Frobenius covariant

From HandWiki

In matrix theory, the Frobenius covariants of a square matrix A are special polynomials of it, namely projection matrices Fi(A) associated with the eigenvalues and eigenvectors of A.[1]: pp.403, 437–8  They are named after the mathematician Ferdinand Frobenius.

Each covariant is a projection on the eigenspace associated with the eigenvalue λi. Frobenius covariants are the coefficients of Sylvester's formula, which expresses a function of a matrix f(A) as a matrix polynomial, namely a linear combination of that function's values on the eigenvalues of A.

Formal definition

Let A be a diagonalizable matrix with eigenvalues λ1, ..., λk.

The Frobenius covariant Fi(A), for i = 1,..., k, is the matrix Fi(A)≡∏j=1j≠ik1λi−λj(A−λjI).

It is essentially the Lagrange polynomial with matrix argument; alternatively, the same unique polynomial can be constructed using Newton polynomial. If the eigenvalue λi is simple, then as an idempotent projection matrix to a one-dimensional subspace, Fi(A) has a unit trace.

Computing the covariants

Ferdinand Georg Frobenius (1849–1917), German mathematician. His main interests were elliptic functions, differential equations, and later group theory.

The Frobenius covariants of a matrix A can be obtained from any eigendecomposition A = SDS−1, where S is non-singular and D is diagonal with Di,i = λi. The matrix S is defined up to multiplication on the right by a diagonal matrix. If A has no multiple eigenvalues, then let ci be the ith right eigenvector of A, that is, the ith column of S; and let ri be the ith left eigenvector of A, namely the ith row of S−1. Then Fi(A) = ci ri. As a projection matrix, the Frobenius covariant satisfies the relation Fi(A)Fj(A)=δijFi(A), which leads to ri cj = δij.

Given that v and w are the right and left vectors satisfying w Fi(A) v Template:≠ 0, the right and left eigenvectors of A may be written as ci = Nci Fi(A) v and ri = Nri w Fi(A). The orthonormality of the eigenvectors gives one constraint for the normalization coefficients. The remaining freedom is related to the choice of representation for the matrix S.

If A has an eigenvalue λi appearing multiple times, then Fi(A) = Σj cj rj, where the sum is over all rows and columns associated with the eigenvalue λi.[1]: p.521 

Example

Consider the two-by-two matrix: A=[1342]. This matrix has two eigenvalues, 5 and −2, which can be found by solving the characteristic equation. By virtue of the Cayley–Hamilton theorem, (A − 5)(A + 2) = 0.

The corresponding eigendecomposition is A=[31/74−1/7][500−2][31/74−1/7]−1=[31/74−1/7][500−2][1/71/74−3]. Hence the Frobenius covariants, manifestly projections, are F1(A)=c1r1=[34][1/71/7]=[3/73/74/74/7]=F12(A)F2(A)=c2r2=[1/7−1/7][4−3]=[4/7−3/7−4/73/7]=F22(A), with F1(A)F2(A)=0,F1(A)+F2(A)=I. Note tr F1(A) = tr F2 (A)= 1, as required.

Generalization

When a square matrix A∈ℂn×n has repeated eigenvalues (such as a defective matrix or a diagonalizable one), the standard definition of Frobenius covariants based on Lagrange polynomials becomes undefined due to division by zero.

In this general case, let λ1,…,λm be the distinct eigenvalues of A, each with an algebraic multiplicity of ni such that n=∑i=1mni. The generalized Frobenius covariants Ai∈ℂn×n are projection matrices Ai2=Ai) that project ℂn onto the generalized eigenspace corresponding to λi along the direct sum of the remaining generalized eigenspaces, satisfying ∑i=1mAi=I.

The generalized Frobenius covariants Ai can be constructed using several equivalent methods.

Construction via Hermite interpolation

To generalize the projection matrices for cases with repeated eigenvalues, one can use Hermite interpolating polynomials. The generalized Frobenius covariant Ai is evaluated by applying the polynomial to A as a matrix polynomial: Ai=hi,0(A), where hi,0(λ) is the degree-0 Hermite interpolating basis polynomial corresponding to node λi satisfying the interpolation conditions: hi,0(j)(λk)=δikδj0(1≤k≤m,0≤j<nk), where δab denotes the Kronecker delta.

Let hi,j(λ) denote the jth order derivative Hermite interpolating basis polynomial corresponding to node λi for 0≤j<ni. In terms of the Jordan–Chevalley decomposition, applying a higher-order basis polynomial hi,j(λ) to matrix A isolates the jth power of the nilpotent matrix Ni=Ai(A−λiI) associated with the Jordan blocks of λi (which is not a projection matrix for j>0). When j=0, the 0th power of the nilpotent matrix acts as the identity on that generalized eigenspace, reducing hi,0(A) to the projection matrix Ai.

The degree-0 Hermite interpolating basis polynomial hi,0(λ) can be constructed using any standard Hermite interpolation method, such as divided differences with confluent nodes, solving a confluent Vandermonde system, or via the Chinese remainder theorem.

Closed-form explicit formula

Based on the explicit formula for degree-0 Hermite interpolating basis polynomials, define the complementary factor polynomial qi(λ) for eigenvalue λi as qi(λ)=∏j≠i(λ−λj)nj. The degree-0 Hermite interpolating basis polynomial hi,0(λ) can be expressed directly in closed form: hi,0(λ)=[∑j=0ni−11j!(1qi(λ))(j)|λ=λi(λ−λi)j]⏟𝒯λini−1[1qi(λ)]qi(λ), where 𝒯ak[f(x)] denotes the kth degree Taylor operator of a function f(x) centered at x=a. Evaluating this degree-0 Hermite interpolating basis polynomial hi,0(λ) at A constructs the generalized Frobenius covariant Ai=hi,0(A).

Construction without Hermite interpolation

Besides constructing Hermite interpolating basis polynomials, the generalized Frobenius covariants Ai can also be constructed directly using several non-interpolation methods.

Jordan normal form

This approach generalizes the eigendecomposition method for distinct eigenvalues described in § Computing the covariants to arbitrary matrices with non-trivial Jordan blocks. If the Jordan normal form decomposition A=PJP−1 is known, Ai can be evaluated directly as Ai=PEiP−1, where Ei is a block diagonal matrix whose blocks corresponding to λi are identity matrices of matching dimensions, with all other entries set to zero.

Nullspace projection

Without finding Jordan chains, the generalized eigenspace 𝒱i=ker((A−λiI)ni) of dimension ni and its complementary space 𝒱ic=ker(∏r≠i(A−λrI)nr) of dimension n−ni decompose the vector space into a direct sum ℂn=𝒱i⊕𝒱ic. In general, 𝒱i and 𝒱ic are not orthogonal unless A is a normal matrix.

Let Qi,1∈ℂn×ni be a matrix whose columns form a basis for 𝒱i, and Qi,2∈ℂn×(n−ni) be a matrix whose columns form a basis for 𝒱ic. Assembling the partitioned basis matrix Qi=[Qi,1Qi,2] yields Ai=Qi[Ini000n−ni]Qi−1.

Holomorphic matrix functional calculus

Under the holomorphic matrix functional calculus, the generalized Frobenius covariant Ai can be defined directly without polynomial interpolation as a contour integral of the resolvent matrix (zI−A)−1: Ai=12πi∮Γi(zI−A)−1dz, where Γi is a simple closed contour in the complex plane that encloses λi in its interior while excluding all other eigenvalues of A.

References

  1. ↑ 1.0 1.1 Roger A. Horn and Charles R. Johnson (1991), Topics in Matrix Analysis. Cambridge University Press, ISBN 978-0-521-46713-1