1
$\begingroup$

I have a Markov chain $\{X_k\}_{k\geq 0}$ on $\mathbb{R}$. The corresponding probability density functions satisfy $$ f_{k+1}(t) = \int_{-\infty}^\infty \Psi(t,\tau)f_k(\tau)\,d\tau,\qquad k=0,1,2,\dots $$ I have an analytic expression for the transition kernel $\Psi$, and let's suppose for the moment that the Markov chain is irreducible, positive recurrent, aperiodic, and Harris. And of course, $\int_{-\infty}^\infty \Psi(t,\tau)\,dt=1$.

I am interested in characterizing the moments of the stationary distribution $\pi$. Specifically:

  • What are sufficient conditions that would ensure the moments of $\pi$ are finite?

  • Is there a way to compute bounds on the moments of $\pi$ if they are finite? I can't do this numerically because $\Psi$ is parameterized; I'm interested in how the moments of $\pi$ vary as a function of these parameters. My first instinct was to try to write $\int_{-\infty}^\infty t^mf_{k+1}(t)\,dt$, substitute the recurrence from above and try to simplify and maybe use Holder's inequality, but I ran into a roadblock: it turns out that $\int_{-\infty}^\infty t^m \Psi(t,\tau)\,dt = \infty$ for all $m\geq 1$, even though the integral is finite for $m=0$. So at this point I have no idea how to proceed.

$\endgroup$
2
  • 1
    $\begingroup$ If your transition probabilities don't have finite moments, why would you expect $\pi$ to have finite moments? $\endgroup$ Oct 13, 2020 at 8:48
  • $\begingroup$ Now that I think of it, I guess that doesn't make much sense... My numerical simulations suggested that $\pi$ exists and has finite moments, but you bring up a good point. I will have to think about this more. $\endgroup$ Oct 13, 2020 at 16:37

1 Answer 1

2
$\begingroup$

You wrote:

I can verify that $\Psi$ is continuously differentiable, $\Psi(t,\tau)>0$ for all $t,\tau\in\mathbb{R}$, and of course, $\int_{-\infty}^\infty \Psi(t,\tau)\,dt=1$.

[...] these properties should be sufficient to guarantee that a stationary distribution $\pi$ exists and is unique, and that $f_k \to \pi$ (in the T.V. sense) for any initial $f_0$.

Of course, this is not so. E.g., if $\Psi(t,s)=g(t-s)$, where $g$ is the standard normal pdf, then (considering, for instance the Fourier transform, one can easily see that) there is no stationary distribution. Also, then for any initial $f_0$ and each real $t$ we have $f_k(t)\to0$ as $k\to\infty$.


You have now added more conditions:

let's suppose for the moment that the Markov chain is irreducible, positive recurrent, aperiodic, and Harris. And of course, $\int_{-\infty}^\infty \Psi(t,\tau)\,dt=1$

saying then the following:

These properties should be sufficient to guarantee that a stationary distribution $\pi$ exists and is unique, and that $f_k \to \pi$ (in the T.V. sense) for any initial $f_0$. Moreover, all moments of $\pi$ are finite and the $m^\text{th}$ moment of $f_k$ converges to the $m^\text{th}$ moment of $\pi$ as $k\to\infty$.

However, the latter conclusion will still fail to hold in general -- because the the state space of the chain can be nonlinearly transformed in an arbitrary manner.

More specifically, suppose (say) that the support set of the stationary distribution $\pi$ of an (irreducible positive recurrent aperiodic Harris) Markov chain $(X_k)$ is not bounded from above, so that $$G(x):=\pi\big((x,\infty)\big)>0$$ for all real $x$. Let then $$Y_k:=f(X_k),$$ where $$f(x):=\int_0^x\frac{du}{G(u)}$$ for real $x$, with $\int_0^x:=-\int_x^0$ for real $x<0$. Then $(Y_k)$ is an (irreducible positive recurrent aperiodic Harris) Markov chain with stationary distribution $\pi_f:=\pi f^{-1}$, the pushforward of $\pi$ under the map $f$. Moreover, \begin{align} \int_{[0,\infty)}y\,\pi_f(dy)&=\int_{[0,\infty)}f(x)\,\pi(dx) \\ &=\int_{[0,\infty)}\pi(dx)\,\int_0^x\frac{du}{G(u)} \\ &=\int_0^\infty\frac{du}{G(u)}\,\int_{(u,\infty)} \pi(dx) \\ &=\int_0^\infty\frac{du}{G(u)}\,G(u)=\infty. \end{align} So, the first moment of $\pi_f$ cannot be finite.

Similarly one can deal with the case when the support set of the stationary distribution $\pi$ has a finite limit point.

$\endgroup$
9
  • $\begingroup$ Thanks -- I was missing some crucial details in my original post. I think my statements only guaranteed irreducibility. I will edit my post to make the question more precise. $\endgroup$ Oct 11, 2020 at 22:12
  • $\begingroup$ @LaurentLessard : This is still not enough. $\endgroup$ Oct 12, 2020 at 18:05
  • $\begingroup$ ok -- I realized I don't know what is needed to ensure finite moments, so I modified the question once more. $\endgroup$ Oct 12, 2020 at 23:15
  • $\begingroup$ Why the downvote for this answer? $\endgroup$ Oct 12, 2020 at 23:43
  • $\begingroup$ Because you didn't actually answer my question. Will happily upvote once question is answered or addressed. $\endgroup$ Oct 13, 2020 at 1:04

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.