A random variable X has a probability density function if there is a function f : R R so that P(a < X b) =

f (x)dx
a

for any a < b. A random variable with a density is called continuous. We remind the reader that the distribution of a continuous random variable is determined by its density function. The goal of this discussion is to discuss the determination of the distribution of the random variable Y = h(X ), where h is a differentiable function and X is a continuous random variable with density function f . We rst consider the case where the following hold: 1. h : D R is differentiable on its domain D, which is an open interval of R. 2. h is one-to-one, which means that h(x) = h(y) implies that x = y. 3. The support of X , dened as (1) is contained in D. For any set S R, dene h(S) = {y : y = f (x) for some x D}. Theorem 1. Suppose that the conditions above hold. Then Y = h(X ) is a continuous random variable with density function (2) fY (y) = fX (h1 (y)) d 1 h (y) 1 {y h(D)} , dy supp( f ) = {x : f (x) > 0}
def

for all points y so that fX is continuous at h1 (y). Proof. Since h is one-to-one, it is either always increasing or always decreasing on D. Assume that h is increasing, the other case is similar. We begin by computing the distribution function of Y : P(Y y) = P(h(X ) y) = P X h1 (y)
h1 ( y)

= =

d 1 h (z)dz dy y d 1 h (z) dz = fX (h1 (z)) dy fX (h1 (z))

1

fX (x)dx by change of variables since h is increasing .

Now suppose that y is such that h1 (y) is a point of continuity for fX . Then by the Fundamental Theorem of Calculus, FY is differentiable at y with derivative fY (y) = d d 1 FY (y) = fX (h1 (y)) h (y) . dy dy

Now we consider the case where h is not one-to-one. We will assume the following: For each y h(D), the set h1 ({y}) = {x : h(x) = y} is a nite set. We recall the following theorem from calculus: Theorem (Inverse Function Theorem). Let h : D R be a differentiable function. Let y = f (x) for some x D. Suppose that f (x) = 0. Then there is an open interval I containing x and an open interval J containing y, so that h restricted to I is one-to-one, and there is a differentiable inverse 1 h x : J I. Thus, for each x h1 ({y}), there is a function gx dened in a neighborhood of x so that h gi (y ) = y for all y in a neighborhood of y, and gi h(x ) = x for all x in a neighborhood of x. Assume that for each x h1({y}), we have h (x) = 0. Now, since h 1({y}) is nite, say equal to {x1 , . . . , xr }, letting gi = gxi there is an interval J containing y on which each of the gi is dened. We can take J small enough so that {gi (J )} are disjoint intervals. We have for a y b with a < b and a, b J , P(a Y b) = P
x h 1 ( { y } )

{X gx([a, b])} P(X gx ([a, b])) f (u)du

b

= = =
x h

x h 1 ( { y } )

xh1 ({y}) gx ([a,b])

( {y}) a

f (gx (s)) gx(s) ds f (gx (s)) gx(s) ds

a x h 1 ( { y } )

x1

x2

x3

F IGURE 1. A many-to-one function Then taking a = y and b = y + y, differentiating with respect to y, and evaluating at y = 0 yields (3) fY (y) =
x h 1 ( { y } )

f (gx (s)) gx (s) .

Figure 1 shows a many-to-one function. Note how a little neighborhood around y maps to neighborhoods surrounding the three points in h1 ({y}). For this y, the sum in (3) will have three terms. Let us consider an example. Suppose that X has an exponential(1) distribution, and let if 0 < x 1 3x 3 1 8 h(x) = 1 5 x 1 if < x < 3 3 15 2 x 8 8 if x 15 . 15 The reader should graph this function. Let 0 < y < 1. The there are three x so that h(x) = y. Namely, 1 x= y 3 1 4 x = y+ 5 15 1 8 x = y+ . 2 15 The three functions on the right of the above equation are then g1(x), g2 (x) and g3 (x). Thus we have ey/3 ey/54/15 ey/28/15 + + . 3 5 2 Here is another example. Suppose that X has the density x f (x) = 2 1 {0 < x < 2} , 2 fY (y) =

and consider the random variable Y = sin X . We will now nd the density of Y . First take y > 0. Then sin1({y}) (0, 2) = {arcsin(y), arcsin(y) + /2} . This follows since, by convention, arcsin(y) is dened to take values in [ 2 , 2 ] for y [1, 1]. Thus using the notation as above, we have g1 (y) = arcsin(y) , g2 (y) = arcsin(y) + Thus we have fY (y) = arcsin(y) 22 1 1 y2 + arcsin(y) + 2 22 1 1 y2 = arcsin(y) + 4 2 1 y2 . 2

Let us review some facts from multivariate calculus. Let h : D R be a one-to-one function, where D Rn and R Rn . We write h(x1 , . . . , xn ) = (h1 (x1 , . . . , xn), . . . , hn(x1 , . . . , xn )) . The total derivative of h at x = (x1 , . . . , xn) is dened as the matrix h h1 h1 1 (x) (x) (x) x1 x2 xn h2 h2 h2 x (x) x2 (x) xn (x) 1 Dh(x) = . . . . . .. . . . . .
hn x1 (x) hn x2 (x)

hn xn (x)

The Jacobian of h at x, which we will denote by Jh (x) is dened as (4) Jh (x) = det Dh(x) .
def

Now the change of variables formula says the following: Let h : D R be a function which has continuous rst partial derivatives. Then f (y)dy =
E h1 ( E )

f (h(x))|Jh (x)|dx .

Now we can state the formula for nding the density of Y = h(X ), where X is a random vector in Rn , and h is a one-to-one function dened on D Rn , where the support of fX is contained in D.

Theorem 2. Let h be as above, and let X be a continuous random vector in Rn with density fX . Then the density of Y is given by fY (y) = fX (h1 (y))|Jh1 (y)| . A fact which is often very useful is that (5) |Jh1 (y)| = 1 |Jh (h1 (y)| .

Let X , Y be independent standard Normal random variables. Let (D, ) = h(X , Y ) = (X 2 + Y 2 , arctan(Y /X )) . Then We have Dh1(d , ) = Thus 1 1 1 cos2 + sin2 = . 2 2 2 Then we have for d > 0 and [0, 2): |Jh1 (d , )| = 1 1 (d cos2 +d sin2 ) 1 1 d 1 e 2 = e 2 2 2 2 2 This shows that D and are independent, and D is exponential(1/2), and is Uniform[0, 2). [Why?] Note we could run this in reverse: Suppose we start with D an exponential(1/2) random variable, and an independent Uniform[0, 2) random variable. Then let g(d , ) = ( d cos , d sin ) . fD, (d , ) = Then g1 (x, y) = h(x, y), where h is dened as above. Now nding the Jacobian of g1 itself can be done, but it is perhaps easier to use (5): 1 1 = = 2. Jg1 (x, y) = Jh (x, y) = Jh1 (h(x, y)) 1/2 Thus,
2 2 2 2 1 1 1 fX ,Y (x, y) = e(x +y )/2 2 = e(x +y )/2 . 2 2 2 Thus g(D, ) gives a pair of independent standard normal random variables. [Why?]

h1(d , ) = ( d cos , d sin ) .

1 cos 2 d 1 sin 2 d

d sin d cos

This gives a method of simulating a pair of Normal random variable. It is relatively easy to simulate a uniform and an exponential random variable. Then applying the function g to them gives a pair of Normal random variables.