Link Search Menu Expand Document

Rayleigh Quotient and Generalized Eigenvalue

Table of contents


1. Representative Norms of Matrix about $\sigma_i$

  1. $l^2$ norm (Spectral norm) : $\lVert A\rVert_2 = \text{max}\dfrac{\lVert A\mathbf{x}\rVert}{\lVert\mathbf{x}\rVert} = \sigma_1$
  2. Frobenius norm : $\lVert A\rVert_F = \sqrt{\sigma_1^2 + \sigma_2^2 + \cdots + \sigma_r^2}$
    • $ \lVert A\rVert_F^2 = \lvert a_{11}\rvert^2 + \lvert a_{12}\rvert^2 + \cdots + \lvert a_{mn}\rvert^2 = \sum_i\sum_j\lvert a_{ij}\rvert^2 $
      : another definition

    • $ \lVert A\rVert_F^2 = tr(A^TA) $
      : $\lambda(AA^T) = \lambda(A^TA) = \sigma(A)^2 $

    • $ \lVert A\rVert_F^2 = \sigma_1^2 + \sigma_2^2 + \cdots + \sigma_r^2 = \lVert\Sigma\rVert_F^2 $
      : $ U, V$ no effect

  3. Nuclear norm : $\lVert A\rVert_N = \sigma_1 + \sigma_2 + \cdots + \sigma_r$
  • $\lVert Q_1A\rVert = \lVert AQ_2\rVert = \lVert A\rVert $
    $\because Q_1AQ_2 = (Q_1U)\Sigma (V^TQ_2) = (Q_1U)\Sigma (Q_2V)^T$ : same singular values

2. Rayleigh Quotient

(1) Definition

\[R(S, \mathbf{x}) = \frac{\mathbf{x}^TS\mathbf{x}}{\mathbf{x}^T\mathbf{x}}\]

$S$ : symmetric (like Hermitian) (①all real eigenvalues, ②orthonormal eigenvectors set)
$\mathbf{x}$ : nonzero

(2) Properties

  • $\max\limits_{\mathbf{x}}R(S, \mathbf{x}) = R(S, \mathbf{q}_1) = \lambda_1$ : the largest eigenvalue
  • $\min\limits_{\mathbf{x}}R(S, \mathbf{x}) = R(S, \mathbf{q}_n) = \lambda_n$ : the smallest eigenvalue
  • critical point $\mathbf{x}=\mathbf{q}_k$ of $R(S, \mathbf{x})$, $\implies R(S, \mathbf{q}_k)=\lambda_k$ : corresponding eigenvalue
  • $S$ : symmetric $\implies S=A^TA \implies \lambda_1(S)=\sigma_1(A)$ : $\text{maxEV}(S=A^TA) = \text{maxSV}^2(A)$

(3) Meaning

EV problem of $S$ $\xrightarrow{\text{change}}$ optimization problem of $R(S,\mathbf{x})$

3. Generalized Rayleigh Quotient

(1) Definition

\[R_M(S, \mathbf{x}) = \frac{\mathbf{x}^TS\mathbf{x}}{\mathbf{x}^TM\mathbf{x}}\]

$S$ : symmetric (like Hermitian) (①all real eigenvalues, ②orthonormal eigenvectors set)
$M$ : (symmetric) positive definite ($\implies$ invertible)
$\mathbf{x}$ : nonzero

(2) Wish

\[\begin{align*} \text{optimization problem of } R_M(S,\mathbf{x}) &\xrightarrow{\text{change}} \text{ EV problem of symmetric } S' \\ R_M(S,\mathbf{x}) &\xrightarrow{\text{change}} R_I(S',\mathbf{x})) \end{align*}\]

(3) Solution

\[R_M(S, \mathbf{x}) = R(H, \mathbf{y})\]

$ H = M^{-1/2}SM^{-1/2} $ : symetric and similar to $S$
$ M = Q\Lambda Q^T \implies M^{1/2} = Q\Lambda^{1/2}Q^T \ $ if $\Lambda > 0$
$ \mathbf{x} = M^{-1/2}\mathbf{y} $

4. Generalized Eigenvalue, Eigenvector

(1) Definition

\[S\mathbf{x} = \lambda M \mathbf{x}\]

$S$ : symetric
$M$ : (symetric) positive definite

(2) Solution

\[(\lambda, \mathbf{x}) S\mathbf{x}=\lambda M\mathbf{x} \iff (\lambda, \mathbf{y})H\mathbf{y} = \lambda \mathbf{y}\]

$ H = M^{-1/2}SM^{-1/2} $ : symetric and similar to $S$, (positive definite if $S$ is so)
$ \lbrace \mathbf{y}_i \rbrace $ : orthonormal eigenvectors set
$ \lbrace \mathbf{x}_i \rbrace = \lbrace M^{-1/2}\mathbf{y}_i \rbrace $ : $M$-orthogonal eigenvectors set

(3) Properties

$ \lbrace \mathbf{x}_i \rbrace $ : M-orthogonal ($\mathbf{x}_1^T M \mathbf{x}_2=0$)