Let $M$ be a positive definite matrix and let $w \in S^{d-1}$ be a unit vector uniformly distributed over the sphere. I want to understand the distribution of the quadratic form $\frac{w^T M^3 w}{w^T M^2 w}$. Since the $w$ is uniformly distributed over the unit $(d-1)$-sphere, the coordinate values are Gaussian distributed. It seems that both $w^T M^3 w$ and $w^T M^2 w$ are Gaussian distributed random variables. The ratio distribution of two Gaussian distributions is a Cauchy distribution, however in this case the distributions have correlation coefficient $\rho = 1$. Every formula for the ratio distribution I've seen has the expression $\frac{1}{\sqrt{1-\rho^2}}$.

Experimentally the distribution of $\frac{w^T M^3 w}{w^T M^2 w}$ behaves very similarly to a Gaussian. Is there a simple expression for the distribution of this random variable or for its mean and variance?


1 Answer 1


$\newcommand{\ep}{\epsilon} \newcommand{\ga}{\gamma} \newcommand{\la}{\lambda} \newcommand{\La}{\Lambda} \newcommand{\Si}{\Sigma} \newcommand{\R}{\mathbb{R}} \newcommand{\E}{\operatorname{\mathsf E}} \newcommand{\PP}{\operatorname{\mathsf P}}$

Let \begin{equation} R:=\frac{w^T M^3 w}{w^T M^2 w}. \end{equation} The distribution of none of the random variables (r.v.'s) $w^T M^3 w$, $w^T M^2 w$, $R$ is Gaussian -- because each of them is positive.

However, $R$ is the ratio of two quadratic forms in iid standard normal r.v.'s $Z_1,\dots,Z_d$. Indeed, by the spherical symmetry of the distribution of the standard normal random vector $Z:=(Z_1,\dots,Z_d)$, the random vector $w$ equals $Z/\|Z\|$ in distribution, where $\|Z\|$ is the Euclidean norm of $Z$. So, we may write $R$ as the ratio of two quadratic forms in $Z_1,\dots,Z_d$: \begin{equation} R=\frac{Q_3}{Q_2},\quad Q_j:=Z^T M^j Z. \end{equation}

Consider now the spectral decomposition $M=O^T\La O$ of the positive definite matrix $M$, where $O$ is an orthogonal matrix and $\La$ is the diagonal matrix with the eigenvalues $\la_1\ge\cdots\ge\la_d>0$ of $M$ on the diagonal. Then \begin{equation} Q_j=Z^T O^T \La^j O Z=V^T \La^j V=\sum_1^d\la_i^j V_i^2, \end{equation} where $V=(V_1,\dots,V_d):=OZ$, so that $V$ is also a standard normal random vector, and hence $V_1,\dots,V_d$ are iid standard normal r.v.'s. Note also that $Q_3=\sum_1^d\la_i^3 V_i^2\le\la_1\sum_1^d\la_i^2 V_i^2=\la_1Q_2$ and, similarly, $Q_3\ge\la_dQ_2$, so that we have the specific bounds on the values of $R$, namely $\la_d\le R\le\la_1$.

In particular, if $\la_1=\la_d=:\la$, then $R=\la$, a constant. To avoid this triviality, let us assume that from now on we have $\la_d<\la_1$.

Then, using transformations of distributions of random vectors (see e.g. formula (10)), one can find integral expressions for the density (pdf) $f_R$ of $R$.

For $d=2$, this pdf can be found explicitly. Indeed, then we can write \begin{equation} R=\frac{\la_1^3 V_1^2+\la_2^3 V_2^2}{\la_1^2 V_1^2+\la_2^2 V_2^2} =\frac{\la_1^3 F+\la_2^3}{\la_1^2 F+\la_2^2}=:\rho(F),\quad F:=\frac{V_1^2}{V_2^2}. \end{equation} Let $u:=\rho^{-1}$, so that \begin{equation} u(r)=\frac{\la_2^2 (r-\la_2)}{\la_1^2 \left(\la_1-r\right)} \end{equation} for $r\in(\la_2,\la_1)$. Note also that $F$ has the $F_{1,1}$ distribution, with pdf $f_F$ given by \begin{equation} f_F(x)=\frac1{\pi(1+x)\sqrt x} \end{equation} for $x>0$. So, by formula (5), the pdf $f_R$ of $R$ is given by \begin{equation} f_R(r)=f_V(u(r))|u'(r)|= \frac{\la_1 \la_2}{\pi \sqrt{\left(\la_1-r\right) \left(r-\la_2\right)} \,[\la_1^2+(\la_1+\la_2) \left(\la_2-r\right)]} \end{equation} for $r\in(\la_2,\la_1)$. Here is the graph of $f_R$ for $\la_1=2$ and $\la_2=1$; the shape is certainly not Gaussian.

enter image description here


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.