Let $(X_n)$ be a martingale. What can be said about the distribution of its maximum over a window of fixed length: $$M_n = \max_{n-10 \leq k \leq n} X_k$$ or about the "range" over a window: $$R_n = \max_{n-10 \leq k \leq n} X_k - \min_{n-10 \leq k \leq n} X_k $$

I know Doob's inequality, but can we give more precise informations about $M_n$ or $R_n$ ? At least when $X_{n+1} - X_n$ has a normal distribution?


In your latter case ($X_{n+1}-X_n$ normal) you can use a large deviation principle for martingales (if $X_{n+1}-X_n$ is bounded it is called Azuma inequality but there are extensions with worse rates for martingale differences with finite exponential moments) together with a naive union bound to get a good estimate. This will work also for estimating $$\max_{k\in [cn,n]}X_k$$ with $c>0$.

  • $\begingroup$ Thanks, I'll have a look. In the case $X_n$ is just a simple random walk with independent increments $+1$ or $-1$ with probability $1/2$, is there a well known distribution for $R_n$ ? It looks roughly like a log-normal... $\endgroup$
    – Basj
    Feb 1 '16 at 16:33
  • 4
    $\begingroup$ It would be a shame to use a naïve union bound as the Hoeffding-Azuma inequality gives the same bound for the maximum of a martingale as for the last term. Of the many tail bounds for martingales, my personal favorite is [1] which directly applies to the maximum. [1] Freedman, David A. "On tail probabilities for martingales." the Annals of Probability (1975): 100-118. $\endgroup$ Feb 7 '16 at 17:45

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.