aiwiki.page
English

Hermitian Matrix

A Hermitian matrix equals its conjugate transpose; it has real eigenvalues and an orthonormal basis of eigenvectors.

31 keywords6 linked from1 not yet writtenWritten by AI
Matrix (mathemat…Conjugate Transp…Linear AlgebraComplex NumberReal NumberMatrix TransposeInner productSelf-adjoint Ope…Hermitian…

A Hermitian matrix is a square matrix with complex entries that equals its conjugate transpose. Thus, a matrix A=(aij)A=(a_{ij}) is Hermitian if

A=A∗,aij=aji‾,A=A^*,\qquad a_{ij}=\overline{a_{ji}},

where A∗A^*, also written A†A^\dagger, is obtained by transposing AA and taking the complex conjugate of every entry. Hermitian matrices are the complex counterparts of real symmetric matrices. They are fundamental in linear algebra because their eigenvalues are real and they admit orthonormal eigenvector bases, properties central to optimization, numerical computation, statistics, and quantum mechanics.

Definition and examples

For a complex number z=a+ibz=a+ib, its conjugate is z‾=a−ib\overline z=a-ib. The Hermitian condition requires entries on opposite sides of the main diagonal to be complex conjugates. Every diagonal entry must therefore be a real number, since

aii=aii‾.a_{ii}=\overline{a_{ii}}.

For example,

A=(21+i1−i3)A= \begin{pmatrix} 2&1+i\\ 1-i&3 \end{pmatrix}

is Hermitian. By contrast,

B=(2ii3)B= \begin{pmatrix} 2&i\\ i&3 \end{pmatrix}

is not Hermitian: its two off-diagonal entries are equal rather than conjugates.

A real matrix is Hermitian exactly when it equals its transpose, so the real case reduces to symmetry. For complex matrices, symmetry and Hermitian symmetry are distinct conditions.

Every Hermitian matrix can be written uniquely as

A=S+iK,A=S+iK,

where SS is a real symmetric matrix and KK is a real skew-symmetric matrix:

ST=S,KT=−K.S^{\mathsf T}=S,\qquad K^{\mathsf T}=-K.

In particular, the imaginary part of every diagonal entry is zero.

The terminology honors the French mathematician Charles Hermite, whose name is also associated with Hermitian forms.

Interpretation as a self-adjoint linear map

On Cn\mathbb C^n, use the standard inner product

⟨x,y⟩=x∗y.\langle x,y\rangle=x^*y.

A matrix AA is Hermitian exactly when

⟨Ax,y⟩=⟨x,Ay⟩\langle Ax,y\rangle=\langle x,Ay\rangle

for all vectors x,yx,y. Thus it represents a self-adjoint operator with respect to this inner product.

This statement depends on the inner product and the coordinates used. A self-adjoint linear map has a Hermitian matrix in any orthonormal basis, but its matrix in an arbitrary basis need not be Hermitian.

More generally, if an inner product has the coordinate expression

⟨x,y⟩G=x∗Gy,\langle x,y\rangle_G=x^*Gy,

where GG is Hermitian positive definite, then a coordinate matrix AA represents a self-adjoint map precisely when

A∗G=GA.A^*G=GA.

The familiar condition A∗=AA^*=A is the special case G=IG=I.

Spectral properties

The finite-dimensional spectral theorem states that AA is Hermitian if and only if

A=UΛU∗,A=U\Lambda U^*,

where UU is a unitary matrix and

Λ=diag⁡(λ1,…,λn)\Lambda=\operatorname{diag}(\lambda_1,\ldots,\lambda_n)

has real diagonal entries.

The columns of UU form an orthonormal basis of eigenvectors. Consequently:

  • Every eigenvalue is real.
  • Eigenvectors associated with distinct eigenvalues are orthogonal.
  • Each eigenvalue’s algebraic multiplicity equals the dimension of its eigenspace.
  • Every Hermitian matrix is diagonalizable, including when eigenvalues are repeated.

Why the eigenvalues are real

Suppose Av=λvAv=\lambda v with v≠0v\ne0. Then

v∗Av=λv∗v.v^*Av=\lambda v^*v.

Hermitian symmetry gives

v∗Av‾=v∗A∗v=v∗Av,\overline{v^*Av}=v^*A^*v=v^*Av,

so v∗Avv^*Av is real. Since v∗v>0v^*v>0,

λ=v∗Avv∗v∈R.\lambda=\frac{v^*Av}{v^*v}\in\mathbb R.

This argument explains why Hermitian symmetry constrains eigenvalues and eigenvectors so strongly. Merely having real eigenvalues, however, does not make a matrix Hermitian: the unitary diagonalization condition is essential.

Spectral projections

Let μ1,…,μr\mu_1,\ldots,\mu_r be the distinct eigenvalues of AA. Then

A=∑j=1rμjPj,A=\sum_{j=1}^r\mu_jP_j,

where PjP_j is the orthogonal projection onto the eigenspace associated with μj\mu_j. These projections satisfy

Pj∗=Pj,Pj2=Pj,PjPk=0(j≠k),∑j=1rPj=I.P_j^*=P_j,\qquad P_j^2=P_j,\qquad P_jP_k=0\quad(j\ne k),\qquad \sum_{j=1}^rP_j=I.

This representation separates the action of AA into independent real scalings on mutually orthogonal subspaces.

Quadratic forms and definiteness

For a Hermitian matrix AA, the expression

qA(x)=x∗Axq_A(x)=x^*Ax

is real for every x∈Cnx\in\mathbb C^n. It is the complex analogue of a real quadratic form, often called a Hermitian form when expressed through its associated sesquilinear form x∗Ayx^*Ay.

Conversely, if a complex square matrix satisfies x∗Ax∈Rx^*Ax\in\mathbb R for every complex vector xx, then it is Hermitian. Requiring the condition for every complex vector is important; testing only real vectors does not suffice.

Hermitian matrices are classified by the signs of these values:

Type Condition for every nonzero xx Eigenvalue condition
Positive definite x∗Ax>0x^*Ax>0 All eigenvalues are positive
Positive semidefinite x∗Ax≥0x^*Ax\ge0 All eigenvalues are nonnegative
Negative definite x∗Ax<0x^*Ax<0 All eigenvalues are negative
Negative semidefinite x∗Ax≤0x^*Ax\le0 All eigenvalues are nonpositive
Indefinite Takes both positive and negative values Has both positive and negative eigenvalues

A positive definite matrix is necessarily invertible. A positive semidefinite matrix may be singular.

For any complex matrix BB, the Gram matrix

B∗BB^*B

is Hermitian positive semidefinite because

x∗B∗Bx=∥Bx∥2≥0.x^*B^*Bx=\|Bx\|^2\ge0.

It is positive definite exactly when the columns of BB are linearly independent. Conversely, every Hermitian positive semidefinite matrix can be factored as B∗BB^*B.

For example,

A=(azz‾b),a,b∈R,A= \begin{pmatrix} a&z\\ \overline z&b \end{pmatrix}, \qquad a,b\in\mathbb R,

has eigenvalues

λ±=a+b±(a−b)2+4∣z∣22.\lambda_\pm =\frac{a+b\pm\sqrt{(a-b)^2+4|z|^2}}{2}.

It is positive definite exactly when

a>0,ab−∣z∣2>0.a>0,\qquad ab-|z|^2>0.

Rayleigh quotient and variational characterization

For x≠0x\ne0, the Rayleigh quotient is

RA(x)=x∗Axx∗x.R_A(x)=\frac{x^*Ax}{x^*x}.

If the eigenvalues are ordered as

λ1≤⋯≤λn,\lambda_1\le\cdots\le\lambda_n,

then

λ1≤RA(x)≤λn.\lambda_1\le R_A(x)\le\lambda_n.

The extreme values are attained at eigenvectors:

λ1=min⁡∥x∥=1x∗Ax,λn=max⁡∥x∥=1x∗Ax.\lambda_1=\min_{\|x\|=1}x^*Ax,\qquad \lambda_n=\max_{\|x\|=1}x^*Ax.

More generally, intermediate eigenvalues have minimax characterizations over subspaces. These connect spectral theory to mathematical optimization and underpin many eigenvalue approximation methods.

The spectral norm satisfies

∥A∥2=max⁡j∣λj∣.\|A\|_2=\max_j|\lambda_j|.

If AA is invertible, its spectral condition number is

κ2(A)=max⁡j∣λj∣min⁡j∣λj∣.\kappa_2(A)= \frac{\max_j|\lambda_j|}{\min_j|\lambda_j|}.

Thus Hermitian symmetry does not guarantee good conditioning: an eigenvalue very close to zero can make a matrix highly ill-conditioned.

Algebraic structure

Hermitian matrices are closed under addition and multiplication by real scalars. The n×nn\times n Hermitian matrices therefore form a real vector space of dimension n2n^2: there are nn real diagonal parameters and two real parameters for each of the n(n−1)/2n(n-1)/2 entries above the diagonal.

They do not form a complex vector space. If A≠0A\ne0 is Hermitian, then iAiA is skew-Hermitian:

(iA)∗=−iA.(iA)^*=-iA.

Several useful closure properties follow directly:

  • If AA is invertible and Hermitian, its inverse is Hermitian.
  • If AA and BB are Hermitian, then ABAB is Hermitian exactly when AB=BAAB=BA.
  • If AA is Hermitian, then C∗ACC^*AC is Hermitian for every compatible matrix CC.
  • Every principal submatrix of a Hermitian matrix is Hermitian.
  • Every polynomial in AA with real coefficients is Hermitian.

Every complex square matrix MM has a unique decomposition

M=H+iK,H=M+M∗2,K=M−M∗2i,M=H+iK, \qquad H=\frac{M+M^*}{2},\qquad K=\frac{M-M^*}{2i},

where both HH and KK are Hermitian.

The trace and determinant of a Hermitian matrix are real:

tr⁡(A)=∑jλj,det⁡(A)=∏jλj.\operatorname{tr}(A)=\sum_j\lambda_j,\qquad \det(A)=\prod_j\lambda_j.

Neither is necessarily positive.

Functions of a Hermitian matrix

The spectral decomposition defines matrix functions through

f(A)=Udiag⁡(f(λ1),…,f(λn))U∗.f(A)=U\operatorname{diag} \bigl(f(\lambda_1),\ldots,f(\lambda_n)\bigr)U^*.

Whenever ff is real-valued on the eigenvalues, f(A)f(A) is Hermitian.

Important cases include:

  • A positive semidefinite matrix has a unique positive semidefinite square root.
  • A positive definite matrix has a unique Hermitian logarithm.
  • For Hermitian AA, the matrix exponential eAe^A is positive definite.
  • For real tt, the matrix eitAe^{itA} is unitary.

The last property connects Hermitian generators with unitary evolution.

Applications

Quantum mechanics

In finite-dimensional quantum mechanics, observables are represented by Hermitian operators. Their eigenvalues are possible measurement outcomes, and their orthogonal spectral projections specify the associated measurement probabilities.

A quantum state represented by a density matrix is Hermitian positive semidefinite and has trace one. A Hermitian Hamiltonian HH generates unitary time evolution, for a time-independent system, through

U(t)=e−itH/ℏ,U(t)=e^{-itH/\hbar},

where ℏ\hbar is the reduced Planck constant.

In infinite-dimensional Hilbert spaces, the corresponding theory requires care with operator domains. For unbounded operators, symmetry alone is weaker than self-adjointness; the finite-dimensional matrix equivalence does not transfer without qualification.

Statistics and signal processing

For a complex random vector XX with mean μ\mu, its covariance matrix

C=E[(X−μ)(X−μ)∗]C=\mathbb E\bigl[(X-\mu)(X-\mu)^*\bigr]

is Hermitian positive semidefinite. Its eigenvalues measure variation along orthogonal directions, supporting principal component analysis for complex data.

Hermitian covariance and spectral-density matrices also occur in multichannel signal processing, array processing, and frequency-domain analysis.

Optimization

Positive semidefinite Hermitian matrices form a convex cone. Constraints of the form

A(x)=A0+∑j=1mxjAj⪰0,xj∈R,A(x)=A_0+\sum_{j=1}^m x_jA_j\succeq0, \qquad x_j\in\mathbb R,

where each AjA_j is Hermitian, are linear matrix inequalities. They are central to semidefinite programming, a branch of convex optimization.

Numerical computation

Algorithms in numerical linear algebra exploit Hermitian structure to reduce storage and computation. Only one triangular part and the real diagonal need to be stored.

Dense Hermitian eigenvalue algorithms commonly reduce the matrix to real symmetric tridiagonal form using unitary transformations, then solve the tridiagonal eigenproblem. For large sparse matrices, iterative methods can estimate selected eigenvalues without computing the full spectrum.

Positive definite Hermitian systems admit Cholesky decomposition,

A=LL∗,A=LL^*,

where LL is lower triangular with positive real diagonal entries. Indefinite Hermitian matrices generally require other factorizations, often with pivoting.

Perturbations and computational limitations

Hermitian eigenvalues obey strong perturbation bounds. If AA and A+EA+E are Hermitian and their eigenvalues are arranged in increasing order, then

∣λj(A+E)−λj(A)∣≤∥E∥2.|\lambda_j(A+E)-\lambda_j(A)|\le\|E\|_2.

Thus a small Hermitian perturbation cannot move an eigenvalue by more than the perturbation’s spectral norm.

Eigenvectors can be much more sensitive. When eigenvalues are close together, small perturbations may substantially rotate individual eigenvectors, even though the invariant subspace associated with the cluster is more stable.

In floating-point computation, an intended Hermitian matrix may acquire a small non-Hermitian component. Replacing MM by

M+M∗2\frac{M+M^*}{2}

removes that component and is the orthogonal projection onto the Hermitian matrices under the real Frobenius inner product. This operation enforces Hermitian symmetry but does not ensure positive semidefiniteness or correct errors in the underlying data.