People often fit the model to data by minimizing J(w)+w'Bw where J is the objective function and B is some matrix. Normally people use diagonal or symmetric positive definite matrices for B, but what happens if you use other types?

Here's a Mathematica notebook using Manipulate functionality to let you visualize the shrinkage that happens with different matrices, assuming J(w)'s Hessian is the identity matrix. Drag the locators to modify eigenvectors of B.

One thing to note is that if B is badly conditioned, strange things happen, for instance for some values of w*, they may get pushed away from zero. For negative-definite or indefinite matrices B you may get all w*'s getting shrunk away from zero, or they may get flipped across origin, where w* is argmax_w J(w)

Positive Definite

Indefinite

Negative Definite

## No comments:

Post a Comment