The Living Thing / Notebooks :

Measure concentration inequalities

On being 80% sure you are only 20% wrong

Welcome to the probability inequality mines!

When something in your process (measurement, estimation) means that you can be pretty sure that a whole bunch of your stuff is damn likely to be somewhere in particular.

This is basic workhorse stuff in univariate probability, and turns out to be yet more essential in multivariate matrix probability, as seen in matrix factorisation, compressive sensing, PAC-bounds and suchlike.


Overviews include

Foundational but impenetrable things I won't read right now: Talagrand's opus that is commonly credited with kicking off the modern fad especially with the chaining method. (Tala95)

Finite sample bounds

These are everywhere in statistics. Special attention will be given here to finite-sample inequalities. Asymptotic normality is so last season. These days we care about finite sample performance, and asymptotic results don't help us there. Apparently I can construct useful bounds using concentration inequalities? One suggested keyword to disambiguate: Ahlswede-Winterfeld bounds?

Basic inequalities

the classics












For the Gaussian distribution. Filed there.

Martingale type



Let us copy from wikipedia:

Heuristically: if we pick complex numbers , and add them together, each multiplied by jointly independent random signs , then the expected value of the sum's magnitude is close to .

Let i.i.d. random variables with for , i.e., a sequence with Rademacher distribution. Let and let . Then

for some constants . It is a simple matter to see that when , and when )0 < p \le 2).

Empirical process theory

Large deviation inequalities, empirical process inequalities, Talagrand chaining method. Berry-Esseen bound.

Matrix Chernoff bounds

Nikhil Srivastava's Discrepancy, Graphs, and the Kadison-Singer Problem has an interesting example of bounds via discrepancy theory (and only indirectly probability). Gros11 is also readable, and gives quantum-mechanical results (i.e. the matrices are complex-valued).

Trop15 summarises:

In recent years, random matrices have come to play a major role in computational mathematics, but most of the classical areas of random matrix theory remain the province of experts. Over the last decade, with the advent of matrix concentration inequalities, research has advanced to the point where we can conquer many (formerly) challenging problems with a page or two of arithmetic.

To read