9.

Two Functions of Two Random

Variables

In the spirit of the previous lecture, let us look at an

immediate generalization: Suppose X and Y are two random
variables with joint p.d.f f XY ( x, y). Given two functions g ( x, y )
and h( x, y ), define the new random variables
Z = g ( X ,Y ) (9-1)
W = h ( X , Y ). (9-2)
How does one determine their joint p.d.f fZW(z, w) ? Obviously
with fZW(z, w) in hand, the marginal p.d.fs fZ (z) and fW (w)
can be easily determined.

1
PILLAI
The procedure is the same as that in (8-3). In fact for given z
and w,
FZW ( z , w ) = P (Z (ξ ) ≤ z , W (ξ ) ≤ w ) = P (g ( X , Y ) ≤ z , h ( X , Y ) ≤ w )
= P (( X , Y ) ∈ D z , w ) = ∫∫ f XY ( x , y ) dxdy , (9-3)
( x , y )∈D z , w

where Dz ,w is the region in the xy plane such that the

inequalities g( x, y) ≤ z and h( x, y ) ≤ w are simultaneously
satisfied.
We illustrate this technique in the next example.
y

D z ,w
D z ,w
x
2
Fig. 9.1 PILLAI
Example 9.1: Suppose X and Y are independent uniformly
distributed random variables in the interval ( 0 , θ ).
Define Z = min( X , Y ), W = max( X , Y ). Determine f ZW ( z , w ).
Solution: Obviously both w and z vary in the interval ( 0 , θ ).
Thus FZW ( z, w) = 0, if z < 0 or w < 0. (9-4)

FZW ( z , w ) = P (Z ≤ z ,W ≤ w ) = P (min( X , Y ) ≤ z , max( X , Y ) ≤ w ) . (9-5)

We must consider two cases: w ≥ z and w < z, since they

give rise to different regions for Dz ,w (see Figs. 9.2 (a)-(b)).
Y Y

( w, w) ( z, z )

( z, z ) y=w ( w, w)

X X

(a ) w ≥ z (b) w < z 3
Fig. 9.2 PILLAI
For w ≥ z , from Fig. 9.2 (a), the region Dz , w is represented
by the doubly shaded area. Thus
FZW ( z , w ) = F XY ( z , w ) + F XY ( w , z ) − F XY ( z , z ) , w ≥ z, (9-6)
and for w < z , from Fig. 9.2 (b), we obtain
FZW ( z , w ) = F XY ( w , w ) , w < z. (9-7)
With
x y xy (9-8)
F XY ( x , y ) = F X ( x ) FY ( y ) = ⋅ = 2 ,
θ θ θ
we obtain
 (2 w − z ) z / θ 2 , 0 < z < w <θ,
FZW ( z, w) =  (9-9)
 w2 /θ 2 , 0 < w < z <θ.
Thus
2 / θ 2 , 0 < z < w <θ,
f ZW ( z, w) =  (9-10)
 0, otherwise . 4
PILLAI
From (9-10), we also obtain
θ 2 z
fZ ( z) = ∫ z
f ZW ( z , w ) dw =  1 − ,
θ θ
0 < z <θ, (9-11)

and
w 2w (9-12)
fW ( w ) = ∫ 0
f ZW ( z , w ) dz =
θ 2
, 0 < w <θ.

If g ( x , y ) and h ( x , y ) are continuous and differentiable

functions, then as in the case of one random variable (see (5-
30)) it is possible to develop a formula to obtain the joint
p.d.f f ZW ( z, w) directly. Towards this, consider the equations
g ( x, y ) = z, h ( x, y ) = w. (9-13)

For a given point (z,w), equation (9-13) can have many

solutions. Let us say
( x1 , y1 ), ( x2 , y2 ), ", ( xn , yn ), 5
PILLAI
represent these multiple solutions such that (see Fig. 9.3)
g ( xi , yi ) = z, h( xi , yi ) = w. (9-14)

y ∆2
w ∆1
•
( x2 , y2 )
w + ∆w •
( x1, y1 ) ∆i
∆w •
( z, w) • ( xi , yi )
z + ∆z x
∆z
•
∆n
z ( xn , yn )
(a) (b)
Fig. 9.3

Consider the problem of evaluating the probability

P(z < Z ≤ z + ∆z, w < W ≤ w + ∆w)
= P(z < g ( X , Y ) ≤ z + ∆z, w < h( X , Y ) ≤ w + ∆w) . (9-15)
6
PILLAI
Using (7-9) we can rewrite (9-15) as
P(z < Z ≤ z + ∆z, w < W ≤ w + ∆w) = f ZW ( z, w)∆z∆w. (9-16)
But to translate this probability in terms of f XY ( x, y ), we need
to evaluate the equivalent region for ∆z ∆w in the xy plane.
Towards this referring to Fig. 9.4, we observe that the point
A with coordinates (z,w) gets mapped onto the point A′ with
coordinates ( xi , yi ) (as well as to other points as in Fig. 9.3(b)).
As z changes to z + ∆z to point B in Fig. 9.4 (a), let B ′
represent its image in the xy plane. Similarly as w changes
to w + ∆w to C, let C ′ represent its image in the xy plane.
w y
D′
C D C′
∆w
∆i
w yi
θ ϕ B′
A B A′
∆z
z z xi x
7
(a) Fig. 9.4 (b) PILLAI
Finally D goes to D ′, and A′B ′C ′D ′ represents the equivalent
parallelogram in the XY plane with area ∆i . Referring back
to Fig. 9.3, the probability in (9-16) can be alternatively
expressed as
∑ P(( X , Y ) ∈ ∆ ) =∑ f
i
i
i
XY ( xi , yi ) ∆ i . (9-17)

Equating (9-16) and (9-17) we obtain

∆i
f ZW ( z, w) = ∑ f XY ( xi , yi ) . (9-18)
i ∆z∆w
To simplify (9-18), we need to evaluate the area ∆i of the
parallelograms in Fig. 9.3 (b) in terms of ∆z∆w. Towards
this, let g1 and h1 denote the inverse transformation in (9-14),
so that
xi = g1 ( z, w), yi = h1 ( z, w). (9-19)
8
PILLAI
As the point (z,w) goes to ( xi , yi ) ≡ A′, the point ( z + ∆z, w) → B′,
the point ( z, w + ∆w) → C′, and the point ( z + ∆z, w + ∆w) → D′.
Hence the respective x and y coordinates of B ′ are given by
∂g1 ∂g
g1 ( z + ∆z , w) = g1 ( z , w) + ∆ z = xi + 1 ∆ z , (9-20)
∂z ∂z
and
∂h1 ∂h1 (9-21)
h1 ( z + ∆z , w) = h1 ( z , w) + ∆ z = yi + ∆ z.
∂z ∂z
Similarly those of C ′ are given by
∂g1 ∂h1 (9-22)
xi + ∆ w, yi + ∆w.
∂w ∂w
The area of the parallelogram A′B ′C ′D ′ in Fig. 9.4 (b) is
given by
∆ i = ( A′B ′)( A′C ′) sin( θ − ϕ )
(9-23)
= ( A′B ′ cos ϕ )( A′C ′ sin θ ) − ( A′B ′ sin ϕ )( A′C ′ cos θ ) . 9
PILLAI
But from Fig. 9.4 (b), and (9-20) - (9-22)
∂g1 ∂h (9-24)
A′B ′ cos ϕ = ∆z , A′C ′ sin θ = 1 ∆w,
∂z ∂w
∂h ∂g
A′B ′ sin ϕ = 1 ∆z , A′C ′ cos θ = 1 ∆w. (9-25)
∂z ∂w
so that
 ∂ g 1 ∂ h1 ∂ g 1 ∂ h1 
∆i =  −  ∆ z∆ w (9-26)
 ∂z ∂w ∂w ∂z 
and
 ∂g1 ∂g1 
 
∆i  ∂ g 1 ∂ h1 ∂ g 1 ∂ h1   ∂z ∂w 
=  −  = det   (9-27)
∆ z∆ w  ∂z ∂w ∂w ∂z  ∂ h1 ∂ h1 

 
 ∂z ∂w 
The right side of (9-27) represents the Jacobian J ( z, w) of
the transformation in (9-19). Thus
10
PILLAI
  ∂g1 ∂g1 
 ∂z ∂w 
 
J ( z, w ) = det  .

(9-28)
∂ h1 ∂ h1 
 
 ∂z ∂w 

Substituting (9-27) - (9-28) into (9-18), we get

1
f ZW ( z, w) = ∑| J ( z, w) | f XY ( xi , yi ) = ∑ f XY ( xi , yi ), (9-29)
i i | J ( xi , yi ) |

since
1
| J ( z, w) | = (9-30)
| J ( xi , yi ) |

where J ( xi , yi ) represents the Jacobian of the original

transformation in (9-13) given by
 ∂g ∂g 
 
 ∂x ∂y 
J ( x i , y i ) = det   . (9-31)
 ∂h ∂h 
 ∂x ∂ y  x = x
11
 i , y = yi PILLAI
Next we shall illustrate the usefulness of the formula in
(9-29) through various examples:
Example 9.2: Suppose X and Y are zero mean independent
Gaussian r.vs with common variance σ 2 .
Define Z = X 2 + Y 2 , W = tan−1(Y / X ), where | w |≤ π / 2.
Obtain f ZW ( z, w).
Solution: Here
1 −( x 2 + y 2 ) / 2σ 2
(9-32)
f XY ( x, y ) = e .
2πσ 2

Since
z = g ( x, y ) = x 2 + y 2 ; w = h( x, y ) = tan −1 ( y / x ), | w |≤ π / 2, (9-33)

if ( x1, y1 ) is a solution pair so is (− x1,− y1 ). From (9-33)

y
= tan w, or y = x tan w. (9-34)
x
12
PILLAI
Substituting this into z, we get

z = x2 + y2 = x 1 + tan2 w = x secw, or x = z cosw. (9-35)

and
(9-36)
y = x tan w = z sin w.

Thus there are two solution sets

x1 = z cos w, y1 = z sin w, x2 = − z cos w, y 2 = − z sin w. (9-37)
We can use (9-35) - (9-37) to obtain J ( z, w). From (9-28)
∂x ∂x
∂z ∂w cos w − z sin w
J (z, w ) = =
sin w z cos w
= z, (9-38)
∂y ∂y
∂z ∂w
so that
| J ( z, w) |= z. (9-39) 13
PILLAI
We can also compute J ( x, y) using (9-31). From (9-33),
x y
x 2
+ y 2
x 2
+ y 2
1 1
J (x, y) = =
x 2
+ y 2
=
z
. (9-40)
− y x
x2 + y 2
x 2
+ y 2

Notice that | J ( z, w) |= 1/ | J ( xi , yi ) |, agreeing with (9-30).

Substituting (9-37) and (9-39) or (9-40) into (9-29), we get
f ZW ( z, w) = z ( f XY ( x1 , y1 ) + f XY ( x2 , y2 ) )
z − z 2 / 2σ 2 π
= e , 0 < z < ∞, | w |< . (9-41)
πσ 2
2
Thus
π /2 z − z 2 / 2σ 2
fZ ( z) = ∫ −π / 2
f ZW ( z , w ) dw = 2 e
σ
, 0 < z < ∞, (9-42)

which represents a Rayleigh r.v with parameter σ 2, and

∞ 1 π
fW ( w ) = ∫ 0
f ZW ( z , w ) dz = , | w |< ,
π 2
(9-43) 14
PILLAI
which represents a uniform r.v in the interval (−π / 2,π / 2).
Moreover by direct computation
f ZW ( z , w ) = f Z ( z ) ⋅ fW ( w ) (9-44)

implying that Z and W are independent. We summarize these

results in the following statement: If X and Y are zero mean
independent Gaussian random variables with common
variance, then X 2 + Y 2 has a Rayleigh distribution and tan−1(Y / X)
has a uniform distribution. Moreover these two derived r.vs
are statistically independent. Alternatively, with X and Y as
independent zero mean r.vs as in (9-32), X + jY represents a
complex Gaussian r.v. But
X + jY = Ze jW , (9-45)

where Z and W are as in (9-33), except that for (9-45) to hold

good on the entire complex plane we must have − π < W < 15π ,
and hence it follows that the magnitude and phase of PILLAI
a complex Gaussian r.v are independent with Rayleigh
and uniform distributions (U ~ ( −π , π ) ) respectively. The
statistical independence of these derived r.vs is an interesting
observation.
Example 9.3: Let X and Y be independent exponential
random variables with common parameter λ.
Define U = X + Y, V = X - Y. Find the joint and marginal
p.d.f of U and V.
Solution: It is given that
1 −( x + y ) / λ
f XY ( x, y ) = e , x > 0, y > 0. (9-46)
λ2

Now since u = x + y, v = x - y, always | v |< u , and there is

only one solution given by
u+v u−v
x= , y= . (9-47)
2 2
Moreover the Jacobian of the transformation is given by 16
PILLAI
 1 1
J ( x, y ) = = −2
1 −1
and hence
1 −u / λ (9-48)
fUV (u, v ) = 2 e , 0 < | v | < u < ∞,
2λ
represents the joint p.d.f of U and V. This gives
u 1 u u −u / λ
fU ( u ) = ∫ fUV (u, v )dv = 2 ∫ e −u / λ dv = e , 0 < u < ∞, (9-49)
−u 2λ −u λ2

and
∞ 1 ∞ 1 −|v|/ λ
fV ( v ) = ∫ fUV (u, v )du = 2 ∫ e − u / λ du = e , − ∞ < v < ∞. (9-50)
|v| 2λ |v| 2λ
Notice that in this case the r.vs U and V are not independent.
As we show below, the general transformation formula in
(9-29) making use of two functions can be made useful even
when only one function is specified. 17
PILLAI
Auxiliary Variables:
Suppose
Z = g ( X , Y ), (9-51)

where X and Y are two random variables. To determine f Z (z )

by making use of the above formulation in (9-29), we can
define an auxiliary variable

W = X or W =Y (9-52)

and the p.d.f of Z can be obtained from f ZW ( z, w) by proper

integration.
Example 9.4: Suppose Z = X + Y and let W = Y so that the
transformation is one-to-one and the solution is given
by y1 = w, x1 = z − w. 18
PILLAI
The Jacobian of the transformation is given by
1 1
J ( x, y ) = =1
0 1
and hence
f ZW ( x, y ) = f XY ( x1 , y1 ) = f XY ( z − w, w)

or
+∞
f Z ( z ) = ∫ f ZW ( z, w)dw = ∫ f XY ( z − w, w)dw, (9-53)
−∞

which agrees with (8.7). Note that (9-53) reduces to the

convolution of f X (z ) and fY (z ) if X and Y are independent
random variables. Next, we consider a less trivial example.
Example 9.5: Let X ∼ U (0,1) and Y ∼ U (0,1) be independent.
Define Z = (− 2 ln X )1 / 2 cos( 2π Y ). (9-54)
19
PILLAI
Find the density function of Z.
Solution: We can make use of the auxiliary variable W = Y
in this case. This gives the only solution to be
− ( z sec( 2πw ) ) 2 / 2
x1 = e , (9-55)
y1 = w, (9-56)

and using (9-28)

∂ x1 ∂ x1 − ( z sec( 2 π w ) ) 2 / 2 ∂ x1
− z sec ( 2 π w ) e
2

∂z ∂w ∂w
J ( z, w ) = = (9-57)
∂ y1 ∂ y1 0 1
∂z ∂w
− ( z sec( 2 π w ) ) 2 / 2
= − z sec ( 2 π w ) e
2
.

Substituting (9-55) - (9-57) into (9-29), we obtain

− ( z sec( 2 π w ) )
2
/2
f ZW ( z , w ) = z sec ( 2 π w ) e
2
,
(9-58)
− ∞ < z < +∞ , 0 < w < 1, 20
PILLAI
and
− ( z tan( 2 π w ) ) / 2
1 1 2

∫ ∫
− z2 / 2
fZ (z) = f ZW ( z , w ) dw = e z sec (2π w ) e
2
dw. (9-59)
0 0

Let u = z tan(2π w ) so that du = 2π z sec 2 (2π w ) dw. Notice

that as w varies from 0 to 1, u varies from − ∞ to + ∞.
Using this in (9-59), we get
1 +∞ du 1
∫
− z2 / 2 −u2 / 2 − z2 / 2
fZ (z) = e e = e , − ∞ < z < ∞, (9-60)
2π − ∞
    π
2
2π
1

which represents a zero mean Gaussian r.v with unit

variance. Thus Z ∼ N ( 0 ,1 ). Equation (9-54) can be used as
a practical procedure to generate Gaussian random variables
from two independent uniformly distributed random
sequences. 21
PILLAI
Example 9.6 : Let X and Y be independent identically distributed
Geometric random variables with
P ( X = k ) = P (Y = K ) = pq k , k = 0, 1, 2,".
(a) Show that min (X , Y ) and X – Y are independent random variables.
(b) Show that min (X , Y ) and max (X , Y ) – min (X , Y ) are also
independent random variables.
Solution: (a) Let
Z = min (X , Y ) , and W = X – Y. (9-61)
Note that Z takes only nonnegative values {0, 1, 2, "}, while W takes
both positive, zero and negative values {0, ± 1, ± 2, "}. We have
P(Z = m, W = n) = P{min (X , Y ) = m, X – Y = n}. But
Y X ≥ Y ⇒ W = X − Y is nonnegative
Z = min( X , Y ) = 
 X X < Y ⇒ W = X − Y is negative.
Thus
P ( Z = m,W = n ) = P{min( X , Y ) = m, X − Y = n, ( X ≥ Y ∪ X < Y )}
= P (min( X , Y ) = m, X − Y = n, X ≥ Y )
22
+ P(min( X , Y ) = m, X − Y = n, X < Y ) (9-62) PILLAI
P ( Z = m,W = n ) = P (Y = m, X = m + n, X ≥ Y )
+ P ( X = m, Y = m − n , X < Y )
 P ( X = m + n ) P(Y = m ) = pq m + n pq m , m ≥ 0, n ≥ 0
= m−n
 P ( X = m ) P (Y = m − n ) = pq m
pq , m ≥ 0, n < 0
= p 2 q 2 m +|n| , m = 0, 1, 2," n = 0, ± 1, ± 2," (9-63)
represents the joint probability mass function of the random variables
Z and W. Also
P ( Z = m ) = ∑ P( Z = m,W = n ) = ∑ p 2 q 2 m q |n|
n n

= p 2 q 2 m (1 + 2q + 2q 2 + ")
2q
= p 2 q 2 m (1 + 1− q ) = pq 2 m (1 + q)
= p (1 + q) q 2 m , m = 0, 1, 2,". (9-64)

Thus Z represents a Geometric random variable since 23

1 − q 2 = p(1 + q), and PILLAI
 ∞ ∞
P(W = n ) = ∑ P ( Z = m, W = n ) = ∑ p 2 q 2 m q|n|
m =0 m =0

= p 2 q|n| (1 + q 2 + q 4 + ") = p 2 q|n| 1

1−q2
= 1+pq q|n| , n = 0, ± 1, ± 2,". (9-65)

Note that
P ( Z = m,W = n ) = P ( Z = m ) P (W = n ), (9-66)

establishing the independence of the random variables Z and W.

The independence of X – Y and min (X , Y ) when X and Y are
independent Geometric random variables is an interesting observation.
(b) Let
Z = min (X , Y ) , R = max (X , Y ) – min (X , Y ). (9-67)
In this case both Z and R take nonnegative integer values 0, 1, 2, ".
Proceeding as in (9-62)-(9-63) we get
24
PILLAI
P{Z = m, R = n} = P{min( X , Y ) = m, max( X ,Y ) − min( X ,Y ) = n, X ≥ Y }
+ P{min( X , Y ) = m, max( X ,Y ) − min( X ,Y ) = n, X < Y }
= P{Y = m, X = m + n, X ≥ Y ) + P ( X = m, Y = m + n, X < Y }
= P{ X = m + n, Y = m, X ≥ Y ) + P ( X = m, Y = m + n, X < Y }
 pq m + n pq m + pq m pq m + n , m = 0, 1, 2,", n = 1, 2,"
= m+n
 pq pq m
, m = 0, 1, 2,", n=0
2 p 2 q 2 m + n , m = 0, 1, 2,", n = 1, 2,"
=  2 2m
 p q , m = 0, 1, 2,", n = 0. (9-68)

Eq. (9-68) represents the joint probability mass function of Z and R

in (9-67). From (9-68),
∞ ∞
P ( Z = m ) = ∑ P{Z = m, R = n} = p q (1 + 2∑ q n ) = p 2 q 2 m 1 + p
n =0
2 2m

n =1
( 2q
)
= p (1 + q) q 2 m , m = 0, 1, 2," (9-69) 25
PILLAI
and
p
 , n=0
∞
 1+ q
P ( R = n ) = ∑ P{Z = m, R = n} = 
2p n
m =0 1+ q q , n = 1, 2,". (9-70)

From (9-68)-(9-70), we get

P ( Z = m, R = n ) = P ( Z = m ) P ( R = n ) (9-71)
which proves the independence of the random variables Z and R
defined in (9-67) as well.

26
PILLAI