Connection between minimum of Frobenius norm of error of low-rank approximation and energy norm of that approximation?

49 Views Asked by At

Suppose we have a Gram matrix $A$ (symmetric, positive semi-definite).

I cannot recall the exact statement, but there was some theorem like:

For a vector $x$ of unit length $||x||_2 = 1$:

$x^* = \arg \min \limits_x || A - \lambda x x^T ||_F = \arg \max \limits_x \sqrt{ x^T A x } = \arg \max \limits_x || x ||_A $, where:

  • $|| A - \lambda x x^T ||_F$ is the Frobenius norm of rank-1 approximation error
  • $|| x ||_A = \sqrt{x^T A x}$ is the energy norm of vector $x$ upon matrix $A$ (or, equivalently, same result, formulated in terms of Rayleigh quotient).

Can you suggest the exact formulation of this result and its name?