2
$\begingroup$

Let $N,k \geq 1$ be two integers and consider the following Markov chain on $[0,k\times N]^k$. It starts at $X^0=(N, N, \ldots, N)$ and $X^{n+1}$ is the realisation of a multinomial distribution with $kN$ trials and probability distribution $(X^n_1, \ldots, X^n_k) / (kN)$. I am interested (for large $N \gg 1$) in the expectation of first time when one of the coordinates of the Markov chain vanishes. A good upper bound would be great.

A scaling argument should show that for any $\epsilon>0$ it takes $O(N)$ steps for one of components to be of order $\epsilon N$. But for one component to be zero, I believe that it should be of order $N \ln(N)$.

For example, for $k=2$, the scaled process $Z^{(N)}_t = N^{-1} \, X^{[Nt]}_1$ behaves (when away from $0$ or $1$) as the diffusion $dZ = \sqrt{Z(1-Z)}\,dW$ so that one can compute the expected hitting time of $0$ or $1$ by noticing that $t \mapsto F(Z_t)-t$ is a martingale with $F(z) = z \ln(z) + (1-z) \ln(1-z)$. This approach does to seem to generalize well for the original Markov chain.

$\endgroup$

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.