A procedure in an inner-product space that converts a finite linearly independent set of vectors into an orthogonal (or orthonormal) set spanning the same subspace by successively subtracting projections.

Principle

For each vector in sequence, remove its projection onto the span of previously orthogonalized vectors, then (optionally) normalize; this enforces pairwise orthogonality while preserving the span.

Demonstration

Given v1 = (1,1,0), v2 = (1,0,1) in R3 with the standard inner product, subtract projection of v2 on v1 to obtain an orthogonal vector u2 = v2 − (⟨v2,v1⟩/⟨v1,v1⟩) v1, then normalize u1,u2 to get an orthonormal basis of their span.

Misapplication

Applying the classical algorithm to a linearly dependent set yields zero vectors and must be handled carefully; numerically, the classical Gram–Schmidt is unstable for nearly dependent vectors, so using it naively in floating-point arithmetic can produce inaccurate orthogonality.

Consequence

Produces an orthogonal or orthonormal basis of the span, enables QR factorization of matrices with independent columns, and provides a constructive route to orthogonal projections and coordinate representations relative to an orthonormal basis.

Reversal

Replacing the sequential projection subtraction by a direct diagonalization (e.g., SVD) or by Householder reflections yields alternate orthonormalizing procedures; reversing Gram–Schmidt would mean combining orthogonal vectors back into the original dependent set, which is not generally unique.

Boundary

Applies to finite sequences in inner-product spaces; full orthonormal bases in infinite-dimensional Hilbert spaces require convergence considerations and orthonormalization of infinite sequences can fail if the sequence is not suitably controlled.

Semantic Tension

Competes with numerically stable orthonormalization methods (modified Gram–Schmidt, Householder, SVD): Gram–Schmidt is conceptually simple and constructive but can be numerically inferior; it is also distinct from statistical orthogonalization like PCA, which is data-driven and involves eigen-decomposition.

Synthesis

A stepwise projection-removal algorithm that transforms a finite independent set into an orthogonal (optionally normalized) basis of its span, foundational for QR factorization and for constructing orthogonal projections, but sensitive to dependence and numerical round-off.