Subspace Alignment is a domain adaptation method that aligns the source and target domains by finding and aligning their respective subspaces—learned through PCA or other dimensionality reduction techniques—so that the source classifier can be applied to target data projected into the aligned subspace. Subspace alignment assumes that domain shift primarily manifests as a rotation or transformation of the feature subspace rather than a change in the underlying data distribution within the subspace.

Why Subspace Alignment Matters in AI/ML: Subspace alignment provides a geometrically interpretable and computationally efficient approach to domain adaptation that captures the intuition that source and target data lie in different low-dimensional subspaces of the same ambient feature space, and alignment is achieved by finding the optimal rotation between them.

• PCA-based subspaces — Source and target feature matrices are decomposed via PCA: X_S ≈ U_S Σ_S V_S^T and X_T ≈ U_T Σ_T V_T^T; the top-d eigenvectors of each domain's covariance matrix define the domain's principal subspace; alignment operates on these subspace bases • Alignment transformation — The alignment matrix M = P_S^T P_T (where P_S, P_T are the d-dimensional PCA bases) maps the source subspace to the target subspace; source features are transformed: x̃_S = P_S M x_S, aligning them with the target's principal directions • Geodesic flow kernel (GFK) — An extension that models the continuous path (geodesic) between source and target subspaces on the Grassmann manifold; features are projected through all intermediate subspaces along this path, providing smoother and more robust alignment • Closed-form solution — Subspace alignment has a simple closed-form solution requiring only PCA and matrix multiplication, with no iterative optimization, no hyperparameter tuning beyond the subspace dimension d, and O(d³) computational cost • Limitations — Assumes domain shift is primarily a linear subspace transformation; fails when domains have fundamentally different feature structures, nonlinear shifts, or when important discriminative features lie outside the top-d principal components

Subspace alignment provides the geometric foundation for understanding domain adaptation as a subspace transformation problem, offering closed-form, interpretation-rich, and computationally efficient adaptation through PCA-based subspace discovery and alignment, establishing the geometric perspective that informs modern deep adaptation methods.