Interactive demo · Foundations of Deep Learning · Appendix A

Eigenvectors and a matrix as a linear map

A 2×2 matrix A moves every point w of the plane to Aw; the unit circle goes to an ellipse. An eigenvector is a direction that A only stretches (by its eigenvalue λ), Au = λu (A.29). For a symmetric matrix the eigenvectors are orthogonal and are the axes of the ellipse, and the sign of wTAw is decided by the eigenvalues (A.45). For a non-symmetric matrix none of this holds: try the presets.

The matrix A = [[a, b], [c, d]]

a b c d

Probe direction w = (cos α, sin α)

Try these

Left: dashed grey, the unit circle; blue, its image; red and green, the eigenvector directions with the vectors u and λu; dotted blue, the axes of the image ellipse (singular vectors times singular values); white, the probe w and its image Aw. Right: wTAw as w goes round half the circle; its maximum and minimum are the eigenvalues of the symmetric part (A + AT)/2 (dashed), which equal those of A (dotted) only when A is symmetric.

In the lecture notes