Normal Matrices and the Common Spectral Theorem at $n=4$
Introduction
A matrix is normal when $A^{\ast}A=AA^{\ast}$, and exactly then is it unitarily diagonalisable; for a normal $A=U\Lambda U^{\ast}$ one has $A^{\ast}=U\bar\Lambda U^{\ast}$, $A^{\mathsf T}=\bar U\Lambda\bar U^{\ast}$ and $\bar A=\bar U\bar\Lambda\bar U^{\ast}$, so the pair of conjugate unitaries diagonalises all four matrices at once, and a single unitary suffices when $A$ is real or symmetric.
The article proves the equivalence of normality with unitary diagonalisability, writes the four diagonal forms, identifies the four spectral families and their constraints on the eigenvalues, exhibits a Jordan block of size four as the failure, and reads the result in the axis-free frame, where the left matrix of a real element is a scalar multiple of an orthogonal matrix.
It assumes The Spectral Theorem: the Hermitian Case at $n=4$ and The Spectral Theorem: the Anti-Hermitian Case at $n=4$. The model at order two is Normal $2\times2$ Matrices and the Common Spectral Theorem; the fine structure of the spectrum is in The Minimal Polynomial and the Jordan Form of a $4\times4$ Matrix.
Normal Matrices
Definition. $A$ is normal when $A^{\ast}A=AA^{\ast}$.
Proposition. Hermitian, anti-Hermitian and unitary matrices are normal; a scalar multiple of a unitary matrix is normal; sums and products of normal matrices need not be normal.
Proof. For $A^{\ast}=\pm A$ the two products $A^{\ast}A$ and $AA^{\ast}$ are equal because $A$ commutes with itself; for $A^{\ast}=A^{-1}$ both products are the identity. For the failure of additivity, the diagonal $D=\operatorname{diag}(1,i,1,i)$ and the symmetric $S=E_{12}+E_{21}$ are both normal, while $D+S$ has the upper-left block $\begin{pmatrix}1&1\\1&i\end{pmatrix}$, for which $X^{\ast}X$ and $XX^{\ast}$ differ in the off-diagonal entries.
Corollary. Normality is not a linear condition, since the normal matrices do not form a subspace; it is a quadratic condition on the entries.
Proof. The display of the failure.
Proposition. $A$ is normal exactly when $\operatorname{tr}(A^{\ast}A)=\sum_k|\lambda_k|^{2}$ and the Frobenius norm of $A$ is the Euclidean norm of the spectrum.
Proof. The Schur triangular form gives $\operatorname{tr}(A^{\ast}A)=\sum_k|\lambda_k|^{2}+\sum_{i Corollary. Normality is the statement that the triangular form is diagonal, hence that the eigenbasis is orthonormal. Proof. The Schur form is triangular in an orthonormal basis, and it is diagonal exactly when the basis consists of eigenvectors. Theorem. $A$ is normal if and only if there is a unitary $U$ and a diagonal $\Lambda$ with $A=U\Lambda U^{\ast}$. Proof. If $A=U\Lambda U^{\ast}$ then $A^{\ast}=U\bar\Lambda U^{\ast}$ and the two products are $U|\Lambda|^{2}U^{\ast}$. Conversely, the Schur form is triangular with the eigenvalues on the diagonal, and the hypothesis of normality forces the strictly upper part to vanish. Corollary. For such $A$,
$$
A^{\ast}=U\bar\Lambda U^{\ast},\qquad A^{\mathsf T}=\bar U\Lambda\bar U^{\ast},\qquad \bar A=\bar U\bar\Lambda\bar U^{\ast},
$$
so the unitary $U$ diagonalises $A$ and $A^{\ast}$, and the conjugate unitary $\bar U$ diagonalises $A^{\mathsf T}$ and $\bar A$. Proof. Transpose $U\Lambda U^{\ast}$ and use $\overline{U^{\ast}}=\bar U^{\mathsf T}=(\bar U)^{\ast}$; conjugate the same identity. Corollary. A single unitary diagonalises all four exactly when $\bar U=U$ up to relabelling, which holds when $A$ is real and symmetric, and more generally when the eigenbasis can be chosen closed under conjugation. Proof. $A^{\mathsf T}=A$ forces $\bar U\Lambda\bar U^{\ast}=U\Lambda U^{\ast}$; for a real symmetric $A$ the eigenvectors can be taken real. The spectral projectors. $A=\sum_k\lambda_kP_k$ with $P_k$ the orthogonal projections onto the eigenspaces, $P_jP_k=\delta_{jk}P_k$, $\sum_kP_k=\mathrm I_4$, and each $P_k$ a polynomial in $A$. Proof. The projections are $U\operatorname{diag}(0,\dots,1,\dots,0)U^{\ast}$, and the Lagrange interpolation polynomials in $A$ produce them. Corollary. $A^{\ast}=\sum_k\bar\lambda_kP_k$ and $A$ is determined by its spectrum together with its spectral projectors, so normality makes the matrix a function of the spectrum and the basis. Proof. The spectral resolution and the orthogonality of the projectors. Theorem. Each of the following is normal with a constraint on the spectrum, and conversely: Proof. From $A=U\Lambda U^{\ast}$ each condition becomes the corresponding condition on the diagonal entries, and conversely. Corollary. The Hermitian and anti-Hermitian families are the fixed sets of the involutions $A\mapsto A^{\ast}$ and $A\mapsto-A^{\ast}$, the unitary family is the fixed set of $A\mapsto(A^{\ast})^{-1}$ on the invertible matrices, and each of the four is a real algebraic subset of the normal set. Proof. The table, read as equations in the spectral data. Corollary. The intersection of the Hermitian and unitary families is the set of involutions $A^{2}=\mathrm I$, of which the Hermitian idempotents are the positive members. Proof. Real spectrum with unit modulus gives $\lambda\in\{\pm1\}$. The four eigenvalues. At order four the spectrum is a multiset of four complex numbers, and the constraint sets are the real line, the imaginary line, the unit circle and the positive half-line, each counted with multiplicity. Proof. The diagonal form and the table. Theorem. A matrix with a Jordan block of size two or more is not normal. Proof. Let $A=\lambda\mathrm I+N$ with $N^{2}=0$, $N\neq0$. Then $A^{\ast}A-AA^{\ast}=N^{\ast}N-NN^{\ast}$, and for $N=E_{12}$ this is $E_{22}-E_{11}\neq0$. Corollary. The nilpotent $J_4=E_{12}+E_{23}+E_{34}$, with $J_4^{4}=0$ and $J_4^{3}=E_{14}\neq0$, is not normal, and no unitary diagonalises it although its only eigenvalue is zero. Proof. $J_4^{\ast}J_4=E_{22}+E_{33}+E_{44}$ and $J_4J_4^{\ast}=E_{11}+E_{22}+E_{33}$, so the defect is $E_{44}-E_{11}$; and a diagonal form would force $J_4=0$, since its only eigenvalue is zero. Corollary. The size-four Jordan block is the extreme failure: its invariant flag has length four, the longest possible at order four. Proof. The flag of $J_4$ is $0\subset\ker J_4\subset\ker J_4^{2}\subset\ker J_4^{3}\subset\mathbb{C}^{4}$, of length four. The deficit. The defect $J_4^{\ast}J_4-J_4J_4^{\ast}=E_{44}-E_{11}$ has operator norm one, and that nonzero value is the measure of the failure: a normal matrix has zero defect, whatever its spectrum. Proof. The two products computed above, and $\lVert E_{44}-E_{11}\rVert=1$ acting on $e_4$. Proposition. In the axis-free frame the left matrix of a real element is a real signed permutation combination and satisfies $L^{\mathsf T}L=N\mathrm I_4$ with $N=Q_t^{2}+B^{2}$, so $L$ is a scalar multiple of an orthogonal matrix and is normal. Proof. $L(\tilde Q)^{\mathsf T}=L(\tilde Q^{\natural})$ and $\tilde Q^{\natural}\tilde Q=\tilde Q\tilde Q^{\natural}=Ne_t$, so both products of $L$ with its transpose are $N\mathrm I_4$. Corollary. The left matrices of the real algebra are normal with spectrum on the circle of radius $\sqrt N$, so the whole real algebra sits inside the normal matrices. Proof. An orthogonal matrix has eigenvalues of modulus one, and multiplying by the scalar $\sqrt N$ scales the moduli. The complex case. For a general complex element the left matrix need not be normal, and it is normal exactly when $\tilde Q^{\natural\ast}\tilde Q=\tilde Q\tilde Q^{\natural\ast}$, that is when the element commutes with its conjugate transpose in the biquaternion algebra. Proof. $L(\tilde Q)^{\ast}=L(\tilde Q^{\natural\ast})$, so $L^{\ast}L=L(\tilde Q^{\natural\ast}\tilde Q)$ and $LL^{\ast}=L(\tilde Q\tilde Q^{\natural\ast})$; the representation is faithful, so the two are equal exactly when the elements are. Corollary. Normality in the algebra is the normality of the left matrix, and the axis-free frame turns the abstract condition into a condition on the four coefficients. Proof. The faithfulness of $L$ and the identity above. The warning. The phrase "the left matrices are normal because they are orthogonal up to a scalar" is true for the real algebra and false for the complex one: over $\mathbb{C}$ a signed permutation with complex coefficients is not a scalar multiple of a unitary matrix. Proof. A complex combination of orthogonal matrices is unitary only if its singular values coincide, which a general combination does not satisfy. Normality $A^{\ast}A=AA^{\ast}$ is equivalent to unitary diagonalisability, which is equivalent to the vanishing of the strictly upper part of the Schur form. For a normal $A=U\Lambda U^{\ast}$ the four matrices $A$, $A^{\ast}$, $A^{\mathsf T}$ and $\bar A$ are diagonalised by the pair $U$, $\bar U$, with $A^{\mathsf T}=\bar U\Lambda\bar U^{\ast}$ and $\bar A=\bar U\bar\Lambda\bar U^{\ast}$; one unitary suffices when the eigenbasis is closed under conjugation, as for a real symmetric matrix. The four spectral families — Hermitian, anti-Hermitian, unitary, positive definite — are the normal matrices whose spectrum lies on the real line, the imaginary line, the unit circle and the positive half-line. The Jordan block of size four is the extreme failure, with a flag of length four. In the axis-free frame the left matrix of a real element satisfies $L^{\mathsf T}L=N\mathrm I_4$ and is normal; for a complex element normality holds exactly when the element commutes with its conjugate transpose.The Common Spectral Theorem
The Four Families and Their Constraints
family
condition
spectral constraint
Hermitian
$A^{\ast}=A$
$\lambda_k\in\mathbb{R}$
anti-Hermitian
$A^{\ast}=-A$
$\lambda_k\in i\mathbb{R}$
unitary
$A^{\ast}=A^{-1}$
$|\lambda_k|=1$
positive definite
$A^{\ast}=A$, $\langle v,Av\rangle>0$
$\lambda_k>0$
The Failure for Non-Normal Matrices
The Axis-Free Reading
Summary
Summary of Notation
Symbol
Meaning
$A^{\ast}A=AA^{\ast}$
normality
$A=U\Lambda U^{\ast}$
the unitary diagonal form
$\bar U\Lambda\bar U^{\ast}$
the diagonal form of the transpose
$J_4$
the nilpotent Jordan block of size four
$N=Q_t^{2}+B^{2}$
the scalar of orthogonality in the axis-free frame
Further Reading