Can a Hermitian matrix contain imaginary numbers?

Yes. Off-diagonal entries can be complex, paired with their conjugates.

Are all real symmetric matrices Hermitian?

Yes, because conjugating real entries leaves them unchanged.

Does Hermitian imply positive definite?

No. The worked diagonal counterexample has a negative eigenvalue while satisfying the Hermitian condition.

AI Glossary / Mathematics and optimization / Advanced

Hermitian

Definition

A square matrix is Hermitian when it equals its conjugate transpose: A = Aᴴ. Taking the conjugate transpose means swapping rows and columns and replacing each complex entry by its complex conjugate. Hermitian matrices generalize real symmetric matrices to complex numbers. They are important because their eigenvalues are real and they admit an orthonormal basis of eigenvectors over the complex numbers.

What it means

Complex numbers contain a real part and an imaginary part. Conjugation changes the sign of the imaginary part. In a Hermitian matrix, entries reflected across the diagonal are conjugates of one another, and diagonal entries are real. This is a structural property, not a requirement that every entry be real or positive. Hermitian matrices appear in complex-valued signal processing, covariance calculations and quadratic forms. Their algebraic structure allows specialized numerical methods, but software generally expects the input to satisfy the required structure.

How it works

For entries a_ij, the condition is a_ij = conjugate(a_ji). In finite-dimensional complex Euclidean space, the spectral theorem gives A = UΛUᴴ with U unitary and Λ real diagonal. The quadratic form zᴴAz is real for every complex vector z when A is Hermitian. Positive semidefiniteness adds the stronger requirement that this real value be nonnegative for all z; it is not part of the Hermitian definition. For real entries, conjugation has no effect and the condition reduces to ordinary symmetry A = Aᵀ.

Three practical examples

A complex-valued signal-processing model can use a Hermitian covariance matrix to represent relationships among channels.
A spectral computation can use a Hermitian eigensolver when the matrix structure is established, obtaining real eigenvalues and orthonormal eigenvectors.
An optimization derivation can examine a Hermitian quadratic form and separately determine whether its eigenvalues make it positive definite, semidefinite or indefinite.

Think of it this way

Think of a table whose reflected entries form matching pairs, except that the imaginary part changes sign across the diagonal. The analogy helps remember the pattern, but it does not explain the spectral theorem or positivity. A visually balanced table can still contain negative eigenvalues, and numerical approximations can slightly disturb an exact Hermitian structure.

Worked example

Let A have rows (2, i) and (−i, 2), where i² = −1. Transposing gives rows (2, −i) and (i, 2); conjugating returns A, so it is Hermitian. It is not equal to its ordinary transpose, showing why the conjugation matters. The characteristic determinant is det(A − λI) = (2 − λ)² − i(−i) = (2 − λ)² − 1. Therefore the eigenvalues are 1 and 3. For eigenvalue 1, the vector (1, i) satisfies A(1, i) = (1, i). For eigenvalue 3, (1, −i) satisfies A(1, −i) = 3(1, −i). Each vector has norm √2, and their complex inner product is zero: conjugate(1, i) dotted with (1, −i) gives 1 + (−i)(−i) = 0. Dividing each by √2 gives orthonormal eigenvectors. Both eigenvalues are positive, so this particular matrix is positive definite. In contrast, the real diagonal matrix with entries 1 and −2 is also Hermitian but indefinite: its quadratic form is positive on (1, 0) and negative on (0, 1). Hermitian structure alone does not establish positivity.

Illustrative case study

In a fictional sensor project, Cypress Signals builds a matrix from complex frequency-domain measurements. A reviewer checks conjugation in the covariance calculation and finds that an ordinary transpose had been used in one path. The team corrects the formula and compares the resulting matrix with its conjugate transpose within a documented numerical tolerance. It separately checks the eigenvalue behavior expected from the model. The case illustrates a structural implementation check, not a claim that the sensor system is accurate or ready for deployment.

Common misconceptions and mistakes

Hermitian and unitary are different properties. Hermitian means equal to the conjugate transpose; unitary means the conjugate transpose is the inverse. A Hermitian matrix need not be invertible, and its eigenvalues need not be positive. Ordinary symmetry is insufficient for complex matrices. Numerical eigensolvers may read only one triangle of a matrix and assume Hermitian structure, so an apparently sensible output does not prove the supplied full array satisfies the definition. Tiny imaginary or asymmetric components require interpretation in light of the construction and numerical precision.

Business use

Use specialized Hermitian routines only after confirming that the mathematical model and data construction warrant them. Check dimensions, conjugation and the relation between the two triangles. Record tolerances for approximate equality and inspect departures large enough to suggest an error. If the matrix is supposed to be a covariance or another positive semidefinite object, verify that additional property separately. Do not silently replace a substantially non-Hermitian matrix by its Hermitian part without understanding how that changes the problem. Structural checks complement, rather than replace, validation of the surrounding model.

Check your understanding

Why is the matrix with off-diagonal entries i and −i Hermitian even though it is not symmetric under ordinary transpose?

Answer: Transposition swaps those entries, and conjugation changes their signs, restoring the original matrix.

Related terms

Conjugate transpose
Symmetric matrix
Unitary matrix
Eigenvalue
positive semidefinite
complex inner product

Learn next

verify the two eigenvector equations directly and normalize the vectors. Then replace both diagonal entries in the worked matrix by zero and calculate its eigenvalues. Explain why the resulting matrix remains Hermitian but loses positive definiteness.

Sources

Sources: Jung and colleagues, The Aalto Dictionary of Machine Learning, Hermitian and conjugate transpose, https://aaltodictionaryofml.github.io/ADictML_Main.pdf ; NumPy, numpy.linalg.eigh, https://numpy.org/doc/stable/reference/generated/numpy.linalg.eigh.html . The complex two-by-two matrix, eigenvectors, inner-product check and fictional sensor review are original.

Previous: Herbrand's theorem Next: Hessian