Mix the first eight principal components of the Olivetti faces, watch reconstruction improve with k, and walk the (z1, z2) plane. Escape exits when embedded.
Mean face μ
Reconstruction · mean + Σ zkvk
Centred residual ‖(x − μ) − (x̂ − μ)‖² / ‖x − μ‖², comparable with the eigenvalue tail.
Tail uses the first 100 components only (hits 0 at k = 100); the pixel residual stays above 0 because the truncated basis cannot recover the full image.