The singular value decomposition writes any matrix, a picture included, as a sum of simple pieces σ·u·vᵀ: a column pattern times a row pattern, weighted by a singular value σ. Sorted from the largest σ down, the first few pieces carry most of the picture, so keeping only them compresses it.