# Category Archives: Course on Concentration Inequalities

## Information theory background I

This is the first of a series of two posts with a goal of providing a background on information theory. For today, the itinerary is as follows: we will start with introducing basic information quantities, proceed with proving  tensorization  statements for them — i.e. that they … Continue reading

## McDiarmid’s inequality

In the previous post of the mini-course, we proved Azuma–Hoeffding inequality. We were able to relax the assumption that    were independent but we were dealing only with their sum. Now we are to demonstrate that martingale method lets also to control more general functions    (now of independent arguments). … Continue reading

## Azuma-Hoeffding inequality

In this post we will further expand the setting in which concentration inequalities are stated. Remember that we started from the arithmetical mean where all    were independent. We first showed that if    are sufficiently “light-tailed” (subgaussian), then    concentrates above its mean, with a remarkable special … Continue reading

Posted in Course on Concentration Inequalities | Tagged | 1 Comment

## Subexponential distributions III — Johnson–Lindenstrauss lemma

In this post we will finish our affair with subexponential distributions. We will consider yet another example where they can be applied, an important statement called Johnson–Lindenstrauss lemma. Essentially, it states that    points in a high-dimensional Euclidean space may be projected in a subspace with dimension only … Continue reading

## Subexponential distributions II — A closer look at Bernstein’s inequality

In the previous post we considered subexponential distributions and concentration inequality, called Berstein inequality, for independent sums of subexponential random variables. In this post we will concentrate more thoroughly on the meaning of parameters    and the concept of  deviation zones.

## Subexponential distributions I — Bernstein inequality and quadratic forms

In this post we move on from  subgaussian  distributions to another important class of distributions called  subexponential.  The simplest and most common example of such distributions is chi-square. As usual, we are interested in tail probability bounds for such distributions, and … Continue reading

## Chernoff bounding is good enough

There is a way to demonstrate that Chernoff bounding is — in some sense — optimal, and it relies to the so-called Sanov’s theorem, which controls the empirical distribution (as a random measure) in terms of the Kullback-Leibler divergence from the … Continue reading