EasyDeepLearn

Derive PCA — what does it optimize?

hard

Answer

  • Given centered X, find orthonormal w that maximizes Var(Xw)  =  w\operatorname{Var}(\mathrm{Xw})\; = \;w' Σ w subject to ||w||=1.
  • Solution: eigenvector of Σ = X'X/n corresponding to the largest eigenvalue.
  • Subsequent components: next largest eigenvalue, orthogonal.
  • Equivalently: minimizes squared reconstruction error over rank-k subspaces (Eckart-Young).
  • Two views, same answer.
Check yourself — multiple choice
  • Random
  • Max Var(Xw)\operatorname{Var}(\mathrm{Xw}) with ||w||=1 → top eigenvector of Σ; equivalently minimizes rank-k reconstruction error (Eckart-Young)
  • Same as clustering
  • Not derivable

PCA: max variance ≡ min reconstruction error; top eigenvectors of Σ.

#dimensionality-reduction#pca

Practise Unsupervised Learning

214 interview questions in this topic.

Related questions