An Application of Doob's Martingale Inequality

Mar 23, 2018

Problem

Suppose we have a sequence of i.i.d. Gaussian random variables $X_t$'s with bounded variance. Let $S_n = \sum_{t = 1}^n X_t$. How can we bound the probability of

\begin{equation} \label{eq:problem} \left \{ \max_{1 \leq t \leq n} S_t > \eps \right \}? \end{equation}

Sub-gaussian random variables

To make this problem more realistic, it is always safe to loose Gaussian random variables to zero mean sub-gaussian random variables.

Intuitively, sub-gaussian r.v.‘s have tails decreasing as fast as those of Gaussian r.v.‘s. Hence, most of the inequalities related to Gaussian r.v.‘s can be safely applied to sub-gaussians without any modifications! You are suggested to refer to Subgaussian random variables: An expository note for more details. Here, we assume $X_t$ is a $\sigma$-subgaussian r.v. which is comparable to a Gaussian r.v. with variance $\sigma^2$.

An elementary property of $\sigma$-subgaussian random variable $X$ is

\begin{equation} \label{eq:ele-prop-subgaussian} \E[ \exp( \lambda X ) ] \leq \exp( \lambda^2 \sigma^2 / 2 ). \end{equation}

Also, it is useful to know the following facts:

  • $S_t$ is $\sqrt{t} \sigma$-subgaussian,
  • and $\Pr( X > \eps ) \leq \exp\left( - \frac{ \eps^2 }{2\sigma^2} \right)$.

A naive way

Obviously, we can use a union bound to give a naive bound.

For each $t$, since $S_t$ is $\sqrt{t} \sigma$-subgaussian we have $\Pr( S_t > \eps ) \leq \exp\left( - \frac{\eps^2}{2 t \sigma^2} \right)$. Via a union bound, we get

\begin{equation*} \Pr\left( \max_{1 \leq t \leq n} S_t > \eps \right) \leq \sum_{t = 1}^n \exp\left( - \frac{\eps^2}{2 t \sigma^2 } \right), \end{equation*}

from which we can see the upper bound is no less then

\begin{equation} \label{eq:union-bound} n \exp\left( - \frac{\eps^2}{2 n \sigma^2 } \right). \end{equation}

This bound is very loose. Later, you will see the reason.

An alternative way

Doob’s martingale inequality

The formal statement of Doob’s martingale inequality can be found in 1. We restate it in the following.

Suppose the sequence $T_1, \dots, T_n$ is a submartingale, taking non-negative values. Then it holds that

\begin{equation} \label{eq:doob-inequality} \Pr\left( \max_{1\leq t \leq n} T_t > \eps \right) \leq \frac{ \E[T_n] }{\eps}. \end{equation}

With this tool in mind, we are now ready to bound $\eqref{eq:problem}$ in another way.

Using standard Chernoff’s method, for any $\lambda > 0$, we have

\begin{align*} \Pr\left( \max_{1 \leq t \leq n} S_t > \eps \right) & = \Pr\left( \max_{1 \leq t \leq n} \exp(\lambda S_t) > \exp( \lambda \eps) \right ). \end{align*}

Since $\E[ \exp( \lambda X_t ) ] \geq \exp( \E[ \lambda X_t )] = 1$, sequence $ \exp(\lambda S_1), \dots, \exp(\lambda S_t)$ is a submartingale. (This is a good exercise. You can validate it by yourself.) By $\eqref{eq:doob-inequality}$, we further have

\begin{align*} \Pr\left( \max_{1 \leq t \leq n} S_t > \eps \right) & \leq \frac{ \E[\exp(\lambda S_n )] }{ \exp(\lambda \eps) } \\
& = \frac{ \prod_{t = 1}^n \E[ \exp(\lambda X_t )] }{ \exp(\lambda \eps) } \\
& \leq \exp\left( \frac{\lambda^2 \sigma^2 n}{2} - \lambda \eps \right), \end{align*}

where the second equality is due to the mutual indenpendency of $X_t$'s, and the last inequality is due to $\eqref{eq:ele-prop-subgaussian}$.

The minimum is achieved when $\lambda = \frac{\eps}{ \sigma^2 n}$. So we finally get

$$ \Pr\left( \max_{1 \leq t \leq n} S_t > \eps \right) \leq \exp\left( - \frac{\eps^2}{2 n \sigma^2 } \right), $$

which is only one $n$-th of $\eqref{eq:union-bound}$!

Note that Lemma 2 in 2 gives the same statement. However, I think it is more direct to use Doob’s martingale inequality.


  1. Doob’s martingale inequality , Wikipedia. ↩︎

  2. Shengjia Zhao, Enze Zhou, Ashish Sabharwal, and Stefano Ermon. Adaptive Concentration Inequalities for Sequential Decision Problems . In NIPS, pages 1343–1351, 2016. ↩︎

mathprobability

Generate New Posts by Shell Scripts

Test

comments powered by Disqus