1
$\begingroup$

Definition. $\mathbf{X} = \begin{bmatrix}X \\ Y\end{bmatrix}$ is multivariate normal (equivalently, $X$ and $Y$ are jointly normal) if there is a vector $\boldsymbol\mu \in \mathbb{R}^2$, a $k \times 2$ matrix $\mathbf{A}$ (with $k \leq 2$), and a random vector $\mathbf{z} \in \mathbb{R}^k$ of $\mathcal{N}(0, 1)$ random variables such that $$\mathbf{X} = \boldsymbol\mu+\mathbf{A}^{T}\mathbf{z}$$ ($T$ denoting the transpose operator) with $\mathbf{A}^{T}\mathbf{A} = \boldsymbol\Sigma$. We write $\mathbf{X} \sim \text{MVN}_{2}(\boldsymbol\mu, \boldsymbol\Sigma)$ if this is the case.

Theorem. Suppose $\mathbf{X} \sim \text{MVN}_{2}(\boldsymbol\mu, \boldsymbol\Sigma)$ with $\boldsymbol\Sigma$ non-singular. Then $\mathbf{X}$ has density $$f_{\mathbf{X}}(\mathbf{x}) = \left[\dfrac{1}{(2\pi)^2\det(\boldsymbol\Sigma)} \right]^{1/2}\exp\left[-\dfrac{1}{2}(\mathbf{x}-\boldsymbol\mu)^{T}\boldsymbol\Sigma^{-1}(\mathbf{x}-\boldsymbol\mu) \right]\text{.}\tag{1}$$

Question. Suppose I have two random variables $X, Y$ with support in $\mathbb{R}^2$ such that their (variance-)covariance matrix is non-singular. I am able to write their joint density in the form of $(1)$ above. Is this sufficient to show that $\mathbf{X} = \begin{bmatrix} X \\ Y \end{bmatrix}$ is multivariate normal?

If more context is needed, let me know. I have not been able to find an answer to this in any searching I've done, and apologize that I haven't been able to provide much work for this. I understand how the Theorem above is proven: use transformation techniques, but the problem with this is that you need to know what the distribution of a random variable already is prior to the transformation, so you can't easily "reverse" this proof.

11/26: I've considered pursuing this by computing the MGF of $\mathbf{X}$ and showing that it matches the MGF of a multivariate normal distribution (invoking the theorem that MGFs uniquely determine the distribution), but this is seems to be a lot more work than it is worth.

$\endgroup$
2
+150
$\begingroup$

But you do know the density of $(X, Y)$, which is all you need. I will write down the general proof for the $n$-dimensional normal distribution.

Since $\Sigma$ is a symmetric matrix, we can diagonalize it orthogonally, i.e. $\Sigma = Q^T \Lambda Q$, where $\Lambda = \operatorname{diag}(\lambda_1, \ldots, \lambda_n)$ are the positive eigenvalues of $\Sigma$. This allows us to define a square root of $\Sigma$ as $A := Q^T \Lambda^{1/2} Q$, where $\Lambda^{1/2} := \operatorname{diag}(\sqrt{\lambda_1}, \ldots, \sqrt{\lambda_n})$.

Note that $A$ is a symmetric matrix that satisfies $$A \Sigma^{-1} A = (Q^t \Lambda^{1/2} Q) (Q^T \Lambda^{-1} Q) (Q^T \Lambda^{1/2} Q) = Q^T \Lambda^{1/2} \Lambda^{-1} \Lambda^{1/2} Q = Q^TQ = I_n.$$

Also, $\det A = \sqrt{\lambda_1\cdots \lambda_n} = \sqrt{\det \Sigma}$ holds.

Now consider the random variable $Z = A^{-1}(X - \mu)$. From the density transformation theorem we know that the density of $Z$ is given by

$$(\det A) \cdot \frac{1}{(2\pi)^{n/2} \sqrt{\det \Sigma}} \exp(-\tfrac{1}{2} z^t A^t \Sigma^{-1} A z) = (2\pi)^{-n/2} \exp(-\tfrac{1}{2} z^t z).$$

But this means that $Z$ consists of independent standard normal distributions, which we wanted to prove.

$\endgroup$

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.