I am reading the book 'Perturbation Theory for Linear Operators' from Kato.

He defines in his Book (Chapter 1 §5) the resolvent for some operator $T: X \to X$ on a finite Banach space $X$ as a operator-valued function $$ R(\zeta) = (T - \zeta )^{-1} . $$ This map is only defined for $\zeta$ that are different from any of the eigenvalues of $T$.

He then wants to investigate the Laurent series of $R(\zeta)$. The Laurent series is defined for any function $$ f: D\subset \mathbb{C} \to \mathbb{C} $$ which is holomorphic on $\{z \in \mathbb{C} \mid r < |z - z_0| < r'\}$ by $$ f(z) = \sum_{n= -\infty}^\infty c_n(z-z_0)^m$$ and $$ c_n = \frac{1}{2 \pi i } \int_{ |z-z_0| = \rho} \frac{f(z)}{(z-z_0)^{n+1}} dz $$ for an arbitrary $r < \rho <r'$.

But how can you investigate the Laurent Series for the resolvent $R$? It is not a map from $$D \subset \mathbb{C} \to \mathbb{C}$$ but from $$D \subset \mathbb{C} \to \mathcal{B}(X).$$ How can one see that the equality in the above equations still holds if I replace $f$ with the operator-value function $R$?

  • $\begingroup$ I thik it's a case of "use of function theory for operator-valued function", as suggested in the note 1 pag. 37 of Kato. $\endgroup$ – Emilio Novati Feb 11 '15 at 14:31
  • $\begingroup$ Since a Laurent series is absolutely convergent in his convergence annulus, I think that it matter the Th. 5.3.2 of: J. Deudonné - Foundations of modern analysys. $\endgroup$ – Emilio Novati Feb 11 '15 at 14:36

For finite-dimensional spaces, you can work in terms of matrices. This provides a smooth transition because you should be used to vector functions of a variable, which is all that a matrix really is. When you integrate a matrix $R(\lambda)=[r_{j,k}(\lambda)]_{j,k}$ where $1 \le j,k \le N$, you can just integrate each entry of the matrix separately. So, the integral of a matrix function is the matrix of integrals of the individual terms of the matrix: $$ \int R(\lambda)d\lambda = \left[\int r_{j,k}(\lambda)d\lambda\right]_{j,k}. $$ Matrices are nice because you know things about inversion of matrices. For example, the inverse of an $N\times N$ matrix $A=[a_{j,k}]$ comes out of the adjunct matrix $A_{\mbox{adj}}=[(-1)^{j+k}A_{k,j}]$ where $A_{j,k}$ is the determinant of the $(N-1)\times(N-1)$ co-factor matrix obtained by eliminating the $j$-th row and $k$-th column of $A$ (note that the swap of $j,k$ in the adjunct matrix for $A$ is deliberate.) This is because $$ AA_{\mbox{adj}}=A_{\mbox{adj}}A = \mbox{det}(A)I, $$ where $I$ is the $N\times N$ identity matrix. Therefore, $A$ is invertible iff $\mbox{det}(A)\ne 0$ and, in that case, the inverse is $$ A^{-1}=\frac{1}{\mbox{det}(A)}A_{\mbox{adj}}. $$ That gives you a way to study the resolvent operator through the characteristic polynomial $p(\lambda)=\mbox{det}(\lambda I- A)=(-1)^{N}\mbox{det}(A-\lambda I)$. The resolvent operator for $A$ is $$ (A-\lambda I)^{-1} = \frac{(-1)^{N}}{p(\lambda)}(A-\lambda I)_{\mbox{adj}} $$ All of the entries of the adjunct matrix $(A-\lambda I)_{\mbox{adj}}$ are formed from determinants of the entries of $(A-\lambda I)$, which means that the following is a matrix whose entries are polynomial $q_{j,k}(\lambda)$ whose orders never exceed $N-1$: $$ (A-\lambda I)_{\mbox{adj}}=[q_{j,k}(\lambda)] $$ By expanding all of the entries out and gathering like powers into separate matrices, you get $$ (A-\lambda I)_{\mbox{adj}}=A_{0}+\lambda A_{1}+\lambda^{2}A_{2}+\cdots+\lambda^{N-1}A_{N-1}, $$ where $A_{j}$ are constant matrices for all $0 \le j \le N-1$. So the resolvent always has the form $$ (A-\lambda I)^{-1} = (-1)^{N}\left[\frac{1}{p}A_{0}+\frac{\lambda}{p}A_{1}+\cdots+\frac{\lambda^{N-1}}{p}A_{N-1}\right]. $$ Now when you integrate, you can treat the matrices $A_{j}$ as constants and just worry about scalar integration of the coefficient functions $\lambda^{n}/p(\lambda)$. Everything reduces to scalar integrals with big fat constants--which are matrices--along for the ride. You can show that $$ q(A) = \frac{1}{2\pi i}\oint_{C} q(\lambda)(\lambda I-A)^{-1}d\lambda = -\frac{1}{2\pi i}\oint_{C} q(\lambda)(A-\lambda I)^{-1}d\lambda $$ for any scalar polynomial $q$ and any simple closed positively-oriented contour enclosing all of the roots of the characteristic polynomial $p$ in its interior. In particular, because $p(\lambda)(A-\lambda I)^{-1}$ has only removable singularities, you get the Cayley-Hamilton theorem right away: $$ p(A) = 0. $$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.