EE5110: Probability Foundations for Electrical Engineers July-November 2015

Lecture 26: Characteristic Functions

Lecturer: Dr. Krishna Jagannathan Scribe: Aseem Sharma and Ajay M.

The characteristic function of a random variable X is defined as

CX (t) =E[eitX ]
=E[cos(tX)] + iE[sin(tX)],

which can also be written as Z

CX (t) = eitx dPX .

If X is a continuous random variable with density function fX (x), then

Z
CX (t) = eitx fX (x) dx.

The advantage with the characteristic function is that it always exists, unlike the moment generating function,
which can be infinite everywhere except s = 0.
Example 1: Let X be an exponential random variable with parameter . Find its characteristic function.
Solution: Recall that for an exponential random variable with parameter , fX (x) = ex . Thus, we have
Z
CX (t) = ex eitx dx
x=0

= .
it
We have evaluated the above integral essentially by pretending that it is a real number. Although this
happens to produce the correct answer in this case, the correct method of evaluating a characteristic function
is by performing contour integration. Indeed, in the next example, it is not possible to obtain the correct
answer by pretending that it is a real number (which is not).
Example 2: Let X be a Cauchy random variable. Find its characteristic function.
Solution: The density function for a Cauchy random variable is
1
fX (x) = .
(1 + x2 )
Therefore,
Z
eitx
CX (t) = dx
(1 + x2 )
x=

=e|t| .

The above expression is not entirely trivial to obtain. Indeed, it requires considering two separate contour
integrals for t > 0 and t < 0, and invoking Cauchys residue theorem to evaluate the contour integrals.
(For details, see http://www.wpressutexas.net/forum/attachment.php?attac hmentid=408&d=1296667390.)

26-1
26-2 Lecture 26: Characteristic Functions

However, it is also possible to obtain the characteristic function of the Cauchy random variable by invoking
a Fourier transform duality trick from your undergraduate signals and systems course. (Do it!)
Recall also that the moment generating function of a Cauchy random variable does not converge anywhere
except at s = 0. On the other hand, we find here that the characteristic function for the Cauchy random
variable exists everywhere. This is essentially because the integral defining the chracteristic function con-
verges absolutely, and hence uniformly, for all t R. Characteristic functions are thus particularly useful in
handling heavy-tailed random variables, for which the corresponding moment generating functions do not
exist.
Let us next discuss some properties of characteristic functions.

26.1 Properties of characteristic functions

26.1.1 Elementary properties

1) If Y = aX + b, CY (t) = eibt CX (at).

2) If X and Y are independent random variables and Z = X + Y , then CZ (t) = CX (t)CY (t).

3) If MX (s) < for s [, ], then CX (t) = MX (it) for all t R.

Example 3: Let X N (0, 1). The moment generating function is

s2
MX (s) = e 2 .

Then, the characteristic function is

t2
CX (t) = MX (it) = e 2 .
2
For a non-standard
 Gaussian, Y N (, ), we can now invoke property 1) and conclude that CY (t) =
2 t2
exp it 2 .

26.1.2 Defining properties

Theorem 26.1 A characteristic function CX (t) satisfies the following properties:

1) CX (0) = 1 and |CX (t) | 1, t R.

2) CX (t) is uniformly continuous on R, i.e., t R, a (h) 0 as h 0 such that

|CX (t + h) CX (t)| (h).

3) CX (t) is a non-negative definite kernel, i.e., for any n, any real t1 , t2 , . . . , tn , and any complex
z1 , z2 , . . . , zn , we have X
zj CX (tj tk )zk 0.
j,k

Proof:
Lecture 26: Characteristic Functions 26-3

1)
Z Z
|CX (t)| = | eitx dPX | |eitx |dPX = 1.

2)

|E[ei(t+h)X ] E[eitX ]| = |E[eitX (eihX 1)]|

E[|eihX 1|].

Let |eihX 1| = y(h) and E[y(h)] = (h). We now need to show that (h) 0 as h 0. Note that
y(h) 0 as h 0. Further,

y(h) = |eihX 1|
q
2 2
= (cos (hX) 1) + (sin (hX))
p
= 2 2 cos (hX)
 
hX
= 2 sin
2
2.

Since y(h) is bounded above by 2, applying DCT, we thus have (h) 0 as h 0.

3)

X XZ
zj CX (tj tk )zk = zj ei(tj tk )X zk dPX
j,k j,k
XZ
= zj eitj X (zk eitk X ) dPX
j,k
X
= E[ zj eitj X (zk eitk X )]
j,k
X
E[ |zj eitj X |2 ]
j

0.

The significance of 3) may not be apparent at a first glance. However, these three properties are considered
as the defining properties of a characteristic function, because these properties are also sufficient for an
arbitrary function to be the characteristic function of some random variable. This important result is known
as Bochners theorem, which is beyond our scope.

Theorem 26.2 (Bochners theorem) A function C() is a characteristic function of a random variable
if and only if it satisfies the properties of theorem 26.1.
26-4 Lecture 26: Characteristic Functions

26.2 Inversion Theorems

The following inverse theorems are presented without proof, since the proofs require some sophisticated
machinery from harmonic analysis and complex variables. Essentially, they state that the CDF of a random
variable can be recovered from the characteristic function.

Theorem 26.3

(i) Let X be a continuous random variable, having a probability density function fX (x) and the corre-
sponding characteristic function be
Z
CX (t) = eitx fX (x)dx. (26.1)

The probability density function, fX (x) can be obtained from the characteristic function as
T
1
Z
fX (x) = lim eitx CX (t)dt, (26.2)
2 T T

at every point where fX (x) is differentiable.

(ii) The sufficient (but not necessary) condition for the existence of a probability density function is that
the characteristic function should be absolutely integrable, i.e.,
Z
|CX (t)|dt < . (26.3)

(iii) Let CX (t) be a valid characteristic function of a random variable X with a cumulative distribution
function FX (x). We define,
 
1
FX (x) = FX (x) + lim FX (y) for some y, (26.4)
2 yx

then

1 eiat eibt
Z
FX (b) FX (a) = lim CX (t)dt a and b. (26.5)
T 2 T

In part (iii) above, the function FX (x) coincides with the CDF FX (x) at all points where the CDF is con-
tinuous. At points of discontinuity, it is easy to see that FX (x) takes the value at the mid-point of the
right and left limits of the CDF. Equation (26.5) says that the function FX (x) can be recovered from the
characteristic function. Finally, since the CDF is right-continuous, we can recover FX (x) from FX (x).

26.3 Moments from the Characteristic Function

Theorem 26.4

dk CX (t)
(i) Let X be a random variable having a characteristic function CX (t). If dtk exists at t = 0, then
Lecture 26: Characteristic Functions 26-5

(a) E[|X k |] < when k is even.

(b) E[|X k 1|] < when k is odd.

(ii) If E[|X k |] < , then

dk CX (t) 

k k
i E[X ] = . (26.6)
dtk t=0
Further,
k
E Xj
 
j
X
(it) + O tk ,


CX (t) = (26.7)
j=0
j!

where the error, O tk means that O tk / tk 0 as t 0.

eitx dPX converges uniformly, we are justified in taking the derivative inside the
R
Note: Since CX (t) =
integral.

26.4 Exercise:
1. [Papoulis] Use characteristic function definition to find the distribution of Y = aX 2 , if X is Gaussian
with zero mean and variance 2 .
2. [Papoulis] Use characteristic function definition to find the distribution of Y = sin(X), if X is uniformly
distributed in (/2, /2).