The Setting
Fix a finite index set , and let be the space of complex square matrices indexed by . On , one considers the matrix energy and its gradient. At positive definite matrices, one seeks Fréchet and one-parameter derivatives of both the matrix square root and the energy on the positive semidefinite cone, while on the boundary one seeks one-parameter derivatives along positive semidefinite paths annihilating a fixed projection. One also seeks a spectral formula and unitary invariance for the Hessian quadratic cost, together with symmetry and pairing bounds for the associated bilinear forms.
Definitions
The cone of positive semidefinite matrices in is the set of matrices satisfying . A matrix is written when it is positive definite. For a map on and a matrix , the Fréchet derivative of at within is the Fréchet derivative over of at within . The real trace of a matrix is the real number .✦ The real Hilbert-Schmidt inner product of matrices is the real number , where denotes the conjugate transpose of .
The matrix energy of a matrix is the real number given by (1), (1)✦ where is the matrix square root. The energy gradient of a matrix is the functional on defined on by (2): (2)
The Sylvester operator associated with is the map sending to (3): (3)✦ When , is invertible with inverse on . The inverse root Sylvester operator of a positive definite matrix is the inverse of on .
The positive definite Hessian pairing of a positive definite matrix and directions is the real number given by (4): (4) The Sylvester penalty of matrices is the real number , and the Sylvester variational functional of matrices is the real number . The root variational cost of is the real number given by (5), (5)✦ The Hessian quadratic cost of matrices is the real number given by (6), and the polarized Hessian bilinear form of matrices is the real number given by (7): (6)✦ (7)✦
The spectral weight of and at indices is the real quantity (8) (with ): (8)✦
The Claim
Theorem 1. Suppose is positive definite. Then the matrix square root is over at within with Fréchet derivative (9), (9)✦ and the matrix energy is Fréchet differentiable over at within with derivative (10): (10)✦
Let be differentiable at with derivative . If and for all in a neighborhood of , then is differentiable at with (11): (11)✦ If is also differentiable at with derivative and is Hermitian, then is differentiable at with (12): (12)✦
Consider with and a path differentiable at with derivative . Suppose while and in a neighborhood of . Then is differentiable at with (13): (13)✦ If is also differentiable at with derivative , while is Hermitian and in a neighborhood of , then is differentiable at with (14): (14)✦
Given with for all and such that whenever , the diagonal matrix with diagonal entries satisfies (15) (with ): (15)✦ For any and , writing gives (16): (16)✦ Whenever is unitary, , and , the Hessian cost satisfies (17): (17)✦
For any positive definite matrix and all , the identities (18) and (19) and the bounds (20) and (21) hold: (18)✦ (19)✦ (20)✦ (21)✦ Take satisfying , , and . If satisfies , then (22) holds, (22)✦ and if also satisfies , then (23) holds: (23)✦
In the paper, the result reads: on the cone of matrices positive definite on a fixed support, is smooth, with and for Hermitian directions on that support; in an orthonormal eigenbasis with positive eigenvalues on the support, ; and in particular, is a positive definite symmetric bilinear form on the real space of supported Hermitian directions with . The theorem above represents a fixed support by a positive semidefinite projection with and using Fréchet and one-parameter derivatives of and , while the spectral formula splits into unitary invariance with diagonal evaluation and the pairing bounds take additive form.
What Is Checked
The claim is stated over the matrix energy , its gradient , the Sylvester operator and inverse , the root and Hessian costs and , the pairings and , and the spectral weight on , for positive definite and positive semidefinite matrices, unitary and diagonal matrices, and differentiable paths with a projection . For these constructions, the real trace and the energy gradient are continuous real-linear functionals on , every Sylvester operator is real-linear, and both (for ) and (for ) are continuous automorphisms of .