Bernstein Inequalities (probability Theory)

In probability theory, Bernstein inequalities give bounds on the probability that the sum of random variables deviates from its mean. In the simplest case, let ''X''₁, ..., ''X''_''n'' be independent Bernoulli random variables taking values +1 and −1 with probability 1/2 (this distribution is also known as the Rademacher distribution), then for every positive $\varepsilon$,

:$\mathbb\left (\left, \frac\sum_^n X_i\ > \varepsilon \right ) \leq 2\exp \left (-\frac \right).$

Bernstein inequalities were proved and published by Sergei Bernstein in the 1920s and 1930s.J.V.Uspensky, "Introduction to Mathematical Probability", McGraw-Hill Book Company, 1937 Later, these inequalities were rediscovered several times in various forms. Thus, special cases of the Bernstein inequalities are also known as the Chernoff bound, Hoeffding's inequality and Azuma's inequality.

Some of the inequalities

1. Let $X_1, \ldots, X_n$ be independent zero-mean random variables. Suppose that $, X_i, \leq M$ almost surely, for all $i.$ Then, for all positive $t$,

:$\mathbb \left (\sum_^n X_i \geq t \right ) \leq \exp \left ( -\frac \right ).$

2. Let $X_1, \ldots, X_n$ be independent zero-mean random variables. Suppose that for some positive real $L$ and every integer $k \geq 2$,

: \mathbb \left X_i^k \right , \right \leq \frac \mathbb \left _i^2\rightL^ k!

Then

:$\mathbb \left (\sum_^n X_i \geq 2t \sqrt \right ) < \exp(-t^2), \qquad \text\quad 0 \leq t \leq \frac\sqrt.$

3. Let $X_1, \ldots, X_n$ be independent zero-mean random variables. Suppose that

: \mathbb \left X_i^k \right , \right \leq \frac \left(\frac\right)^

for all integer $k \geq 4.$ Denote

: A_k = \sum \mathbb \left X_i^k\right

Then,

: \mathbb \left( \left, \sum_^n X_j - \frac \\geq \sqrt \, t \left 1 + \frac \right\right) < 2 \exp (- t^2), \qquad \text \quad 0 < t \leq \frac.

4. Bernstein also proved generalizations of the inequalities above to weakly dependent random variables. For example, inequality (2) can be extended as follows. Let $X_1, \ldots, X_n$ be possibly non-independent random variables. Suppose that for all integers $i>0$,

:\begin
\mathbb \left. \left X_1, \ldots, X_ \right &= 0, \\
\mathbb \left. \left X_1, \ldots, X_ \right &\leq R_i \mathbb \left X_i^2 \right \\
\mathbb \left. \left X_1, \ldots, X_ \right &\leq \tfrac \mathbb \left. \left X_1, \ldots, X_ \right L^ k!
\end

Then

:$\mathbb \left( \sum_^n X_i \geq 2t \sqrt \right) < \exp(-t^2), \qquad \text\quad 0 < t \leq \frac \sqrt.$

More general results for martingales can be found in Fan et al. (2015).

Proofs

The proofs are based on an application of Markov's inequality to the random variable

:$\exp \left ( \lambda \sum_^n X_j \right ),$

for a suitable choice of the parameter $\lambda > 0$.

Generalizations

The Bernstein inequality could be generalized to Gaussian random matrices. Let $G = g^H A g + 2 \operatorname(g^H a)$ be a scalar where $A$ is a complex Hermitian matrix and $a$ is complex vector of size $N$. The vector $g \sim \mathcal(0,I)$ is a Gaussian vector of size $N$. Then for any $\sigma \geq 0$, we have

:$\mathbb \left( G \leq \operatorname(A) - \sqrt\sqrt - \sigma s^-(A) \right) < \exp(-\sigma),$
where $\operatorname$ is the vectorization operation. Also, $s^- (A) = \max(\lambda_(-A),0)$ and $\lambda_(-A)$ is the maximum eigenvalue of $-A$. The proof is detailed in here. Another similar inequality is also formulated as
:$\mathbb \left( G \geq \operatorname(A) + \sqrt\sqrt + \sigma s^+(A) \right) < \exp(-\sigma),$
where $s^+(A) = \max(\lambda_(A),0)$ and $\lambda_(A)$ is the maximum eigenvalue of $A$.

See also

*Concentration inequality - a summary of tail-bounds on random variables.
*Hoeffding's inequality

References

(according to: S.N.Bernstein, Collected Works, Nauka, 1964)

A modern translation of some of these results can also be found in

{{DEFAULTSORT:Bernstein Inequalities (Probability Theory)
Probabilistic inequalities