5
$\begingroup$

I am reading Pinelis "An approach to inequalities for the distributions of infinite -dimensional martingales" and cannot follow his proof of Theorem 3:

Let $(f_n)$ be a martingale in a separable Banach space $(\mathcal{X},||~||)$, $\mathcal{X} = L^p, p \geq 2$ and $f^*=\sup\{||f_n||\}$.

Theorem 3 says

\begin{align} P(f^*>r) \leq 2\exp\big(-r^2/2(p-1)\big), \quad r \geq 0 \end{align}

and Pinelis writes

One can compare the last inequality with \begin{align} P(f^*>r) \leq (r+1)\exp(-r^2/2) \end{align} in Kallenberg ans Sztencel (1991) proved for $\mathcal{X} = L^2$.

Does he refer to

\begin{align} P(f^*>r) \leq \frac{1+r}{1+rc}\exp\big(-\frac{r}{2c} \ln(1+rc)\big), \quad r \geq 0 \end{align} from Kallenberg and Sztencel (1991)? If so, I still cannot see the path he takes to prove his theorem.

Kallenberg and Sztencel (1991): Some dimension-free features of vector-valued martingales

$\endgroup$
  • 1
    $\begingroup$ It looks to me like Pinelis is talking about Kallenberg and Sztencel's Theorem 5.3. $\endgroup$ – Nate Eldredge Sep 5 '15 at 19:59
13
$\begingroup$

As written in my paper [1], the inequality $$P(f^*>r) \le 2\exp\big(-r^2/2(p-1)\big) $$ in Theorem 3 in [1] for martingales in $\mathcal{X}=L^p$ can be compared with the inequality $$ P(f^*>r) \le C (r+1)\exp(-r^2/2) $$ with an unidentified absolute constant $C>0$, given (as noted in the comment by Nate Eldredge) in Theorem 5.3 of [Kallenberg and Sztencel], proved for $\mathcal{X} = L^2$.

As for the inequality $$P(f^*>r) \le C \frac{1+r}{1+rc}\exp\big(-\frac{r}{2c} \ln(1+rc)\big) $$ with an absolute constant $C>0$, which is inequality (5.2) of [Kallenberg and Sztencel], it can be compared with the inequality in Theorem 2 of [1].

More general results were given in [3].

Also, you wrote: "I still cannot see the path he takes to prove his theorem." Can you specify the steps that seem unclear?

| cite | improve this answer | |
$\endgroup$

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.