Section 9.7 The Hille-Yosida Theorem
The Hille-Yosida Theorem is the first result that will answer the question: When is an operator a generator of a \(C^0\)-semigroup?Theorem 9.7.1.
The operator \(A:X\to X\) is generator of a \(C^0\)-semigroup \(\{T(t)\}\) with exponential bound
\begin{equation}
\|T(t)\| \leq M e^{\omega t},\tag{9.7.1}
\end{equation}
if and only if - \(D(A)\) is dense in \(X\) and \(A\) is closed.
- Each \(\lambda>\omega\) is in the resolvent set \(\rho(A)\) and the resolvent is estimated as\begin{equation*} \|R_\lambda(A)^n\| \leq \frac{M}{(\lambda-\omega)^n}. \end{equation*}
Proof.
\begin{equation*}
I_n(\lambda) x := \frac{1}{(n-1)!}\int_0^\infty t^{n-1} e^{-\lambda t} T(t) x \, dt, \qquad \lambda\in \CC, \mbox{Re}\lambda>\omega, n\geq 1.
\end{equation*}
We will show that
\begin{equation*}
I_n(\lambda) = (-R_\lambda(A))^n\qquad \mbox{and} \qquad \|I_n(\lambda)\|\leq \frac{M}{(\mbox{Re}\lambda-\omega)^n }.
\end{equation*}
The inequality is straight-forward by using the exponential bound and integration by parts \((n-1)\)-times:
\begin{align}
\|I_n(\lambda) \| \amp\leq \amp \frac{M}{(n-1)!}\int_0^\infty \left| t^{n-1} e^{-\lambda t} e^{\omega t} \right| \, dt\nonumber\tag{9.7.2}\\
\amp=\amp \frac{M}{(n-1)!}\int_0^\infty t^{n-1} \left| e^{(\omega-\lambda) t} \right| \, dt\nonumber\tag{9.7.3}\\
\amp=\amp \frac{M}{(n-1)!}\int_0^\infty t^{n-1} e^{(\omega-\mbox{Re}\lambda) t} \, dt\nonumber\tag{9.7.4}\\
\amp=\amp \frac{M(-1)^{n-1}}{(\omega-\mbox{Re}\lambda)^{(n-1)}}\int_0^\infty e^{(\omega-\mbox{Re}\lambda) t} \, dt\nonumber\tag{9.7.5}\\
\amp=\amp \frac{M(-1)^{n-1}}{(\omega-\mbox{Re}\lambda)^{n}} (-1)\tag{9.7.6}\\
\amp=\amp \frac{M}{(\mbox{Re}\lambda - \omega)^n}.\tag{9.7.7}
\end{align}
For \(x\in D(A)\) and \(n>1\) we find
\begin{align}
I_n(\lambda) A x \amp=\amp \frac{1}{(n-1)!} \int_0^\infty t^{n-1} e^{-\lambda t} T(t) A x \, dt \nonumber\notag\\
\amp=\amp \frac{1}{(n-1)!} \int_0^\infty t^{n-1} e^{-\lambda t}\frac{d}{dt} (T(t) x) \, dt \nonumber\notag\\
\amp=\amp - \frac{1}{(n-2)!} \int_0^\infty t^{n-2} e^{-\lambda t} T(t) x \, dt + \frac{\lambda}{(n-1)!} \int_0^\infty t^{n-1} e^{-\lambda t} T(t) x \, dt \nonumber\notag\\
\amp=\amp -I_{n-1}(\lambda)x + \lambda I_n(\lambda) x.\tag{9.7.8}
\end{align}
Similarly, for \(n=1\) we get
\begin{align}
I_1(\lambda) Ax \amp=\amp \int_0^\infty e^{-\lambda t} T(t) A x \, dt \nonumber\tag{9.7.9}\\
\amp=\amp \int_0^\infty e^{-\lambda t} \frac{d}{dt} ( T(t) x) \, dt \nonumber\tag{9.7.10}\\
\amp=\amp - e^0 T(0) x +\lambda \int_0^\infty e^{-\lambda t} T(t) x \, dt \nonumber\tag{9.7.11}\\
\amp=\amp - x +\lambda I_1(\lambda) x.\tag{9.7.12}
\end{align}
Hence we get recursive relations between the \(I_n\text{.}\) From (Display Mathematics [STRUCT].[NUM]) we find
\begin{equation*}
I_n(\lambda) (Ax-\lambda x) = - I_{n-1}(\lambda) x
\end{equation*}
and since \(T\) and \(A\) commute, we also have
\begin{equation*}
(Ax-\lambda x) I_n(\lambda) = - I_{n-1}(\lambda) x.
\end{equation*}
Moreover
\begin{equation*}
I_1(\lambda)(A-\lambda I) = - I = (A-\lambda I) I_1(\lambda).
\end{equation*}
This implies that
\begin{equation}
R_\lambda(A)x = (A-\lambda I)^{-1}x = - I_1(\lambda)x = -\int_0^\infty e^{-\lambda t} T(t) x dt.\label{ResolventLaplace}\tag{9.7.13}
\end{equation}
This equation is an important relation in its own right, as it describes the resolvent as a Laplace transform of the semigroup! Starting with \(I_1\) we then argue recursively to find
\begin{equation*}
I_n(\lambda) = (-R_\lambda(A))^n.
\end{equation*}
Step 3: Item 1. and 2. are sufficient (Item 1.+2. \(\Longrightarrow\) semigroup). Now we come back to one of the ideas that we discussed in the introduction to this chapter; the idea to define an operator exponential as an inverse exponential limit
\begin{equation*}
T_n(t) := \left(I-\frac{t}{n}A \right)^{-n}.
\end{equation*}
We show \(T_n\to T\text{.}\) We rewrite \(T_n(t)\) as
\begin{equation*}
T_n(t) =\left(\frac{t}{n}\right)^{-n} \left(-R_{\frac{n}{t}} (A) \right)^n,
\end{equation*}
which is bounded by Item 2. for \(\frac{n}{t}>\omega\text{.}\) This is particularly bounded for large \(n\) and small \(t\text{.}\) To show that \(T_n(0)=I\text{,}\) it is sufficient to show that
\begin{equation*}
\lim_{t\to 0} \left(I-\frac{t}{n}A\right)^{-1} = I,
\end{equation*}
since the \(n\)-time iteration would then also be the identity. We use the relation for \(u\in D(A)\)
\begin{equation*}
u = \left(I-\frac{t}{n} A \right)^{-1}\left( I-\frac{t}{n} A\right) u = \left(I-\frac{t}{n} A\right)^{-1} u - \frac{t}{n} \left( I-\frac{t}{n} A\right)^{-1} A u.
\end{equation*}
Then
\begin{equation*}
\left\| \left(I-\frac{t}{n} A\right)^{-1} u - u \right\| = \frac{t}{n} \left\|\left(I-\frac{t}{n} A\right)^{-1} A u \right\| \leq C t \|Au\| \to 0,
\end{equation*}
for \(t\to 0\text{,}\) since \(\|Au\|\) is bounded for \(u\in D(A)\text{.}\) Hence \(T_n(0) u= u\) for a dense subset \(D(A)\text{,}\) and since \(T_n(0)\) is continuous, we have \(T_n(0)=I\) on \(X\text{.}\) \\ Next we want to show that indeed
\begin{equation*}
T(t) = \lim_{n\to \infty} \left(I-\frac{t}{n} A\right)^{-n}.
\end{equation*}
We compute
\begin{equation}
\frac{d}{dt} T_n(t) = -n \left(I-\frac{t}{n} A\right)^{-n-1}\left(-\frac{A}{n}\right) = A\left(I-\frac{t}{n} A\right)^{-n-1}.\label{HYvier}\tag{9.7.14}
\end{equation}
Again we argue for a dense subset, but this time for \(D(A^2)\subset D(A)\subset X\text{,}\) where it can be shown that \(D(A^2)\) is also dense by repeating the arguments in Theorem (Theorem 9.3.3). Let \(u\in D(A^2)\text{.}\) We show that \(\{T_n\}\) form a Cauchy sequence.
\begin{align}
\amp\amp T_n(t) u - T_m(t) u \nonumber\tag{9.7.15}\\
\amp=\amp \int_0^t \frac{d}{ds} \left( T_m(t-s) T_n(s) u \right) ds \nonumber\tag{9.7.16}\\
\amp=\amp \int_0^t \left(-\dot T_m(t-s) T_n(s) u + T_m(t-s) \dot T_n(s) u\right)ds\nonumber\tag{9.7.17}\\
\amp=\amp \int_0^t \left(-A\left(I-\frac{t-s}{m} A\right)^{-m-1} \left(I-\frac{s}{n} A\right)^{-n} u +\left(I-\frac{t-s}{m}A\right)^{-m} A \left(I-\frac{s}{n} A\right)^{-n-1} u \right) ds \nonumber\tag{9.7.18}\\
\amp=\amp \int_0^t \left(I-\frac{t-s}{m} A\right)^{-m-1} \left(I-\frac{s}{n} A\right)^{-n-1} \left[-A\left(I-\frac{s}{n}A\right)+ \left(I-\frac{t-s}{m} A\right)A \right]u ds \nonumber\tag{9.7.19}\\
\amp=\amp \int_0^t \underbrace{\left(I-\frac{t-s}{m} A\right)^{-m-1}}_{\mbox{bounded}} \underbrace{\left(I-\frac{s}{n} A\right)^{-n-1}}_{\mbox{bounded}} \left[\frac{s}{n} - \frac{t-s}{m} \right] A^2 u \; ds.\tag{9.7.20}
\end{align}
Then
\begin{align*}
\|T_n(t) u - T_m(t) u\|\amp\leq \amp C\|A^2 u\|\int_0^t \frac{s}{n} + \frac{t-s}{m} ds\\
\amp=\amp \frac{C t^2}{2} \left[ \frac{1}{n} +\frac{1}{m} \right] \|A^2 u\|.
\end{align*}
The last term is bounded, since \(u\in D(A^2)\) and it goes to zero for \(n,m\to\infty\text{.}\) Hence
\begin{equation*}
T(t) =\lim_{n\to \infty} T_n(t)
\end{equation*}
exists. \noindent Step 4: It remains to show that this \(T(t)\) is indeed our semigroup and that \(A\) is its infinitesimal generator. We know already that \(\lim_{t\to 0} T_n(t) = I\text{,}\) hence \(T(t)\) is strongly continuous. For the semigroup property we use equation (Display Mathematics [STRUCT].[NUM]) for \(n=m\text{,}\) without the integral, to obtain
\begin{equation*}
\frac{d}{ds} \left( T_n(t-s) T_n(s) u\right) = \frac{2s-t}{n} \left(I-\frac{t-s}{n} A\right)^{-n-1} \left( I-\frac{s}{n} A\right)^{-n-1} A^2 u.
\end{equation*}
In the limit as \(n\to \infty\) this becomes
\begin{equation*}
\frac{d}{ds} T(t-s) T(t) = 0 .
\end{equation*}
Hence \(T(t-s)T(s)\) is independent of \(s\text{.}\) Then it equals the value at \(s=0\text{,}\) which is \(T(t)\text{.}\) Finally, taking the limit as \(n\to \infty\) in ((9.7.14)) leads to
\begin{equation*}
\frac{d}{dt} T(t) = A T(t).
\end{equation*}
Corollary 9.7.2.
Let \(A\) be the generator of a strongly continuous semigroup \(T(t)\text{,}\) then
\begin{equation}
T(t) = \lim_{n\to \infty} \left( I-\frac{t}{n} A\right)^{-n} \tag{9.7.21}
\end{equation}
and
\begin{equation}
R_\lambda(A) = -\int_0^\infty e^{-\lambda t} T(t) dt.\label{ResolventLaplace2}\tag{9.7.22}
\end{equation}
Remark 9.7.3.
The formula for the semigroup (Corollary 9.7.2) is reminiscent to the discrete implicit Euler scheme for a numerical solver of \(\dot u = A u\text{.}\) Consider a small time increment \(\Delta t = \frac{t}{n}\text{.}\) Then the implicit Euler scheme is
\begin{equation*}
\frac{u(t+\Delta t) - u(t) }{\Delta t} = A u(t+\Delta t),
\end{equation*}
which gives
\begin{equation*}
\left(\frac{1}{\Delta t} I - A\right) u(t+\Delta t) = \frac{u(t)}{\Delta t}.
\end{equation*}
This is written as
\begin{equation*}
u(t+\Delta t) = (I-\Delta t A)^{-1} u(t).
\end{equation*}
An iteration of this scheme and taking \(\Delta t \to 0\) gives us something like (Corollary 9.7.2).Definition 9.7.4.
If \(T(t)\) is a strongly continuous semigroup with generator \(A\text{,}\) then we write
\begin{equation*}
T(t) = e^{At} .
\end{equation*}
Remark 9.7.5.
In the semigroup triangle Figure 9.3 we are able to add the connection from \(A\) to \(T\) and also the connection from \(T\) to \(R_\lambda\text{.}\) The only missing link is the arrow from \(R_\lambda\) to \(T\text{.}\) This can be filled in for analytic semigroups, which we develop in Section Section 9.11. \tikzstyle{semigr} = [ellipse, draw, fill=red!30, text width=3cm, text centered, minimum height=4em] \tikzstyle{gener} = [ellipse, draw, fill=yellow!40, text width=3cm, text centered, minimum height=4em] \tikzstyle{resol} = [ellipse, draw, fill=blue!20, text width=3cm, text centered, minimum height=4em] \tikzstyle{arrow} =[thick,->, >=stealth]
\begin{equation*}
\{T(t)\}_{t\geq 0}
\end{equation*}
Generator
\begin{equation*}
(A, D(A))
\end{equation*}
Resolvent
\begin{equation*}
R_\lambda(A)
\end{equation*}
\(\displaystyle T(t) = e^{At}\)\(\displaystyle Ax = \lim_{t\to 0^+} \frac{T(t)x - x}{t} \)\(\displaystyle R_\lambda(A) = -\int_0^\infty e^{-\lambda t} T(t) dt \)\(R_\lambda(A) = (A-\lambda I)^{-1}\)\(A = R_\lambda(A)^{-1} + \lambda I\)semigroup triangle