Section 8.5 Fredholm Alternatives
Definition 8.5.1.
A linear operator \(A:X\to Y\) has finite rank, if its range is finite dimensional.Proposition 8.5.2.
If \(A\) has finite rank on a Hilbert space \(H\text{,}\) and if for \(\lambda\neq 0\)
\begin{equation*}
\inf\{\|(A-\lambda I) x\|: \|x\|=1\} =0,
\end{equation*}
then \(\lambda\in \sigma_p(A)\text{.}\)Proof.
\begin{equation*}
\|(A-\lambda I) x_n\|\to 0, \qquad \mbox{for}\quad n\to \infty.
\end{equation*}
Since the range of \(A\) is finite, there exists a convergent subsequence in the range of \(A\text{:}\)
\begin{equation*}
\|A x_{n_j}-y\| \to 0, \qquad \mbox{for}\quad n\to \infty.
\end{equation*}
Then
\begin{equation*}
x_{n_j} = \frac{1}{\lambda} \Bigl(\underbrace{(\lambda I - A) x_{n_j}}_{\to 0 } + \underbrace{A x_{n_j}}_{\to y} \Bigr) \to \frac{1}{\lambda} y.
\end{equation*}
For the norm of \(x_{n_j}\) we have
\begin{equation*}
1=\|x_{n_j}\| \to 1 = \frac{1}{\lambda}\|y\|, \qquad \mbox{ and} \qquad \|y\|\neq 0.
\end{equation*}
Furthermore,
\begin{equation*}
A x_{n_j} \to y \qquad \mbox{and} \qquad A x_{n_j} \to \frac{1}{\lambda} A y,
\end{equation*}
which implies that \(Ay = \lambda y\) and \(\lambda\in \sigma_p(A)\text{.}\)Corollary 8.5.3.
Let \(H\) be a Hilbert space and \(A:H\to H\) a linear map with finite rank. Then for each \(\lambda\in \CC\) one of the following cases holds- \(\displaystyle \lambda\in \rho(A)\)
- \(\lambda\in \sigma_p(A)\) is an eigenvalue of finite multiplicity.
Proof.
[cross-reference to target(s) "p:finiterank" missing or not unique], there exists a \(c>0\) such that
\begin{equation*}
\|(A-\lambda I) x\| \geq c\|x\| \qquad \mbox{for all} \quad x\in H.
\end{equation*}
Consider \(y\in \overline{\mbox{Range}{(A-\lambda I)}}\text{.}\) Then there exists a sequence \(\{x_n\}_n\) with
\begin{equation*}
(A-\lambda I) x_n \to y, \qquad\mbox{as}\quad n\to\infty.
\end{equation*}
Then
\begin{equation*}
\|x_n - x_m\| \leq c^{-1} \|(A-\lambda I) x_n - (A-\lambda I) x_m\|
\end{equation*}
and \(\{x_n\}_n\) is a Cauchy sequence with a limit in \(H\text{:}\) \(x_n \to x\text{.}\) Consequently, \((A-\lambda I) x = y\) and \(R_\lambda(A)y\) exists. Hence \(\mbox{Range}(A-\lambda I) = \mbox{domain}(R_\lambda(A))\) is closed. Since \(\lambda\not\in \sigma_p(A)\) it is also not in \(\sigma_p(A^*)\) and
\begin{equation*}
\mbox{Range}(A-\lambda I) = (\mbox{ker}(A-\lambda I)^*)^\perp = H.
\end{equation*}
Now let
\begin{equation*}
T y := \{\mbox{ unique element } x \mbox{ such that } \quad (A-\lambda I)x =y\}.
\end{equation*}
Then \((A-\lambda I) Ty = y \) and
\begin{equation*}
c \|Ty\| \leq \|(A-\lambda I) Ty\| = \|y\| ,
\end{equation*}
which implies that \(T\) is bounded. In fact \(T=(A-\lambda I)^{-1}\) so \(\lambda\in \rho(A). \) If \(\lambda\in \sigma_p(A)\) then the eigenspace must be finite dimensional, since \(A\) has finite rank.Theorem 8.5.4.
Let \(G\subset \CC\) and consider an analytic map \(\lambda \mapsto F(\lambda)\in \mathcal{L}(H)\text{,}\) \(\lambda\in G\text{.}\) Assume \(F(\lambda)\) is of finite rank and that
\begin{equation*}
\mbox{Range}(F(\lambda)) \subset M, \qquad \mbox{with} \quad \mbox{dim}(M)<\infty.
\end{equation*}
Then one of the two alternatives holds. Either - \((I-F(\lambda))^{-1} \) exists for no \(\lambda\in G\text{,}\) or
- \((I-F(\lambda))^{-1} \) exists for every \(\lambda \in G\backslash S\text{,}\) where \(S\) is a discrete subset of \(G\text{.}\) In this case the map \(\lambda\mapsto (I-F(\lambda))^{-1}\) is analytic in \(G\backslash S\) and if \(\lambda\in S\) then \(F(\lambda)\Phi = \Phi\) has a finite dimensional family of solutions.
Proof.
\begin{equation*}
F(\lambda) \phi = \sum_{i=1}^n (\gamma_i(\lambda), \phi) \psi_i,
\end{equation*}
and the maps \(\lambda\mapsto \gamma_i(\lambda)\) are analytic in \(G\text{.}\) We define
\begin{equation*}
\Lambda_{ij}(\lambda ) := (\gamma_i(\lambda), \psi_j).
\end{equation*}
The inverse \((I- F(\lambda))^{-1}\) does not exist when \(F(\lambda)\phi = \phi\) has a nontrivial solution. We write this equation in components for \(\phi = \sum_{i=1}^n a_i \psi_i\) as
\begin{equation*}
\phi = \sum_{i=1}^n a_i \psi_i = F(\lambda) \phi = \sum_{i=1}^n \left(\gamma_i(\lambda), \sum_{j=1}^n a_j \psi_j\right) \psi_i.
\end{equation*}
Using orthogonality of the eigenfunctions, we find
\begin{equation*}
a_i = \sum_{j=1}^n (\gamma_i(\lambda), \psi_i) a_j = \sum_{j=1}^n \Lambda_{ij} a_j,
\end{equation*}
which we write in matrix notation as
\begin{equation*}
(I-\Lambda(\lambda))a = 0 ,\qquad a=(a_1,\dots, a_n), \quad \Lambda= (\Lambda_{ij}).
\end{equation*}
For a non-trivial solution we need
\begin{equation*}
d(\lambda)=\det(I-\Lambda(\lambda)) =0.
\end{equation*}
By the assumption \(d(\lambda)\) is analytic in \(\lambda\in G\text{,}\) hence \(d(\lambda)\) is either identically zero in \(G\) (case 1.), or the zeroes form a discrete set. Finally, since \(F(\lambda)\) has finite rank, we can only have finitely many solutions of \(F(\lambda)\phi = \phi\text{.}\)Theorem 8.5.5.
Let \(G\subset \CC\) and consider an analytic map \(\lambda \mapsto B(\lambda)\in \mathcal{L}(H)\text{,}\) \(\lambda\in G\text{.}\) Assume \(B(\lambda)\) is compact for each \(\lambda\in G\text{.}\) Then one of the two alternatives holds. Either- \((I-B(\lambda))^{-1} \) exists for no \(\lambda\in G\text{,}\) or
- \((I-B(\lambda))^{-1} \) exists for every \(\lambda \in G\backslash S\text{,}\) where \(S\) is a discrete subset of \(G\text{.}\) In this case the map \(\lambda\mapsto (I-B(\lambda))^{-1}\) is analytic in \(G\backslash S\) and if \(\lambda\in S\) then \(B(\lambda)\Phi = \Phi\) has a finite dimensional family of solutions.
Proof.
Example 8.5.6.
\begin{equation*}
u \mapsto Ku = \int_0^1 e^{-\gamma(x-y)} u(y) dy
\end{equation*}
in
\begin{equation*}
D(K) = (L^2(0,1))_+ = \{ u\in L^2(0,1): u\geq 0\}.
\end{equation*}
Since the kernel of the integral operator
\begin{equation*}
k(x,y) = e^{-\gamma(x-y)} \in L^2((0,1)\times (0,1)),
\end{equation*}
the operator \(K\) is a compact Hilbert-Schmidt operator. Hence it has a discrete spectrum of eigenvalues with a possible limit point at \(0\text{.}\) To solve \(Ku = \lambda u\text{,}\) that is
\begin{equation*}
\int_0^1 e^{-\gamma(x-y)} u(y) dy = \lambda u(x),
\end{equation*}
we multiply by the kernel \(e^{-\gamma(z-x)}\) and integrate:
\begin{align*}
\int_0^1 e^{-\gamma(z-x)}\int_0^1 e^{-\gamma(x-y)} u(y) \, dy\, dx \amp=\amp \lambda \int_0^1 e^{-\gamma(z-x)} u(x) dx\\
\amp=\amp \lambda^2 u(z).
\end{align*}
On the other hand we compute directly
\begin{align*}
\int_0^1 e^{-\gamma(z-x)}\int_0^1 e^{-\gamma(x-y)} u(y) \, dy\, dx \amp=\amp \int_0^1 \int_0^1 e^{-\gamma z+\gamma y} u(y) dy\, dx\\
\amp=\amp \int_0^1 e^{-\gamma(z-y)} u(y) dy\\
\amp=\amp \lambda u(z).
\end{align*}
Hence for \(u\neq 0\) we have \(\lambda^2 = \lambda\text{,}\) which implies \(\lambda =0\) or \(\lambda =1\text{.}\) If \(\lambda =0\text{,}\) then we have
\begin{equation*}
\int_0^1 e^{-\gamma(x-y)} u(y) dy =0
\end{equation*}
and since \(u\geq 0\) this implies that \(u=0\) a.e. in \([0,1]\text{.}\) Since we work in \(L^2\) this means \(u=0\text{,}\) and \(u\) is not an eigenfunction. Hence \(\lambda=0\) is not an eigenvalue. If \(\lambda = 1\text{,}\) then we need to solve
\begin{equation*}
\int_0^1e^{-\gamma(x-y)} u(y) dy = u(x).
\end{equation*}
We try \(u(x) = e^{-\gamma x}\text{:}\)
\begin{equation*}
\int_0^1e^{-\gamma(x-y)} u(y) dy = e^{-\gamma x} \int_0^1 e^{\gamma y} e^{-\gamma y} dy = e^{-\gamma x} = u(x).
\end{equation*}
Hence \(u(x) = e^{-\gamma x}\) is an eigenfunction of \(K\) with eigenvalue \(1\text{.}\) The spectrum is
\begin{equation*}
\sigma(K) = \{1\}.
\end{equation*}
Moreover, since \(k(x,y)>0\) we can apply the Krein-Rutman Theorem \cite{KreinRutman} (not covered here), and find that \(\lambda=1\) has multiplicity one with a unique positive eigenfucntion \(u(x) = e^{-\gamma x}\text{.}\) Theorem 8.5.7.
Assume \(A:H\to H\) is a compact linear map. Then \(\sigma(A)\) is compact and the only possible limit point is \(\lambda=0\text{.}\) Furthermore, given \(\lambda \in \CC\backslash\{0\}\) there are two alternatives. Either- \(\lambda\in \rho(A)\text{,}\) or
- \(\lambda\in \sigma_p(A)\) is an eigenvalue of finite multiplicity.
Proof.
Theorem 8.5.8.
Let \(A:H\to H\) be linear, compact and self-adjoint. Then all eigenvalues \(\lambda_i\) are real and there exists an orthonormal set of eigenfucntions \(\{\psi_i\}_i\) such that \(A\) has a spectral representation
\begin{equation*}
A\phi = \sum_{i=1}^\infty \lambda_i(\phi, \psi_i) \psi_i.
\end{equation*}
Proof.
\begin{equation*}
M:= \overline{\mbox{span}\{\psi_i, i=1, \dots, \infty\}}
\end{equation*}
and we show that \(\mbox{Range}(A)\subset M\text{.}\) First, since \(A\) is compact, we have \(AM=M\text{.}\) For \(\phi\in M^\perp\) we have that \((A\phi, \psi) = (\phi, A\psi) = 0\) for all \(\psi\in M\text{.}\) Hence \(A\phi\in M^\perp\) as well. This implies that both \(M\) and \(M^\perp\) are invariant under \(A\text{.}\) We define
\begin{equation*}
\hat A := A\Big|_{M^\perp}
\end{equation*}
and we will show that \(\hat A=0\text{.}\) Indeed, \(\hat A\) is self adjoint and compact (same as \(A\)), and each spectral value is an eigenvalue. But each eigenvector belongs already to \(M\text{.}\) Hence \(\sigma_p(\hat A) =\emptyset\text{.}\) Hence \(\hat A=0\text{.}\) It follows that \(\{\psi_i\}_i\) is a basis of \(\mbox{Range}(A)\) and we write
\begin{equation*}
A(u) =\sum_{i=1}^\infty (\psi_i, Au) \psi_i = \sum_{i=1}^\infty (A\psi_i, u) \psi_i = \sum_{i=1}^\infty \lambda_i(\psi_i, u) \psi_i.
\end{equation*}
