1 i c) mod mod m is short for modulo m or the remainder when divided by m. m 14

15 Since the mod m operation is used, all Z i s will be between 0 and m-1. To get the U(0,1) random observations each Z i generated is divided by m. U 1 Z1 m Z m 2, U 2, So are the U i s really U(0,1) random observations? 15

16 Let m=63, a=22, c=4 and Z 0 =19. Generate the first five U(0,1) observations. 16

17 i 22*Z i Z i U i i 22*Z i Z i U i i 22*Z i Z i U i i 22*Z i Z i U i

18 What will happen after the 63 rd number is generated? m, a, and c are the parameters of the random number generator. There can be an infinite number of different implementations of a LCG. The values used for m, a, and c determine whether the generator is good or bad. 18

19 The example LCG demonstrates cycling in the prior table. Since m=63, it can generate at most 63 numbers before it repeats the same sequence. This small random number generator has full period since it generates all possible (m=63) numbers before cycling. A long period (full if possible) is desirable since more observations can be generated before cycling. No gaps. 19

20 The example generator has full period but bad statistical properties (next slide). A good random number generator will have values for m, a, and c such that full or close to full period is obtained, as well as good statistical properties. Crystal Ball m = a = c = 0 Period =

21 Theorem (Hull and Dobell 1962) The LCG Z i = (az i-1 + c) mod m has full period if and only if the following three conditions hold. 1. The only positive integer that exactly divides both m and c is If q is a prime number that divides m, then q divides a If 4 divides m, then 4 divides a-1. The parameters of the LCG dictate the period length of the LCG as well as other properties of the numbers generated. 21

22 Example: Using the multiplicative congruential method, find the period of the generator for a = 13, m = 2 6, and X 0 = 1, 2, 3, and 4. The solution is given in next slide. When the seed is 1 and 3, the sequence has period 16. However, a period of length eight is achieved when the seed is 2 and a period of length four occurs when the seed is 4. 22

23 Period Determination Using Various seeds i X i X i X i X i

24 Types of LCGs When c = 0, the LCG is called a multiplicative generator. When c 0, the LCG is called a mixed generator. Most LCGs implemented are multiplicative Can t have full period. How is m selected. A large period is desired m=2 31 (based on a 32 bit word size). With m=2 31 it has been proven that the period can be at most 2 29 (25% of the values are cycled and gaps may be present). 24

25 m has been selected as the largest prime number less than m=2 31, which is 2,147,483,647 = A period of m-1 can be guaranteed if the parameter a is primitive element modulo m. Selections for a that are primitive element modulo m and have been in use are: 16, ,360,016 25

26 LCGs are a special case of the form Z i = g(z i-1, Z i-2,...) (mod m), U i = Z i /m, for some function g Examples: g(z i-1 ) = az i-1 + c LCG g(z i-1, Z i-2,..., Z i-q ) = a 1 Z i-1 + a 2 Z i a q Z i-q multiple recursive generator g(z i-1 ) = a'z 2 i-1 + az i-1 + c quadratic CG g(z i-1, Z i-2 ) = Z i-1 + Z i-2 Fibonacci (bad) 26

27 Composite Generators Combine two (or more) individual generators in some way. Differencing LCGs Z 1i and Z 2i from LCGs with different moduli Let Z i = (Z 1i Z 2i ) (mod m); U i = Z i / m Very good statistical properties Very portable (micros, different languages) Wichmann/Hill Use three LCGs to get U 1i, U 2i, and U 3i sequences Let U i = fractional part of U 1i + U 2i + U 3i Long period, good statistics, portability 27

28 Tausworthe Generators Originated in cryptography Generate sequence of bits b1, b2, b3,... via congruence bi = (bi-r + bi-q) (mod 2) = Various algorithms to group bits into Ui s Can achieve very long periods Theoretical appeal: for properly chosen parameters, can prove that over a cycle, mean» 1/2 (as for true U(0,1)) variance» 1/12 (as for true U(0,1)) autocorrelation» 0 (as for true IID sequence) 28