Projection onto the Fixed-Rank Tangent Space

Consider the set of matrices of rank exactly ,

Let have compact SVD

where and have orthonormal columns, and is invertible. Write

Remark

All orthogonality statements use the Frobenius inner product .

Theorem (Tangent and normal projections)

For every , the orthogonal projections onto the tangent and normal spaces at are

Block characterization of tangent directions

Choose orthonormal complements and , and define

In these orthonormal bases, has the block form

A matrix is tangent at if there is a differentiable rank- curve such that

Write the curve and the direction in these bases as

At , and the other three blocks vanish. Since is invertible, so is for sufficiently small . Block elimination gives

Because the curve has rank , its Schur complement vanishes:

The relation gives

Since and remains bounded, the Schur-complement identity gives . Comparing this with yields .

For the converse, take arbitrary blocks , , and of the appropriate dimensions, and set

Define

Its Schur complement is zero, so has rank for small . Since ,

Hence the tangent directions are exactly those with a zero lower-right block.

Tangent and normal blocks

The Frobenius inner product is unchanged by multiplication with the orthogonal matrices and , so the four blocks are mutually orthogonal. Since tangent directions have a zero lower-right block, the tangent space is

where , , and . Its orthogonal complement is

where . Thus the tangent projection keeps the first three blocks, while the normal projection keeps only the lower-right block. For any ,

Subtracting the normal component from gives

The projectors depend only on the column spaces of and , not on .