1. Definition and Basic Properties
1.1 Conjugate transpose and the Hermitian condition
Let \(A\in \mathbb{C}^{n\times n}\). The conjugate transpose \(A^\*\) is defined by \((A^\*)_{ij}=\overline{A_{ji}}\). The matrix \(A\) is Hermitian if it satisfies \[ A=A^\*. \] Equivalently, \(A_{ij}=\overline{A_{ji}}\) for all indices \(i,j\). In particular, diagonal entries obey \(A_{ii}=\overline{A_{ii}}\), so each \(A_{ii}\) is real.
A useful way to read the condition is entrywise: off-diagonal elements occur in conjugate pairs across the diagonal, while the diagonal is constrained to real values.
1.2 Examples: real symmetric vs. complex Hermitian
If \(A\) is real, then conjugation does nothing. In this case the Hermitian condition becomes \(A=A^T\), so real symmetric matrices are exactly the real instance of Hermitian matrices.
In the complex setting, Hermitian matrices can contain non-real entries. For example, \[ \begin{pmatrix} 2 & 1-i\\ 1+i & -3 \end{pmatrix} \] is Hermitian because the entry \(1-i\) in the upper-right is conjugate to \(1+i\) in the lower-left, and both diagonal entries are real.
1.3 Closure properties and algebraic operations
Hermitian matrices form a real vector space: if \(A\) and \(B\) are Hermitian and \(\alpha,\beta\in\mathbb{R}\), then \(\alpha A+\beta B\) is Hermitian. Complex scalars do not preserve Hermitian symmetry in general; multiplying by \(i\) typically produces a skew-Hermitian matrix.
A closely related closure statement involves products with unitaries: if \(U\) is unitary (\(U^\*U=I\)) and \(A\) is Hermitian, then \(U^\*AU\) is Hermitian. This invariance under unitary congruence is central to diagonalization results.
Sums and scalar combinations over \(\mathbb{R}\) therefore behave well, while general matrix products of Hermitian matrices need not be Hermitian.
1.4 Relationships with normal matrices
A matrix \(N\) is normal if it commutes with its conjugate transpose: \(NN^\*=N^\*N\). Every Hermitian matrix is normal because \(A=A^\*\) implies \(AA^\*=A^2=A^\*A\).
Hermitian matrices represent one of the key subclasses of normal matrices. For normal matrices, eigenvalues may be complex, but for Hermitian matrices additional structure forces eigenvalues to be real and provides an especially clean spectral theorem.
2. Spectral Theorem and Diagonalization
2.1 Eigenvalues of Hermitian matrices are real
If \(A\) is Hermitian and \(v\neq 0\) satisfies \(Av=\lambda v\), then \[ \lambda=\frac{v^\*Av}{v^\*v}. \] The numerator \(v^\*Av\) is real because \(A=A^\*\), hence \(v^\*Av=\overline{v^\*Av}\). Therefore \(\lambda\in\mathbb{R}\). This real-eigenvalue property is one of the defining analytical advantages of Hermitian matrices.
2.2 Orthogonality of eigenvectors
For Hermitian \(A\), eigenvectors corresponding to distinct eigenvalues are orthogonal with respect to the standard complex inner product. If \(Av=\lambda v\) and \(Aw=\mu w\) with \(\lambda\neq \mu\), then \[ v^\*Aw = \mu\, v^\*w,\quad \text{but also}\quad v^\*Aw=(A v)^\*w=\lambda\, v^\*w, \] so \((\lambda-\mu)v^\*w=0\) and hence \(v^\*w=0\). This orthogonality supports the construction of orthonormal eigenbases.
2.3 Unitary diagonalization (spectral decomposition)
The spectral theorem states that a Hermitian matrix can be diagonalized by a unitary matrix. Concretely, there exists unitary \(U\) and a real diagonal matrix \(D\) such that \[ A=U D U^\*. \] The diagonal entries of \(D\) are the eigenvalues of \(A\). Columns of \(U\) form an orthonormal basis of eigenvectors. This diagonalization is not merely formal: it turns many matrix questions into scalar ones applied to eigenvalues.
2.4 Functional calculus for Hermitian matrices
Because Hermitian matrices are unitarily diagonalizable, one can define \(f(A)\) for suitable real-valued functions \(f\) applied to eigenvalues. If \(A=U D U^\*\) and \(D=\mathrm{diag}(\lambda_1,\dots,\lambda_n)\), then \[ f(A)=U\,\mathrm{diag}(f(\lambda_1),\dots,f(\lambda_n))\,U^\*. \] This framework underlies formulas for matrix exponentials, square roots (for positive semidefinite matrices), and other spectral functions.
3. Quadratic Forms and Inner-Product Viewpoint
3.1 Hermitian forms \(x^\*Ax\)
For a Hermitian matrix \(A\), the expression \[ q_A(x)=x^\*Ax \] defines a real-valued quadratic form on \(\mathbb{C}^n\) (real-valued for all \(x\)). The real-valued nature follows from conjugate symmetry: \[ \overline{x^\*Ax} = x^\*A^\*x = x^\*Ax. \] Quadratic forms convert algebraic properties of \(A\) into scalar behavior that can be studied via geometry and optimization.
3.2 Geometric meaning via orthonormal bases
When an orthonormal eigenbasis is chosen, the quadratic form becomes transparent. If \(A=U D U^\*\) and \(y=U^\*x\), then \[
| x^\*Ax = y^\*Dy = \sum_{k=1}^n \lambda_k | y_k | ^2. |
|---|
\] Thus, eigenvalues act as weights governing how the quadratic form responds along each eigendirection. In particular, the sign pattern of eigenvalues corresponds to the “shape” of the quadratic form.
3.3 Rayleigh quotient and extremal principles
A common normalization is the Rayleigh quotient \[ R_A(x)=\frac{x^\*Ax}{x^\*x},\quad x\neq 0. \]
| For Hermitian \(A\), \(R_A(x)\) is real and equals the weighted average of eigenvalues with weights \( | y_k | ^2\) from the eigenbasis representation. As \(x\) varies, the Rayleigh quotient attains its extremal values at eigenvectors. |
|---|
3.4 Min–max characterizations of eigenvalues
Hermitian eigenvalues can be characterized using constrained maximization and minimization on subspaces. Let \(\lambda_1\ge \cdots \ge \lambda_n\) be eigenvalues ordered decreasingly. Then the top eigenvalue satisfies \[
| \lambda_1=\max_{\|x\|=1} x^\*Ax, |
|---|
\] and more generally each \(\lambda_k\) can be recovered by optimizing the Rayleigh quotient over \((k)\)-dimensional subspaces, with orthogonality constraints to previously selected eigenspaces. These min–max principles are fundamental in both theory and algorithms.
4. Positive Definite and Semidefinite Hermitian Matrices
4.1 Definitions and equivalent criteria
A Hermitian matrix \(A\) is positive semidefinite (PSD) if \[ x^\*Ax \ge 0\quad \text{for all } x\in\mathbb{C}^n. \] It is positive definite (PD) if the inequality is strict for all nonzero \(x\): \[ x^\*Ax >0\quad \text{for all } x\neq 0. \] For Hermitian matrices, these conditions are equivalent to spectral constraints: \(A\) is PSD (resp. PD) exactly when all eigenvalues are nonnegative (resp. positive).
4.2 Cholesky decomposition (existence and implications)
For PD Hermitian matrices, a factorization exists of the form \[ A = L L^\* \] where \(L\) is lower triangular with positive diagonal entries. This is the Cholesky decomposition. It provides a computationally efficient method to encode definiteness and to solve linear systems.
For PSD matrices, a related factorization may exist but may require modifications such as using reduced forms or allowing factors with zero diagonal entries, reflecting rank deficiency.
4.3 Connection to norms and distances
When \(A\) is PSD, the quadratic form \(x^\*Ax\) behaves like a squared seminorm. If \(A\) is PD, it defines a genuine norm: \[
| \|x\|_A = \sqrt{x^\*Ax}. |
|---|
\] Such norms induce geometry: they stretch and rotate space according to the eigenvectors of \(A\), with scaling given by square roots of eigenvalues. Distances and projection operations in numerical linear algebra often leverage these weighted norms.
4.4 Spectral characterization of definiteness
From the eigen-decomposition \(A=U D U^\*\), definiteness is determined by the signs of \(\lambda_i\). Specifically:
- PSD: \( \lambda_i\ge 0\) for all \(i\).
- PD: \( \lambda_i>0\) for all \(i\).
- Rank and null space: zero eigenvalues correspond exactly to directions \(x\) with \(x^\*Ax=0\).
This characterization helps in both theoretical proofs and practical checks that reduce matrix questions to verifying properties of eigenvalues.
5. Decompositions and Canonical Forms
5.1 Spectral decomposition as a sum of projectors
An eigen-decomposition can be written not only as a diagonalization but as a sum over spectral projectors. If \(\lambda_1,\dots,\lambda_r\) are distinct eigenvalues and \(P_k\) is the orthogonal projector onto the eigenspace for \(\lambda_k\), then \[ A=\sum_{k=1}^r \lambda_k P_k. \] The projectors satisfy \(P_kP_\ell=0\) for \(k\neq \ell\) and \(\sum_k P_k=I\). This form isolates algebraic effects by eigenspaces and is useful for deriving identities involving \(A\) and functions of \(A\).
5.2 Polarization identity and recovering bilinear forms
The quadratic form determines the associated sesquilinear (conjugate-linear in one argument) form. For Hermitian \(A\), one can recover \(x^\*Ay\) from evaluations of \(q_A\) using polarization identities. In complex inner-product spaces, this step is important because it connects a quadratic expression to a full bilinear interaction between vectors.
Conceptually, knowing how \(x^\*Ax\) behaves for all \(x\) pins down all mixed terms \(x^\*Ay\), ensuring that the quadratic form fully encodes the matrix \(A\) when \(A\) is Hermitian.
5.3 Rank, null space, and inertia
The rank of \(A\) equals the number of nonzero eigenvalues. The null space \(\ker(A)\) is the direct sum of eigenspaces for eigenvalue \(0\). Because Hermitian matrices have real eigenvalues, one can also define the inertia: the counts of positive, zero, and negative eigenvalues.
Inertia is invariant under unitary congruence for Hermitian matrices, reflecting that the sign structure of eigenvalues is a robust characteristic of the quadratic form.
5.4 Principal submatrices and interlacing (overview level)
Hermitian matrices often appear in ways that require studying sub-blocks, such as principal submatrices. Eigenvalues of principal submatrices relate to those of the full matrix through eigenvalue interlacing phenomena. Informally, when you take a smaller principal submatrix, its eigenvalues “fit between” those of the larger matrix. This principle provides bounds and supports inductive reasoning in many proofs and algorithms.
6. Computational Aspects
6.1 Checking Hermitian structure numerically
| In computation, a matrix intended to be Hermitian may suffer from rounding errors. A standard check involves comparing \(A\) to \(A^\*\): if the norm of \(A-A^\*\) is small relative to \(\|A\|\), the matrix is approximately Hermitian. Alternatively, one can enforce Hermitian symmetry by replacing \(A\) with \(\frac12(A+A^\*)\), which preserves the intended eigenstructure up to numerical precision. |
|---|
6.2 Eigenvalue algorithms for Hermitian matrices
Eigenvalues and eigenvectors of Hermitian matrices are computed reliably using specialized methods that exploit symmetry, such as:
- reduction to tridiagonal form via unitary similarity,
- QR-type iterations tailored for tridiagonal Hermitian matrices,
- divide-and-conquer or other modern variants for efficiency.
These algorithms benefit from real eigenvalues and orthogonality properties, enabling stable computation of spectral data.
6.3 Conditioning and stability considerations
The sensitivity of eigenvalues to perturbations depends on eigenvalue separation. For Hermitian matrices, the eigenvector condition improves in the presence of well-separated eigenvalues, while nearly repeated eigenvalues can lead to larger changes in eigenvectors (though the invariant subspace remains relatively stable). Numerical stability is supported by the orthonormal eigenbasis structure and by using unitary transformations that do not amplify errors in the same way non-orthogonal operations might.
6.4 Applications of unitary transformations in computation
Many numerical procedures for Hermitian matrices rely on unitary transformations because they preserve inner products and norms. Unitary similarity reduces the matrix to simpler forms while maintaining key spectral properties. This strategy underlies stable implementations of eigenvalue solvers, matrix decompositions, and projection methods in high-dimensional problems.
7. Applications and Context (Mathematical)
7.1 Linear operators on complex inner-product spaces
Hermitian matrices represent self-adjoint linear operators in finite-dimensional complex inner-product spaces. In this operator viewpoint, many abstract properties become concrete: the reality of eigenvalues corresponds to self-adjointness, and orthonormal eigenbases reflect diagonalizability by unitary maps. This dual perspective—matrix and operator—is common throughout functional analysis in its finite-dimensional analogs.
7.2 Hermitian matrices in optimization high-level
In optimization, Hermitian matrices naturally model curvature in quadratic objective functions and constraints. Since the definiteness of the matrix controls whether a quadratic function is convex, PSD or PD Hermitian matrices serve as fundamental cases in guaranteeing existence and uniqueness properties of minima (under standard assumptions). Spectral information also yields step-size and feasibility criteria in methods that use second-order approximations.
7.3 Least-squares and projections (high-level)
Least-squares problems often lead to systems involving Hermitian matrices, such as normal equations (in contexts where they are formed) or Gram matrices. The projection of a vector onto a subspace can be expressed using Hermitian projectors, and the orthonormal basis of eigenvectors clarifies how error components decompose across eigenspaces. These connections make Hermitian structure useful for both conceptual understanding and numerical algorithms.
7.4 Matrix inequalities involving Hermitian matrices (overview level)
A large set of inequalities in matrix analysis is formulated for Hermitian matrices, including bounds on eigenvalues, trace-based comparisons, and relationships between operator norms and quadratic forms. Since the eigenvalues of Hermitian matrices are real, many scalar inequalities extend to matrix statements through spectral methods. These tools are widely used in analysis of algorithms, stability studies, and theoretical investigations of operator behavior.