An explanation of the definition of the characteristic function of the bivariate stable distribution

470 Views Asked by At

The multivariate stable distribution is defined as follows (see Nolan 2013 or wikipedia). Let $\mathbb{S}$ be the unit sphere in $\mathbb R^d\colon \mathbb{S} = \{u \in \mathbb R^d\colon|u| = 1\}$. A random vector $X$ has a multivariate stable distribution, that is $X \sim S(\alpha, \Lambda, \delta)$, if the joint characteristic function of $X$ is:

$$\operatorname{E} \exp(i u^T X) = \exp \left\{-\int \limits_{s \in \mathbb S}\left\{|u^Ts|^\alpha + i \nu (u^Ts, \alpha) \right\} \, \Lambda(ds) + i u^T\delta\right\}$$

For $y\in\mathbb R$,

$$\nu(y,\alpha) =\begin{cases} -\mathbf{sign}(y) \tan(\pi \alpha / 2)|y|^\alpha & \alpha \ne 1, \\ (2/\pi)y \ln |y| & \alpha=1. \end{cases}$$

$\delta \in \mathbb R^d$ is the shift parameter and $\Lambda$ is a spectral measure of the distribution. $0<\alpha<2$ is the stability parameter.

My question is this. Could someone please explain to me (in simple, intuitive terms) what $u$ is? It appears throughout the first equation, but I don't know what it is. What is its relationship with $s$?

Nolan, John P., Multivariate elliptically contoured stable distributions: theory and estimation, Comput. Stat. 28, No. 5, 2067-2089 (2013). ZBL1306.65118.

1

There are 1 best solutions below

3
On BEST ANSWER

Let me make an analogy to a characteristic function of a real valued random variable.

Given a random variable $X$, its characteristic function is the map $\varphi_X \colon \mathbb{R} \rightarrow \mathbb{C}$ defined as:

$$ \varphi_X(u) \, \colon = \mathbf{E} \left[ e^{ i u X} \right].$$

Note that since the expectation is taken over the variable $X$, this becomes a function of $u$ only. Characteristic functions are useful, since if you have two random variables $X, \, Y$ and you can show that

$$\varphi_X(u) = \varphi_Y(u), \qquad \forall \, x,\,y \, \in \mathbb{R}$$

then it follows that $X$ and $Y$ have the same distribution.

Now returning back to your context: in your case the variable does not take values in $\mathbb{R}$, but rather in $\mathbb{R}^d$. The equivalent form of a characteristic function for a variable $X \in \mathbb{R}^d$ is a function $\varphi_X \colon \mathbb{R}^d \rightarrow \mathbb{C}$, where now:

$$\varphi_X( u) \, \colon = \mathbf{E} \left[ e^{i u^T X} \right],$$

note that we can take the expectation as $u^T X = \sum_{i=1}^d u_i X_i \in \mathbb{R}$ (since expectations only make sense for `1-dimensional' variables).

As in the case of $d = 1$, if you can show that given two variables $X, Y$ taking values in $\mathbb{R}^d$ such that

$$\varphi_X(u) = \varphi_Y(u), \qquad \forall \, x, \, y \in \mathbb{R}^d$$

then it follows that the variables are equal in distribution.

Further Introduction to Characteristic Functions

In this I go into some more detail to motivate characteristic functions. I return to the scenario that $X$ is a real valued variable, and $\varphi_X \colon \mathbb{R} \rightarrow \mathbb{C}$.

A common question in probability / statistics is: given two random variables / sets of data, how do I know if they have the same distribution.

A first approach might be to check: do they have the same expectation (mean, if doing statistics)? So suppose we have variables $X, Y$ we might say

$$ \text{Is } \mathbf{E}[X] = \mathbf{E}[Y] \text{ true?}$$

If its not then we'd conclude that they're definitely do not have the same distribution.

But, is the converse true? If $X, \, Y$ have the same mean, do they have the same distribution? The answer is no.

Example

Suppose $X \sim N(0,1)$ (a normal distribution with mean $0$ and standard deviation $1$), and $Y \sim \text{Unif}[-1,1]$ (a uniform variable on the interval $[-1,1]$). Both variables have the property

$$ \mathbf{E}[X] = \mathbf{E}[Y] = 0$$

However they definitely don't have the same distribution (since $\mathbf{P}[X > 1] > 0$, whilst $\mathbf{P}[Y > 1] = 0$).

So we might then ask: "If two random variables have the same mean, and the same variance... Do they have the same distribution?". Again the answer to this is no.

Example

Suppose $X \sim N(1,1)$ and $Y \sim \text{Poi}(1)$ (a Poisson variable with mean $1$). Then

$$\mathbf{E}[X] = \mathbf{E}[Y] = 1, \qquad \text{Var}(X) = \text{Var}(Y) = 1$$

Similarly we can construct examples for any given integer $k \geq 0$ that show that

$$\mathbf{E}[ X^k] = \mathbf{E}[Y^k]$$

and yet $X$ and $Y$ do not have the same distribution. We call $\mathbf{E}[X^k]$ the $k-$th moment of $X$.

However... It turns out (and this requires proving) that if you can show that all moments of the distributions agree: then they have the same distribution. i.e. if

$$\mathbf{E}[X^k] = \mathbf{E}[Y^k], \qquad \forall \, k \geq 0$$

then $X \sim Y$.

How does this relate to the characteristic function?

A similar statement from real analysis says that a power series uniquely determines a function. So the insight is: if we create a power series with the moments of $X$ as coefficients, and then we find a variable $Y$ with the same power series... Then: uniqueness of power series implies they have the same coefficients, i.e. the same moments, and therefore have the same distribution!

So we construct the power series:

$$ f(u) = \mathbf{E}[X^0] + \mathbf{E}[X^1] u + \mathbf{E}[X^2] \frac{u^2}{2!} + \mathbf{E}[X^3] \frac{u^3}{3!} + \cdots$$

We introduce the parameter $u$ purely as an artificial way to make a function from a sequence of moments $\{\mathbf{E}[X^k]\}_{k \geq 0}$. Using linearity of expectations, and the convenience of dividing the moments by factorials,we get:

$$ f(u) = \mathbf{E}\left[ 1 + uX + \frac{u^2X^2}{2}+ \frac{u^3X^3}{3!} + \cdots \right] = \mathbf{E}[ e^{uX} ]$$

The function $f$ is almost the characteristic function (its missing the $i$), and is known as the moment generating function. The above argument therefore says that if we have two random variables $X,Y$ and they have the same moment generating function, then they have the same distribution.

Introduction of the parameter $i$, so $f(iu)$, changes very little: the claim is still true that $\varphi = f(iu)$ uniquely determines the distribution. Really the addition of $i$ is historic, in as much as $\varphi$ is equivalent to the Fourier transform in real analysis.

So in short: the moments of a random variable uniquely determine the distribution. Therefore, a power series with the moments as coefficients uniquely determines the distribution. The characteristic function is exactly that: a power series, in some parameter $u$, with the moments of a random variable as coefficients.

The value of $u$ itself is not of interest: $u$ is a way to encode a sequence of information (the coefficients), as a function over a continuous parameter.