The Setting
Consider a linear map from to real symmetric matrices () and the symmetric matrices formed from the identity matrix and the outer product of a unit vector with nonzero and . Because orthogonal projections onto kernels can jump discontinuously wherever the rank drops, the question is when the rank stabilizes locally so that the orthogonal projection onto the kernel of extends smoothly to an open set in and differentiates along directions tangent to the sphere.
More generally, for any smooth family of self-adjoint operators on a finite-dimensional real inner product space parametrized by a subset of a real normed space, one asks for local rank constancy and a smooth local extension of the orthogonal projection onto the kernel.
Definitions
For a nonnegative integer and a set , the space of square matrices over is the set of matrices with entries in indexed by rows and columns.✦ For a nonnegative integer , Euclidean space is the -dimensional real Euclidean vector space with standard inner product and norm ✦, and the space of real symmetric matrices is the subspace of self-adjoint matrices in .✦ For real normed spaces and , the space is the real normed space of bounded linear operators from to , and for a real normed space , the space of bounded linear operators is . For a finite-dimensional real normed space and an operator , the rank of is the dimension of the image of . For a nonnegative integer and a symmetric matrix , the rank of is the rank of the bounded linear operator on induced by .
Given a nonnegative integer and a vector , the outer product is the square matrix with entry ✦, and the symmetric outer product is the symmetric matrix with entry .✦ For a finite-dimensional real inner product space and an operator , the orthogonal projection onto the kernel of is the orthogonal projection of onto . For a nonnegative integer and a matrix , the kernel projection is the square matrix representing the orthogonal projection of onto the kernel of the Euclidean linear map defined by .✦
Decompose vectors as with and . The unit sphere is the set , and the domain is the subset . For a vector , the orthogonal complement of is the subspace . For a nonnegative integer , a linear map , the identity matrix , and a vector , the kernel projection is . For vectors and a scalar , the spherical tangent curve is given by (1): (1)✦
The Gram projection of a bounded linear operator between real Hilbert spaces and , with adjoint , is the bounded linear operator defined by (2): (2)✦ where is the inverse of when is invertible and otherwise.
The Claim
Theorem 1. Let be a real normed space, let be a finite-dimensional real inner product space, let be a nonempty subset, and let be a smooth () map on such that is self-adjoint for every . Then there exist and an open neighborhood of satisfying (3): (3)✦ There also exist , an open neighborhood of , and a map that is on such that (4) holds: (4)✦ Simultaneously, there exist , an open neighborhood of , and a map that is on satisfying (5): (5)✦
For any nonnegative integer and linear map , there exist and an open neighborhood of satisfying (6): (6)✦ For an integer and a linear map , there exist , an open neighborhood of , and a map that is on such that (7) holds: (7)✦ For an integer and a linear map , there exist and a linear map satisfying (8) for every : (8)✦
In the paper, the result reads: on , the matrix family (the paper’s ) attains its maximum rank at some , and on a neighborhood of in it has constant rank and (or if ) is smooth.
What Is Checked
The claims are stated over a smooth map from a real normed space to bounded linear operators on a finite-dimensional real inner product space that is self-adjoint on a nonempty subset , a bounded linear operator between real Hilbert spaces, and a linear map for a nonnegative integer (with in (7) and (8)) applied to on and together with the kernel projections and the spherical tangent curves along .