4
$\begingroup$

Suppose, these are two pictures of plots of conditional probability density of a set of samples.

enter image description here

From the pictures, I can tell that,

(1) the left plot has non-zero covariance between the two dimensions(i.e. off diagonal elements are non-zero), and the means for each dimensions are below 0.

(2) On the right, the covariance matrix is diagonal(i.e. covariance=0. i.e. is spherical), mean=0.

Am I correct?

If YES, what mathematical operation can I apply to the left-plot to make it look like the right one?

$\endgroup$
5
$\begingroup$

Without commenting on whether you are correct, I can tell you the following.

Let $\mathbf{X}$ be a $d$-dimensional random vector:

$$\mathbf{X}=(X_{1},X_{2},\ldots,X_{d})'$$

The joint distribution function can be expressed as:

$$F_{X}(x)=F_{X}(x_{1},\ldots,x_{d})=\text{Pr}(\mathbf{X}\leq \mathbf{x})=\text{Pr}(X_{1}\leq x_{1},\ldots,X_{d}\leq x_{d})$$

With marginal distribution of $X_{i}$:

$$F_{i}(x_{i})=\text{Pr}(X_{i}\leq x_{i})=F_{X}(\infty,\ldots,x_{i},\ldots,\infty)$$

We can say that $\mathbf{X}$ has an elliptical distribution if:

$$\mathbf{X}\overset{d}{=}\mathbf{\mu}+\mathbf{A}\mathbf{Y}$$

where $\mathbf{Y}\sim S_{k}(\psi)$, $\mathbf{A}\in \mathbb{R}^{d\times k}$ and $\mathbf{\mu}\in\mathbb{R}^{d}$. $S_{k}(\psi)$ means a $k$-dimensional spherical distribution with characteristic generator $\psi$.

Essentially, elliptical distributions are obtained via multivariate affine transformations of spherical distributions. This can be illustrated by the characteristic function $\phi_{X}$:

$$\phi_{X}(t)=\mathbb{E}\big(e^{it'\mathbf{X}}\big)=\mathbb{E}\big(e^{it'(\mathbf{\mu}+\mathbf{A}\mathbf{Y})}\big)=e^{it'\mathbf{\mu}}\mathbb{E}\big(e^{i(\mathbf{A}'t)'\mathbf{Y}}\big)=e^{it'\mathbf{\mu}}\psi(t'\mathbf{\Sigma} t)$$

where $\mathbf{\Sigma}=\mathbf{A}\mathbf{A}'$, which is a positive semidefinite matrix. The elliptical distribution is denoted by:

$$\mathbf{X}\sim E_{d}(\mathbf{\mu},\mathbf{\Sigma},\psi)$$

where $\mu$, $\Sigma$ and $\psi$ are referred to as the location vector, dispersion matrix and characteristic generator of the distribution, respectively. $E_{d}$ denotes a $d$-dimensional ellipitical distribution.

Thus, to transform from an elliptical distribution to a spherical one, the following transformation can be applied:

$$\mathbf{X}\sim E_{d}(\mathbf{\mu},\mathbf{\Sigma},\psi)\Longleftrightarrow \mathbf{\Sigma}^{-\tfrac{1}{2}}(\mathbf{X}-\mathbf{\mu}) \sim S_{d}(\psi)$$

This is analogous to the type of transformation you would do to a univariate normal random variable, $N(\mu,\sigma^{2})$, to obtain a standard normal random variable, $N(0,1)$:

$$Z\sim N(\mu,\sigma^{2})\Longleftrightarrow \sigma^{-1}(Z-\mu) \sim N(0,1)$$

$\endgroup$
  • $\begingroup$ Please be kind to give explanation to you symbols. What is d,k,j,i,t,Ed etc. $\endgroup$ – ttnphns Nov 17 '16 at 6:50
  • $\begingroup$ Thank you very much. Is there any specific name for the transformation operation you showed? $\endgroup$ – user366312 Nov 17 '16 at 6:56
  • $\begingroup$ @anonymous I'm not sure if there's any specific name for it. It's just an affine transformation, similar to what you'd do to a normal distribution to get a standard normal distribution: $X\sim N(\mu,\sigma^{2})\Longleftrightarrow \frac{X-\mu}{\sigma}\sim N(0,1)$ $\endgroup$ – Ed P Nov 17 '16 at 7:07
  • $\begingroup$ @ttnphns I was pressed for time when you asked. Notation has been added, hopefully I covered everything. $\endgroup$ – Ed P Nov 17 '16 at 23:34
  • $\begingroup$ @StatsPlease a nice simplified answer. $\endgroup$ – Subhash C. Davar Nov 21 '16 at 13:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.