San José State University
Department of Economics

applet-magic.com
Thayer Watkins
Silicon Valley
& Tornado Alley
USA

Infinitely Divisible Random Variables
and Their Characteristic Functions

The analysis of stochastic processes involving variables with Lévy stable distribution has become more common-place. The formula for the characteristic function of such stable distributions is well known but the mathematical analysis involved in its derivation and proof is not so readily available. Such analyses and proofs are available but they are in monographs which are so old that libraries relegate them to their storage facilities. The purpose of this page is to present some of that mathematical analysis.

The theory of stable distribution was worked out with a fairly elaborate mathematical structure. One key element of that structure is the theory of infinitely divisible random variables. A random variable z would be said to be divisible if it could be represented as the sum of two independent random variables with identically distributions; i.e.,

z = z1 + z2

where the probability distributions of z1 and z2 are identical. A variable z is said to be infinitely divisible if all values of n from 1 up it can be represented as the sum of n independent, identically distributed randome variables.

The condition for infinite divisibility can be more easily expressed in terms of characteristic functions. If Φ(ω) is the characteristic function of the distribution of z and Φn(ω) is the characteristic function of the common distribution of the n summands in

z = z1 + z2 + ..... + zn

then

Φ(ω) = [Φn(ω)]n

This means that

Φn(ω) = [Φ(ω)]1/n

The Φn(ω) must be a valid characteristic function; i.e., it must be derived from a valid probabilitity distribution. There are certain properties that a characteristic function must have, such as that Φ(0) = 1 and that it must be continuous but these are not enough to guarantee that a function is the characteristic function of a valid probability distribution, one that has non-negative values for all values of the random variable.

Before showing that there are some characteristic functions that are not infinitely divisible it is perhaps appropriate to show that there are some that are. Consider the normal distributions. The characteristic function for a normal variable is of the form

Φ(ω) = exp(iμω - (σ2/2)ω2)

where μ is the mean value and σ is the standard deviation.

The characteristic function for the identically distributed n summands of a normal variable is then:

Φn(ω) = [Φ(ω)]1/n
= exp(i(μ/n)ω - (σ2/(2n)ω2)
= exp(i(μ/n)ω - (((σ/√n)2/2)ω2)

The last line is the characteristic function of a normal variable with a mean of μ/n and standard deviation of σ/√n. This holds true for all and n and thus a normal variable is infinitely divisible.

To see that not all random variables are infinitely divisible one does not have to look any further than a discrete random variable that can take on only two distinct values, say b1 and b2. For this variable to be infinitely divisible it would at least have to be divisible; i.e., representable as the sum of two independent, identically distributed random variables. Suppose such a distribution existed. If the variable could take on only one values then there would be no way to obtain the two values of b1 and b2. Suppose the variable could take on two values, say a1 and a2. Then the possible sums of two such random variables are a1 and a1, a1 and a2, and a2 and a2. This precludes the possibility that the sum can have exactly two values, b1 and b2. Thus a random variable that can take on two and only two distinct values cannot be divisible, much less infinitely divisible.

For a random variable that can take on only three distinct values divisibility is also precluded but for a different reason from the above. Let a random have the allowed values of b1, b2 and b3 with probabilities of p1, p2and p3, respectively and where p1+p2+p3 = 1. If this variable were to be represented as the sum of two independent, identically distributed variables those variables could only take on two values, say a1 and a2, with probabilities p and q, respectively. The possible values of the sum of the two variables are 2a1, a1+a2, 2a2. Values of a1 and a2 can be found such that:

2a1 = b1
a1 + a2 = b2
2a2 = b3

if and only if b2 = (b1+b3). Suppose that is the case. The probabilities of 2a1, a1+a2, and 2a2 are p2, 2pq and q2, respectively. The conditions

p2 = p1
2pq = p2
q2 = p3

can be satisfied only if p2 = 2(p1p3)1/2. Any three-valued random variable for which this is not true cannot be divisible, much less infinitely divisible. Thus in general discrete valued random variables cannot be infinitely divisible.

Given the conditions to be satisfied it somewhat surprising that any random variable would be infinitely divisible, but we know from the previous that at least normal variables are infinitely divisible. Actually many of the more statistically significant types of random variables are infinitely divisible.

The conditions for infinite divisibility can be equivalently (and more conveniently) expressed in terms of the log-characteristic functions; i.e.,

log(Φn(ω)) = log(Φ(ω))/n

Some of the infinitely divisible types of random variables are

Some Infinitely Divisible Types of Random Variables
TypeLog-Characteristic FunctionDivided Log-Characteristic Function
Normaliμω - (σ2/2)ω2 i(μ/n)ω - ((σ/√n)2/2)ω2
Poissonλ(e-1) (λ/n)(e-1)
Cauchyiμω-a|ω| i(μ/n)ω-(a/n)|ω|

Canonical Representations of the
Log-Characteristic Function of an
Infinitely Divisible Random Variable

The log-characteristic function of an infinitely divisible random variable can be represented by the following:

Kolmogorov's Formula

log(Φ(ω)) = iδω
+ ∫(eiωu - 1 - iωu)(1/u2)dK(u)

where K(u) is a nondecreasing function such that K(-∞) = 0.

Formula of Lévy and Khintchine

log(Φ(ω)) = iδω +
∫(eiωu - 1 - iωu/(1+u2))((1+u2)/u2)dG(u)

where G(u) is a nondecreasing function of bounded variation.

(To be continued.)


HOME PAGE OF applet-magic
HOME PAGE OF Thayer Watkins