13  Similarity and Diagonalization

Similarity

Definition 13.1 Two square matrices \mathbf{A}, \mathbf{B} \in \mathbb{C}^{n \times n} are similar if there exists a non-singular matrix \mathbf{P} such that

\mathbf{A} = \mathbf{P} \mathbf{B} \mathbf{P}^{-1}.

\mathbf{P}^{-1} \mathbf{A} \mathbf{P} = \mathbf{B},

Theorem 13.1 Similar matrices have the same eigenvalues.

Suppose \lambda_{1}, \dots, \lambda_{n} are eigenvalues of \mathbf{B}.

\begin{aligned} \text{det} (\mathbf{B} - \lambda \mathbf{I}) & = 0 \\ \text{det} (\mathbf{P}^{-1} \mathbf{A} \mathbf{P} - \lambda \mathbf{I}) & = 0 \\ \text{det} (\mathbf{P}^{-1} \mathbf{A} \mathbf{P} - \lambda \mathbf{P}^{-1} \mathbf{P}) & = 0 \\ \text{det} (\mathbf{P}^{-1} (\mathbf{A} - \lambda \mathbf{I}) \mathbf{P}) & = 0 \\ \text{det} (\mathbf{P}^{-1}) \text{det} (\mathbf{A} - \lambda \mathbf{I}) \text{det} (\mathbf{P}) & = 0 \\ \text{det} (\mathbf{P}^{-1}) \text{det} (\mathbf{P}) \text{det} (\mathbf{A} - \lambda \mathbf{I}) & = 0 \\ \text{det} (\mathbf{A} - \lambda \mathbf{I}) & = 0 & [\text{det} (\mathbf{P}^{-1}) = \frac{1}{\text{det} (\mathbf{P})}] \\ \end{aligned}

Unitarily (orthogonally) similar

Definition 13.2 Two square matrices \mathbf{A}, \mathbf{B} \in \mathbb{C}^{n \times n} are unitarily (orthogonally) similar if there exists an unitary (orthogonal) matrix \mathbf{U} such that

\mathbf{A} = \mathbf{U} \mathbf{B} \mathbf{U}^{-1},

which, according to the property of orthogonal matrix, can also be written as

\mathbf{A} = \mathbf{U} \mathbf{B} \mathbf{U}^{H},

Diagonalization

Definition 13.3 A square matrix \mathbf{A} \in \mathbb{C}^{n \times n} is diagonalizable if \mathbf{A} is similar to a diagonal matrix:

\mathbf{A} = \mathbf{P} \mathbf{\Lambda} \mathbf{P}^{-1}

where \mathbf{\Lambda} is a diagonal matrix.

Diagonalization and eigensystems

Theorem 13.2 A square matrix \mathbf{A} \in \mathbb{C}^{n \times n} is diagonalizable if and only if \mathbf{A} has n linearly independent eigenvectors

\mathbf{A} = \mathbf{P} \mathbf{\Lambda} \mathbf{P}^{-1},

where

  • the columns of \mathbf{P} \in \mathbb{R}^{n \times n} are n linearly independent eigenvectors,

  • and the diagonal values of \mathbf{\Lambda} are corresponding eigenvalues.

We first prove that If \mathbf{A} is diagonalizable, then \mathbf{A} has n linearly independent eigenvectors.

First note that

\begin{aligned} \mathbf{A} & = \mathbf{P} \mathbf{\Lambda} \mathbf{P}^{-1} \\ \mathbf{A} \mathbf{P} & = \mathbf{P} \mathbf{\Lambda} \\ \end{aligned}

Since \mathbf{\Lambda} is a diagonal matrix with \lambda_{1}, \dots, \lambda_{n} on its diagonal,

\mathbf{A} \mathbf{P}_{*, i} = \mathbf{P}_{*, i} \lambda_{i} = \lambda_{i} \mathbf{P}_{*, i}.

Thus, (\lambda, \mathbf{P}_{*, i}) is an eigenpair of \mathbf{A}, and \mathbf{A} has n such eigenpairs.

Since \mathbf{P} is non-singular (full-rank), \mathbf{P} has independent columns. Thus, \mathbf{A} has n independent eigenvectors.

Then we prove that if \mathbf{A} has n linearly independent eigenvectors, then \mathbf{A} is diagonalizable.

Assume the columns of \mathbf{P} \in \mathbf{C}^{n \times n} are the eigenvectors of \mathbf{A}, and \lambda_{1}, \dots, \lambda_{n} are corresponding eigenvalues,

\mathbf{A} \mathbf{P}_{*, i} = \lambda_{i} \mathbf{P}_{*, i} \quad \forall i = 1, \dots, n.

By rewriting \lambda_{1}, \dots, \lambda_{n} as diagonals for the diagonal matrix \mathbf{\Lambda},

\mathbf{A} \mathbf{P} = \mathbf{P} \mathbf{\Lambda}.

Since the columns of the \mathbf{P} are linearly independent, \mathbf{P} has full rank and thus there exists the inverse of \mathbf{P}

\begin{aligned} \mathbf{A} \mathbf{P} & = \mathbf{P} \mathbf{\Lambda} \\ \mathbf{P}^{-1} \mathbf{A} \mathbf{P} & = \mathbf{\Lambda} \end{aligned}

which shows that \mathbf{A} is similar to \mathbf{\Lambda} and thus is diagonalizable.

Diagonalization and multiplicities

Theorem 13.3 A square matrix \mathbf{A} \times \mathbb{C}^{n \times n} is diagonalizable if and only if

\mathrm{geo mult}_{\mathbf{A}} (\lambda_{i}) = \mathrm{alg mult}_{\mathbf{A}} (\lambda_{i}),

for each \lambda in the spectrum \sigma (\mathbf{A}).

TODO

Corollary 13.1 If no eigenvalue of \mathbf{A} is repeated, then \mathbf{A} is diagonalizable.

TODO