2
$\begingroup$

We have a centered Gaussian process $X_{t}$ where we have exact equality $$E[X_{t}X_{s}]=a_{1}-a_{2}|t-s|$$ for $|t-s|<\epsilon_{0}\ll \frac{a_{1}}{a_{2}}$ and $a_{i}>0$.

Q: I am curious if there is any other concrete Gaussian process $(Y_{s})_{s\in [0,\epsilon_{0}]}$ out there with the same exact covariance when $|t-s|<\epsilon_{0}$ for some $\epsilon_{0}>0$ (not asymptotical behaviour with error term, but exact equality).

It will be interesting if $Y_{t}$ is in terms of some known process like a functional of Brownian motion or a stationary solution of some SDE.

We are not concerned with $Y_{t}$ having different distribution than $X_{t}$(even though they do looked as Gaussian processes over $t\in [0,\epsilon']$ for $\epsilon'$ small enough). Our main concern is if such covariances have been studied in the literature or if we can devise one.

Some idea: start from $Y_{t}=\int_{0}^{t}f(r,t)dW_{r}$ and try to find a deterministic $f(r,t)$ with the desired covariance: by Ito isometry $\int_{0}^{s}f(r,s+h)f(r,s)ds=a_{1}-a_{2}h$.

Our process

Let $X_{\epsilon}(x)\sim N(0,\ln\frac{1}{\epsilon})$ with covariance:

enter image description here

For simplicity above we suppressed the $\epsilon$ and just let $X_{t}:=X_{\epsilon}(t)$.

Our particular process. Consider the hyperbolic measure $\lambda:=\frac{1}{y^{2}}dx dy$ in the upper half-plane and a White noise process W indexed by Borel sets of finite hyperbolic area:

$$\{A\subset \mathbb{H}: \lambda(A)<\infty; \sup_{(x,y),(x',y')\in A}|x-x'|<\infty\}$$

with covariance:

$$E[W(A_{1})W(A_{2})]:=\lambda(A_{1}\cap A_{2}).$$

Then let $X_{t}=W(V_{\epsilon}+t)$ for

$$V_{\epsilon}:=\{(x,y)\in \mathbb{H}: x\in [-1/4,1/4]\text{ and }max(2|x|,\epsilon)\leq y<1/2\}.$$

$\endgroup$

2 Answers 2

0
$\begingroup$

By Pólya’s theorem, any even real-valued function $f$ on $\mathbb R$ with $f(\infty-)=0$ which is convex on $[0,\infty)$ is positive definite. So, any such function is the (auto)covariance function of a stationary Gaussian process; see e.g. Section "Properties of the Autocovariance Function", page 2.

Now just take any two different functions, $f_1$ and $f_2$, of the Pólya class such that $f_2(t)=1-|t|=f_2(t)$ for $|t|\le1/2$. Then the corresponding stationary Gaussian processes, say $(X_{1,t})$ and $(X_{2,t})$, with the covariance functions $f_1$ and $f_2$ will have different distributions. Therefore, these two processes will be different from each other.


To be more specific, note first here that, by vertical and horizontal re-scaling, without loss of generality $a_1=a_2=1$, so that $$EX_sX_t=1-|t-s|\quad\text{if}\quad|t-s|\le u, \tag{1}$$ where $u\in(0,1)$. Let then $$Y_t:=B_{t+1}-B_t=\int_t^{t+1}dB_s,$$ where $(B_t)_{t\in\mathbb R}$ is the standard Brownian motion with $B_0=0$. Then $$EY_sY_t=1-|t-s|\quad\text{if}\quad|t-s|\le 1$$ (with $EY_sY_t=0$ if $|t-s|>1$), so that $$EY_sY_t=EX_sX_t\quad\text{if}\quad|t-s|\le u,$$ as desired.

For more examples, take any $h\in(0,1)$ and let $$U_t:=\frac1{\sqrt2}\,(Y_{(1-h)t}+Z_{(1+h)t}),$$ where $(Z_t)$ is an independent copy of the Gaussian process $(Y_t)$. Then $$EU_sU_t=1-|t-s|=EY_sY_t \quad\text{if}\quad|t-s|\le1/(1+h)$$ and hence $$EU_sU_t=EX_sX_t \quad\text{if}\quad|t-s|\le\min[u,1/(1+h)],$$ as desired.

$\endgroup$
16
  • $\begingroup$ Thank you Iosif. The focus of the question was on finding some known process with the above mentioned covariance. Also I didn't write $\leq u$, I wrote $<u$. The distribution comment was just auxiliary, not the focus of the question. $\endgroup$ Commented Apr 19, 2020 at 20:41
  • $\begingroup$ If you like in your setting, you should think that the underlying interval we study is $[0,u]$ for $u<\frac{1}{2}$ and so we get the same distribution there by uniqueness of covariance for Gaussian processes. $\endgroup$ Commented Apr 19, 2020 at 20:47
  • $\begingroup$ @OOESCoupling : Your question was "if there is any other Gaussian process $Y_s$ out there with the same exact covariance when $|t-s|<\epsilon_0$ for some $\epsilon_0>0$". This question was fully answered by just modifying the process you had with such a covariance and thus using no additional knowledge. Your comment about the distinction between $<u$ and $\le u$ looks strange to me, since the covariance function of any 2nd-order process continuous at 0 is continuous everywhere, by Bochner's theorem. Anyhow, now I have given another version of the answer, based on Pólya's theorem. $\endgroup$ Commented Apr 19, 2020 at 21:57
  • $\begingroup$ Thank you Iosif. We are only looking at them as Gaussian processes restricted over $[0,\frac{1}{2}]$ ignoring the rest. Not as processes over the entire positive line, which one can devise many examples as you wrote. $\endgroup$ Commented Apr 19, 2020 at 22:30
  • 1
    $\begingroup$ @OOESCoupling : You keep changing the question. It was "if there is any other Gaussian process [with certain general properties]". Now, after getting full answers to the previous versions of your questions, you change the question to "It will be interesting if $Y_t$ is in terms of some known process like a functional of Brownian motion or a stationary solution of some SDE." An appropriate thing to do in such situations is, not to invalidate valid answers, but to post the additional questions separately. Otherwise, what did I spend my time for? $\endgroup$ Commented Apr 20, 2020 at 0:13
0
$\begingroup$

By vertical and horizontal re-scaling, without loss of generality $a_1=a_2=1$, so that $$EX_sX_t=1-|t-s|\quad\text{if}\quad|t-s|\le u, \tag{1}$$ where $u\in(0,1)$. Take any $h\in(0,1)$ and let $$U_t:=\frac1{\sqrt2}\,(X_{(1-h)t}+Y_{(1+h)t}),$$ where $(Y_t)$ is an independent copy of your Gaussian process $(X_t)$. Then $$EU_sU_t=1-|t-s|=EX_sX_t \tag{2}$$ if $|t-s|\le u/(1+h)$, as desired.

Moreover, letting $u$ take the largest possible value such that (1) still holds, we will see that the first equality in (2) will fail to hold if $|t-s|=u$, which shows that the distribution of $(U_t)$ is different from that of $(X_t)$. Therefore, the process $(U_t)$ is different from the process $(X_t)$.

$\endgroup$

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.