Lecture 12 | Convex Optimization I (Stanford)

[@shiv093]

0:23 Experiment design(recap) 8:10 derivation of the dual 10:18 Geometric problems (minimum volume ellipsoid around a set) 35:20 Maximum volume inscribed ellipsoid 48:40 Efficiency of ellipsoidal approximations 1:00:16 Centering 1:04:28 Analytic center of a set of inequalities 1:07:25 analytic center of linear inequalities 1:08:33 Linear discrimination

[@michaelmellinger2324]

50:35 ellipsoids are universal approximators of convex sets 54:55 We don’t care about sums of squares of things. It’s just because we can do it. And that’s the only class I took so far

[@heizilyu]

the fact that we don't assume A to be symmetric and pd (20:00) can be more easily shown if ||Av+b||_2 is expanded and we do symmetric decomposition of the matrix A^TA.

[@MaxWasserman1]

Around @54:00 he talks about justifying e.g. 2-norm on pitch-rate in the objective in helicopter design. He then goes on to talk about actually finding the 'acceptable' convex set C of outcomes (pitch rate, rms, etc) via simulation, survey, etc and suggests we find the maximum volume inscribed ellipsoid E_max = {x || Ax + b||_2

[@shupengwei9419]

10:29 geometric problems

[@mrweisu]

For the outliers problem, just fit a R10 Normal distribution with unknown mean and variance. Sort the likelihood from small to large! Easier than peeling ellipsoids!

[@annawilson3824]

1:05:00

[@toddflanagan5531]

Ch. 8