9
$\begingroup$

I am working on a topic outside my main research area, so I am afraid I am reproving obvious results, so I would like to ask for a reference. Google didn't help, mostly because I am looking for formulas.

Let $v$ be an $n$-variate Gaussian random variable, with $E[v]=0$ and $E[vv^\top]=I$. If I am not mistaken, it follows (with a bit of work) from Isserlis' theorem that for any $a,b\in\mathbb{R}^n$ $$ E[(a^\top v)^2(b^\top v)^2] = (a^\top a)(b^\top b) + 2 (a^\top b)^2. $$ Q1. Do you have a reference for this identity?

Now let us consider a generalization: $v_t$, for $t\in \mathbb{Z}$, is a $n$-variate random process, i.e., each $v_t$ is a vector-valued random variable, and $E[v_t]=0$, $E[v_tv_t^\top]=I$, $E[v_tv_s^\top]=0$ for $s\neq t$. Let $L$ be a shift operator on the sequence, i.e., $Lv_t = v_{t-1}$, and $a(L)$ be a polynomial in it with coefficients $a_i\in\mathbb{R}^n$, i.e., $a(L)v_t = a_0 v_t + a_1v_{t-1} + \dots + a_d v_{t-d}$. Then, I imagine that there should be a result of the kind $$E[(a(L)^\top v)^2(b(L)^\top v)^2] = \text{the coefficient of $L^0$ in } (a(L)^\top a(L^{-1}))(b(L)^\top b(L^{-1})) + 2 (a(L)^\top b(L^{-1}))^2.$$

Q1. Is a formula of this kind true / known? Where can I find references on this topic and this kind of algebra on processes?

$\endgroup$

1 Answer 1

8
$\begingroup$

The first equality you mention is a special case of Wick's formula or diagram formula. Suppose that you have a Gaussian random vector $X=(X_1,\dotsc, X_n)$ that is centered, i.e., $\newcommand{\bE}{\mathbb{E}}$ $\newcommand{\bR}{\mathbb{R}}$

$$ \bE[X_i]=0,\;\;\forall i=1,\dotsc, n. $$

A special case of Wick's formula computes $\bE[X_1\dotsc X_n]$. Note that this expectation is $0$ if $n$ is odd so we only need to consider the case $n=2k$.

A Feynman graph associated to $\{X_1,\dotsc, X_{2k}\}$ is then a graph with vertex set $\{X_1,\dotsc, X_{2k}\}$ and exactly $k$-edges so no two edges have a vertex in common. If $\newcommand{\be}{\boldsymbol{e}}$ $\be$ is an edge of a Feynman graph with endpoints $X_i, X_j$ then we define its weight to be the covariance $\bE[X_iX_j]$. The weight of a Feynman graph $\Gamma$ is then

$$w(\Gamma):=\prod_{\be \in E(\Gamma)}w(\be), $$

where $E(\Gamma)$ denotes the set of edges of $\Gamma$. Wick's formula states that

$$ \bE[X_1X_2\dotsc X_{2k}]=\sum_{\Gamma} w(\Gamma) $$

where the summation is carried over the set of Feynman graphs with vertex set $\{X_1,\dotsc, X_{2k}\}$.

Here is how it applies to your case. Suppose that $Y=(Y_1,Y_2)$ is a centered Gaussian vector. Then the random vector

$$ X=(X_1,X_2,X_3,X_4)=(Y_1,Y_1,Y_2,Y_2) $$

is also a centered random Gaussian vector and

$$ \bE[X_1X_2X_3X_4]=\bE[Y_1^2Y_2^2]. $$

In this case there are $3$ Feyman graphs:

$$ \Gamma_1,\;\;E(\Gamma_1)= \big\{ (1,2), (3,4)\big\},\;\; w(\Gamma_1)= \bE[Y_1^2]\bE[Y_2^2], $$

$$\Gamma_2,\;\;E(\Gamma_2)=\big\{ (1,3), (2,4)\big\} ,\;\; w(\Gamma_2)=\bE[Y_1Y_2]^2, $$

$$\Gamma_2,\;\;E(\Gamma_2)=\big\{ (1,3), (2,3)\big\} ,\;\; w(\Gamma_3)=w(\Gamma_2)=\bE[Y_1Y_2]^2. $$

Wick's formula then implies

$$ \bE[Y_1^2Y_2^2]= \bE[Y_1^2]\bE[Y_2^2]+2\bE[Y_1Y_2]^2. $$

The result you mention in $Q_1$ corresponds to $Y_1= a^\top v$, $Y_2=b^\top v$.

The fact that $(Y_1,Y_2)$ is a Gaussian vector follows from the following property of Gaussian vectors: if $V=(V_1,\dotsc, V_n)$ is a Gaussian vector and $A: \bR^n\to\bR^m$ is a linear map, then $AV$ is an $m$-dimensional Gaussian vector.

If we choose $A: \bR^n\to \bR^2$, $V\mapsto (Y_1,Y_2)=(a^\top V, b^\top V)$ then you see that $(Y_1,Y_2)$ is Gaussian.

For the generalization you mention, there are plenty to say if the process is Gaussian.

For details see Sections 1 and 2 of these notes. For many more details see Janson's book Gaussian Hilbert Spaces

$\endgroup$
2
  • 1
    $\begingroup$ Note that this formula of Wick was published by Isserlis in 1916 or 1918. $\endgroup$ Nov 6, 2016 at 11:15
  • 2
    $\begingroup$ To paraphrase Arnold, it is called Wick's formula because Wick was not the first to have discovered it. $\endgroup$ Nov 6, 2016 at 11:30

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.