Writing
I use these notes to work through mathematical ideas from the problem they solve to the final derivation. Every page is native to this site and supports TeX equations, so nothing depends on an external notebook viewer.
Reparameterization gradients
The parameter of interest sits inside a sampling distribution. Two unbiased estimators expose why a change of variables can reduce gradient variance.
Why optimize log likelihood?
A likelihood product and its logarithm share a maximizer. The logarithm gives additive derivatives and stable numerical computation.
Convergence of policy iteration
Why the values produced by policy iteration increase at every step and do not get stuck at a local optimum.
Assumptions in linear regression
What the linearity, independence, normality, and constant-variance assumptions are for, and what still holds when they are violated.
Why the residual variance estimator is unbiased
A matrix derivation that the residual sum of squares divided by n minus p minus 1 is an unbiased estimator of the error variance.