Must a symmetric matrix be square?

Yes. Equality with its transpose requires matching row and column dimensions.

Can a symmetric matrix have repeated eigenvalues?

Yes. Within a repeated-eigenvalue subspace, the orthonormal eigenbasis need not be unique.

Is a diagonal real matrix symmetric?

Yes. All off-diagonal entries are zero and therefore mirror exactly.

AI Glossary / Mathematics and optimization / Intermediate

Symmetric matrix

Definition

A real symmetric matrix is a square matrix that equals its transpose. Its entry in row i, column j matches the entry in row j, column i, so values mirror one another across the main diagonal.

What it means

Symmetry expresses a relationship between paired positions, not equal values everywhere. A symmetric matrix can contain different diagonal entries, negative values and zeros. It can be dense or sparse.

Covariance matrices and matrices of pairwise inner products are familiar AI examples. Their symmetry reflects the fact that exchanging the order of the two quantities leaves the corresponding covariance or inner product unchanged.

Symmetry alone does not make a matrix a valid covariance matrix. Covariance also requires positive semidefiniteness. A symmetric matrix can have negative eigenvalues, and it can be singular or invertible. These are separate properties that should be checked rather than inferred from the mirrored layout.

How it works

For a matrix A, the transpose Aᵀ exchanges rows and columns. Symmetry means A = Aᵀ, or equivalently Aij = Aji for every pair of indices. A rectangular matrix cannot satisfy this equality because its transpose has a different shape.

The real symmetric spectral theorem provides an orthonormal basis of eigenvectors and real eigenvalues. This permits A = QΛQᵀ, where Q has orthonormal columns and Λ is diagonal. The decomposition describes independent perpendicular directions along which the matrix scales vectors.

If every eigenvalue is nonnegative, the symmetric matrix is positive semidefinite; if every eigenvalue is positive, it is positive definite. Negative eigenvalues correspond to directions with a negative quadratic form.

Numerical routines may assume symmetry and read only one triangle of the input. Tiny roundoff differences can be assessed with tolerances, but a substantial disagreement between triangles should trigger investigation rather than silent acceptance.

Three practical examples

A principal-component analysis uses a symmetric covariance matrix to identify orthogonal directions of variation.
A kernel method works with a symmetric Gram matrix, while separately checking the required positive-semidefinite structure.
An optimization analysis uses a symmetric Hessian when the objective has suitably continuous second derivatives.

Think of it this way

Think of a table listing a mutual relationship between every pair of people. If the relationship is defined identically in either order, the table mirrors across its diagonal.

The analogy concerns a symmetric relation. Many real relationships, such as messages sent from one person to another, are directional and do not produce symmetric tables.

Worked example

Symmetric does not mean positive

Let A have rows (2, 1) and (1, 2). It is symmetric because its off-diagonal entries both equal one.

Multiplying A by (1, 1) gives (3, 3), so that direction has eigenvalue 3. Multiplying A by (1, −1) gives (1, −1), giving eigenvalue 1. Normalizing these two perpendicular vectors produces an orthonormal eigenbasis.

For x = (u, v), the quadratic form is xᵀAx = 2u² + 2uv + 2v². It also equals (3/2)(u + v)² + (1/2)(u − v)², which is positive for every nonzero x. This A is positive definite as well as symmetric.

Now take B with rows (0, 1) and (1, 0). B is also symmetric, but its eigenvalues are 1 and −1. For x = (1, −1), xᵀBx = −2. B is therefore not positive semidefinite and cannot serve as a covariance matrix. The mirrored entries establish symmetry; they do not establish the sign of the quadratic form.

Illustrative case study

A fictional recommendation team assembles a pairwise similarity table from two processing jobs. It passes the result to a routine designed for symmetric matrices, which reads only the lower triangle. The upper triangle contains materially different values from an earlier data version.

The team first reconciles the data versions rather than treating the solver output as proof that the table was correct. It then verifies symmetry and checks the additional conditions required by its chosen model.

The lesson is that a mathematical assumption can become an interface contract. A routine optimized for that contract may not validate every property automatically, so data checks must match what the computation expects.

Common misconceptions and mistakes

Symmetric means all entries are positive. Negative entries are allowed.
Symmetry guarantees positive eigenvalues. It guarantees real eigenvalues for real symmetric matrices, not positivity.
Every symmetric matrix is invertible. A zero eigenvalue makes it singular.
A nearly symmetric matrix can always be fixed by averaging it with its transpose. That operation changes the matrix and may hide an upstream problem.
Complex symmetric and Hermitian mean the same thing. Hermitian symmetry uses conjugate transpose, which matters for complex entries.

Business use

Symmetric structure can make matrix calculations more efficient and their results easier to interpret. Before relying on it, verify that the modeled relationship should actually be symmetric and that the stored values respect that relationship within justified tolerances.

For covariance or kernel applications, check the additional mathematical requirements. Preserve units and ordering so that row and column meanings align. If symmetrization is used to address numerical noise, document why the discrepancy is considered harmless and assess whether the change affects the downstream result.

Check your understanding

Why can B in the worked example not be a covariance matrix?

Answer: Although it is symmetric, it gives a negative quadratic form in one direction. Covariance matrices must be positive semidefinite, so directional variances cannot be negative.

Related terms

Transpose
Eigenvalue
Covariance matrix
orthogonal matrix
positive semidefinite matrix
Hermitian matrix

Learn next

Read positive semidefiniteness and eigendecomposition. Verify both eigenvector calculations above, then construct a symmetric matrix with a zero eigenvalue. Explain why symmetry, positivity and invertibility answer different questions.

Eigenvalue

Sources

Aalto Dictionary of Machine Learning, symmetric matrix: https://aaltodictionaryofml.github.io/ADictML_Main.pdf. Definition.

NumPy, numpy.linalg.eigh: https://numpy.org/doc/stable/reference/generated/numpy.linalg.eigh.html. Symmetric/Hermitian eigensolver behavior and triangle selection. Matrix calculations and the fictional data-version incident are original examples.

Previous: Symmetric distribution Next: Symmetric multiprocessing