Best Low-Rank Approximation

Let have singular value decomposition

where . For , define the truncated SVD

Remark

Throughout, matrices are real and all orthogonality statements refer to the Frobenius inner product

Theorem (Eckart–Young–Mirsky)

Among all matrices of rank at most , is a best approximation to for every unitarily invariant norm. In particular,

and

The general result follows from Mirsky's singular-value inequality. The two common cases admit short direct proofs.

Frobenius-norm proof

Take any with , and let be the orthogonal projector onto the column space of . Since , the two terms in

are orthogonal in the Frobenius inner product. Hence

Also, and are orthogonal because

where . Therefore

Using the SVD of ,

For any and ,

The diagonal identity again uses . Thus the summands are mutually orthogonal, and

Set . Since an orthogonal projector is positive semidefinite and contractive,

Extend to an orthonormal basis of , and let

Since is orthogonal, cyclicity of the trace gives

Since is a symmetric idempotent matrix, it is orthogonally diagonalizable with eigenvalues in . Its rank counts the eigenvalues equal to , while its trace sums them. Hence

Therefore

Under these constraints, moving weight from a later index to an earlier one cannot decrease . Since , the maximum is attained at

and therefore

Combining this estimate with gives

Equality holds for , which proves the Frobenius-norm result.

If , this Frobenius-norm minimizer is unique. If , rotations inside the corresponding singular subspaces give other minimizers.

Spectral-norm proof

Let

Since , the intersection contains a nonzero vector. Choose a unit vector in this intersection. Then and

It follows that . On the other hand,

so . This gives the spectral-norm result.

Unlike the Frobenius-norm minimizer when , the spectral-norm minimizer need not be unique. Indeed, replacing the retained singular values in by nonzero values satisfying

still gives a rank- approximation with spectral-norm error .