Orthogonal Similarity and Orthogonal Diagonalization of Matrices
Definition1
Orthogonal similarity
For two square matrices $A$ and $B$, if there exists an orthogonal matrix $P$ that satisfies the following, then $A$ and $B$ are said to be orthogonally similar.
$$ A = P^{\mathsf{T}}BP $$
Orthogonal diagonalization
A square matrix $A$ is said to be orthogonally diagonalizable (or $P$ is said to orthogonally diagonalize $A$) if it is orthogonally similar to some diagonal matrix $D$. In other words, $A$ is orthogonally diagonalizable if there exists an orthogonal matrix $P$ such that the following holds.
$$ A = P^{\mathsf{T}}DP $$
Remarks
Since an orthogonal matrix satisfies $P^{\mathsf{T}} = P^{-1}$, if $A$ and $B$ are orthogonally similar then $A$ and $B$ are similar.
Theorem
(a) If orthogonally diagonalizable then diagonalizable. The converse does not hold.
For a real square matrix $A$ the following conditions are equivalent.
(a) $A$ is orthogonally diagonalizable.
(b) $A$ has an orthonormal set consisting of $n$ eigenvectors.
(c) $A$ is a symmetric matrix.
As a corollary of (a) $\implies$ (b), if $A$ is orthogonally diagonalizable and $A = P^{\mathsf{T}} D P$, then each diagonal entry $d_{ii}$ of $D$ is an eigenvalue of $A$ corresponding to the $i$-th column vector of $P$.
Proof
(a) $\implies$ (b)
Assume the matrix $A$ in $n \times n$ is orthogonally diagonalizable. Then there exists an orthogonal matrix $P$ satisfying
$$ D = P^{-1} A P $$
Let the column vectors of $P$ be $\mathbf{p}_{i}$. Then $\mathbf{p}_{i}$ are linearly independent eigenvectors. Moreover, since $P$ is orthogonal, the column vectors form an orthonormal set.
■
From the above proof we obtain the following equation.
$$ PD = AP \implies \begin{bmatrix} d_{11} \mathbf{p}_{1} & \cdots & d_{nn} \mathbf{p}_{n} \end{bmatrix} = \begin{bmatrix} A \mathbf{p}_{1} & \cdots & A \mathbf{p}_{n} \end{bmatrix} $$
Therefore, the $i$-th diagonal entry $d_{ii}$ of $D$ is an eigenvalue of $A$ corresponding to the $i$-th column vector $\mathbf{p}_{i}$ of $P$.
(b) $\implies$ (c)
Suppose $A$ has an orthonormal set $\left\{ \mathbf{p}_{i} \right\}$ consisting of $n$ eigenvectors. The matrix $P = \begin{bmatrix} \mathbf{p}_{1} & \mathbf{p}_{2} & \cdots & \mathbf{p}_{n} \end{bmatrix}$ having these as column vectors diagonalizes $A$.
$$ A = P^{-1}DP $$
Since $\left\{ \mathbf{p}_{i} \right\}$ is an orthonormal set, $P$ is an orthogonal matrix.
$$ A = P^{\mathsf{T}}DP $$
Hence the following holds, and $A$ is a symmetric matrix.
$$ A^{\mathsf{T}} = (P^{\mathsf{T}}DP)^{\mathsf{T}} = P^{\mathsf{T}}D^{\mathsf{T}}P = P^{\mathsf{T}}DP = A $$
■
(c) $\implies$ (a)
If $A$ is symmetric then it is a normal matrix, and by the spectral theorem $A$ is unitarily diagonalizable. Unitarily diagonalizable implies orthogonally diagonalizable.
■
Howard Anton, Elementary Linear Algebra: Aplications Version (12th Edition, 2019), p408-409 ↩︎