[WIP] The Multiplicative Weights Update Algorithm

is interesting and insane and beautiful and ugly.
A. R. Sricharan
(09 Jun 2023)

Estimated reading time: ~1 minutes with ~176 words.

The Multiplicative Weights Update algorithm! It’s extremely useful, I’ve used it, and I don’t “get” it. If I see one more writeup that motivates the algorithm using expert predictions I’m gonna flip my shit. I’m not a (terminally) online person, and while I understand that it’s useful in the online setting, I. Don’t. Care. Writing up something about it is how I hope to finally understand this goddamn algorithm. There’s way too many ways of seeing that I have read up on, and I have too many things in my mind. I need to offload some of it.

0.1 View 1: Solve LPs approximately

This is the most useful/least intuitive view for me. It’s great that it solves LPs. How does it do it? You want to find an x such that Axb. What you can actually do is solve ptAxptb for any probability distribution p. Great, keep solving such inequalities for different p. This is MWU. After a while, the average of all previous solutions is an approximate solution, Ax¯b+ϵ.

To be finished at some point of time..