Bernstein's Inequality
Bernstein’s inequality is an important concentration inequality. In this post we motivate, state and prove a “maximal inequality” version which I think is clearer than the usual formulation.
Introduction
Concentration inequalities are central to probability theory, mathematical statistics and theoretical machine learning, providing a mathematical framework to the notion that “with enough samples you eventually get the right answer.” More precisely, they provide bounds on the typical deviations that a random variable makes from its expected value. Bernstein’s inequality allows us to control the size of a sum of independent zero-mean random variables where variance and almost sure bounds on the summands are available.
In this post we state and prove Bernstein’s inequality, and also demonstrate its approximate optimality by establishing two different lower bounds. The Julia code for the simulations is available on GitHub.
Motivation
In many applications we need to control the maximum of a collection of random variables. For example, we might be proving uniform convergence of a statistical estimator, establishing consistency for a binary classifier under empirical risk minimization, or controlling the regret of a reinforcement learning algorithm. Understanding how this maximum behaves as a function of the number of variables is essential (Figure 1). In this post we focus on maximal inequalities for sums of independent random variables.
Setup
We propose the following setup which is widely applicable in practice.
-
$X_{i j}$ are real-valued random variables for $1 \leq i \leq n$ and $1 \leq j \leq d$.
-
$X_{1j}, \ldots, X_{n j}$ are independent and identically distributed (i.i.d.) for each $j$.
-
$\E[X_{1j}] = 0$ for each $j$.
-
$\max_{1 \leq j \leq d} \V[X_{1j}] \leq \sigma^2$.
-
$\max_{1 \leq j \leq d} |X_{1j}| \leq M$ almost surely (a.s.).
-
We will provide expectation bounds for the variable $\max_{1 \leq j \leq d} \left| \sum_{i=1}^n X_{i j} \right|$.
A brief discussion is in order. Firstly, the mean-zero property and the variance bound tell us that $\max_{1 \leq j \leq d} \E\big[\left| \sum_{i=1}^n X_{i j} \right|\big] \leq \sqrt{n\sigma^2}$. However in order to put the maximum inside the expectation, we need finer control on the tails of the summands, attained by imposing the almost sure bound. Note that we do not make any assumptions regarding the dependencies between different values of $j$.
Bernstein’s inequality
Now let’s state the main result, a maximal version of Bernstein’s inequality. The proof is provided later in the post, to avoid distracting from the discussion.
Theorem (Bernstein's maximal inequality)
For each $1 \leq j \leq d$, let $X_{1j}, \ldots, X_{n j}$ be i.i.d. real-valued random variables. Suppose $\E[X_{1j}] = 0$ for each $j$, $\max_{1 \leq j \leq d} \V[X_{1j}] \leq \sigma^2$ and $\max_{1 \leq j \leq d} |X_{1j}| \leq M$ a.s. Then $$ \E\left[ \max_{1 \leq j \leq d} \left| \sum_{i=1}^n X_{i j} \right| \right] \leq \sqrt{2 n \sigma^2 \log 2d} + \frac{M}{3} \log 2d. $$Why this formulation?
If you have seen Bernstein’s inequality before, you may notice a few differences between this version and the usual formulation, including:
-
This result is stated for the maximum of $d$ random variables. This is to highlight the dependence of the resulting bound on $d$.
-
This version is stated as an expectation rather than a tail probability to avoid notational complexity, but can be easily strengthened to include the tail bound.
-
The terms on the right hand side are rather complicated and perhaps unfamiliar, but allow us to more directly parse the bound on the maximum. The more standard version of Bernstein’s inequality makes it difficult to read off this value.
Interpreting the bound
The resulting bound of $\sqrt{2 n \sigma^2 \log 2d} + \frac{M}{3} \log 2d$ consists of two terms which are worth discussing separately.
-
The first term is $\sqrt{2 n \sigma^2 \log 2d}$, which depends on $n$ and $\sigma^2$ but not $M$, and has a sub-Gaussian-type dependence on the dimension. This is the bound obtained if we assume that each $X_{1j}$ is $\sigma^2$-sub-Gaussian, and this term corresponds to the central limit theorem for $\frac{1}{\sqrt{n \sigma^2}}\sum_{i=1}^n X_{i j}$.
-
The second term is $\frac{M}{3} \log 2d$ and depends on $M$ but not $n$ or $\sigma^2$. This is a sub-exponential-type tail which captures rare event phenomena associated with bounded random variables.
It is worth remarking at this point that Bennett’s inequality provides a further refinement of Bernstein’s inequality, but the difference is minor in many applications.
Approximate optimality
In this section we provide two explicit examples which show why each of the two terms discussed above are necessary. It is somewhat remarkable that these examples are so easy to find, providing a straightforward demonstration of the near-optimality of Bernstein’s maximal inequality.
Example 1: central limit theorem
Let $X_{i j} = \pm \sigma$ with equal probability be i.i.d. for $1 \leq i \leq n$ and $1 \leq j \leq d$. Note $\E[X_{i j}] = 0$, $\V[X_{i j}] = \sigma^2$ and $|X_{i j}| = \sigma$ a.s., so Bernstein’s inequality gives
\[\begin{align*} \E\left[ \max_{1 \leq j \leq d} \left| \sum_{i=1}^n X_{i j} \right| \right] &\leq \sqrt{2 n \sigma^2 \log 2d} + \frac{\sigma}{3} \log 2d, \end{align*}\]and hence for fixed $d \geq 1$
\[\begin{align*} \limsup_{n \to \infty} \E\left[ \max_{1 \leq j \leq d} \left| \frac{1}{\sqrt{n \sigma^2}} \sum_{i=1}^n X_{i j} \right| \right] &\leq \sqrt{2 \log 2d}. \end{align*}\]However we also have by the central limit theorem that
\[\begin{align*} \frac{1}{\sqrt{n \sigma^2}} \sum_{i=1}^n (X_{i1}, \ldots, X_{id}) \rightsquigarrow (Z_1, \ldots, Z_d) \end{align*}\]as $n \to \infty$, where $Z_j \sim \cN(0,1)$ are i.i.d. So by a Gaussian lower bound in the appendix,
\[\begin{align*} \lim_{n \to \infty} \, \E\left[ \max_{1 \leq j \leq d} \left| \frac{1}{\sqrt{n \sigma^2}} \sum_{i=1}^n X_{i j} \right| \right] = \E\left[ \max_{1 \leq j \leq d} |Z_j| \right] \geq \frac{1}{2} \sqrt{\log d} \end{align*}.\]Thus the first term in Bernstein’s maximal inequality is unimprovable up to constants.
Example 2: Poisson weak convergence
Now let $X_{i j} = M\left(1 - \frac{1}{n}\right)$ with probability $1/n$ and $-\frac{M}{n}$ with probability $1 - 1/n$ be i.i.d. for $1 \leq i \leq n$ and $1 \leq j \leq d$. Note that $\E[X_{i j}] = 0$, $\V[X_{i j}] = \frac{n-1}{n^2} M^2$ and $|X_{i j}| \leq \frac{n-1}{n} M$ a.s., so Bernstein’s inequality gives
\[\E\left[ \max_{1 \leq j \leq d} \left| \sum_{i=1}^n X_{i j} \right| \right] \leq \sqrt{2 M^2 \log 2d} + \frac{M}{3} \log 2d,\]and hence
\[\limsup_{d \to \infty} \limsup_{n \to \infty} \E\left[ \max_{1 \leq j \leq d} \left| \frac{1}{M \log d} \sum_{i=1}^n X_{i j} \right| \right] \leq \frac{1}{3}.\]However also note the binomial distribution limit
\[\begin{align*} \P\left(\sum_{i=1}^n \left(\frac{X_{i j}}{M} + \frac{1}{n}\right) = k\right) &= \frac{n!}{k!(n-k)!} \left(\frac{1}{n}\right)^k \left(1 - \frac{1}{n}\right)^{n-k} \to \frac{1}{e k!} \end{align*}\]as $n \to \infty$. Thus we have the Poisson weak convergence
\[\begin{align*} \frac{1}{M} \sum_{i=1}^n (X_{i1}, \ldots, X_{id}) + (1, \ldots, 1) \rightsquigarrow (Z_1, \ldots, Z_d) \end{align*}\]as $n \to \infty$ where $Z_j \sim \Pois(1)$ are i.i.d. So by a Poisson lower bound in the appendix,
\[\begin{align*} \liminf_{d \to \infty} \lim_{n \to \infty} \E\left[ \max_{1 \leq j \leq d} \left| \frac{\log \log d}{M \log d} \sum_{i=1}^n X_{i j} \right| \right] = \liminf_{d \to \infty} \frac{\log \log d}{\log d} \left( \E\left[\max_{1 \leq j \leq d} Z_j \right] - 1 \right) \geq \frac{1}{6}. \end{align*}\]Hence the second term in Bernstein’s inequality is tight up to a factor of $\log \log d$. This factor diverges so slowly that Bernstein’s inequality is practically optimal in many applications. For example, $\log \log d \geq 6$ already requires ${d > 10^{175}}$, far more than the number of particles in the universe!
References
-
Four lectures on probabilistic methods for data science by Roman Vershynin
-
The University of Oxford’s course on Algorithmic Foundations of Learning, taught by Patrick Rebeschini in 2018
-
Princeton University’s course on Probability in High Dimension, taught by Ramon van Handel in 2021
-
A note on the distribution of the maximum of a set of Poisson random variables by K. M. Briggs, L. Song and T. Prellberg, 2009
-
A note on Poisson maxima by A.C. Kimber, 1983
Appendix: proofs
We begin by proving the main result of this post.
Proof of Bernstein's maximal inequality
We first bound the moment generating function of $X_{i j}$. Let $t > 0$ and note that by the mean-zero property and the variance and almost sure bounds, $$ \begin{align*} \E\left[ e^{t X_{i j}} \right] &= 1 + \sum_{k=2}^\infty \frac{t^k \E[X_{i j}^k]}{k!} \leq 1 + t^2 \sigma^2 \sum_{k=2}^\infty \frac{t^{k-2} M^{k-2}}{k!}. \end{align*} $$ Now since $k! \geq 2 \cdot 3^{k-2}$ for all $k \geq 2$ and $1 + x \leq e^x$ for all $x$, we have for $t < 3/M$ $$ \begin{align*} \E\left[ e^{t X_{i j}} \right] &\leq 1 + \frac{t^2 \sigma^2}{2} \sum_{k=2}^\infty \left( \frac{t M}{3} \right)^{k-2} = 1 + \frac{t^2 \sigma^2/2}{1 - t M/3} \leq \exp\left( \frac{t^2 \sigma^2/2}{1 - t M/3} \right). \end{align*} $$ Now we bound the expected maximum, using Jensen's inequality on the convex logarithm function, to see $$ \begin{align*} \E\left[ \max_{1 \leq j \leq d} \sum_{i=1}^n X_{i j} \right] &\leq \frac{1}{t} \log \E\left[ \exp \max_{1 \leq j \leq d} \sum_{i=1}^n t X_{i j} \right] \\ &\leq \frac{1}{t} \log \E\left[ \sum_{j=1}^d \exp \sum_{i=1}^n t X_{i j} \right] \\ &\leq \frac{1}{t} \log \big( d \, \E\left[ \exp t X_{i j} \right]^n \big) \\ &\leq \frac{1}{t} \log d + \frac{n \sigma^2 t / 2}{1 - Mt/3}. \end{align*} $$ Minimizing the bound using calculus by selecting $t = \frac{2}{2M/3 + \sqrt{2n \sigma^2 / \log d}}$ gives $$ \begin{align*} \E\left[ \max_{1 \leq j \leq d} \sum_{i=1}^n X_{i j} \right] &\leq \sqrt{2 n \sigma^2 \log d} + \frac{M}{3} \log d. \end{align*} $$ Finally we set $X_{i (d+j)} = -X_{i j}$ for $1 \leq j \leq d$ to see that $$ \begin{align*} \E\left[ \max_{1 \leq j \leq d} \left| \sum_{i=1}^n X_{i j} \right| \right] &\leq \sqrt{2 n \sigma^2 \log 2d} + \frac{M}{3} \log 2d. \end{align*} $$Next we prove the Gaussian lower bound used in Example 1.
Lemma (Gaussian lower bound)
Let $Z_1, \ldots, Z_d$ be i.i.d. $\cN(0,1)$ random variables. Then $$ \E\left[ \max_{1 \leq j \leq d} |Z_j| \right] \geq \frac{1}{2} \sqrt{\log d}. $$Proof
For any $t > 0$, we have by Markov's inequality $$ \E\left[ \max_{1 \leq j \leq d} |Z_j| \right] \geq t \, \P\left( \max_{1 \leq j \leq d} |Z_j| \geq t \right) = t \left(1 - \left(1 - \P\left(|Z_j| \geq t \right) \right)^d \right). $$ Now note that by the Gaussian density function and since $s^2 \leq 2(s-t)^2 + 2t^2$, $$ \begin{align*} \P\left(|Z_j| \geq t \right) &= \sqrt\frac{2}{\pi} \int_t^\infty e^{-s^2/2} \diff{s} \geq \sqrt\frac{2}{\pi} e^{-t^2} \int_t^\infty e^{-(s-t)^2} \diff{s} \geq e^{-t^2}. \end{align*} $$ Hence because $1-x \leq e^{-x}$, $$ \E\left[ \max_{1 \leq j \leq d} |Z_j| \right] \geq t \left(1 - \left(1 - e^{-t^2} \right)^d \right) \geq t \left(1 - \exp\left(-d e^{-t^2} \right) \right). $$ Finally set $t = \sqrt{\log d}$ to see $$ \E\left[ \max_{1 \leq j \leq d} |Z_j| \right] \geq \sqrt{\log d} \left(1 - 1/e \right) \geq \frac{1}{2}\sqrt{\log d}. $$Finally we establish the Poisson lower bound used in Example 2.