Hoeffding's inequality

In probability theory, Hoeffding's inequality provides an upper bound on the probability that the sum of bounded independent random variables deviates from its expected value by more than a certain amount. Hoeffding's inequality was proven by Wassily Hoeffding in 1963.
Hoeffding's inequality is a special case of the Azuma–Hoeffding inequality and McDiarmid's inequality. It is similar to the Chernoff bound, but tends to be less sharp, in particular when the variance of the random variables is small. It is similar to, but incomparable with, one of Bernstein's inequalities.

Statement

Let be independent random variables such that almost surely. Consider the sum of these random variables,
Then Hoeffding's theorem states that, for all,
Here is the expected value of.
Note that the inequalities also hold when the have been obtained using sampling without replacement; in this case the random variables are not independent anymore. A proof of this statement can be found in Hoeffding's paper. For slightly better bounds in the case of sampling without replacement, see for instance the paper by.

Generalization

Let be independent observations such that and. Let. Then, for any,

Special Case: Bernoulli RVs

Suppose and for all i. This can occur when X_i are independent Bernoulli random variables, though they need not be identically distributed. Then we get the inequality
or equivalently,
for all. This is a version of the additive Chernoff bound which is more general, since it allows for random variables that take values between zero and one, but also weaker, since the Chernoff bound gives a better tail bound when the random variables have small variance.

General case of bounded from above random variables

Hoeffding's inequality can be extended to the case of bounded from above random variables.
Let be independent random variables such that and almost surely.
Denote by
Hoeffding's inequality for bounded from above random variables states that for all,
In particular, if for all,
then for all,

General case of sub-Gaussian random variables

The proof of Hoeffding's inequality can be generalized to any sub-Gaussian distribution. Recall that a random variable is called sub-Gaussian, if
for some. For any bounded variable, for for some sufficiently large. Then for all so taking yields
for. So every bounded variable is sub-Gaussian.
For a random variable, the following norm is finite if and only if is sub-Gaussian:
Then let be independent sub-Gaussian random variables, the general version of the Hoeffding's inequality states that:
where c > 0 is an absolute constant.
Equivalently, we can define sub-Gaussian distributions by variance proxy, defined as follows. If there exists some such that for all, then is called a variance proxy, and the smallest such is called the optimal variance proxy, and denoted by. In this form, Hoeffding's inequality states

Proof

We quote:
The proof of Hoeffding's inequality then follows similarly to concentration inequalities like Chernoff bounds.
This proof easily generalizes to the case for sub-Gaussian distributions with variance proxy.

Usage

Confidence intervals

Hoeffding's inequality can be used to derive confidence intervals. We consider a coin that shows heads with probability and tails with probability. We toss the coin times, generating samples . The expected number of times the coin comes up heads is. Furthermore, the probability that the coin comes up heads at least times can be exactly quantified by the following expression:
where is the number of heads in coin tosses.
When for some, Hoeffding's inequality bounds this probability by a term that is exponentially small in :
Since this bound holds on both sides of the mean, Hoeffding's inequality implies that the number of heads that we see is concentrated around its mean, with exponentially small tail.
Thinking of as the "observed" mean, this probability can be interpreted as the level of significance for a confidence interval around of size 2:
Finding for opposite inequality sign in the above, i.e. that violates inequality but not equality above, gives us:
Therefore, we require at least samples to acquire a -confidence interval.
Hence, the cost of acquiring the confidence interval is sublinear in terms of confidence level and quadratic in terms of precision. Note that there are more efficient methods of estimating a confidence interval.