svm-model-fully-determined-by-support-vectors

IN derived (depth 1)

Created 2026-06-21T09:59:01+00:00 · Reviewed 2026-06-21T15:37:01+00:00

The entire SVM model — decision boundary, weight vector, and bias — is fully determined by the support vectors alone; all other training points are irrelevant to the learned classifier.

Justifications

SL — Decision boundary, weights, and bias all depend exclusively on support vectors

Antecedents (all must be IN):

IN svm-decision-boundary-depends-only-on-support-vectors — The SVM decision boundary depends only on the support vectors (points closest to the boundary), not on all training points, making SVMs memory-efficient at prediction time.
IN svm-weight-vector-linear-combination-support-vectors — The SVM weight vector w = sum(c_i * y_i * x_i) is a linear combination of support vectors only; most dual coefficients c_i are zero.
IN svm-bias-recovery-from-support-vector — The SVM bias term b is recovered from any support vector on the margin boundary using b = w^T x_i - y_i, exploiting that y_i^(-1) = y_i for labels in {-1, +1}.

Dependents

These beliefs depend on this one:

IN svm-mathematical-coherence-sparsity-and-equivalence — SVMs exhibit notable mathematical coherence — the model is fully determined by a sparse subset of training points (support vectors), and the soft-margin optimization admits three equivalent formulations (slack variables with margin constraints, hinge loss ERM with Tikhonov regularization, and the C-parameter tradeoff), providing both computational sparsity and multiple theoretical perspectives on the same underlying optimization.
IN svm-rare-complete-ml-framework — SVMs represent a notably coherent framework in ML — three decades of incremental development produced convex optimization with global optimality guarantees, kernel-enabled nonlinearity, and a model fully determined by a sparse subset of training points — a degree of mathematical closure that few other learning paradigms achieve.