1 Introduction

Let A be the generator of a bounded \(C_0\)-semigroup \((e^{At})_{t\ge 0}\) on a Hilbert space. Suppose that the inverse \(A^{-1}\) of A exists and also generates a \(C_0\)-semigroup \((e^{A^{-1}t})_{t\ge 0}\). The problem we study first is to estimate the decay rate of \(\Vert e^{A^{-1}t}A^{-1}\Vert \) under additional assumptions on stability of \((e^{At})_{t\ge 0}\). This is a variant of the so-called inverse generator problem raised by deLaubenfels [9]: “Let A be the generator of a bounded \(C_0\)-semigroup on a Banach space. Assume that there exists an inverse \(A^{-1}\) as a closed, densely-defined operator. Does \(A^{-1}\) also generate a bounded \(C_0\)-semigroup?” Positive answers to the inverse generator problem have been given for bounded holomorphic \(C_0\)-semigroups on Banach spaces [9] and for contraction \(C_0\)-semigroups on Hilbert spaces [29]. Negative examples can be found in [20] with respect to the generation property of \(A^{-1}\) and in [29] with respect to the boundedness property of \((e^{A^{-1}t})_{t\ge 0}\); see also [11, 16] for other counter-examples. For bounded \(C_0\)-semigroups on Hilbert spaces, the answer to the inverse generator problem is still unknown, but it has been shown in [28] that if A generates an exponentially stable \(C_0\)-semigroup on Hilbert spaces, then \(\Vert e^{A^{-1}t}\Vert = O(\log t)\) as \(t \rightarrow \infty \), that is, there exist \(M > 0\) and \(t_0 > 1\) such that \(\Vert e^{A^{-1}t}\Vert \le M \log t\) for all \(t \ge t_0\). This result has been extended to bounded \(C_0\)-semigroups with boundedly invertible generators in [5]. We refer the reader to the survey article [14] for more detailed discussion of the inverse generator problem.

It is known that if A generates an exponentially stable \(C_0\)-semigroup on a Banach space, then

$$\begin{aligned} \big \Vert e^{A^{-1}t} A^{-k}\big \Vert = O \left( \frac{1}{t^{k/2 - 1/4}} \right) \qquad (t \rightarrow \infty ) \end{aligned}$$
(1.1)

for all \(k \in {\mathbb {N}}\). This norm-estimate has been obtained in [29] for \(k=1\) and in [10] for \(k \ge 2\). In this paper, we obtain an analogous estimate for the Hilbert case: if A generates an exponentially stable \(C_0\)-semigroup on a Hilbert space, then

$$\begin{aligned} \big \Vert e^{A^{-1}t} A^{-k}\big \Vert = O \left( \frac{1}{t^{k/2}} \right) \qquad (t \rightarrow \infty ) \end{aligned}$$
(1.2)

for all \(k \in {\mathbb {N}}\). To obtain the estimate (1.1) of the Banach case, an integral representation of \((e^{A^{-1}t})_{t \ge 0}\) (see [29, Lemma 3.2] and [16, Theorem 1]) has been applied. In contrast, to obtain the estimate (1.2) of the Hilbert case, we employ a functional-calculus approach based on the \({\mathcal {B}}\)-calculus introduced in [5]. Another class of \(C_0\)-semigroups we study are polynomially stable \(C_0\)-semigroups with parameter \(\beta >0\) on Hilbert spaces, where \((e^{At})_{t\ge 0}\) is called a polynomially stable \(C_0\)-semigroup with parameter \(\beta >0\) if \((e^{At})_{t\ge 0}\) is bounded and if \(\Vert e^{At}(I-A)^{-1}\Vert = O(t^{-1/\beta })\) as \(t \rightarrow \infty \). By a functional-calculus approach as in the case for exponentially stable \(C_0\)-semigroups, we show that if, in addition, A is a normal operator, then

$$\begin{aligned} \big \Vert e^{A^{-1}t} A^{-k}\big \Vert = O \left( \frac{1}{t^{k/(2+\beta )}} \right) \qquad (t \rightarrow \infty ) \end{aligned}$$
(1.3)

for all \(k \in {\mathbb {N}}\). We also present simple examples for which the norm-estimates (1.2) and (1.3) cannot be improved. Moreover, we prove that the (not necessarily normal) generator A of a polynomially stable \(C_0\)-semigroup with any parameter on a Hilbert space satisfies \(\sup _{t \ge 0} \Vert e^{A^{-1}t}A^{-1}\Vert < \infty \), inspired by the Lyapunov-equation technique developed in [28].

Next, we study the Crank–Nicolson discretization scheme with variable stepsizes. Let A be the generator of a bounded \(C_0\)-semigroup on a Hilbert space, and let \(\tau _n\), \(n \in {\mathbb {N}}_0:= \{ 0,1,2,\ldots \}\), satisfy \(\tau _{\min } \le \tau _n \le \tau _{\max }\) for some \(\tau _{\max } \ge \tau _{\min } >0\). Consider the time-varying difference equation

$$\begin{aligned} x_{n+1} = A_d(\tau _n)x_n,\quad n \in {\mathbb {N}}_0, \end{aligned}$$
(1.4)

where \(A_d(\tau )\) is the Cayley transform of A with parameter \(\tau >0\), i.e.,

$$\begin{aligned} A_d(\tau ):= \left( I + \frac{\tau }{2}A \right) \left( I - \frac{\tau }{2}A \right) ^{-1}. \end{aligned}$$

It is our aim to estimate the decay rate of the solution \((x_n)_{n \in {\mathbb {N}}}\) with smooth initial data \(x_0 \in D(A)\).

For a constant stepsize \(\tau >0\), upper bounds on the growth rate of \((A_d(\tau )^n)_{n \in {\mathbb {N}}_0}\) have been obtained for bounded \(C_0\)-semigroups and exponentially stable \(C_0\)-semigroups on Banach spaces; see [6, 15]. The relation between \((e^{A^{-1}t})_{t \ge 0}\) and \((A_d(\tau )^n)_{n \in {\mathbb {N}}_0}\) in terms of growth rates has been investigated in [15]. As in the inverse generator problem, \((A_d(\tau )^n)_{n \in {\mathbb {N}}_0}\) is bounded whenever \(C_0\)-semigroups are bounded holomorphic on Banach spaces [7]. Furthermore, it is well known that for contraction \(C_0\)-semigroups on Hilbert spaces, \(A_d(\tau )\) is a contraction; see [22]. It is still open whether \((A_d(\tau )^n)_{n \in {\mathbb {N}}_0}\) is bounded for every bounded \(C_0\)-semigroup on a Hilbert space. However, it has been shown independently in [1, 13, 17] that if A and \(A^{-1}\) both generate bounded \(C_0\)-semigroups on a Hilbert space, then \((A_d(\tau )^n)_{n \in {\mathbb {N}}_0}\) is bounded. If in addition \((e^{At})_{t\ge 0}\) is strongly stable, then \((A_d(\tau )^n)_{n \in {\mathbb {N}}_0}\) is strongly stable [17]. Without the assumption on \(A^{-1}\), the norm-estimate \(\Vert A_d(\tau )^n\Vert = O(\log n)\) as \(n \rightarrow \infty \) given in [13] remains the best so far in the Hilbert case. For polynomially stable \(C_0\)-semigroups with parameter \(\beta >0\) on Hilbert spaces, the following estimates have been obtained in [25]:

  • \( \Vert A_d(\tau )^nA^{-1}\Vert = O\bigg (\left( \dfrac{\log n}{n}\right) ^{1/(2+\beta )}\bigg ) \) if \((A_d(\tau )^n)_{n \in {\mathbb {N}}_0}\) is bounded.

  • \(\Vert A_d(\tau )^nA^{-1}\Vert = O\left( \dfrac{1}{n^{1/(2+\beta )}} \right) \) if A is normal.

For more information on asymptotics of \((A_d(\tau )^n)_{n \in {\mathbb {N}}_0}\), we refer to the survey [14].

Some of the above results have been extended to the case of variable stepsizes. For bounded \(C_0\)-semigroups on Banach spaces, the Crank–Nicolson scheme with variable stepsizes has the same growth bound as that with constant stepsizes; see [2]. It has been proved that the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the time-varying difference equation (1.4) is bounded if A generates a bounded holomorphic \(C_0\)-semigroups on a Banach space [21, 23, 24] or if A and \(A^{-1}\) generate bounded \(C_0\)-semigroups on a Hilbert space [23]. The stability property, i.e., \( \lim _{n \rightarrow \infty } x_n = 0 \) for all initial data \(x_0\), has been also obtained under some additional assumptions in [23].

We show that if A and \(A^{-1}\) are the generators of an exponentially stable \(C_0\)-semigroup and a bounded \(C_0\)-semigroup on a Hilbert space, respectively, then the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the time-varying difference equation (1.4) with \(x_0 \in D(A)\) satisfies

$$\begin{aligned} \Vert x_n\Vert = O\left( \sqrt{\frac{\log n}{n}} \right) \qquad (n \rightarrow \infty ). \end{aligned}$$
(1.5)

Moreover, if A is a normal operator generating a polynomially stable \(C_0\)-semigroup with parameter \(\beta >0\) on a Hilbert space, then

$$\begin{aligned} \Vert x_n\Vert = O\left( \frac{1}{n^{1/(2+\beta )}} \right) \qquad (n \rightarrow \infty ) \end{aligned}$$
(1.6)

for all \(x_0 \in D(A)\). These estimates (1.5) and (1.6) follow from the combination of the norm-estimates (1.2) and (1.3) of \(e^{A^{-1}t}A^{-1}\) and the Lyapunov-equation technique developed in [23]. In the constant case \(\tau _n \equiv \tau >0\), we give a simple example of an exponentially stable \(C_0\)-semigroup showing that the decay rate \(\Vert A_d(\tau )^n A^{-1}\Vert = O(1/\sqrt{n})\) cannot in general be improved. The estimate (1.5) includes the logarithmic factor \(\sqrt{\log n}\), but its necessity remains open. On the other hand, the decay rate given in (1.6) is the same as that obtained for constant stepsizes in [25], and one cannot in general replace \(n^{-1/(2+\beta )}\) in (1.6) by functions with better decay rates.

This paper is organized as follows. In Sect. 2, we give long-time norm-estimates for \(e^{A^{-1}t}A^{-1}\). We consider first exponentially stable \(C_0\)-semigroups and then polynomially stable \(C_0\)-semigroups. In Sect. 3, the decay rate estimates of \(\Vert e^{A^{-1}t}A^{-1}\Vert \) are utilized in order to examine the quantified asymptotic behavior of the Crank–Nicolson scheme with smooth initial data.

Notation Let \({\mathbb {C}}_+:= \{ \lambda \in {\mathbb {C}}: \mathop {\textrm{Re}}\nolimits \lambda >0 \}\), \(\overline{{\mathbb {C}}}_+:= \{ \lambda \in {\mathbb {C}}: \mathop {\textrm{Re}}\nolimits \lambda \ge 0 \}\), \({\mathbb {C}}_-:= \{ \lambda \in {\mathbb {C}}: \mathop {\textrm{Re}}\nolimits \lambda <0 \}\), and \(i {\mathbb {R}}:= \{ i \eta : \eta \in {\mathbb {R}} \}\). Let \({\mathbb {N}}_0\) and \({\mathbb {R}}_+\) be the set of nonnegative integers and the set of nonnegative real numbers, respectively. For real-valued functions fg on \(J \subset {\mathbb {R}}\), we write

$$\begin{aligned} f(t) = O\big (g(t)\big )\qquad (t \rightarrow \infty ) \end{aligned}$$

if there exist constants \(M >0\) and \(t_0 \in J\) such that \(f(t) \le M g(t)\) for all \(t \in J\) satisfying \(t \ge t_0\). The gamma function is denoted by \(\Gamma \), i.e.,

$$\begin{aligned} \Gamma (\alpha ) = \int ^{\infty }_0 t^{\alpha -1} e^{-t}dt,\quad \alpha > 0. \end{aligned}$$

Let H be a Hilbert space. The inner product of H is denoted by \(\langle \cdot , \cdot \rangle \). The space of bounded linear operators on H is denoted by \({\mathcal {L}}(H)\). For a linear operator A on H, let D(A) and \(\sigma (A)\) denote the domain and the spectrum of A, respectively. For a densely-defined linear operator A on H, we denote by \(A^*\) the Hilbert space adjoint of A. If A is the generator of a bounded \(C_0\)-semigroup on a Hilbert space and if A is injective, the fractional power \((-A)^{\alpha }\) of \(-A\) is defined by the sectorial functional calculus for \(\alpha \in {\mathbb {R}}\) as in [18, Chapter 3]. Let \(\ell ^2\) be the space of complex-valued square-summable sequences \(x = (x_n)_{n \in {\mathbb {N}}}\) endowed with the norm \(\Vert x\Vert := \sqrt{\sum _{n=1}^{\infty } |x_n|^2}\).

2 Norm-Estimates of Semigroups Generated by Inverse Generators

The aim of this section is to derive a norm-estimate for \(e^{A^{-1}t} (-A)^{-\alpha }\) with \(\alpha >0\). First we assume that A is the generator of an exponentially stable \(C_0\)-semigroup on a Hilbert space. Next we focus on the generator A of a polynomially stable \(C_0\)-semigroup on a Hilbert space.

2.1 Inverses of Generators of Exponentially Stable Semigroups

Let A be the generator of an exponentially stable \(C_0\)-semigroup on a Hilbert space. Then A is invertible, and \(A^{-1}\) generates a \(C_0\)-semigroup \((e^{A^{-1}t})_{t \ge 0}\). To obtain a norm-estimate for \(e^{A^{-1}t} (-A)^{-\alpha }\), we employ the \({\mathcal {B}}\)-calculus introduced in [5].

2.1.1 Basic Facts on \(\varvec{{\mathcal {B}}}\)-Calculus

Let \({\mathcal {B}}\) be the space of all holomorphic functions f on \({\mathbb {C}}_+\) such that

$$\begin{aligned} \Vert f\Vert _{{\mathcal {B}}_0}:= \int ^{\infty }_0 \sup _{\eta \in {\mathbb {R}}} |f'(\xi +i\eta )|d\xi < \infty . \end{aligned}$$

We recall elementary properties of functions f in \({\mathcal {B}}\). The proof can be found in [5, Proposition 2.2].

Proposition 2.1

For \(f \in {\mathcal {B}}\), the following statements hold:

  1. (i)

    \(f(\infty ):= \lim _{\mathop {\textrm{Re}}\nolimits z \rightarrow \infty } f(z)\) exists in \({\mathbb {C}}\).

  2. (ii)

    f is bounded on \({\mathbb {C}}_+\), and \(\Vert f\Vert _{\infty }:= \sup _{z \in {\mathbb {C}}_+} |f(z)| \le |f(\infty )| + \Vert f\Vert _{{\mathcal {B}}_0} \).

  3. (iii)

    \(f(i\eta ):= \lim _{\xi \rightarrow 0 +} f(\xi +i\eta )\) exists, uniformly for \(\eta \in {\mathbb {R}}\).

A norm on \({\mathcal {B}}\) is defined by

$$\begin{aligned} \Vert f\Vert _{{\mathcal {B}}}:= \Vert f\Vert _{\infty } + \Vert f\Vert _{{\mathcal {B}}_0}, \end{aligned}$$

which is equivalent to each norm of the form

$$\begin{aligned} |f(z_0)| + \Vert f\Vert _{{\mathcal {B}}_0}\qquad (z_0 \in \overline{{\mathbb {C}}}_+ \cup \{ \infty \}). \end{aligned}$$

The space \({\mathcal {B}}\) equipped with the norm \(\Vert \cdot \Vert _{{\mathcal {B}}}\) is a Banach algebra by [5, Proposition 2.3].

Let \(\text {M}({\mathbb {R}}_+)\) be the Banach algebra of all bounded Borel measures on \({\mathbb {R}}_+\) under convolution, endowed with the norm \(\Vert \mu \Vert _{\text {M}({\mathbb {R}}_+)}:= |\mu |({\mathbb {R}}_+)\), where \(|\mu |\) is the total variation of \(\mu \). For \(\mu \in \text {M}({\mathbb {R}}_+)\), the Laplace transform of \(\mu \) is the function

$$\begin{aligned} {\mathcal {L}}\mu :\overline{{\mathbb {C}}}_+ \rightarrow {\mathbb {C}},\quad ({\mathcal {L}}\mu )(z) = \int _{{\mathbb {R}}_+} e^{-zt} \mu (dt). \end{aligned}$$

We define

$$\begin{aligned} {{\mathcal {L}}}{{\mathcal {M}}}:= \{ {\mathcal {L}}\mu : \mu \in \text {M}({\mathbb {R}}_+) \}. \end{aligned}$$

Then the space \({{\mathcal {L}}}{{\mathcal {M}}}\) endowed with the norm \(\Vert {\mathcal {L}}\mu \Vert _{\text {HP}}:= \Vert \mu \Vert _{\text {M}({\mathbb {R}}_+)}\) becomes a Banach algebra. As shown on [5, p. 42], \({{\mathcal {L}}}{{\mathcal {M}}}\) is a subspace of \({\mathcal {B}}\) with continuous inclusion.

Let \(-B\) be the generator of a bounded \(C_0\)-semigroup \((e^{-Bt})_{t \ge 0}\) on a Hilbert space H. Define \(K:= \sup _{t \ge 0} \Vert e^{-Bt}\Vert \). Using Plancherel’s theorem and the Cauchy-Scwartz inequality, we obtain

$$\begin{aligned} \int _{{\mathbb {R}}} \left| \langle (\xi + i \eta +B)^{-2}x,y \rangle \right| d\eta \le \frac{\pi K^2}{\xi } \Vert x\Vert \, \Vert y\Vert \end{aligned}$$

for all \(x,y \in H\) and \(\xi >0\); see [5, Example 4.1]. From this observation, we define

$$\begin{aligned}{} & {} \langle f(B)x,y \rangle \nonumber \\{} & {} := f(\infty ) \langle x,y \rangle - \frac{2}{\pi } \int ^{\infty }_0 \xi \int _{{\mathbb {R}}} \langle (\xi - i\eta + B)^{-2} x, y \rangle f'(\xi +i\eta ) d\eta d\xi \end{aligned}$$
(2.1)

for \(f \in {\mathcal {B}}\) and \(x,y \in H\). This definition yields a bounded functional calculus; see [5, Theorem 4.4] for the proof.

Theorem 2.2

Let \(-B\) be the generator of a bounded \(C_0\)-semigroup \((e^{-Bt})_{t \ge 0}\) on a Hilbert space H, and let f(B) be defined as in (2.1). Then the following statements hold:

  1. (i)

    The map \(\Phi _B:f \mapsto f(B)\) is a bounded algebra homomorphism from \({\mathcal {B}}\) into \({\mathcal {L}}(H)\).

  2. (ii)

    If \(f = {\mathcal {L}}\mu \in {{\mathcal {L}}}{{\mathcal {M}}}\), then f(B) coincides with the operator

    $$\begin{aligned} x \mapsto \int _{{\mathbb {R}}_+} e^{-Bt}x \mu (dt) \end{aligned}$$

    as defined in the Hille-Phillips calculus.

  3. (iii)

    If \(K:= \sup _{t \ge 0} \Vert e^{-Bt}\Vert \), then

    $$\begin{aligned} \Vert f(B)\Vert \le |f(\infty )| + 2K^2\Vert f\Vert _{{\mathcal {B}}_0} \le 2K^2\Vert f\Vert _{{\mathcal {B}}} \end{aligned}$$

    for all \(f \in {\mathcal {B}}\).

The map

$$\begin{aligned} \Phi _B: {\mathcal {B}} \rightarrow {\mathcal {L}}(H),\quad f \mapsto f(B) \end{aligned}$$

is called the \({\mathcal {B}}\)-calculus for B.

2.1.2 Norm-Estimate by \(\varvec{{\mathcal {B}}}\)-Calculus

By Theorem 2.2, the norm-estimate for a function \(f \in {\mathcal {B}}\) implies the norm-estimate for the corresponding operator f(B). In [5, Lemma 3.4], the estimate

$$\begin{aligned} \Vert h_t\Vert _{{\mathcal {B}}} = O\left( \log t\right) \qquad (t \rightarrow \infty ),\qquad \text {where}\ h_t(z) := e^{-t/(z+1)}, \end{aligned}$$
(2.2)

has been derived for the norm-estimate of \(e^{A^{-1}t}\). To estimate \(\Vert e^{A^{-1}t} (-A)^{-\alpha }\Vert \) for \(\alpha >0\), the following result is used.

Lemma 2.3

For \(t,\alpha > 0\), define

$$\begin{aligned} f_{t,\alpha }(z):= \frac{e^{-t/(z+1)}}{(z+1)^\alpha },\quad z \in {\mathbb {C}}_+. \end{aligned}$$

Then \(f_{t,\alpha } \in {\mathcal {B}}\) for all \(t,\alpha >0\). Moreover,

$$\begin{aligned} \Vert f_{t,\alpha }\Vert _{{\mathcal {B}}_0} = O\left( \frac{1}{t^{\alpha /2}}\right) \qquad (t \rightarrow \infty ) \end{aligned}$$
(2.3)

for each \(\alpha >0\).

Proof

For \(t >0\) and \(\alpha > 1\), define

$$\begin{aligned} F_\alpha (t):= \int ^{\infty }_0 \sup _{\eta \in {\mathbb {R}}} \left| \frac{e^{-t/(\xi +i\eta +1)}}{(\xi +i\eta +1)^\alpha }\right| d\xi . \end{aligned}$$

Since

$$\begin{aligned} f_{t,\alpha }'(z) = \frac{te^{-t/(z+1)} }{(z+1)^{\alpha +2}} - \frac{\alpha e^{-t/(z+1)} }{(z+1)^{\alpha +1}} \end{aligned}$$

for all \(z \in {\mathbb {C}}_+\), we obtain

$$\begin{aligned} \int ^{\infty }_0 \sup _{\eta \in {\mathbb {R}}} |f_{t,\alpha }'(\xi +i\eta )|d\xi \le tF_{\alpha +2}(t) + \alpha F_{\alpha +1}(t) \end{aligned}$$
(2.4)

for all \(t,\alpha > 0\).

Let \(\alpha >1\). Define

$$\begin{aligned} g_{t,s,\alpha }(r):= \frac{e^{-ts/(s^2+r)}}{(s^2+r)^{\alpha /2}} \end{aligned}$$

for \(t>0\), \(s >1\), and \(r \ge 0\). Then

$$\begin{aligned} F_{\alpha }(t) = \int ^{\infty }_1 \sup _{r \ge 0} g_{t,s,\alpha }(r)ds \end{aligned}$$

for all \(t > 0\). Put \(c:= \left( \alpha /(2e)\right) ^{\alpha /2}\). We have that for each \(s > 1\),

$$\begin{aligned} \sup _{r \ge 0} g_{t,s,\alpha }(r) = {\left\{ \begin{array}{ll} g_{t,s,\alpha }(0) = \dfrac{e^{-t/s}}{s^\alpha }, &{} 0 < t \le \dfrac{\alpha s}{2}, \vspace{6pt}\\ g_{t,s,\alpha }\left( \dfrac{2ts}{\alpha }-s^2 \right) = \dfrac{c}{(ts)^{\alpha /2}}, &{} t \ge \dfrac{\alpha s}{2} > \dfrac{\alpha }{2}. \end{array}\right. } \end{aligned}$$

If \(0 < t \le \alpha /2\), then

$$\begin{aligned} \int ^{\infty }_1 \sup _{r \ge 0} g_{t,s,\alpha }(r)ds&= \int ^{\infty }_1 \frac{e^{-t/s}}{s^\alpha }ds = \frac{1}{t^{\alpha -1}} \int ^t_0 u^{\alpha -2}e^{-u} du \le \frac{\Gamma (\alpha -1)}{t^{\alpha -1}}. \end{aligned}$$

If \(t > \alpha /2\), then

$$\begin{aligned} \int ^{\infty }_1 \sup _{r \ge 0} g_{t,s,\alpha }(r)ds&= \frac{c}{t^{\alpha /2}} \int ^{2t/\alpha }_{1} \frac{1}{s^{\alpha /2}}ds + \int ^{\infty }_{2t/\alpha } \frac{e^{-t/s}}{s^\alpha } ds. \end{aligned}$$

We write the first term on the right-hand side as

$$\begin{aligned} \frac{c}{t^{\alpha /2}} \int ^{2t/\alpha }_{1} \frac{1}{s^{\alpha /2}}ds&= {\left\{ \begin{array}{ll} \dfrac{c \log t}{t}, &{}{} \alpha = 2, \\ \dfrac{c}{\alpha /2-1} \left( \dfrac{1}{t^{\alpha /2}} - \dfrac{(\alpha /2)^{\alpha /2-1}}{t^{\alpha -1}} \right) , &{}{} \alpha \not =2. \end{array}\right. }\end{aligned}$$

The second term satisfies

$$\begin{aligned} \int ^{\infty }_{2t/\alpha } \frac{e^{-t/s}}{s^\alpha } ds = \frac{1}{t^{\alpha -1}} \int ^{\alpha /2}_0 u^{\alpha -2}e^{-u} du \le \frac{\Gamma (\alpha -1)}{t^{\alpha -1}}. \end{aligned}$$

Therefore, \((0\le )\, F_\alpha (t) < \infty \) for each \(t >0\), and

$$\begin{aligned} F_\alpha (t) = {\left\{ \begin{array}{ll} O\left( \dfrac{1}{t^{\alpha -1}}\right) , &{}{} 1<\alpha < 2, \\ O\left( \dfrac{\log t}{t}\right) , &{}{} \alpha = 2,\\ O\left( \dfrac{1}{t^{\alpha /2}}\right) , &{}{} \alpha >2. \end{array}\right. } \end{aligned}$$
(2.5)

From the estimate (2.4), we obtain \(f_{t,\alpha } \in {\mathcal {B}}\) for all \(t,\alpha >0\). Since (2.5) yields

$$\begin{aligned} tF_{\alpha +2}(t) + \alpha F_{\alpha +1}(t) = O\left( \frac{1}{t^{\alpha /2}}\right) \qquad (t \rightarrow \infty ) \end{aligned}$$

for all \(\alpha >0\), we obtain the desired estimate (2.3). \(\square \)

In [5, Corollary 5.7], the \({\mathcal {B}}\)-calculus and the norm-estimate (2.2) of \(h_t(z) = e^{-t/(z+1)}\) have been used to obtain the norm-estimate for \(e^{A^{-1}t}\). Analogously, the following theorem gives the norm-estimate for \(e^{A^{-1}t}(-A)^{-\alpha }\) with \(\alpha > 0\).

Theorem 2.4

Let A be the generator of an exponentially stable \(C_0\)-semigroup \((e^{At})_{t \ge 0}\) on a Hilbert space H. Then

$$\begin{aligned} \big \Vert e^{A^{-1}t} (-A)^{-\alpha }\big \Vert = O\left( \frac{1}{t^{\alpha /2}} \right) \qquad (t \rightarrow \infty ) \end{aligned}$$
(2.6)

for all \(\alpha > 0\).

Proof

There exist constants \(K,\omega >0\) such that

$$\begin{aligned} \big \Vert e^{tA}\big \Vert \le Ke^{-\omega t} \end{aligned}$$

for all \(t \ge 0\). We may assume that \(\omega = 1\) by replacing A by \(\omega ^{-1}A\) and t by \(\omega t\). Note that \(B:= -A-I\) generates a bounded \(C_0\)-semigroup \((e^{-tB})_{t \ge 0}\) with \(\sup _{t \ge 0} \Vert e^{-tB}\Vert \le K\).

Take \(t,\alpha > 0\) arbitrarily, and consider the functions

$$\begin{aligned} h_t(z):= e^{-t/(z+1)},\quad r_\alpha (z):= \frac{1}{(z+1)^{\alpha }},\quad z \in {\mathbb {C}}_+. \end{aligned}$$

Then \(h_t\) belongs to \({{\mathcal {L}}}{{\mathcal {M}}} \subset {\mathcal {B}}\); see [5, Example 2.12]. From the definition of the gamma function, we have that for all \(z >0\),

$$\begin{aligned} \Gamma (\alpha ) = \int ^{\infty }_0 t^{\alpha -1} e^{-t}dt = (z+1)^{\alpha } \int ^{\infty }_0 t^{\alpha -1} e^{-(z+1)t}dt. \end{aligned}$$

Hence, the uniqueness theorem for holomorphic functions yields

$$\begin{aligned} r_{\alpha } (z) = \frac{1}{\Gamma (\alpha )} \int ^{\infty }_0 t^{\alpha -1}e^{-(z+1)t} dt \end{aligned}$$

for all \(z\in {\mathbb {C}}_+\). This means that \(r_{\alpha }\) is the Laplace transform of the function

$$\begin{aligned} t\mapsto \frac{t^{\alpha -1} e^{-t}}{\Gamma (\alpha )},\quad t > 0. \end{aligned}$$

Hence \(r_{\alpha }\) also belongs to \({{\mathcal {L}}}{{\mathcal {M}}} \subset {\mathcal {B}}\).

Theorem 2.2 (i) shows that the function \(f_{t,\alpha } =h_t r_{\alpha }\) satisfies

$$\begin{aligned} f_{t,\alpha }(B) = h_t(B)r_{\alpha }(B). \end{aligned}$$

We obtain

$$\begin{aligned} h_t(B) = e^{A^{-1}t} \end{aligned}$$

as shown in the proof of [5, Corollary 5.7]. By Theorem 2.2 (ii), \(r_{\alpha }(B)\) coincides with the operator defined as the Hille-Phillips calculus. This and [18, Proposition 3.3.5] give

$$\begin{aligned} r_{\alpha }(B) = (I+B)^{-\alpha } = (-A)^{-\alpha }. \end{aligned}$$

Therefore,

$$\begin{aligned} f_{t,\alpha }(B) = e^{A^{-1}t}(-A)^{-\alpha }. \end{aligned}$$

Using Theorem 2.2 (iii), we have

$$\begin{aligned} \big \Vert e^{A^{-1}t}(-A)^{-\alpha }\big \Vert \le |f_{t,\alpha }(\infty )| + 2K^2 \Vert f_{t,\alpha }\Vert _{{\mathcal {B}}_0} = 2K^2 \Vert f_{t,\alpha }\Vert _{{\mathcal {B}}_0}. \end{aligned}$$

Thus, the desired estimate (2.6) holds by Lemma 2.3. \(\square \)

From the next example, we see that the norm-estimate (2.6) cannot be improved in general.

Example 2.5

Let \(\gamma >0\) and set \(\lambda _k:= -\gamma + ik\) for \(k \in {\mathbb {N}}\). Define an operator A on \(\ell ^2\) by

$$\begin{aligned} (A\zeta _k)_{k \in {\mathbb {N}}}:= (\lambda _k \zeta _k)_{n \in {\mathbb {N}}} \end{aligned}$$

with domain

$$\begin{aligned} D(A):= \{ \zeta = (\zeta _k)_{k \in {\mathbb {N}}} \in \ell ^2: (\lambda _k \zeta _k)_{k \in {\mathbb {N}}} \in \ell ^2 \}. \end{aligned}$$

Then

$$\begin{aligned} \big \Vert e^{A^{-1}t} (-A)^{-\alpha }\big \Vert = \sup _{k \in {\mathbb {N}}} \frac{e^{t \mathop {\textrm{Re}}\nolimits \lambda _k/ |\lambda _k|^2}}{|\lambda _k|^\alpha } = \sup _{k \in {\mathbb {N}}} \frac{e^{-\gamma t/(\gamma ^2+k^2)}}{(\gamma ^2+k^2)^{\alpha /2}} \end{aligned}$$

for all \(t,\alpha >0\). Put \(f(w):= w^{-\alpha /2}e^{-\gamma t/w}\) for \(w \ge \gamma ^2+1\). Then

$$\begin{aligned} \sup _{w \ge \gamma ^2+1} f(w) = f\left( \frac{2\gamma t}{\alpha } \right) = \left( \frac{\alpha }{2e\gamma t} \right) ^{\alpha /2} \end{aligned}$$

for all \(t \ge \alpha (\gamma ^2+1)/(2\gamma )\). From this, it follows that for all \(k \in {\mathbb {N}}\) and \(t = \alpha (\gamma ^2+k^2)/(2\gamma )\),

$$\begin{aligned} \big \Vert e^{A^{-1}t} (-A)^{-\alpha }\big \Vert = \left( \frac{\alpha }{2e\gamma t} \right) ^{\alpha /2}. \end{aligned}$$

Hence

$$\begin{aligned} \liminf _{t \rightarrow \infty } t^{\alpha /2} \big \Vert e^{A^{-1}t} (-A)^{-\alpha }\big \Vert \ge \left( \frac{\alpha }{2e\gamma } \right) ^{\alpha /2}, \end{aligned}$$

and the estimate (2.6) cannot be improved for this diagonal operator. Moreover, we see that the rate of polynomial decay of \(\Vert e^{A^{-1}t} (-A)^{-\alpha }\Vert \) does not depend on the exponential growth bound \(-\gamma \) of the \(C_0\)-semigroup \((e^{tA})_{t \ge 0}\).

2.2 Inverses of Generators of Polynomially Stable Semigroups

We recall the definition of polynomially stable \(C_0\)-semigroups.

Definition 2.6

A \(C_0\)-semigroup \((e^{At})_{t\ge 0}\) on a Hilbert space is polynomially stable with parameter \(\beta >0\) if \((e^{At})_{t\ge 0}\) is bounded and if

$$\begin{aligned} \big \Vert e^{At}(I-A)^{-1}\big \Vert = O\left( \frac{1}{t^{1/\beta }} \right) \qquad (t \rightarrow \infty ). \end{aligned}$$
(2.7)

A \(C_0\)-semigroup \((e^{At})_{t\ge 0}\) is simply called polynomially stable if it is polynomially stable with some parameter \(\beta >0\).

Let A be the generator of a polynomially stable semigroup on a Hilbert space H. By [4, Theorem 1.1], \(i{\mathbb {R}} \cap \sigma (A) = \emptyset \). Then A is invertible, and therefore \(A^{-1}\) generates a \(C_0\)-semigroup on H. When A is normal, we obtain the rate of decay of \(\Vert e^{A^{-1}t}(-A)^{-\alpha }\Vert \) as in the case of exponentially stable \(C_0\)-semigroups. We also show that \(\sup _{t \ge 0}\Vert e^{A^{-1}t}A^{-1}\Vert < \infty \) without assuming that A is normal.

2.2.1 Case Where Generators are Normal

When the generator A is a normal operator on a Hilbert space, a spectral condition equivalent to polynomial decay is known. The proof can be found in [3, Proposition 4.1].

Proposition 2.7

Let H be a Hilbert space and let \(A:D(A)\subset H \rightarrow H\) be a normal operator. Assume that \(\sigma (A) \subset {\mathbb {C}}_-\). For a fixed \(\beta >0\), the \(C_0\)-semigroup \((e^{At})_{t\ge 0}\) satisfies

$$\begin{aligned} \big \Vert e^{At}A^{-1}\big \Vert = O\left( \frac{1}{t^{1/\beta }} \right) \qquad (t \rightarrow \infty ) \end{aligned}$$

if and only if there exist \(C,\delta >0\) such that

$$\begin{aligned} |\mathop {\textrm{Im}}\nolimits \lambda | \ge \frac{C}{|\mathop {\textrm{Re}}\nolimits \lambda |^{1/\beta }} \end{aligned}$$

for all \(\lambda \in \sigma (A)\) with \(\mathop {\textrm{Re}}\nolimits \lambda \ge -\delta \).

The next result gives an estimate for the rate of decay of \(\Vert e^{A^{-1}t} (-A)^{-\alpha }\Vert \).

Proposition 2.8

Let H be a Hilbert space, and let \(A:D(A)\subset H \rightarrow H\) be a normal operator generating a polynomially stable \(C_0\)-semigroup \((e^{At})_{t \ge 0}\) with parameter \(\beta >0\) on H. Then

$$\begin{aligned} \big \Vert e^{A^{-1}t} (-A)^{-\alpha }\big \Vert = O\left( \frac{1}{t^{\alpha /(2+\beta )}}\right) \qquad (t \rightarrow \infty ) \end{aligned}$$
(2.8)

for all \(\alpha > 0\).

Proof

For all nonzero \(\lambda \in {\mathbb {C}}\), we obtain

$$\begin{aligned} \lambda I - A^{-1} = \left( A - \frac{1}{\lambda } I \right) \lambda A^{-1}, \end{aligned}$$

which implies that \(1/\lambda \in \sigma (A)\) if and only if \(\lambda \in \sigma (A^{-1})\). The normality of A implies that of \(A^{-1}\) by [27, Theorem 5.42]. Fix \(t,\alpha >0\), and define

$$\begin{aligned} f(\lambda ):= {\left\{ \begin{array}{ll} e^{-\lambda t}\lambda ^\alpha , &{} \lambda \in {\mathbb {C}} {\setminus } (-\infty ,0], \\ 0, &{} \lambda = 0. \end{array}\right. } \end{aligned}$$

Then \(f(-A^{-1}) = e^{A^{-1}t} (-A)^{-\alpha }\). Moreover, \(f(-A^{-1})\) is normal and

$$\begin{aligned} f\big (\sigma (-A^{-1})\big ) = \sigma \big (f(-A^{-1})\big ); \end{aligned}$$

see, e.g., [19, Theorem 4.5]. Therefore,

$$\begin{aligned} \big \Vert e^{A^{-1}t}(-A)^{-\alpha }\big \Vert = \sup _{\lambda \in \sigma (-A^{-1}) } |f(\lambda )| = \sup _{\lambda \in \sigma (A) } \frac{e^{t \mathop {\textrm{Re}}\nolimits \lambda /|\lambda |^2}}{|\lambda |^{\alpha }}. \end{aligned}$$

By Proposition 2.7, there exist \(C,\delta >0\) such that \(|\mathop {\textrm{Im}}\nolimits \lambda | \ge C|\mathop {\textrm{Re}}\nolimits \lambda |^{-1/\beta }\) for all \(\lambda \in \sigma (A)\) with \(|\mathop {\textrm{Re}}\nolimits \lambda | \le \delta \). If \(\lambda \in \sigma (A)\) satisfies \(|\mathop {\textrm{Re}}\nolimits \lambda | \le \delta \), then

$$\begin{aligned} |\mathop {\textrm{Re}}\nolimits \lambda | \ge \frac{C^{\beta }}{|\lambda |^{\beta }}, \end{aligned}$$

and therefore

$$\begin{aligned} \frac{e^{t \mathop {\textrm{Re}}\nolimits \lambda /|\lambda |^2}}{ |\lambda |^{\alpha } } \le \sup _{s >0} \frac{e^{-C^{\beta } t/s^{2+\beta }}}{ s^{\alpha }} = \left( \frac{\alpha }{(2+\beta )eC^{\beta }t} \right) ^{\frac{\alpha }{2+\beta }}. \end{aligned}$$

On the other hand, for all \(\lambda \in \sigma (A)\) with \(|\mathop {\textrm{Re}}\nolimits \lambda | > \delta \), we obtain

$$\begin{aligned} \frac{e^{t \mathop {\textrm{Re}}\nolimits \lambda /|\lambda |^2}}{ |\lambda |^{\alpha } } \le \sup _{s >0} \frac{e^{-\delta t /s^2} }{ s^{\alpha } } = \left( \frac{\alpha }{2e \delta t} \right) ^{\frac{\alpha }{2}}. \end{aligned}$$

Thus, the norm-estimate (2.8) holds. \(\square \)

As in the case of exponentially stable \(C_0\)-semigroups, we present an example of an operator on \(\ell ^2\) showing that the norm-estimate (2.8) cannot be in general improved.

Example 2.9

Set \(\lambda _k:= -1/k+ ik\) for \(k \in {\mathbb {N}}\). Define an operator A on \(\ell ^2\) by

$$\begin{aligned} (A\zeta _k)_{k \in {\mathbb {N}}}:= (\lambda _k \zeta _k)_{k \in {\mathbb {N}}}\end{aligned}$$

with domain

$$\begin{aligned} D(A):= \{ \zeta = (\zeta _k)_{k\in {\mathbb {N}}} \in \ell ^2: (\lambda _k \zeta _k)_{k \in {\mathbb {N}}} \in \ell ^2 \}. \end{aligned}$$

By Proposition 2.7, A is the generator of a polynomially stable \(C_0\)-semigroup with parameter \(\beta = 1\). Let \(t,\alpha >0\). We have

$$\begin{aligned} \big \Vert e^{A^{-1}t} (-A)^{-\alpha } \big \Vert = \sup _{k \in {\mathbb {N}}} \frac{e^{t \mathop {\textrm{Re}}\nolimits \lambda _k/ |\lambda _k|^2}}{|\lambda _k|^\alpha } = \sup _{k \in {\mathbb {N}}} \frac{e^{-kt/(k^4+1)}}{(1/k^2+k^2)^{\alpha /2}}. \end{aligned}$$

Moreover,

$$\begin{aligned} \frac{e^{-kt/(k^4+1)}}{(1/k^2+k^2)^{\alpha /2}} \ge \frac{e^{-t/k^3}}{2^{\alpha /2} k^\alpha } \end{aligned}$$

for all \(k \in {\mathbb {N}}\). Put \(f(w):= w^{-\alpha }e^{-t/w^3}\) for \(w \ge 1\). Then

$$\begin{aligned} \sup _{w \ge 1} f(w) = f\left( \left( \frac{3t}{\alpha }\right) ^{1/3} \right) = \left( \frac{\alpha }{3et}\right) ^{\alpha /3} \end{aligned}$$

for all \(t \ge \alpha /3\). This implies that for all \(k \in {\mathbb {N}}\) and \(t = \alpha k^3 /3\),

$$\begin{aligned} \big \Vert e^{A^{-1}t} (-A)^{-\alpha }\big \Vert \ge \frac{1}{2^{\alpha /2}} \left( \frac{\alpha }{3et}\right) ^{\alpha /3}. \end{aligned}$$

Therefore

$$\begin{aligned} \liminf _{t \rightarrow \infty } t^{\alpha /3} \big \Vert e^{A^{-1}t} (-A)^{-\alpha }\big \Vert \ge \frac{1}{2^{\alpha /2}} \left( \frac{\alpha }{3e}\right) ^{\alpha /3}, \end{aligned}$$

and the estimate (2.8) cannot be improved for this operator A.

2.2.2 Norm-Estimate Through Lyapunov Equations

We show that if A is the generator of a polynomially stable \(C_0\)-semigroup on a Hilbert space, then \(\sup _{t \ge 0}\Vert e^{A^{-1}t} A^{-1}\Vert < \infty \). To this end, we use the following results on Lyapunov equations; see [8, Theorem 4.1.3, Corollary 6.5.1, and Theorem 6.5.2] for the proofs.

Lemma 2.10

Let A be the generator of an exponentially stable semigroup \((e^{At})_{t \ge 0}\) on a Hilbert space H. Then there exists a unique self-adjoint, positive operator \(P \in {\mathcal {L}}(H)\) such that \(PD(A) \subset D(A^*)\) and

$$\begin{aligned} A^*P + P A = -I\quad \hbox { on}\ D(A). \end{aligned}$$
(2.9)

Moreover, the operator P is given by

$$\begin{aligned} Px = \int ^{\infty }_0 \big (e^{At} \big )^* e^{At}x dt \end{aligned}$$
(2.10)

for all \(x \in H\).

To obtain \(\sup _{t \ge 0}\Vert e^{A^{-1}t} A^{-1}\Vert < \infty \), we shall use the fact that \(P \in {\mathcal {L}}(H)\) defined by (2.10) satisfies the Lyapunov equation (2.9).

Lemma 2.11

Let A be the generator of a \(C_0\)-semigroup on a Hilbert space H, and let C be a bounded linear operator from H to another Hilbert space Y. If there exists a self-adjoint, non-negative operator \(P \in {\mathcal {L}}(H)\) such that \(PD(A) \subset D(A^*)\) and

$$\begin{aligned} A^*P + PA = -C^*C\quad \hbox { on}\ D(A), \end{aligned}$$

then

$$\begin{aligned} \int ^{\infty }_0 \big \Vert Ce^{At}x\big \Vert ^2 dt \le \langle x, Px \rangle \end{aligned}$$

for all \(x \in H\).

We shall need the estimate for adjoint operators in the following lemma; see [28, Lemma 2.1] for the proof.

Lemma 2.12

Let A be the generator of a bounded \(C_0\)-semigroup \((e^{At})_{t \ge 0}\) on a Hilbert space H, and let \(K:= \sup _{t \ge 0} \Vert e^{At}\Vert \). Then for all \(x \in H\) and \(\xi ,\gamma >0\),

$$\begin{aligned} \int ^{\infty }_0 \big \Vert e^{(\gamma A-\xi I)^{-1}t}(\gamma A - \xi I)^{-1} x \big \Vert ^2 dt \le \frac{K^2 \Vert x\Vert ^2}{2\xi }. \end{aligned}$$
(2.11)

The same estimate holds for the adjoint.

We are now in a position to prove that \(\sup _{t \ge 0} \Vert e^{A^{-1}t} A^{-1}\Vert < \infty \) if A is the generator of a polynomially stable \(C_0\)-semigroup. Actually, we will make the slightly weaker assumption that \(\Vert e^{At} (I-A)^{-1}\Vert = O((\log t)^{-\beta })\) as \(t\rightarrow \infty \) for some \(\beta >1\). Notice that if the bounded \(C_0\)-semigroup \((e^{At})_{t \ge 0}\) satisfies this estimate, then \(i{\mathbb {R}} \cap \sigma (A) = \emptyset \) by [4, Theorem 1.1]. The proof of the following theorem is inspired by the arguments in the proof of [28, Theorem 2.2].

Theorem 2.13

Let A be the generator of a bounded \(C_0\)-semigroup on a Hilbert space H. If there exists \(\beta >1\) such that

$$\begin{aligned} \big \Vert e^{At} (I-A)^{-1}\big \Vert = O\left( \frac{1}{(\log t)^{\beta }} \right) \qquad (t \rightarrow \infty ), \end{aligned}$$
(2.12)

then

$$\begin{aligned} \sup _{t \ge 0} \big \Vert e^{A^{-1}t} A^{-1}\big \Vert < \infty . \end{aligned}$$
(2.13)

In particular, if A is the generator of a polynomially stable \(C_0\)-semigroup on H, then the estimate (2.13) holds.

Proof

The proof consists of three steps. In the first step, we prove an estimate analogous to (2.11) for bounded \(C_0\)-semigroups satisfying (2.12). In the next step, we obtain the norm-estimate of

$$\begin{aligned} e^{(A-\xi _1I)^{-1}t}x - e^{(A-\xi _2I)^{-1}t} x \end{aligned}$$

for \(x \in D(A)\) and \(0< \xi _1,\xi _2 < \delta _0\), where \(\delta _0 >0\) is some sufficiently small constant. This estimate is used to compare \(e^{(A-\delta I)^{-1}t}x\) and \(e^{(A-\delta e^{-N}I)^{-1}t}x\) for \(x \in D(A)\), \(N \in {\mathbb {N}}\), and \(0< \delta < \delta _0\). Letting \(N \rightarrow \infty \), we show the desired estimate (2.13) in the last step.

Step 1: Let A be the generator of a bounded \(C_0\)-semigroup \((e^{At})_{t\ge 0}\) such that (2.12) holds for some \(\beta > 1\). We write \(K:= \sup _{t \ge 0} \Vert e^{At}\Vert \). By (2.12), there exist \(M_0 >0\) and \(t_0 > 1\) such that

$$\begin{aligned} \big \Vert e^{A t}A^{-1}\big \Vert \le \frac{M_0}{(\log t)^{\beta }} \end{aligned}$$
(2.14)

for all \(t \ge t_0\). Choose \(t_1 > \max \{t_0,e^{2\beta }\}\). There exists \(M_1 >0\) such that

$$\begin{aligned} \int _{t_1}^{\infty } \frac{e^{-\xi t}}{(\log t)^{2\beta }} dt \le \frac{M_1}{\xi |\log \xi |^{2\beta }} \end{aligned}$$
(2.15)

for all \(0< \xi < 1/t_1\); see [26, Lemma 4.2]. For \(\xi >0\), define the operator \(P(\xi ) \in {\mathcal {L}}(H)\) by

$$\begin{aligned} P(\xi ) x:= \int ^{\infty }_0 e^{-2\xi t}\big (e^{At} \big )^* e^{At}x dt,\quad x \in H. \end{aligned}$$

From the estimates (2.14) and (2.15), we obtain

$$\begin{aligned} \langle x, P(\xi ) x\rangle&= \int ^{t_1}_0 e^{-2\xi t} \big \Vert e^{At}x\big \Vert ^2 dt + \int ^{\infty }_{t_1} e^{-2\xi t} \big \Vert e^{A t}x\big \Vert ^2 dt \\&\le t_1 K^2 \Vert x\Vert ^2 + \frac{M_0^2 M_1 \Vert Ax\Vert ^2 }{2 \xi |\log (2\xi )|^{2 \beta }} \end{aligned}$$

for all \(x \in D(A)\) and \(0< \xi < 1/(2t_1) =: \delta _0\). Hence there exists \(K_1 >0\) such that

$$\begin{aligned} \sup _{0< \xi < \delta _0} \xi |\log \xi |^{2\beta } \langle x,P(\xi )x \rangle \le K_1 \Vert Ax\Vert ^2 \end{aligned}$$
(2.16)

for all \(x \in D(A)\).

By Lemma 2.10, \(P(\xi )D(A) \subset D(A^*)\) and

$$\begin{aligned} (A- \xi I)^*P(\xi ) + P(\xi )(A - \xi I) = -I\quad \hbox { on}\ D(A) \end{aligned}$$
(2.17)

for all \(\xi >0\). Since (2.17) yields

$$\begin{aligned} (A^*- \xi I)^{-1}P(\xi ) + P(\xi )(A - \xi I)^{-1} = -(A^*- \xi I)^{-1}(A - \xi I)^{-1}, \end{aligned}$$

we have from Lemma 2.11 and the estimate (2.16) that

$$\begin{aligned} \int ^{\infty }_0 \big \Vert e^{(A-\xi I)^{-1}t} (A- \xi I)^{-1}x \big \Vert ^2 dt \le \langle x,P(\xi )x \rangle \le \frac{K_1 \Vert Ax\Vert ^2}{\xi |\log \xi |^{2\beta }} \end{aligned}$$
(2.18)

for all \(x \in D(A)\) and \(0< \xi < \delta _0\). Moreover, by Lemma 2.12, there exists \(K_2 >0\) such that

$$\begin{aligned} \int ^{\infty }_0 \big \Vert e^{(A^*-\xi I)^{-1}t} (A^*- \xi I)^{-1}y \big \Vert ^2 dt \le \frac{K_2 \Vert y\Vert ^2}{\xi } \end{aligned}$$
(2.19)

for all \(y \in H\) and \(\xi >0\).

Step 2: Take \(0< \xi _1,\xi _2 < \delta _0\). The variation of constants formula yields

$$\begin{aligned}&e^{(A-\xi _1 I)^{-1}t} x - e^{(A-\xi _2 I)^{-1}t}x \\&\quad = \int ^t_0 e^{(A-\xi _1 I)^{-1}(t-s)} \big ( (A-\xi _1 I)^{-1} - (A-\xi _2 I)^{-1} \big ) e^{(A-\xi _2 I)^{-1}s} xds \end{aligned}$$

for all \(x \in H\) and \(t \ge 0\). Moreover, the resolvent equation gives

$$\begin{aligned} (A-\xi _1 I)^{-1} - (A-\xi _2 I)^{-1} = (\xi _1 - \xi _2) (A-\xi _1 I)^{-1} (A-\xi _2 I)^{-1}. \end{aligned}$$

Therefore, we obtain

$$\begin{aligned}&e^{(A-\xi _1 I)^{-1}t} x - e^{(A-\xi _2 I)^{-1}t}x \\&\quad =(\xi _1 - \xi _2) \int ^t_0 e^{(A-\xi _1 I)^{-1}(t-s)} (A-\xi _1 I)^{-1} (A-\xi _2 I)^{-1}e^{(A-\xi _2 I)^{-1}s}xds \end{aligned}$$

for all \(x \in H\) and \(t \ge 0\). Using the estimates (2.18) and (2.19), we have that for all \(x \in D(A)\), \(y \in H\), and \(t \ge 0\),

$$\begin{aligned}&\left| \left\langle y,~ e^{(A-\xi _1 I)^{-1}t} x - e^{(A-\xi _2 I)^{-1}t} x \right\rangle \right| \\&\quad = |\xi _1 - \xi _2|\, \bigg | \int ^t_0 \Big \langle e^{(A^*-\xi _1 I)^{-1}(t-s)} (A^*- \xi _1 I)^{-1}y,\\&\hspace{100pt} e^{(A-\xi _2 I)^{-1}s} (A- \xi _2 I)^{-1}x \Big \rangle ds \bigg | \\&\quad \le |\xi _1 - \xi _2| \left( \int ^{\infty }_0 \big \Vert e^{(A^*-\xi _1 I)^{-1}t} (A^*- \xi _1 I)^{-1}y \big \Vert ^2 dt \right) ^{1/2} \\&\hspace{100pt} \times \left( \int ^{\infty }_0 \big \Vert e^{(A-\xi _2 I)^{-1}t} (A- \xi _2 I)^{-1}x \big \Vert ^2 dt \right) ^{1/2} \\&\quad \le \frac{ K_0|\xi _1 - \xi _2| }{ \sqrt{\xi _1\xi _2} \, |\log \xi _2|^{\beta }} \Vert Ax\Vert \, \Vert y\Vert , \end{aligned}$$

where \(K_0:= \sqrt{K_1K_2}\). Hence

$$\begin{aligned} \big \Vert e^{(A-\xi _1 I)^{-1}t} x - e^{(A-\xi _2 I)^{-1}t} x \big \Vert \le \frac{K_0|\xi _1 - \xi _2| }{ \sqrt{\xi _1\xi _2} \, |\log \xi _2|^{\beta } }\Vert Ax\Vert \end{aligned}$$
(2.20)

for all \(x \in D(A)\) and \(t \ge 0\).

Step 3: Let \(N \in {\mathbb {N}}\) and \(0<\delta< \delta _0\, (< 1)\). Substituting \(\xi _1 = \delta e^{-n+1}\) and \(\xi _2 = \delta e^{-n}\), \(n=1,2,\dots ,N\), into the estimate (2.20), we obtain

$$\begin{aligned} \big \Vert e^{(A- \delta I)^{-1}t} x - e^{(A- \delta e^{-N} I)^{-1}t} x \big \Vert&= \left\| \sum _{n=1}^N e^{(A- \delta e^{-n+1} I)^{-1}t} x - e^{(A-\delta e^{-n} I)^{-1}t} x \right\| \\&\le K_0 \Vert Ax\Vert \sum _{n=1}^N \frac{|e^{-n+1} - e^{-n}| }{\sqrt{e^{-n+1} e^{-n}}\, |\log (\delta e^{-n})|^{\beta }} \\&\le \frac{(e-1)K_0}{\sqrt{e} } \Vert Ax\Vert \sum _{n=1}^{N} \frac{1}{|n - \log \delta |^{\beta }} \end{aligned}$$

for all \(x \in D(A)\) and \(t \ge 0\). From the assumption \(\beta > 1\), we obtain

$$\begin{aligned} K_3:= \sum _{n=1}^{\infty } \frac{1}{|n - \log \delta |^{\beta }} < \infty . \end{aligned}$$

Since

$$\begin{aligned} \lim _{N \rightarrow \infty } \big \Vert e^{(A-\delta e^{-N} I)^{-1}t} - e^{A^{-1}t} \big \Vert = 0 \end{aligned}$$

for each \(t \ge 0\), it follows that

$$\begin{aligned} \big \Vert e^{(A- \delta I)^{-1}t} x - e^{A^{-1}t} x \big \Vert \le \frac{(e-1)K_0K_3}{\sqrt{e} } \Vert Ax\Vert \end{aligned}$$

for all \(x \in D(A)\) and \(t \ge 0\). This gives

$$\begin{aligned} \sup _{t\ge 0} \big \Vert e^{(A-\delta I)^{-1}t} A^{-1} - e^{A^{-1}t} A^{-1} \big \Vert < \infty . \end{aligned}$$

Since \(A- \delta I\) generates an exponentially stable \(C_0\)-semigroup, Theorem 2.4 implies that

$$\begin{aligned} \sup _{t \ge 0 }\big \Vert e^{(A- \delta I)^{-1}t} A^{-1}\big \Vert < \infty . \end{aligned}$$

Thus,

$$\begin{aligned}&\sup _{t \ge 0} \big \Vert e^{A^{-1}t} A^{-1} \big \Vert \\&\quad \le \sup _{t \ge 0} \big \Vert e^{(A-\delta I)^{-1}t} A^{-1} \big \Vert + \sup _{t \ge 0}\big \Vert e^{(A- \delta I)^{-1}t} A^{-1} - e^{A^{-1}t} A^{-1} \big \Vert < \infty . \end{aligned}$$

\(\square \)

3 Crank–Nicolson Scheme with Smooth Initial Data

Let A be the generator of a bounded \(C_0\)-semigroup on a Hilbert space H. For \(\tau > 0\), we define

$$\begin{aligned} A_d(\tau ):= \left( I + \frac{\tau }{2}A \right) \left( I - \frac{\tau }{2}A \right) ^{-1}. \end{aligned}$$
(3.1)

Let \((\tau _n)_{n \in {\mathbb {N}}_0}\) be a sequence of strictly positive real numbers. We consider the time-varying difference equation

$$\begin{aligned} x_{n+1} = A_d(\tau _n) x_n,\quad n \in {\mathbb {N}}_0;\qquad x_0 \in H. \end{aligned}$$
(3.2)

In this section, we study the decay rate of the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the difference equation (3.2) with smooth initial data.

3.1 Generators of Exponentially Stable Semigroups

Let H be a Hilbert space, and let \(A:D(A) \subset H \rightarrow H\) be injective. Suppose that A and \(A^{-1}\) generate bounded \(C_0\)-semigroups on H. Take \(\xi >0\). By Lemma 2.10, there exist unique self-adjoint, positive operators \(P(\xi ), Q(\xi ) \in {\mathcal {L}}(H)\) such that \(P(\xi )D(A) \subset D(A^*)\), \(Q(\xi )D(A^{-1}) \subset D((A^{-1})^*)\), and

$$\begin{aligned} (A- \xi I)^*P(\xi ) + P(\xi )(A - \xi I)= & {} -I\quad \hbox { on}\ D(A), \end{aligned}$$
(3.3a)
$$\begin{aligned} (A^{-1} - \xi I)^*Q(\xi ) + Q(\xi )(A^{-1} - \xi I)= & {} -I\quad \hbox { on}\ D(A^{-1}). \end{aligned}$$
(3.3b)

The operators \(P(\xi )\) and \(Q(\xi )\) are given by

$$\begin{aligned} P(\xi ) x= & {} \int ^{\infty }_0 e^{-2\xi t} (e^{At})^* e^{At}x dt, \end{aligned}$$
(3.4a)
$$\begin{aligned} Q(\xi ) x= & {} \int ^{\infty }_0 e^{-2\xi t} (e^{A^{-1}t})^* e^{A^{-1}t}x dt \end{aligned}$$
(3.4b)

for all \(x \in X\). To estimate the decay rate of the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the difference equation (3.2), we shall use that the operators \(P(\xi )\) and \(Q(\xi )\) defined as in (3.4) solve the Lypapunov equations (3.3).

In the proofs of [23, Lemmas 2.1 and 2.2], the following result has been obtained from the Lypapunov equations (3.3).

Lemma 3.1

Let A be the generator of a bounded \(C_0\)-semigroup on a Hilbert space H. Suppose that A is injective and that \(A^{-1}\) also generate a bounded \(C_0\)-semigroup on H. Let \(0< \tau _{\min } \le \tau _n \le \tau _{\max } < \infty \) for all \(n \in {\mathbb {N}}_0\), and let \(P(\xi ), Q(\xi ) \in {\mathcal {L}}(H)\) be given by (3.4) for \(\xi >0\). Define \(R(r) \in {\mathcal {L}}(H)\) by

$$\begin{aligned} R(r):= \frac{2}{\tau _{\min }}P\left( \frac{\xi _r}{\tau _{\max }} \right) + 2\tau _{\max } Q(\tau _{\min } \xi _r),\quad r \in (0,1), \end{aligned}$$

where

$$\begin{aligned} \xi _r:= \frac{1-r^2}{2(r^2+1)}. \end{aligned}$$

Then there exists a constant \(M >0\) such that the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the difference equation (3.2) satisfies

$$\begin{aligned} |(n+1)r^n \langle y, x_n \rangle | \le \frac{M \Vert y\Vert }{\sqrt{1-r}} \sqrt{\langle x_0, R(r)x_0 \rangle } \end{aligned}$$

for all \(x_0, y \in X\), \(n \in {\mathbb {N}}\), and \(r \in (0,1)\).

Now we estimate \(\langle x,P(\xi ) x \rangle \) and \(\langle x,Q(\xi ) x \rangle \), by using the integral representations (3.4). Suppose that A is the generator of an exponentially stable \(C_0\)-semigroup \((e^{At})_{t \ge 0}\). Let \(K_0 \ge 1\) and \(\omega >0\) satisfy \(\Vert e^{At}\Vert \le K_0e^{-\omega t}\) for all \(t \ge 0\). Then the operator \(P(\xi )\) given by (3.4a) satisfies

$$\begin{aligned} \langle x, P(\xi )x \rangle \le K_0^2\Vert x\Vert ^2 \int ^{\infty }_0 e^{-2(\omega + \xi )t}dt \le \frac{K_0^2 \Vert x\Vert ^2}{2\omega } \end{aligned}$$

for all \(x \in X\) and \(\xi >0\). Hence

$$\begin{aligned} \sup _{\xi >0} \langle x, P(\xi )x \rangle \le K_1 \Vert x\Vert ^2 \end{aligned}$$
(3.5)

for all \(x \in X\), where \(K_1:= K_0^2/(2\omega )\). Using the norm-estimate (2.6) for \(e^{A^{-1}t}(-A)^{-\alpha }\) with \(0 < \alpha \le 1\), we estimate \(\langle x, Q(\xi )x \rangle \) for \(x \in D((-A)^{\alpha })\) in the next lemma.

Lemma 3.2

Let A be the generator of an exponentially stable \(C_0\)-semigroup \((e^{At})_{t \ge 0}\) on a Hilbert space H. Suppose that \(A^{-1}\) generates a bounded \(C_0\)-semigroup \((e^{A^{-1}t})_{t \ge 0}\) on H. Then \(Q(\xi ) \in {\mathcal {L}}(H)\) given by (3.4b) for \(\xi >0\) satisfies the following statements:

  1. (i)

    For each \(0< \alpha < 1\), there exists \(K>0\) such that

    $$\begin{aligned} \sup _{0< \xi < 1} \xi ^{1-\alpha } \langle x, Q(\xi )x \rangle \le K \Vert (-A)^{\alpha }x\Vert ^2 \end{aligned}$$
    (3.6)

    for all \(x \in D((-A)^{\alpha })\)

  2. (ii)

    There exists \(K>0\) such that

    $$\begin{aligned} \sup _{0< \xi < 1/2} \frac{ \langle x, Q(\xi )x \rangle }{\log (1/\xi )} \le K \Vert Ax\Vert ^2 \end{aligned}$$
    (3.7)

    for all \(x \in D(A)\).

Proof

Let \(0 < \alpha \le 1\) and \( M_1:= \sup _{t \ge 0}\Vert e^{A^{-1} t}\Vert < \infty \). By Theorem 2.4, there exist \(M_2, t_0 >0\) such that

$$\begin{aligned} \big \Vert e^{A^{-1}t}x\big \Vert \le \frac{M_2 \Vert (-A)^{\alpha }x\Vert }{t^{\alpha /2}} \end{aligned}$$

for all \(x \in D((-A)^{\alpha })\) and \(t \ge t_0\). Hence

$$\begin{aligned} \langle x, Q(\xi ) x\rangle&= \int ^{t_0}_0 e^{-2\xi t} \big \Vert e^{A^{-1}t}x\big \Vert ^2 dt + \int ^{\infty }_{t_0} e^{-2\xi t}\big \Vert e^{A^{-1}t}x\big \Vert ^2 dt \\&\le t_0 M_1^2 \Vert x\Vert ^2 + M_2^2 \Vert (-A)^{\alpha }x\Vert ^2 \int ^{\infty }_{t_0} \frac{e^{-2\xi t} }{t^{\alpha }} dt \end{aligned}$$

for all \(x \in D((-A)^{\alpha })\) and \(\xi >0\). When \(0< \alpha < 1\), we have

$$\begin{aligned} \int ^{\infty }_{t_0} \frac{e^{-2\xi t} }{t^{\alpha }} dt \le \int ^{\infty }_0 \frac{e^{-2\xi t} }{t^{\alpha }} dt = \frac{\Gamma (1-\alpha )}{(2\xi )^{1-\alpha }}. \end{aligned}$$

Hence there is \(K>0\) such that (3.6) holds for all \(x \in D((-A)^{\alpha })\).

Next we consider the case \(\alpha = 1\). The exponential integral satisfies

$$\begin{aligned} \int _\tau ^{\infty } \frac{e^{-t}}{t} dt \le e^{-\tau }\log \left( 1+ \frac{1}{\tau } \right) \end{aligned}$$

for all \(\tau >0\); see [12, inequality (5)]. Therefore,

$$\begin{aligned} \int _{t_0}^{\infty } \frac{e^{-2\xi t}}{t} dt = \int _{2\xi t_0}^{\infty } \frac{e^{-t}}{t} dt \le e^{-2\xi t_0} \log \left( 1 + \frac{1}{2\xi t_0} \right) \end{aligned}$$

for all \(\xi >0\). There exists \(c>0\) such that for all \(0< \xi <1/2\),

$$\begin{aligned} \log \left( 1 + \frac{1}{2\xi t_0} \right) \le c\log \left( \frac{1}{\xi } \right) . \end{aligned}$$

Thus, (3.7) holds for all \(x \in D(A)\) and some suitable constant \(K >0\). \(\square \)

Combining Lemma 3.1 with the estimates (3.5)–(3.7), we obtain estimates for the decay rate of the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the difference equation (3.2) with smooth initial data.

Theorem 3.3

Let A be the generator of an exponentially stable \(C_0\)-semigroup \((e^{At})_{t \ge 0}\) on a Hilbert space H. Suppose that \(A^{-1}\) generates a bounded \(C_0\)-semigroup \((e^{A^{-1}t})_{t \ge 0}\) on H. If \(0< \inf _{n \in {\mathbb {N}}_0} \tau _n \le \sup _{n \in {\mathbb {N}}_0} \tau _n < \infty \), then there exists \(K>0\) such that the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the difference equation (3.2) satisfies the following statements:

  1. (i)

    For each \(0< \alpha < 1\), there exists \(K>0\) such that

    $$\begin{aligned} \Vert x_n\Vert \le \frac{K}{n^{\alpha /2}} \Vert (-A)^\alpha x_0\Vert \end{aligned}$$
    (3.8)

    for all \(x_0 \in D((-A)^{\alpha })\) and \(n \ge 1\).

  2. (ii)

    There exists \(K>0\) such that

    $$\begin{aligned} \Vert x_n\Vert \le K \sqrt{\frac{\log n}{n}} \Vert Ax_0\Vert \end{aligned}$$
    (3.9)

    for all \(x_0 \in D(A)\) and \(n \ge 2\).

Proof

Set

$$\begin{aligned} \tau _{\min }:= \inf _{n \in {\mathbb {N}}_0} \tau _n >0,\quad \tau _{\max }:= \sup _{n \in {\mathbb {N}}_0} \tau _n < \infty , \end{aligned}$$

and let the operator \(R(r) \in {\mathcal {L}}(H)\) be as in Lemma 3.1. Then there is \(M>0\) such that the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the difference equation (3.2) satisfies

$$\begin{aligned} \Vert x_n\Vert \le \frac{M \sqrt{\langle x_0, R(r)x_0 \rangle } }{(n+1)r^n \sqrt{1-r}} \end{aligned}$$
(3.10)

for all \(x_0 \in X\), \(n \in {\mathbb {N}}\), and \(r \in (0,1)\).

Let \(0 < \alpha \le 1\), and define

$$\begin{aligned} f_{\alpha }(\xi ):= {\left\{ \begin{array}{ll} \xi ^{\alpha -1}, &{}{} 0<\alpha <1, \\ \log \left( \dfrac{1}{\xi } \right) , &{}{} \alpha = 1 \end{array}\right. } \end{aligned}$$

for \(0< \xi < 1\). By the estimates (3.5)–(3.7), there exist \(K_1,K_2 >0\) and \(r_0\in (0,1)\) such that

$$\begin{aligned} \langle x_0, R(r)x_0 \rangle \le \frac{2K_1}{\tau _{\min }} \Vert x_0\Vert ^2 + 2\tau _{\max } K_2 \Vert (-A)^{\alpha }x_0\Vert ^2f_{\alpha }(\tau _{\min } \xi _r) \end{aligned}$$
(3.11)

for all \(x_0 \in D((-A)^\alpha )\) and \(r \in (r_0,1)\). Put \(r = n/(n+1)\) for \(n \in {\mathbb {N}}\). Then

$$\begin{aligned} \xi _r = \frac{1-r^2}{2(r^2+1)} = \frac{2n+1}{2(2n^2+2n+1)} \end{aligned}$$

and

$$\begin{aligned} \frac{1}{(n+1)r^n \sqrt{1-r}} = \frac{(1+1/n)^n}{\sqrt{n+1}}. \end{aligned}$$

Combining the estimates (3.10) and (3.11), we have that there exist \(K_4 >0\) and \(n_0 \in {\mathbb {N}}\) such that

$$\begin{aligned} \Vert x_n\Vert \le K_4 \sqrt{\frac{f_{\alpha } (1/n)}{n}} \Vert (-A)^{\alpha }x_0\Vert \end{aligned}$$

for all \(x_0 \in D((-A)^{\alpha })\) and \(n \ge n_0 + 1\). By definition,

$$\begin{aligned} \sqrt{\frac{f_{\alpha } (1/n)}{n}} = {\left\{ \begin{array}{ll} \dfrac{1}{n^{\alpha /2}}, &{} 0<\alpha <1, \\ \sqrt{\dfrac{\log n }{n}}, &{} \alpha = 1 \end{array}\right. } \end{aligned}$$

for all \(n \ge 2\). Since

$$\begin{aligned} \Vert A_d(\tau )\Vert = \left\| 2\left( I - \frac{\tau }{2}A \right) ^{-1} - I \right\| \le c \end{aligned}$$

for all \(\tau \in [\tau _{\min },\tau _{\max }]\) and some \(c \ge 1\), it follows that \(\Vert x_n\Vert \le c^{n_0} \Vert x_0\Vert \) for all \(x_0 \in X\) and \(0\le n \le n_0\). Thus, we obtain the desired conclusion. \(\square \)

We compare the norm-estimates (3.8) and (3.9) with those in the time-invariant case \(\tau _n \equiv 2\), using a simple example.

Example 3.4

Let the operator A on \(\ell ^2\) be as in Example 2.5. Define \(A_d:= A_d(2)= (I+A)(I-A)^{-1}\) and let \(0 < \alpha \le 1\). To obtain the decay rate of the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the time-invariant difference equation

$$\begin{aligned} x_{n+1} = A_d x_n,\quad n \in {\mathbb {N}}_0;\qquad x_0 \in D((-A)^{\alpha }), \end{aligned}$$

we estimate \(\Vert A_d^{n}(-A)^{-\alpha }\Vert \) for \(n \in {\mathbb {N}}_0\).

We have that for all \(n \in {\mathbb {N}}_0\),

$$\begin{aligned} \big \Vert A_d^{n^2}(-A)^{-\alpha }\big \Vert = \sup _{k \in {\mathbb {N}}} \left| \frac{1+\lambda _k}{1-\lambda _k} \right| ^{n^2} |\lambda _k|^{-\alpha } \ge \left| \frac{1+\lambda _n}{1-\lambda _n} \right| ^{n^2} |\lambda _n|^{-\alpha }. \end{aligned}$$

Moreover,

$$\begin{aligned} \left| \frac{1+\lambda _n}{1-\lambda _n} \right| ^{n^2} \rightarrow e^{-2\gamma },\quad n^{\alpha } |\lambda _n|^{-\alpha } \rightarrow 1 \end{aligned}$$

as \(n \rightarrow \infty \). Hence

$$\begin{aligned} \liminf _{n \rightarrow \infty } n^{\alpha } \Vert A_d^{n^2} (-A)^{-\alpha }\Vert \ge e^{-2\gamma }. \end{aligned}$$

This implies that the norm-estimate (3.8) for the case \(0< \alpha <1\) is optimal in the sense that one cannot obtain any better rates in general.

Next we show that

$$\begin{aligned} \Vert A_d^{n}A^{-1}\Vert = O\left( \frac{1}{\sqrt{n}} \right) \qquad (n \rightarrow \infty ). \end{aligned}$$
(3.12)

For all \(n \in {\mathbb {N}}\),

$$\begin{aligned} \Vert A_d^{n}A^{-1}\Vert = \sup _{k \in {\mathbb {N}}} \left| \frac{1+\lambda _k}{1-\lambda _k} \right| ^{n} |\lambda _k|^{-1} \le \sqrt{\sup _{w \ge \gamma ^2 + 1} f_n(w)}, \end{aligned}$$

where

$$\begin{aligned} f_n(w):= \left( \frac{w+1-2\gamma }{w+1+2\gamma } \right) ^n \frac{1}{w}. \end{aligned}$$

A simple calculation shows that for all sufficiently large \(n \in {\mathbb {N}}\),

$$\begin{aligned} \sup _{w \ge \gamma ^2 + 1} f_n(w) = f_n(w_n), \end{aligned}$$

where

$$\begin{aligned} w_n:= 2\sqrt{\gamma ^2n^2+\gamma ^2-\gamma n} + 2\gamma n - 1. \end{aligned}$$

Since

$$\begin{aligned} \left( \frac{w_n+1-2\gamma }{w_n+1+2\gamma } \right) ^n \rightarrow e^{-1},\quad \frac{n}{w_n} \rightarrow \frac{1}{4\gamma } \end{aligned}$$

as \(n \rightarrow \infty \), we conclude that the estimate (3.12) holds. It is not clear, in general, whether one can remove the logarithmic term \(\sqrt{\log n}\) in the estimate (3.9) for the case \(\alpha =1\).

3.2 Normal Generators of Polynomially Stable Semigroups

Suppose that A is a normal operator on a Hilbert space H and generates a polynomially stable \(C_0\)-semigroup. As in the case of exponentially stable \(C_0\)-semigroups, one can obtain an estimate for the decay rate of the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the difference equation (3.2) with smooth initial data. Note that \(A^{-1}\) generates a contraction \(C_0\)-semigroup on H; see [29, Lemma 4.1].

The following lemma gives estimates analogous to those in Lemma 3.2.

Lemma 3.5

Let H be a Hilbert space and let \(A:D(A)\subset H \rightarrow H\) be a normal operator generating a polynomially stable \(C_0\)-semigroup \((e^{At})_{t \ge 0}\) with parameter \(\beta >0\) on H. Then \(Q(\xi ) \in {\mathcal {L}}(H)\) given by (3.4b) for \(\xi >0\) satisfies the following statements:

  1. (i)

    For each \(0< \alpha < 1+\beta /2\), there exists \(K>0\) such that

    $$\begin{aligned} \sup _{0< \xi < 1} \xi ^{1-\frac{2\alpha }{2+\beta }} \langle x, Q(\xi )x \rangle \le K \Vert (-A)^{\alpha }x\Vert ^2 \end{aligned}$$
    (3.13)

    for all \(x \in D((-A)^{\alpha })\).

  2. (ii)

    There exists \(K>0\) such that

    $$\begin{aligned} \sup _{0< \xi < 1/2} \frac{ \langle x, Q(\xi )x \rangle }{\log (1/\xi )} \le K \Vert (-A)^{1+\beta /2}x\Vert ^2 \end{aligned}$$
    (3.14)

    for all \(x \in D((-A)^{1+\beta /2})\).

Proof

Let \(0< \alpha \le 1+\beta /2\), and put

$$\begin{aligned} {{\widetilde{\alpha }}}:= \frac{2\alpha }{2+\beta } \le 1. \end{aligned}$$

By Proposition 2.8, there exist \(M, t_0 >0\) such that

$$\begin{aligned} \big \Vert e^{A^{-1}t}x\big \Vert \le \frac{M \Vert (-A)^{\alpha }x\Vert }{t^{{{\widetilde{\alpha }}} / 2}} \end{aligned}$$
(3.15)

for all \(x \in D((-A)^{\alpha })\) and \(t \ge t_0\). The rest of the proof is quite similar to that of Lemma 3.2, and hence we omit it. \(\square \)

If A is the generator of a polynomially stable \(C_0\)-semigroup \((e^{At})_{t \ge 0}\) with parameter \(\beta >0\), then [3, Proposition 3.1] shows that for all \(\alpha >0\), there exist \(M,t_0 >0\) such that

$$\begin{aligned} \big \Vert e^{A t}x\big \Vert \le \frac{M \Vert (-A)^{\alpha } x\Vert }{t^{\alpha /\beta }} \end{aligned}$$

for all \(x \in D((-A)^{\alpha })\) and \(t \ge t_0\). The decay rate \(t^{-\alpha /\beta }\) is faster than the decay rate of \(\Vert e^{A^{-1}t}x\Vert \) in (3.15). Hence, the operator \(P(\xi )\) given by (3.4a) satisfies the same estimates as \(Q(\xi )\).

Using the estimates on \(P(\xi )\) and \(Q(\xi )\), we derive a norm-estimate for the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the difference equation (3.2). Since this result can be obtained by the same arguments as in the proof of Theorem 3.3, we omit the proof.

Proposition 3.6

Let H be a Hilbert space and let \(A:D(A)\subset H \rightarrow H\) be a normal operator generating a polynomially stable \(C_0\)-semigroup \((e^{At})_{t \ge 0}\) with parameter \(\beta >0\) on H. If \(0< \inf _{n \in {\mathbb {N}}_0} \tau _n \le \sup _{n \in {\mathbb {N}}_0} \tau _n < \infty \), then the solution \((x_n)_{n \in {\mathbb {N}}_0}\) of the difference equation (3.2) satisfies the following statements:

  1. (i)

    For each \(0< \alpha < 1 + \beta /2\), there exists \(K>0\) such that

    $$\begin{aligned} \Vert x_n\Vert \le \frac{K}{n^{\alpha /(2+\beta )}} \Vert (-A)^\alpha x_0\Vert \end{aligned}$$
    (3.16)

    for all \(x_0 \in D((-A)^{\alpha })\) and \(n \ge 1\).

  2. (ii)

    There exists \(K>0\) such that

    $$\begin{aligned} \Vert x_n\Vert \le K \sqrt{\frac{\log n}{n}} \Vert (-A)^{1+\beta /2}x_0\Vert \end{aligned}$$
    (3.17)

    for all \(x_0 \in D((-A)^{1+\beta /2})\) and \(n \ge 2\).

Example 3.7

Consider again the operator A on \(\ell ^2\) in Example 2.9. We recall that the parameter \(\beta \) of polynomial decay is given by \(\beta = 1\). Define \(A_d:= A_d(2)= (I+A)(I-A)^{-1}\). It has been shown in [25, Example 4.6] that the estimate \(\Vert A_d^n (-A)^{-3/m}\Vert = O(n^{-1/m})\) is optimal for all \(m \in {\mathbb {N}}\). Hence we see from the case \(m =3\) that the estimate (3.16) cannot in general be improved. Although \(\Vert A_d^n (-A)^{-3/2}\Vert = O(1/\sqrt{n})\) can be deduced from the case \(m=2\), it is open whether the logarithmic term \(\sqrt{\log n}\) in the norm-estimate (3.17) can be omitted.