6
$\begingroup$

Consider a Gaussian Process $g\sim GP(\mu,k)$ with mean zero $\mu\equiv0$ and continues covariance $k(t_1,t_2)=k(|t_1-t_2|)$ defined on the interval $A=[0,T]$. I'd like to make no assumptions about g(0).

What can be said on the infimum of $g(t)$? For example, what is the distribution of $m=\inf_A g(t)$? What is the mean? What is the median?

Related and perhaps simpler version would be to ask what is $Pr(\{g(t)>c\ :\ \forall t\in A\})=Pr(m>c)$? Or at least what is $Pr(m>0)$, the probability of no zero crossing in $[0,T]$?

Note that unlike related question I'm not interested in the asymptotic behavior for large $c$ but rather what is the probability of infimum near 0 (or how $Pr(m>c)$ behave for small $c$).

EDITED: Per Ylvisaker 1965 we know the expected number of zeros in $A$ is $\frac{T}{\pi}\sqrt{-k''(0)}$. But I couldn't see how it help get $Pr(m>0)$.

$\endgroup$

2 Answers 2

4
$\begingroup$

I'll summarize what I've learned.

Denote a Gaussian Process $g\sim GP(\mu,k)$ with mean zero $\mu\equiv0$ and continues covariance $k(t_{1},t_{2})=k(|t_{1}-t_{2}|)$ defined on the interval $[0,T]$ and also $m=\inf_{t\in[0,T]}g(t)$.

I could not find any result regarding the pdf (or cdf) of $m$, not even its mean. Also $P(m>c)$ yielded no results except for the case $c=0$, where from symmetry $P(m>0)=\frac{1}{2}p(0,T)$ with $p(n,T)$ is the probability of $g(t)$ having exactly $n$ zeros in $[0,T]$. This latter quantity is discussed extensively in the literature on “zero crossing intervals” and can be calculated numerically, as we shall now explain.

As already mensioned, the mean number of zeros in time unit is given by Rice' Formula $\beta=\sqrt{-k''(0)}/\pi$ (when $k$ is normalized, $k(0)=1$). Also denote $P_{n}(\tau)$ the probability density for of the event that the $n+1$th zero crossing occuring at interval $\tau$ after the first zero crossing. McFadden's paper from 1958, The axis-crossing intervals of random functions--II show that $U(\tau)=\sum_{0}^{\infty}P_{n}(\tau)$, the probability of some zero crossing at interval $\tau$ after a zero crossing have a (complicated) closed-form expression $U(\tau)=\frac{2}{\pi}(\frac{1}{H(\tau)}-\tan^{-1}H(\tau))R(\tau)$ (eq. 20-21 there) with $R(\tau)=\frac{1}{2\beta\pi}\frac{k(\tau)k'(\tau)^{2}+(1-k(\tau)^{2})k''(\tau)}{(1-k(\tau)^{2})^{3/2}}$ and $H(\tau)=\frac{k(\tau)k'(\tau)^{2}+(1-k(\tau)^{2})k''(\tau)}{\sqrt{[k'(\tau)^{2}+(1-k(\tau)^{2})k''(0)]^{2}-[k(\tau)k'(\tau)^{2}+(1-k(\tau)^{2})k''(\tau)]^{2}}}$. Now $P_{0}(\tau)$ can be calculating numerically from the convolutional equation $P_{0}(\tau)=\frac{1}{2}[U(\tau)+R(\tau)]-\frac{1}{2}[U(\tau)-R(\tau)]*P_{0}(\tau)$ (eq. 47 there). Then $p(0,T)$ can be calculated numerically by integrating $p(0,T)=1+\beta[\int_{0}^{T}d\tau\int_{0}^{\tau}dlP_{0}(l)-\tau]$.

$\endgroup$
1
  • $\begingroup$ Since $\inf(X_t)=-\sup(-X_t)$ and the gaussian process is zero mean, we have $\inf(X_t)=-\sup(-X_t)=-\sup(X_t)$. Then all results on $\sup(X_t)$ can be adapted to here. Is it? For example, from chaining method we could know $E(\sup(X_t))$ thus, we can know $E(\inf(X_t))=-E(\sup(X_t))$ $\endgroup$
    – tony
    Jul 12, 2023 at 15:34
3
$\begingroup$

Googling "Stationary Gaussian process" and "level crossing" or "first passage" brings up a lot of relevant-looking material. This one seems to address your question most directly out of the ones I peeked at: Zero Crossing Probabilities for Gaussian Stationary Processes by G. F. Newell and M. Rosenblatt. The authors look at the quantity $H_X(T) = P[X(t)>0\text{ for all } 0\le t \le T]$ and establish some upper and lower bounds.

$\endgroup$
2
  • $\begingroup$ Thanks. If I got it right it is a rather general result, in the sense that the dependence on T is bounded from above and below, establishing the decay depend on T in the exponent. Their result don't depend on the correlation function, hinting its only a coarse description. I'm looking for something which will allow to establish which properties of the correlation function control this value. $\endgroup$
    – Uri Cohen
    Aug 6, 2015 at 19:13
  • $\begingroup$ Their result depends only on how $k(t)$ decays for large $t$. It is indeed a coarse description. If you need something stronger, you will have to search further. Perhaps this could be useful: jstor.org/stable/3481558 $\endgroup$ Aug 6, 2015 at 19:21

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.