7
$\begingroup$

Let $$Y =\begin{pmatrix} Y_1 \\ Y_2 \end{pmatrix} \sim N(0, \Sigma) \quad \Sigma=\begin{pmatrix} \sigma_{11} & \sigma_{12}\\ \sigma_{21} & \sigma_{22} \end{pmatrix}$$

Show that $$\mathbb E(Y_1^2Y_2^2)=\sigma_{11}\sigma_{22} + 2\sigma_{12}^2. $$ I tried to turn this in all the ways but I need $\mathbb {Var}(Y_1Y_2)$ that I don't have. In the solution I was given, it only uses the fact that "the distribution is Gaussian".

$\endgroup$

5 Answers 5

8
$\begingroup$

Since all Gaussian variables have a kurtosis of $3,$ any Gaussian variable with variance $\sigma^2$ has a central fourth moment of $3\sigma^4.$ In particular, when $Z$ is a zero-mean Gaussian,

$$E[Z^4] = 3\sigma^4.$$

We will apply this below to the Gaussian variables $Z = Y_1\pm Y_2$ by observing (from the symmetry of $\Sigma$) that $$\operatorname{Var}(Y_1\pm Y_2) = \sigma_{11} + \sigma_{22} \pm 2\sigma_{12}.$$

Emulating the method I describe at https://stats.stackexchange.com/a/267021/919 (based on the Polarization Identity for $n=4$ with $x_1 = x_2 = Y_1$ and $x_3 = x_4 = Y_2$), write

$$12Y_1^2Y_2^2 = (Y_1+Y_2)^4 + (Y_1-Y_2)^4 - 2(Y_1^4) - 2(Y_2^4).$$

Use linearity of expectation to compute

$$\begin{aligned} 12 E[Y_1^2 Y_2^2] &= E[(Y_1+Y_2)^4] + E[(Y_1-Y_2)^4] - 2E[Y_1^4] - 2E[Y_2^4]\\ & = 3(\sigma_{11} + \sigma_{22} + 2\sigma_{12})^2 + 3(\sigma_{11} + \sigma_{22} - 2\sigma_{12})^2 - 2(3\sigma_{11}^2) - 2(3\sigma_{22}^2)\\ &= 12\sigma_{11}\sigma_{22} + 24\sigma_{12}^2. \end{aligned}.$$

Dividing both sides by $12$ gives the result in the question.

$\endgroup$
8
$\begingroup$

Use the Characteristic function. It is well known that,

$$\varphi_Y(s) := \mathbb E\left[e^{is^\intercal Y}\right] = e^{-\frac12s^\intercal \Sigma s}.$$

So \begin{align} \frac{\partial^2 \varphi_Y}{\partial^2 s_1} = \mathbb E\left[-Y_1^2e^{is^\intercal Y}\right] = \frac{\partial}{\partial s_1}\left[\left(-\sigma_{11}s_1-\sigma_{12}s_2\right)e^{-\frac12s^\intercal \Sigma s}\right] = \left(-\sigma_{11} + \left(\sigma_{11}s_1 + \sigma_{12}s_2\right)^2\right)e^{-\frac12s^\intercal \Sigma s} \end{align}

Put $s_1 = 0$ and you will have:

$$-\mathbb E\left[Y_1^2 e^{is_2Y_2}\right] = \left(-\sigma_{11} + \sigma_{12}^2s_2^2\right)e^{-\frac12\sigma_{22}s_2^2}$$

and then,

$$\mathbb E\left[Y_1^2 Y_2^2e^{is_2Y_2}\right] = \frac{\partial^2}{\partial s_2^2}\mathbb E\left[-Y_1^2 e^{is_2Y_2}\right] = \frac{\partial}{\partial s_2}\left[\left(2\sigma_{12}^2s_2 +\sigma_{11}\sigma_{22}s_2 - \sigma_{12}^2 \sigma_{22}s_2^3\right)e^{-\frac12\sigma_{22}s_2^2}\right] = \left(2\sigma_{12}^2 + \sigma_{11}\sigma_{22} -3\sigma_{12}^2\sigma_{22}s^2_2-2\sigma_{12}^2\sigma_{22}s^3_2 - \sigma_{11}\sigma^2_{22}s^2_2 + \sigma_{12}^2 \sigma_{22}^2s_2^4\right)e^{-\frac12\sigma_{22}s_2^2}$$

Put $s_2 = 0$ and you will have:

$$\mathbb E\left[Y_1^2Y_2^2\right] = 2\sigma_{12}^2 + \sigma_{11}\sigma_{22}$$


It is worth it to mention that with this method you can compute all:

$$\mathbb E\left[Y_1^kY_2^\ell\right]$$

$\endgroup$
7
$\begingroup$

You can use the conditional distributions, which are univariate normal.

For example, $Y_2$ given $Y_1$ is normal with mean $$E\left[Y_2\mid Y_1\right]=\frac{\sigma_{12}}{\sigma_{11}} Y_1$$ and variance $$\operatorname{Var}\left[Y_2\mid Y_1\right]=\left(1-\frac{\sigma_{12}^2}{\sigma_{11}\sigma_{22}}\right)\sigma_{22}$$

So, by law of total expectation,

\begin{align} E\left[Y_1^2Y_2^2\right]&=E\,E\left[Y_1^2Y_2^2\mid Y_1\right] \\&=E\left[Y_1^2 E\left[Y_2^2\mid Y_1\right]\right] \\&=E\left[Y_1^2\left\{\operatorname{Var}\left[Y_2\mid Y_1\right]+(E\left[Y_2\mid Y_1\right])^2\right\}\right] \end{align}

You would need the fourth moment of $Y_1$, which is $E\left[Y_1^4\right]=3\sigma^2_{11}$.

$\endgroup$
5
$\begingroup$

Represent $Y_2$ by $\frac{\sigma_{12}}{\sigma_{11}}\cdot Y_1 + \sqrt{\sigma_{22} - \frac{\sigma_{12}^2}{\sigma_{11}^2} } \cdot Z$, where $Z \sim \mathop{\mathcal N}\left(0,1\right), Z \perp \!\!\! \perp Y_1.$
Now, $$ {Y_1}^2 \cdot \left( \frac{\sigma_{12}}{\sigma_{11}} \cdot Y_1 + \sqrt{\sigma_{22} - \frac{\sigma_{12}^2}{\sigma_{11}^2} } \cdot Z \right)^2 \\ = {Y_1}^4 \cdot \frac{\sigma_{12}^2}{\sigma_{11}^2} + 2 \cdot {Y_1}^3 \cdot Z \cdot \frac{\sigma_{12}}{\sigma_{11}} \cdot \sqrt{\sigma_{22} - \frac{\sigma_{12}^2}{\sigma_{11}^2} } + {Y_1}^2 \cdot Z^2 \cdot \left( \sigma_{22} - \frac{\sigma_{12}^2}{\sigma_{11}^2} \right). $$ Utilizing independence of $Y_1$ and $Z$ (and linearity) in calculating the expectation of the above term gives \begin{align} \mathop{\mathbb E}\left({Y_1}^2{Y_2}^2\right) &= \mathop{\mathbb E}\left({Y_1}^4 \right) \cdot \frac{\sigma_{12}^2}{\sigma_{11}^2} + \mathop{\mathbb E}\left(Z\right)\cdot \left(\dots\right) + \mathop{\mathbb E}\left({Y_1}^2\right) \cdot \mathop{\mathbb E}\left(Z^2\right) \cdot \left( \sigma_{22} - \frac{\sigma_{12}^2}{\sigma_{11}^2} \right) \\ &= 3 \sigma_{11}^2 \cdot \frac{\sigma_{12}^2}{\sigma_{11}^2} + 0 + \sigma_{11} \cdot 1 \cdot \left( \sigma_{22} - \frac{\sigma_{12}^2}{\sigma_{11}^2} \right) \\ &= \sigma_{11} \sigma_{22} + 2 \sigma_{12}^2. \end{align}

$\endgroup$
5
$\begingroup$

One way to approach this problem is exploiting the Cholesky decomposition of $\Sigma$. Let $\Sigma = LL^\top$ be the Cholesky decomposition of $\Sigma$, where $L = \begin{bmatrix} a & 0 \\ b & c \end{bmatrix}$ is a lower triangular matrix. $Y$ then has the same distribution of $LX$, where $X = \begin{bmatrix} X_1 \\ X_2 \end{bmatrix} \sim N(0, I_{(2)})$. Therefore, $Y_1 \overset{d}{=} aX_1$, $Y_2 \overset{d}{=} bX_1 + cX_2$, whence \begin{align} & E[Y_1^2Y_2^2] = E[a^2X_1^2(b^2X_1^2 + 2bcX_1X_2 + c^2X_2^2)] \\ =& a^2b^2E[X_1^4] +2a^2bcE[X_1^3X_2] + a^2c^2E[X_1^2X_2^2] \\ =& 3a^2b^2 + a^2c^2. \tag{1}\label{1} \end{align} In the last step, we used $E[X_1^4] = 3, E[X_1^3X_2] = E[X_1^3]E[X_2] = 0$ and $E[X_1^2X_2^2] = E[X_1^2]E[X_2^2] = 1$, which are the results of $X \sim N(0, I_{(2)})$.

Next, expanding the decomposition $LL^\top = \Sigma$ and comparing entries on both sides yields $a^2 = \sigma_{11}$, $ab = \sigma_{12}$ and $b^2 + c^2 = \sigma_{22}$. Hence $\eqref{1}$ reduces to \begin{align} E[Y_1^2Y_2^2] = 2a^2b^2 + a^2(b^2 + c^2) = 2\sigma_{12}^2 + \sigma_{11}\sigma_{22}, \end{align} as desired.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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