317 lines
9.9 KiB
TeX
317 lines
9.9 KiB
TeX
\documentclass[../../script.tex]{subfiles}
|
|
% !TEX root = ../../script.tex
|
|
|
|
\begin{document}
|
|
\section{Eigenvalue problems}
|
|
|
|
\begin{defi}
|
|
Let $A \in \field^{n \times n}$. Then $\lambda \in \field$ is called an eigenvalue of $A$, if
|
|
\[
|
|
\exists v \in \field^n, ~v \ne 0: ~~Av = \lambda v
|
|
\]
|
|
Such a vector $v$ is called eigenvector. We call
|
|
\[
|
|
\set[Av = \lambda v]{v \in \field^n} =: E_{\lambda}
|
|
\]
|
|
eigenspace belonging to $\lambda$.
|
|
\end{defi}
|
|
|
|
\begin{eg}
|
|
Let
|
|
\[
|
|
A = \begin{pmatrix}
|
|
2 & 1 & -1 \\
|
|
0 & 1 & 0 \\
|
|
0 & 0 & 1
|
|
\end{pmatrix}
|
|
\]
|
|
Then
|
|
\begin{align*}
|
|
A \cdot \begin{pmatrix}
|
|
1 \\ 0 \\ 0
|
|
\end{pmatrix}
|
|
&= \begin{pmatrix}
|
|
2 \\ 0 \\ 0
|
|
\end{pmatrix}
|
|
= 2 \cdot \begin{pmatrix}
|
|
1 \\ 0 \\ 0
|
|
\end{pmatrix} \\
|
|
A \cdot \begin{pmatrix}
|
|
1 \\ -1 \\ 0
|
|
\end{pmatrix}
|
|
&= \begin{pmatrix}
|
|
1 \\ -1 \\ 0
|
|
\end{pmatrix}
|
|
= 1 \cdot \begin{pmatrix}
|
|
1 \\ -1 \\ 0
|
|
\end{pmatrix} \\
|
|
A \cdot \begin{pmatrix}
|
|
1 \\ 0 \\ 1
|
|
\end{pmatrix}
|
|
&= \begin{pmatrix}
|
|
1 \\ 0 \\ 1
|
|
\end{pmatrix}
|
|
= 1 \cdot \begin{pmatrix}
|
|
1 \\ 0 \\ 1
|
|
\end{pmatrix}
|
|
\end{align*}
|
|
The eigenspaces are
|
|
\begin{align*}
|
|
E_2 &= \set[\kappa \in \realn]{\kappa \cdot \begin{pmatrix}1\\0\\0\end{pmatrix}} \\
|
|
E_1 &= \set[\kappa, \rho \in \realn]{\kappa \cdot \begin{pmatrix}1\\-1\\0\end{pmatrix} + \rho \cdot \begin{pmatrix}1\\0\\1\end{pmatrix}}
|
|
= \spn\set{\begin{pmatrix}1\\-1\\0\end{pmatrix}, \begin{pmatrix}1\\0\\1\end{pmatrix}}
|
|
\end{align*}
|
|
\end{eg}
|
|
|
|
\begin{rem}
|
|
The eigenspace to an eigenvalue $\lambda$ is a linear subspace.
|
|
\end{rem}
|
|
|
|
\begin{rem}
|
|
We want to find $\lambda \in \field$, $v \in \field^n$ such that
|
|
\[
|
|
Av = \lambda v \iff (\underbrace{A - \lambda I}_{\in \field^{n \times n}}) v = 0
|
|
\]
|
|
If $(A - \lambda I)$ is invertible, then $v = 0$. So the interesting case is when $(A - \lambda I)$ not invertible.
|
|
\[
|
|
(A - \lambda I) \text{ not invertible} \iff \det(A - \lambda I) = 0
|
|
\]
|
|
This determinant is called the characteristic polynomial. This polynomial has degree $n$, and the eigenvalues are the roots of that polynomial.
|
|
So let $\lambda$ be an eigenvalue of $A$, then
|
|
\[
|
|
(A - \lambda I) v = 0
|
|
\]
|
|
is a linear equation system for the components of $v$.
|
|
\end{rem}
|
|
|
|
\begin{eg}
|
|
Let
|
|
\[
|
|
A = \begin{pmatrix}
|
|
0 & 1 \\ -1 & 0
|
|
\end{pmatrix}
|
|
\in \cmpln^{2 \times 2}
|
|
\]
|
|
The characteristic polynomial is
|
|
\[
|
|
\det(A - \lambda I) = \begin{vmatrix}
|
|
-\lambda & 1 \\ -1 & -\lambda
|
|
\end{vmatrix}
|
|
= \lambda^2 + 1
|
|
\]
|
|
Its roots are
|
|
\begin{multicols}{2}
|
|
\noindent
|
|
\[ \lambda_1 = i \]
|
|
\[ \lambda_2 = -i \]
|
|
\end{multicols}
|
|
\noindent To find the eigenvector belonging to $\lambda_1$, we declare $v_1 = (x, y) \in \cmpln^2$ and solve the linear equation system
|
|
\begin{multicols}{2}\noindent
|
|
\[ (A - \lambda_1 I) v_1 = 0 \]
|
|
\begin{align*}
|
|
-ix + 1y &= 0 \\
|
|
-1x - iy &= 0 \\
|
|
\end{align*}
|
|
\end{multicols}
|
|
\noindent It has the solutions $x = -i$ and $y = 1$, so
|
|
\[
|
|
v_1 = \begin{pmatrix}
|
|
-i \\ 1
|
|
\end{pmatrix}
|
|
\]
|
|
Doing the same for $v_2$ yields
|
|
\[
|
|
v_2 = \begin{pmatrix}
|
|
i \\ 1
|
|
\end{pmatrix}
|
|
\]
|
|
It is to be noted that the eigenvectors aren't unique (multiples of eigenvectors are also eigenvectors).
|
|
\end{eg}
|
|
|
|
\begin{eg}
|
|
Let $D$ be a diagonal matrix, with the diagonal entries $\lambda_j$. Then
|
|
\[
|
|
\det(D - \lambda I) = \begin{vmatrix}
|
|
\lambda_1 - \lambda & & & \\
|
|
& \lambda_2 - \lambda & & \\
|
|
& & \ddots & \\
|
|
& & & \lambda_n - \lambda \\
|
|
\end{vmatrix}
|
|
\]
|
|
The roots (eigenvalues) are $\lambda_1, \lambda_2, \cdots, \lambda_n$, and the eigenvectors are $De_i = \lambda_i e_i$.
|
|
\end{eg}
|
|
|
|
\begin{defi}
|
|
$A \in \field^{n \times n}$ is called diagonalizable if there exists a basis of $\field^n$ that consists of eigenvectors.
|
|
\end{defi}
|
|
|
|
\begin{thm}
|
|
A matrix $A \in \field^{n \times n}$ is diagonalizable, if and only if there exists a diagonal matrix $D$ and a invertible matrix $T$ such that
|
|
\[
|
|
D = \inv{T}AT
|
|
\]
|
|
\end{thm}
|
|
\begin{proof}
|
|
Let $e_1, e_2, \cdots, e_n$ be the canonical basis of $\field^n$. Define $TD\inv{T} = A$, and let $\lambda_1, \cdots, \lambda_n$ be the diagonal entries of $D$.
|
|
Then we know that
|
|
\begin{equation}
|
|
De_i = \lambda_ie_i, ~~\forall i \in \set{1, \cdots n}
|
|
\end{equation}
|
|
Since $T$ is invertible, the $Te_1, \cdots Te_n$ form a basis.
|
|
\begin{equation}
|
|
A(Te_i) = T(\inv{T}AT)e_i = TDe_i = T\lambda_i e_i = \lambda_i (Te_i)
|
|
\end{equation}
|
|
Therefore $Te_i$ is an eigenvector of $A$ to the eigenvalue $\lambda_i$. Now let $v_1, \cdots, v_n$ be a basis of $\field^n$ and
|
|
\begin{equation}
|
|
Av_i = \lambda_iv_i, ~~\lambda_1, \cdots, \lambda_n \in \field^n
|
|
\end{equation}
|
|
Write write $v_1, \cdots, v_n$ as the columns of a matrix, therefore
|
|
\begin{subequations}
|
|
\begin{equation}
|
|
T = (v_1, v_2, \cdots, v_n)
|
|
\end{equation}
|
|
\begin{equation}
|
|
D = \begin{pmatrix}
|
|
\lambda_1 & & \\
|
|
& \vdots & \\
|
|
& & \lambda_n
|
|
\end{pmatrix}
|
|
\end{equation}
|
|
\end{subequations}
|
|
So $Te_i = v_i$, and thus
|
|
\begin{equation}
|
|
A(Te_i) = Av_i = \lambda_iv_i = \lambda_i(Te_i) = T\lambda_ie_i = TDe_i
|
|
\end{equation}
|
|
This means that $(AT - TD)e_i = 0$, $\forall i \in \set{1, \cdots, n}$.
|
|
\begin{equation}
|
|
\implies AT = TD
|
|
\end{equation}
|
|
$T$ is invertible (left as an exercise for the reader), and thus
|
|
\begin{equation}
|
|
\implies \inv{T}AT = D
|
|
\end{equation}
|
|
\end{proof}
|
|
|
|
\begin{eg}
|
|
\begin{enumerate}[(i)]
|
|
\item Let
|
|
\[
|
|
A = \begin{pmatrix}
|
|
0 & 1 \\ -1 & 0
|
|
\end{pmatrix}
|
|
\]
|
|
The eigenvalues and eigenvectors are
|
|
\begin{multicols}{2}\noindent
|
|
\[
|
|
A \cdot \begin{pmatrix}
|
|
-i \\ 1
|
|
\end{pmatrix}
|
|
=
|
|
i \begin{pmatrix}
|
|
-i \\ 1
|
|
\end{pmatrix}
|
|
\]
|
|
\[
|
|
A \cdot \begin{pmatrix}
|
|
i \\ 1
|
|
\end{pmatrix}
|
|
=
|
|
-i \begin{pmatrix}
|
|
i \\ 1
|
|
\end{pmatrix}
|
|
\]
|
|
\end{multicols}
|
|
\noindent Therefore
|
|
\[
|
|
T = \begin{pmatrix}
|
|
-i & i \\ 1 & 1
|
|
\end{pmatrix}
|
|
\]
|
|
which has the inverse
|
|
\[
|
|
\inv{T} = \frac{1}{2} \begin{pmatrix}
|
|
i & 1 \\ -i & 1
|
|
\end{pmatrix}
|
|
\]
|
|
Finally,
|
|
\[
|
|
\inv{T}AT = \frac{1}{2}
|
|
\begin{pmatrix}
|
|
i & 1 \\ -i & 1
|
|
\end{pmatrix}
|
|
\begin{pmatrix}
|
|
1 & 1 \\ i & -i
|
|
\end{pmatrix}
|
|
= \frac{1}{2}
|
|
\begin{pmatrix}
|
|
2i & 0 \\ 0 & -2i
|
|
\end{pmatrix}
|
|
=
|
|
\begin{pmatrix}
|
|
i & 0 \\ 0 & -i
|
|
\end{pmatrix}
|
|
\]
|
|
This is a diagonal matrix, therefore $A$ is diagonalizable.
|
|
|
|
\item The matrix
|
|
\[
|
|
\begin{pmatrix}
|
|
0 & 1 \\ 0 & 0
|
|
\end{pmatrix}
|
|
\]
|
|
is not diagonalizable since its only eigenvector is $(1, 0)^T$.
|
|
\end{enumerate}
|
|
\end{eg}
|
|
|
|
\begin{rem}
|
|
For diagonal matrices the following is true
|
|
\[
|
|
\begin{pmatrix}
|
|
\lambda_1 & & & \\
|
|
& \lambda_2 & & \\
|
|
& & \ddots & \\
|
|
& & & \lambda_3 \\
|
|
\end{pmatrix}^k
|
|
=
|
|
\begin{pmatrix}
|
|
\lambda_1^k & & & \\
|
|
& \lambda_2^k & & \\
|
|
& & \ddots & \\
|
|
& & & \lambda_3^k \\
|
|
\end{pmatrix}
|
|
\]
|
|
If $\inv{T}AT = D$ (where $D$ is a diagonal matrix), then
|
|
\[
|
|
D^k = (\inv{T}AT)^k = \underbrace{\inv{T}AT \cdot \inv{T}AT \cdot \cdots}_{k \text{ times}} = \inv{T}A^kT
|
|
\]
|
|
\[
|
|
\implies A^k = TD^k\inv{T}
|
|
\]
|
|
\end{rem}
|
|
|
|
\begin{thm}
|
|
Let $A \in \realn^{n \times n}$ be a symmetric matrix, i.e. $A = A^T$. (Or if $A \in \cmpln^{n \times n}$ a self-adjoint matrix $A = A^H$).
|
|
Then $A$ has an orthonormal basis consisting of eigenvectors and is diagonalizable.
|
|
\end{thm}
|
|
\begin{proof}
|
|
Let $\lambda \in \cmpln$ be an eigenvalue of $A \in \field^{n \times n}$ with eigenvector $v \in \field^n$ and $A = A^H$. Then
|
|
\begin{equation}
|
|
\lambda\innerproduct{v}{v} = \innerproduct{v}{\lambda v} = \innerproduct{v}{Av} = \innerproduct{A^Hv}{v} = \innerproduct{Av}{v} = \innerproduct{\lambda v}{v} = \conj{\lambda} \innerproduct{v}{v}
|
|
\end{equation}
|
|
Therefore
|
|
\begin{equation}
|
|
(\lambda - \conj{\lambda})\underbrace{\innerproduct{v}{v}}_0 = 0
|
|
\end{equation}
|
|
\begin{equation}
|
|
\implies (\lambda - \conj{\lambda}) = 0 \implies \lambda = \conj{\lambda} \implies \lambda \in \realn
|
|
\end{equation}
|
|
Now let $\lambda, \rho \in \realn$ be eigenvalues to the eigenvectors $v, w$, and require $\lambda \ne \rho$. Then
|
|
\begin{equation}
|
|
\rho\innerproduct{v}{w} = \innerproduct{v}{Aw} = \innerproduct{Av}{w} = \conj{\lambda}\innerproduct{v}{w} = \lambda\innerproduct{v}{w}
|
|
\end{equation}
|
|
And thus
|
|
\begin{equation}
|
|
\underbrace{(\rho - \lambda)}_{\ne 0} \underbrace{\innerproduct{v}{w}}_{=0} = 0 \implies v \perp w
|
|
\end{equation}
|
|
\end{proof}
|
|
\end{document} |