Symmetric Matrices and Their Properties

Core Idea

A symmetric matrix satisfies Aᵀ = A; equivalently, its (i,j) entry equals its (j,i) entry for all i, j. Symmetric matrices arise naturally from inner products, quadratic forms, and covariance matrices in statistics. Two fundamental properties distinguish them: all eigenvalues of a real symmetric matrix are real, and eigenvectors corresponding to distinct eigenvalues are orthogonal. These properties make symmetric matrices far better behaved than general matrices and are the hypotheses of the Spectral Theorem. A symmetric matrix A is positive definite if xᵀAx > 0 for all nonzero x, equivalent to all eigenvalues being positive.

How It's Best Learned

Verify that the eigenvalues of specific symmetric 2×2 and 3×3 matrices are real and that eigenvectors for distinct eigenvalues are orthogonal via dot product. Contrast with a non-symmetric matrix having complex eigenvalues (e.g., a rotation matrix).

Common Misconceptions

• Not all matrices with real eigenvalues are symmetric; symmetry is a sufficient but not necessary condition.
• Positive definite matrices have all positive eigenvalues, but not all symmetric matrices with positive diagonal entries are positive definite.
• Symmetry is a property of the matrix in a fixed basis; it is NOT preserved under arbitrary change of basis.

Explainer

You already know from matrix transpose that Aᵀ flips a matrix across its main diagonal, swapping row i with column i. A symmetric matrix is one that survives this flip unchanged: Aᵀ = A. This means the (i,j) entry always equals the (j,i) entry — the matrix looks the same above and below the diagonal. This is more than a curiosity: symmetric matrices arise constantly from inner products and quadratic forms, and their symmetry unlocks a remarkable set of structural properties.

Consider where symmetric matrices come from in practice. If you compute AᵀA for any matrix A, the result is always symmetric: (AᵀA)ᵀ = AᵀAᵀᵀ = AᵀA. Covariance matrices in statistics are symmetric for exactly this reason. Distance and similarity matrices are symmetric because distance from x to y equals distance from y to x. Any matrix of second derivatives (the Hessian in calculus) is symmetric when mixed partials are equal. These aren't coincidences — they reflect the underlying symmetry of the measurement being captured.

The most powerful consequence of symmetry concerns eigenvalues and eigenvectors. For a general square matrix, eigenvalues can be complex numbers even when all matrix entries are real — a rotation matrix, for instance, has complex eigenvalues. Symmetric matrices are different: all eigenvalues of a real symmetric matrix are guaranteed to be real. Even more strikingly, eigenvectors corresponding to *distinct* eigenvalues are guaranteed to be orthogonal — they point in perpendicular directions. This orthogonality follows directly from the symmetry: if Av = λv and Aw = μw with λ ≠ μ, then computing vᵀAw two ways using the symmetry forces v·w = 0.

Positive definiteness adds one more level of structure. A symmetric matrix A is positive definite if xᵀAx > 0 for every nonzero vector x. Geometrically, this quadratic form is always positive — A "opens upward" in every direction. The equivalent eigenvalue condition is clean: all eigenvalues are positive. Positive semidefinite relaxes this to ≥ 0. Covariance matrices are always positive semidefinite; they're positive definite when no data point is an exact linear combination of the others. These properties — real eigenvalues, orthogonal eigenvectors, positive definiteness — are precisely the hypotheses of the Spectral Theorem, which says symmetric matrices can be decomposed into orthogonal eigenvector directions. That decomposition is foundational to principal component analysis, quadratic optimization, and much of applied linear algebra.