The Spectral Theorem

Core Idea

The Spectral Theorem states that every real symmetric matrix A is orthogonally diagonalizable: there exists an orthogonal matrix Q (with Qᵀ = Q⁻¹) and diagonal matrix D such that A = QDQᵀ. The columns of Q are orthonormal eigenvectors of A and D contains the corresponding real eigenvalues. This is stronger than ordinary diagonalization in two ways: the diagonalizing matrix Q is orthogonal (not just invertible), and real eigenvectors always exist. The spectral decomposition A = Σ λᵢuᵢuᵢᵀ writes A as a sum of rank-1 orthogonal projections, one per eigenvalue, revealing the geometric structure of the transformation.

How It's Best Learned

Orthogonally diagonalize a 2×2 symmetric matrix, verify Q is orthogonal, and reconstruct A = QDQᵀ. Then interpret the eigenvectors as the 'principal axes' of the quadratic form xᵀAx — an ellipse aligned with the eigenvectors.

Common Misconceptions

• The Spectral Theorem applies to SYMMETRIC matrices; it does not hold for general (non-symmetric) matrices, which may not be diagonalizable or may require complex eigenvectors.
• 'Spectral' refers to the spectrum (set of eigenvalues) of A, not to visible-light spectra.
• Orthogonal diagonalization requires Gram-Schmidt within each eigenspace when an eigenvalue has geometric multiplicity > 1; distinct eigenvalue eigenvectors are already orthogonal but may need normalization.

Explainer

You already know that a matrix can be diagonalized — written as PDP⁻¹ — when it has enough linearly independent eigenvectors. The Spectral Theorem says something far stronger holds for symmetric matrices: not just diagonalizable, but *orthogonally* diagonalizable. The diagonalizing matrix Q is not merely invertible — its columns are mutually perpendicular unit vectors. This means Qᵀ = Q⁻¹, so the decomposition A = QDQᵀ is both elegant and computationally stable.

Why does symmetry force this? Two deep facts work together. First, all eigenvalues of a real symmetric matrix are real — no complex numbers arise, even though the characteristic polynomial could in principle have complex roots. Second, eigenvectors corresponding to *distinct* eigenvalues are always orthogonal to each other. This is a theorem, not a coincidence: if Au = λu and Av = μv with λ ≠ μ, then ⟨u, v⟩ = 0 follows from the symmetry condition uᵀAv = (Au)ᵀv. Symmetric matrices represent transformations that act like "pure stretching" along perpendicular principal axes, with no rotational mixing.

The spectral decomposition A = Σ λᵢ uᵢuᵢᵀ is the most revealing form. Each term λᵢ uᵢuᵢᵀ is a rank-1 matrix that projects any vector onto the axis uᵢ and scales by λᵢ. When you multiply Av for any vector v, each projection extracts the component of v along uᵢ, scales it by λᵢ, and all the scaled components reassemble. The matrix acts as an independent stretch along each eigenvector axis — no cross-coupling between axes.

This principal axes interpretation becomes concrete with quadratic forms xᵀAx. A positive definite symmetric matrix defines an ellipsoid, and the eigenvectors of A give the orientation of that ellipsoid's principal axes while the eigenvalues give the stretching factors. Rotating to the eigenvector basis eliminates all cross-terms. This is the mathematical core of principal component analysis (PCA) in data science, spectral methods in graph theory, and the quantum mechanical treatment of observables.