1
$\begingroup$

I am interested in upper and lower bounding probability that one component of a multivariate Gaussian exceeds all others. For instance, say we have a multivariate Gaussian RV $X \sim N(\mu, \Sigma)$ where $X \in \mathbb{R}^d$. I am interested in the probability $$ \Pr\{ X_1 \gt X_j \text{ for all } j\ge 2 \}. $$

My understanding from Kenneth Train's book on discrete choice modeling (ch. 5, equation 5.1) is that this computation is intractable when the number of variables in the multivariate Gaussian distribution is greater than 2.

However, it's easy to upper bound this probability as $$ \Pr\bigl\{ X_1 \gt X_j \text{ for all } j\ge 2 \bigr\} ~\le~ \min\bigl\{\, \Pr(X_1 \gt X_2),\, \Pr(X_1 \gt X_2), \,\ldots, \, \Pr(X_1 \gt X_d) \,\bigr\} $$ where the probability of each of the Gaussian two-way comparisons on the right hand side is easy to compute (see here for example). The bound is due to the fact that the intersection of all the constituent events must have a probability no bigger than each of them individually.

However, I haven't been able to come up with a useful lower bound for the same and am hoping someone here may have some useful suggestions on how to go about finding one.

$\endgroup$
4
  • $\begingroup$ This question is impossible to answer unless and until you fully specify what you mean by "a useful lower bound". $\endgroup$ Commented Jun 23 at 13:33
  • $\begingroup$ Hi losif, that’s fair enough. By useful I simply mean a lower bound that’s better than “0” $\endgroup$
    – ted
    Commented Jun 23 at 17:13
  • 2
    $\begingroup$ @Alf : The Gaussian correlation inequality applies to symmetric convex sets, whereas the corresponding sets here are not symmetric. $\endgroup$ Commented Jun 23 at 20:22
  • $\begingroup$ Thanks, Iosif, of course you’re right! $\endgroup$
    – Alf
    Commented Jun 24 at 6:16

1 Answer 1

2
$\begingroup$

$\newcommand{\R}{\mathbb R}\newcommand{\Si}{\Sigma}$We have to bound
\begin{equation*} p:=P(X_1>\max_{j=2}^d X_j) \end{equation*} from below.

Note that
\begin{equation*} p=(2\pi)^{-d/2}|\Si|^{-1/2} \int_{\R^d}dx\,1(x_1>\max_{j=2}^d x_j)\exp(-(x-\mu)^\top\Si^{-1}(x-\mu)/2), \end{equation*} where $|\Si|$ is the determinant of $\Si$, and \begin{equation*} (x-\mu)^\top\Si^{-1}(x-\mu)\le\|\Si^{-1}\|\,|x-\mu|^2 \\ \le\|\Si^{-1}\|(|x|+|\mu|)^2 \le\|\Si^{-1}\| \Big(\frac{|x|^2}{1-t}+\frac{|\mu|^2}t\Big) \tag{10}\label{10} \end{equation*} for any $t\in(0,1)$, where $\|\Si^{-1}\|$ is the spectral norm of $\Si^{-1}$ and $|\cdot|$ is the Euclidean norm of vectors in $\R^d$; to verify the last inequality in \eqref{10}, minimize the latter expression there wrt $t\in(0,1)$. So, \begin{equation*} p\ge |\Si|^{-1/2}\exp\Big(-\frac{\|\Si^{-1}\|\,|\mu|^2}{2t}\Big) \Big(\frac{\|\Si^{-1}\|}{1-t}\Big)^{-d/2}q, \end{equation*} where \begin{equation*} q:=P(Z_1>\max_{j=2}^d Z_j), \end{equation*} $(Z_1,\dots,Z_d)\sim N\Big(0,\dfrac{1-t}{\|\Si^{-1}\|}\,I_d\Big)$, and $I_d$ is the $d\times d$ identity matrix, so that, by symmetry, $q=1/d$.

Thus, \begin{equation*} p\ge L_t:= \frac1{|\Si|^{1/2} \exp\Big(\dfrac{\|\Si^{-1}\|\,|\mu|^2}{2t}\Big) \Big(\dfrac{\|\Si^{-1}\|}{1-t}\Big)^{d/2}\,d}. \end{equation*}


In particular, if $\mu=0$ and $\Si=s^2I_d$ for some real $s>0$, then, letting $t\downarrow0$, we have $L_t\to 1/d=P(X_1>\max_{j=2}^d X_j)=p$, so that the lower bound $L_t$ on $p$ is sharp in this case.

$\endgroup$
1
  • 1
    $\begingroup$ @ted : I have added details on the upper-bounding the Mahalanobis distance. $\endgroup$ Commented Jun 23 at 21:32

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .