Chapter 10 Generating Functions 10.1 Generating Functions for Discrete Distribu- tions So far we have considered in detail only the two most important attributes of a random variable, namely, the mean and the variance. We have seen how these attributes enter into the fundamental limit theorems of probability, as well as into all sorts of practical calculations. We have seen that the mean and variance of a random variable contain important information about the random variable, or, more precisely, about the distribution function of that variable. Now we shall see that the mean and variance do not contain all the available information about the density function of a random variable. To begin with, it is easy to give examples of different distribution functions which have the same mean and the same variance. For instance, suppose X and Y are random variables, with distributions p X =  12 3456 01/41/2001/4  , p Y =  1234 56 1/4001/21/40  . Then with these choices, we have E(X)=E(Y )=7/2 and V (X)=V (Y )=9/4, and yet certainly p X and p Y are quite different density functions. This raises a question: If X is a random variable with range {x 1 ,x 2 , } of at most countable size, and distribution function p = p X , and if we know its mean µ = E(X) and its variance σ 2 = V (X), then what else do we need to know to determine p completely? Moments A nice answer to this question, at least in the case that X has finite range, can be given in terms of the moments of X, which are numbers defined as follows: 365 366 CHAPTER 10. GENERATING FUNCTIONS µ k = kth moment of X = E(X k ) = ∞  j=1 (x j ) k p(x j ) , provided the sum converges. Here p(x j )=P (X = x j ). In terms of these moments, the mean µ and variance σ 2 of X are given simply by µ = µ 1 , σ 2 = µ 2 − µ 2 1 , so that a knowledge of the first two moments of X gives us its mean and variance. But a knowledge of all the moments of X determines its distribution function p completely. Moment Generating Functions To see how this comes about, we introduce a new variable t, and define a function g(t) as follows: g(t)=E(e tX ) = ∞  k=0 µ k t k k! = E  ∞  k=0 X k t k k!  = ∞  j=1 e tx j p(x j ) . We call g(t) the moment generating function for X, and think of it as a convenient bookkeeping device for describing the moments of X. Indeed, if we differentiate g(t) n times and then set t =0,wegetµ n : d n dt n g(t)     t=0 = g (n) (0) = ∞  k=n k! µ k t k−n (k −n)! k!      t=0 = µ n . It is easy to calculate the moment generating function for simple examples. 10.1. DISCRETE DISTRIBUTIONS 367 Examples Example 10.1 Suppose X has range {1, 2, 3, ,n} and p X (j)=1/n for 1 ≤ j ≤ n (uniform distribution). Then g(t)= n  j=1 1 n e tj = 1 n (e t + e 2t + ···+ e nt ) = e t (e nt − 1) n(e t − 1) . If we use the expression on the right-hand side of the second line above, then it is easy to see that µ 1 = g  (0) = 1 n (1+2+3+···+ n)= n +1 2 , µ 2 = g  (0) = 1 n (1+4+9+···+ n 2 )= (n + 1)(2n +1) 6 , and that µ = µ 1 =(n +1)/2 and σ 2 = µ 2 − µ 2 1 =(n 2 − 1)/12. ✷ Example 10.2 Suppose now that X has range {0, 1, 2, 3, ,n} and p X (j)=  n j  p j q n−j for 0 ≤ j ≤ n (binomial distribution). Then g(t)= n  j=0 e tj  n j  p j q n−j = n  j=0  n j  (pe t ) j q n−j =(pe t + q) n . Note that µ 1 = g  (0) = n(pe t + q) n−1 pe t   t=0 = np , µ 2 = g  (0) = n(n − 1)p 2 + np , so that µ = µ 1 = np, and σ 2 = µ 2 − µ 2 1 = np(1 − p), as expected. ✷ Example 10.3 Suppose X has range {1, 2, 3, } and p X (j)=q j−1 p for all j (geometric distribution). Then g(t)= ∞  j=1 e tj q j−1 p = pe t 1 − qe t . 368 CHAPTER 10. GENERATING FUNCTIONS Here µ 1 = g  (0) = pe t (1 − qe t ) 2     t=0 = 1 p , µ 2 = g  (0) = pe t + pqe 2t (1 − qe t ) 3     t=0 = 1+q p 2 , µ = µ 1 =1/p, and σ 2 = µ 2 − µ 2 1 = q/p 2 , as computed in Example 6.26. ✷ Example 10.4 Let X have range {0, 1, 2, 3, } and let p X (j)=e −λ λ j /j! for all j (Poisson distribution with mean λ). Then g(t)= ∞  j=0 e tj e −λ λ j j! = e −λ ∞  j=0 (λe t ) j j! = e −λ e λe t = e λ(e t −1) . Then µ 1 = g  (0) = e λ(e t −1) λe t    t=0 = λ, µ 2 = g  (0) = e λ(e t −1) (λ 2 e 2t + λe t )    t=0 = λ 2 + λ, µ = µ 1 = λ, and σ 2 = µ 2 − µ 2 1 = λ. The variance of the Poisson distribution is easier to obtain in this way than directly from the definition (as was done in Exercise 6.2.30). ✷ Moment Problem Using the moment generating function, we can now show, at least in the case of a discrete random variable with finite range, that its distribution function is com- pletely determined by its moments. Theorem 10.1 Let X be a discrete random variable with finite range {x 1 ,x 2 , ,x n } , and moments µ k = E(X k ). Then the moment series g(t)= ∞  k=0 µ k t k k! converges for all t to an infinitely differentiable function g(t). Proof. We know that µ k = n  j=1 (x j ) k p(x j ) . 10.1. DISCRETE DISTRIBUTIONS 369 If we set M = max |x j |, then we have |µ k |≤ n  j=1 |x j | k p(x j ) ≤ M k · n  j=1 p(x j )=M k . Hence, for all N we have N  k=0     µ k t k k!     ≤ N  k=0 (M|t|) k k! ≤ e M|t| , which shows that the moment series converges for all t. Since it is a power series, we know that its sum is infinitely differentiable. This shows that the µ k determine g(t). Conversely, since µ k = g (k) (0), we see that g(t) determines the µ k . ✷ Theorem 10.2 Let X be a discrete random variable with finite range {x 1 ,x 2 , , x n }, distribution function p, and moment generating function g. Then g is uniquely determined by p, and conversely. Proof. We know that p determines g, since g(t)= n  j=1 e tx j p(x j ) . In this formula, we set a j = p(x j ) and, after choosing n convenient distinct values t i of t,wesetb i = g(t i ). Then we have b i = n  j=1 e t i x j a j , or, in matrix notation B = MA . Here B =(b i ) and A =(a j ) are column n-vectors, and M =(e t i x j )isann × n matrix. We can solve this matrix equation for A: A = M −1 B , provided only that the matrix M is invertible (i.e., provided that the determinant of M is different from 0). We can always arrange for this by choosing the values t i = i − 1, since then the determinant of M is the Vandermonde determinant det       111··· 1 e tx 1 e tx 2 e tx 3 ··· e tx n e 2tx 1 e 2tx 2 e 2tx 3 ··· e 2tx n ··· e (n−1)tx 1 e (n−1)tx 2 e (n−1)tx 3 ··· e (n−1)tx n       370 CHAPTER 10. GENERATING FUNCTIONS of the e x i , with value  i<j (e x i −e x j ). This determinant is always different from 0 if the x j are distinct. ✷ If we delete the hypothesis that X have finite range in the above theorem, then the conclusion is no longer necessarily true. Ordinary Generating Functions In the special but important case where the x j are all nonnegative integers, x j = j, we can prove this theorem in a simpler way. In this case, we have g(t)= n  j=0 e tj p(j) , and we see that g(t)isapolynomial in e t . If we write z = e t , and define the function h by h(z)= n  j=0 z j p(j) , then h(z) is a polynomial in z containing the same information as g(t), and in fact h(z)=g(log z) , g(t)=h(e t ) . The function h(z) is often called the ordinary generating function for X. Note that h(1) = g(0)=1,h  (1) = g  (0) = µ 1 , and h  (1) = g  (0) −g  (0) = µ 2 −µ 1 . It follows from all this that if we know g(t), then we know h(z), and if we know h(z), then we can find the p(j) by Taylor’s formula: p(j) = coefficient of z j in h(z) = h (j) (0) j! . For example, suppose we know that the moments of a certain discrete random variable X are given by µ 0 =1, µ k = 1 2 + 2 k 4 , for k ≥ 1 . Then the moment generating function g of X is g(t)= ∞  k=0 µ k t k k! =1+ 1 2 ∞  k=1 t k k! + 1 4 ∞  k=1 (2t) k k! = 1 4 + 1 2 e t + 1 4 e 2t . 10.1. DISCRETE DISTRIBUTIONS 371 This is a polynomial in z = e t , and h(z)= 1 4 + 1 2 z + 1 4 z 2 . Hence, X must have range {0, 1, 2}, and p must have values {1/4, 1/2, 1/4}. Properties Both the moment generating function g and the ordinary generating function h have many properties useful in the study of random variables, of which we can consider only a few here. In particular, if X is any discrete random variable and Y = X + a, then g Y (t)=E(e tY ) = E(e t(X+a) ) = e ta E(e tX ) = e ta g X (t) , while if Y = bX, then g Y (t)=E(e tY ) = E(e tbX ) = g X (bt) . In particular, if X ∗ = X − µ σ , then (see Exercise 11) g x ∗ (t)=e −µt/σ g X  t σ  . If X and Y are independent random variables and Z = X + Y is their sum, with p X , p Y , and p Z the associated distribution functions, then we have seen in Chapter 7 that p Z is the convolution of p X and p Y , and we know that convolution involves a rather complicated calculation. But for the generating functions we have instead the simple relations g Z (t)=g X (t)g Y (t) , h Z (z)=h X (z)h Y (z) , that is, g Z is simply the product of g X and g Y , and similarly for h Z . To see this, first note that if X and Y are independent, then e tX and e tY are independent (see Exercise 5.2.38), and hence E(e tX e tY )=E(e tX )E(e tY ) . 372 CHAPTER 10. GENERATING FUNCTIONS It follows that g Z (t)=E(e tZ )=E(e t(X+Y ) ) = E(e tX )E(e tY ) = g X (t)g Y (t) , and, replacing t by log z, we also get h Z (z)=h X (z)h Y (z) . Example 10.5 If X and Y are independent discrete random variables with range {0, 1, 2, ,n} and binomial distribution p X (j)=p Y (j)=  n j  p j q n−j , and if Z = X + Y , then we know (cf. Section 7.1) that the range of X is {0, 1, 2, ,2n} and X has binomial distribution p Z (j)=(p X ∗ p Y )(j)=  2n j  p j q 2n−j . Here we can easily verify this result by using generating functions. We know that g X (t)=g Y (t)= n  j=0 e tj  n j  p j q n−j =(pe t + q) n , and h X (z)=h Y (z)=(pz + q) n . Hence, we have g Z (t)=g X (t)g Y (t)=(pe t + q) 2n , or, what is the same, h Z (z)=h X (z)h Y (z)=(pz + q) 2n = 2n  j=0  2n j  (pz) j q 2n−j , from which we can see that the coefficient of z j is just p Z (j)=  2n j  p j q 2n−j . ✷ 10.1. DISCRETE DISTRIBUTIONS 373 Example 10.6 If X and Y are independent discrete random variables with the non-negative integers {0, 1, 2, 3, } as range, and with geometric distribution func- tion p X (j)=p Y (j)=q j p, then g X (t)=g Y (t)= p 1 − qe t , and if Z = X + Y , then g Z (t)=g X (t)g Y (t) = p 2 1 − 2qe t + q 2 e 2t . If we replace e t by z,weget h Z (z)= p 2 (1 − qz) 2 = p 2 ∞  k=0 (k +1)q k z k , and we can read off the values of p Z (j) as the coefficient of z j in this expansion for h(z), even though h(z) is not a polynomial in this case. The distribution p Z is a negative binomial distribution (see Section 5.1). ✷ Here is a more interesting example of the power and scope of the method of generating functions. Heads or Tails Example 10.7 In the coin-tossing game discussed in Example 1.4, we now consider the question “When is Peter first in the lead?” Let X k describe the outcome of the kth trial in the game X k =  +1, if kth toss is heads, −1, if kth toss is tails. Then the X k are independent random variables describing a Bernoulli process. Let S 0 = 0, and, for n ≥ 1, let S n = X 1 + X 2 + ···+ X n . Then S n describes Peter’s fortune after n trials, and Peter is first in the lead after n trials if S k ≤ 0 for 1 ≤ k<nand S n =1. Now this can happen when n = 1, in which case S 1 = X 1 = 1, or when n>1, in which case S 1 = X 1 = −1. In the latter case, S k = 0 for k = n −1, and perhaps for other k between 1 and n. Let m be the least such value of k; then S m = 0 and 374 CHAPTER 10. GENERATING FUNCTIONS S k < 0 for 1 ≤ k<m. In this case Peter loses on the first trial, regains his initial position in the next m − 1 trials, and gains the lead in the next n − m trials. Let p be the probability that the coin comes up heads, and let q =1−p. Let r n be the probability that Peter is first in the lead after n trials. Then from the discussion above, we see that r n =0, if n even, r 1 = p (= probability of heads in a single toss), r n = q(r 1 r n−2 + r 3 r n−4 + ···+ r n−2 r 1 ) , if n>1,nodd. Now let T describe the time (that is, the number of trials) required for Peter to take the lead. Then T is a random variable, and since P (T = n)=r n , r is the distribution function for T . We introduce the generating function h T (z) for T : h T (z)= ∞  n=0 r n z n . Then, by using the relations above, we can verify the relation h T (z)=pz + qz(h T (z)) 2 . If we solve this quadratic equation for h T (z), we get h T (z)= 1 ±  1 − 4pqz 2 2qz = 2pz 1 ∓  1 − 4pqz 2 . Of these two solutions, we want the one that has a convergent power series in z (i.e., that is finite for z = 0). Hence we choose h T (z)= 1 −  1 − 4pqz 2 2qz = 2pz 1+  1 − 4pqz 2 . Now we can ask: What is the probability that Peter is ever in the lead? This probability is given by (see Exercise 10) ∞  n=0 r n = h T (1) = 1 −  1 − 4pq 2q = 1 −|p − q| 2q =  p/q, if p<q, 1, if p ≥ q, so that Peter is sure to be in the lead eventually if p ≥ q. How long will it take? That is, what is the expected value of T ? This value is given by E(T )=h  T (1) =  1/(p − q), if p>q, ∞, if p = q. [...]... 2 10 0 Z8 0 1 0 3 3 0 3 1 0 0 9 0 0 4 0 0 0 3 13 0 Z9 0 2 0 2 5 0 3 0 0 0 12 0 0 4 0 0 0 3 16 0 Z10 0 3 0 2 8 0 3 0 0 0 12 0 0 6 0 0 0 3 17 0 Z11 0 3 0 1 6 0 4 0 0 0 13 0 0 4 0 0 0 3 15 0 Z12 0 6 0 1 6 0 6 0 0 0 15 0 0 5 0 0 0 2 18 0 Profit -5 0 250 -1 00 50 250 -1 00 300 -5 0 -1 00 -5 0 750 -5 0 -5 0 200 -5 0 -5 0 -5 0 50 850 -5 0 Table 10. 4: Simulation of chain letter (finite distribution case) 392 CHAPTER 10. .. 10 9 6 0 Z6 8 0 0 0 0 1 0 0 0 17 0 3 0 0 0 0 11 5 9 0 Z7 11 0 0 0 0 2 0 0 0 14 0 1 0 0 0 0 9 7 10 0 Z8 9 0 0 0 0 4 0 0 0 11 0 0 0 0 0 0 11 9 13 0 Z9 7 0 0 0 0 9 0 0 0 11 0 0 0 0 0 0 12 8 0 0 Z10 6 0 0 0 0 7 0 0 0 10 0 0 0 0 0 0 14 8 0 0 Z11 6 0 0 0 0 9 0 0 0 16 0 0 0 0 0 0 13 6 0 0 Z12 5 0 0 0 0 7 0 0 0 25 0 0 0 0 0 0 10 3 0 0 Profit 200 -5 0 -5 0 -5 0 -1 00 300 0 -5 0 0 1300 -1 00 -5 0 -1 00 0 50 -5 0 550 100 ... there are 100 0 offspring, the offspring of the next generation are the result of 100 0 chance events, and it will take a while to simulate these 100 0 experiments However, since the final result is 388 CHAPTER 10 GENERATING FUNCTIONS 3 2.5 2 1.5 1 0.5 5 10 15 20 25 Figure 10. 4: Simulation of Zn /mn for the Keyfitz example the sum of 100 0 independent experiments we can use the Central Limit Theorem to replace... involved has a probability p of selling the letter Then you will receive 50 dollars with probability p and another 50 dollars if the letter is sold to 12 people, since then your name would have risen to the top of the list This occurs with probability p12 , and so your expected winnings are 100 + 50p + 50p12 Thus the chain in this situation is a highly unfavorable game It would be more reasonable to allow... and represents the probability that the process will die out 10. 2 BRANCHING PROCESSES Generation 1 2 3 4 5 6 7 8 9 10 11 12 383 Probability of dying out 2 312 385203 437116 475879 505878 529713 549035 564949 578225 589416 598931 Table 10. 1: Probability of dying out p0 p1 p2 p3 p4 p5 p6 p7 p8 p9 p10 = 2092 = 2584 = 2360 = 1593 = 0828 = 0357 = 0133 = 0042 = 0011 = 0002 = 0000 Table 10. 2: Distribution... process dies out with probability 1 If m > 1 then d < 1 and the process dies out with probability d 2 We shall often want to know the probability that a branching process dies out by a particular generation, as well as the limit of these probabilities Let dn be 382 CHAPTER 10 GENERATING FUNCTIONS y 1 y=z y = h(z) p 0 d0 = 0 d1 d2 d 3 z d 1 Figure 10. 3: Geometric determination of d the probability of dying... gives 50 dollars to the person from whom the letter was purchased and then sends 50 dollars to the person whose name is at the top of the list The buyer then crosses off the name at the top of the list and adds her own name at the bottom in each letter before it is sold again Let us first assume that the buyer may sell the letter only to a single person If you buy the letter you will want to compute your... cz (1 − cz)2 386 CHAPTER 10 GENERATING FUNCTIONS and m = h (1) = b (1 − c)2 We know that if m ≤ 1 the process will surely die out and d = 1 To find the probability d when m > 1 we must find a root d < 1 of the equation z = h(z) , or bz b + 1 − c 1 − cz This leads us to a quadratic equation We know that z = 1 is one solution The other is found to be 1−b−c d= c(1 − c) z =1− It is easy to verify that d... us re-examine the Keyfitz data to see if a distribution of the type considered in Example 10. 11 could reasonably be used as a model for this population We would have to estimate from the data the parameters b and c for the formula pk = bck−1 Recall that m= b (1 − c)2 (10. 7) and the probability d that the process dies out is d= 1−b−c c(1 − c) Solving Equation 10. 7 and 10. 8 for b and c gives c= 7 T... mistake He assumed that 1 was the only solution to Equation 10. 3 To examine this question more carefully, we first note that solutions to Equation 10. 3 represent intersections of the graphs of y=z and y = h(z) = p0 + p1 z + p2 z 2 + · · · Thus we need to study the graph of y = h(z) We note that h(0) = p0 Also, h (z) = p1 + 2p2 z + 3p3 z 2 + · · · , (10. 4) and h (z) = 2p2 + 3 · 2p3 z + 4 · 3p4 z 2 . mistake. He assumed that 1 was the only solution to Equation 10. 3. To examine this question more carefully, we first note that solutions to Equation 10. 3 represent intersections of the graphs of y. functions for p and p  . 376 CHAPTER 10. GENERATING FUNCTIONS 6 Let p be the probability distribution p =  01 2 01/32/3  , and let p n = p ∗ p ∗···∗p be the n-fold convolution of p with itself. (a). there is no way to load two dice so that the probability that a given sum will turn up when they are tossed is the same for all sums (i.e., that all outcomes are equally likely). 10 Show that if h(z)= 1