2.2 The Gamma Distribution In this section we look at some of the basic properties of gamma random variables; see Hogg and Tanis [6]. A random variable X is said to have a gamma distribution with parameters m > 0 and > 0 if its probability density function has the form f(t) = f(t; m,) = (1) tm-1e-t/ m (m) 0 if t > 0 if t < 0 In this case we shall say X is a gamma random variable with parameters m and and write X ~ (m,). Sometimes m is called the shape parameter and the scale parameter. In general, m might not be an integer. Gamma random variables are used to model a number of physical quantities. Some examples are 1. The time it takes for something to occur, e.g. a lifetime or the service time in a queue. 2. The rate at which some physical quantity is accumulating during a certain period of time, e.g. the excess water flowing into a dam during a certain period of time due to rain or the amount of grain harvested during a certain season. Sometimes it is convenient to use = 1/ as a parameter instead of . The pdf then has the form (2) f(t) = f(t; m,) = mtm-1e-t (m) for t > 0. We shall also write X ~ (m,1/) in this case. It should be clear from the context whether f(t; m,) or f(t; m,) stands for (1) or (2). 2.2 - 1 Proposition 1. If m > 0 and > 0 then m-1 -t/ t m e dt (m) (3) = 1 0 which confirms that f(t) defined by (1) is a valid density function. 0 0 m-1 -t/ um-1e-u t e m-1 -u Proof. m dt = du = 1 since (m) = (m) u e du. (m) 0 Proposition 2. If X has a gamma distribution with parameters m and , then the mean of X is m-1 -t/ t e X = E(X) = t m (m) dt = m (4) 0 Proof. t m -t/ m-1 -t/ t e t m e dt = m (m) dt (m) 0 0 tme-t/ = m m+1 dt = m f(t;m+1,) dt = m (m+1) 0 0 where we have used (m+1) = m (m) and (3). Proposition 3. If X has a gamma distribution with parameters m and , then the expected value of X2 is (5) m-1 -t/ t e E(X ) = t2 m (m) dt 2 = m(m+1)2 0 The variance of X is (6) (X)2 = E((X - X)2) = m2 The standard deviation of X is 2.2 - 2 X = (7) m tm+1e-t/ t e 2 t e 2 Proof. t dt = m dt = m(m+1) m+2 dt = m (m) (m) (m+2) m-1 -t/ 0 m+1 -t/ 0 0 2 m(m+1) f(t;m+2,) dt = m(m+1) where we have used (m+2) = m(m+1) (m) and 2 0 (3). This proves (5). Since E((X - X)2) = E(X2) - X2 the formula (6) follows from (4) and (5). (7) follows from (6). Proposition 4. If f(t) is given by (1) then for t > 0 one has [(m - 1) - t] tm-2e-t/ m+1 (m) (8) f '(t) = (9) f(t) has a single local maximum at t = (m - 1) if m > 1. (10) f(t) is strictly decreasing for t > 0 if m 1 Proof. (8) is a straightforward computation and (9) and (10) follow from (8). Proposition 5. Assume X has a gamma distribution with parameters m and and let Y = cX for some positive number c. Then Y has a gamma distribution with parameters m and c. Proof. If f(t) given by (1) is the density function of X then the density function of Y is tm-1e-t/(c) (c) (m) 0 m (1/c)f(t/c) = if t > 0 if t < 0 which is equal to f(t; m,c). 2.2 - 3 Proposition 5. If X and Y are independent gamma random variables and X has parameters m and and Y has parameters q and , then X + Y is a gamma random variable with parameters m + q and . Proof. We first show that 1 (12) um-1(1-u)q-1 du = B(m, q) 0 where (12) B(m, q) = /2 (m) (q) 2m-1 2q-1 = 2 cos (t) sin (t) dt (m +q ) 0 0 0 m-1 q-1 -(t+s) is the beta function. To see this first note that (m) (q) = dsdt. Make t s e the change of variables r = t + s and u = t/(t+s). Then t = ru and s = r(1-u) and dsdt = rdudr and the first quadrant in the st-plane gets mapped into the strip {(r,u): 0 < r < , 0 1 (ru) < u < 1}. So (m) (q) = m-1 0 1 um-1(1-u)q-1 du (r(1-u)) e rdudr = (m +q ) q-1 -r 0 0 and (12) follows. Next we show that (13) m-1 q-1 t t * (m) (q) = tm+q-1 (m +q) t To see this note that t m-1 *t q-1 sm-1(t-s)q-1 ds. Make the change of variables s = tu. = 0 1 1 m-1 q-1 m+q-1 m-1 q-1 m+q-1 (m) (q) We get tm-1 * tq-1 = (tu) (t-tu) tdu = t u (1-u) du = t (m +q ) 0 0 and (13) follows. It follows from (11) and (13) that tm+q-1e-t/ tm-1e-t/ tq-1e-t/ f(t; m,) * f(t; q,) = m = f(t; m+q,) * q = m+q (m +q) (m) (q) and the propostion follows. 2.2 - 4 Proposition 6. If X has a gamma distribution with parameters m and = 1/, then the Laplace transform L(s) and moment generating function M(r) of X are given by (14) L(s) = (15) M(r) = 1 = (1 + s)m ( + s)m 1 (1 - r) m = ( - s)m Proof. One has 0 0 -st -m m-1 -t/ -1 -m m-1 -(s+1/)t L(s) = [ (m)]-1 dt e t e dt = [ (m)] t e If one makes the change of variables u = (s + 1/)t one obtains -m(u/(s + 1/))m-1e-u (1/(s + 1/))du L(s) = [ (m)] -1 0 1 um-1e-u du = = (1/(1 + s)) [ (m)] (1 + s)m m -1 0 This proves (14). (15) follows from the (14) and the fact that M(r) = L(-r). Let t t (16) m-1 -s/ s e f(s;m,) ds = m G(t) = G(t;m,) = ds (m) 0 0 be the cummulative distribution function of the gamma random variable X ~ (m,) and let (17) m-1 -s/ s e H(t) = H(t;m,) = 1 - G(t) = m (m) ds t be the complementary distribution function (or survival function). Let (18) m-1 -s m(t) = s e ds t be the upper incomplete gamma function and t (19) m-1 -s m(t) = s e ds 0 be the lower incomplete gamma function. 2.2 - 5 Proposition 7. (t/) (20) G(t;m,) = m (m) (21) H(t;m,) = m(t/) (m) Proof. (20) follows by making the change of variables u = s/ in (16) and (21) follows by making the change of variables u = s/ in (17). 2.2 - 6