A factorization of a matrix A into A = U Σ V^* where U and V are unitary (orthogonal in real case) matrices and Σ is a diagonal (rectangular) matrix of nonnegative singular values, exposing rank, range, and nullspace structure.

Principle

SVD provides orthonormal input and output bases that diagonalize the action of A up to nonnegative scalings: A maps the i-th right singular vector to the i-th left singular vector scaled by the i-th singular value. It exists for every matrix and is unique up to sign/phase for repeated singular values.

Demonstration

For A = [[3,0],[0,2]] an SVD is U=I, Σ=diag(3,2), V=I. In principal component analysis, the SVD of a centered data matrix identifies orthogonal principal directions and their variances (squared singular values). In least-squares, the SVD yields a stable solver and pseudoinverse by inverting nonzero singular values.

Misapplication

Using SVD results blindly on data without centering or normalization (misinterpreting global mean as signal), or truncating singular values without checking the effect on relevant observables. Confusing SVD with eigendecomposition of A (not A^*A) for non-symmetric matrices.

Consequence

SVD yields best low-rank approximations in both spectral and Frobenius norms, computes the Moore–Penrose pseudoinverse, exposes numerical rank and nullspace for regularization, and stabilizes linear inverse solutions.

Reversal

Treating the decomposition as requiring diagonalization by the same basis on both sides (i.e., assuming A is unitarily diagonalizable) — the reversed notion is eigendecomposition, which fails for most nonnormal or rectangular matrices.

Boundary

Applies to all finite matrices and compact operators on Hilbert spaces; extensions to tensors require different multilinear decompositions. SVD assumes an inner-product structure and produces orthonormal factors, so it excludes decompositions using nonorthogonal bases.

Semantic Tension

Tension exists between SVD and eigendecomposition/polar decomposition: all relate to operator structure but impose different basis constraints (orthonormal vs invariant directions) and different existence/uniqueness properties; practitioners must choose the decomposition that matches the problem structure.

Synthesis

SVD factorizes any matrix into orthonormal input/output bases and a diagonal of nonnegative scalings (singular values), revealing rank, energy distribution, and stable ways to invert or approximate the operator.