Suppose that $X_n$ is a martingale with $X_0 = 0$ and $EX^2_n < \infty$. Show that $$P\left(\max_{1\leq m \leq n} {X_m} \geq \lambda\right) \leq \frac{EX^2_n}{EX^2_n+\lambda^2}$$ by using the fact that $(X_n+c)^2$ is a submartinagle.

So far, I have tried to manipulate $P(\max_{1\leq m \leq n} X_m \geq \lambda) $ into the form of $$P\left(\max_{1\leq m \leq n} (X_m + c)^2 \geq \lambda^2 + EX^2_n\right)$$ by finding an appropriate $c$ that yields the inequality, but it does not seem to work. How should I proceed?

The problem is from Durrett Exercise 5.4.5.

  • $\begingroup$ Does this help that for every martingale, $M_i$, we have $\Pr(\max_{i}|M_i|>\lambda)\leq \frac{E|M_n|^p}{\lambda^p}$? where $p\geq 1$ and $\lambda >0$. $\endgroup$ – Math-fun Jan 29 '16 at 14:27
  • $\begingroup$ Yes, so ultimately I want to find the expression in the form of $P(\max_m (X_m+c)^2 \leq \lambda^2)$ that gives the desired bound, but it does not seem to walk. $\endgroup$ – Sam Gates Jan 29 '16 at 14:44

Observe that

$$\max_{1\leq m \leq n} X_m \geq \lambda \Leftrightarrow \max_{1\leq m \leq n} X_m + c \geq \lambda + c \implies \max_{1\leq m \leq n} \lvert X_m + c \rvert \geq \lambda + c$$

$$\max_{1\leq m \leq n} \lvert X_m + c \rvert \geq \lambda + c \Leftrightarrow \max_{1\leq m \leq n} \lvert X_m + c \rvert^2 \geq (\lambda + c)^2$$

The second relation is true if $\lambda + c \geq 0$. For now, suppose that this is the case.

Then you have $$P\{\max_{1\leq m \leq n} X_m \geq \lambda\} \leq P\{\max_{1\leq m \leq n} \lvert X_m + c \rvert^2 \geq (\lambda + c)^2\}$$

Apply Doob's inequality to the RHS to obtain

$$P\{\max_{1\leq m \leq n} \lvert X_m + c \rvert^2 \geq (\lambda + c)^2\} \leq \frac{1}{(\lambda + c)^2}E[\lvert X_n + c \rvert^2] = \frac{1}{(\lambda + c)^2}(E[X_n^2] + c^2)$$

On the last equality I made use of the fact that $E[X_n] = E[X_0] = 0$ (martingale property). Now we want to find a lower bound for this last expression by choosing the optimal $c$. Differentiating this expression with respect to $c$ and equating it to $0$ gives the optimal $c$ as $\frac{E[X_n^2]}{\lambda}$. Substitute this $c$ into $\frac{1}{(\lambda + c)^2}(E[X_n^2] + c^2)$ to finish the exercise.

Note that the optimal $c$ satisfies the assumption we made $\lambda + c \geq 0$ if $\lambda \geq 0$. That should be stated somewhere in the question. However, I think that the $\max$ expressions should run from $0$ instead of $1$, in which case $\lambda \geq 0$ becomes a natural starting point.


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.