I Calculus of Variations on Kullback-Liebler Divergence

Master1022
Messages
590
Reaction score
116
TL;DR Summary
How to use calculus of variations on KL-divergence
Hi,

This isn't a homework question, but a side task given in a machine learning class I am taking.

Question: Using variational calculus, prove that one can minimize the KL-divergence by choosing ##q## to be equal to ##p##, given a fixed ##p##.

Attempt:

Unfortunately, I have never seen calculus of variations (it was suggested that we teach ourselves). I have been trying to watch some videos online, but I mainly just see references to Euler-Lagrange equations which I don't think are of much relevance here (please correct me if I am wrong) and not much explanation of the functional derivatives.

Nonetheless, I think this shouldn't be too hard, but am struggling to understand how to use the tools.

If we start with the definition of the KL-divergence we get:
\text{KL}[p||q] = \int p(x) log(\frac{p(x)}{q(x)}) dx = I

Would it be possible for anyone to help me get started on the path? I am not sure how to proceed really after I write down ## \frac{\delta I}{\delta q} ##?

Thanks in advance
 
Mathematics news on Phys.org
Euler Lagrange is what you want, but you also have to worry about the conditions that you have on q that come from it being a probability distribution, namely that the integral is 1 and it's always nonnegative. I think the integral constraint is the important part

http://liberzon.csl.illinois.edu/teaching/cvoc/node38.html

Has some notes on how to add constraints to the euler Lagrange equations.
 
Maybe you should start with fixed p, and then try to get optimal q to maximize or minimize KL.
 
Insights auto threads is broken atm, so I'm manually creating these for new Insight articles. In Dirac’s Principles of Quantum Mechanics published in 1930 he introduced a “convenient notation” he referred to as a “delta function” which he treated as a continuum analog to the discrete Kronecker delta. The Kronecker delta is simply the indexed components of the identity operator in matrix algebra Source: https://www.physicsforums.com/insights/what-exactly-is-diracs-delta-function/ by...
Fermat's Last Theorem has long been one of the most famous mathematical problems, and is now one of the most famous theorems. It simply states that the equation $$ a^n+b^n=c^n $$ has no solutions with positive integers if ##n>2.## It was named after Pierre de Fermat (1607-1665). The problem itself stems from the book Arithmetica by Diophantus of Alexandria. It gained popularity because Fermat noted in his copy "Cubum autem in duos cubos, aut quadratoquadratum in duos quadratoquadratos, et...
I'm interested to know whether the equation $$1 = 2 - \frac{1}{2 - \frac{1}{2 - \cdots}}$$ is true or not. It can be shown easily that if the continued fraction converges, it cannot converge to anything else than 1. It seems that if the continued fraction converges, the convergence is very slow. The apparent slowness of the convergence makes it difficult to estimate the presence of true convergence numerically. At the moment I don't know whether this converges or not.
Back
Top