§10 Transformation of random variables

§10.1 General principle

10.1.1 Let X1 , . . . , Xn be n random variables. For each i = 1, . . . , n, let yi = yi (x1 , . . . , xn ) be
a real-valued transformation of (x1 , . . . , xn ) ∈ supp(X1 , . . . , Xn ). Assume that the mapping
(x1 , . . . , xn ) 7→ (y1 , . . . , yn ) is one-to-one between the supports of (X1 , . . . , Xn ) and (Y1 , . . . , Yn ),
where Yi = yi (X1 , . . . , Xn ), i = 1, . . . , n. Thus, we may map (Y1 , . . . , Yn ) to (X1 , . . . , Xn ) by
an inverse transformation Xi = xi (Y1 , . . . , Yn ), i = 1, . . . , n.
To simplify presentation, we use henceforth a matrix notation:
           
X1 Y)
x1 (Y Y1 X)
y1 (X x1 y1
 ..   ..   ..   ..   ..   .. 
Y ) =  .  , Y =  .  = y (X
X =  .  = x (Y X ) =  .  , x =  .  , y =  .  , etc.
Xn Y)
xn (Y Yn X)
yn (X xn yn

x) of X , it is of interest to derive the joint probability

Given the joint probability function fX (x
function fY (yy ) of Y .

x), the corresponding joint mass function of Y is

10.1.2 For discrete X with joint mass function fX (x




fY (yy ) = P(Y
Y = y ) = P x (Y
Y ) = x (yy ) = P X = x (yy ) = fX x (yy ) , y ∈ supp(Y
Y ).

10.1.3 For continuous X with joint pdf fX (x x) and assuming differentiability of y (·) [hence also of
x (·)], the corresponding joint pdf of Y is

fY (yy ) = fX x (yy )  det x 0 (yy ),

 
y ∈ supp(Y
Y ),

where  
∂x1 (yy )/∂y1 · · · ∂x1 (yy )/∂yn
∂x  .. .. ..
x 0 (yy ) = , .

. .
∂yy

∂xn (yy )/∂y1 · · · ∂xn (yy )/∂yn
is the Jacobian matrix of transformation.

§10.2 Transformation of more than one random variables: examples

10.2.1 For independent X1 , X2 with Xi ∼ Poisson (λi ) (i = 1, 2), define Y1 = X1 and Y2 = X1 + X2 .

67
 Here " # " #
x1 y1
x) =
y (x ⇒ x (yy ) =
x1 + x2 y2 − y1
Joint mass function of Y = [Y1 , Y2 ]> :

e−λ1 λy11 e−λ2 λy22 −y1 

fY (yy ) = fX x (yy ) = fX1 (y1 )fX2 (y2 − y1 ) = 1 y1 , y2 − y1 = 0, 1, 2, . . .
y1 ! (y2 − y1 )!
−(λ1 +λ2 ) y2
 
e λ2 y2
(λ1 /λ2 )y1 1 y2 = 0, 1, 2, . . . , y1 = 0, . . . , y2 .

=
y2 ! y1

Conditional mass function of Y1 given Y2 = y2 :

 
y2
(λ1 /λ2 )y1 1 y1 = 0, . . . , y2

fY1 |Y2 (y1 |y2 ) ∝ fY (yy ) ∝
y1
  y1  y2 −y1
y2 λ1 λ2 
⇒ fY1 |Y2 (y1 |y2 ) = 1 y1 = 0, . . . , y2
y1 λ1 + λ2 λ1 + λ2
 
λ1
⇒ Y1 |Y2 = y2 ∼ Binomial y2 , .
λ1 + λ2

Marginal mass function of Y2 :

e−(λ1 +λ2 ) (λ1 + λ2 )y2 

fY2 (y2 ) = fY (yy )/fY1 |Y2 (y1 |y2 ) = 1 y2 = 0, 1, 2, . . .
y2 !
⇒ Y2 ∼ Poisson (λ1 + λ2 ).

10.2.2 Consider independent X1 , X2 with Xi ∼ Gamma (αi , λ) (i = 1, 2), so that

2  αi αi −1 −λxi 
Y λ x ie λα1 +α2 xα1 1 −1 xα2 2 −1 e−λ(x1 +x2 )
fX (x1 , x2 ) = = (x1 , x2 > 0).
i=1
Γ(αi ) Γ(α1 )Γ(α2 )

Define, for a constant c > 0, Y1 = c(X1 + X2 ) and Y2 = X1 /(X1 + X2 ). Here

" # " #
c(x1 + x2 ) y1 y2 /c
x) =
y (x ⇒ x (yy ) = .
x1 /(x1 + x2 ) y1 (1 − y2 )/c

Thus, the Jacobian matrix is

" # " #
1 ∂(y y
1 2 )/∂y 1 ∂(y y
1 2 )/∂y 2 1 y2 y 1
x 0 (yy ) = = ⇒ det x 0 (yy ) = −c−2 y1 .
c ∂y1 (1 − y2 )/∂y1 ∂y1 (1 − y2 )/∂y2 c 1 − y2 −y1

68
 Joint pdf of (Y1 , Y2 ) is

fY (yy ) = fX x (yy )  det x 0 (yy )

 
α −1
λα1 +α2 (y1 y2 /c)α1 −1 y1 (1 − y2 )/c 2 e−(λ/c) y1 y1
  

= 2
1 y1 > 0, y2 ∈ [0, 1]
Γ(α1 )Γ(α2 ) c
α1 +α2 α1 +α2 −1 −(λ/c) y1
(λ/c) y1 e Γ(α1 + α2 ) α1 −1
y2 (1 − y2 )α2 −1 1 y2 ∈ [0, 1] .

= 1 {y1 > 0} ×
Γ(α1 + α2 ) Γ(α1 )Γ(α2 )
It follows that Y1 , Y2 are independent with


Y1 ∼ Gamma α1 + α2 , λ/c , Y2 ∼ Beta (α1 , α2 ).

10.2.3 Given joint pdf of X = (X1 , X2 , X3 ):

fX (x1 , x2 , x3 ) = 40 x1 x3 1 {x1 , x2 , x3 ≥ 0, x1 + x3 ≤ 1, x2 + x3 ≤ 1}.

Define Y1 = X1 /(1 − X3 ), Y2 = X2 /(1 − X3 ) and Y3 = 1 − X3 .

Here    
x1 /(1 − x3 ) y1 y3
x) = x2 /(1 − x3 ) ⇒ x (yy ) =  y2 y3  .
y (x
   

1 − x3 1 − y3
Thus, the Jacobian matrix is
   
∂(y1 y3 )/∂y1 ∂(y1 y3 )/∂y2 ∂(y1 y3 )/∂y3 y3 0 y1
x 0 (yy ) =  ∂(y2 y3 )/∂y1 ∂(y2 y3 )/∂y2 ∂(y2 y3 )/∂y3  =  0 y3 y2 
   

∂(1 − y3 )/∂y1 ∂(1 − y3 )/∂y2 ∂(1 − y3 )/∂y3 0 0 −1

⇒ det x 0 (yy ) = −y32 .

Joint pdf of (Y1 , Y2 , Y3 ) is

fY (yy ) = fX x (yy )  det x 0 (yy )

 

= 40 y1 y33 (1 − y3 ) 1 y1 , y2 , y3 ∈ [0, 1]


= 2y1 1 {0 ≤ y1 ≤ 1} 1 {0 ≤ y2 ≤ 1} 20y33 (1 − y3 ) 1 {0 ≤ y3 ≤ 1} .



It follows that Y1 , Y2 , Y3 are independent with marginal pdf’s

fY1 (u) = 2u 1 {0 ≤ u ≤ 1}, fY2 (u) = 1 {0 ≤ u ≤ 1}, fY3 (u) = 20u3 (1 − u) 1 {0 ≤ u ≤ 1}.

In particular, Y2 ∼ U [0, 1].

69
 §10.3 *** More challenges ***
10.3.1 Let (X, Y ) be a random coordinate pair uniformly distributed over the quadrangle with vertices
(0, 0), (a, 0), (a, 1) and (2a, 1), where a > 0. Find the means and variances of X and Y and their
correlation.

10.3.2 Suppose that U has a uniform distribution over [0, 1], V has an exponential distribution of unit
rate, and that U and V are independent.

(a) Find the conditional density function and expectation of U + V given that U = V .
(b) Find the conditional density function and expectation of U + V given that U ≤ V .

70