Suppose $$X \sim \mathcal N_n(\text{diag}(\Sigma), \sigma^2 \Sigma)$$ where $\Sigma$ may be allowed to be low rank, and let $Y = \min_i > X_i$.

What can be said about $P\left(Y \geq 0\right)$?

In general, I know that the exact distributions of Gaussian order statistics can be intractable, such as this math.se Q&A and the discussion here, but I'm hoping that the relationship between the mean and covariance matrix may lead to some simplification, or how I don't need the distribution of $Y$ but rather just the probability that it is greater than zero. The $X_i$ not being iid prevents me from using the usual things I know for examining minima and maxima, but I'm still hoping something can be done aside from numerical integration/simulation given values of $\Sigma$ and $\sigma$. I'd be very interested in approximations too.

The context on this and the unusual mean vector come from a now-deleted question on stats.se that essentially asked the following:

If we have $$X\sim\mathcal N_k(\mathbf 0, \sigma^2 I)$$ and nonrandom nonzero vectors $z_1,\dots,z_n\in\mathbb R^k$, what is the probability that $$\|X\|^2 \leq \|X-z_i\|^2 \text{ for all } i?$$

$$\|X-z_i\|^2 = \|X\|^2 - 2 X^Tz_i + \|z_i\|^2$$

so the question is equivalent to $P(\|z_i\|^2- 2 X^Tz_i \geq 0 \text{ for all }i)$. I collected the $z_i$ into the columns of a $k\times n$ matrix $Z$ so I can write the random variables in question as an affine transformation of $X$ via $$ \text{diag}(Z^TZ) - 2 Z^TX \sim \mathcal N_n(\text{diag}(Z^TZ), 4\sigma^2 Z^TZ) $$ and I want the probability that this random vector is all non-negative, so this led me to the question I asked. The factored form of $\Sigma$ here is why I want to allow for possibly low rank covariance matrices since I could have $k \leq n$.

  • $\begingroup$ Do you really intend that the mean of the multinormal should be $\text{diag}(\Sigma)$ or is that a typo? Looks strange ... $\endgroup$ – kjetil b halvorsen Aug 23 '20 at 4:24
  • $\begingroup$ @kjetilbhalvorsen yeah I do mean that $\endgroup$ – jld Aug 23 '20 at 15:50
  • $\begingroup$ Interesting --- out of curiosity, what is the context of the problem? $\endgroup$ – kjetil b halvorsen Aug 23 '20 at 19:33
  • $\begingroup$ @kjetilbhalvorsen I've just updated with the context $\endgroup$ – jld Aug 24 '20 at 13:14

At least in the simplest non-trivial case, there are some tractable formulas. Let $$\Sigma=\begin{pmatrix}a & b\\b & d\end{pmatrix}, \text{ where }a>0,\ d>0,\ ad-b^2>0$$ The probability is given by a messy integral of the form $$P(Y\ge0)=\int_0^\infty\!\!\!\int_0^\infty f_{a,b,d,\sigma}(x_1,x_2)\,dx_1\, dx_2$$ However, the integral turns out nicely when $b=0$, giving an expression with the regularized incomplete gamma function $Q$: $$P(Y\ge0)|_{b=0}=\frac14 \left(Q\left(\frac12,\frac{a}{2\sigma^2}\right)-2\right)\! \left(Q\left(\frac12,\frac{d}{2\sigma^2}\right)-2\right) $$ The integral also turns out nicely when we differentiate by $b$ under the integral sign: $$ \frac{dP(Y\ge0)}{db}= \int_0^\infty\!\!\!\int_0^\infty \frac{df_{a,b,d,\sigma}(x_1,x_2)}{db}\,dx_1\, dx_2 = \frac{ \exp\left(\frac{-ad(a+d-2b)}{2\sigma^2(ad-b^2)}\right)}{2\pi\sqrt{ad-b^2}} $$ This lets us calculate the probability for any $a,B,d,\sigma$ as $$P(Y\ge0) = P(Y\ge0)|_{b=0}+\int_{b=0}^B \frac{dP(Y\ge0)}{db} db$$ Then:

  • We can check, e.g., that this correctly gives $P(Y\ge0)=0.4794$ when $a=1,\, b=1/2,\, d=2,\, \sigma=3$.
  • This formula is a computational improvement over the messy integral, replacing an infinite integral in two dimensions with one-dimensional integrals over finite regions.
  • The formula allows asymptotic analysis, e.g. for calculating a limiting probability as $b$ approaches $\sqrt{ad}$, where the messy approach of $\iint\lim_{b\to\sqrt{ad}}f_{a,b,d,\sigma}$ is ill-defined.

Some variant of this approach may also work for $n>2$, especially in the low-rank situation mentioned in the post. Feynman was famous for solving problems like this.


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.