SVD: Any matrix A = UΣVᵀ. Rank-k approx uses top k singular values: Aₖ = Σᵢ₌₁ᵏ σᵢ uᵢ vᵢᵀ. Compression ratio = k(m+n+1)/(mn). The singular values represent image "frequencies" — larger σ = more important structure.