Introduction
The stability of a linear evolution equation tells us what happens eventually, while its transient tells us what happens on the way, and for non-normal generators the two can be very far apart. In a standard finite dimensional case, i.e. for a matrix A ∈ C d × d A \in\mathbb{C}^{d \times d} A ∈ C d × d the transient growth is bounded by the Kreiss constant K 1 ( A ) : = sup Re z > 0 Re ( z ) ∥ ( z − A ) − 1 ∥ \mathcal{K}_1(A) := \sup_{\operatorname{Re} z>0} \operatorname{Re}(z) \lVert(z-A)^{-1}\rVert K 1 ( A ) := sup Re z > 0 Re ( z ) ∥( z − A ) − 1 ∥ , and the continuous-time Kreiss matrix theorem
K 1 ( A ) ≤ sup t ≥ 0 ∥ e t A ∥ ≤ e d K 1 ( A ) (1) \mathcal{K}_1(A) \le\sup_{t\geq0}\lVert e^{tA}\rVert\le e d\,\mathcal{K}_1(A) \tag*{(1)} K 1 ( A ) ≤ t ≥ 0 sup ∥ e t A ∥ ≤ e d K 1 ( A ) (1)
states that the peak of the transient growth is controlled from both sides by a single resolvent quantity, with a factor linear in the dimension [29 , 30 , 31 , 44 , 47 , 48 ] . This two-sided estimate is the backbone of the pseudospectral theory of transient growth [47 ] , and its infinite-dimensional limits are by now fully understood: on Banach spaces the Kreiss resolvent condition gives no bound on the semigroup at all, while on Hilbert spaces it gives sublinear growth in time [1 , 2 , 9 , 13 ] .
The purpose of this paper is to develop the corresponding theory for fractional Cauchy problems
∂ t α u = L u , u ( 0 ) = u 0 , 0 < α ≤ 1 , (2) \partial_t^\alpha u = Lu,\qquad u(0)=u_0,\qquad0<\alpha\leq1, \tag*{(2)} ∂ t α u = Lu , u ( 0 ) = u 0 , 0 < α ≤ 1 , (2)
with the Caputo derivative ∂ t α \partial_t^\alpha ∂ t α and a closed operator L L L on a Banach space X X X . Such problems describe evolutions with memory: subdiffusive transport in heterogeneous or crowded media arises as the scaling limit of continuous-time random walks with heavy-tailed waiting times [33 , 34 , 36 ] , the time-fractional Schrödinger equation was proposed as a quantum model with memory [37 ] , and fractional relaxation and oscillation equations are standard models in viscoelasticity and control [17 , 32 , 38 ] . The solution of (2) is u ( t ) = E α ( t α L ) u 0 u(t)=E_\alpha(t^\alpha L)u_0 u ( t ) = E α ( t α L ) u 0 , where E α ( t α L ) E_\alpha(t^\alpha L) E α ( t α L ) is the Mittag–Leffler family generated by L L L [4 , 40 ] , given for bounded or sectorial L L L by the Mittag–Leffler function E α ( z ) = ∑ k ≥ 0 z k Γ ( 1 + α k ) E_\alpha(z)=\sum_{k\geq0}\frac{z^k}{\Gamma(1+\alpha k)} E α ( z ) = ∑ k ≥ 0 Γ ( 1 + α k ) z k .
Already at the level of definitions, these families differ from semigroups in ways that call for a theory of their own. The map t ↦ E α ( t α L ) t\mapsto E_\alpha(t^\alpha L) t ↦ E α ( t α L ) has no semigroup property, so arguments that propagate a bound from short to long times are not available. The stability region is larger: for a matrix L L L the family is bounded if and only if the spectrum avoids the open sector Σ α : = { z ≠ 0 : ∣ Arg z ∣ < α π 2 } \Sigma_\alpha:=\{z\ne0:|\operatorname{Arg}z|<\frac{\alpha\pi}{2}\} Σ α := { z = 0 : ∣ Arg z ∣ < 2 α π } and the eigenvalues on its boundary are semisimple [32 ] . Since Σ α \Sigma_\alpha Σ α is a proper subsector of the right half-plane when α < 1 \alpha<1 α < 1 , modes that grow exponentially in physical time can be perfectly stable under the fractional clock, and the subordination formula, which writes E α ( t α L ) E_\alpha(t^\alpha L) E α ( t α L ) as an average of e s L e^{sL} e s L over a random time [4 ] , then gives no information, because it averages an exponentially growing semigroup.
The envelope t ↦ ∥ E α ( t α L ) ∥ t\mapsto\lVert E_\alpha(t^\alpha L)\rVert t ↦ ∥ E α ( t α L )∥ in general has a short-time, an intermediate and a long-time regime, each governed by a different resolvent quantity. In particular, in Hilbert spaces for the semigroup case α = 1 \alpha=1 α = 1 , it is well understood that numerical and spectral abscissa govern short-time and long-time behavior, respectively, see e.g. [47 ] . While this remains also the case in the fractional setting (we review these essentially known results in Section Appendix B.3), the characterization of intermediate behavior is much more involved problem. Indeed, whenever the operator L L L is normal, while the envelope of the semigroup ∥ e t L ∥ \lVert e^{tL}\rVert ∥ e t L ∥ is sharply bounded by one, in the fractional setting ∣ E α ∣ |E_\alpha| ∣ E α ∣ exceeds 1 / α 1/\alpha 1/ α at points of the boundary rays of Σ α \Sigma_\alpha Σ α , implying that even a normal generator with spectrum on these rays amplifies envelope by more than 1 / α 1/\alpha 1/ α , thus provoking transient growth. Hence, in this paper we focus on answering the question:
Which resolvent quantity controls sup t ≥ 0 ∥ E α ( t α L ) ∥ \sup_{t\ge0}\|E_\alpha(t^\alpha L)\| sup t ≥ 0 ∥ E α ( t α L ) ∥ , and with which constants?
In answering it, we particularly focus on the finite-rank operators on infinite-dimensional spaces, which is motivated by recent works on data-driven estimators of transfer operators and generators of dynamical systems [21 , 23 , 24 , 26 ] , including estimators of slow modes of molecular systems [11 ] , and their pseudospectra can be computed at a cost governed by the rank [27 ] .
Related work
The continuous-time Kreiss matrix theorem goes back to Kreiss [29 , 30 ] ; the linear dependence on the dimension was obtained by LeVeque and Trefethen [31 ] , the constant e e e through Spijker’s [44 , 48 , 50 ] , and Kraaijevanger [28 ] showed that the linear order is sharp, for powers as well as for the exponential; see also Spijker et al. [43 ] , and for general domains Toh and Trefethen [46 ] . Eisner and Zwart [13 ] showed that the Kreiss condition gives no growth bound on Banach spaces, that it gives ∥ e t L ∥ = O ( t ) \|e^{tL}\|=O(t) ∥ e t L ∥ = O ( t ) on Hilbert spaces, and that growth of order t γ t^\gamma t γ occurs for every γ < 1 \gamma<1 γ < 1 , using Fourier multipliers on L 2 L^2 L 2 spaces with power weights of class A 2 A_2 A 2 . The Hilbert-space bound was improved to O ( t log t ) O\left(\frac{t}{\sqrt{\log t}}\right) O ( l o g t t ) by Arnold [1 ] , extended to general resolvent growth by Dell’Oro [9 ] and Dell’Oro et al. [10 ] , and very recently to O ( t 1 − ε ) O(t^{1-\varepsilon}) O ( t 1 − ε ) with ε \varepsilon ε depending on the Kreiss constant [2 ] ; see also Rozendaal and Veraar [41 ] . All these arguments use the semigroup law. On the side of spectral sets, the numerical range is a ( 1 + 2 ) (1+\sqrt{2}) ( 1 + 2 ) -spectral set [8 ] , von Neumann’s inequality makes a half-plane containing the numerical range a spectral set with constant one [45 , 49 ] , and Crouzeix [7 ] interpolates between the two for unbounded convex sets.
Mittag–Leffler families, also called α \alpha α -times resolvent families or solution operators, are the resolvent families of Volterra equations with the kernel t α − 1 Γ ( α ) \frac{t^{\alpha-1}}{\Gamma(\alpha)} Γ ( α ) t α − 1 [4 , 40 ] . Their generation theory, subordination and asymptotics are classical [4 , 17 , 40 ] , and resolvent characterisations of growth bounds and of uniform stability were obtained recently by Ghavidel and Shahali [16 ] and Mei and Li [35 ] , in the spirit of the Hille–Yosida and Gearhart–Prüss theorems. These results are qualitative or asymptotic, and we are not aware of quantitative two-sided bounds of Kreiss type for Mittag–Leffler families beyond the matrix case conjectured in Kostić and Uzelac Bujišić [22 ] . For matrices, the spectral criterion of Matignon [32 ] characterises boundedness, and fractional pseudospectra have been used to localise spectra and to study stability of fractional systems [12 , 42 ] .
Organisation
For the reader’s convince, we start with the summerize of our main results as four Theorems A–D in § 2 . Then, in § 3 we review facts on the Mittag–Leffler functions, Mittag–Leffler operator families, and prove basic properties of the newly introduced the fractional Kreiss constant. Based on these preliminary results, § 4 proves Theorem B , and § 5 the finite-rank core of Theorem A. § 6 treats normal operators and the numerical range, and § 8 completes Theorem A for generators with a slow component and analyses prototype (a). § 9 proves Theorem C , § 10 describes the mechanisms and time scales of the transient behavior, § 11 revisits the three prototypes introduced in § 2 , and § 12 discusses open problems. The appendix contains the table of notation and the proofs of auxiliary results, and the Supplementary Material the numerical methods, the complete tables, the study of random finite-rank operators and a description of the code.
Main contributions in a nutshell
Recalling that the Laplace transform of E α ( t α L ) E_\alpha(t^\alpha L) E α ( t α L ) is ω α − 1 ( ω α − L ) − 1 \omega^{\alpha-1}(\omega^\alpha-L)^{-1} ω α − 1 ( ω α − L ) − 1 , somewhat natural way to define the fractional Kreiss constant is
K α ( L ) : = sup Re ω > 0 Re ( ω ) ∥ ω α − 1 ( ω α − L ) − 1 ∥ = sup z ∈ Σ α a α ( z ) ∥ ( z − L ) − 1 ∥ , (3) \mathcal{K}_\alpha(L) := \sup_{\operatorname{Re}\omega>0} \operatorname{Re}(\omega)\left\|\omega^{\alpha-1}(\omega^\alpha-L)^{-1}\right\| = \sup_{z\in\Sigma_\alpha} a_\alpha(z)\left\|(z-L)^{-1}\right\|, \tag*{(3)} K α ( L ) := Re ω > 0 sup Re ( ω ) ω α − 1 ( ω α − L ) − 1 = z ∈ Σ α sup a α ( z ) ( z − L ) − 1 , (3)
where a α ( z ) : = ∣ z ∣ cos ( Arg z α ) a_\alpha(z):=|z|\cos\left(\frac{\operatorname{Arg}z}{\alpha}\right) a α ( z ) := ∣ z ∣ cos ( α Arg z ) ; it reduces to K 1 \mathcal{K}_1 K 1 for α = 1 \alpha=1 α = 1 . The second expression replaces the classical weight Re z \operatorname{Re}z Re z on the right half-plane by a weight on Σ α \Sigma_\alpha Σ α that vanishes on the boundary rays, so K α ( L ) \mathcal{K}_\alpha(L) K α ( L ) is finite for operators that are unstable as generators of semigroups. Up to a factor between 1 1 1 and 1 α \frac{1}{\alpha} α 1 , a α ( z ) a_\alpha(z) a α ( z ) is the distance from z z z to the boundary of Σ α \Sigma_\alpha Σ α (Theorem 3.6(c)), so K α ( L ) \mathcal{K}_\alpha(L) K α ( L ) measures how fast the resolvent may blow up as z z z approaches the stability region from inside the sector of instability. A constant of this type, for matrices and in pseudospectral form, was proposed in the precursor [22 ] , where a fractional Kreiss matrix theorem was conjectured; the present paper proves this conjecture in a stronger form, for all real α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] , on Banach spaces and with the rank in place of the dimension.
Our main object is a generator with a finite slow component, i.e. a system whose mid and long-term dynamics is determined by finitely many eigenvalues that are well separated from the fast time-scales. More precisely, a projection P ∈ B ( X ) P\in\mathcal{B}(X) P ∈ B ( X ) of finite rank m m m is a slow component of L L L if it commutes with L L L , the part of L L L in ( I − P ) X (I-P)X ( I − P ) X generates a bounded Mittag–Leffler family, spectra of P L PL P L and ( I − P ) L (I-P)L ( I − P ) L are disjoint with the former being closer to stability boundary.
As we first show the transient growth of such a generator is carried out by the m m m slow modes and by the geometry of the splitting, and both are seen by the deflated Kreiss constant
K α ( L ; P ) : = sup z ∈ Σ α a α ( z ) ∥ ( z − L ) − 1 P ∥ , \mathcal{K}_\alpha(L;P):=\sup_{z\in\Sigma_\alpha}a_\alpha(z)\left\|(z-L)^{-1}P\right\|, K α ( L ; P ) := z ∈ Σ α sup a α ( z ) ( z − L ) − 1 P ,
the Kreiss quantity of the compressed resolvent.
Theorem A (Theorems 5.1 , 5.4 and 7.1 ). Let 0 < α ≤ 1 0<\alpha\leq1 0 < α ≤ 1 , let X ≠ { 0 } X\neq\{0\} X = { 0 } be a Banach space, let L L L be a closed, densely defined operator on X X X , and put
κ α , n : = e 2 ( 3 + 4 n ( 2 − α ) ) , κ α , n ′ : = e 2 ( 1 + 4 n ( 2 − α ) ) . \kappa_{\alpha,n}:=\frac{\mathrm{e}}{2}\left(3+4n(2-\alpha)\right),\qquad\kappa'_{\alpha,n}:=\frac{\mathrm{e}}{2}\left(1+4n(2-\alpha)\right). κ α , n := 2 e ( 3 + 4 n ( 2 − α ) ) , κ α , n ′ := 2 e ( 1 + 4 n ( 2 − α ) ) . (i) If L L L generates a bounded Mittag–Leffler family, then K α ( L ) ≤ sup t ≥ 0 ∥ E α ( t α L ) ∥ \mathcal{K}_\alpha(L)\leq\sup_{t\geq0}\left\|E_\alpha(t^\alpha L)\right\| K α ( L ) ≤ sup t ≥ 0 ∥ E α ( t α L ) ∥ .
(ii) If P P P is a slow component of rank m m m , then L L L generates a Mittag–Leffler family,
K α ( L ; P ) ≤ sup t ≥ 0 ∥ E α ( t α L ) P ∥ ≤ κ α , m K α ( L ; P ) , \mathcal{K}_\alpha(L;P)\leq\sup_{t\geq0}\left\|E_\alpha(t^\alpha L)P\right\|\leq\kappa_{\alpha,m}\mathcal{K}_\alpha(L;P), K α ( L ; P ) ≤ t ≥ 0 sup ∥ E α ( t α L ) P ∥ ≤ κ α , m K α ( L ; P ) , sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ κ α , m K α ( L ; P ) + sup t ≥ 0 ∥ E α ( t α L ) ( I − P ) ∥ , \sup_{t\geq0}\left\|E_\alpha(t^\alpha L)\right\|\leq\kappa_{\alpha,m}\mathcal{K}_\alpha(L;P)+\sup_{t\geq0}\left\|E_\alpha(t^\alpha L)(I-P)\right\|, t ≥ 0 sup ∥ E α ( t α L ) ∥ ≤ κ α , m K α ( L ; P ) + t ≥ 0 sup ∥ E α ( t α L ) ( I − P ) ∥ , and κ α , m \kappa_{\alpha,m} κ α , m may be replaced by κ α , m ′ \kappa'_{\alpha,m} κ α , m ′ when 0 0 0 is not an eigenvalue of the slow block.
(iii) If L L L is bounded with rank L ≤ n \operatorname{rank} L \le n rank L ≤ n , then sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ κ α , n K α ( L ) \sup_{t\ge0}\|E_{\alpha}(t^{\alpha}L)\| \le\kappa_{\alpha,n}\mathcal{K}_{\alpha}(L) sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ κ α , n K α ( L ) .
The constants depend on α \alpha α and on the rank, never on the dimension of X X X , and they are uniform in α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] .
Part (iii) is the core case of (ii): a finite-rank operator with finite Kreiss constant splits as X = Ran L ⊕ ker L X = \operatorname{Ran} L \oplus\ker L X = Ran L ⊕ ker L , its slow component is the projection onto Ran L \operatorname{Ran} L Ran L along ker L \ker L ker L , and the stationary contribution of the kernel is absorbed into the estimate without paying the norm of the oblique complementary projection. Part (ii) applies to unbounded generators that may have infinite rank, need not generate a C 0 C_0 C 0 -semigroup and may have spectrum in the right half-plane, as in prototype (a), and the fast part may be controlled on Banach spaces by sectoriality or on Hilbert spaces by the numerical range.
Second, we note that the upper bounds rest on a Bernstein-type inequality that can be of independent interest. Inverting the Laplace transform after one integration by parts reduces the problem to the length of the curve y ↦ h F ( a + i y ) y \mapsto h_F(a+iy) y ↦ h F ( a + i y ) , h F ( ω ) : = ω − 1 F ( ω α ) h_F(\omega) := \omega^{-1}F(\omega^\alpha) h F ( ω ) := ω − 1 F ( ω α ) , where F F F is a rational function of degree at most n n n built from a matrix element of the resolvent. Let R α , n \mathcal{R}_{\alpha,n} R α , n denote the nonzero rational functions F = F 1 F 0 F = \frac{F_1}{F_0} F = F 0 F 1 in lowest terms with deg F 1 ≤ deg F 0 ≤ n \deg F_1 \le\deg F_0 \le n deg F 1 ≤ deg F 0 ≤ n and no poles in Σ α \Sigma_\alpha Σ α .
Theorem B. (Theorem 4.1 ). Let 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 , n ∈ N 0 n \in\mathbb{N}_0 n ∈ N 0 , F ∈ R α , n F \in\mathcal{R}_{\alpha,n} F ∈ R α , n and a > 0 a > 0 a > 0 . Then
2 sup y ∈ R ∣ h F ( a + i y ) ∣ ≤ ∫ R ∣ h F ′ ( a + i y ) ∣ d y ≤ π ( 3 + 4 n ( 2 − α ) ) sup y ∈ R ∣ h F ( a + i y ) ∣ , 2\sup_{y\in\mathbb{R}}|h_F(a+iy)| \le\int_{\mathbb{R}}|h'_F(a+iy)|\,\mathrm{d}y \le\pi\bigl(3+4n(2-\alpha)\bigr)\sup_{y\in\mathbb{R}}|h_F(a+iy)|, 2 y ∈ R sup ∣ h F ( a + i y ) ∣ ≤ ∫ R ∣ h F ′ ( a + i y ) ∣ d y ≤ π ( 3 + 4 n ( 2 − α ) ) y ∈ R sup ∣ h F ( a + i y ) ∣ , and the constant improves to π ( 1 + 4 n ( 2 − α ) ) \pi(1+4n(2-\alpha)) π ( 1 + 4 n ( 2 − α )) when F ( 0 ) = 0 F(0)=0 F ( 0 ) = 0 .
For α = 1 \alpha=1 α = 1 the function h F h_F h F is rational and the classical route is Spijker’s lemma, which counts the intersections of algebraic curves [44 , 50 ] . For irrational α \alpha α the curve { ( a + i y ) α : y ∈ R } \{(a+iy)^\alpha: y\in\mathbb{R}\} {( a + i y ) α : y ∈ R } is not algebraic and such counts are not available. We replace them by a bound on the total turning of arg h F ′ \arg h'_F arg h F ′ , which follows from the convexity of the image of a vertical line under ω ↦ ω α \omega\mapsto\omega^\alpha ω ↦ ω α : its tangent turns monotonically through an angle ( 1 − α ) π (1-\alpha)\pi ( 1 − α ) π , and this forces the argument of ( a + i y ) α − ζ (a+iy)^\alpha-\zeta ( a + i y ) α − ζ to vary by at most ( 2 − α ) π (2-\alpha)\pi ( 2 − α ) π for every point ζ \zeta ζ off the curve.
Next, we obtain that rank dependence in Theorem A cannot be removed, and show that one family of weighted spaces on the circle suffices to see it. For 0 < γ < 1 0 < \gamma< 1 0 < γ < 1 let L γ 2 L^2_\gamma L γ 2 be the space L 2 ( T , ∣ tan ( x 2 ) ∣ γ d x ) L^2(\mathbb{T},|\tan(\frac{x}{2})|^\gamma\,\mathrm{d}x) L 2 ( T , ∣ tan ( 2 x ) ∣ γ d x ) , and let D α D^\alpha D α be the Weyl fractional derivative, the Fourier multiplier with symbol ( i k ) α (ik)^\alpha ( ik ) α .
Theorem C. (Theorems 8.2 to 8.5). Let 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 and 0 < γ < 1 0 < \gamma< 1 0 < γ < 1 . The rank-2 N 2N 2 N truncations B N B_N B N of N − α D α N^{-\alpha}D^\alpha N − α D α to the trigonometric polynomials of degree at most N N N satisfy sup N K α ( B N ) < ∞ \sup_N\mathcal{K}_\alpha(B_N)<\infty sup N K α ( B N ) < ∞ and sup t ≥ 0 ∥ E α ( t α B N ) ∥ ≥ c N γ \sup_{t\ge0}\|E_\alpha(t^\alpha B_N)\|\ge cN^\gamma sup t ≥ 0 ∥ E α ( t α B N ) ∥ ≥ c N γ on L γ 2 L^2_\gamma L γ 2 . Consequently the best constant in Theorem A(iii) grows at least like n γ n^\gamma n γ for every γ < 1 \gamma<1 γ < 1 ; there is a bounded operator on a Hilbert space with K α ( L ) < ∞ \mathcal{K}_\alpha(L)<\infty K α ( L ) < ∞ and lim sup t → ∞ t − γ ∥ E α ( t α L ) ∥ > 0 \limsup_{t\to\infty}t^{-\gamma}\|E_\alpha(t^\alpha L)\|>0 lim sup t → ∞ t − γ ∥ E α ( t α L ) ∥ > 0 ; and D α D^\alpha D α itself satisfies K α ( D α ) < ∞ \mathcal{K}_\alpha(D^\alpha)<\infty K α ( D α ) < ∞ on L γ 2 L^2_\gamma L γ 2 but generates no Mittag–Leffler family.
The growth in this construction is oscillatory: on the boundary rays E α ( ( i s ) α ) = α − 1 e − i s E_\alpha((is)^\alpha)=\alpha^{-1}e^{-is} E α (( i s ) α ) = α − 1 e − i s up to an error of bounded variation, so the Mittag–Leffler family of D α D^\alpha D α contains a translation, and translation by π \pi π is unbounded on L γ 2 L^2_\gamma L γ 2 because it moves the zero of the weight onto its pole. For α = 1 \alpha=1 α = 1 this recovers the weighted Fourier multipliers of Eisner and Zwart [13 ] .
Then, we focus on Hilbert spaces where the fast part of the dynamics is often controlled by its numerical range, and in this setting give a sharp answer with the fractional theory. Let M α ∗ : = sup z ∉ Σ α ∣ E α ( z ) ∣ M_\alpha^* := \sup_{z\notin\Sigma_\alpha}|E_\alpha(z)| M α ∗ := sup z ∈ / Σ α ∣ E α ( z ) ∣ , which is the best uniform bound for normal generators with spectrum outside Σ α \Sigma_\alpha Σ α and satisfies M α ∗ > 1 α M_\alpha^* > \frac{1}{\alpha} M α ∗ > α 1 for α < 1 \alpha< 1 α < 1 .
Theorem D (Theorem 6.4 ). Let L L L be a closed, densely defined operator on a Hilbert space whose numerical range does not meet Σ α \Sigma_\alpha Σ α and whose resolvent set meets Σ α \Sigma_\alpha Σ α . Then L L L generates a Mittag–Leffler family and sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ M α ∗ \sup_{t\geq0}\|E_\alpha(t^\alpha L)\| \leq M_\alpha^* sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ M α ∗ , with equality for suitable normal operators. No spectral-set estimate on a convex set containing the numerical range yields a better uniform bound, whatever its constant.
The generation statement covers operators that generate no C 0 C_0 C 0 -semigroup, such as the generator of prototype (b), and the proof is by separation and Sz.-Nagy’s dilation theorem. The bound is consistent with, but not covered by, Crouzeix’s recent estimates for unbounded convex sets [7 ] : its constant is that of a half-plane, where his spectral constant equals one, but in the extremal case the functions E α ( t α ⋅ ) E_\alpha(t^\alpha\cdot) E α ( t α ⋅ ) have no limit at infinity and fall outside his function class. The last statement of Theorem D is elementary once the right geometric lemma is in place, and it explains why the smaller constants that Crouzeix [7 ] obtains for sectors cannot help for the uniform bound: the union over t t t of the dilated sets generates a cone on which ∣ E α ∣ |E_\alpha| ∣ E α ∣ has the same supremum as on a half-plane. Combined with Theorem A, Theorem D gives explicit certificates for generators whose fast part has its numerical range outside Σ α \Sigma_\alpha Σ α .
Finally, to show that the fractional transient has a structure of its own, note that normal operators amplify by M α ∗ M_\alpha^* M α ∗ , a factor larger than 1 α \frac{1}{\alpha} α 1 that tends to one as α ↑ 1 \alpha\uparrow1 α ↑ 1 . On the other hand, as we describe in § 10 , strongly non-normal Jordan blocks, which produce the dimension-dependent amplification e n 2 π e\sqrt{\frac{n}{2\pi}} e 2 π n in (1) , produce for each fixed α < 1 \alpha< 1 α < 1 only a bounded amplification ratio. This is because the Mittag–Leffler distribution spreads the peak of s n − 1 e − s ( n − 1 ) ! \frac{s^{n-1}e^{-s}}{(n-1)!} ( n − 1 )! s n − 1 e − s over a range of times, and the bound deteriorates as α ↑ 1 \alpha\uparrow1 α ↑ 1 , as it must, suggesting that the growth exhibited in Theorem C is oscillatory rather than nilpotent in origin. In addition to this informative Jordan block example, for the three prototypes bellow, in § 11 we compare both sides of Theorem A with the bounds that are otherwise available: the semigroup route through subordination, the numerical-range bounds of Crouzeix and Palencia [8 ] and Crouzeix [7 ] , deflation bounds that estimate the slow block and the splitting separately, and, where a complete eigenbasis exists, the eigenvector bound.
(a) Subdiffusion driven by a self-excited oscillator. A two-dimensional oscillatory mode with eigenvalues e ± i θ e^{\pm i\theta} e ± i θ , α π 2 < θ < π 2 \frac{\alpha\pi}{2}<\theta<\frac{\pi}{2} 2 α π < θ < 2 π , grows exponentially as an ordinary differential equation but is stable under the fractional clock; when it feeds a source into a subdiffusive medium, the generator is unbounded, it has a two-dimensional slow component, its semigroup grows exponentially, and its numerical range meets Σ α \Sigma_\alpha Σ α .
(b) The time-fractional Schrödinger equation. The generator e − i α π 2 A V e^{-i\frac{\alpha\pi}{2}}A_V e − i 2 α π A V , with A V A_V A V a self-adjoint Schrödinger operator, is normal and generates no C 0 C_0 C 0 -semigroup because A V A_V A V is unbounded above, yet its Mittag–Leffler family is uniformly bounded and amplifies by exactly the sharp constant of normal operators.
(c) Time-changed Markov dynamics. Expectations of a Markov process run on the clock of an inverse stable subordinator evolve by E α ( t α L ) E_{\alpha}(t^{\alpha}L) E α ( t α L ) , with L L L the generator of the process. Linear chains of compartments are strongly non-normal, and in data-driven practice the generator is replaced by a finite-rank estimate on an infinite-dimensional function space [21 , 23 , 25 , 26 ] , whose number of retained modes, and not the dimension of the ambient space, is the natural parameter.
In prototype (a) the semigroup and numerical-range bounds are all infinite, while the deflated Kreiss constant follows the propagator within 13% for every coupling strength in all computed cases; in prototype (b) the amplification equals M α ∗ M_{\alpha}^{*} M α ∗ exactly, although no semigroup exists; and in prototype (c) the classical amplification ratio of linear compartment chains grows like the square root of their length while the fractional one stays small, and for low-rank generators learned from data the slow-component quantities are computed at a cost set by the rank of the estimator, while a spurious eigenvalue in Σ α \Sigma_{\alpha} Σ α is flagged by an infinite deflated Kreiss constant. The numerical methods, the complete data and a statistical study of random finite-rank operators are collected in a separate Supplementary Material, since none of the results of the paper depends on them.
Preliminary results
Notation
All powers are principal, \omega^{\beta}:=e^{\beta\Log\omega} with Arg ∈ ( − π , π ] \operatorname{Arg}\in(-\pi,\pi] Arg ∈ ( − π , π ] , and 0 β : = 0 0^{\beta}:=0 0 β := 0 for β > 0 \beta>0 β > 0 . We write H : = { ω ∈ C : Re ω > 0 } \mathbb{H}:=\{\omega\in\mathbb{C}:\operatorname{Re}\omega>0\} H := { ω ∈ C : Re ω > 0 } for the open right half-plane and, for 0 < θ ≤ π 0<\theta\le\pi 0 < θ ≤ π , Σ ( θ ) : = { z ∈ C ∖ { 0 } : ∣ Arg z ∣ < θ } \Sigma(\theta):=\{z\in\mathbb{C}\setminus\{0\}:|\operatorname{Arg}z|<\theta\} Σ ( θ ) := { z ∈ C ∖ { 0 } : ∣ Arg z ∣ < θ } for the open sector of half-opening θ \theta θ . Two sets appear throughout,
Σ α : = Σ ( α π 2 ) and Λ α : = C ∖ Σ α , \Sigma_{\alpha}:=\Sigma\left(\frac{\alpha\pi}{2}\right)\qquad\text{and}\qquad\Lambda_{\alpha}:=\mathbb{C}\setminus\Sigma_{\alpha}, Σ α := Σ ( 2 α π ) and Λ α := C ∖ Σ α ,
namely the open sector of instability and the closed fractional stability region; Λ α \Lambda_{\alpha} Λ α contains 0 0 0 , the left half-plane and the two boundary rays { r e ± i α π / 2 : r ≥ 0 } \{re^{\pm i\alpha\pi/2}:r\ge0\} { r e ± i α π /2 : r ≥ 0 } . For 0 < α < 2 0<\alpha<2 0 < α < 2 the map ω ↦ ω α \omega\mapsto\omega^{\alpha} ω ↦ ω α is a biholomorphism of H \mathbb{H} H onto Σ α \Sigma_{\alpha} Σ α , and the weight
a α ( z ) : = ∣ z ∣ cos ( Arg z α ) ( z ∈ Σ α ) a_{\alpha}(z):=|z|\cos\left(\frac{\operatorname{Arg}z}{\alpha}\right)\qquad(z\in\Sigma_{\alpha}) a α ( z ) := ∣ z ∣ cos ( α Arg z ) ( z ∈ Σ α )
satisfies a α ( ω α ) = Re ( ω ) ∣ ω ∣ α − 1 ( ω ∈ H ) a_{\alpha}(\omega^{\alpha})=\operatorname{Re}(\omega)|\omega|^{\alpha-1}\qquad(\omega\in\mathbb{H}) a α ( ω α ) = Re ( ω ) ∣ ω ∣ α − 1 ( ω ∈ H ) .
It is positive on Σ α \Sigma_{\alpha} Σ α , extends continuously by 0 to ∂ Σ α \partial\Sigma_{\alpha} ∂ Σ α , and coincides with Re z \operatorname{Re}z Re z when α = 1 \alpha=1 α = 1 . Throughout, X X X is a complex Banach space with dual X ∗ X^{*} X ∗ and pairing ⟨ x , x ∗ ⟩ : = x ∗ ( x ) \langle x,x^{*}\rangle:=x^{*}(x) ⟨ x , x ∗ ⟩ := x ∗ ( x ) , H H H is a complex Hilbert space, B ( X ) \mathcal{B}(X) B ( X ) is the algebra of bounded operators, and L L L denotes a closed, densely defined operator with domain D ( L ) D(L) D ( L ) , resolvent set ρ ( L ) \rho(L) ρ ( L ) and spectrum σ ( L ) \sigma(L) σ ( L ) ; bounded operators are included. An operator L L L on H H H is dissipative if Re ⟨ L v , v ⟩ ≤ 0 \operatorname{Re}\langle Lv,v\rangle\le0 Re ⟨ Lv , v ⟩ ≤ 0 for v ∈ D ( L ) v\in D(L) v ∈ D ( L ) , and m-dissipative if moreover ( 0 , ∞ ) ⊂ ρ ( L ) (0,\infty)\subset\rho(L) ( 0 , ∞ ) ⊂ ρ ( L ) . A table of notation is given in Section Appendix A.
The Mittag–Leffler function
For α > 0 \alpha>0 α > 0 the Mittag–Leffler function E α ( z ) = ∑ k ≥ 0 z k Γ ( 1 + α k ) E_{\alpha}(z)=\sum_{k\ge0}\frac{z^{k}}{\Gamma(1+\alpha k)} E α ( z ) = ∑ k ≥ 0 Γ ( 1 + α k ) z k is entire and real on the real axis, so E α ( z ) ‾ = E α ( z ‾ ) \overline{E_{\alpha}(z)}=E_{\alpha}(\overline{z}) E α ( z ) = E α ( z ) ; moreover E 1 ( z ) = e z E_{1}(z)=e^{z} E 1 ( z ) = e z and E 1 2 ( z ) = e z 2 erfc ( − z ) E_{\frac{1}{2}}(z)=e^{z^{2}}\operatorname{erfc}(-z) E 2 1 ( z ) = e z 2 erfc ( − z ) [17 ] . All later estimates rest on its behaviour at infinity: it grows like α − 1 exp ( z 1 / α ) \alpha^{-1}\exp(z^{1/\alpha}) α − 1 exp ( z 1/ α ) inside Σ α \Sigma_{\alpha} Σ α , where z 1 / α z^{1/\alpha} z 1/ α has positive real part, it oscillates with modulus close to 1 α \frac{1}{\alpha} α 1 on the boundary rays, where z 1 α z^{\frac{1}{\alpha}} z α 1 is purely imaginary, and it decays algebraically in the interior of Λ α \Lambda_\alpha Λ α .
For 0 < α < 1 0 < \alpha< 1 0 < α < 1 the function x ↦ E α ( − x ) x \mapsto E_\alpha(-x) x ↦ E α ( − x ) is moreover the Laplace transform of a probability distribution, so that the fractional dynamics is an average of exponential ones. Let Θ α \Theta_\alpha Θ α be a positive random variable with E e − λ Θ α = e − λ α \mathbb{E}e^{-\lambda\Theta_\alpha} = e^{-\lambda^\alpha} E e − λ Θ α = e − λ α (the one-sided α \alpha α -stable law), g α g_\alpha g α its density, and m α ( u ) = α − 1 u − 1 − 1 α g α ( u − 1 α ) m_\alpha(u) = \alpha^{-1}u^{-1-\frac{1}{\alpha}}g_\alpha(u^{-\frac{1}{\alpha}}) m α ( u ) = α − 1 u − 1 − α 1 g α ( u − α 1 ) the density of Y α : = Θ α − α Y_\alpha:= \Theta_\alpha^{-\alpha} Y α := Θ α − α , the Mittag–Leffler distribution.
Since the supremum of ∣ E α ∣ |E_\alpha| ∣ E α ∣ over the stability region is the constant that governs normal operators, recall Theorem D , we record its basic properties here. Let
M α ∗ : = sup z ∈ Λ α ∣ E α ( z ) ∣ , M α ( θ ) : = sup r ≥ 0 ∣ E α ( r e i θ ) ∣ ( α π 2 ≤ θ ≤ π ) . M_\alpha^* := \sup_{z\in\Lambda_\alpha}|E_\alpha(z)|,\qquad M_\alpha(\theta) := \sup_{r\ge0}|E_\alpha(re^{i\theta})| \quad\left(\frac{\alpha\pi}{2}\le\theta\le\pi\right). M α ∗ := z ∈ Λ α sup ∣ E α ( z ) ∣ , M α ( θ ) := r ≥ 0 sup ∣ E α ( r e i θ ) ∣ ( 2 α π ≤ θ ≤ π ) .
For α = 1 \alpha=1 α = 1 we have M 1 ( θ ) = 1 M_1(\theta)=1 M 1 ( θ ) = 1 for π 2 ≤ θ ≤ π \frac{\pi}{2}\le\theta\le\pi 2 π ≤ θ ≤ π and M 1 ∗ = 1 M_1^*=1 M 1 ∗ = 1 .
The next lemma collects the facts on E α E_\alpha E α that we use, like the fact that ∣ E α ∣ |E_\alpha| ∣ E α ∣ is largest on the boundary rays of Σ α \Sigma_\alpha Σ α , where it exceeds 1 / α 1/\alpha 1/ α ; this gives the sharp constant for normal and numerical-range generators in § 6 and the transient growth of normal operators.
Lemma 3.1. Let 0 < α < 1 0<\alpha<1 0 < α < 1 . Then M α ∗ = M α ( α π 2 ) < ∞ M_\alpha^*=M_\alpha\left(\frac{\alpha\pi}{2}\right)<\infty M α ∗ = M α ( 2 α π ) < ∞ , the function θ ↦ M α ( θ ) \theta\mapsto M_\alpha(\theta) θ ↦ M α ( θ ) is nonincreasing, M α ( θ ) = 1 M_\alpha(\theta)=1 M α ( θ ) = 1 for π 2 ≤ θ ≤ π \frac{\pi}{2}\le\theta\le\pi 2 π ≤ θ ≤ π , lim θ ↓ α π 2 M α ( θ ) = M α ∗ \lim_{\theta\downarrow\frac{\alpha\pi}{2}}M_\alpha(\theta)=M_\alpha^* lim θ ↓ 2 α π M α ( θ ) = M α ∗ , and M α ∗ > 1 α M_\alpha^*>\frac{1}{\alpha} M α ∗ > α 1 . More generally, for every closed sector Σ ‾ \overline{\Sigma} Σ with vertex 0 0 0 and opening less than 2 π 2\pi 2 π that does not meet Σ α \Sigma_\alpha Σ α , the supremum of ∣ E α ∣ |E_\alpha| ∣ E α ∣ over Σ ‾ \overline{\Sigma} Σ is the larger of the values M α ( ∣ θ ∣ ) M_\alpha(|\theta|) M α ( ∣ θ ∣ ) over the arguments θ ∈ ( − π , π ] \theta\in(-\pi,\pi] θ ∈ ( − π , π ] of its two boundary rays.
Proof. By Theorem Appendix B.1(c), E α E_\alpha E α is bounded on Λ α \Lambda_\alpha Λ α . A closed sector of opening less than 2 π 2\pi 2 π that does not meet Σ α \Sigma_\alpha Σ α is contained in Λ α \Lambda_\alpha Λ α , so E α E_\alpha E α is bounded and continuous on it and holomorphic in its interior. By the Phragmén–Lindelöf principle for bounded holomorphic functions in sectors [6 Chapter VI] , its supremum equals the supremum over the two boundary rays, which is the last claim; by conjugation symmetry the value on the ray of argument θ \theta θ is M α ( ∣ θ ∣ ) M_\alpha(|\theta|) M α ( ∣ θ ∣ ) . Applied to the sectors { θ ≤ ∣ Arg z ∣ ≤ π } ∪ { 0 } \{\theta\le|\operatorname{Arg}z|\le\pi\}\cup\{0\} { θ ≤ ∣ Arg z ∣ ≤ π } ∪ { 0 } , α π 2 ≤ θ < π \frac{\alpha\pi}{2}\le\theta<\pi 2 α π ≤ θ < π , it shows that M α M_\alpha M α is nonincreasing and that M α ∗ = M α ( α π 2 ) M_\alpha^*=M_\alpha\left(\frac{\alpha\pi}{2}\right) M α ∗ = M α ( 2 α π ) . For Re z ≤ 0 \operatorname{Re}z\le0 Re z ≤ 0 the integral ∫ 0 ∞ m α ( u ) e z u d u \int_0^\infty m_\alpha(u)e^{zu}\,du ∫ 0 ∞ m α ( u ) e z u d u converges absolutely, it is holomorphic in { Re z < 0 } \{\operatorname{Re}z<0\} { Re z < 0 } and continuous on { Re z ≤ 0 } \{\operatorname{Re}z\le0\} { Re z ≤ 0 } , and it equals E α E_\alpha E α on ( − ∞ , 0 ] (-\infty,0] ( − ∞ , 0 ] by Theorem Appendix B.1(d); hence it equals E α E_\alpha E α on { Re z ≤ 0 } \{\operatorname{Re}z\le0\} { Re z ≤ 0 } , so ∣ E α ( z ) ∣ ≤ 1 = E α ( 0 ) |E_\alpha(z)|\le1=E_\alpha(0) ∣ E α ( z ) ∣ ≤ 1 = E α ( 0 ) there and M α ( θ ) = 1 M_\alpha(\theta)=1 M α ( θ ) = 1 for θ ≥ π 2 \theta\ge\frac{\pi}{2} θ ≥ 2 π . As a supremum of continuous functions, M α M_\alpha M α is lower semicontinuous, which together with monotonicity gives the limit. Finally, on the ray z = r e i α π 2 z=re^{i\frac{\alpha\pi}{2}} z = r e i 2 α π we have z 1 α = i r 1 α z^{\frac{1}{\alpha}}=ir^{\frac{1}{\alpha}} z α 1 = i r α 1 , and Theorem Appendix B.1(a) gives E α ( z ) = α − 1 e i r 1 α − r − 1 e − i α π 2 Γ ( 1 − α ) + O ( r − 2 ) E_\alpha(z)=\alpha^{-1}e^{ir^{\frac{1}{\alpha}}}-\frac{r^{-1}e^{-i\frac{\alpha\pi}{2}}}{\Gamma(1-\alpha)}+O(r^{-2}) E α ( z ) = α − 1 e i r α 1 − Γ ( 1 − α ) r − 1 e − i 2 α π + O ( r − 2 ) . Along a sequence r k → ∞ r_k\to\infty r k → ∞ with e i r k 1 α = − e − i α π 2 e^{ir_k^{\frac{1}{\alpha}}}=-e^{-i\frac{\alpha\pi}{2}} e i r k α 1 = − e − i 2 α π this yields ∣ E α ( z ) ∣ = α − 1 + r k − 1 Γ ( 1 − α ) + O ( r k − 2 ) |E_\alpha(z)|=\alpha^{-1}+\frac{r_k^{-1}}{\Gamma(1-\alpha)}+O(r_k^{-2}) ∣ E α ( z ) ∣ = α − 1 + Γ ( 1 − α ) r k − 1 + O ( r k − 2 ) , which exceeds α − 1 \alpha^{-1} α − 1 for large k k k .
Numerically, α M α ∗ = 1.326 \alpha M_\alpha^*=1.326 α M α ∗ = 1.326 , 1.245 1.245 1.245 , 1.171 1.171 1.171 , 1.068 1.068 1.068 and 1.034 1.034 1.034 for α = 0.1 \alpha=0.1 α = 0.1 , 0.3 0.3 0.3 , 0.5 0.5 0.5 , 0.8 0.8 0.8 and 0.9 0.9 0.9 ; the maximum is attained at moderate arguments, r ∗ 1 α ≈ 2 r_*^{\frac{1}{\alpha}}\approx2 r ∗ α 1 ≈ 2 to 2.6 2.6 2.6 , and not in the asymptotic regime (Table 1 ).
Table 1: Constants of the theory for selected orders. M α ∗ M_\alpha^* M α ∗ is the sharp bound for normal operators, attained at s ∗ = r ∗ 1 / α s_* = r_*^{1/\alpha} s ∗ = r ∗ 1/ α on the boundary ray; α M α ∗ \alpha M_\alpha^* α M α ∗ is the ratio of the two sides of Theorem A for normal operators with spectrum on ∂ Σ α \partial\Sigma_\alpha ∂ Σ α ; ϱ 2 ( α ) \varrho_2(\alpha) ϱ 2 ( α ) is the Jordan-block limit of Theorem 9.1; κ α , 1 \kappa_{\alpha,1} κ α , 1 and κ α , 1 ′ \kappa'_{\alpha,1} κ α , 1 ′ are the constants of Theorem A for rank one. Further orders, including irrational ones, are listed in Table S1 of the Supplementary Material.
Table 1.
Mittag–Leffler families
We write k β ( t ) : = t β − 1 Γ ( β ) k_\beta(t):=\frac{t^{\beta-1}}{\Gamma(\beta)} k β ( t ) := Γ ( β ) t β − 1 for β > 0 \beta>0 β > 0 and t > 0 t>0 t > 0 . The natural notion of a solution operator for (2) is defined through the Laplace transform, which for bounded L L L is computed termwise from the power series and for unbounded L L L becomes the definition.
Definition 3.2. Let 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 . The operator L L L generates a Mittag–Leffler family if there are a strongly continuous function S : [ 0 , ∞ ) → B ( X ) S : [0,\infty) \to\mathcal{B}(X) S : [ 0 , ∞ ) → B ( X ) and w ≥ 0 w \ge0 w ≥ 0 such that
∫ 0 ∞ e − w t ∥ S ( t ) ∥ d t < ∞ \int_{0}^{\infty} e^{-wt}\lVert S(t)\rVert\,\mathrm{d}t < \infty ∫ 0 ∞ e − w t ∥ S ( t )∥ d t < ∞ and, for every real λ > w \lambda> w λ > w , λ α ∈ ρ ( L ) \lambda^\alpha\in\rho(L) λ α ∈ ρ ( L ) and
∫ 0 ∞ e − λ t S ( t ) x d t = λ α − 1 ( λ α − L ) − 1 x ( x ∈ X ) . (4) \int_{0}^{\infty} e^{-\lambda t}S(t)x\,\mathrm{d}t = \lambda^{\alpha-1}(\lambda^\alpha-L)^{-1}x \qquad(x \in X). \tag*{(4)} ∫ 0 ∞ e − λ t S ( t ) x d t = λ α − 1 ( λ α − L ) − 1 x ( x ∈ X ) . (4) By the uniqueness theorem for the Laplace transform, S S S is determined by L L L , and we write E α ( t α L ) : = S ( t ) E_\alpha(t^\alpha L) := S(t) E α ( t α L ) := S ( t ) .
For α = 1 \alpha= 1 α = 1 these are the generators of C 0 C_0 C 0 -semigroups and E 1 ( t L ) = e t L E_1(tL) = e^{tL} E 1 ( t L ) = e t L . For 0 < α < 1 0 < \alpha< 1 0 < α < 1 , E α ( t α L ) E_\alpha(t^\alpha L) E α ( t α L ) is the resolvent family of the Volterra equation with kernel k α k_\alpha k α , also called the α \alpha α -times resolvent family or solution operator of L L L [16 ] . If x ∈ D ( L ) x \in D(L) x ∈ D ( L ) , then λ α − 1 ( λ α − L ) − 1 x = λ − 1 x + λ − α ⋅ λ α − 1 ( λ α − L ) − 1 L x \lambda^{\alpha-1}(\lambda^\alpha-L)^{-1}x = \lambda^{-1}x+\lambda^{-\alpha}\cdot\lambda^{\alpha-1}(\lambda^\alpha-L)^{-1}Lx λ α − 1 ( λ α − L ) − 1 x = λ − 1 x + λ − α ⋅ λ α − 1 ( λ α − L ) − 1 Lx , and uniqueness of the Laplace transform turns (4) into
E α ( t α L ) x = x + ∫ 0 t k α ( t − s ) E α ( s α L ) L x d s ( t ≥ 0 , x ∈ D ( L ) ) , (5) E_\alpha(t^\alpha L)x=x+\int_0^t k_\alpha(t-s)E_\alpha(s^\alpha L)Lx\,\mathrm{d}s \qquad(t\ge0,\ x\in D(L)), \tag*{(5)} E α ( t α L ) x = x + ∫ 0 t k α ( t − s ) E α ( s α L ) Lx d s ( t ≥ 0 , x ∈ D ( L )) , (5)
the integrated form of (2) ; in particular E α ( 0 ⋅ L ) = I E_\alpha(0\cdot L)=I E α ( 0 ⋅ L ) = I by density.
In the following two lemmas, whose proofs we provide in Section Appendix B for completeness, we summarize main facts on Mittag–Leffler families generated by bounded and sectorial operators.
Lemma 3.3. Let 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 .
(a) If L ∈ B ( X ) L \in\mathcal{B}(X) L ∈ B ( X ) , then L L L generates a Mittag–Leffler family, E α ( t α L ) = ∑ k ≥ 0 t α k L k Γ ( 1 + α k ) E_\alpha(t^\alpha L)=\sum_{k\ge0}\frac{t^{\alpha k}L^k}{\Gamma(1+\alpha k)} E α ( t α L ) = ∑ k ≥ 0 Γ ( 1 + α k ) t α k L k , the map t ↦ E α ( t α L ) t\mapsto E_\alpha(t^\alpha L) t ↦ E α ( t α L ) is norm continuous, ∥ E α ( t α L ) ∥ ≤ E α ( t α ∥ L ∥ ) \lVert E_\alpha(t^\alpha L)\rVert\le E_\alpha(t^\alpha\lVert L\rVert) ∥ E α ( t α L )∥ ≤ E α ( t α ∥ L ∥) , and (4) holds for every ω ∈ C \omega\in\mathbb{C} ω ∈ C with Re ω > ∥ L ∥ 1 / α \operatorname{Re}\omega>\lVert L\rVert^{1/\alpha} Re ω > ∥ L ∥ 1/ α in place of λ \lambda λ , the integral converging absolutely in norm.
(b) (Subordination.) Let 0 < α < 1 0 < \alpha< 1 0 < α < 1 , and let L L L generate a C 0 C_0 C 0 -semigroup with ∥ e s L ∥ ≤ M e ϖ s \lVert e^{sL}\rVert\le M e^{\varpi s} ∥ e s L ∥ ≤ M e ϖ s for some M ≥ 1 M\ge1 M ≥ 1 and ϖ ≥ 0 \varpi\ge0 ϖ ≥ 0 . Then L L L generates a Mittag–Leffler family,
E α ( t α L ) x = ∫ 0 ∞ m α ( u ) e t α u L x d u ( t ≥ 0 , x ∈ X ) , E_\alpha(t^\alpha L)x=\int_0^\infty m_\alpha(u)e^{t^\alpha uL}x\,\mathrm{d}u \qquad(t\ge0,\ x\in X), E α ( t α L ) x = ∫ 0 ∞ m α ( u ) e t α uL x d u ( t ≥ 0 , x ∈ X ) , and
sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ sup s ≥ 0 ∥ e s L ∥ . \sup_{t\ge0}\lVert E_\alpha(t^\alpha L)\rVert\le\sup_{s\ge0}\lVert e^{sL}\rVert. t ≥ 0 sup ∥ E α ( t α L )∥ ≤ s ≥ 0 sup ∥ e s L ∥ . (c) If X = H X=H X = H and L L L is m m m -dissipative, then ∥ E α ( t α L ) ∥ ≤ 1 \lVert E_\alpha(t^\alpha L)\rVert\le1 ∥ E α ( t α L )∥ ≤ 1 for all t ≥ 0 t\ge0 t ≥ 0 .
Part (b) is the subordination principle [4 ] : E α ( t α L ) E_\alpha(t^\alpha L) E α ( t α L ) averages the semigroup over the random time t α Y α t^\alpha Y_\alpha t α Y α . It yields no bound when σ ( L ) \sigma(L) σ ( L ) meets the right half-plane, and it does not apply when L L L generates no C 0 C_0 C 0 -semigroup. Such operators are the main object of this paper, and the next lemma covers the most important class of them.
Lemma 3.4 (Sectorial generators). Let 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 and α π 2 < ψ ≤ π \frac{\alpha\pi}{2}<\psi\le\pi 2 α π < ψ ≤ π . Suppose that Σ ( ψ ) ⊂ ρ ( L ) \Sigma(\psi)\subset\rho(L) Σ ( ψ ) ⊂ ρ ( L ) and
s ψ ′ ( L ) : = sup z ∈ Σ ( ψ ′ ) ∥ z ( z − L ) − 1 ∥ < ∞ for every ψ ′ ∈ ( α π 2 , ψ ) . s_{\psi'}(L):=\sup_{z\in\Sigma(\psi')}\lVert z(z-L)^{-1}\rVert<\infty\qquad\text{for every }\psi'\in\left(\frac{\alpha\pi}{2},\psi\right). s ψ ′ ( L ) := z ∈ Σ ( ψ ′ ) sup ∥ z ( z − L ) − 1 ∥ < ∞ for every ψ ′ ∈ ( 2 α π , ψ ) .
Fix ψ ′ ∈ ( α π 2 , min { ψ , α π } ) \psi' \in\left(\frac{\alpha\pi}{2},\min\{\psi,\alpha\pi\}\right) ψ ′ ∈ ( 2 α π , min { ψ , α π } ) , put φ : = ψ ′ α ∈ ( π 2 , π ) \varphi:= \frac{\psi'}{\alpha} \in\left(\frac{\pi}{2},\pi\right) φ := α ψ ′ ∈ ( 2 π , π ) , and let H α φ H_{\alpha\varphi} H α φ be the Hankel contour consisting of the ray { ϱ e − i φ : ϱ ≥ 1 } \{\varrho e^{-i\varphi}:\varrho\ge1\} { ϱ e − i φ : ϱ ≥ 1 } traversed towards e − i φ e^{-i\varphi} e − i φ , the arc { e i ϑ : ∣ ϑ ∣ ≤ φ } \{e^{i\vartheta}:|\vartheta|\le\varphi\} { e i ϑ : ∣ ϑ ∣ ≤ φ } traversed counterclockwise, and the ray { ϱ e i φ : ϱ ≥ 1 } \{\varrho e^{i\varphi}:\varrho\ge1\} { ϱ e i φ : ϱ ≥ 1 } traversed away from e i φ e^{i\varphi} e i φ . Then L L L generates a Mittag--Leffler family,
E α ( t α L ) = 1 2 π i ∫ H α φ e ω t ω α − 1 ( ω α − L ) − 1 d ω ( t > 0 ) , (6) E_\alpha(t^\alpha L)=\frac{1}{2\pi i}\int_{H_{\alpha\varphi}}e^{\omega t}\omega^{\alpha-1}(\omega^\alpha-L)^{-1}\,d\omega\qquad(t>0), \tag*{(6)} E α ( t α L ) = 2 π i 1 ∫ H α φ e ω t ω α − 1 ( ω α − L ) − 1 d ω ( t > 0 ) , (6)
with absolute convergence in norm, and
sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ c φ s ψ ′ ( L ) , c φ : = 1 2 π ∫ H α φ ∣ e ω ∣ ∣ ω ∣ − 1 ∣ d ω ∣ . \sup_{t\ge0}\left\|E_\alpha(t^\alpha L)\right\|\le c_\varphi s_{\psi'}(L),\qquad c_\varphi:=\frac{1}{2\pi}\int_{H_{\alpha\varphi}}|e^\omega||\omega|^{-1}\,|d\omega|. t ≥ 0 sup ∥ E α ( t α L ) ∥ ≤ c φ s ψ ′ ( L ) , c φ := 2 π 1 ∫ H α φ ∣ e ω ∣∣ ω ∣ − 1 ∣ d ω ∣.
For α = 1 \alpha=1 α = 1 this is the classical construction of bounded analytic semigroups. For α < 1 \alpha<1 α < 1 and ψ ≤ π 2 \psi\le\frac{\pi}{2} ψ ≤ 2 π the operator L L L may have spectrum in the right half-plane, and then it need not generate a C 0 C_0 C 0 -semigroup.
The novel fractional Kreiss constant
Definition 3.5. For a closed operator L L L on X X X with Σ α ⊂ ρ ( L ) \Sigma_\alpha\subset\rho(L) Σ α ⊂ ρ ( L ) , define the fractional Kreiss constant to be
K α ( L ) : = sup ω ∈ H Re ( ω ) ∥ ω α − 1 ( ω α − L ) − 1 ∥ = sup z ∈ Σ α a α ( z ) ∥ ( z − L ) − 1 ∥ . \mathcal{K}_\alpha(L):=\sup_{\omega\in\mathbb{H}}\operatorname{Re}(\omega)\left\|\omega^{\alpha-1}(\omega^\alpha-L)^{-1}\right\|=\sup_{z\in\Sigma_\alpha}a_\alpha(z)\left\|(z-L)^{-1}\right\|. K α ( L ) := ω ∈ H sup Re ( ω ) ω α − 1 ( ω α − L ) − 1 = z ∈ Σ α sup a α ( z ) ( z − L ) − 1 .
If Σ α ⊄ ρ ( L ) \Sigma_\alpha\not\subset\rho(L) Σ α ⊂ ρ ( L ) we put K α ( L ) : = ∞ \mathcal{K}_\alpha(L):=\infty K α ( L ) := ∞ .
The two expressions coincide by the substitution z = ω α z=\omega^\alpha z = ω α and the identity for a α ( ω α ) a_\alpha(\omega^\alpha) a α ( ω α ) . The first one shows that K α ( L ) \mathcal{K}_\alpha(L) K α ( L ) is the classical Kreiss quantity sup Re ω > 0 Re ( ω ) ∥ S ^ ( ω ) ∥ \sup_{\operatorname{Re}\omega>0}\operatorname{Re}(\omega)\|\widehat{S}(\omega)\| sup Re ω > 0 Re ( ω ) ∥ S ( ω ) ∥ of the Laplace transform S ^ \widehat{S} S of the family, and the second one shows how the resolvent of L L L enters: near the boundary rays of Σ α \Sigma_\alpha Σ α the weight vanishes linearly, so eigenvalues on these rays are allowed as long as they are semisimple. First, we will prove some of the properties of the fractional Kreiss constant.
Lemma 3.6. Let X ≠ { 0 } X\ne\{0\} X = { 0 } and let L L L be closed and densely defined.
(a) K α ( L ) ≥ 1 \mathcal{K}_\alpha(L)\ge1 K α ( L ) ≥ 1 .
(b) K α ( L ) ≤ K 1 ( L ) \mathcal{K}_\alpha(L)\le\mathcal{K}_1(L) K α ( L ) ≤ K 1 ( L ) , and the inequality can be strict with K 1 ( L ) = ∞ \mathcal{K}_1(L)=\infty K 1 ( L ) = ∞ .
(c) For z ∈ Σ α z\in\Sigma_\alpha z ∈ Σ α , dist ( z , ∂ Σ α ) ≤ a α ( z ) ≤ α − 1 dist ( z , ∂ Σ α ) \operatorname{dist}(z,\partial\Sigma_\alpha)\le a_\alpha(z)\le\alpha^{-1}\operatorname{dist}(z,\partial\Sigma_\alpha) dist ( z , ∂ Σ α ) ≤ a α ( z ) ≤ α − 1 dist ( z , ∂ Σ α ) .
(d) (Pseudospectral form.) Let σ ε ( L ) : = { z : ∥ ( z − L ) − 1 ∥ > ε − 1 } \sigma_\varepsilon(L):=\{z:\|(z-L)^{-1}\|>\varepsilon^{-1}\} σ ε ( L ) := { z : ∥ ( z − L ) − 1 ∥ > ε − 1 } , with ∥ ( z − L ) − 1 ∥ : = ∞ \|(z-L)^{-1}\|:=\infty ∥ ( z − L ) − 1 ∥ := ∞ on σ ( L ) \sigma(L) σ ( L ) , and a α , ε ( L ) : = sup { a α ( z ) : z ∈ σ ε ( L ) ∩ Σ α } a_{\alpha,\varepsilon}(L):=\sup\{a_\alpha(z):z\in\sigma_\varepsilon(L)\cap\Sigma_\alpha\} a α , ε ( L ) := sup { a α ( z ) : z ∈ σ ε ( L ) ∩ Σ α } , with sup ∅ : = 0 \sup\varnothing:=0 sup ∅ := 0 . Then
K α ( L ) = sup ε > 0 a α , ε ( L ) ε . \mathcal{K}_\alpha(L)=\sup_{\varepsilon>0}\frac{a_{\alpha,\varepsilon}(L)}{\varepsilon}. K α ( L ) = ε > 0 sup ε a α , ε ( L ) . (e) (Scaling.) K α ( c L ) = K α ( L ) \mathcal{K}_\alpha(cL)=\mathcal{K}_\alpha(L) K α ( c L ) = K α ( L ) for every c > 0 c>0 c > 0 .
Proof. (a) We may assume K : = K α ( L ) < ∞ \mathcal{K}:=\mathcal{K}_{\alpha}(L)<\infty K := K α ( L ) < ∞ . Since a α ( λ ) = λ a_{\alpha}(\lambda)=\lambda a α ( λ ) = λ for λ > 0 \lambda>0 λ > 0 , we have ∥ λ ( λ − L ) − 1 ∥ ≤ K \|\lambda(\lambda-L)^{-1}\|\leq\mathcal{K} ∥ λ ( λ − L ) − 1 ∥ ≤ K . For x ∈ D ( L ) x\in D(L) x ∈ D ( L ) , ∥ λ ( λ − L ) − 1 x − x ∥ = ∥ ( λ − L ) − 1 L x ∥ ≤ K ∥ L x ∥ λ → 0 \|\lambda(\lambda-L)^{-1}x-x\|=\|(\lambda-L)^{-1}Lx\|\leq\frac{\mathcal{K}\|Lx\|}{\lambda}\to0 ∥ λ ( λ − L ) − 1 x − x ∥ = ∥ ( λ − L ) − 1 Lx ∥ ≤ λ K ∥ Lx ∥ → 0 as λ → ∞ \lambda\to\infty λ → ∞ . By density and the uniform bound, λ ( λ − L ) − 1 x → x \lambda(\lambda-L)^{-1}x\to x λ ( λ − L ) − 1 x → x for every x x x , whence K ≥ lim inf λ → ∞ ∥ λ ( λ − L ) − 1 ∥ ≥ 1 \mathcal{K}\geq\liminf_{\lambda\to\infty}\|\lambda(\lambda-L)^{-1}\|\geq1 K ≥ lim inf λ → ∞ ∥ λ ( λ − L ) − 1 ∥ ≥ 1 .
(b) If z = ∣ z ∣ e i ϑ z=|z|e^{i\vartheta} z = ∣ z ∣ e i ϑ with ∣ ϑ ∣ < α π 2 |\vartheta|<\frac{\alpha\pi}{2} ∣ ϑ ∣ < 2 α π , then ∣ ϑ ∣ α ≥ ∣ ϑ ∣ \frac{|\vartheta|}{\alpha}\geq|\vartheta| α ∣ ϑ ∣ ≥ ∣ ϑ ∣ , so a α ( z ) ≤ Re z a_{\alpha}(z)\leq\operatorname{Re}z a α ( z ) ≤ Re z , and Σ α ⊂ H \Sigma_{\alpha}\subset H Σ α ⊂ H . For L = e i θ L=e^{i\theta} L = e i θ on X = C X=\mathbb{C} X = C with α π 2 < θ < π 2 \frac{\alpha\pi}{2}<\theta<\frac{\pi}{2} 2 α π < θ < 2 π we have K 1 ( L ) = ∞ \mathcal{K}_{1}(L)=\infty K 1 ( L ) = ∞ , while ∣ z − e i θ ∣ ≥ dist ( z , ∂ Σ α ) |z-e^{i\theta}|\geq\operatorname{dist}(z,\partial\Sigma_{\alpha}) ∣ z − e i θ ∣ ≥ dist ( z , ∂ Σ α ) for z ∈ Σ α z\in\Sigma_{\alpha} z ∈ Σ α , so K α ( L ) ≤ 1 α \mathcal{K}_{\alpha}(L)\leq\frac{1}{\alpha} K α ( L ) ≤ α 1 by (c).
(c) Let δ : = α π 2 − ∣ Arg z ∣ ∈ ( 0 , α π 2 ] \delta:=\frac{\alpha\pi}{2}-|\operatorname{Arg}z|\in\left(0,\frac{\alpha\pi}{2}\right] δ := 2 α π − ∣ Arg z ∣ ∈ ( 0 , 2 α π ] . Then dist ( z , ∂ Σ α ) = ∣ z ∣ sin δ \operatorname{dist}(z,\partial\Sigma_{\alpha})=|z|\sin\delta dist ( z , ∂ Σ α ) = ∣ z ∣ sin δ and a α ( z ) = ∣ z ∣ sin ( δ α ) a_{\alpha}(z)=|z|\sin\left(\frac{\delta}{\alpha}\right) a α ( z ) = ∣ z ∣ sin ( α δ ) . Since δ ≤ α π 2 ≤ π 2 \delta\leq\frac{\alpha\pi}{2}\leq\frac{\pi}{2} δ ≤ 2 α π ≤ 2 π , we have sin δ ≤ sin ( δ α ) \sin\delta\leq\sin\left(\frac{\delta}{\alpha}\right) sin δ ≤ sin ( α δ ) , and the function g ( δ ) : = α − 1 sin δ − sin ( δ α ) g(\delta):=\alpha^{-1}\sin\delta-\sin\left(\frac{\delta}{\alpha}\right) g ( δ ) := α − 1 sin δ − sin ( α δ ) satisfies g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 and g ′ ( δ ) = α − 1 ( cos δ − cos ( δ α ) ) ≥ 0 g'(\delta)=\alpha^{-1}\left(\cos\delta-\cos\left(\frac{\delta}{\alpha}\right)\right)\geq0 g ′ ( δ ) = α − 1 ( cos δ − cos ( α δ ) ) ≥ 0 .
(d) If z ∈ σ ε ( L ) ∩ Σ α z\in\sigma_{\varepsilon}(L)\cap\Sigma_{\alpha} z ∈ σ ε ( L ) ∩ Σ α , then a α ( z ) ε < a α ( z ) ∥ ( z − L ) − 1 ∥ \frac{a_{\alpha}(z)}{\varepsilon}<a_{\alpha}(z)\|(z-L)^{-1}\| ε a α ( z ) < a α ( z ) ∥ ( z − L ) − 1 ∥ . Conversely, if ∥ ( z − L ) − 1 ∥ = ε 0 − 1 \|(z-L)^{-1}\|=\varepsilon_{0}^{-1} ∥ ( z − L ) − 1 ∥ = ε 0 − 1 with z ∈ Σ α z\in\Sigma_{\alpha} z ∈ Σ α , then z ∈ σ ε ( L ) z\in\sigma_{\varepsilon}(L) z ∈ σ ε ( L ) for every ε > ε 0 \varepsilon>\varepsilon_{0} ε > ε 0 , and a α ( z ) ε → a α ( z ) ∥ ( z − L ) − 1 ∥ \frac{a_{\alpha}(z)}{\varepsilon}\to a_{\alpha}(z)\|(z-L)^{-1}\| ε a α ( z ) → a α ( z ) ∥ ( z − L ) − 1 ∥ as ε ↓ ε 0 \varepsilon\downarrow\varepsilon_{0} ε ↓ ε 0 . If z 0 ∈ σ ( L ) ∩ Σ α z_{0}\in\sigma(L)\cap\Sigma_{\alpha} z 0 ∈ σ ( L ) ∩ Σ α , then both sides are infinite.
(e) The map z ↦ c z z\mapsto cz z ↦ cz preserves Σ α \Sigma_{\alpha} Σ α , a α ( c z ) = c a α ( z ) a_{\alpha}(cz)=ca_{\alpha}(z) a α ( cz ) = c a α ( z ) and ∥ ( c z − c L ) − 1 ∥ = c − 1 ∥ ( z − L ) − 1 ∥ \|(cz-cL)^{-1}\|=c^{-1}\|(z-L)^{-1}\| ∥ ( cz − c L ) − 1 ∥ = c − 1 ∥ ( z − L ) − 1 ∥ .
Part (d) relates K α \mathcal{K}_{\alpha} K α to the fractional pseudospectra of Šanca et al. [42 ] : it measures how far the ε \varepsilon ε -pseudospectrum reaches into Σ α \Sigma_{\alpha} Σ α , in the units of a α a_{\alpha} a α , relative to ε \varepsilon ε . Part (e) shows that K α \mathcal{K}_{\alpha} K α is a dimensionless quantity, which is why a uniform bound on K α \mathcal{K}_{\alpha} K α carries no information about the size of L L L ; this observation will be used in § 9 .
Remark 3.7 (Other normalisations). Since Re ω ≤ ∣ ω ∣ \operatorname{Re}\omega\leq|\omega| Re ω ≤ ∣ ω ∣ , we have that the quantity K α + ( L ) : = sup ω ∈ H ( Re ω ) α ∥ ( ω α − L ) − 1 ∥ \mathcal{K}_{\alpha}^{+}(L):=\sup_{\omega\in H}(\operatorname{Re}\omega)^{\alpha}\|(\omega^{\alpha}-L)^{-1}\| K α + ( L ) := sup ω ∈ H ( Re ω ) α ∥ ( ω α − L ) − 1 ∥ dominates K α ( L ) \mathcal{K}_{\alpha}(L) K α ( L ) , but it is not equivalent to boundedness of E α ( t α L ) E_{\alpha}(t^{\alpha}L) E α ( t α L ) . For 0 < α < 1 0<\alpha<1 0 < α < 1 and the scalar operator L = e i α π 2 L=e^{i\frac{\alpha\pi}{2}} L = e i 2 α π , sup t ∣ E α ( t α L ) ∣ = M α ∗ \sup_{t}|E_{\alpha}(t^{\alpha}L)|=M_{\alpha}^{*} sup t ∣ E α ( t α L ) ∣ = M α ∗ by Theorem 3.1 and K α ( L ) ≤ 1 α \mathcal{K}_{\alpha}(L)\leq\frac{1}{\alpha} K α ( L ) ≤ α 1 by Theorem 3.6(c), while the points z = e i ( α π 2 − δ ) = ω α z=e^{i(\frac{\alpha\pi}{2}-\delta)}=\omega^{\alpha} z = e i ( 2 α π − δ ) = ω α , ω = e i ( π 2 − δ α ) \omega=e^{i(\frac{\pi}{2}-\frac{\delta}{\alpha})} ω = e i ( 2 π − α δ ) , give K α + ( L ) ≥ sin α ( δ α ) δ → ∞ \mathcal{K}_{\alpha}^{+}(L)\geq\frac{\sin^{\alpha}(\frac{\delta}{\alpha})}{\delta}\to\infty K α + ( L ) ≥ δ s i n α ( α δ ) → ∞ as δ ↓ 0 \delta\downarrow0 δ ↓ 0 . For the weight Re ω \operatorname{Re}\omega Re ω of Theorem 3.5, by contrast, finiteness is necessary for boundedness of the family (Theorem 5.1 ).
Slow components
We say that a projection P ∈ B ( X ) P\in\mathcal{B}(X) P ∈ B ( X ) commutes with L L L if P D ( L ) ⊂ D ( L ) PD(L)\subset D(L) P D ( L ) ⊂ D ( L ) and P L x = L P x PLx=LPx P Lx = L P x for x ∈ D ( L ) x\in D(L) x ∈ D ( L ) . The part of L L L in P X PX P X is the restriction of L L L to D ( L ) ∩ P X D(L)\cap PX D ( L ) ∩ P X ; it is closed and densely defined in P X PX P X , because P D ( L ) ⊂ D ( L ) ∩ P X PD(L)\subset D(L)\cap PX P D ( L ) ⊂ D ( L ) ∩ P X and L x = L P x = P L x Lx=LPx=PLx Lx = L P x = P Lx for x ∈ D ( L ) ∩ P X x\in D(L)\cap PX x ∈ D ( L ) ∩ P X .
The idea behind the following definition is that the transient growth of many generators is produced by finitely many modes, while the remaining dynamics is bounded for a structural reason. The word slow refers to the role of these modes in the transient rather than to the size of their eigenvalues: a slow component collects the finitely many modes that can produce growth, and its complement carries a bounded family.
Definition 3.8. Let L L L be closed and densely defined on X X X . A slow component of rank m ≥ 1 m\geq1 m ≥ 1 is a projection P ∈ B ( X ) P\in\mathcal{B}(X) P ∈ B ( X ) of rank m m m that commutes with L L L and such that the part L Q L_{Q} L Q of L L L in Q X QX QX , Q : = I − P Q:=I-P Q := I − P , generates a Mittag–Leffler family on Q X QX QX with sup t ≥ 0 ∥ E α ( t α L Q ) ∥ < ∞ \sup_{t\geq0}\|E_{\alpha}(t^{\alpha}L_{Q})\|<\infty sup t ≥ 0 ∥ E α ( t α L Q ) ∥ < ∞ . We call L P L_{P} L P , the part of L L L in P X PX P X , the slow block , and
K α ( L ; P ) : = sup z ∈ Σ α a α ( z ) ∥ ( z − L P ) − 1 P ∥ \mathcal{K}_{\alpha}(L;P):=\sup_{z\in\Sigma_{\alpha}}a_{\alpha}(z)\|(z-L_{P})^{-1}P\| K α ( L ; P ) := z ∈ Σ α sup a α ( z ) ∥ ( z − L P ) − 1 P ∥
the deflated Kreiss constant , with K α ( L ; P ) : = ∞ \mathcal{K}_{\alpha}(L;P):=\infty K α ( L ; P ) := ∞ if σ ( L P ) ∩ Σ α ≠ ∅ \sigma(L_{P})\cap\Sigma_{\alpha}\ne\emptyset σ ( L P ) ∩ Σ α = ∅ ; the norm is that of B ( X ) \mathcal{B}(X) B ( X ) . By Theorem 3.9(a) below, ( z − L P ) − 1 P = ( z − L ) − 1 P (z-L_{P})^{-1}P=(z-L)^{-1}P ( z − L P ) − 1 P = ( z − L ) − 1 P for z ∈ ρ ( L ) z\in\rho(L) z ∈ ρ ( L ) .
The typical candidate for a slow component is the Riesz projection P = ( 2 π i ) − 1 ∮ Γ ( z − L ) − 1 d z P=(2\pi i)^{-1}\oint_{\Gamma}(z-L)^{-1}\,dz P = ( 2 π i ) − 1 ∮ Γ ( z − L ) − 1 d z onto finitely many isolated eigenvalues of finite algebraic multiplicity, enclosed by a positively oriented contour Γ \Gamma Γ that encloses no other point of σ ( L ) \sigma(L) σ ( L ) ; then P P P has finite rank, P X ⊂ D ( L ) PX\subset D(L) P X ⊂ D ( L ) , P P P commutes with L L L , σ ( L P ) \sigma(L_{P}) σ ( L P ) consists of the enclosed eigenvalues and σ ( L Q ) \sigma(L_{Q}) σ ( L Q ) is the rest of σ ( L ) \sigma(L) σ ( L ) [20 Theorem III.6.17] . It is a slow component as soon as L Q L_{Q} L Q generates a bounded family, for instance by Theorem 3.4 or, on Hilbert spaces, by the numerical-range criterion of § 6 .
Lemma 3.9. Let P P P be a projection of finite rank m ≥ 1 m\ge1 m ≥ 1 commuting with L L L , and let Q : = I − P Q:=I-P Q := I − P .
(a) P X ⊂ D ( L ) PX\subset D(L) P X ⊂ D ( L ) , L P ∈ B ( P X ) L_{P}\in\mathcal{B}(PX) L P ∈ B ( P X ) , and σ ( L ) = σ ( L P ) ∪ σ ( L Q ) \sigma(L)=\sigma(L_{P})\cup\sigma(L_{Q}) σ ( L ) = σ ( L P ) ∪ σ ( L Q ) ; for z ∈ ρ ( L ) z\in\rho(L) z ∈ ρ ( L ) , ( z − L ) − 1 P = ( z − L P ) − 1 P (z-L)^{-1}P=(z-L_{P})^{-1}P ( z − L ) − 1 P = ( z − L P ) − 1 P and ( z − L ) − 1 Q = ( z − L Q ) − 1 Q (z-L)^{-1}Q=(z-L_{Q})^{-1}Q ( z − L ) − 1 Q = ( z − L Q ) − 1 Q .
(b) If P = U V P=UV P = U V with U : C m → X U:\mathbb{C}^{m}\to X U : C m → X injective, V : X → C m V:X\to\mathbb{C}^{m} V : X → C m surjective and V U = I m VU=I_{m} V U = I m , and if Λ ∈ C m × m \Lambda\in\mathbb{C}^{m\times m} Λ ∈ C m × m is the matrix of L P L_{P} L P in this factorisation, then ( z − L P ) − 1 P = U ( z − Λ ) − 1 V (z-L_{P})^{-1}P=U(z-\Lambda)^{-1}V ( z − L P ) − 1 P = U ( z − Λ ) − 1 V for z ∈ ρ ( L P ) z\in\rho(L_{P}) z ∈ ρ ( L P ) .
(c) max { ∥ P ∥ , K α ( L P ) } ≤ K α ( L ; P ) ≤ ∥ P ∥ K α ( L P ) \max\{\lVert P\rVert,\mathcal{K}_{\alpha}(L_{P})\}\le\mathcal{K}_{\alpha}(L;P)\le\lVert P\rVert\mathcal{K}_{\alpha}(L_{P}) max {∥ P ∥ , K α ( L P )} ≤ K α ( L ; P ) ≤ ∥ P ∥ K α ( L P ) , and if Σ α ⊂ ρ ( L ) \Sigma_{\alpha}\subset\rho(L) Σ α ⊂ ρ ( L ) , then K α ( L P ) ≤ K α ( L ) \mathcal{K}_{\alpha}(L_{P})\le\mathcal{K}_{\alpha}(L) K α ( L P ) ≤ K α ( L ) . If X = H X=H X = H and P P P is an orthogonal projection, then K α ( L ; P ) = K α ( L P ) \mathcal{K}_{\alpha}(L;P)=\mathcal{K}_{\alpha}(L_{P}) K α ( L ; P ) = K α ( L P ) .
Proof. (a) Since D ( L ) D(L) D ( L ) is dense, P D ( L ) PD(L) P D ( L ) is a dense subspace of the finite-dimensional space P X PX P X , hence P X = P D ( L ) ⊂ D ( L ) PX=PD(L)\subset D(L) P X = P D ( L ) ⊂ D ( L ) , and L y = L P y = P L y ∈ P X Ly=L_{P}y=PLy\in PX L y = L P y = P L y ∈ P X for y ∈ P X y\in PX y ∈ P X . If z ∈ ρ ( L P ) ∩ ρ ( L Q ) z\in\rho(L_{P})\cap\rho(L_{Q}) z ∈ ρ ( L P ) ∩ ρ ( L Q ) , then R : = ( z − L P ) − 1 P + ( z − L Q ) − 1 Q R:=(z-L_{P})^{-1}P+(z-L_{Q})^{-1}Q R := ( z − L P ) − 1 P + ( z − L Q ) − 1 Q maps X X X into D ( L ) D(L) D ( L ) and satisfies ( z − L ) R = I (z-L)R=I ( z − L ) R = I and R ( z − L ) x = x R(z-L)x=x R ( z − L ) x = x for x ∈ D ( L ) x\in D(L) x ∈ D ( L ) , so z ∈ ρ ( L ) z\in\rho(L) z ∈ ρ ( L ) . Conversely, applying ( z − L ) − 1 (z-L)^{-1} ( z − L ) − 1 to P ( z − L ) x = ( z − L ) P x P(z-L)x=(z-L)Px P ( z − L ) x = ( z − L ) P x shows that ( z − L ) − 1 (z-L)^{-1} ( z − L ) − 1 commutes with P P P and Q Q Q , so its restrictions invert z − L P z-L_{P} z − L P and z − L Q z-L_{Q} z − L Q .
(b) For y ∈ P X y\in PX y ∈ P X we have L P y = U Λ V y L_{P}y=U\Lambda Vy L P y = U Λ V y , and U ( z − Λ ) − 1 V U(z-\Lambda)^{-1}V U ( z − Λ ) − 1 V inverts z − L P z-L_{P} z − L P on P X PX P X .
(c) The lower bound ∥ P ∥ \lVert P\rVert ∥ P ∥ holds because λ ( λ − L P ) − 1 P → P \lambda(\lambda-L_{P})^{-1}P\to P λ ( λ − L P ) − 1 P → P as λ → ∞ \lambda\to\infty λ → ∞ and a α ( λ ) = λ a_{\alpha}(\lambda)=\lambda a α ( λ ) = λ ; the lower bound K α ( L P ) \mathcal{K}_{\alpha}(L_{P}) K α ( L P ) because y = P y y=Py y = P y for y ∈ P X y\in PX y ∈ P X ; the upper bound because ∥ ( z − L P ) − 1 P ∥ ≤ ∥ ( z − L P ) − 1 ∥ ∥ P ∥ \lVert(z-L_{P})^{-1}P\rVert\le\lVert(z-L_{P})^{-1}\rVert\lVert P\rVert ∥( z − L P ) − 1 P ∥ ≤ ∥( z − L P ) − 1 ∥ ∥ P ∥ ; and K α ( L P ) ≤ K α ( L ) \mathcal{K}_{\alpha}(L_{P})\le\mathcal{K}_{\alpha}(L) K α ( L P ) ≤ K α ( L ) by (a). In the orthogonal case ∥ ( z − L P ) − 1 P ∥ = ∥ ( z − L P ) − 1 ∥ \lVert(z-L_{P})^{-1}P\rVert=\lVert(z-L_{P})^{-1}\rVert ∥( z − L P ) − 1 P ∥ = ∥( z − L P ) − 1 ∥ .
By (b), K α ( L ; P ) \mathcal{K}_{\alpha}(L;P) K α ( L ; P ) is a Kreiss quantity of an operator of rank at most m m m that involves only the slow block and the two factors of the projection, so it can be evaluated at a cost governed by m m m and by the cost of applying U U U and V V V . By (c) it lies between the Kreiss constant of the slow block and the same constant multiplied by the obliqueness ∥ P ∥ \lVert P\rVert ∥ P ∥ of the splitting; in prototype (a), analysed in § 8 and § 11 , the propagator follows K α ( L ; P ) \mathcal{K}_{\alpha}(L;P) K α ( L ; P ) closely although ∥ P ∥ \lVert P\rVert ∥ P ∥ is large.
A Bernstein-type inequality on vertical lines
Before stating the main inequality that is the foundation of the theorem for the fractional Kreiss case we explain where it comes from. Let L L L be bounded with σ ( L ) ∩ Σ α = ∅ \sigma(L)\cap\Sigma_{\alpha}=\emptyset σ ( L ) ∩ Σ α = ∅ , and let x ∈ X x \in X x ∈ X , x ∗ ∈ X ∗ x^* \in X^* x ∗ ∈ X ∗ be fixed. Since the main aim is to find a suitable upper-bound for the norm of the function E α ( t α L ) E_\alpha(t^\alpha L) E α ( t α L ) , it is sensible to consider the scalar function and its Laplace transform
g ( t ) : = ⟨ E α ( t α L ) x , x ∗ ⟩ , g ^ ( ω ) = ω α − 1 ⟨ ( ω α − L ) − 1 x , x ∗ ⟩ , g(t) := \langle E_\alpha(t^\alpha L)x, x^* \rangle, \qquad\widehat{g}(\omega) = \omega^{\alpha-1}\langle(\omega^\alpha-L)^{-1}x,x^*\rangle, g ( t ) := ⟨ E α ( t α L ) x , x ∗ ⟩ , g ( ω ) = ω α − 1 ⟨( ω α − L ) − 1 x , x ∗ ⟩ ,
respectively, which will, by the choice of the fractional Kreiss constant, control ∣ g ^ ∣ |\widehat{g}| ∣ g ∣ on vertical lines:
∣ g ^ ( a + i y ) ∣ ≤ K α ( L ) ∥ x ∥ ∥ x ∗ ∥ a . |\widehat{g}(a+iy)| \leq\frac{\mathcal{K}_\alpha(L)\|x\|\|x^*\|}{a}. ∣ g ( a + i y ) ∣ ≤ a K α ( L ) ∥ x ∥∥ x ∗ ∥ .
To recover g ( t ) g(t) g ( t ) we will invert the Laplace transform along the line Re ω = a \operatorname{Re}\omega=a Re ω = a . The direct inversion integral does not converge absolutely, because g ^ \widehat{g} g decays only like ∣ ω ∣ − 1 |\omega|^{-1} ∣ ω ∣ − 1 , but after one integration by parts it does (see (13) in § 5 ), and it gives
∣ g ( t ) ∣ ≤ e a t 2 π t ∫ R ∣ g ^ ′ ( a + i y ) ∣ d y . |g(t)| \leq\frac{e^{at}}{2\pi t}\int_{\mathbb{R}}|\widehat{g}'(a+iy)|\,dy. ∣ g ( t ) ∣ ≤ 2 π t e a t ∫ R ∣ g ′ ( a + i y ) ∣ d y .
The integral on the right is the length of the curve y ↦ g ^ ( a + i y ) y\mapsto\widehat{g}(a+iy) y ↦ g ( a + i y ) , while the Kreiss constant controls its distance from the origin. The main task is to compare the two, that is, to show that a curve of this special form cannot be much longer than its maximal modulus. If L L L has rank n n n , the function z ↦ z ⟨ ( z − L ) − 1 x , x ∗ ⟩ z\mapsto z\langle(z-L)^{-1}x,x^*\rangle z ↦ z ⟨( z − L ) − 1 x , x ∗ ⟩ is rational of degree at most n n n (Theorem 5.2), so g ^ = h F \widehat{g}=h_F g = h F with
h F ( ω ) : = ω − 1 F ( ω α ) ( ω ∈ H ) h_F(\omega):=\omega^{-1}F(\omega^\alpha)\qquad(\omega\in\mathbb{H}) h F ( ω ) := ω − 1 F ( ω α ) ( ω ∈ H )
and F F F in the class R α , n \mathcal{R}_{\alpha,n} R α , n of nonzero rational functions F = F 1 F 0 F=\frac{F_1}{F_0} F = F 0 F 1 , with coprime polynomials F 1 , F 0 F_1,F_0 F 1 , F 0 , such that
deg F 1 ≤ deg F 0 ≤ n and F 0 ( z ) ≠ 0 for all z ∈ Σ α . \deg F_1\leq\deg F_0\leq n\qquad\text{and}\qquad F_0(z)\ne0\text{ for all }z\in\Sigma_\alpha. deg F 1 ≤ deg F 0 ≤ n and F 0 ( z ) = 0 for all z ∈ Σ α .
Zeros of F 0 F_0 F 0 at 0 0 0 or on ∂ Σ α \partial\Sigma_\alpha ∂ Σ α are allowed. For F ∈ R α , n F\in\mathcal{R}_{\alpha,n} F ∈ R α , n and a > 0 a>0 a > 0 we consider the quotient of length and maximal modulus,
J F ( a ) : = ∫ R ∣ h F ′ ( a + i y ) ∣ d y sup y ∈ R ∣ h F ( a + i y ) ∣ . \mathcal{J}_F(a):=\frac{\int_{\mathbb{R}}|h'_F(a+iy)|\,dy}{\sup_{y\in\mathbb{R}}|h_F(a+iy)|}. J F ( a ) := sup y ∈ R ∣ h F ( a + i y ) ∣ ∫ R ∣ h F ′ ( a + i y ) ∣ d y .
The main aim will be to bound this quotient and to do so, we will prove the following Bernstein-type theorem:
Theorem 4.1 (Bernstein-type inequality). Let 0 < α ≤ 1 0<\alpha\leq1 0 < α ≤ 1 , n ∈ N 0 n\in\mathbb{N}_0 n ∈ N 0 and F ∈ R α , n F\in\mathcal{R}_{\alpha,n} F ∈ R α , n . For every a > 0 a>0 a > 0 the quotient J F ( a ) \mathcal{J}_F(a) J F ( a ) is well defined and
2 ≤ J F ( a ) ≤ π ( 3 + 4 n ( 2 − α ) ) . 2\leq\mathcal{J}_F(a)\leq\pi\bigl(3+4n(2-\alpha)\bigr). 2 ≤ J F ( a ) ≤ π ( 3 + 4 n ( 2 − α ) ) . If F F F is regular at 0 0 0 with F ( 0 ) = 0 F(0)=0 F ( 0 ) = 0 , then J F ( a ) ≤ π ( 1 + 4 n ( 2 − α ) ) \mathcal{J}_F(a)\leq\pi\bigl(1+4n(2-\alpha)\bigr) J F ( a ) ≤ π ( 1 + 4 n ( 2 − α ) ) .
We call it Bernstein-type because it bounds an L 1 L^1 L 1 norm of the derivative by the supremum of the function, with a constant that depends only on a degree. For α = 1 \alpha=1 α = 1 , h F h_F h F is rational of degree at most n + 1 n+1 n + 1 , and Spijker’s lemma [44 , 50 ] , transported from the circle to the line by a Möbius map, which preserves the degree, gives the better constant 2 π ( n + 1 ) 2\pi(n+1) 2 π ( n + 1 ) .
That proof counts the intersections of the curve with straight lines by Bézout’s theorem. However for α < 1 \alpha< 1 α < 1 the curve { ( a + i y ) α : y ∈ R } \{(a+iy)^\alpha: y \in\mathbb{R}\} {( a + i y ) α : y ∈ R } is not algebraic when α \alpha α is irrational, and when α = p q \alpha= \frac{p}{q} α = q p is rational its degree grows with q q q , so such count will not be uniform in α \alpha α . The proof that we will present avoids intersection counts altogether: it bounds the total turning of the tangent of the curve y ↦ h F ( a + i y ) y \mapsto h_F(a+iy) y ↦ h F ( a + i y ) , and this bound is uniform in α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] .
Our proof technique has two ingredients. The first is classical: the length of a plane curve is the average, over all directions, of the total variation of its projections onto lines, and a projection changes monotonicity only when the tangent of the curve is orthogonal to the direction of projection, so the length is controlled by the maximal modulus times the total turning of the tangent. The second ingredient, which is where the fractional order enters, is a geometric bound on that turning, uniform in α \alpha α , which we discuss in detail first.
Throughout, for a continuous function u u u on an open interval I ⊆ R I \subseteq\mathbb{R} I ⊆ R we write:
Var I ( u ) \operatorname{Var}_I(u) Var I ( u ) for its total variation;
N u ( p ) : = # { y ∈ I : u ( y ) = p } ∈ N 0 ∪ { ∞ } N_u(p) := \#\{y \in I : u(y) = p\} \in\mathbb{N}_0 \cup\{\infty\} N u ( p ) := # { y ∈ I : u ( y ) = p } ∈ N 0 ∪ { ∞ } for its counting function;
N u [ c ] ( ϑ ) : = # { y ∈ I : u ( y ) ∈ ϑ + c Z } N_u^{[c]}(\vartheta) := \#\{y \in I : u(y) \in\vartheta+ c\mathbb{Z}\} N u [ c ] ( ϑ ) := # { y ∈ I : u ( y ) ∈ ϑ + c Z } , for c > 0 c > 0 c > 0 , for its counting function modulo c c c .
There is a link between the stated quantities which comes from the Banach’s indicatrix theorem [3 ] , extended from compact intervals to I I I by monotone convergence, and by Tonelli’s theorem applied to N u [ c ] ( ϑ ) = ∑ k ∈ Z N u ( ϑ + k c ) N_u^{[c]}(\vartheta) = \sum_{k\in\mathbb{Z}} N_u(\vartheta+ kc) N u [ c ] ( ϑ ) = ∑ k ∈ Z N u ( ϑ + k c ) ,
Var I ( u ) = ∫ R N u ( p ) d p = ∫ 0 c N u [ c ] ( ϑ ) d ϑ ( c > 0 ) , (7) \operatorname{Var}_I(u) = \int_{\mathbb{R}} N_u(p)\,dp = \int_0^c N_u^{[c]}(\vartheta)\,d\vartheta\qquad(c > 0), \tag*{(7)} Var I ( u ) = ∫ R N u ( p ) d p = ∫ 0 c N u [ c ] ( ϑ ) d ϑ ( c > 0 ) , (7)
and Var I ( u ) = ∫ I ∣ u ′ ∣ \operatorname{Var}_I(u) = \int_I |u'| Var I ( u ) = ∫ I ∣ u ′ ∣ if u ∈ C 1 ( I ) u \in C^1(I) u ∈ C 1 ( I ) . We also use the elementary fact that a C 1 C^1 C 1 function whose derivative has k < ∞ k < \infty k < ∞ zeros takes every value at most k + 1 k+1 k + 1 times, since it is strictly monotone between consecutive critical points.
For α = 1 \alpha= 1 α = 1 the image { 1 + i y : y ∈ R } \{1+iy : y \in\mathbb{R}\} { 1 + i y : y ∈ R } of a vertical line is a straight line, and the argument of 1 + i y − ζ 1+iy-\zeta 1 + i y − ζ seen from any point ζ \zeta ζ off the line changes by exactly π \pi π . For α < 1 \alpha< 1 α < 1 the image
C α : = { ( 1 + i y ) α : y ∈ R } ⊂ Σ α C_\alpha:= \{(1+iy)^\alpha: y \in\mathbb{R}\} \subset\Sigma_\alpha C α := {( 1 + i y ) α : y ∈ R } ⊂ Σ α
bends, and the following lemma shows that a curve whose tangent turns monotonically through an angle Θ \Theta Θ is seen from any external point under an angle of variation at most π + Θ \pi+\Theta π + Θ . This geometric fact replaces the algebraic intersection counts of the case α = 1 \alpha= 1 α = 1 .
Lemma 4.2 (Turning lemma). Let γ ∈ C 1 ( I ; C ) \gamma\in C^1(I;\mathbb{C}) γ ∈ C 1 ( I ; C ) with γ ′ ≠ 0 \gamma' \ne0 γ ′ = 0 on an open interval I I I , and suppose that a continuous lift T T T of arg γ ′ \arg\gamma' arg γ ′ is either constant or strictly monotone, with range contained in an open interval J J J of length Θ ≤ π \Theta\le\pi Θ ≤ π ; when T T T is constant, J J J is replaced by its single value and Θ : = 0 \Theta:= 0 Θ := 0 . Let ζ ∈ C ∖ γ ( I ) \zeta\in\mathbb{C}\setminus\gamma(I) ζ ∈ C ∖ γ ( I ) and let β ζ \beta_\zeta β ζ be a continuous lift of s ↦ arg ( γ ( s ) − ζ ) s \mapsto\arg(\gamma(s)-\zeta) s ↦ arg ( γ ( s ) − ζ ) . Then Var I ( β ζ ) ≤ π + Θ \operatorname{Var}_I(\beta_\zeta) \le\pi+\Theta Var I ( β ζ ) ≤ π + Θ .
Proof. For τ ∈ [ 0 , π ) \tau\in[0,\pi) τ ∈ [ 0 , π ) let ℓ τ : = ζ + R e i τ \ell_\tau:= \zeta+ \mathbb{R}e^{i\tau} ℓ τ := ζ + R e i τ be the line through ζ \zeta ζ in direction e i τ e^{i\tau} e i τ , and consider f τ ( s ) : = Im ( e − i τ ( γ ( s ) − ζ ) ) f_\tau(s) := \operatorname{Im}(e^{-i\tau}(\gamma(s)-\zeta)) f τ ( s ) := Im ( e − i τ ( γ ( s ) − ζ )) , whose zeros are the parameters with γ ( s ) ∈ ℓ τ \gamma(s) \in\ell_\tau γ ( s ) ∈ ℓ τ . Since f τ ′ ( s ) = ∣ γ ′ ( s ) ∣ sin ( T ( s ) − τ ) f_\tau'(s) = |\gamma'(s)|\sin(T(s)-\tau) f τ ′ ( s ) = ∣ γ ′ ( s ) ∣ sin ( T ( s ) − τ ) , the critical points of f τ f_\tau f τ are the parameters with T ( s ) ∈ τ + π Z T(s) \in\tau+ \pi\mathbb{Z} T ( s ) ∈ τ + π Z . If T T T is strictly monotone, there is at most one such parameter for each element of $(\tau+\pi\mathbb{Z})\cap J, and this set has at most one element because J J J is open of length at most π \pi π ; if T T T is constant, γ \gamma γ parametrises a segment, a half-line or a line, which meets ℓ τ \ell_\tau ℓ τ at most once except for the single direction parallel to it. Hence, for almost every τ \tau τ , γ \gamma γ meets ℓ τ \ell_\tau ℓ τ at most twice, and at most once if ( τ + π Z ) ∩ J = ∅ (\tau+\pi\mathbb{Z})\cap J=\varnothing ( τ + π Z ) ∩ J = ∅ ; the set of τ ∈ [ 0 , π ) \tau\in[0,\pi) τ ∈ [ 0 , π ) with ( τ + π Z ) ∩ J ≠ ∅ (\tau+\pi\mathbb{Z})\cap J\ne\varnothing ( τ + π Z ) ∩ J = ∅ has measure at most Θ \Theta Θ . Now arg ( γ ( s ) − ζ ) ∈ τ + 2 π Z \arg(\gamma(s)-\zeta)\in\tau+2\pi\mathbb{Z} arg ( γ ( s ) − ζ ) ∈ τ + 2 π Z exactly when γ ( s ) \gamma(s) γ ( s ) lies on the open half-line ζ + ( 0 , ∞ ) e i τ \zeta+(0,\infty)e^{i\tau} ζ + ( 0 , ∞ ) e i τ . If K ( τ ) K(\tau) K ( τ ) denotes the number of such parameters, then K ( τ ) + K ( τ + π ) K(\tau)+K(\tau+\pi) K ( τ ) + K ( τ + π ) is the number of intersections of γ \gamma γ with ℓ τ \ell_\tau ℓ τ , because γ ( s ) ≠ ζ \gamma(s)\ne\zeta γ ( s ) = ζ . By (7) with c = 2 π c=2\pi c = 2 π ,
Var I ( β ζ ) = ∫ 0 2 π K ( τ ) d τ = ∫ 0 π ( K ( τ ) + K ( τ + π ) ) d τ ≤ 2 Θ + ( π − Θ ) = π + Θ . \operatorname{Var}_{I}(\beta_\zeta)=\int_{0}^{2\pi}K(\tau)\,d\tau=\int_{0}^{\pi}\left(K(\tau)+K(\tau+\pi)\right)\,d\tau\le2\Theta+(\pi-\Theta)=\pi+\Theta. Var I ( β ζ ) = ∫ 0 2 π K ( τ ) d τ = ∫ 0 π ( K ( τ ) + K ( τ + π ) ) d τ ≤ 2Θ + ( π − Θ ) = π + Θ.
We apply the lemma to C α \mathcal{C}_\alpha C α . With y = tan φ y=\tan\varphi y = tan φ , ∣ φ ∣ < π 2 |\varphi|<\frac{\pi}{2} ∣ φ ∣ < 2 π , the curve is parametrised by w ( φ ) = ( 1 + i tan φ ) α = sec α φ e i α φ w(\varphi)=(1+i\tan\varphi)^\alpha=\sec^\alpha\varphi e^{i\alpha\varphi} w ( φ ) = ( 1 + i tan φ ) α = sec α φ e i α φ , and after some algebra w ′ ( φ ) = α sec α + 1 φ e i T ( φ ) w'(\varphi)=\alpha\sec^{\alpha+1}\varphi e^{iT(\varphi)} w ′ ( φ ) = α sec α + 1 φ e i T ( φ ) with T ( φ ) = π 2 − ( 1 − α ) φ T(\varphi)=\frac{\pi}{2}-(1-\alpha)\varphi T ( φ ) = 2 π − ( 1 − α ) φ . This lift is constant for α = 1 \alpha=1 α = 1 and strictly decreasing for α < 1 \alpha<1 α < 1 , with range ( α π 2 , π − α π 2 ) \left(\frac{\alpha\pi}{2},\pi-\frac{\alpha\pi}{2}\right) ( 2 α π , π − 2 α π ) of length ( 1 − α ) π (1-\alpha)\pi ( 1 − α ) π . Consequently, for every ζ ∉ C α \zeta\notin\mathcal{C}_\alpha ζ ∈ / C α and every continuous lift β ζ \beta_\zeta β ζ of y ↦ arg ( ( 1 + i y ) α − ζ ) y\mapsto\arg((1+iy)^\alpha-\zeta) y ↦ arg (( 1 + i y ) α − ζ ) ,
Var R ( β ζ ) ≤ ( 2 − α ) π . (8) \operatorname{Var}_{\mathbb{R}}(\beta_\zeta)\le(2-\alpha)\pi. \tag*{(8)} Var R ( β ζ ) ≤ ( 2 − α ) π . (8)
The bound is attained: seen from a point ζ > 1 \zeta>1 ζ > 1 inside the convex region bounded by C α \mathcal{C}_\alpha C α , for instance ζ = 2 \zeta=2 ζ = 2 , the argument of ( 1 + i y ) α − ζ (1+iy)^\alpha-\zeta ( 1 + i y ) α − ζ decreases monotonically from − α π 2 -\frac{\alpha\pi}{2} − 2 α π through ± π \pm\pi ± π to α π 2 \frac{\alpha\pi}{2} 2 α π , whereas seen from the origin it only sweeps the angle α π \alpha\pi α π between the asymptotic directions of the curve.
Proof of Theorem 4.1 . Fix F = F 1 F 0 ∈ R α , n F=\frac{F_1}{F_0}\in\mathcal{R}_{\alpha,n} F = F 0 F 1 ∈ R α , n , write q : = deg F 0 ≤ n q:=\deg F_0\le n q := deg F 0 ≤ n and h : = h F h:=h_F h := h F , and put D F ( z ) : = − F ( z ) + α z F ′ ( z ) D_F(z):=-F(z)+\alpha zF'(z) D F ( z ) := − F ( z ) + α z F ′ ( z ) . The proof is organized in four components focus on: (i) structure of h h h , (ii) length controlled by turning, (iii) turning of the derivative h ′ h' h ′ and (iv) lines with critical points.
(i) Since F F F is holomorphic on Σ α = { ω α : ω ∈ H } \Sigma_\alpha=\{\omega^\alpha:\omega\in\mathbb{H}\} Σ α = { ω α : ω ∈ H } and ω − 1 ω α − 1 = ω − 2 ω α \omega^{-1}\omega^{\alpha-1}=\omega^{-2}\omega^\alpha ω − 1 ω α − 1 = ω − 2 ω α , the product rule gives
h ′ ( ω ) = ω − 2 D F ( ω α ) ( ω ∈ H ) . (9) h'(\omega)=\omega^{-2}D_F(\omega^\alpha)\qquad(\omega\in\mathbb{H}). \tag*{(9)} h ′ ( ω ) = ω − 2 D F ( ω α ) ( ω ∈ H ) . (9)
Moreover D F = G F 0 2 D_F=\frac{G}{F_0^2} D F = F 0 2 G with G : = − F 1 F 0 + α z ( F 1 ′ F 0 − F 1 F 0 ′ ) G:=-F_1F_0+\alpha z(F_1'F_0-F_1F_0') G := − F 1 F 0 + α z ( F 1 ′ F 0 − F 1 F 0 ′ ) of degree at most 2 q 2q 2 q , because deg F 1 ≤ q \deg F_1\le q deg F 1 ≤ q ; after cancelling common factors, D F = G N G D D_F=\frac{G_N}{G_D} D F = G D G N with d N : = deg G N ≤ 2 q d_N:=\deg G_N\le2q d N := deg G N ≤ 2 q , d D : = deg G D ≤ 2 q d_D:=\deg G_D\le2q d D := deg G D ≤ 2 q and G D G_D G D dividing F 0 2 F_0^2 F 0 2 , so D F D_F D F has no poles in Σ α \Sigma_\alpha Σ α . The function h h h does not vanish identically on any vertical line, since F ≢ 0 F\not\equiv0 F ≡ 0 on the open set Σ α \Sigma_\alpha Σ α , and D F ≠ 0 D_F\ne0 D F = 0 , since otherwise h h h would be constant and h ( t ) = t − 1 F ( t α ) → 0 h(t)=t^{-1}F(t^\alpha)\to0 h ( t ) = t − 1 F ( t α ) → 0 as t → + ∞ t\to+\infty t → + ∞ . As F F F and D F D_F D F are bounded near infinity, there are C , y ∗ > 0 C,y_*>0 C , y ∗ > 0 with ∣ h ( a + i y ) ∣ ≤ C ∣ y ∣ |h(a+iy)|\le\frac{C}{|y|} ∣ h ( a + i y ) ∣ ≤ ∣ y ∣ C and ∣ h ′ ( a + i y ) ∣ ≤ C ∣ y ∣ 2 |h'(a+iy)|\le\frac{C}{|y|^2} ∣ h ′ ( a + i y ) ∣ ≤ ∣ y ∣ 2 C for all a > 0 a>0 a > 0 and ∣ y ∣ ≥ y ∗ |y|\ge y_* ∣ y ∣ ≥ y ∗ . Hence the length ℓ a : = ∫ R ∣ h ′ ( a + i y ) ∣ d y \ell_a:=\int_{\mathbb{R}}|h'(a+iy)|\,dy ℓ a := ∫ R ∣ h ′ ( a + i y ) ∣ d y and the maximal modulus h a ∗ : = sup y ∣ h ( a + i y ) ∣ h_a^*:=\sup_y|h(a+iy)| h a ∗ := sup y ∣ h ( a + i y ) ∣ are finite and positive, and ℓ a ≥ 2 h a ∗ \ell_a\ge2h_a^* ℓ a ≥ 2 h a ∗ because the curve y ↦ h ( a + i y ) y\mapsto h(a+iy) y ↦ h ( a + i y ) starts and ends at 0 0 0 ; this is the lower bound of the theorem. Finally, with F a ( z ) : = F ( a α z ) ∈ R α , n F_a(z):=F(a^\alpha z)\in\mathcal{R}_{\alpha,n} F a ( z ) := F ( a α z ) ∈ R α , n we have h F a ( a ω ) = a − 1 h F a ( ω ) h_{F_a}(a\omega)=a^{-1}h_{F_a}(\omega) h F a ( aω ) = a − 1 h F a ( ω ) , so J F ( a ) = J F a ( 1 ) \mathcal{J}_F(a)=\mathcal{J}_{F_a}(1) J F ( a ) = J F a ( 1 ) and F a ( 0 ) = F ( 0 ) F_a(0)=F(0) F a ( 0 ) = F ( 0 ) , and it suffices to consider a = 1 a=1 a = 1 .
(ii) Suppose that h ′ h' h ′ has no zero on { Re ω = 1 } \{\operatorname{Re}\omega=1\} { Re ω = 1 } , and let η \eta η be a continuous lift of y ↦ arg h ′ ( 1 + i y ) y\mapsto\arg h'(1+iy) y ↦ arg h ′ ( 1 + i y ) . We claim that
ℓ 1 ≤ h 1 ∗ ( π + Var R ( η ) ) . (10) \ell_1\le h_1^*(\pi+\operatorname{Var}_{\mathbb{R}}(\eta)). \tag*{(10)} ℓ 1 ≤ h 1 ∗ ( π + Var R ( η )) . (10)
Put γ ( y ) : = h ( 1 + i y ) \gamma(y):=h(1+iy) γ ( y ) := h ( 1 + i y ) and u ϑ : = Re ( e − i ϑ γ ) u_{\vartheta}:=\operatorname{Re}(e^{-i\vartheta}\gamma) u ϑ := Re ( e − i ϑ γ ) for ϑ ∈ [ 0 , π ) \vartheta\in[0,\pi) ϑ ∈ [ 0 , π ) . Since ∫ 0 π ∣ cos ( φ − ϑ ) ∣ d ϑ = 2 \int_0^\pi|\cos(\varphi-\vartheta)|\,d\vartheta=2 ∫ 0 π ∣ cos ( φ − ϑ ) ∣ d ϑ = 2 for every φ \varphi φ , Tonelli’s theorem gives the Cauchy–Crofton formula 2 ℓ 1 = ∫ 0 π Var R ( u ϑ ) d ϑ 2\ell_1=\int_0^\pi\operatorname{Var}_{\mathbb{R}}(u_{\vartheta})\,d\vartheta 2 ℓ 1 = ∫ 0 π Var R ( u ϑ ) d ϑ . A direct computation gives u ϑ ′ ( y ) = − ∣ h ′ ( 1 + i y ) ∣ sin ( η ( y ) − ϑ ) u_{\vartheta}'(y)=-|h'(1+iy)|\sin(\eta(y)-\vartheta) u ϑ ′ ( y ) = − ∣ h ′ ( 1 + i y ) ∣ sin ( η ( y ) − ϑ ) , so the critical points of u ϑ u_{\vartheta} u ϑ are the solutions of η ( y ) ∈ ϑ + π Z \eta(y)\in\vartheta+\pi\mathbb{Z} η ( y ) ∈ ϑ + π Z , and there are N η [ π ] ( ϑ ) N_{\eta}^{[\pi]}(\vartheta) N η [ π ] ( ϑ ) of them. Since ∣ u ϑ ∣ ≤ h 1 ∗ |u_{\vartheta}|\le h_1^* ∣ u ϑ ∣ ≤ h 1 ∗ , (7) and the elementary counting fact give Var R ( u ϑ ) = ∫ − h 1 ∗ h 1 ∗ N u ϑ ( p ) d p ≤ 2 h 1 ∗ ( N η [ π ] ( ϑ ) + 1 ) \operatorname{Var}_{\mathbb{R}}(u_{\vartheta})=\int_{-h_1^*}^{h_1^*}N_{u_{\vartheta}}(p)\,dp\le2h_1^*(N_{\eta}^{[\pi]}(\vartheta)+1) Var R ( u ϑ ) = ∫ − h 1 ∗ h 1 ∗ N u ϑ ( p ) d p ≤ 2 h 1 ∗ ( N η [ π ] ( ϑ ) + 1 ) . Integrating over ϑ \vartheta ϑ and using (7) with c = π c=\pi c = π gives 2 ℓ 1 ≤ 2 h 1 ∗ ( π + Var R ( η ) ) 2\ell_1\le2h_1^*(\pi+\operatorname{Var}_{\mathbb{R}}(\eta)) 2 ℓ 1 ≤ 2 h 1 ∗ ( π + Var R ( η )) .
(iii) Under the assumption of Step 2, write D F ( z ) = c 0 ∏ i ( z − a i ) k i ∏ j ( z − b j ) − l j D_F(z)=c_0\prod_i(z-a_i)^{k_i}\prod_j(z-b_j)^{-l_j} D F ( z ) = c 0 ∏ i ( z − a i ) k i ∏ j ( z − b j ) − l j with c 0 ≠ 0 c_0\ne0 c 0 = 0 , distinct zeros a i a_i a i of multiplicities k i k_i k i and distinct poles b j b_j b j of orders l j l_j l j , so that ∑ i k i = d N \sum_i k_i=d_N ∑ i k i = d N and ∑ j l j = d D \sum_j l_j=d_D ∑ j l j = d D . The poles lie outside Σ α ⊃ C α \Sigma_\alpha\supset C_\alpha Σ α ⊃ C α by Step 1, and a zero on C α C_\alpha C α would produce a zero of h ′ h' h ′ on the line by (9) ; so no factor vanishes on C α C_\alpha C α . By (9) ,
η ( y ) : = Arg c 0 − 2 arctan y + ∑ i k i β a i ( y ) − ∑ j l j β b j ( y ) \eta(y):=\operatorname{Arg}c_0-2\arctan y+\sum_i k_i\beta_{a_i}(y)-\sum_j l_j\beta_{b_j}(y) η ( y ) := Arg c 0 − 2 arctan y + i ∑ k i β a i ( y ) − j ∑ l j β b j ( y )
is a continuous lift of arg h ′ ( 1 + i y ) \arg h'(1+iy) arg h ′ ( 1 + i y ) , with β ζ \beta_\zeta β ζ as in (8) . Subadditivity of the variation, Var R ( arctan ) = π \operatorname{Var}_{\mathbb{R}}(\arctan)=\pi Var R ( arctan ) = π and (8) give Var R ( η ) ≤ 2 π + ( d N + d D ) ( 2 − α ) π ≤ 2 π + 4 q ( 2 − α ) π \operatorname{Var}_{\mathbb{R}}(\eta)\le2\pi+(d_N+d_D)(2-\alpha)\pi\le2\pi+4q(2-\alpha)\pi Var R ( η ) ≤ 2 π + ( d N + d D ) ( 2 − α ) π ≤ 2 π + 4 q ( 2 − α ) π , and (10) gives J F ( 1 ) ≤ π ( 3 + 4 q ( 2 − α ) ) \mathcal{J}_F(1)\le\pi(3+4q(2-\alpha)) J F ( 1 ) ≤ π ( 3 + 4 q ( 2 − α )) .
If F ( 0 ) = 0 F(0)=0 F ( 0 ) = 0 , then F 0 ( 0 ) ≠ 0 F_0(0)\ne0 F 0 ( 0 ) = 0 and F 1 ( 0 ) = 0 F_1(0)=0 F 1 ( 0 ) = 0 , so F = z F ~ 1 F 0 F=z\frac{\widetilde F_1}{F_0} F = z F 0 F 1 with deg F ~ 1 ≤ q − 1 \deg\widetilde F_1\le q-1 deg F 1 ≤ q − 1 , and after some algebra D F ( z ) = z D ~ ( z ) D_F(z)=z\widetilde D(z) D F ( z ) = z D ( z ) with D ~ = G ~ F 0 2 \widetilde D=\frac{\widetilde G}{F_0^2} D = F 0 2 G , G ~ : = − ( 1 − α ) F ~ 1 F 0 + α z ( F ~ 1 ′ F 0 − F ~ 1 F 0 ′ ) \widetilde G:=-(1-\alpha)\widetilde F_1F_0+\alpha z(\widetilde F_1'F_0-\widetilde F_1F_0') G := − ( 1 − α ) F 1 F 0 + α z ( F 1 ′ F 0 − F 1 F 0 ′ ) of degree at most 2 q − 1 2q-1 2 q − 1 . Then h ′ ( ω ) = ω α − 2 D ~ ( ω α ) h'(\omega)=\omega^{\alpha-2}\widetilde D(\omega^\alpha) h ′ ( ω ) = ω α − 2 D ( ω α ) , the prefactor contributes ( 2 − α ) π (2-\alpha)\pi ( 2 − α ) π instead of 2 π 2\pi 2 π , the factors contribute at most ( 4 q − 1 ) ( 2 − α ) π (4q-1)(2-\alpha)\pi ( 4 q − 1 ) ( 2 − α ) π , and Var R ( η ) ≤ 4 q ( 2 − α ) π \operatorname{Var}_{\mathbb{R}}(\eta)\le4q(2-\alpha)\pi Var R ( η ) ≤ 4 q ( 2 − α ) π , so that J F ( 1 ) ≤ π ( 1 + 4 q ( 2 − α ) ) \mathcal{J}_F(1)\le\pi(1+4q(2-\alpha)) J F ( 1 ) ≤ π ( 1 + 4 q ( 2 − α )) .
(iv) By (9) , h ′ ( ω ) = 0 h'(\omega)=0 h ′ ( ω ) = 0 if and only if G N ( ω α ) = 0 G_N(\omega^\alpha)=0 G N ( ω α ) = 0 , and since ω ↦ ω α \omega\mapsto\omega^\alpha ω ↦ ω α is injective on H \mathbb{H} H , h ′ h' h ′ has at most 2 n 2n 2 n zeros in H \mathbb{H} H . By Step 1, J F ( a ) = J F a ( 1 ) \mathcal{J}_F(a)=\mathcal{J}_{F_a}(1) J F ( a ) = J F a ( 1 ) , so the bounds of Step 3 hold for all but at most 2 n 2n 2 n values of a a a . It remains to see that a ↦ J F ( a ) a\mapsto\mathcal{J}_F(a) a ↦ J F ( a ) is continuous. On K : = { a + i y : a ∈ [ a 1 , a 2 ] , ∣ y ∣ ≤ y ∗ } K:=\{a+iy:a\in[a_1,a_2],\ |y|\le y_*\} K := { a + i y : a ∈ [ a 1 , a 2 ] , ∣ y ∣ ≤ y ∗ } , 0 < a 1 < a 2 0<a_1<a_2 0 < a 1 < a 2 , the functions h h h and h ′ h' h ′ are jointly continuous and bounded, because the image of K K K under ω ↦ ω α \omega\mapsto\omega^\alpha ω ↦ ω α is a compact subset of Σ α \Sigma_\alpha Σ α and poles of F F F on ∂ Σ α \partial\Sigma_\alpha ∂ Σ α stay at positive distance from it. Dominated convergence with the majorant C ′ 1 { ∣ y ∣ ≤ y ∗ } + C ∣ y ∣ − 2 1 { ∣ y ∣ > y ∗ } C'1_{\{|y|\le y_*\}}+C|y|^{-2}1_{\{|y|>y_*\}} C ′ 1 { ∣ y ∣ ≤ y ∗ } + C ∣ y ∣ − 2 1 { ∣ y ∣ > y ∗ } gives continuity of a ↦ ℓ a a\mapsto\ell_a a ↦ ℓ a , and the uniform tail bound on h h h gives continuity of a ↦ h a ∗ > 0 a\mapsto h_a^*>0 a ↦ h a ∗ > 0 . This completes the proof. □
Remark 4.3 (The range 1 < α < 2 1<\alpha<2 1 < α < 2 ). For 1 < α < 2 1<\alpha<2 1 < α < 2 the map ω ↦ ω α \omega\mapsto\omega^\alpha ω ↦ ω α is still injective on H \mathbb{H} H , and the tangent angle T ( φ ) = π 2 + ( α − 1 ) φ T(\varphi)=\frac{\pi}{2}+(\alpha-1)\varphi T ( φ ) = 2 π + ( α − 1 ) φ of C α C_\alpha C α is increasing with range of length ( α − 1 ) π (\alpha-1)\pi ( α − 1 ) π . The turning lemma then gives Var R ( β ζ ) ≤ α π \operatorname{Var}_{\mathbb{R}}(\beta_\zeta)\le\alpha\pi Var R ( β ζ ) ≤ α π for each factor, the prefactor ( 1 + i y ) − 2 (1+iy)^{-2} ( 1 + i y ) − 2 still contributes 2 π 2\pi 2 π , and the same proof yields J F ( a ) ≤ π ( 3 + 4 n α ) \mathcal{J}_F(a)\le\pi(3+4n\alpha) J F ( a ) ≤ π ( 3 + 4 n α ) . We do not pursue this range, because for 1 < α < 2 1<\alpha<2 1 < α < 2 the family E α ( t α L ) E_\alpha(t^\alpha L) E α ( t α L ) solves (2) only with u ′ ( 0 ) = 0 u'(0)=0 u ′ ( 0 ) = 0 , and several results below, among them the subordination formula, do not extend.
The finite-rank core
This section proves Theorem A for bounded operators of finite rank, which is the heart of the matter, while later in § 8 we prove the result for generators with a slow component by reduction to this case. The strategy proceeds in three steps: (i) a lower bound by the Kreiss constant holds for every bounded family; (ii) for a finite-rank operator, every matrix element of the resolvent is a rational function of degree at most the rank, and the operator splits into a slow part acting through a small matrix and a stationary kernel; and finally (iii) the inversion of the Laplace transform along a vertical line, after one integration by parts, turns the upper bound into the Bernstein-type inequality proven in § 4 .
The lower bound
The lower bound in Theorem A holds for every generator of a bounded family and requires no finiteness assumption. Its proof shows that the Laplace transform of a bounded family is the resolvent, transported to H \mathbb{H} H , on the whole half-plane and not only on the real half-line where Theorem 3.2 prescribes it.
Theorem 5.1. Let 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 , let X ≠ { 0 } X \ne\{0\} X = { 0 } , and let L L L be closed, densely defined and generate a Mittag–Leffler family with C L : = sup t ≥ 0 ∥ E α ( t α L ) ∥ < ∞ C_L := \sup_{t \ge0} \lVert E_\alpha(t^\alpha L)\rVert< \infty C L := sup t ≥ 0 ∥ E α ( t α L )∥ < ∞ . Then Σ α ⊂ ρ ( L ) \Sigma_\alpha\subset\rho(L) Σ α ⊂ ρ ( L ) ,
ω α − 1 ( ω α − L ) − 1 x = ∫ 0 ∞ e − ω t E α ( t α L ) x d t ( ω ∈ H , x ∈ X ) , \omega^{\alpha-1}(\omega^\alpha-L)^{-1}x=\int_0^\infty e^{-\omega t}E_\alpha(t^\alpha L)x\,\mathrm{d}t \qquad(\omega\in\mathbb{H},\ x\in X), ω α − 1 ( ω α − L ) − 1 x = ∫ 0 ∞ e − ω t E α ( t α L ) x d t ( ω ∈ H , x ∈ X ) , and K α ( L ) ≤ C L K_\alpha(L) \le C_L K α ( L ) ≤ C L .
Proof. Write S ( t ) : = E α ( t α L ) S(t) := E_\alpha(t^\alpha L) S ( t ) := E α ( t α L ) . For ω ∈ H \omega\in\mathbb{H} ω ∈ H and x ∈ X x \in X x ∈ X , the proof of the lower bound basically follows from the Laplace theorem:
Re ( ω ) ∥ ω α − 1 ( ω α − L ) − 1 x ∥ = Re ( ω ) ∥ ∫ 0 ∞ e − ω t S ( t ) x d t ∥ ≤ Re ( ω ) ∫ 0 ∞ e − t Re ω d t C L ∥ x ∥ = C L ∥ x ∥ . (11) \operatorname{Re}(\omega)\left\lVert\omega^{\alpha-1}(\omega^\alpha-L)^{-1}x\right\rVert=\operatorname{Re}(\omega)\left\lVert\int_0^\infty e^{-\omega t}S(t)x\,\mathrm{d}t\right\rVert\le\operatorname{Re}(\omega)\int_0^\infty e^{-t\operatorname{Re}\omega}\,\mathrm{d}t\,C_L\lVert x\rVert=C_L\lVert x\rVert. \tag*{(11)} Re ( ω ) ω α − 1 ( ω α − L ) − 1 x = Re ( ω ) ∫ 0 ∞ e − ω t S ( t ) x d t ≤ Re ( ω ) ∫ 0 ∞ e − t Re ω d t C L ∥ x ∥ = C L ∥ x ∥ . (11)
Taking the supremum over ω ∈ H \omega\in\mathbb{H} ω ∈ H and ∥ x ∥ ≤ 1 \lVert x\rVert\le1 ∥ x ∥ ≤ 1 gives K α ( L ) ≤ C L K_\alpha(L) \le C_L K α ( L ) ≤ C L by Definition 3.5. It remains to justify the equality in (11) , which is the Laplace formula of the theorem and includes the claim ω α ∈ ρ ( L ) \omega^\alpha\in\rho(L) ω α ∈ ρ ( L ) .
Since ∥ S ( t ) ∥ ≤ C L \lVert S(t)\rVert\le C_L ∥ S ( t )∥ ≤ C L , the integral S ^ ( ω ) : = ∫ 0 ∞ e − ω t S ( t ) d t \widehat{S}(\omega) := \int_0^\infty e^{-\omega t}S(t)\,\mathrm{d}t S ( ω ) := ∫ 0 ∞ e − ω t S ( t ) d t converges absolutely for every ω ∈ H \omega\in\mathbb{H} ω ∈ H . It defines a holomorphic function S ^ : H → B ( X ) \widehat{S} : \mathbb{H} \to B(X) S : H → B ( X ) , being locally bounded and strongly holomorphic. The map ω ↦ ω α \omega\mapsto\omega^\alpha ω ↦ ω α is a biholomorphism of H \mathbb{H} H onto Σ α \Sigma_\alpha Σ α , so
R ( z ) : = z ( 1 − α ) / α S ^ ( z 1 / α ) R(z):=z^{(1-\alpha)/\alpha}\widehat{S}\left(z^{1/\alpha}\right) R ( z ) := z ( 1 − α ) / α S ( z 1/ α )
is holomorphic on Σ α \Sigma_\alpha Σ α . By (4), R ( z ) = ( z − L ) − 1 R(z)=(z-L)^{-1} R ( z ) = ( z − L ) − 1 for real z > w α z>w^\alpha z > w α .
Let U U U be the connected component of the open set ρ ( L ) ∩ Σ α \rho(L)\cap\Sigma_\alpha ρ ( L ) ∩ Σ α that contains ( w α , ∞ ) (w^\alpha,\infty) ( w α , ∞ ) . By the identity theorem, R ( z ) = ( z − L ) − 1 R(z)=(z-L)^{-1} R ( z ) = ( z − L ) − 1 on U U U . Suppose z ∗ ∈ Σ α z_*\in\Sigma_\alpha z ∗ ∈ Σ α is a boundary point of U U U . Then z ∗ ∈ σ ( L ) z_*\in\sigma(L) z ∗ ∈ σ ( L ) , and as z → z ∗ z\to z_* z → z ∗ within U U U ,
∥ R ( z ) ∥ = ∥ ( z − L ) − 1 ∥ ≥ dist ( z , σ ( L ) ) − 1 → ∞ . \lVert R(z)\rVert=\lVert(z-L)^{-1}\rVert\ge\operatorname{dist}\left(z,\sigma(L)\right)^{-1}\to\infty. ∥ R ( z )∥ = ∥( z − L ) − 1 ∥ ≥ dist ( z , σ ( L ) ) − 1 → ∞.
This contradicts the continuity of R R R at z ∗ z_* z ∗ . Hence U = Σ α U=\Sigma_\alpha U = Σ α , that is, Σ α ⊂ ρ ( L ) \Sigma_\alpha\subset\rho(L) Σ α ⊂ ρ ( L ) . Moreover,
ω α − 1 ( ω α − L ) − 1 = ω α − 1 R ( ω α ) = S ^ ( ω ) \omega^{\alpha-1}(\omega^\alpha-L)^{-1}=\omega^{\alpha-1}R(\omega^\alpha)=\widehat{S}(\omega) ω α − 1 ( ω α − L ) − 1 = ω α − 1 R ( ω α ) = S ( ω )
for every ω ∈ H \omega\in\mathbb{H} ω ∈ H , which is the equality in (11) . ∎
Finite-rank operators and their slow projection
A finite-rank operator L = U V L=UV L = U V , with U : C n → X U:\mathbb{C}^{n}\to X U : C n → X and V : X → C n V:X\to\mathbb{C}^{n} V : X → C n , acts nontrivially only through the small matrix L ^ : = V U \widehat{L}:=VU L := V U , and the identity ( z − U V ) − 1 = z − 1 ( I + U ( z − V U ) − 1 V ) (z-UV)^{-1}=z^{-1}(I+U(z-VU)^{-1}V) ( z − U V ) − 1 = z − 1 ( I + U ( z − V U ) − 1 V ) reduces its resolvent to that of L ^ \widehat{L} L . The next lemma collects what we need from this reduction, and it identifies the slow component of a finite-rank operator.
Lemma 5.2. Let L ∈ B ( X ) L\in\mathcal{B}(X) L ∈ B ( X ) have rank n ≥ 1 n\ge1 n ≥ 1 . Choose a basis x 1 , … , x n x_1,\ldots,x_n x 1 , … , x n of Ran L \operatorname{Ran}L Ran L , write L x = ∑ j = 1 n φ j ( x ) x j Lx=\sum_{j=1}^{n}\varphi_j(x)x_j Lx = ∑ j = 1 n φ j ( x ) x j with φ j ∈ X ∗ \varphi_j\in X^* φ j ∈ X ∗ , and let U : C n → X U:\mathbb{C}^{n}\to X U : C n → X , U c : = ∑ j c j x j Uc:=\sum_j c_jx_j U c := ∑ j c j x j , and V : X → C n V:X\to\mathbb{C}^{n} V : X → C n , V x : = ( φ j ( x ) ) j Vx:=(\varphi_j(x))_j V x := ( φ j ( x ) ) j , so that L = U V L=UV L = U V with U U U injective and V V V surjective. Put L ^ : = V U ∈ C n × n \widehat{L}:=VU\in\mathbb{C}^{n\times n} L := V U ∈ C n × n .
(a) σ ( L ) ∖ { 0 } = σ ( L ^ ) ∖ { 0 } \sigma(L)\setminus\{0\}=\sigma(\widehat{L})\setminus\{0\} σ ( L ) ∖ { 0 } = σ ( L ) ∖ { 0 } , and for z ∉ σ ( L ^ ) ∪ { 0 } z\notin\sigma(\widehat{L})\cup\{0\} z ∈ / σ ( L ) ∪ { 0 } , ( z − L ) − 1 = z − 1 ( I + U ( z − L ^ ) − 1 V ) (z-L)^{-1}=z^{-1}(I+U(z-\widehat{L})^{-1}V) ( z − L ) − 1 = z − 1 ( I + U ( z − L ) − 1 V ) .
(b) If K α ( L ) < ∞ \mathcal{K}_{\alpha}(L)<\infty K α ( L ) < ∞ , then L ^ \widehat{L} L is invertible, Π : = U L ^ − 1 V \Pi:=U\widehat{L}^{-1}V Π := U L − 1 V is a projection with Ran Π = Ran L \operatorname{Ran}\Pi=\operatorname{Ran}L Ran Π = Ran L and ker Π = ker L \ker\Pi=\ker L ker Π = ker L , so that X = Ran L ⊕ ker L X=\operatorname{Ran}L\oplus\ker L X = Ran L ⊕ ker L , and L Π = Π L = L L\Pi=\Pi L=L L Π = Π L = L . Moreover
E α ( t α L ) = U E α ( t α L ^ ) L ^ − 1 V + ( I − Π ) ( t ≥ 0 ) . E_{\alpha}(t^{\alpha}L)=U E_{\alpha}(t^{\alpha}\widehat{L})\widehat{L}^{-1}V+(I-\Pi)\qquad(t\geq0). E α ( t α L ) = U E α ( t α L ) L − 1 V + ( I − Π ) ( t ≥ 0 ) . (c) For x ∈ X x\in X x ∈ X and x ∗ ∈ X ∗ x^*\in X^* x ∗ ∈ X ∗ the function F ( z ) : = z ⟨ ( z − L ) − 1 x , x ∗ ⟩ F(z):=z\langle(z-L)^{-1}x,x^*\rangle F ( z ) := z ⟨( z − L ) − 1 x , x ∗ ⟩ extends to a proper rational function whose reduced denominator has degree at most n n n and whose poles lie in σ ( L ^ ) \sigma(\widehat{L}) σ ( L ) . If K α ( L ) < ∞ \mathcal{K}_{\alpha}(L)<\infty K α ( L ) < ∞ , then F ∈ R α , n ∪ { 0 } F\in\mathcal{R}_{\alpha,n}\cup\{0\} F ∈ R α , n ∪ { 0 } , F F F is regular at 0 0 0 , and F ( 0 ) = ⟨ ( I − Π ) x , x ∗ ⟩ F(0)=\langle(I-\Pi)x,x^*\rangle F ( 0 ) = ⟨( I − Π ) x , x ∗ ⟩ .
Proof. (a) The identity is verified by multiplying out, and it shows that every z ∉ σ ( L ^ ) ∪ { 0 } z\notin\sigma(\widehat{L})\cup\{0\} z ∈ / σ ( L ) ∪ { 0 } lies in ρ ( L ) \rho(L) ρ ( L ) ; conversely, if L ^ c = z c \widehat{L}c=zc L c = z c with z ≠ 0 z\ne0 z = 0 and c ≠ 0 c\ne0 c = 0 , then L ( U c ) = z U c L(Uc)=zUc L ( U c ) = z U c and U c ≠ 0 Uc\ne0 U c = 0 .
(b) Suppose L ^ \widehat{L} L is singular, so that ( s − L ^ ) − 1 (s-\widehat{L})^{-1} ( s − L ) − 1 has a pole at s = 0 s=0 s = 0 and ∥ ( s − L ^ ) − 1 ∥ ≥ c 0 s − 1 \|(s-\widehat{L})^{-1}\|\geq c_0s^{-1} ∥ ( s − L ) − 1 ∥ ≥ c 0 s − 1 for small s > 0 s>0 s > 0 . Since U U U is injective and V V V is surjective, there is c 1 > 0 c_1>0 c 1 > 0 with ∥ U M V ∥ ≥ c 1 ∥ M ∥ \|UMV\|\geq c_1\|M\| ∥ U M V ∥ ≥ c 1 ∥ M ∥ for all M ∈ C n × n M\in\mathbb{C}^{n\times n} M ∈ C n × n , and (a) gives s ∥ ( s − L ) − 1 ∥ ≥ c 1 ∥ ( s − L ^ ) − 1 ∥ − 1 → ∞ s\|(s-L)^{-1}\|\geq c_1\|(s-\widehat{L})^{-1}\|-1\to\infty s ∥ ( s − L ) − 1 ∥ ≥ c 1 ∥ ( s − L ) − 1 ∥ − 1 → ∞ as s ↓ 0 s\downarrow0 s ↓ 0 .
Since a α ( s ) = s a_{\alpha}(s)=s a α ( s ) = s , this contradicts K α ( L ) < ∞ \mathcal{K}_{\alpha}(L)<\infty K α ( L ) < ∞ . Hence L ^ \widehat{L} L is invertible, and
Π 2 = U L ^ − 1 L ^ L ^ − 1 V = Π , L Π = U V U L ^ − 1 V = L Π L = U L ^ − 1 V U V = L \Pi^2=U\widehat{L}^{-1}\widehat{L}\widehat{L}^{-1}V=\Pi,\quad L\Pi=UVU\widehat{L}^{-1}V=L\Pi L=U\widehat{L}^{-1}VUV=L Π 2 = U L − 1 L L − 1 V = Π , L Π = U V U L − 1 V = L Π L = U L − 1 V U V = L
The range of Π \Pi Π is Ran U = Ran L \operatorname{Ran}U=\operatorname{Ran}L Ran U = Ran L , and Π x = 0 \Pi x=0 Π x = 0 if and only if V x = 0 Vx=0 V x = 0 , that is, L x = 0 Lx=0 Lx = 0 . Finally, L k = U L ^ k − 1 V L^k=U\widehat{L}^{k-1}V L k = U L k − 1 V for k ≥ 1 k\geq1 k ≥ 1 , so
E α ( t α L ) = I + U ( ∑ k ≥ 1 t α k L ^ k − 1 Γ ( 1 + α k ) ) V = I + U ( E α ( t α L ^ ) − I ) L ^ − 1 V , E_{\alpha}(t^{\alpha}L)=I+U\left(\sum_{k\geq1}\frac{t^{\alpha k}\widehat{L}^{k-1}}{\Gamma(1+\alpha k)}\right)V=I+U\left(E_{\alpha}(t^{\alpha}\widehat{L})-I\right)\widehat{L}^{-1}V, E α ( t α L ) = I + U ( k ≥ 1 ∑ Γ ( 1 + α k ) t α k L k − 1 ) V = I + U ( E α ( t α L ) − I ) L − 1 V ,
which is the stated formula.
(c) By (a), F ( z ) = ⟨ x , x ∗ ⟩ + ℓ T ( z − L ^ ) − 1 k F(z)=\langle x,x^*\rangle+\ell^{\mathsf{T}}(z-\widehat{L})^{-1}k F ( z ) = ⟨ x , x ∗ ⟩ + ℓ T ( z − L ) − 1 k with k : = V x k:=Vx k := V x and ℓ i : = ⟨ x i , x ∗ ⟩ \ell_i:=\langle x_i,x^*\rangle ℓ i := ⟨ x i , x ∗ ⟩ . By Cramer’s rule,
F ( z ) = ⟨ x , x ∗ ⟩ det ( z − L ^ ) + ℓ T adj ( z − L ^ ) k det ( z − L ^ ) , F(z)=\frac{\langle x,x^*\rangle\det(z-\widehat{L})+\ell^{\mathsf{T}}\operatorname{adj}(z-\widehat{L})k}{\det(z-\widehat{L})}, F ( z ) = det ( z − L ) ⟨ x , x ∗ ⟩ det ( z − L ) + ℓ T adj ( z − L ) k ,
which is proper with denominator of degree n n n ; cancelling common factors preserves properness, and the poles lie in σ ( L ^ ) \sigma(\widehat{L}) σ ( L ) . If K α ( L ) < ∞ \mathcal{K}_{\alpha}(L)<\infty K α ( L ) < ∞ , then σ ( L ) ∩ Σ α = ∅ \sigma(L)\cap\Sigma_{\alpha}=\varnothing σ ( L ) ∩ Σ α = ∅ , so σ ( L ^ ) ∩ Σ α = ∅ \sigma(\widehat{L})\cap\Sigma_{\alpha}=\varnothing σ ( L ) ∩ Σ α = ∅ by (a) and 0 ∉ Σ α 0\notin\Sigma_{\alpha} 0 ∈ / Σ α , and L ^ \widehat{L} L is invertible by (b), so F F F is regular at 0 0 0 with F ( 0 ) = ⟨ x , x ∗ ⟩ − ℓ T L ^ − 1 V x = ⟨ ( I − Π ) x , x ∗ ⟩ F(0)=\langle x,x^*\rangle-\ell^{\mathsf{T}}\widehat{L}^{-1}Vx=\langle(I-\Pi)x,x^*\rangle F ( 0 ) = ⟨ x , x ∗ ⟩ − ℓ T L − 1 V x = ⟨( I − Π ) x , x ∗ ⟩ . ∎
By (b), the family acts on Ran L \operatorname{Ran} L Ran L through the n × n n \times n n × n matrix L ^ \widehat{L} L and leaves the kernel of L L L fixed. The two subspaces need not be orthogonal, so the propagator of the slow part is U E α ( t α L ^ ) L ^ − 1 V = E α ( t α L ) Π U E_{\alpha}(t^{\alpha}\widehat{L})\widehat{L}^{-1}V = E_{\alpha}(t^{\alpha}L)\Pi U E α ( t α L ) L − 1 V = E α ( t α L ) Π , an operator on all of X X X whose norm feels the angle between Ran L \operatorname{Ran} L Ran L and ker L \operatorname{ker} L ker L . This obliqueness of the splitting already shows up on both sides of Theorem A.
Remark 5.3 (Obliqueness). If K α ( L ) < ∞ K_{\alpha}(L) < \infty K α ( L ) < ∞ , then ∥ I − Π ∥ ≤ K α ( L ) \lVert I-\Pi\rVert\le K_{\alpha}(L) ∥ I − Π ∥ ≤ K α ( L ) . Indeed, by Theorem 5.2(a), s ( s − L ) − 1 = I + U ( s − L ^ ) − 1 V → I − U L ^ − 1 V = I − Π s(s-L)^{-1} = I + U(s-\widehat{L})^{-1}V \to I-U\widehat{L}^{-1}V = I-\Pi s ( s − L ) − 1 = I + U ( s − L ) − 1 V → I − U L − 1 V = I − Π as s ↓ 0 s \downarrow0 s ↓ 0 , and a α ( s ) = s a_{\alpha}(s)=s a α ( s ) = s . If moreover every eigenvalue of L ^ \widehat{L} L lies in the interior of Λ α \Lambda_{\alpha} Λ α , then E α ( t α L ^ ) → 0 E_{\alpha}(t^{\alpha}\widehat{L}) \to0 E α ( t α L ) → 0 , because in the Jordan form of L ^ \widehat{L} L the entries are multiples of t α j E α ( j ) ( t α λ ) = O ( t − α ) t^{\alpha j}E_{\alpha}^{(j)}(t^{\alpha}\lambda)=O(t^{-\alpha}) t α j E α ( j ) ( t α λ ) = O ( t − α ) by Theorem Appendix B.1(c) (trivially for α = 1 \alpha=1 α = 1 ), and Theorem 5.2(b) gives E α ( t α L ) → I − Π E_{\alpha}(t^{\alpha}L)\to I-\Pi E α ( t α L ) → I − Π in norm. A strongly oblique splitting therefore forces a large Kreiss constant and a large stationary state, whatever the spectrum of L ^ \widehat{L} L .
The upper bound
Theorem 5.4. Let 0 < α ≤ 1 0<\alpha\le1 0 < α ≤ 1 , let X ≠ { 0 } X\ne\{0\} X = { 0 } be a Banach space, and let L ∈ B ( X ) L\in\mathcal{B}(X) L ∈ B ( X ) have rank at most n n n . Then
K α ( L ) ≤ sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ κ α , n K α ( L ) , κ α , n = e 2 ( 3 + 4 n ( 2 − α ) ) . (12) K_{\alpha}(L)\le\sup_{t\ge0}\lVert E_{\alpha}(t^{\alpha}L)\rVert\le\kappa_{\alpha,n}K_{\alpha}(L),\qquad\kappa_{\alpha,n}=\frac{e}{2}\bigl(3+4n(2-\alpha)\bigr). \tag*{(12)} K α ( L ) ≤ t ≥ 0 sup ∥ E α ( t α L )∥ ≤ κ α , n K α ( L ) , κ α , n = 2 e ( 3 + 4 n ( 2 − α ) ) . (12) In particular, sup t ∥ E α ( t α L ) ∥ < ∞ \sup_t\lVert E_{\alpha}(t^{\alpha}L)\rVert<\infty sup t ∥ E α ( t α L )∥ < ∞ if and only if K α ( L ) < ∞ K_{\alpha}(L)<\infty K α ( L ) < ∞ .
Proof. The lower bound is Theorem 5.1 . For the upper bound we may assume K : = K α ( L ) < ∞ K:=K_{\alpha}(L)<\infty K := K α ( L ) < ∞ , so that σ ( L ) ∩ Σ α = ∅ \sigma(L)\cap\Sigma_{\alpha}=\varnothing σ ( L ) ∩ Σ α = ∅ . Since t = 0 t=0 t = 0 the claim reads 1 ≤ κ α , n K 1\le\kappa_{\alpha,n}K 1 ≤ κ α , n K , which holds by Theorem 3.6(a), and if L = 0 L=0 L = 0 , then E α ( t α L ) = I E_{\alpha}(t^{\alpha}L)=I E α ( t α L ) = I , we might assume that 1 ≤ rank L ≤ n 1\le\operatorname{rank}L\le n 1 ≤ rank L ≤ n . Then fix unit vectors x ∈ X x\in X x ∈ X and x ∗ ∈ X ∗ x^*\in X^* x ∗ ∈ X ∗ .
Define the function g : R ≥ 0 → C g:\mathbb{R}_{\ge0}\to\mathbb{C} g : R ≥ 0 → C to be
g ( t ) : = ⟨ E α ( t α L ) x , x ∗ ⟩ g(t):=\langle E_{\alpha}(t^{\alpha}L)x,x^*\rangle g ( t ) := ⟨ E α ( t α L ) x , x ∗ ⟩
and the function g ^ : H → C \widehat{g}:\mathbb{H}\to\mathbb{C} g : H → C as
g ^ ( ω ) : = ω α − 1 ⟨ ( ω α − L ) − 1 x , x ∗ ⟩ for ω ∈ H . \widehat{g}(\omega):=\omega^{\alpha-1}\langle(\omega^\alpha-L)^{-1}x,x^*\rangle\qquad\text{for }\omega\in\mathbb{H}. g ( ω ) := ω α − 1 ⟨( ω α − L ) − 1 x , x ∗ ⟩ for ω ∈ H .
Since { ω α : ω ∈ H } = Σ α ⊂ ρ ( L ) \{\omega^\alpha:\omega\in\mathbb{H}\}=\Sigma_\alpha\subset\rho(L) { ω α : ω ∈ H } = Σ α ⊂ ρ ( L ) , g ^ \widehat{g} g is holomorphic on H \mathbb{H} H , and by Theorem 5.2(c) it equals h F h_F h F , with the corresponding F ∈ R α , n ∪ { 0 } F\in\mathcal{R}_{\alpha,n}\cup\{0\} F ∈ R α , n ∪ { 0 } being F ( z ) = z ⟨ ( z − L ) − 1 x , x ∗ ⟩ F(z)=z\langle(z-L)^{-1}x,x^*\rangle F ( z ) = z ⟨( z − L ) − 1 x , x ∗ ⟩ . Let R : = ( ω α − L ) − 1 R:=(\omega^\alpha-L)^{-1} R := ( ω α − L ) − 1 , then the following holds:
g ^ ′ ( ω ) = ( α − 1 ) ω α − 2 ⟨ R x , x ∗ ⟩ − α ω 2 α − 2 ⟨ R 2 x , x ∗ ⟩ , \widehat{g}'(\omega)=(\alpha-1)\omega^{\alpha-2}\langle Rx,x^*\rangle-\alpha\omega^{2\alpha-2}\langle R^2x,x^*\rangle, g ′ ( ω ) = ( α − 1 ) ω α − 2 ⟨ R x , x ∗ ⟩ − α ω 2 α − 2 ⟨ R 2 x , x ∗ ⟩ ,
and ∥ R ∥ ≤ 2 ∣ ω ∣ − α \lVert R\rVert\le2|\omega|^{-\alpha} ∥ R ∥ ≤ 2∣ ω ∣ − α for ∣ ω ∣ α ≥ 2 ∥ L ∥ |\omega|^\alpha\ge2\lVert L\rVert ∣ ω ∣ α ≥ 2 ∥ L ∥ , so that ∣ g ^ ′ ( ω ) ∣ ≤ 4 ∥ x ∥ ∥ x ∗ ∥ ∣ ω ∣ − 2 |\widehat{g}'(\omega)|\le4\lVert x\rVert\lVert x^*\rVert|\omega|^{-2} ∣ g ′ ( ω ) ∣ ≤ 4 ∥ x ∥ ∥ x ∗ ∥ ∣ ω ∣ − 2 .
Let c : = ∥ L ∥ 1 / α c:=\lVert L\rVert^{1/\alpha} c := ∥ L ∥ 1/ α . By Theorem 3.3(a),
∫ 0 ∞ e − ω t g ( t ) d t = g ^ ( ω ) (13) \int_0^\infty e^{-\omega t}g(t)\,\mathrm{d}t=\widehat{g}(\omega) \tag*{(13)} ∫ 0 ∞ e − ω t g ( t ) d t = g ( ω ) (13)
for Re ω > c \operatorname{Re}\omega>c Re ω > c , with absolute convergence, and differentiation under the integral gives
∫ 0 ∞ e − ω t g ( t ) d t = − g ^ ′ ( ω ) . \int_{0}^{\infty}e^{-\omega t}g(t)\,\mathrm{d}t=-\widehat{g}'(\omega). ∫ 0 ∞ e − ω t g ( t ) d t = − g ′ ( ω ) .
For b > c b>c b > c the function t ↦ e − b t g ( t ) 1 [ 0 , ∞ ) ( t ) t\mapsto e^{-bt}g(t)\mathbf{1}_{[0,\infty)}(t) t ↦ e − b t g ( t ) 1 [ 0 , ∞ ) ( t ) is continuous and integrable on R \mathbb{R} R , its Fourier transform y ↦ − g ^ ′ ( b + i y ) y\mapsto-\widehat{g}'(b+iy) y ↦ − g ′ ( b + i y ) is continuous and integrable. The Fourier inversion theorem gives the formula below on the line Re ω = b \operatorname{Re}\omega=b Re ω = b . For 0 < a < b 0<a<b 0 < a < b , Cauchy’s theorem on the rectangles with vertices a ± i Y a\pm iY a ± iY , b ± i Y b\pm iY b ± iY , Y ≥ ( 2 ∥ L ∥ ) 1 / α Y\ge(2\lVert L\rVert)^{1/\alpha} Y ≥ ( 2 ∥ L ∥ ) 1/ α , whose horizontal sides contribute at most ( b − a ) e b t 4 ∥ x ∥ ∥ x ∗ ∥ Y − 2 (b-a)e^{bt}4\lVert x\rVert\lVert x^*\rVert Y^{-2} ( b − a ) e b t 4 ∥ x ∥ ∥ x ∗ ∥ Y − 2 , moves it to every line Re ω = a > 0 \operatorname{Re}\omega=a>0 Re ω = a > 0 :
⟨ E α ( t α L ) x , x ∗ ⟩ = − 1 2 π i t ∫ a − i ∞ a + i ∞ e ω t g ^ ′ ( ω ) d ω ( t > 0 , a > 0 ) , (14) \langle E_{\alpha}(t^{\alpha}L)x,x^*\rangle=-\frac{1}{2\pi i\,t}\int_{a-i\infty}^{a+i\infty}e^{\omega t}\widehat{g}'(\omega)\,\mathrm{d}\omega\qquad(t>0,\ a>0), \tag*{(14)} ⟨ E α ( t α L ) x , x ∗ ⟩ = − 2 π i t 1 ∫ a − i ∞ a + i ∞ e ω t g ′ ( ω ) d ω ( t > 0 , a > 0 ) , (14)
with an absolutely convergent integral. The integration by parts hidden in the factor t t t is what makes the integrand integrable, and it is the reason why the variation of h F h_F h F , and not h F h_F h F itself, enters the bound. Apart from the identification g ^ = h F \widehat{g}=h_F g = h F , Steps 1 and 2 use only that L ∈ B ( X ) L\in\mathcal{B}(X) L ∈ B ( X ) and Σ α ⊂ ρ ( L ) \Sigma_\alpha\subset\rho(L) Σ α ⊂ ρ ( L ) , so (13) holds for every such operator and all x ∈ X x\in X x ∈ X , x ∗ ∈ X ∗ x^*\in X^* x ∗ ∈ X ∗ .
Otherwise (13) and Theorem 4.1 give, for every a > 0 a>0 a > 0 and t > 0 t>0 t > 0 ,
∣ g ( t ) ∣ ≤ e a t 2 π t ∫ R ∣ h F ′ ( a + i y ) ∣ d y ≤ e a t 2 π t π ( 3 + 4 n ( 2 − α ) ) sup y ∈ R ∣ h F ( a + i y ) ∣ . |g(t)|\le\frac{e^{at}}{2\pi t}\int_{\mathbb{R}}|h_F'(a+iy)|\,\mathrm{d}y\le\frac{e^{at}}{2\pi t}\pi\bigl(3+4n(2-\alpha)\bigr)\sup_{y\in\mathbb{R}}|h_F(a+iy)|. ∣ g ( t ) ∣ ≤ 2 π t e a t ∫ R ∣ h F ′ ( a + i y ) ∣ d y ≤ 2 π t e a t π ( 3 + 4 n ( 2 − α ) ) y ∈ R sup ∣ h F ( a + i y ) ∣.
Since ∣ h F ( ω ) ∣ ≤ ∥ ω α − 1 ( ω α − L ) − 1 ∥ ≤ κ Re ω |h_F(\omega)|\le\|\omega^{\alpha-1}(\omega^\alpha-L)^{-1}\|\le\frac{\kappa}{\operatorname{Re}\omega} ∣ h F ( ω ) ∣ ≤ ∥ ω α − 1 ( ω α − L ) − 1 ∥ ≤ Re ω κ , we obtain
∣ g ( t ) ∣ ≤ 3 + 4 n ( 2 − α ) 2 ⋅ e a t a t ⋅ κ . |g(t)|\le\frac{3+4n(2-\alpha)}{2}\cdot\frac{e^{at}}{at}\cdot\kappa. ∣ g ( t ) ∣ ≤ 2 3 + 4 n ( 2 − α ) ⋅ a t e a t ⋅ κ .
The choice a = 1 t a=\frac{1}{t} a = t 1 minimises e a t a t \frac{e^{at}}{at} a t e a t , with value e e e , and the supremum over x x x and x ∗ x^* x ∗ gives (12) .
Remark 5.5. (i) Theorem 5.4 holds on every Banach space and uses no Hilbert-space structure, and only the rank of L L L enters, not the dimension of X X X : for L ∈ C d × d L\in\mathbb{C}^{d\times d} L ∈ C d × d with rank L = n < d \operatorname{rank}L=n<d rank L = n < d the pole of the resolvent at 0 0 0 is absorbed by the factor z z z in F ( z ) = z ⟨ ( z − L ) − 1 x , x ∗ ⟩ F(z)=z\langle(z-L)^{-1}x,x^*\rangle F ( z ) = z ⟨( z − L ) − 1 x , x ∗ ⟩ . For α = 1 \alpha=1 α = 1 , Spijker’s lemma replaces π ( 3 + 4 n ) \pi(3+4n) π ( 3 + 4 n ) by 2 π ( n + 1 ) 2\pi(n+1) 2 π ( n + 1 ) and the same argument gives sup t ∥ e t L ∥ ≤ e ( n + 1 ) K 1 ( L ) \sup_t\|e^{tL}\|\le e(n+1)\mathcal{K}_1(L) sup t ∥ e t L ∥ ≤ e ( n + 1 ) K 1 ( L ) for operators of rank n n n on Banach spaces. Theorem 4.1 replaces Spijker’s lemma for every α < 1 \alpha<1 α < 1 , where no algebraic argument is available. (ii) The constant is κ α , n \kappa_{\alpha,n} κ α , n and not κ α , n ′ \kappa'_{\alpha,n} κ α , n ′ because F ( 0 ) = ⟨ ( I − Π ) x , x ∗ ⟩ F(0)=\langle(I-\Pi)x,x^*\rangle F ( 0 ) = ⟨( I − Π ) x , x ∗ ⟩ need not vanish: the fixed kernel contributes a stationary term to every matrix element. If one estimates the slow part E α ( t α L ) Π E_\alpha(t^\alpha L)\Pi E α ( t α L ) Π separately, Theorem 7.1 below gives the refined constant κ α , n ′ \kappa'_{\alpha,n} κ α , n ′ for it, at the price of the term ∥ I − Π ∥ \|I-\Pi\| ∥ I − Π∥ for the kernel; by Theorem 5.3 this term never exceeds the left-hand side of (12) . (iii) By Theorem 3.6(b), K α ( L ) \mathcal{K}_\alpha(L) K α ( L ) may be finite while K 1 ( L ) = ∞ \mathcal{K}_1(L)=\infty K 1 ( L ) = ∞ . Then e t L e^{tL} e t L is unbounded, and it grows exponentially when σ ( L ) \sigma(L) σ ( L ) meets H \mathbb{H} H , so neither subordination nor (1) gives any information about E α ( t α L ) E_\alpha(t^\alpha L) E α ( t α L ) .
Hilbert spaces: normal operators and the numerical range
On a Hilbert space the fast part of a generator is often controlled by its numerical range, and the question is what the numerical range can say about the Mittag–Leffler family. For semigroups the answer is the Lumer–Phillips theorem: a dissipative generator has a contractive semigroup. For α < 1 \alpha< 1 α < 1 the family of a dissipative generator is still contractive (Theorem 3.3(c)), but the stability region Λ α \Lambda_\alpha Λ α is larger than the left half-plane, and the relevant question becomes what happens when the numerical range avoids only the sector Σ α \Sigma_\alpha Σ α . We first settle the normal case, which identifies the sharp constant.
Normal operators
Proposition 6.1 (Normal operators). Let 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 and let L L L be a normal operator on H H H , possibly unbounded, with spectral measure P L P_L P L and σ ( L ) ⊂ Λ α \sigma(L) \subset\Lambda_\alpha σ ( L ) ⊂ Λ α . Then L L L generates a Mittag–Leffler family,
E α ( t α L ) = ∫ σ ( L ) E α ( t α z ) d P L ( z ) , ∥ E α ( t α L ) ∥ = sup z ∈ σ ( L ) ∣ E α ( t α z ) ∣ , E_\alpha(t^\alpha L) = \int_{\sigma(L)} E_\alpha(t^\alpha z)\,dP_L(z), \qquad\|E_\alpha(t^\alpha L)\| = \sup_{z\in\sigma(L)} |E_\alpha(t^\alpha z)|, E α ( t α L ) = ∫ σ ( L ) E α ( t α z ) d P L ( z ) , ∥ E α ( t α L ) ∥ = z ∈ σ ( L ) sup ∣ E α ( t α z ) ∣ , (4) holds for every λ > 0 \lambda> 0 λ > 0 , and 1 ≤ K α ( L ) ≤ 1 / α 1 \le\mathcal{K}_\alpha(L) \le1/\alpha 1 ≤ K α ( L ) ≤ 1/ α , sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ M α ∗ \sup_{t\ge0}\|E_\alpha(t^\alpha L)\| \le M_\alpha^* sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ M α ∗ , independently of dim H \dim H dim H . Both bounds are sharp: if σ ( L ) \sigma(L) σ ( L ) contains a nonzero point of ∂ Σ α \partial\Sigma_\alpha ∂ Σ α , then K α ( L ) = 1 \mathcal{K}_\alpha(L) = 1 K α ( L ) = 1 and sup t ∥ E α ( t α L ) ∥ = M α ∗ \sup_t \|E_\alpha(t^\alpha L)\| = M_\alpha^* sup t ∥ E α ( t α L ) ∥ = M α ∗ . In particular, for 0 < α < 1 0 < \alpha< 1 0 < α < 1 normal operators exhibit transient growth, and the ratio of the two sides of Theorem A equals α M α ∗ > 1 \alpha M_\alpha^* > 1 α M α ∗ > 1 for them.
Proof. Since t α Λ α = Λ α t^\alpha\Lambda_\alpha= \Lambda_\alpha t α Λ α = Λ α and E α E_\alpha E α is bounded and continuous on Λ α \Lambda_\alpha Λ α , the operators S ( t ) : = ∫ E α ( t α z ) d P L ( z ) S(t) := \int E_\alpha(t^\alpha z)\,dP_L(z) S ( t ) := ∫ E α ( t α z ) d P L ( z ) are bounded, with norm equal to the supremum of ∣ E α ( t α ⋅ ) ∣ |E_\alpha(t^\alpha\cdot)| ∣ E α ( t α ⋅ ) ∣ over the support σ ( L ) \sigma(L) σ ( L ) of P L P_L P L , and S S S is strongly continuous by dominated convergence. For z ∈ Λ α z \in\Lambda_\alpha z ∈ Λ α the functions λ ↦ ∫ 0 ∞ e − λ t E α ( t α z ) d t \lambda\mapsto\int_0^\infty e^{-\lambda t}E_\alpha(t^\alpha z)\,dt λ ↦ ∫ 0 ∞ e − λ t E α ( t α z ) d t and λ ↦ λ α − 1 ( λ α − z ) − 1 \lambda\mapsto\lambda^{\alpha-1}(\lambda^\alpha-z)^{-1} λ ↦ λ α − 1 ( λ α − z ) − 1 are holomorphic on H \mathbb{H} H , since the integrand is bounded and λ α ∈ Σ α ∌ z \lambda^\alpha\in\Sigma_\alpha\not\ni z λ α ∈ Σ α ∋ z , and they agree for λ > ∣ z ∣ 1 / α \lambda> |z|^{1/\alpha} λ > ∣ z ∣ 1/ α by Theorem 3.3(a), hence for all λ > 0 \lambda> 0 λ > 0 . Fubini’s theorem with respect to the complex measures ⟨ P L ( ⋅ ) x , y ⟩ \langle P_L(\cdot)x,y\rangle ⟨ P L ( ⋅ ) x , y ⟩ gives (4) , and ∥ S ( t ) ∥ ≤ M α ∗ \|S(t)\| \le M_\alpha^* ∥ S ( t ) ∥ ≤ M α ∗ .
For z ∈ Σ α z \in\Sigma_\alpha z ∈ Σ α we have ∥ ( z − L ) − 1 ∥ = dist ( z , σ ( L ) ) − 1 ≤ dist ( z , ∂ Σ α ) − 1 \|(z-L)^{-1}\| = \operatorname{dist}(z,\sigma(L))^{-1} \le\operatorname{dist}(z,\partial\Sigma_\alpha)^{-1} ∥ ( z − L ) − 1 ∥ = dist ( z , σ ( L ) ) − 1 ≤ dist ( z , ∂ Σ α ) − 1 , so K α ( L ) ≤ 1 / α \mathcal{K}_\alpha(L) \le1/\alpha K α ( L ) ≤ 1/ α by Theorem 3.6(c), and K α ( L ) ≥ 1 \mathcal{K}_\alpha(L) \ge1 K α ( L ) ≥ 1 by Theorem 3.6(a). If λ = r e i α π / 2 ∈ σ ( L ) \lambda= r e^{i\alpha\pi/2} \in\sigma(L) λ = r e i α π /2 ∈ σ ( L ) with r > 0 r > 0 r > 0 (the lower ray is treated in the same way, by conjugation symmetry), the points z = λ + s n z = \lambda+ s\mathbf{n} z = λ + s n , with n \mathbf{n} n the inner unit normal of the boundary ray at λ \lambda λ , satisfy ∥ ( z − L ) − 1 ∥ ≥ s − 1 \|(z-L)^{-1}\| \ge s^{-1} ∥ ( z − L ) − 1 ∥ ≥ s − 1 and a α ( z ) / s → 1 / α a_\alpha(z)/s \to1/\alpha a α ( z ) / s → 1/ α as s ↓ 0 s \downarrow0 s ↓ 0 , by the computation in the proof of Theorem 3.6(c). Moreover t α λ t^\alpha\lambda t α λ runs over the whole boundary ray as t t t varies, so sup t ∥ E α ( t α L ) ∥ ≥ M α ( α π / 2 ) = M α ∗ \sup_t\|E_\alpha(t^\alpha L)\| \ge M_\alpha(\alpha\pi/2) = M_\alpha^* sup t ∥ E α ( t α L ) ∥ ≥ M α ( α π /2 ) = M α ∗ by Theorem 3.1. □
Remark 6.2 (Generators without a semigroup). Let 0 < α < 1 0 < \alpha< 1 0 < α < 1 , let A ≠ 0 A \ne0 A = 0 be a nonnegative self-adjoint operator and α π 2 ≤ θ < π 2 \frac{\alpha\pi}{2} \le\theta< \frac{\pi}{2} 2 α π ≤ θ < 2 π . Then L : = e ± i θ A L := e^{\pm i\theta}A L := e ± i θ A is normal with spectrum on a ray in Λ α \Lambda_\alpha Λ α , so it generates a Mittag–Leffler family with sup t ∥ E α ( t α L ) ∥ = M α ( θ ) \sup_t\|E_\alpha(t^\alpha L)\| = M_\alpha(\theta) sup t ∥ E α ( t α L ) ∥ = M α ( θ ) and K α ( L ) ≤ 1 / α \mathcal{K}_\alpha(L) \le1/\alpha K α ( L ) ≤ 1/ α ; if A A A is unbounded, then Re σ ( L ) \operatorname{Re}\sigma(L) Re σ ( L ) is unbounded above and L L L generates no C 0 C_0 C 0 -semigroup [14 ] . The time-fractional Schrödinger equation i α ∂ t α u = A V u i^\alpha\partial_t^\alpha u = A_Vu i α ∂ t α u = A V u of Naber [37 ] , with a Schrödinger operator A V A_V A V , corresponds to L = i − α A V = e − i α π / 2 A V L = i^{-\alpha}A_V = e^{-i\alpha\pi/2}A_V L = i − α A V = e − i α π /2 A V . If A V A_V A V is self-adjoint, bounded below and unbounded above, the positive part of its spectrum is mapped onto the lower boundary ray of Σ α \Sigma_\alpha Σ α and its negative eigenvalues onto the ray of argument π − α π 2 \pi-\frac{\alpha\pi}{2} π − 2 α π , so the solution operators are uniformly bounded with sup t ∥ E α ( t α L ) ∥ = M α ∗ \sup_t\|E_\alpha(t^\alpha L)\| = M_\alpha^* sup t ∥ E α ( t α L ) ∥ = M α ∗ exactly, although e t L e^{tL} e t L does not exist. This is prototype (b) of the introduction, to which § 11.2 returns.
The numerical range
For a closed, densely defined operator L L L on H H H let W ( L ) : = { ⟨ L v , v ⟩ : v ∈ D ( L ) , ∥ v ∥ = 1 } W(L):=\{\langle Lv,v\rangle:v\in D(L),\ \lVert v\rVert=1\} W ( L ) := {⟨ Lv , v ⟩ : v ∈ D ( L ) , ∥ v ∥ = 1 } be its numerical range, which is convex, and let C W C_W C W be the closed convex cone generated by it, i.e. the closure of { s w : s ≥ 0 , w ∈ W ( L ) } \{sw:s\ge0,\ w\in W(L)\} { s w : s ≥ 0 , w ∈ W ( L )} . If W ( L ) W(L) W ( L ) does not meet the open cone Σ α \Sigma_\alpha Σ α , neither does C W C_W C W . The following geometric lemma shows that a half-plane, with spectral-set constant being equal to one, is as good as any convex set for the functions E α ( t α ⋅ ) E_\alpha(t^\alpha\cdot) E α ( t α ⋅ ) .
Lemma 6.3. Let 0 < α ≤ 1 0<\alpha\le1 0 < α ≤ 1 and let C ≠ C C\ne\mathbb{C} C = C be a closed convex cone with vertex 0 0 0 that does not meet Σ α \Sigma_\alpha Σ α . There is a closed half-plane Ω 0 \Omega_0 Ω 0 with 0 ∈ ∂ Ω 0 0\in\partial\Omega_0 0 ∈ ∂ Ω 0 , C ⊂ Ω 0 C\subset\Omega_0 C ⊂ Ω 0 and Ω 0 ∩ Σ α = ∅ \Omega_0\cap\Sigma_\alpha=\varnothing Ω 0 ∩ Σ α = ∅ such that sup z ∈ Ω 0 ∣ E α ( z ) ∣ = sup z ∈ C ∣ E α ( z ) ∣ \sup_{z\in\Omega_0}|E_\alpha(z)|=\sup_{z\in C}|E_\alpha(z)| sup z ∈ Ω 0 ∣ E α ( z ) ∣ = sup z ∈ C ∣ E α ( z ) ∣ . This common value is the larger of M α ( ∣ θ ∣ ) M_\alpha(|\theta|) M α ( ∣ θ ∣ ) over the arguments θ ∈ ( − π , π ] \theta\in(-\pi,\pi] θ ∈ ( − π , π ] of the boundary rays of C C C , and it is at most M α ∗ M_\alpha^* M α ∗ .
Proof. For α = 1 \alpha=1 α = 1 every such cone lies in the closed left half-plane, on which ∣ e z ∣ ≤ 1 |e^z|\le1 ∣ e z ∣ ≤ 1 , so we let α < 1 \alpha<1 α < 1 ; we may also assume C ≠ { 0 } C\ne\{0\} C = { 0 } . If C C C is a line, it bounds a closed half-plane disjoint from Σ α \Sigma_\alpha Σ α , on which the supremum is attained on the line by Theorem 3.1. Otherwise C = { r e i θ : r ≥ 0 , θ 1 ≤ θ ≤ θ 2 } C=\{re^{i\theta}:r\ge0,\ \theta_1\le\theta\le\theta_2\} C = { r e i θ : r ≥ 0 , θ 1 ≤ θ ≤ θ 2 } with α π 2 ≤ θ 1 − θ 2 ≤ 2 π − α π 2 \frac{\alpha\pi}{2}\le\theta_1-\theta_2\le2\pi-\frac{\alpha\pi}{2} 2 α π ≤ θ 1 − θ 2 ≤ 2 π − 2 α π and θ 2 − θ 1 ≤ π \theta_2-\theta_1\le\pi θ 2 − θ 1 ≤ π . Let d ( θ ) : = min { θ , 2 π − θ } d(\theta):=\min\{\theta,2\pi-\theta\} d ( θ ) := min { θ , 2 π − θ } and δ 0 : = min { d ( θ 1 ) , d ( θ 2 ) } ≥ α π 2 \delta_0:=\min\{d(\theta_1),d(\theta_2)\}\ge\frac{\alpha\pi}{2} δ 0 := min { d ( θ 1 ) , d ( θ 2 )} ≥ 2 α π . By Theorem 3.1, sup C ∣ E α ∣ = M α ( δ 0 ) \sup_C|E_\alpha|=M_\alpha(\delta_0) sup C ∣ E α ∣ = M α ( δ 0 ) , and for φ ∈ [ α π 2 , π − α π 2 ] \varphi\in[\frac{\alpha\pi}{2},\pi-\frac{\alpha\pi}{2}] φ ∈ [ 2 α π , π − 2 α π ] the half-plane Ω φ : = { r e i θ : r ≥ 0 , φ ≤ θ ≤ φ + π } \Omega_\varphi:=\{re^{i\theta}:r\ge0,\ \varphi\le\theta\le\varphi+\pi\} Ω φ := { r e i θ : r ≥ 0 , φ ≤ θ ≤ φ + π } does not meet Σ α \Sigma_\alpha Σ α and satisfies sup Ω φ ∣ E α ∣ = M α ( min { φ , π − φ } ) \sup_{\Omega_\varphi}|E_\alpha|=M_\alpha(\min\{\varphi,\pi-\varphi\}) sup Ω φ ∣ E α ∣ = M α ( min { φ , π − φ }) . If δ 0 ≥ π 2 \delta_0\ge\frac{\pi}{2} δ 0 ≥ 2 π , then C ⊂ Ω π 2 C\subset\Omega_{\frac{\pi}{2}} C ⊂ Ω 2 π and both suprema equal 1. If δ 0 < π 2 \delta_0<\frac{\pi}{2} δ 0 < 2 π , choose
φ ∈ [ max { θ 2 − π , δ 0 } , min { θ 1 , π − δ 0 } ] . \varphi\in\left[\max\{\theta_2-\pi,\delta_0\},\ \min\{\theta_1,\pi-\delta_0\}\right]. φ ∈ [ max { θ 2 − π , δ 0 } , min { θ 1 , π − δ 0 } ] .
an interval that is nonempty because θ 2 − π ≤ θ 1 \theta_2-\pi\le\theta_1 θ 2 − π ≤ θ 1 , θ 2 ≤ 2 π − d ( θ 2 ) ≤ 2 π − δ 0 \theta_2\le2\pi-d(\theta_2)\le2\pi-\delta_0 θ 2 ≤ 2 π − d ( θ 2 ) ≤ 2 π − δ 0 , δ 0 ≤ d ( θ 1 ) ≤ θ 1 \delta_0\le d(\theta_1)\le\theta_1 δ 0 ≤ d ( θ 1 ) ≤ θ 1 and δ 0 < π 2 \delta_0<\frac{\pi}{2} δ 0 < 2 π . Then φ ≤ θ 1 \varphi\le\theta_1 φ ≤ θ 1 and φ + π ≥ θ 2 \varphi+\pi\ge\theta_2 φ + π ≥ θ 2 , so C ⊂ Ω φ C\subset\Omega_\varphi C ⊂ Ω φ , and min { φ , π − φ } ≥ δ 0 \min\{\varphi,\pi-\varphi\}\ge\delta_0 min { φ , π − φ } ≥ δ 0 , so sup Ω φ ∣ E α ∣ ≤ M α ( δ 0 ) = sup C ∣ E α ∣ \sup_{\Omega_\varphi}|E_\alpha|\le M_\alpha(\delta_0)=\sup_C|E_\alpha| sup Ω φ ∣ E α ∣ ≤ M α ( δ 0 ) = sup C ∣ E α ∣ by monotonicity of M α M_\alpha M α . The reverse inequality holds because C ⊂ Ω φ C\subset\Omega_\varphi C ⊂ Ω φ .
Numerical range
Theorem 6.4 (Numerical range). Let 0 < α ≤ 1 0<\alpha\le1 0 < α ≤ 1 and let L L L be a closed, densely defined operator on H H H with W ( L ) ∩ Σ α = ∅ W(L)\cap\Sigma_\alpha=\varnothing W ( L ) ∩ Σ α = ∅ and ρ ( L ) ∩ Σ α ≠ ∅ \rho(L)\cap\Sigma_\alpha\ne\varnothing ρ ( L ) ∩ Σ α = ∅ .
(a) Σ α ⊂ ρ ( L ) \Sigma_\alpha\subset\rho(L) Σ α ⊂ ρ ( L ) , ∥ ( z − L ) − 1 ∥ ≤ dist ( z , W ( L ) ) − 1 \lVert(z-L)^{-1}\rVert\le\operatorname{dist}(z,W(L))^{-1} ∥( z − L ) − 1 ∥ ≤ dist ( z , W ( L ) ) − 1 for z ∈ Σ α z\in\Sigma_\alpha z ∈ Σ α , and 1 ≤ K α ( L ) ≤ 1 α 1\le K_\alpha(L)\le\frac{1}{\alpha} 1 ≤ K α ( L ) ≤ α 1 .
∥ E α ( t α L ) ∥ ≤ sup z ∈ ∂ Ω ∣ E α ( t α z ) ∣ ( t ≥ 0 ) . (15) \lVert E_\alpha(t^\alpha L)\rVert\le\sup_{z\in\partial\Omega}|E_\alpha(t^\alpha z)|\qquad(t\ge0). \tag*{(15)} ∥ E α ( t α L )∥ ≤ z ∈ ∂ Ω sup ∣ E α ( t α z ) ∣ ( t ≥ 0 ) . (15) sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ sup z ∈ C W ∣ E α ( z ) ∣ ≤ M α ∗ , \sup_{t\ge0}\lVert E_\alpha(t^\alpha L)\rVert\le\sup_{z\in C_W}|E_\alpha(z)|\le M_\alpha^*, t ≥ 0 sup ∥ E α ( t α L )∥ ≤ z ∈ C W sup ∣ E α ( z ) ∣ ≤ M α ∗ , and for every closed convex cone C ≠ C C\ne\mathbb{C} C = C disjoint from Σ α \Sigma_\alpha Σ α the first bound is attained by a normal operator whose numerical range generates C C C . If W ( L ) ⊂ { Re z ≤ − δ } W(L)\subset\{\operatorname{Re}z\le-\delta\} W ( L ) ⊂ { Re z ≤ − δ } with δ ≥ 0 \delta\ge0 δ ≥ 0 , then ∥ E α ( t α L ) ∥ ≤ G α ( δ t α ) \lVert E_\alpha(t^\alpha L)\rVert\le G_\alpha(\delta t^\alpha) ∥ E α ( t α L )∥ ≤ G α ( δ t α ) , where G α ( s ) : = sup y ∈ R ∣ E α ( − s + i y ) ∣ G_\alpha(s):=\sup_{y\in\mathbb{R}}|E_\alpha(-s+iy)| G α ( s ) := sup y ∈ R ∣ E α ( − s + i y ) ∣ . In particular, if L L L is dissipative, then K α ( L ) = sup t ∥ E α ( t α L ) ∥ = 1 K_\alpha(L)=\sup_t\lVert E_\alpha(t^\alpha L)\rVert=1 K α ( L ) = sup t ∥ E α ( t α L )∥ = 1 .
For 0 < α < 1 0<\alpha<1 0 < α < 1 we have E α ( − s ) ≤ G α ( s ) ≤ min { 1 , C α s } E_\alpha(-s)\le G_\alpha(s)\le\min\{1,\frac{C_\alpha}{s}\} E α ( − s ) ≤ G α ( s ) ≤ min { 1 , s C α } for s > 0 s>0 s > 0 , by Theorem 3.1 and Theorem Appendix B.1(b), so a uniformly dissipative generator satisfies ∥ E α ( t α L ) ∥ ≤ min { 1 , C α δ − 1 t − α } \|E_\alpha(t^\alpha L)\|\le\min\{1,C_\alpha\delta^{-1}t^{-\alpha}\} ∥ E α ( t α L ) ∥ ≤ min { 1 , C α δ − 1 t − α } , and the order t − α t^{-\alpha} t − α is attained by L = − δ L=-\delta L = − δ . For α = 1 \alpha=1 α = 1 , G 1 ( s ) = e − s G_1(s)=e^{-s} G 1 ( s ) = e − s and (15) reduces to the Lumer–Phillips bound.
Proof. (a) For z ∉ W ( L ) ‾ z\notin\overline{W(L)} z ∈ / W ( L ) and unit vectors v ∈ D ( L ) v\in D(L) v ∈ D ( L ) , ∥ ( z − L ) v ∥ ≥ ∣ ⟨ ( z − L ) v , v ⟩ ∣ ≥ dist ( z , W ( L ) ) \|(z-L)v\|\ge|\langle(z-L)v,v\rangle|\ge\operatorname{dist}(z,W(L)) ∥ ( z − L ) v ∥ ≥ ∣ ⟨( z − L ) v , v ⟩ ∣ ≥ dist ( z , W ( L )) , so z − L z-L z − L is injective with closed range, and by Kato [20 Theorem V.3.2] the codimension of its range is constant on each connected component of C ∖ W ( L ) ‾ \mathbb{C}\setminus\overline{W(L)} C ∖ W ( L ) . The open connected set Σ α \Sigma_\alpha Σ α does not meet W ( L ) ‾ \overline{W(L)} W ( L ) and meets ρ ( L ) \rho(L) ρ ( L ) , so it lies in a component on which z − L z-L z − L is bijective. This gives Σ α ⊂ ρ ( L ) \Sigma_\alpha\subset\rho(L) Σ α ⊂ ρ ( L ) and the resolvent bound. Since dist ( z , W ( L ) ) ≥ dist ( z , Λ α ) = dist ( z , ∂ Σ α ) \operatorname{dist}(z,W(L))\ge\operatorname{dist}(z,\Lambda_\alpha)=\operatorname{dist}(z,\partial\Sigma_\alpha) dist ( z , W ( L )) ≥ dist ( z , Λ α ) = dist ( z , ∂ Σ α ) for z ∈ Σ α z\in\Sigma_\alpha z ∈ Σ α , the bound K α ( L ) ≤ 1 α \mathcal{K}_\alpha(L)\le\frac{1}{\alpha} K α ( L ) ≤ α 1 follows from Theorem 3.6(c).
(b) Write Ω = { z : Re ( e − i χ z ) ≤ − δ } \Omega=\{z:\operatorname{Re}(e^{-i\chi}z)\le-\delta\} Ω = { z : Re ( e − i χ z ) ≤ − δ } . Since Ω ∩ Σ α = ∅ \Omega\cap\Sigma_\alpha=\varnothing Ω ∩ Σ α = ∅ and Σ α \Sigma_\alpha Σ α is a cone, replacing z z z by r z rz r z and letting r → 0 r\to0 r → 0 and r → ∞ r\to\infty r → ∞ shows that δ ≥ 0 \delta\ge0 δ ≥ 0 and ∣ χ ∣ ≤ ( 1 − α ) π 2 |\chi|\le\frac{(1-\alpha)\pi}{2} ∣ χ ∣ ≤ 2 ( 1 − α ) π . Put T : = e − i χ L + δ T:=e^{-i\chi}L+\delta T := e − i χ L + δ with D ( T ) = D ( L ) D(T)=D(L) D ( T ) = D ( L ) , so that Re ⟨ T v , v ⟩ ≤ 0 \operatorname{Re}\langle Tv,v\rangle\le0 Re ⟨ T v , v ⟩ ≤ 0 for v ∈ D ( L ) v\in D(L) v ∈ D ( L ) . The open half-plane V : = { z : Re ( e − i χ z ) > − δ } V:=\{z:\operatorname{Re}(e^{-i\chi}z)>-\delta\} V := { z : Re ( e − i χ z ) > − δ } is connected, does not meet W ( L ) ‾ \overline{W(L)} W ( L ) , and contains 1 ∈ Σ α ⊂ ρ ( L ) 1\in\Sigma_\alpha\subset\rho(L) 1 ∈ Σ α ⊂ ρ ( L ) because cos χ > 0 \cos\chi>0 cos χ > 0 ; by the argument in (a), V ⊂ ρ ( L ) V\subset\rho(L) V ⊂ ρ ( L ) . For λ > 0 \lambda>0 λ > 0 the point e i χ ( λ − δ ) e^{i\chi}(\lambda-\delta) e i χ ( λ − δ ) lies in V V V , so λ − T = e − i χ ( e i χ ( λ − δ ) − L ) \lambda-T=e^{-i\chi}(e^{i\chi}(\lambda-\delta)-L) λ − T = e − i χ ( e i χ ( λ − δ ) − L ) is bijective. Thus T T T is m-dissipative and generates a C 0 C_0 C 0 -semigroup of contractions. By the dilation theorem of Sz.-Nagy [45 ] there are a Hilbert space K ⊃ H K\supset H K ⊃ H and a strongly continuous unitary group ( U ( s ) ) s ∈ R (U(s))_{s\in\mathbb{R}} ( U ( s ) ) s ∈ R on K K K such that e s T = P H U ( s ) ∣ H e^{sT}=P_HU(s)|_H e s T = P H U ( s ) ∣ H for s ≥ 0 s\ge0 s ≥ 0 , where P H P_H P H is the orthogonal projection onto H H H . Let G G G be the skew-adjoint generator of ( U ( s ) ) (U(s)) ( U ( s )) . Taking Laplace transforms,
( ς − T ) − 1 h = P H ( ς − G ) − 1 h ( h ∈ H , Re ς > 0 ) . (16) (\varsigma-T)^{-1}h=P_H(\varsigma-G)^{-1}h\qquad(h\in H,\ \operatorname{Re}\varsigma>0). \tag*{(16)} ( ς − T ) − 1 h = P H ( ς − G ) − 1 h ( h ∈ H , Re ς > 0 ) . (16)
The operator N : = e i χ ( G − δ ) N:=e^{i\chi}(G-\delta) N := e i χ ( G − δ ) is normal on K K K with σ ( N ) ⊂ e i χ ( i R − δ ) = ∂ Ω ⊂ Λ α \sigma(N)\subset e^{i\chi}(i\mathbb{R}-\delta)=\partial\Omega\subset\Lambda_\alpha σ ( N ) ⊂ e i χ ( i R − δ ) = ∂ Ω ⊂ Λ α . By Theorem 6.1, N N N generates a Mittag–Leffler family with ∥ E α ( t α N ) ∥ ≤ sup z ∈ ∂ Ω ∣ E α ( t α z ) ∣ \|E_\alpha(t^\alpha N)\|\le\sup_{z\in\partial\Omega}|E_\alpha(t^\alpha z)| ∥ E α ( t α N ) ∥ ≤ sup z ∈ ∂ Ω ∣ E α ( t α z ) ∣ . Put S ( t ) : = P H E α ( t α N ) ∣ H S(t):=P_HE_\alpha(t^\alpha N)|_H S ( t ) := P H E α ( t α N ) ∣ H , which is strongly continuous and satisfies (15) . Let λ > 0 \lambda>0 λ > 0 and ς : = e − i χ λ α + δ \varsigma:=e^{-i\chi}\lambda^\alpha+\delta ς := e − i χ λ α + δ , so that Re ς = λ α cos χ + δ > 0 \operatorname{Re}\varsigma=\lambda^\alpha\cos\chi+\delta>0 Re ς = λ α cos χ + δ > 0 , λ α − N = e i χ ( ς − G ) \lambda^\alpha-N=e^{i\chi}(\varsigma-G) λ α − N = e i χ ( ς − G ) and ς − T = e − i χ ( λ α − L ) \varsigma-T=e^{-i\chi}(\lambda^\alpha-L) ς − T = e − i χ ( λ α − L ) . By Theorem 6.1 and (16) , for h ∈ H h\in H h ∈ H ,
∫ 0 ∞ e − λ t S ( t ) h d t = λ α − 1 e − i χ P H ( ς − G ) − 1 h = λ α − 1 e − i χ ( ς − T ) − 1 h = λ α − 1 ( λ α − L ) − 1 h , \int_0^\infty e^{-\lambda t}S(t)h\,dt=\lambda^{\alpha-1}e^{-i\chi}P_H(\varsigma-G)^{-1}h=\lambda^{\alpha-1}e^{-i\chi}(\varsigma-T)^{-1}h=\lambda^{\alpha-1}(\lambda^\alpha-L)^{-1}h, ∫ 0 ∞ e − λ t S ( t ) h d t = λ α − 1 e − i χ P H ( ς − G ) − 1 h = λ α − 1 e − i χ ( ς − T ) − 1 h = λ α − 1 ( λ α − L ) − 1 h ,
so L L L generates the Mittag–Leffler family S S S .
(c) By Theorem 6.3 there is a closed half-plane Ω 0 ⊃ C W ⊃ W ( L ) \Omega_0\supset C_W\supset W(L) Ω 0 ⊃ C W ⊃ W ( L ) with 0 ∈ ∂ Ω 0 0\in\partial\Omega_0 0 ∈ ∂ Ω 0 , Ω 0 ∩ Σ α = ∅ \Omega_0\cap\Sigma_\alpha=\varnothing Ω 0 ∩ Σ α = ∅ and sup Ω 0 ∣ E α ∣ = sup C W ∣ E α ∣ \sup_{\Omega_0}|E_\alpha|=\sup_{C_W}|E_\alpha| sup Ω 0 ∣ E α ∣ = sup C W ∣ E α ∣ . Since t α ∂ Ω 0 = ∂ Ω 0 t^\alpha\partial\Omega_0=\partial\Omega_0 t α ∂ Ω 0 = ∂ Ω 0 , (15) with Ω = Ω 0 \Omega=\Omega_0 Ω = Ω 0 gives the first bound, and the second is Theorem 6.3. For sharpness, let Γ \Gamma Γ be the union of the boundary rays of a given cone C C C and let L L L be multiplication by z z z on L 2 ( Γ , ∣ d z ∣ ) L^2(\Gamma,|dz|) L 2 ( Γ , ∣ d z ∣ ) . It is normal with σ ( L ) = Γ \sigma(L)=\Gamma σ ( L ) = Γ , the cone generated by W ( L ) W(L) W ( L ) is C C C , and by Theorem 6.1 and Theorem 3.1 its family has supremum sup z ∈ Γ ∣ E α ( z ) ∣ = sup C ∣ E α ∣ \sup_{z\in\Gamma}|E_\alpha(z)|=\sup_C|E_\alpha| sup z ∈ Γ ∣ E α ( z ) ∣ = sup C ∣ E α ∣ . For the second claim take Ω = { Re z ≤ − δ } \Omega=\{\operatorname{Re}z\le-\delta\} Ω = { Re z ≤ − δ } and note t α ∂ Ω = { Re z = − δ t α } t^\alpha\partial\Omega=\{\operatorname{Re}z=-\delta t^\alpha\} t α ∂ Ω = { Re z = − δ t α } . If L L L is dissipative, the case δ = 0 \delta=0 δ = 0 and M α ( π 2 ) = 1 M_\alpha(\frac{\pi}{2})=1 M α ( 2 π ) = 1 give ∥ E α ( t α L ) ∥ ≤ 1 \|E_\alpha(t^\alpha L)\|\le1 ∥ E α ( t α L ) ∥ ≤ 1 , and 1 ≤ K α ( L ) ≤ sup t ∥ E α ( t α L ) ∥ 1\le\mathcal{K}_\alpha(L)\le\sup_t\|E_\alpha(t^\alpha L)\| 1 ≤ K α ( L ) ≤ sup t ∥ E α ( t α L ) ∥ by (a) and Theorem 5.1 .
Estimate (15) says that a closed half-plane containing W ( L ) W(L) W ( L ) and disjoint from Σ α \Sigma_\alpha Σ α is a spectral set with constant one for the functions E α ( t α ⋅ ) E_\alpha(t^\alpha\cdot) E α ( t α ⋅ ) , which is von Neumann’s inequality for a half-plane [49 ] . In addition, we note that Crouzeix and Palencia [8 ] proved that the closure of the numerical range of a bounded operator is a ( 1 + 2 ) (1+\sqrt{2}) ( 1 + 2 ) -spectral set, and Crouzeix [7 Theorem 3] showed that an unbounded closed convex set Ω ≠ C \Omega\ne\mathbb{C} Ω = C containing the numerical range, with aperture 2 β 2\beta 2 β , the maximal opening of a sector contained in Ω \Omega Ω , is a K ( β ) K(\beta) K ( β ) -spectral set for the rational functions bounded on Ω \Omega Ω and for the holomorphic functions on Ω \Omega Ω that are continuous on Ω ‾ \overline{\Omega} Ω and have a limit at infinity, with K ( β ) = 1 − β / π + 2 − 4 β / π + β 2 / π 2 K(\beta)=1-\beta/\pi+\sqrt{2-4\beta/\pi+\beta^2/\pi^2} K ( β ) = 1 − β / π + 2 − 4 β / π + β 2 / π 2
his Section 3 extends this to closed, densely defined operators.
Generators with a finite slow component
We now remove the assumption that the whole operator has finite rank, and allow the remaining dynamics to be unbounded and of infinite rank, provided it is bounded for a structural reason such as dissipativity or a numerical range outside Σ α \Sigma_\alpha Σ α . A slow component (Theorem 3.8) isolates finitely many modes, and the rank dependence of Theorem 5.4 is confined to them. The geometry of the splitting, which may be very oblique, is carried by the deflated Kreiss constant K α ( L ; P ) K_\alpha(L;P) K α ( L ; P ) and by the fast propagator E α ( t α L ) ( I − P ) E_\alpha(t^\alpha L)(I-P) E α ( t α L ) ( I − P ) , both of which can be computed.
Theorem 7.1 (Kreiss theorem with a finite slow component). Let 0 < α ≤ 1 0<\alpha\le1 0 < α ≤ 1 , let X ≠ { 0 } X\ne\{0\} X = { 0 } , let L L L be closed and densely defined on X X X , and let P ∈ B ( X ) P\in\mathcal{B}(X) P ∈ B ( X ) be a projection of finite rank m ≥ 1 m\ge1 m ≥ 1 commuting with L L L , with slow block L P L_P L P and Q : = I − P Q:=I-P Q := I − P .
(i) The slow part of the propagator satisfies
K α ( L ; P ) ≤ sup t ≥ 0 ∥ E α ( t α L P ) P ∥ ≤ κ α , m K α ( L ; P ) , (17) K_\alpha(L;P)\le\sup_{t\ge0}\lVert E_\alpha(t^\alpha L_P)P\rVert\le\kappa_{\alpha,m}K_\alpha(L;P), \tag*{(17)} K α ( L ; P ) ≤ t ≥ 0 sup ∥ E α ( t α L P ) P ∥ ≤ κ α , m K α ( L ; P ) , (17) and κ α , m \kappa_{\alpha,m} κ α , m may be replaced by κ α , m ′ \kappa'_{\alpha,m} κ α , m ′ if 0 ∉ σ ( L P ) 0\notin\sigma(L_P) 0 ∈ / σ ( L P ) . If L L L generates a Mittag–Leffler family, then E α ( t α L ) P = E α ( t α L P ) P E_\alpha(t^\alpha L)P=E_\alpha(t^\alpha L_P)P E α ( t α L ) P = E α ( t α L P ) P .
(ii) If P P P is a slow component, then L L L generates a Mittag–Leffler family, E α ( t α L ) = E α ( t α L P ) P + E α ( t α L Q ) Q E_\alpha(t^\alpha L)=E_\alpha(t^\alpha L_P)P+E_\alpha(t^\alpha L_Q)Q E α ( t α L ) = E α ( t α L P ) P + E α ( t α L Q ) Q , and with C Q : = sup t ≥ 0 ∥ E α ( t α L ) Q ∥ C_Q:=\sup_{t\ge0}\lVert E_\alpha(t^\alpha L)Q\rVert C Q := sup t ≥ 0 ∥ E α ( t α L ) Q ∥ ,
K α ( L ) ≤ sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ κ α , m K α ( L ; P ) + C Q , (18) K_\alpha(L)\le\sup_{t\ge0}\lVert E_\alpha(t^\alpha L)\rVert\le\kappa_{\alpha,m}K_\alpha(L;P)+C_Q, \tag*{(18)} K α ( L ) ≤ t ≥ 0 sup ∥ E α ( t α L )∥ ≤ κ α , m K α ( L ; P ) + C Q , (18) with κ α , m ′ \kappa'_{\alpha,m} κ α , m ′ in place of κ α , m \kappa_{\alpha,m} κ α , m if 0 ∉ σ ( L P ) 0\notin\sigma(L_P) 0 ∈ / σ ( L P ) .
(iii) If X = H X=H X = H and P P P is an orthogonal slow component, then K α ( L ; P ) = K α ( L P ) K_\alpha(L;P)=K_\alpha(L_P) K α ( L ; P ) = K α ( L P ) and
sup t ≥ 0 ∥ E α ( t α L ) ∥ = max { sup t ≥ 0 ∥ E α ( t α L P ) ∥ , sup t ≥ 0 ∥ E α ( t α L Q ) ∥ } . \sup_{t\ge0}\lVert E_\alpha(t^\alpha L)\rVert=\max\left\{\sup_{t\ge0}\lVert E_\alpha(t^\alpha L_P)\rVert,\sup_{t\ge0}\lVert E_\alpha(t^\alpha L_Q)\rVert\right\}. t ≥ 0 sup ∥ E α ( t α L )∥ = max { t ≥ 0 sup ∥ E α ( t α L P )∥ , t ≥ 0 sup ∥ E α ( t α L Q )∥ } .
Proof. (i) Write S P ( t ) : = E α ( t α L P ) P S_P(t):=E_\alpha(t^\alpha L_P)P S P ( t ) := E α ( t α L P ) P and C : = sup t ∥ S P ( t ) ∥ C:=\sup_t\lVert S_P(t)\rVert C := sup t ∥ S P ( t )∥ . By Theorem 3.9(a), L P ∈ B ( P X ) L_P\in\mathcal{B}(PX) L P ∈ B ( P X ) , and ∥ E α ( t α L P ) ∥ B ( P X ) ≤ ∥ S P ( t ) ∥ \lVert E_\alpha(t^\alpha L_P)\rVert_{\mathcal{B}(PX)}\le\lVert S_P(t)\rVert ∥ E α ( t α L P ) ∥ B ( P X ) ≤ ∥ S P ( t )∥ because y = P y y=Py y = P y for y ∈ P X y\in PX y ∈ P X . By Theorem 3.3(a), S P S_P S P is norm continuous and
∫ 0 ∞ e − ω t S P ( t ) d t = ω α − 1 ( ω α − L P ) − 1 P ( Re ω > ∥ L P ∥ 1 / α ) . (19) \int_0^\infty e^{-\omega t}S_P(t)\,dt=\omega^{\alpha-1}(\omega^\alpha-L_P)^{-1}P\qquad\left(\operatorname{Re}\omega>\lVert L_P\rVert^{1/\alpha}\right). \tag*{(19)} ∫ 0 ∞ e − ω t S P ( t ) d t = ω α − 1 ( ω α − L P ) − 1 P ( Re ω > ∥ L P ∥ 1/ α ) . (19)
Lower bound. We may assume C < ∞ C < \infty C < ∞ . Then Σ α ⊂ ρ ( L P ) \Sigma_\alpha\subset\rho(L_P) Σ α ⊂ ρ ( L P ) by Theorem 5.1 applied on P X PX P X , both sides of (19) are holomorphic on H \mathbb{H} H and agree there, and the estimate
Re ( ω ) ∥ ∫ 0 ∞ e − ω t S P ( t ) d t ∥ ≤ C \operatorname{Re}(\omega)\left\|\int_0^\infty e^{-\omega t}S_P(t)\,dt\right\| \le C Re ( ω ) ∫ 0 ∞ e − ω t S P ( t ) d t ≤ C
gives K α ( L ; P ) ≤ C K_\alpha(L;P) \le C K α ( L ; P ) ≤ C .
Upper bound. We may assume K : = K α ( L ; P ) < ∞ K := K_\alpha(L;P) < \infty K := K α ( L ; P ) < ∞ , so that σ ( L P ) ∩ Σ α = ∅ \sigma(L_P) \cap\Sigma_\alpha= \varnothing σ ( L P ) ∩ Σ α = ∅ . For t = 0 t = 0 t = 0 the claim reads ∥ P ∥ ≤ κ α , m K \|P\| \le\kappa_{\alpha,m}K ∥ P ∥ ≤ κ α , m K , which holds by Theorem 3.9(c). Let t > 0 t > 0 t > 0 and let x ∈ X x \in X x ∈ X and x ∗ ∈ X ∗ x^* \in X^* x ∗ ∈ X ∗ be unit vectors. Since L P ∈ B ( P X ) L_P \in B(PX) L P ∈ B ( P X ) and Σ α ⊂ ρ ( L P ) \Sigma_\alpha\subset\rho(L_P) Σ α ⊂ ρ ( L P ) , (13) applies on the Banach space P X PX P X , with L P L_P L P , P x Px P x and the restriction of x ∗ x^* x ∗ in place of L L L , x x x and x ∗ x^* x ∗ , and gives
⟨ S P ( t ) x , x ∗ ⟩ = − 1 2 π i t ∫ a − i ∞ a + i ∞ e ω t g ^ ′ ( ω ) d ω , g ^ ( ω ) : = ω α − 1 ⟨ ( ω α − L P ) − 1 P x , x ∗ ⟩ , \left\langle S_P(t)x,x^*\right\rangle=-\frac{1}{2\pi i t}\int_{a-i\infty}^{a+i\infty}e^{\omega t}\widehat{g}'(\omega)\,d\omega,\qquad\widehat{g}(\omega):=\omega^{\alpha-1}\left\langle(\omega^\alpha-L_P)^{-1}Px,x^*\right\rangle, ⟨ S P ( t ) x , x ∗ ⟩ = − 2 π i t 1 ∫ a − i ∞ a + i ∞ e ω t g ′ ( ω ) d ω , g ( ω ) := ω α − 1 ⟨ ( ω α − L P ) − 1 P x , x ∗ ⟩ ,
for every a > 0 a > 0 a > 0 . Put F ( z ) : = z ⟨ ( z − L P ) − 1 P x , x ∗ ⟩ F(z):=z\left\langle(z-L_P)^{-1}Px,x^*\right\rangle F ( z ) := z ⟨ ( z − L P ) − 1 P x , x ∗ ⟩ , so that g ^ = h F \widehat{g}=h_F g = h F . With the factorisation of Theorem 3.9(b), F ( z ) = z ℓ T ( z − Λ ) − 1 k F(z)=z\ell^{\mathsf T}(z-\Lambda)^{-1}k F ( z ) = z ℓ T ( z − Λ ) − 1 k with k : = V x k:=Vx k := V x and ℓ i : = ⟨ U e i , x ∗ ⟩ \ell_i:=\left\langle Ue_i,x^*\right\rangle ℓ i := ⟨ U e i , x ∗ ⟩ , and Cramer’s rule shows that F F F is a proper rational function with denominator det ( z − Λ ) \det(z-\Lambda) det ( z − Λ ) of degree m m m , whose poles lie in σ ( L P ) \sigma(L_P) σ ( L P ) and hence outside Σ α \Sigma_\alpha Σ α . Thus F ∈ R α , m ∪ { 0 } F\in\mathcal{R}_{\alpha,m}\cup\{0\} F ∈ R α , m ∪ { 0 } . If 0 ∉ σ ( L P ) 0\notin\sigma(L_P) 0 ∈ / σ ( L P ) , then F F F is regular at 0 0 0 and F ( 0 ) = 0 F(0)=0 F ( 0 ) = 0 . If F ≡ 0 F\equiv0 F ≡ 0 , then ⟨ S P ( t ) x , x ∗ ⟩ = 0 \left\langle S_P(t)x,x^*\right\rangle=0 ⟨ S P ( t ) x , x ∗ ⟩ = 0 . Otherwise Theorem 4.1 gives
∣ ⟨ S P ( t ) x , x ∗ ⟩ ∣ ≤ e a t 2 π t ∫ R ∣ h F ′ ( a + i y ) ∣ d y ≤ c 2 π e a t t sup y ∈ R ∣ h F ( a + i y ) ∣ . \left|\left\langle S_P(t)x,x^*\right\rangle\right|\le\frac{e^{at}}{2\pi t}\int_{\mathbb{R}}\left|h_F'(a+iy)\right|\,dy\le\frac{c}{2\pi}\frac{e^{at}}{t}\sup_{y\in\mathbb{R}}\left|h_F(a+iy)\right|. ∣ ⟨ S P ( t ) x , x ∗ ⟩ ∣ ≤ 2 π t e a t ∫ R ∣ h F ′ ( a + i y ) ∣ d y ≤ 2 π c t e a t y ∈ R sup ∣ h F ( a + i y ) ∣ .
with c = π ( 3 + 4 m ( 2 − α ) ) c=\pi(3+4m(2-\alpha)) c = π ( 3 + 4 m ( 2 − α )) , or c = π ( 1 + 4 m ( 2 − α ) ) c=\pi(1+4m(2-\alpha)) c = π ( 1 + 4 m ( 2 − α )) if 0 ∉ σ ( L P ) 0\notin\sigma(L_P) 0 ∈ / σ ( L P ) , and ∣ h F ( ω ) ∣ ≤ ∥ ω α − 1 ( ω α − L P ) − 1 P ∥ ≤ K Re ω \left|h_F(\omega)\right|\le\left\|\omega^{\alpha-1}(\omega^\alpha-L_P)^{-1}P\right\|\le\frac{K}{\operatorname{Re}\omega} ∣ h F ( ω ) ∣ ≤ ω α − 1 ( ω α − L P ) − 1 P ≤ Re ω K . The choice a = 1 t a=\frac{1}{t} a = t 1 and the supremum over x x x and x ∗ x^* x ∗ give ∥ S P ( t ) ∥ ≤ c e 2 π K \|S_P(t)\|\le\frac{ce}{2\pi}K ∥ S P ( t ) ∥ ≤ 2 π ce K , which is the claim.
Finally, if L L L generates a Mittag–Leffler family, then
∫ 0 ∞ e − λ t E α ( t α L ) P x d t = λ α − 1 ( λ α − L ) − 1 P x = λ α − 1 ( λ α − L P ) − 1 P x \int_0^\infty e^{-\lambda t}E_\alpha(t^\alpha L)Px\,dt=\lambda^{\alpha-1}(\lambda^\alpha-L)^{-1}Px=\lambda^{\alpha-1}(\lambda^\alpha-L_P)^{-1}Px ∫ 0 ∞ e − λ t E α ( t α L ) P x d t = λ α − 1 ( λ α − L ) − 1 P x = λ α − 1 ( λ α − L P ) − 1 P x
for large λ \lambda λ by Theorem 3.9(a), and uniqueness of the Laplace transform together with (19) gives E α ( t α L ) P = S P ( t ) E_\alpha(t^\alpha L)P=S_P(t) E α ( t α L ) P = S P ( t ) .
(ii) Let S ( t ) : = E α ( t α L P ) P + E α ( t α L Q ) Q S(t):=E_\alpha(t^\alpha L_P)P+E_\alpha(t^\alpha L_Q)Q S ( t ) := E α ( t α L P ) P + E α ( t α L Q ) Q . This family is strongly continuous, and ∫ 0 ∞ e − w t ∥ S ( t ) ∥ d t < ∞ \int_0^\infty e^{-wt}\|S(t)\|\,dt<\infty ∫ 0 ∞ e − w t ∥ S ( t ) ∥ d t < ∞ for large w w w . For large λ > 0 \lambda>0 λ > 0 the Laplace transforms of the two terms are λ α − 1 ( λ α − L P ) − 1 P x \lambda^{\alpha-1}(\lambda^\alpha-L_P)^{-1}Px λ α − 1 ( λ α − L P ) − 1 P x and λ α − 1 ( λ α − L Q ) − 1 Q x \lambda^{\alpha-1}(\lambda^\alpha-L_Q)^{-1}Qx λ α − 1 ( λ α − L Q ) − 1 Q x , and by Theorem 3.9(a) their sum is λ α − 1 ( λ α − L ) − 1 x \lambda^{\alpha-1}(\lambda^\alpha-L)^{-1}x λ α − 1 ( λ α − L ) − 1 x . Hence L L L generates S S S , and since P Q = Q P = 0 PQ=QP=0 P Q = QP = 0 ,
S ( t ) P = E α ( t α L P ) P , S ( t ) Q = E α ( t α L Q ) Q . S(t)P=E_\alpha(t^\alpha L_P)P,\qquad S(t)Q=E_\alpha(t^\alpha L_Q)Q. S ( t ) P = E α ( t α L P ) P , S ( t ) Q = E α ( t α L Q ) Q .
The upper bound in (18) follows from ∥ S ( t ) ∥ ≤ ∥ S ( t ) P ∥ + ∥ S ( t ) Q ∥ \|S(t)\|\le\|S(t)P\|+\|S(t)Q\| ∥ S ( t ) ∥ ≤ ∥ S ( t ) P ∥ + ∥ S ( t ) Q ∥ and (i), and the lower bound is trivial if the supremum is infinite and otherwise follows from Theorem 5.1 .
(iii) In the orthogonal case S ( t ) S(t) S ( t ) is the orthogonal direct sum of E α ( t α L P ) E_\alpha(t^\alpha L_P) E α ( t α L P ) and E α ( t α L Q ) E_\alpha(t^\alpha L_Q) E α ( t α L Q ) , and K α ( L ; P ) = K α ( L P ) K_\alpha(L;P)=K_\alpha(L_P) K α ( L ; P ) = K α ( L P ) by Theorem 3.9(c). □
Theorem 5.4 should be compared with the special case of Theorem 7.1 in which the fast part is trivial: for a bounded L L L of rank n n n with K α ( L ) < ∞ K_\alpha(L)<\infty K α ( L ) < ∞ , the projection Π \Pi Π of Theorem 5.2 is a slow component with L Q = 0 L_Q=0 L Q = 0 on ker L \ker L ker L , so (18) gives sup t ∥ E α ( t α L ) ∥ ≤ κ α , n ′ K α ( L ; Π ) + ∥ I − Π ∥ \sup_t\|E_\alpha(t^\alpha L)\|\le\kappa'_{\alpha,n}K_\alpha(L;\Pi)+\|I-\Pi\| sup t ∥ E α ( t α L ) ∥ ≤ κ α , n ′ K α ( L ; Π ) + ∥ I − Π∥ . Theorem 5.4 avoids the term ∥ I − Π ∥ \|I-\Pi\| ∥ I − Π∥ altogether, at the price of the constant κ α , n \kappa_{\alpha,n} κ α , n instead of κ α , n ′ \kappa'_{\alpha,n} κ α , n ′ , and in this sense it absorbs the stationary kernel into the Kreiss constant.
The fast part can be controlled on every Banach space by sectoriality, and on Hilbert spaces by its numerical range; in both cases Theorem 7.1 becomes fully explicit.
Corollary 7.2. Let P P P be a projection of finite rank m ≥ 1 m \ge1 m ≥ 1 commuting with L L L , and suppose that the fast part L Q L_Q L Q satisfies the hypothesis of Theorem 3.4 on Q X QX QX for some ψ ∈ ( α π 2 , π ] \psi\in\left(\frac{\alpha\pi}{2},\pi\right] ψ ∈ ( 2 α π , π ] , with ψ ′ \psi' ψ ′ and φ \varphi φ as there. Then P P P is a slow component, L L L generates a Mittag–Leffler family, and
K α ( L ) ≤ sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ κ α , m K α ( L ; P ) + ∥ Q ∥ c φ s ψ ′ ( L Q ) . \mathcal{K}_{\alpha}(L) \le\sup_{t\geq0}\lVert E_{\alpha}(t^{\alpha}L)\rVert\le\kappa_{\alpha,m}\mathcal{K}_{\alpha}(L;P)+\lVert Q\rVert c_{\varphi}s_{\psi'}(L_Q). K α ( L ) ≤ t ≥ 0 sup ∥ E α ( t α L )∥ ≤ κ α , m K α ( L ; P ) + ∥ Q ∥ c φ s ψ ′ ( L Q ) . with κ α , m ′ \kappa'_{\alpha,m} κ α , m ′ in place of κ α , m \kappa_{\alpha,m} κ α , m if 0 ∉ σ ( L P ) 0 \notin\sigma(L_P) 0 ∈ / σ ( L P ) . Neither L L L nor L Q L_Q L Q needs to generate a C 0 C_0 C 0 -semigroup.
Proof. By Theorem 3.4, L Q L_Q L Q generates a Mittag–Leffler family bounded by c φ s ψ ′ ( L Q ) c_{\varphi}s_{\psi'}(L_Q) c φ s ψ ′ ( L Q ) , so P P P is a slow component, and Theorem 7.1 (ii) applies with C Q ≤ ∥ Q ∥ c φ s ψ ′ ( L Q ) C_Q \le\lVert Q\rVert c_{\varphi}s_{\psi'}(L_Q) C Q ≤ ∥ Q ∥ c φ s ψ ′ ( L Q ) .
Corollary 7.3. Let X = H X=H X = H and let P P P be a projection of finite rank m ≥ 1 m\geq1 m ≥ 1 commuting with L L L , with Q = I − P Q=I-P Q = I − P . Suppose that the numerical range W ( L Q ) W(L_Q) W ( L Q ) , taken in the Hilbert space Q H QH Q H , does not meet Σ α \Sigma_{\alpha} Σ α , and that ρ ( L Q ) ∩ Σ α ≠ ∅ \rho(L_Q)\cap\Sigma_{\alpha}\ne\varnothing ρ ( L Q ) ∩ Σ α = ∅ . Let C Q W C_Q^W C Q W be the closed convex cone generated by W ( L Q ) W(L_Q) W ( L Q ) . Then P P P is a slow component, L L L generates a Mittag–Leffler family, and
K α ( L ) ≤ sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ κ α , m K α ( L ; P ) + ∥ Q ∥ sup z ∈ C Q W ∣ E α ( z ) ∣ ≤ κ α , m K α ( L ; P ) + ∥ Q ∥ M α ∗ . \mathcal{K}_{\alpha}(L) \le\sup_{t\geq0}\lVert E_{\alpha}(t^{\alpha}L)\rVert\le\kappa_{\alpha,m}\mathcal{K}_{\alpha}(L;P)+\lVert Q\rVert\sup_{z\in C_Q^W}|E_{\alpha}(z)| \le\kappa_{\alpha,m}\mathcal{K}_{\alpha}(L;P)+\lVert Q\rVert M_{\alpha}^{*}. K α ( L ) ≤ t ≥ 0 sup ∥ E α ( t α L )∥ ≤ κ α , m K α ( L ; P ) + ∥ Q ∥ z ∈ C Q W sup ∣ E α ( z ) ∣ ≤ κ α , m K α ( L ; P ) + ∥ Q ∥ M α ∗ . with κ α , m ′ \kappa'_{\alpha,m} κ α , m ′ in place of κ α , m \kappa_{\alpha,m} κ α , m if 0 ∉ σ ( L P ) 0\notin\sigma(L_P) 0 ∈ / σ ( L P ) ; if W ( L Q ) W(L_Q) W ( L Q ) lies in the closed left half-plane, the last term is ∥ Q ∥ \lVert Q\rVert ∥ Q ∥ . If P P P is orthogonal, then sup t ∥ E α ( t α L ) ∥ ≤ max { κ α , m K α ( L P ) , sup z ∈ C Q W ∣ E α ∣ } \sup_t\lVert E_{\alpha}(t^{\alpha}L)\rVert\leq\max\{\kappa_{\alpha,m}\mathcal{K}_{\alpha}(L_P),\sup_{z\in C_Q^W}|E_{\alpha}|\} sup t ∥ E α ( t α L )∥ ≤ max { κ α , m K α ( L P ) , sup z ∈ C Q W ∣ E α ∣ } .
Proof. By Theorem 6.4 applied to L Q L_Q L Q on Q H QH Q H , L Q L_Q L Q generates a Mittag–Leffler family with sup t ∥ E α ( t α L Q ) ∥ ≤ sup z ∈ C Q W ∣ E α ∣ ≤ M α ∗ \sup_t\lVert E_{\alpha}(t^{\alpha}L_Q)\rVert\leq\sup_{z\in C_Q^W}|E_{\alpha}|\leq M_{\alpha}^{*} sup t ∥ E α ( t α L Q )∥ ≤ sup z ∈ C Q W ∣ E α ∣ ≤ M α ∗ , and the bound is $1$ if W ( L Q ) ⊂ { Re z ≤ 0 } W(L_Q)\subset\{\operatorname{Re}z\leq0\} W ( L Q ) ⊂ { Re z ≤ 0 } . The claims follow from Theorem 7.1 , since $C_Q\leq\lVert Q\rVert\sup_t\lVert E_{\alpha}(t^{\alpha}L_Q)\rVert.
In (18) both quantities on the right-hand side refer to the splitting rather than to the two parts alone. Estimating them separately gives back the cruder bound
sup t ≥ 0 ∥ E α ( t α L ) ∥ ≤ ∥ P ∥ κ α , m K α ( L P ) + ∥ Q ∥ sup t ≥ 0 ∥ E α ( t α L Q ) ∥ , (20) \sup_{t\geq0}\lVert E_{\alpha}(t^{\alpha}L)\rVert\leq\lVert P\rVert\kappa_{\alpha,m}\mathcal{K}_{\alpha}(L_P)+\lVert Q\rVert\sup_{t\geq0}\lVert E_{\alpha}(t^{\alpha}L_Q)\rVert, \tag*{(20)} t ≥ 0 sup ∥ E α ( t α L )∥ ≤ ∥ P ∥ κ α , m K α ( L P ) + ∥ Q ∥ t ≥ 0 sup ∥ E α ( t α L Q )∥ , (20)
in which the non-normality between the slow modes and the remainder appears as a separate factor ∥ P ∥ \lVert P\rVert ∥ P ∥ . This factor reflects a property of L L L and not of a choice. If P P P is the Riesz projection onto an isolated part of the spectrum and P ′ P' P ′ is another projection commuting with L L L with P ′ X = P X P'X=PX P ′ X = P X , then P ′ P' P ′ commutes with every resolvent and hence with P P P , so P ′ = P P ′ = P ′ P = P P'=PP'=P'P=P P ′ = P P ′ = P ′ P = P : among projections with the given range, the splitting is unique and so is its obliqueness.
We conclude this section by a showing that no estimate in terms of the slow block and the fast part alone can hold true, since the obliqueness might the dominating part of the transient growth. To that end, we consider Recalling Prototype (a): a two-dimensional oscillatory mode a ( t ) ∈ C 2 a(t)\in\mathbb{C}^{2} a ( t ) ∈ C 2 , a linearised self-excited oscillator with growth rate cos θ > 0 \cos\theta>0 cos θ > 0 and frequency sin θ \sin\theta sin θ , which feeds a source into a subdiffusive medium on ( 0 , π ) (0,\pi) ( 0 , π ) with absorbing boundaries, both running on the same fractional clock:
∂ t α a = L 1 a , ∂ t α v = ∂ x x v + μ ( a 1 b 1 + a 2 b 2 ) , L 1 = ( cos θ − sin θ sin θ cos θ ) , α π 2 < θ < π 2 , (21) \partial_t^{\alpha}a=L_1a,\qquad\partial_t^{\alpha}v=\partial_{xx}v+\mu\left(a_1b_1+a_2b_2\right),\qquad L_1=\begin{pmatrix}\cos\theta&-\sin\theta\\ \sin\theta&\cos\theta\end{pmatrix},\qquad\frac{\alpha\pi}{2}<\theta<\frac{\pi}{2}, \tag*{(21)} ∂ t α a = L 1 a , ∂ t α v = ∂ xx v + μ ( a 1 b 1 + a 2 b 2 ) , L 1 = ( cos θ sin θ − sin θ cos θ ) , 2 α π < θ < 2 π , (21)
with linearly independent source profiles b 1 , b 2 ∈ L 2 ( 0 , π ) b_1,b_2 \in L^2(0,\pi) b 1 , b 2 ∈ L 2 ( 0 , π ) and a coupling strength μ ≥ 0 \mu\ge0 μ ≥ 0 . On H : = C 2 ⊕ L 2 ( 0 , π ) H := \mathbb{C}^2 \oplus L^2(0,\pi) H := C 2 ⊕ L 2 ( 0 , π ) the generator is L ( a , v ) = ( L 1 a , A v + μ S a ) L(a,v)=(L_1a,Av+\mu Sa) L ( a , v ) = ( L 1 a , A v + μ S a ) , where A A A is the Dirichlet Laplacian with D ( A ) = H 2 ( 0 , π ) ∩ H 0 1 ( 0 , π ) D(A)=H^2(0,\pi)\cap H^1_0(0,\pi) D ( A ) = H 2 ( 0 , π ) ∩ H 0 1 ( 0 , π ) , S a : = a 1 b 1 + a 2 b 2 Sa:=a_1b_1+a_2b_2 S a := a 1 b 1 + a 2 b 2 , and D ( L ) = C 2 ⊕ D ( A ) D(L)=\mathbb{C}^2\oplus D(A) D ( L ) = C 2 ⊕ D ( A ) . Let λ ± = e ± i θ \lambda_\pm=e^{\pm i\theta} λ ± = e ± i θ be the eigenvalues of L 1 L_1 L 1 , with eigenvectors e ± e_\pm e ± , and let Y 0 : C 2 → L 2 ( 0 , π ) Y_0:\mathbb{C}^2\to L^2(0,\pi) Y 0 : C 2 → L 2 ( 0 , π ) be the solution of the Sylvester equation A Y 0 − Y 0 L 1 = − S AY_0-Y_0L_1=-S A Y 0 − Y 0 L 1 = − S , that is, Y 0 e ± = ( λ ± − A ) − 1 S e ± Y_0e_\pm=(\lambda_\pm-A)^{-1}Se_\pm Y 0 e ± = ( λ ± − A ) − 1 S e ± ; it exists because σ ( A ) ⊂ ( − ∞ , − 1 ] \sigma(A)\subset(-\infty,-1] σ ( A ) ⊂ ( − ∞ , − 1 ] does not meet { λ ± } \{\lambda_\pm\} { λ ± } .
Proposition 7.4. Let 0 < α < 1 0<\alpha<1 0 < α < 1 and suppose that Y 0 Y_0 Y 0 is injective, with smallest singular value s 0 > 0 s_0>0 s 0 > 0 and norm y 0 y_0 y 0 . Then L L L is closed and densely defined, ∥ e t L ∥ ≥ e t cos θ \|e^{tL}\|\ge e^{t\cos\theta} ∥ e t L ∥ ≥ e t c o s θ , and W ( L ) W(L) W ( L ) meets Σ α \Sigma_\alpha Σ α for every μ ≥ 0 \mu\ge0 μ ≥ 0 . The Riesz projection P ( a , v ) = ( a , μ Y 0 a ) P(a,v)=(a,\mu Y_0a) P ( a , v ) = ( a , μ Y 0 a ) onto { λ ± } \{\lambda_\pm\} { λ ± } is a slow component of rank 2 with fast part L Q = A L_Q=A L Q = A , and for every μ ≥ 0 \mu\ge0 μ ≥ 0 :
(a) ∥ P ∥ = ∥ Q ∥ = ( 1 + μ 2 y 0 2 ) 1 2 \|P\|=\|Q\|=(1+\mu^2y_0^2)^{\frac12} ∥ P ∥ = ∥ Q ∥ = ( 1 + μ 2 y 0 2 ) 2 1 , C Q = ∥ Q ∥ C_Q=\|Q\| C Q = ∥ Q ∥ and sup t ∥ E α ( t α L Q ) ∥ = 1 \sup_t\|E_\alpha(t^\alpha L_Q)\|=1 sup t ∥ E α ( t α L Q ) ∥ = 1 ;
(b) ∥ P ∥ ≤ K α ( L ; P ) ≤ ∥ P ∥ K α ( L 1 ) \|P\|\le\mathcal{K}_\alpha(L;P)\le\|P\|\mathcal{K}_\alpha(L_1) ∥ P ∥ ≤ K α ( L ; P ) ≤ ∥ P ∥ K α ( L 1 ) , and K α ( L P ) ≤ y 0 s 0 K α ( L 1 ) \mathcal{K}_\alpha(L_P)\le\frac{y_0}{s_0}\mathcal{K}_\alpha(L_1) K α ( L P ) ≤ s 0 y 0 K α ( L 1 ) , a bound independent of μ \mu μ ;
(c) c 1 μ ≤ K α ( L ) ≤ sup t ∥ E α ( t α L ) ∥ ≤ κ α , 2 ′ K α ( L 1 ) ∥ P ∥ + ∥ P ∥ c_1\mu\le\mathcal{K}_\alpha(L)\le\sup_t\|E_\alpha(t^\alpha L)\|\le\kappa'_{\alpha,2}\mathcal{K}_\alpha(L_1)\|P\|+\|P\| c 1 μ ≤ K α ( L ) ≤ sup t ∥ E α ( t α L ) ∥ ≤ κ α , 2 ′ K α ( L 1 ) ∥ P ∥ + ∥ P ∥ , with c 1 : = ∥ ( 1 − A ) − 1 S ( 1 − L 1 ) − 1 ∥ > 0 c_1:=\|(1-A)^{-1}S(1-L_1)^{-1}\|>0 c 1 := ∥ ( 1 − A ) − 1 S ( 1 − L 1 ) − 1 ∥ > 0 .
In particular all of K α ( L ) \mathcal{K}_\alpha(L) K α ( L ) , K α ( L ; P ) \mathcal{K}_\alpha(L;P) K α ( L ; P ) , C Q C_Q C Q and sup t ∥ E α ( t α L ) ∥ \sup_t\|E_\alpha(t^\alpha L)\| sup t ∥ E α ( t α L ) ∥ grow linearly in μ \mu μ , while K α ( L P ) \mathcal{K}_\alpha(L_P) K α ( L P ) and sup t ∥ E α ( t α L Q ) ∥ \sup_t\|E_\alpha(t^\alpha L_Q)\| sup t ∥ E α ( t α L Q ) ∥ stay bounded, so no estimate of the form sup t ∥ E α ( t α L ) ∥ ≤ C ( α , m ) ( K α ( L P ) + sup t ∥ E α ( t α L Q ) ∥ ) \sup_t\|E_\alpha(t^\alpha L)\|\le C(\alpha,m)\bigl(\mathcal{K}_\alpha(L_P)+\sup_t\|E_\alpha(t^\alpha L_Q)\|\bigr) sup t ∥ E α ( t α L ) ∥ ≤ C ( α , m ) ( K α ( L P ) + sup t ∥ E α ( t α L Q ) ∥ ) can hold.
Proof. L L L is a bounded perturbation of L 1 ⊕ A L_1\oplus A L 1 ⊕ A , hence closed, densely defined and the generator of a C 0 C_0 C 0 -semigroup, and the eigenvector ( e + , μ Y 0 e + ) (e_+,\mu Y_0e_+) ( e + , μ Y 0 e + ) for λ + \lambda_+ λ + gives ∥ e t L ∥ ≥ e t cos θ \|e^{tL}\|\ge e^{t\cos\theta} ∥ e t L ∥ ≥ e t c o s θ . For unit a ∈ C 2 a\in\mathbb{C}^2 a ∈ C 2 , ⟨ L ( a , 0 ) , ( a , 0 ) ⟩ = ⟨ L 1 a , a ⟩ \langle L(a,0),(a,0)\rangle=\langle L_1a,a\rangle ⟨ L ( a , 0 ) , ( a , 0 )⟩ = ⟨ L 1 a , a ⟩ , and W ( L 1 ) W(L_1) W ( L 1 ) is the segment [ λ − , λ + ] [\lambda_-,\lambda_+] [ λ − , λ + ] , which contains cos θ ∈ Σ α \cos\theta\in\Sigma_\alpha cos θ ∈ Σ α . With Y : = μ Y 0 Y:=\mu Y_0 Y := μ Y 0 , one checks directly that T : a ↦ ( a , Y a ) T:a\mapsto(a,Ya) T : a ↦ ( a , Y a ) maps C 2 \mathbb{C}^2 C 2 onto the invariant subspace for λ ± \lambda_\pm λ ± , that { 0 } ⊕ L 2 ( 0 , π ) \{0\}\oplus L^2(0,\pi) { 0 } ⊕ L 2 ( 0 , π ) is invariant, and that P = T V P=TV P = T V with V ( a , v ) : = a V(a,v):=a V ( a , v ) := a , so that P P P is the Riesz projection, L P T = T L 1 L_PT=TL_1 L P T = T L 1 and L Q = A L_Q=A L Q = A on Q H = { 0 } ⊕ L 2 ( 0 , π ) QH=\{0\}\oplus L^2(0,\pi) Q H = { 0 } ⊕ L 2 ( 0 , π ) .
(a) In the decomposition H = C 2 ⊕ L 2 H=\mathbb{C}^2\oplus L^2 H = C 2 ⊕ L 2 , P = ( I 0 Y 0 ) P=\begin{pmatrix}I&0\\Y&0\end{pmatrix} P = ( I Y 0 0 ) and Q = ( 0 0 − Y I ) Q=\begin{pmatrix}0&0\\-Y&I\end{pmatrix} Q = ( 0 − Y 0 I ) , whose norms both equal ( 1 + ∥ Y ∥ 2 ) 1 2 (1+\|Y\|^2)^{\frac12} ( 1 + ∥ Y ∥ 2 ) 2 1 . Since A A A is self-adjoint and negative, ∥ E α ( t α A ) ∥ ≤ 1 \|E_\alpha(t^\alpha A)\|\le1 ∥ E α ( t α A ) ∥ ≤ 1 with equality at t = 0 t=0 t = 0 , and E α ( t α L ) Q = ( 0 0 − E α ( t α A ) Y E α ( t α A ) ) E_\alpha(t^\alpha L)Q=\begin{pmatrix}0&0\\-E_\alpha(t^\alpha A)Y&E_\alpha(t^\alpha A)\end{pmatrix} E α ( t α L ) Q = ( 0 − E α ( t α A ) Y 0 E α ( t α A ) ) has norm at most ∥ Q ∥ \|Q\| ∥ Q ∥ , with equality at t = 0 t=0 t = 0 .
(b) The lower bound is Theorem 3.9(c). Since ( z − L ) − 1 P = T ( z − L 1 ) − 1 V (z-L)^{-1}P=T(z-L_1)^{-1}V ( z − L ) − 1 P = T ( z − L 1 ) − 1 V , ∥ V ∥ = 1 \|V\|=1 ∥ V ∥ = 1 and ∥ T ∥ = ∥ P ∥ \|T\|=\|P\| ∥ T ∥ = ∥ P ∥ , the upper bound follows. On P H PH P H , with the norm of H H H , L P = T L 1 T − 1 L_P=TL_1T^{-1} L P = T L 1 T − 1 , where ∥ T ∥ = ( 1 + μ 2 y 0 2 ) 1 2 \|T\|=(1+\mu^2y_0^2)^{\frac12} ∥ T ∥ = ( 1 + μ 2 y 0 2 ) 2 1 and ∥ T − 1 ∥ = ( 1 + μ 2 s 0 2 ) − 1 2 \|T^{-1}\|=(1+\mu^2s_0^2)^{-\frac12} ∥ T − 1 ∥ = ( 1 + μ 2 s 0 2 ) − 2 1 ; hence K α ( L P ) ≤ ∥ T ∥ ∥ T − 1 ∥ K α ( L 1 ) \mathcal{K}_\alpha(L_P)\le\|T\|\|T^{-1}\|\mathcal{K}_\alpha(L_1) K α ( L P ) ≤ ∥ T ∥∥ T − 1 ∥ K α ( L 1 ) , and 1 + μ 2 y 0 2 1 + μ 2 s 0 2 ≤ y 0 2 s 0 2 \frac{1+\mu^2y_0^2}{1+\mu^2s_0^2}\le\frac{y_0^2}{s_0^2} 1 + μ 2 s 0 2 1 + μ 2 y 0 2 ≤ s 0 2 y 0 2 because y 0 ≥ s 0 y_0\ge s_0 y 0 ≥ s 0 .
(c) Since a α ( 1 ) = 1 a_\alpha(1)=1 a α ( 1 ) = 1 and ( 1 − L ) − 1 ( a , 0 ) = ( ( 1 − L 1 ) − 1 a , μ ( 1 − A ) − 1 S ( 1 − L 1 ) − 1 a ) (1-L)^{-1}(a,0)=\bigl((1-L_1)^{-1}a,\mu(1-A)^{-1}S(1-L_1)^{-1}a\bigr) ( 1 − L ) − 1 ( a , 0 ) = ( ( 1 − L 1 ) − 1 a , μ ( 1 − A ) − 1 S ( 1 − L 1 ) − 1 a ) , we get K α ( L ) ≥ ∥ ( 1 − L ) − 1 ∥ ≥ c 1 μ \mathcal{K}_\alpha(L)\ge\|(1-L)^{-1}\|\ge c_1\mu K α ( L ) ≥ ∥ ( 1 − L ) − 1 ∥ ≥ c 1 μ , and c 1 > 0 c_1>0 c 1 > 0 because ( 1 − A ) − 1 (1-A)^{-1} ( 1 − A ) − 1 is injective and S ≠ 0 S\ne0 S = 0 . The middle inequality is Theorem 5.1 , and the upper bound is Theorem 7.1 (ii) with κ α , 2 ′ \kappa'_{\alpha,2} κ α , 2 ′ , since 0 ∉ σ ( L P ) 0\notin\sigma(L_P) 0 ∈ / σ ( L P ) , together with (a) and (b). □ \square □
The injectivity of Y 0 Y_0 Y 0 is a nondegeneracy condition on the source profiles. For the profiles b 1 = π − 1 / 2 1 ( 0 , π ) b_1=\pi^{-1/2}\mathbf{1}_{(0,\pi)} b 1 = π − 1/2 1 ( 0 , π ) and b 2 b_2 b 2 the normalised parabola x ( π − x ) x(\pi-x) x ( π − x ) used in § 11.1 , a computation with 2 ⋅ 10 4 2\cdot10^4 2 ⋅ 1 0 4 sine modes gives y 0 ≈ 0.78 y_0\approx0.78 y 0 ≈ 0.78 and s 0 ≈ 0.030 s_0\approx0.030 s 0 ≈ 0.030 for θ = π 3 \theta=\frac{\pi}{3} θ = 3 π , so the bound of (b) on K α ( L P ) \mathcal{K}_\alpha(L_P) K α ( L P ) is about 26 K α ( L 1 ) 26\mathcal{K}_\alpha(L_1) 26 K α ( L 1 ) , uniformly in μ \mu μ . For eigenfunction profiles b 1 = 2 π sin x b_1=\sqrt{\frac{2}{\pi}}\sin x b 1 = π 2 sin x and b 2 = 2 π sin 2 x b_2=\sqrt{\frac{2}{\pi}}\sin2x b 2 = π 2 sin 2 x injectivity follows from an explicit 2 × 2 2\times2 2 × 2 determinant.
Optimality of the rank dependence
We now show that the rank cannot be removed from Theorem A, even on Hilbert spaces, and that in infinite dimensions a finite fractional Kreiss constant implies neither boundedness of the family nor even its existence. All examples live on one fixed Hilbert space of functions on the circle, and they are truncations of one natural operator, the Weyl fractional derivative.
On the boundary rays of Σ α \Sigma_\alpha Σ α the Mittag–Leffler function behaves like an exponential of an imaginary argument: E α ( ( i s ) α ) = α − 1 e i s + b α ( s ) E_\alpha((is)^\alpha)=\alpha^{-1}e^{is}+b_\alpha(s) E α (( i s ) α ) = α − 1 e i s + b α ( s ) , with an error b α b_\alpha b α that is bounded and of bounded variation (see (23) below). Consequently, if A A A is a Fourier multiplier with symbol i k ik ik , so that e t A e^{tA} e t A is the translation f ↦ f ( ⋅ + t ) f\mapsto f(\cdot+t) f ↦ f ( ⋅ + t ) , and B B B is the multiplier with symbol ( i k ) α (ik)^\alpha ( ik ) α , whose spectrum lies on ∂ Σ α \partial\Sigma_\alpha ∂ Σ α , then
E α ( t α B ) = α − 1 e t A + (multiplier with symbol b α ( t k ) ) . E_\alpha(t^\alpha B)=\alpha^{-1}e^{tA}+\text{(multiplier with symbol }b_\alpha(tk)\text{)}. E α ( t α B ) = α − 1 e t A + (multiplier with symbol b α ( t k ) ) .
On L 2 ( T ) L^2(\mathbb{T}) L 2 ( T ) this is harmless, because translations are unitary, but on a weighted space translation may amplify: if the weight is small near x = 0 x=0 x = 0 and large near x = ± π x=\pm\pi x = ± π , a function concentrated near 0 0 0 , such as the Dirichlet kernel, has a small norm, while its translate by π \pi π is concentrated where the weight is large and has a large norm. Translation by π \pi π thus amplifies, and this amplification survives in E α ( t α B ) . E_\alpha(t^\alpha B). E α ( t α B ) . On the other hand, the resolvent of B B B is a multiplier whose symbol ( z − ( i k ) α ) − 1 (z-(ik)^\alpha)^{-1} ( z − ( ik ) α ) − 1 varies slowly along the boundary rays, and on weighted spaces with weights of Muckenhoupt class A 2 A_2 A 2 such multipliers are controlled by the Riesz projections, uniformly in the number of frequencies. The Kreiss constant stays bounded while the propagator grows, and the growth is limited only by the weight, which must remain in A 2 A_2 A 2 . For α = 1 \alpha=1 α = 1 this is the mechanism of Eisner and Zwart [13 ] .
We first identify T = R / 2 π Z \mathbb{T}=\mathbb{R}/2\pi\mathbb{Z} T = R /2 π Z with ( − π , π ] (-\pi,\pi] ( − π , π ] , write e k ( x ) : = e i k x e_k(x):=e^{ikx} e k ( x ) := e ik x for k ∈ Z k\in\mathbb{Z} k ∈ Z , and let P \mathcal{P} P be the space of trigonometric polynomials. For 0 < γ < 1 0<\gamma<1 0 < γ < 1 let
w γ ( x ) : = ∣ tan ( x 2 ) ∣ γ , L γ 2 : = L 2 ( T , w γ d x ) , ∥ f ∥ γ 2 : = ∫ − π π ∣ f ( x ) ∣ 2 w γ ( x ) d x . w_\gamma(x):=\left|\tan\left(\frac{x}{2}\right)\right|^\gamma,\qquad L_\gamma^2:=L^2(\mathbb{T},w_\gamma\,dx),\qquad\|f\|_\gamma^2:=\int_{-\pi}^{\pi}|f(x)|^2w_\gamma(x)\,dx. w γ ( x ) := tan ( 2 x ) γ , L γ 2 := L 2 ( T , w γ d x ) , ∥ f ∥ γ 2 := ∫ − π π ∣ f ( x ) ∣ 2 w γ ( x ) d x .
The weight behaves like ∣ x 2 ∣ γ \left|\frac{x}{2}\right|^\gamma 2 x γ near 0 0 0 and like ( 2 π − ∣ x ∣ ) γ \left(\frac{2}{\pi-|x|}\right)^\gamma ( π − ∣ x ∣ 2 ) γ near ± π \pm\pi ± π . Since γ < 1 \gamma<1 γ < 1 , both w γ w_\gamma w γ and w γ − 1 w_\gamma^{-1} w γ − 1 are integrable, so P ⊂ L γ 2 ⊂ L 1 ( T ) \mathcal{P}\subset L_\gamma^2\subset L^1(\mathbb{T}) P ⊂ L γ 2 ⊂ L 1 ( T ) by the Cauchy–Schwarz inequality, Fourier coefficients f ^ ( k ) \widehat f(k) f ( k ) of f ∈ L γ 2 f\in L_\gamma^2 f ∈ L γ 2 are defined and depend continuously on f f f , and P \mathcal{P} P is dense in L γ 2 L_\gamma^2 L γ 2 . For j ∈ Z j\in\mathbb{Z} j ∈ Z let R j R_j R j be the Riesz projection ∑ k c k e k ↦ ∑ k ≥ j c k e k \sum_k c_ke_k\mapsto\sum_{k\geq j}c_ke_k ∑ k c k e k ↦ ∑ k ≥ j c k e k on P \mathcal{P} P , and put
r γ : = sup j ∈ Z sup { ∥ R j f ∥ γ : f ∈ P , ∥ f ∥ γ ≤ 1 } . r_\gamma:=\sup_{j\in\mathbb{Z}}\sup\{\|R_jf\|_\gamma:f\in\mathcal{P},\ \|f\|_\gamma\leq1\}. r γ := j ∈ Z sup sup { ∥ R j f ∥ γ : f ∈ P , ∥ f ∥ γ ≤ 1 } .
For a sequence m = ( m k ) k ∈ Z m=(m_k)_{k\in\mathbb{Z}} m = ( m k ) k ∈ Z of bounded variation, Var Z ( m ) : = ∑ k ∈ Z ∣ m k − m k − 1 ∣ < ∞ \operatorname{Var}_{\mathbb{Z}}(m):=\sum_{k\in\mathbb{Z}}|m_k-m_{k-1}|<\infty Var Z ( m ) := ∑ k ∈ Z ∣ m k − m k − 1 ∣ < ∞ , the limits m ± ∞ m_{\pm\infty} m ± ∞ exist, and we let T m T_m T m denote the Fourier multiplier T m e k : = m k e k T_me_k:=m_ke_k T m e k := m k e k on P \mathcal{P} P . The following lemma contains everything we need about multipliers on L γ 2 L_\gamma^2 L γ 2 .
Lemma 8.1 (Weighted multipliers). Let 0 < γ < 1 0<\gamma<1 0 < γ < 1 .
(a) w γ ∈ A 2 ( T ) w_\gamma\in A_2(\mathbb{T}) w γ ∈ A 2 ( T ) and r γ < ∞ r_\gamma<\infty r γ < ∞ . Consequently every R j R_j R j extends to a bounded projection on L γ 2 L_\gamma^2 L γ 2 , and so does the Fourier partial sum projection S N : = R − N − R N + 1 S_N:=R_{-N}-R_{N+1} S N := R − N − R N + 1 onto V N : = span { e k : ∣ k ∣ ≤ N } V_N:=\operatorname{span}\{e_k:|k|\leq N\} V N := span { e k : ∣ k ∣ ≤ N } , with ∥ S N ∥ ≤ 2 r γ \|S_N\|\leq2r_\gamma ∥ S N ∥ ≤ 2 r γ .
(b) If Var Z ( m ) < ∞ \operatorname{Var}_{\mathbb{Z}}(m)<\infty Var Z ( m ) < ∞ , then ∥ T m f ∥ γ ≤ ( ∣ m − ∞ ∣ + r γ Var Z ( m ) ) ∥ f ∥ γ \|T_m f\|_\gamma\leq\left(|m_{-\infty}|+r_\gamma\operatorname{Var}_{\mathbb{Z}}(m)\right)\|f\|_\gamma ∥ T m f ∥ γ ≤ ( ∣ m − ∞ ∣ + r γ Var Z ( m ) ) ∥ f ∥ γ for f ∈ P f\in\mathcal{P} f ∈ P , so T m T_m T m extends to a bounded operator on L γ 2 L^2_\gamma L γ 2 with T m f ^ ( k ) = m k f ^ ( k ) \widehat{T_m f}(k)=m_k\widehat{f}(k) T m f ( k ) = m k f ( k ) .
(c) If m k = g ( ζ k ) m_k=g(\zeta_k) m k = g ( ζ k ) , where the points ζ k \zeta_k ζ k lie in the order of k k k along a locally rectifiable path Γ \Gamma Γ , possibly unbounded, and g g g is holomorphic in a neighbourhood of Γ \Gamma Γ , then Var Z ( m ) ≤ ∫ Γ ∣ g ′ ( ζ ) ∣ ∣ d ζ ∣ \operatorname{Var}_{\mathbb{Z}}(m)\leq\int_\Gamma|g'(\zeta)|\,|d\zeta| Var Z ( m ) ≤ ∫ Γ ∣ g ′ ( ζ ) ∣ ∣ d ζ ∣ .
Proof. (a) Near 0 0 0 and near ± π \pm\pi ± π the weight is comparable to the power weights ∣ x ∣ γ |x|^\gamma ∣ x ∣ γ and ∣ x ∓ π ∣ − γ |x\mp\pi|^{-\gamma} ∣ x ∓ π ∣ − γ , which belong to A 2 A_2 A 2 since γ < 1 \gamma<1 γ < 1 , and the integrability of w γ ± 1 w_\gamma^{\pm1} w γ ± 1 controls long arcs, so w γ ∈ A 2 ( T ) w_\gamma\in A_2(\mathbb{T}) w γ ∈ A 2 ( T ) . The theorem of Hunt et al. [19 ] gives the boundedness of the conjugate function f ↦ f ~ f\mapsto\widetilde{f} f ↦ f on L γ 2 L^2_\gamma L γ 2 , hence of R 0 f = 1 2 ( f + i f ~ ) + 1 2 f ^ ( 0 ) e 0 R_0f=\frac{1}{2}(f+i\widetilde{f})+\frac{1}{2}\widehat{f}(0)e_0 R 0 f = 2 1 ( f + i f ) + 2 1 f ( 0 ) e 0 ; since multiplication by e j e_j e j is an isometry of L γ 2 L^2_\gamma L γ 2 and R j = e j R 0 e − j R_j=e_jR_0e_{-j} R j = e j R 0 e − j , the bound is uniform in j j j . The details are routine and given in Section Appendix C.
(b) Summation by parts gives T m f = m − ∞ f + ∑ j ∈ Z ( m j − m j − 1 ) R j f T_mf=m_{-\infty}f+\sum_{j\in\mathbb{Z}}(m_j-m_{j-1})R_jf T m f = m − ∞ f + ∑ j ∈ Z ( m j − m j − 1 ) R j f for f ∈ P f\in\mathcal{P} f ∈ P , a series that converges absolutely because Var Z ( m ) < ∞ \operatorname{Var}_{\mathbb{Z}}(m)<\infty Var Z ( m ) < ∞ and ∥ R j f ∥ γ ≤ r γ ∥ f ∥ γ \|R_jf\|_\gamma\leq r_\gamma\|f\|_\gamma ∥ R j f ∥ γ ≤ r γ ∥ f ∥ γ . To verify the identity, note that R j f = f R_jf=f R j f = f for j ≤ min { k : f ^ ( k ) ≠ 0 } j\leq\min\{k:\widehat{f}(k)\neq0\} j ≤ min { k : f ( k ) = 0 } and R j f = 0 R_jf=0 R j f = 0 for j > max { k : f ^ ( k ) ≠ 0 } j>\max\{k:\widehat{f}(k)\neq0\} j > max { k : f ( k ) = 0 } , so the terms with small j j j telescope. This gives the bound; the formula for the Fourier coefficients follows by continuity.
(c) ∣ g ( ζ j ) − g ( ζ j − 1 ) ∣ ≤ ∫ Γ j ∣ g ′ ∣ ∣ d ζ ∣ |g(\zeta_j)-g(\zeta_{j-1})|\leq\int_{\Gamma_j}|g'|\,|d\zeta| ∣ g ( ζ j ) − g ( ζ j − 1 ) ∣ ≤ ∫ Γ j ∣ g ′ ∣ ∣ d ζ ∣ , where Γ j \Gamma_j Γ j is the part of Γ \Gamma Γ between ζ j − 1 \zeta_{j-1} ζ j − 1 and ζ j \zeta_j ζ j .
The symbols we need are evaluations of resolvents along the boundary of Σ α \Sigma_\alpha Σ α , and their variation is controlled by an elementary estimate (Section Appendix C): for a ray R θ : = { s e i θ : s ≥ 0 } R_\theta:=\{se^{i\theta}:s\geq0\} R θ := { s e i θ : s ≥ 0 } and z ∉ R θ z\notin R_\theta z ∈ / R θ ,
∫ 0 ∞ d s ∣ z − s e i θ ∣ 2 ≤ π dist ( z , R θ ) . (22) \int_0^\infty\frac{ds}{|z-se^{i\theta}|^2}\leq\frac{\pi}{\operatorname{dist}(z,R_\theta)}. \tag*{(22)} ∫ 0 ∞ ∣ z − s e i θ ∣ 2 d s ≤ dist ( z , R θ ) π . (22)
Next, the construction needs two quantitative facts, one about the Mittag–Leffler function and one about the weight. Both are direct computations, and we state them here and give the details in Section Appendix C.
The first one makes the heuristic of the previous subsection precise. Let Φ α ( s ) : = E α ( ( i s ) α ) \Phi_\alpha(s):=E_\alpha((is)^\alpha) Φ α ( s ) := E α (( i s ) α ) for s ∈ R s\in\mathbb{R} s ∈ R and b α ( s ) : = Φ α ( s ) − α − 1 e i s b_\alpha(s):=\Phi_\alpha(s)-\alpha^{-1}e^{is} b α ( s ) := Φ α ( s ) − α − 1 e i s . Then b 1 ≡ 0 b_1\equiv0 b 1 ≡ 0 , so that b 1 ∗ = 0 b_1^*=0 b 1 ∗ = 0 in the notation below, and for 0 < α < 1 0<\alpha<1 0 < α < 1 the function b α b_\alpha b α is continuous with
b α ∗ : = sup R ∣ b α ∣ + Var R ( b α ) < ∞ . (23) b_\alpha^*:=\sup_{\mathbb{R}}|b_\alpha|+\operatorname{Var}_{\mathbb{R}}(b_\alpha)<\infty. \tag*{(23)} b α ∗ := R sup ∣ b α ∣ + Var R ( b α ) < ∞. (23)
For large ∣ s ∣ |s| ∣ s ∣ this follows from Theorem Appendix B.1(c), which gives ∣ b α ( s ) ∣ = O ( ∣ s ∣ − α ) |b_\alpha(s)|=O(|s|^{-\alpha}) ∣ b α ( s ) ∣ = O ( ∣ s ∣ − α ) and ∣ b α ′ ( s ) ∣ = O ( ∣ s ∣ − 1 − α ) |b_\alpha'(s)|=O(|s|^{-1-\alpha}) ∣ b α ′ ( s ) ∣ = O ( ∣ s ∣ − 1 − α ) on the boundary ray, and for small ∣ s ∣ |s| ∣ s ∣ from the power series.
The second one quantifies the amplification produced by translation. We use the Dirichlet kernel
D N : = ∑ ∣ k ∣ ≤ N e k , D N ( x ) = sin ( ( N + 1 2 ) x ) sin ( x 2 ) . D_N:=\sum_{|k|\leq N}e_k,\qquad D_N(x)=\frac{\sin((N+\frac{1}{2})x)}{\sin(\frac{x}{2})}. D N := ∣ k ∣ ≤ N ∑ e k , D N ( x ) = sin ( 2 x ) sin (( N + 2 1 ) x ) .
Since ∣ D N ∣ |D_N| ∣ D N ∣ is of size N N N on an interval of length about 1 N \frac{1}{N} N 1 around 0 0 0 and bounded by ∣ sin ( x 2 ) ∣ − 1 |\sin(\frac{x}{2})|^{-1} ∣ sin ( 2 x ) ∣ − 1 elsewhere, one finds ∥ D N ∥ γ 2 ≍ N 1 − γ \|D_N\|_\gamma^2\asymp N^{1-\gamma} ∥ D N ∥ γ 2 ≍ N 1 − γ , whereas the translate places this peak near ± π \pm\pi ± π , where the weight is of size N γ N^\gamma N γ , so that ∥ D N ( ⋅ + π ) ∥ γ 2 ≳ N 1 + γ \lVert D_N(\cdot+\pi)\rVert_\gamma^2 \gtrsim N^{1+\gamma} ∥ D N ( ⋅ + π ) ∥ γ 2 ≳ N 1 + γ . Estimating the two weighted integrals gives, for all N ∈ N N\in\mathbb{N} N ∈ N ,
∥ D N ( ⋅ + π ) ∥ γ ≥ c γ N γ ∥ D N ∥ γ , c γ : = ( 2 ⋅ 3 2 γ − 2 ( 1 − γ ) Q γ ) 1 2 , (24) \lVert D_N(\cdot+\pi)\rVert_\gamma\ge c_\gamma N^\gamma\lVert D_N\rVert_\gamma,\qquad c_\gamma:=\left(\frac{2\cdot3^{2\gamma-2}}{(1-\gamma)Q_\gamma}\right)^{\frac{1}{2}}, \tag*{(24)} ∥ D N ( ⋅ + π ) ∥ γ ≥ c γ N γ ∥ D N ∥ γ , c γ := ( ( 1 − γ ) Q γ 2 ⋅ 3 2 γ − 2 ) 2 1 , (24)
where Q γ : = 2 1 + γ + 2 π 2 1 − γ + 4 ∫ π 2 π tan γ ( y 2 ) d y Q_\gamma:=\frac{2}{1+\gamma}+\frac{2\pi^2}{1-\gamma}+4\int_{\frac{\pi}{2}}^\pi\tan^\gamma\left(\frac{y}{2}\right)\,dy Q γ := 1 + γ 2 + 1 − γ 2 π 2 + 4 ∫ 2 π π tan γ ( 2 y ) d y .
Finally, we proceed to the construction. Fix α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] and γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) . The Weyl fractional derivative of order α \alpha α on L γ 2 L_\gamma^2 L γ 2 is the Fourier multiplier with symbol ( i k ) α = ∣ k ∣ α e i sgn ( k ) α π 2 (ik)^\alpha=|k|^\alpha e^{i\operatorname{sgn}(k)\frac{\alpha\pi}{2}} ( ik ) α = ∣ k ∣ α e i sgn ( k ) 2 α π ,
D α f : = g whenever f , g ∈ L γ 2 and g ^ ( k ) = ( i k ) α f ^ ( k ) for all k ∈ Z , D^\alpha f:=g\quad\text{whenever }f,g\in L_\gamma^2\text{ and }\widehat{g}(k)=(ik)^\alpha\widehat{f}(k)\text{ for all }k\in\mathbb{Z}, D α f := g whenever f , g ∈ L γ 2 and g ( k ) = ( ik ) α f ( k ) for all k ∈ Z ,
a closed, densely defined operator, since the Fourier coefficients are continuous on L γ 2 L_\gamma^2 L γ 2 and P ⊂ D ( D α ) \mathcal{P}\subset D(D^\alpha) P ⊂ D ( D α ) . Its eigenvalues ( i k ) α (ik)^\alpha ( ik ) α lie on the two boundary rays of Σ α \Sigma_\alpha Σ α , and Σ α ⊂ ρ ( D α ) \Sigma_\alpha\subset\rho(D^\alpha) Σ α ⊂ ρ ( D α ) by Theorem 8.5 below. For N ∈ N N\in\mathbb{N} N ∈ N we let
B N : = N − α D α S N , ℓ N ( k ) : = { ( i k N ) α , ∣ k ∣ ≤ N , 0 , ∣ k ∣ > N , B_N:=N^{-\alpha}D^\alpha S_N,\qquad\ell_N(k):=
\begin{cases}
\left(\frac{ik}{N}\right)^\alpha,& |k|\le N,\\
0,& |k|>N,
\end{cases} B N := N − α D α S N , ℓ N ( k ) := { ( N ik ) α , 0 , ∣ k ∣ ≤ N , ∣ k ∣ > N ,
so that B N B_N B N is the multiplier with symbol ℓ N \ell_N ℓ N . It is a bounded operator on L γ 2 L_\gamma^2 L γ 2 of rank 2 N 2N 2 N , since ℓ N ( 0 ) = 0 \ell_N(0)=0 ℓ N ( 0 ) = 0 .
Theorem 8.2. Let 0 < α ≤ 1 0<\alpha\le1 0 < α ≤ 1 and 0 < γ < 1 0<\gamma<1 0 < γ < 1 . For every N ∈ N N\in\mathbb{N} N ∈ N , the operator B N ∈ B ( L γ 2 ) B_N\in\mathcal{B}(L_\gamma^2) B N ∈ B ( L γ 2 ) has rank 2 N 2N 2 N and satisfies
(i) σ ( B N ) = { ℓ N ( k ) : ∣ k ∣ ≤ N } ⊂ ∂ Σ α \sigma(B_N)=\{\ell_N(k):|k|\le N\}\subset\partial\Sigma_\alpha σ ( B N ) = { ℓ N ( k ) : ∣ k ∣ ≤ N } ⊂ ∂ Σ α and ∥ B N ∥ ≤ 4 r γ \lVert B_N\rVert\le4r_\gamma ∥ B N ∥ ≤ 4 r γ ;
(ii) K α ( B N ) ≤ 1 + 4 π r γ α K_\alpha(B_N)\le\frac{1+4\pi r_\gamma}{\alpha} K α ( B N ) ≤ α 1 + 4 π r γ ;
(iii) ∥ E α ( ( π N ) α B N ) ∥ ≥ α − 1 c γ N γ − ( 1 + r γ ) b α ∗ \lVert E_\alpha((\pi N)^\alpha B_N)\rVert\ge\alpha^{-1}c_\gamma N^\gamma-(1+r_\gamma)b_\alpha^* ∥ E α (( π N ) α B N )∥ ≥ α − 1 c γ N γ − ( 1 + r γ ) b α ∗ .
Proof. (i) B N B_N B N is diagonal in the basis ( e k ) (e_k) ( e k ) of P \mathcal{P} P with eigenvalues ℓ N ( k ) \ell_N(k) ℓ N ( k ) , and ℓ N ( k ) = 0 \ell_N(k)=0 ℓ N ( k ) = 0 for k = 0 k=0 k = 0 and for ∣ k ∣ > N |k|>N ∣ k ∣ > N . As k k k increases from − ∞ -\infty − ∞ to ∞ \infty ∞ , the points ℓ N ( k ) \ell_N(k) ℓ N ( k ) stay at 0 0 0 , jump to e − i α π 2 e^{-i\frac{\alpha\pi}{2}} e − i 2 α π at k = − N k=-N k = − N , move along ∂ Σ α \partial\Sigma_\alpha ∂ Σ α to 0 0 0 at k = 0 k=0 k = 0 and on to e i α π 2 e^{i\frac{\alpha\pi}{2}} e i 2 α π at k = N k=N k = N , and jump back to 0 0 0 . They lie in this order on the closed path Γ \Gamma Γ that runs along the segment [ 0 , e − i α π 2 ] [0,e^{-i\frac{\alpha\pi}{2}}] [ 0 , e − i 2 α π ] and the segment [ 0 , e i α π 2 ] [0,e^{i\frac{\alpha\pi}{2}}] [ 0 , e i 2 α π ] twice each, which has length 4. Theorem 8.1 (b),(c) with g ( ζ ) = ζ g(\zeta)=\zeta g ( ζ ) = ζ and m − ∞ = 0 m_{-\infty}=0 m − ∞ = 0 gives ∥ B N ∥ ≤ 4 r γ \lVert B_N\rVert\le4r_\gamma ∥ B N ∥ ≤ 4 r γ . If z ∉ { ℓ N ( k ) : ∣ k ∣ ≤ N } z\notin\{\ell_N(k):|k|\le N\} z ∈ / { ℓ N ( k ) : ∣ k ∣ ≤ N } , the symbol ( z − ℓ N ( k ) ) − 1 (z-\ell_N(k))^{-1} ( z − ℓ N ( k ) ) − 1 takes finitely many values and is constant for ∣ k ∣ > N |k|>N ∣ k ∣ > N , so it has finite variation, and by Theorem 8.1 (b) the corresponding multiplier inverts z − B N z-B_N z − B N ; this gives the spectrum.
(ii) Let z ∈ Σ α z\in\Sigma_\alpha z ∈ Σ α . Then z − B N z-B_N z − B N is the multiplier with symbol z − ℓ N ( k ) ≠ 0 z-\ell_N(k)\ne0 z − ℓ N ( k ) = 0 , and its inverse is T m T_m T m with m k = g ( ℓ N ( k ) ) m_k=g(\ell_N(k)) m k = g ( ℓ N ( k )) , g ( ζ ) : = ( z − ζ ) − 1 g(\zeta):=(z-\zeta)^{-1} g ( ζ ) := ( z − ζ ) − 1 , which is holomorphic near Γ ⊂ ∂ Σ α \Gamma\subset\partial\Sigma_\alpha Γ ⊂ ∂ Σ α . Here ∣ m − ∞ ∣ = ∣ z ∣ − 1 ≤ dist ( z , ∂ Σ α ) − 1 |m_{-\infty}|=|z|^{-1}\le\operatorname{dist}(z,\partial\Sigma_\alpha)^{-1} ∣ m − ∞ ∣ = ∣ z ∣ − 1 ≤ dist ( z , ∂ Σ α ) − 1 , and by Theorem 8.1 (c) and (22) , each segment of Γ \Gamma Γ being traversed twice,
Var Z ( m ) ≤ ∫ Γ ∣ d ζ ∣ ∣ z − ζ ∣ ≤ 2 π dist ( z , R α π 2 ) + 2 π dist ( z , R − α π 2 ) ≤ 4 π dist ( z , ∂ Σ α ) . \operatorname{Var}_{\mathbb{Z}}(m)\le\int_\Gamma\frac{|d\zeta|}{|z-\zeta|}\le\frac{2\pi}{\operatorname{dist}(z,R_{\frac{\alpha\pi}{2}})}+\frac{2\pi}{\operatorname{dist}(z,R_{-\frac{\alpha\pi}{2}})}\le\frac{4\pi}{\operatorname{dist}(z,\partial\Sigma_\alpha)}. Var Z ( m ) ≤ ∫ Γ ∣ z − ζ ∣ ∣ d ζ ∣ ≤ dist ( z , R 2 α π ) 2 π + dist ( z , R − 2 α π ) 2 π ≤ dist ( z , ∂ Σ α ) 4 π .
Hence ∥ ( z − B N ) − 1 ∥ ≤ 1 + 4 π r γ dist ( z , ∂ Σ α ) \|(z-B_N)^{-1}\| \le\frac{1+4\pi r_\gamma}{\operatorname{dist}(z,\partial\Sigma_\alpha)} ∥ ( z − B N ) − 1 ∥ ≤ dist ( z , ∂ Σ α ) 1 + 4 π r γ , and Theorem 3.6(c) gives (ii).
(iii) Since B N B_N B N is diagonal, E α ( t α B N ) E_\alpha(t^\alpha B_N) E α ( t α B N ) is the multiplier with symbol E α ( t α ℓ N ( k ) ) E_\alpha(t^\alpha\ell_N(k)) E α ( t α ℓ N ( k )) , and t α ℓ N ( k ) = ( i t k N ) α t^\alpha\ell_N(k)=\left(\frac{itk}{N}\right)^\alpha t α ℓ N ( k ) = ( N i t k ) α for ∣ k ∣ ≤ N |k|\le N ∣ k ∣ ≤ N . On V N V_N V N it therefore acts as the multiplier with symbol Φ α ( t k N ) = α − 1 e i t k N + b α ( t k N ) \Phi_\alpha\left(\frac{tk}{N}\right)=\alpha^{-1}e^{\frac{itk}{N}}+b_\alpha\left(\frac{tk}{N}\right) Φ α ( N t k ) = α − 1 e N i t k + b α ( N t k ) . The first part maps D N D_N D N to α − 1 D N ( ⋅ + t N ) \alpha^{-1}D_N\left(\cdot+\frac{t}{N}\right) α − 1 D N ( ⋅ + N t ) . For the second, the points t k N \frac{tk}{N} N t k , ∣ k ∣ ≤ N |k|\le N ∣ k ∣ ≤ N , increase with k k k , so the restriction of the symbol to ∣ k ∣ ≤ N |k|\le N ∣ k ∣ ≤ N , extended by constants outside, has modulus at most sup ∣ b α ∣ \sup|b_\alpha| sup ∣ b α ∣ and variation at most Var R ( b α ) \operatorname{Var}_{\mathbb R}(b_\alpha) Var R ( b α ) ; by Theorem 8.1 (a),(b) and (22) it maps D N D_N D N to a function of norm at most ( 1 + r γ ) b α ∗ ∥ D N ∥ γ (1+r_\gamma)b_\alpha^*\|D_N\|_\gamma ( 1 + r γ ) b α ∗ ∥ D N ∥ γ . Taking t = π N t=\pi N t = π N and using (23), ∥ E α ( ( π N ) α B N ) D N ∥ γ ≥ ( α − 1 c γ N γ − ( 1 + r γ ) b α ∗ ) ∥ D N ∥ γ \|E_\alpha((\pi N)^\alpha B_N)D_N\|_\gamma\ge\left(\alpha^{-1}c_\gamma N^\gamma-(1+r_\gamma)b_\alpha^*\right)\|D_N\|_\gamma ∥ E α (( π N ) α B N ) D N ∥ γ ≥ ( α − 1 c γ N γ − ( 1 + r γ ) b α ∗ ) ∥ D N ∥ γ .
For n ∈ N n\in\mathbb N n ∈ N let κ α opt ( n ) \kappa_\alpha^{\operatorname{opt}}(n) κ α opt ( n ) denote the least constant κ \kappa κ such that sup t ∥ E α ( t α L ) ∥ ≤ κ K α ( L ) \sup_t\|E_\alpha(t^\alpha L)\|\le\kappa K_\alpha(L) sup t ∥ E α ( t α L ) ∥ ≤ κ K α ( L ) for all Hilbert spaces H ≠ { 0 } H\ne\{0\} H = { 0 } and all L ∈ B ( H ) L\in B(H) L ∈ B ( H ) with rank L ≤ n \operatorname{rank}L\le n rank L ≤ n . By Theorem 5.4 , κ α opt ( n ) ≤ κ α , n \kappa_\alpha^{\operatorname{opt}}(n)\le\kappa_{\alpha,n} κ α opt ( n ) ≤ κ α , n .
Corollary 8.3. Let α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] and γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) . For every n ≥ 2 n\ge2 n ≥ 2 ,
κ α opt ( n ) ≥ c γ ⌊ n 2 ⌋ γ − α ( 1 + r γ ) b α ∗ 1 + 4 π r γ , in particular lim inf n → ∞ κ α opt ( n ) n γ ≥ 2 − γ c γ 1 + 4 π r γ > 0. \kappa_\alpha^{\operatorname{opt}}(n)\ge\frac{c_\gamma\left\lfloor\frac{n}{2}\right\rfloor^\gamma-\alpha(1+r_\gamma)b_\alpha^*}{1+4\pi r_\gamma},\qquad\text{in particular}\qquad\liminf_{n\to\infty}\frac{\kappa_\alpha^{\operatorname{opt}}(n)}{n^\gamma}\ge\frac{2^{-\gamma}c_\gamma}{1+4\pi r_\gamma}>0. κ α opt ( n ) ≥ 1 + 4 π r γ c γ ⌊ 2 n ⌋ γ − α ( 1 + r γ ) b α ∗ , in particular n → ∞ lim inf n γ κ α opt ( n ) ≥ 1 + 4 π r γ 2 − γ c γ > 0. Hence n γ ≲ κ α opt ( n ) ≲ n n^\gamma\lesssim\kappa_\alpha^{\operatorname{opt}}(n)\lesssim n n γ ≲ κ α opt ( n ) ≲ n for every γ < 1 \gamma<1 γ < 1 .
Proof. Take N : = ⌊ n 2 ⌋ N:=\left\lfloor\frac{n}{2}\right\rfloor N := ⌊ 2 n ⌋ , so that rank B N = 2 N ≤ n \operatorname{rank}B_N=2N\le n rank B N = 2 N ≤ n , and combine Theorem 8.2 (ii) and (iii).
Corollary 8.4. Let H : = ⨁ N ≥ 1 L γ 2 H:=\bigoplus_{N\ge1}L_\gamma^2 H := ⨁ N ≥ 1 L γ 2 (Hilbert direct sum) and L : = ⨁ N ≥ 1 B N L:=\bigoplus_{N\ge1}B_N L := ⨁ N ≥ 1 B N . Then L ∈ B ( H ) L\in B(H) L ∈ B ( H ) , σ ( L ) ∩ Σ α = ∅ \sigma(L)\cap\Sigma_\alpha=\varnothing σ ( L ) ∩ Σ α = ∅ , K α ( L ) ≤ 1 + 4 π r γ α K_\alpha(L)\le\frac{1+4\pi r_\gamma}{\alpha} K α ( L ) ≤ α 1 + 4 π r γ , and
lim sup t → ∞ t − γ ∥ E α ( t α L ) ∥ ≥ α − 1 π − γ c γ > 0. \limsup_{t\to\infty}t^{-\gamma}\|E_\alpha(t^\alpha L)\|\ge\alpha^{-1}\pi^{-\gamma}c_\gamma>0. t → ∞ lim sup t − γ ∥ E α ( t α L ) ∥ ≥ α − 1 π − γ c γ > 0. In particular K α ( L ) < ∞ K_\alpha(L)<\infty K α ( L ) < ∞ does not imply sup t ∥ E α ( t α L ) ∥ < ∞ \sup_t\|E_\alpha(t^\alpha L)\|<\infty sup t ∥ E α ( t α L ) ∥ < ∞ on Hilbert spaces.
Proof. By Theorem 8.2 (i), ∥ L ∥ ≤ 4 r γ \|L\|\le4r_\gamma ∥ L ∥ ≤ 4 r γ . For z ∈ Σ α z\in\Sigma_\alpha z ∈ Σ α the inverses ( z − B N ) − 1 (z-B_N)^{-1} ( z − B N ) − 1 are uniformly bounded by 1 + 4 π r γ dist ( z , ∂ Σ α ) \frac{1+4\pi r_\gamma}{\operatorname{dist}(z,\partial\Sigma_\alpha)} dist ( z , ∂ Σ α ) 1 + 4 π r γ , as shown in the proof of Theorem 8.2 (ii), so their direct sum inverts z − L z-L z − L and the bound on K α ( L ) K_\alpha(L) K α ( L ) follows. Since E α ( t α L ) = ⨁ N E α ( t α B N ) E_\alpha(t^\alpha L)=\bigoplus_N E_\alpha(t^\alpha B_N) E α ( t α L ) = ⨁ N E α ( t α B N ) , the choice t = π N t=\pi N t = π N in Theorem 8.2 (iii) gives the claim.
The untruncated operator shows that in infinite dimensions the fractional Kreiss condition does not even guarantee the existence of the family.
Proposition 8.5. Let 0 < α ≤ 1 0<\alpha\le1 0 < α ≤ 1 and 0 < γ < 1 0<\gamma<1 0 < γ < 1 . Then Σ α ⊂ ρ ( D α ) \Sigma_\alpha\subset\rho(D^\alpha) Σ α ⊂ ρ ( D α ) and K α ( D α ) ≤ 2 π r γ α K_\alpha(D^\alpha)\le\frac{2\pi r_\gamma}{\alpha} K α ( D α ) ≤ α 2 π r γ on L γ 2 L_\gamma^2 L γ 2 , but D α D^\alpha D α generates no Mittag–Leffler family on L γ 2 L_\gamma^2 L γ 2 . For α = 1 \alpha=1 α = 1 , the derivative D = d d x D=\frac{d}{dx} D = d x d satisfies the Kreiss resolvent condition on L γ 2 L_\gamma^2 L γ 2 but generates no C 0 C_0 C 0 -semigroup.
Proof. Let z ∈ Σ α z\in\Sigma_\alpha z ∈ Σ α and m k : = ( z − ( i k ) α ) − 1 m_k:=(z-(ik)^\alpha)^{-1} m k := ( z − ( ik ) α ) − 1 . The points ( i k ) α (ik)^\alpha ( ik ) α run in the order of k k k along the lower boundary ray from infinity to 0 0 0 and then along the upper one to infinity, so m − ∞ = 0 m_{-\infty}=0 m − ∞ = 0 and, by Theorem 8.1 (c) and (21), Var Z ( m ) ≤ 2 π dist ( z , ∂ Σ α ) \operatorname{Var}_{\mathbb Z}(m)\le\frac{2\pi}{\operatorname{dist}(z,\partial\Sigma_\alpha)} Var Z ( m ) ≤ dist ( z , ∂ Σ α ) 2 π . The bounded operator T m T_m T m maps L γ 2 L_\gamma^2 L γ 2 into D ( D α ) D(D^\alpha) D ( D α ) , and comparing Fourier coefficients gives ( z − D α ) T m = I (z-D^\alpha)T_m=I ( z − D α ) T m = I and T m ( z − D α ) f = f T_m(z-D^\alpha)f=f T m ( z − D α ) f = f for f ∈ D ( D α ) f\in D(D^\alpha) f ∈ D ( D α ) . Hence z ∈ ρ ( D α ) z\in\rho(D^\alpha) z ∈ ρ ( D α ) , ∥ ( z − D α ) − 1 ∥ ≤ 2 π r γ dist ( z , ∂ Σ α ) \|(z-D^\alpha)^{-1}\|\le\frac{2\pi r_\gamma}{\operatorname{dist}(z,\partial\Sigma_\alpha)} ∥ ( z − D α ) − 1 ∥ ≤ dist ( z , ∂ Σ α ) 2 π r γ , and Theorem 3.6(c) gives the bound on K α ( D α ) K_\alpha(D^\alpha) K α ( D α ) .
Suppose that D α D^\alpha D α generates a Mittag–Leffler family S S S . For f ∈ P f\in\mathcal{P} f ∈ P and λ \lambda λ large,
∫ 0 ∞ e − λ t S ( t ) f d t = ∑ k λ α − 1 λ α − ( i k ) α f ^ ( k ) e k , \int_0^\infty e^{-\lambda t}S(t)f\,dt=\sum_k\frac{\lambda^{\alpha-1}}{\lambda^\alpha-(ik)^\alpha}\widehat{f}(k)e_k, ∫ 0 ∞ e − λ t S ( t ) f d t = k ∑ λ α − ( ik ) α λ α − 1 f ( k ) e k ,
and the scalar identity ∫ 0 ∞ e − λ t E α ( t α ζ ) d t = λ α − 1 ( λ α − ζ ) − 1 \int_0^\infty e^{-\lambda t}E_\alpha(t^\alpha\zeta)\,dt=\lambda^{\alpha-1}(\lambda^\alpha-\zeta)^{-1} ∫ 0 ∞ e − λ t E α ( t α ζ ) d t = λ α − 1 ( λ α − ζ ) − 1 for ζ ∈ Λ α \zeta\in\Lambda_\alpha ζ ∈ Λ α (proof of Theorem 6.1) and uniqueness of the Laplace transform give
S ( t ) f = ∑ k Φ α ( t k ) f ^ ( k ) e k = α − 1 f ( ⋅ + t ) + T b ( t ) f , b ( t ) k : = b α ( t k ) . S(t)f=\sum_k\Phi_\alpha(tk)\widehat{f}(k)e_k=\alpha^{-1}f(\mathord\cdot+t)+T_b(t)f,\qquad b(t)_k:=b_\alpha(tk). S ( t ) f = k ∑ Φ α ( t k ) f ( k ) e k = α − 1 f ( ⋅ + t ) + T b ( t ) f , b ( t ) k := b α ( t k ) .
By Theorem 8.1 (b) and (23) , ∥ T b ( t ) f ∥ γ ≤ ( 1 + r γ ) b α ∗ ∥ f ∥ γ \|T_{b(t)}f\|_\gamma\le(1+r_\gamma)b^\ast_\alpha\|f\|_\gamma ∥ T b ( t ) f ∥ γ ≤ ( 1 + r γ ) b α ∗ ∥ f ∥ γ , so ∥ f ( ⋅ + π ) ∥ γ ≤ α ( ∥ S ( π ) ∥ + ( 1 + r γ ) b α ∗ ) ∥ f ∥ γ \|f(\mathord\cdot+\pi)\|_\gamma\le\alpha(\|S(\pi)\|+(1+r_\gamma)b^\ast_\alpha)\|f\|_\gamma ∥ f ( ⋅ + π ) ∥ γ ≤ α ( ∥ S ( π ) ∥ + ( 1 + r γ ) b α ∗ ) ∥ f ∥ γ for all f ∈ P f\in\mathcal{P} f ∈ P , which contradicts (24) with f = D N f=D_N f = D N and N → ∞ N\to\infty N → ∞ . For α = 1 \alpha=1 α = 1 , b 1 ≡ 0 b_1\equiv0 b 1 ≡ 0 and S ( t ) S(t) S ( t ) would be the translation by t t t . □
By Theorem 3.6(e), K α ( c L ) = K α ( L ) \mathcal{K}_\alpha(cL)=\mathcal{K}_\alpha(L) K α ( c L ) = K α ( L ) for c > 0 c>0 c > 0 , so the absence of generation could also be produced by rescaling the blocks of Theorem 8.4; Theorem 8.5 shows that it happens for a natural differential operator.
The constants of Theorem 8.2 are far from sharp, and it is instructive to see how the construction behaves in practice. Figure 1 shows the Kreiss constants K α ( B N ) \mathcal{K}_\alpha(B_N) K α ( B N ) and the ratios of the two sides of Theorem A, computed on V N V_N V N through the Gram matrix of the weight, for N N N up to 256, that is, for ranks up to 512. The Kreiss constants stay between 1.4 and 16, while the ratios grow at a rate close to N γ N^\gamma N γ , with fitted exponents 0.49 for γ = 0.5 \gamma=0.5 γ = 0.5 and 0.84 for γ = 0.9 \gamma=0.9 γ = 0.9 , and reach 138 at N = 256 N=256 N = 256 . The curves for different α \alpha α nearly coincide, because the translation term α − 1 D N ( ⋅ + t N ) \alpha^{-1}D_N(\mathord\cdot+\frac{t}{N}) α − 1 D N ( ⋅ + N t ) in the proof of Theorem 8.2 dominates. The amplification is therefore a property of the weighted geometry, which the fractional clock transmits essentially unchanged. The complete data are in Section S4 of the Supplementary Material.
Figure 1. The truncated fractional derivatives B N B_N B N on L γ 2 L^2_\gamma L γ 2 (Theorem 8.2 ). (a) Computed K α ( B N ) \mathcal{K}_\alpha(B_N) K α ( B N ) ; it is bounded by Theorem 8.2 (ii), and for γ = 0.9 \gamma=0.9 γ = 0.9 the computed values are still increasing slowly at N = 256 N=256 N = 256 . (b) The ratio sup t ∥ E α ( t α B N ) ∥ K α ( B N ) \frac{\sup_t\|E_\alpha(t^\alpha B_N)\|}{\mathcal{K}_\alpha(B_N)} K α ( B N ) s u p t ∥ E α ( t α B N ) ∥ grows at a rate close to N γ N^\gamma N γ (dotted guides); the curves for the three values of α \alpha α nearly coincide.
Together with Theorem 5.4 , Theorem 8.3 leaves open only whether the exponent one is attained, which for α = 1 \alpha=1 α = 1 is the case by the examples of Kraaijevanger [28 ] .
Mechanisms of fractional transient growth
The two-sided estimates of Theorem A say how large the transient peak is; this section explains where it comes from and when it occurs. Several of its features have no counterpart for semigroups: normal operators amplify, modes that are unstable for the semigroup are stable for the family, and nilpotent non-normality, the classical source of dimension-dependent growth, is tamed by the Mittag–Leffler distribution for each fixed α < 1 \alpha< 1 α < 1 .
Normal amplification and fractionally stable modes
By Theorem 6.1, a normal generator with spectrum on a boundary ray of Σ α \Sigma_\alpha Σ α amplifies by exactly M α ∗ M_\alpha^* M α ∗ , while its Kreiss constant equals 1 / α 1/\alpha 1/ α , so the ratio of the two sides of Theorem A equals α M α ∗ \alpha M_\alpha^* α M α ∗ . This ratio decreases from 1.33 at α = 0.1 \alpha= 0.1 α = 0.1 to 1 at α = 1 \alpha= 1 α = 1 (Table 1 ). The amplification originates on the boundary ray, where E α ( r e i α π / 2 ) E_\alpha(re^{i\alpha\pi/2}) E α ( r e i α π /2 ) is the sum of the unimodular oscillation α − 1 e i r 1 / α \alpha^{-1}e^{ir^{1/\alpha}} α − 1 e i r 1/ α and a slowly decaying correction of order r − 1 r^{-1} r − 1 , and the two add constructively at moderate r r r (Figure 2 (a)). If the spectrum moves away from the boundary ray into the interior of Λ α \Lambda_\alpha Λ α , the amplification decreases monotonically to 1, which is reached at the imaginary axis (Theorem 3.1 and Figure 2 (b)).
Figure 2. (a) The function s ↦ α ∣ E α ( ( i s ) α ) ∣ s \mapsto\alpha|E_{\alpha}((is)^{\alpha})| s ↦ α ∣ E α (( i s ) α ) ∣ on the boundary ray of Σ α \Sigma_{\alpha} Σ α overshoots 1 before it settles into the oscillation α − 1 e i s \alpha^{-1}e^{is} α − 1 e i s (Theorem 3.1). (b) The amplification M α ( θ ) M_{\alpha}(\theta) M α ( θ ) of normal operators with spectrum on the ray of argument θ \theta θ decreases from M α ∗ M_{\alpha}^{*} M α ∗ at θ = α π 2 \theta= \frac{\alpha\pi}{2} θ = 2 α π to 1 at θ = π 2 \theta= \frac{\pi}{2} θ = 2 π . (c) The Jordan-block limits ϱ n ( α ) \varrho_n(\alpha) ϱ n ( α ) of Theorem 9.1: they grow like e n 2 π e\sqrt{\frac{n}{2\pi}} e 2 π n for α = 1 \alpha= 1 α = 1 and stay below ϑ α ( n n − 1 ) n \vartheta_{\alpha}\left(\frac{n}{n-1}\right)^n ϑ α ( n − 1 n ) n for α < 1 \alpha< 1 α < 1 .
The stability region of the family contains the sector { α π / 2 ≤ ∣ Arg z ∣ < π / 2 } \{\alpha\pi/2 \leq|\operatorname{Arg} z| < \pi/2\} { α π /2 ≤ ∣ Arg z ∣ < π /2 } of the right half-plane, where the semigroup grows exponentially. The simplest example is the oscillator L = ( cos θ − sin θ sin θ cos θ ) L = \begin{pmatrix} \cos\theta& -\sin\theta\\ \sin\theta& \cos\theta\end{pmatrix} L = ( cos θ sin θ − sin θ cos θ ) with α π / 2 < θ < π / 2 \alpha\pi/2 < \theta< \pi/2 α π /2 < θ < π /2 , whose eigenvalues e ± i θ e^{\pm i\theta} e ± i θ have positive real part. Then ∥ e t L ∥ = e t cos θ \lVert e^{tL}\rVert= e^{t\cos\theta} ∥ e t L ∥ = e t c o s θ and K 1 ( L ) = ∞ \mathcal{K}_1(L) = \infty K 1 ( L ) = ∞ , while K α ( L ) ≤ 1 / α \mathcal{K}_\alpha(L) \leq1/\alpha K α ( L ) ≤ 1/ α and sup t ≥ 0 ∥ E α ( t α L ) ∥ = M α ( θ ) ≤ M α ∗ \sup_{t \geq0}\lVert E_\alpha(t^\alpha L)\rVert= M_\alpha(\theta) \leq M_\alpha^* sup t ≥ 0 ∥ E α ( t α L )∥ = M α ( θ ) ≤ M α ∗ , since L L L is normal. Adding a non-normal coupling changes the size of both sides of Theorem A, but not their ratio by much; this is prototype (a), analysed in Theorem 7.4 and computed in § 11.1 . Note that in this situation the numerical range of L L L contains cos θ ∈ Σ α \cos\theta\in\Sigma_\alpha cos θ ∈ Σ α , so neither Theorem 6.4 nor any of the spectral-set bounds discussed after it applies.
Jordan blocks
Let Z n ∈ C n × n Z_n \in\mathbb{C}^{n \times n} Z n ∈ C n × n be the nilpotent shift, with ones on the superdiagonal, and J n , μ : = − I + μ Z n J_{n,\mu} := -I + \mu Z_n J n , μ := − I + μ Z n with μ > 0 \mu> 0 μ > 0 , acting on C n \mathbb{C}^n C n with the Euclidean norm. For α = 1 \alpha= 1 α = 1 these blocks are the standard source of the dimension dependence in (1) : as μ → ∞ \mu\to\infty μ → ∞ , the ratio sup t ∥ e t J n , μ ∥ K 1 ( J n , μ ) \frac{\sup_t \lVert e^{tJ_{n,\mu}}\rVert}{\mathcal{K}_1(J_{n,\mu})} K 1 ( J n , μ ) s u p t ∥ e t J n , μ ∥ tends to n n e 1 − n ( n − 1 ) ! ∼ e n 2 π \frac{n^n e^{1-n}}{(n-1)!} \sim e\sqrt{\frac{n}{2\pi}} ( n − 1 )! n n e 1 − n ∼ e 2 π n . Put
Δ n ( α ) : = sup x ≥ 0 x n − 1 ∣ E α ( n − 1 ) ( − x ) ∣ ( n − 1 ) ! , ϱ n ( α ) : = n n ( n − 1 ) n − 1 Δ n ( α ) . \Delta_n(\alpha) := \sup_{x \ge0} \frac{x^{n-1}|E_{\alpha}^{(n-1)}(-x)|}{(n-1)!}, \qquad\varrho_n(\alpha) := \frac{n^n}{(n-1)^{n-1}}\Delta_n(\alpha). Δ n ( α ) := x ≥ 0 sup ( n − 1 )! x n − 1 ∣ E α ( n − 1 ) ( − x ) ∣ , ϱ n ( α ) := ( n − 1 ) n − 1 n n Δ n ( α ) .
Proposition 9.1. Let n ≥ 2 n \ge2 n ≥ 2 and 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 .
(a) sup z ∈ Σ α a α ( z ) ∣ 1 + z ∣ − n = ( n − 1 ) n − 1 n − n \sup_{z \in\Sigma_{\alpha}} a_{\alpha}(z)|1+z|^{-n} = (n-1)^{n-1}n^{-n} sup z ∈ Σ α a α ( z ) ∣1 + z ∣ − n = ( n − 1 ) n − 1 n − n , and lim μ → ∞ sup t ∥ E α ( t α J n , μ ) ∥ K α ( J n , μ ) = ϱ n ( α ) ≥ 1 \lim_{\mu\to\infty}\frac{\sup_t\|E_{\alpha}(t^{\alpha}J_{n,\mu})\|}{K_{\alpha}(J_{n,\mu})} = \varrho_n(\alpha) \ge1 lim μ → ∞ K α ( J n , μ ) s u p t ∥ E α ( t α J n , μ ) ∥ = ϱ n ( α ) ≥ 1 .
(b) ϱ n ( 1 ) = n n e 1 − n ( n − 1 ) ! \varrho_n(1) = \frac{n^n e^{1-n}}{(n-1)!} ϱ n ( 1 ) = ( n − 1 )! n n e 1 − n ; in particular ϱ n ( 1 ) ∼ e n 2 π \varrho_n(1) \sim e\sqrt{\frac{n}{2\pi}} ϱ n ( 1 ) ∼ e 2 π n as n → ∞ n \to\infty n → ∞ .
(c) If 0 < α < 1 0 < \alpha< 1 0 < α < 1 , then ϱ n ( α ) ≤ ϑ α ( n n − 1 ) n ≤ 4 ϑ α \varrho_n(\alpha) \le\vartheta_{\alpha}\left(\frac{n}{n-1}\right)^n \le4\vartheta_{\alpha} ϱ n ( α ) ≤ ϑ α ( n − 1 n ) n ≤ 4 ϑ α , with ϑ α \vartheta_{\alpha} ϑ α as in Theorem Appendix B.1(e).
Proof. (a) For z ∈ Σ α z \in\Sigma_{\alpha} z ∈ Σ α we have a α ( z ) ≤ Re z a_{\alpha}(z) \le\operatorname{Re} z a α ( z ) ≤ Re z and ∣ 1 + z ∣ ≥ 1 + Re z |1+z| \ge1+\operatorname{Re} z ∣1 + z ∣ ≥ 1 + Re z , so a α ( z ) ∣ 1 + z ∣ − n ≤ x ( 1 + x ) − n a_{\alpha}(z)|1+z|^{-n} \le x(1+x)^{-n} a α ( z ) ∣1 + z ∣ − n ≤ x ( 1 + x ) − n with x = Re z > 0 x=\operatorname{Re}z>0 x = Re z > 0 . The right-hand side is at most ( n − 1 ) n − 1 n − n (n-1)^{n-1}n^{-n} ( n − 1 ) n − 1 n − n , with equality at x = 1 n − 1 x=\frac{1}{n-1} x = n − 1 1 , and z = 1 n − 1 z=\frac{1}{n-1} z = n − 1 1 attains this value. Let τ : = t α \tau:= t^{\alpha} τ := t α . Since Z n n = 0 Z_n^n=0 Z n n = 0 and − τ I -\tau I − τ I commutes with Z n Z_n Z n , Taylor’s formula gives E α ( τ J n , μ ) = ∑ k = 0 n − 1 E α ( k ) ( − τ ) ( τ μ ) k Z n k k ! E_{\alpha}(\tau J_{n,\mu}) = \sum_{k=0}^{n-1}\frac{E_{\alpha}^{(k)}(-\tau)(\tau\mu)^k Z_n^k}{k!} E α ( τ J n , μ ) = ∑ k = 0 n − 1 k ! E α ( k ) ( − τ ) ( τ μ ) k Z n k . For α < 1 \alpha<1 α < 1 , differentiating Theorem Appendix B.1(d) under the integral sign gives E α ( k ) ( − τ ) = ∫ 0 ∞ u k e − τ u m α ( u ) d u ≥ 0 E_{\alpha}^{(k)}(-\tau) = \int_0^{\infty}u^k e^{-\tau u}m_{\alpha}(u)\,du \ge0 E α ( k ) ( − τ ) = ∫ 0 ∞ u k e − τ u m α ( u ) d u ≥ 0 , hence
τ k E α ( k ) ( − τ ) k ! = ∫ 0 ∞ ( τ u ) k e − τ u k ! m α ( u ) d u ≤ sup s ≥ 0 s k e − s k ! ≤ 1 , (25) \frac{\tau^k E_{\alpha}^{(k)}(-\tau)}{k!} = \int_0^{\infty}\frac{(\tau u)^k e^{-\tau u}}{k!}m_{\alpha}(u)\,du \le\sup_{s \ge0}\frac{s^k e^{-s}}{k!} \le1, \tag*{(25)} k ! τ k E α ( k ) ( − τ ) = ∫ 0 ∞ k ! ( τ u ) k e − τ u m α ( u ) d u ≤ s ≥ 0 sup k ! s k e − s ≤ 1 , (25)
which is trivial for α = 1 \alpha=1 α = 1 . Consequently, uniformly in τ \tau τ ,
∥ μ 1 − n E α ( τ J n , μ ) − τ n − 1 E α ( n − 1 ) ( − τ ) Z n n − 1 ( n − 1 ) ! ∥ ≤ ∑ k < n − 1 μ k − n + 1 , \left\|\mu^{1-n}E_{\alpha}(\tau J_{n,\mu})-\frac{\tau^{n-1}E_{\alpha}^{(n-1)}(-\tau)Z_n^{n-1}}{(n-1)!}\right\| \le\sum_{k<n-1}\mu^{k-n+1}, μ 1 − n E α ( τ J n , μ ) − ( n − 1 )! τ n − 1 E α ( n − 1 ) ( − τ ) Z n n − 1 ≤ k < n − 1 ∑ μ k − n + 1 ,
and μ 1 − n sup t ∥ E α ( t α J n , μ ) ∥ → Δ n ( α ) \mu^{1-n}\sup_t\|E_{\alpha}(t^{\alpha}J_{n,\mu})\| \to\Delta_n(\alpha) μ 1 − n sup t ∥ E α ( t α J n , μ ) ∥ → Δ n ( α ) because ∥ Z n n − 1 ∥ = 1 \|Z_n^{n-1}\|=1 ∥ Z n n − 1 ∥ = 1 . Similarly ( z − J n , μ ) − 1 = ∑ k = 0 n − 1 μ k Z n k ( z + 1 ) − k − 1 (z-J_{n,\mu})^{-1}=\sum_{k=0}^{n-1}\mu^k Z_n^k(z+1)^{-k-1} ( z − J n , μ ) − 1 = ∑ k = 0 n − 1 μ k Z n k ( z + 1 ) − k − 1 , and since a α ( z ) ≤ Re z < ∣ 1 + z ∣ a_{\alpha}(z)\le\operatorname{Re}z<|1+z| a α ( z ) ≤ Re z < ∣1 + z ∣ and ∣ 1 + z ∣ ≥ 1 |1+z|\ge1 ∣1 + z ∣ ≥ 1 on Σ α \Sigma_{\alpha} Σ α , the same argument gives μ 1 − n K α ( J n , μ ) → ( n − 1 ) n − 1 n − n \mu^{1-n}K_{\alpha}(J_{n,\mu})\to(n-1)^{n-1}n^{-n} μ 1 − n K α ( J n , μ ) → ( n − 1 ) n − 1 n − n . The ratio is at least 1 by Theorem 5.1 .
(b) Δ n ( 1 ) = sup x x n − 1 e − x ( n − 1 ) ! = ( n − 1 ) n − 1 e 1 − n ( n − 1 ) ! \Delta_n(1)=\sup_x \frac{x^{n-1}e^{-x}}{(n-1)!}=\frac{(n-1)^{n-1}e^{1-n}}{(n-1)!} Δ n ( 1 ) = sup x ( n − 1 )! x n − 1 e − x = ( n − 1 )! ( n − 1 ) n − 1 e 1 − n , and Stirling’s formula gives the asymptotics.
(c) Substituting s = x u s=xu s = xu in (25) ,
x n − 1 E α ( n − 1 ) ( − x ) ( n − 1 ) ! = ∫ 0 ∞ s n − 2 e − s ( n − 1 ) ! ⋅ s x m α ( s x ) d s ≤ ϑ α ( n − 2 ) ! ( n − 1 ) ! = ϑ α n − 1 , \frac{x^{n-1}E_{\alpha}^{(n-1)}(-x)}{(n-1)!}=\int_0^\infty\frac{s^{n-2}e^{-s}}{(n-1)!}\cdot\frac{s}{x}m_\alpha\left(\frac{s}{x}\right)\,ds\leq\vartheta_\alpha\frac{(n-2)!}{(n-1)!}=\frac{\vartheta_\alpha}{n-1}, ( n − 1 )! x n − 1 E α ( n − 1 ) ( − x ) = ∫ 0 ∞ ( n − 1 )! s n − 2 e − s ⋅ x s m α ( x s ) d s ≤ ϑ α ( n − 1 )! ( n − 2 )! = n − 1 ϑ α ,
so ϱ n ( α ) ≤ ϑ α n n ( n − 1 ) n \varrho_n(\alpha)\leq\frac{\vartheta_\alpha n^n}{(n-1)^{n}} ϱ n ( α ) ≤ ( n − 1 ) n ϑ α n n , and ( n n − 1 ) n ≤ 4 \left(\frac{n}{n-1}\right)^n\leq4 ( n − 1 n ) n ≤ 4 for n ≥ 2 n\geq2 n ≥ 2 .
In the limit of strong non-normality, Jordan blocks thus produce only a bounded amplification ratio when α < 1 \alpha<1 α < 1 , for instance ϱ n ( 1 2 ) ≤ 2 e π ( n n − 1 ) n → 2 e π ≈ 1.316 \varrho_n\left(\frac{1}{2}\right)\leq\sqrt{\frac{2}{e\pi}}\left(\frac{n}{n-1}\right)^n\to\sqrt{\frac{2e}{\pi}}\approx1.316 ϱ n ( 2 1 ) ≤ e π 2 ( n − 1 n ) n → π 2 e ≈ 1.316 . The reason is the averaging in Theorem 3.3(b): the factor s n − 1 e − s ( n − 1 ) ! \frac{s^{n-1}e^{-s}}{(n-1)!} ( n − 1 )! s n − 1 e − s , whose sharp peak of height of order n − 1 2 n^{-\frac{1}{2}} n − 2 1 produces the growth when α = 1 \alpha=1 α = 1 , is smeared out by the Mittag–Leffler distribution. The bound is not uniform in α \alpha α , since ϑ α → ∞ \vartheta_\alpha\to\infty ϑ α → ∞ as α ↑ 1 \alpha\uparrow1 α ↑ 1 : the moments E Y α k = k ! Γ ( 1 + α k ) \mathbb{E}Y_\alpha^k=\frac{k!}{\Gamma(1+\alpha k)} E Y α k = Γ ( 1 + α k ) k ! tend to 1 1 1 , so Y α → 1 Y_\alpha\to1 Y α → 1 in distribution, which is incompatible with a uniform bound on u m α ( u ) u\,m_\alpha(u) u m α ( u ) near u = 1 u=1 u = 1 . The values in Figure 2 (c), listed in tbl:S2 of the Supplementary Material, show the transition: ϱ 16 \varrho_{16} ϱ 16 equals 1.08, 1.28, 2.18 and 4.32 for α = 0.3 \alpha=0.3 α = 0.3 , 0.5, 0.8 and 1. These limits concern Jordan blocks in the regime of strong non-normality only, so they do not exclude other nilpotent mechanisms, but they are consistent with the construction of Theorem C , where the growth is produced by oscillatory boundary modes; § 11.3 shows the same taming in the means of time-changed compartment chains.
The prototypes revisited
We return to the three prototypes of § 2 and compare, for each of them, the propagator sup t ∥ E α ( t α L ) ∥ \sup_t\lVert E_\alpha(t^\alpha L)\rVert sup t ∥ E α ( t α L )∥ with the two sides of Theorem A and with the bounds that are otherwise available. These are the semigroup route (B1), sup t ∥ E α ( t α L ) ∥ ≤ sup s ∥ e s L ∥ \sup_t\lVert E_\alpha(t^\alpha L)\rVert\leq\sup_s\lVert e^{sL}\rVert sup t ∥ E α ( t α L )∥ ≤ sup s ∥ e s L ∥ by Theorem 3.3(b), which is infinite whenever σ ( L ) \sigma(L) σ ( L ) meets H \mathbb{H} H or L L L generates no semigroup; the numerical-range bounds (B2), K sup C W ∣ E α ∣ K\sup_{C_W}|E_\alpha| K sup C W ∣ E α ∣ with K = 1 + 2 K=1+\sqrt{2} K = 1 + 2 [8 ] , K = K ( β ) K=K(\beta) K = K ( β ) [7 ] or K = 1 K=1 K = 1 (Theorem 6.4 ), all infinite when W ( L ) W(L) W ( L ) meets Σ α \Sigma_\alpha Σ α ; and crude deflation (B3), which estimates the slow block and the obliqueness of the splitting separately, as in (20) . The Supplementary Material also reports (B4), Theorem A with the dimension in place of the rank, and an eigenvector bound (B5) is used where a complete eigenbasis is available. Quantities attached to slow components are computed through the reductions of Theorems 3.9 and 5.2, at a cost governed by m m m , and Kreiss constants are maximised over polar grids in Σ α \Sigma_\alpha Σ α with local refinement, so the reported values are accurate lower estimates rather than certified upper bounds. The numerical methods, the complete tables and the code are described in the Supplementary Material, which also reports a statistical study of 16200 random operators of rank at most 32 on spaces of dimension up to 10 5 10^5 1 0 5 : there the ratio of the two sides of Theorem A never exceeds 5.6, nor 17% of the corresponding κ α , n \kappa_{\alpha,n} κ α , n , and is essentially independent of the dimension for d ≥ 10 3 d\geq10^3 d ≥ 1 0 3 (Section S3).
Prototype (a): the oscillator-driven subdiffusion
For the generator of Theorem 7.4 with α = 1 2 \alpha=\frac{1}{2} α = 2 1 , θ = π 3 \theta=\frac{\pi}{3} θ = 3 π and the profiles b 1 = π − 1 2 1 ( 0 , π ) b_1=\pi^{-\frac{1}{2}}\mathbf{1}_{(0,\pi)} b 1 = π − 2 1 1 ( 0 , π ) and b 2 b_2 b 2 the normalised parabola x ( π − x ) x(\pi-x) x ( π − x ) , all quantities are computed in the sine basis, with up to 10 4 10^4 1 0 4 modes, at which the reported digits no longer change. By Theorem 7.4 the semigroup bound (B1) and every numerical-range bound (B2) are infinite for every coupling strength, so among the bounds that do not require an eigenbasis only the slow component gives information. For (B3) we evaluate (20) with the constant κ 1 2 , 2 ′ \kappa'_{\frac{1}{2},2} κ 2 1 , 2 ′ , computing K α ( L P ) \mathcal{K}_{\alpha}(L_P) K α ( L P ) in the norm of H H H through the similarity L P = T L 1 T − 1 L_P = T L_1 T^{-1} L P = T L 1 T − 1 of the proof of Theorem 7.4. Since this operator has a complete eigenbasis, we also report the eigenvector bound (B5), cond ( S ) max λ ∈ σ ( L ) sup t ∣ E α ( t α λ ) ∣ \operatorname{cond}(S)\max_{\lambda\in\sigma(L)}\sup_t\lvert E_{\alpha}(t^{\alpha}\lambda)\rvert cond ( S ) max λ ∈ σ ( L ) sup t ∣ E α ( t α λ )∣ , where S S S maps the standard basis to an eigenbasis with normalised vectors.
Figure 3 shows the result. The envelope rises with the initial slopes of Theorem Appendix B.3(a), peaks near t = 1 t=1 t = 1 and decays like t − 1 2 t^{-\frac{1}{2}} t − 2 1 , as Theorem Appendix B.3(d) predicts, since L L L satisfies its hypotheses with ψ = θ \psi=\theta ψ = θ . The four quantities K α ( L ) \mathcal{K}_{\alpha}(L) K α ( L ) , K α ( L ; P ) \mathcal{K}_{\alpha}(L;P) K α ( L ; P ) , C Q = ∥ Q ∥ C_Q=\lVert Q\rVert C Q = ∥ Q ∥ and sup t ∥ E α ( t α L ) ∥ \sup_t\lVert E_{\alpha}(t^{\alpha}L)\rVert sup t ∥ E α ( t α L )∥ grow linearly in μ \mu μ , while K α ( L 1 ) = 1.229 \mathcal{K}_{\alpha}(L_1)=1.229 K α ( L 1 ) = 1.229 and the bound 1 on the fast part do not move; at μ = 100 \mu=100 μ = 100 , for instance, ∥ P ∥ = 77.7 \lVert P\rVert=77.7 ∥ P ∥ = 77.7 , K α ( L ) = 67.3 \mathcal{K}_{\alpha}(L)=67.3 K α ( L ) = 67.3 , K α ( L ; P ) = 89.7 \mathcal{K}_{\alpha}(L;P)=89.7 K α ( L ; P ) = 89.7 and sup t ∥ E α ( t α L ) ∥ = 100.9 \sup_t\lVert E_{\alpha}(t^{\alpha}L)\rVert=100.9 sup t ∥ E α ( t α L )∥ = 100.9 . The propagator exceeds the deflated Kreiss constant by at most 13% for every coupling strength, and by at most 12% and 4% for α = 0.3 \alpha=0.3 α = 0.3 and α = 0.8 \alpha=0.8 α = 0.8 , so in this model K α ( L ; P ) \mathcal{K}_{\alpha}(L;P) K α ( L ; P ) is a faithful indicator of the transient although the semigroup grows exponentially and the numerical range meets the sector of instability. The lower bound K α ( L ) \mathcal{K}_{\alpha}(L) K α ( L ) is within a factor 1.5 of the propagator, and the upper bound of Theorem A, 1662 at μ = 100 \mu=100 μ = 100 , is 16 to 17 times larger, a gap that reflects the worst-case constant κ 1 2 , 2 ′ = 17.7 \kappa'_{\frac{1}{2},2}=17.7 κ 2 1 , 2 ′ = 17.7 rather than this operator; the quotients of Theorem 4.1 for concrete rational functions stay below 40% of their bound (Table S3). The crude bound (B3) is much larger: K α ( L P ) \mathcal{K}_{\alpha}(L_P) K α ( L P ) grows to 21.4 at μ = 100 \mu=100 μ = 100 , below the bound y 0 s 0 K α ( L 1 ) ≈ 32 \frac{y_0}{s_0}\mathcal{K}_{\alpha}(L_1)\approx32 s 0 y 0 K α ( L 1 ) ≈ 32 of Theorem 7.4(b), and (20) multiplies it once more by ∥ P ∥ \lVert P\rVert ∥ P ∥ , so that the obliqueness of the splitting is counted twice and (B3) exceeds our bound by a factor of 18 at μ = 100 \mu=100 μ = 100 . The deflated constant counts it once, since K α ( L ; P ) ≤ ∥ P ∥ K α ( L 1 ) \mathcal{K}_{\alpha}(L;P)\leq\lVert P\rVert\mathcal{K}_{\alpha}(L_1) K α ( L ; P ) ≤ ∥ P ∥ K α ( L 1 ) . The eigenvector bound (B5) is about twice the propagator and is the sharpest of all bounds for this operator, but it needs a complete and well-conditioned eigenbasis of the whole generator, which is not available for non-diagonalisable or nearly defective generators, whereas the slow-component bound uses only the slow block, the splitting and a bound on the fast part.
Figure 3. Prototype (a) with α = 1 2 \alpha= \frac{1}{2} α = 2 1 and θ = π 3 \theta= \frac{\pi}{3} θ = 3 π . (a) The envelope t ↦ ∥ E 1 2 ( t 1 2 L ) ∥ t \mapsto\lVert E_{\frac{1}{2}}(t^{\frac{1}{2}}L)\rVert t ↦ ∥ E 2 1 ( t 2 1 L )∥ for three coupling strengths: initial slope, transient peak near t = 1 t=1 t = 1 and algebraic decay of order t − 1 2 t^{-\frac{1}{2}} t − 2 1 ; dotted lines mark K 1 2 ( L ; P ) \mathcal{K}_{\frac{1}{2}}(L;P) K 2 1 ( L ; P ) and dashed lines the upper bound of Theorem A. (b) Propagator, deflated Kreiss constant and upper bounds as functions of μ \mu μ : “ours” is the upper bound of Theorem A, B3′ the crude deflation bound (B3) and B5 the eigenvector bound; the semigroup and numerical-range bounds are infinite.
Prototype (b): the time-fractional Schrödinger equation
In this prototype the values follow from the theory alone. Let A V = − d 2 d x 2 + V A_{\mathrm{V}}=-\frac{d^2}{dx^2}+V A V = − d x 2 d 2 + V be a Schrödinger operator that is self-adjoint, bounded below and unbounded above, with at least one positive spectral value, for instance the operator on R \mathbb{R} R with the Pöschl–Teller potential V ( x ) = − 6 sech 2 x V(x)=-6\operatorname{sech}^2x V ( x ) = − 6 sech 2 x , whose bound states are − 4 -4 − 4 and − 1 -1 − 1 , or its restriction to a large interval with Dirichlet conditions. For L = e − i α π 2 A V L=e^{-i\frac{\alpha\pi}{2}}A_{\mathrm{V}} L = e − i 2 α π A V , Theorem 6.2 and Theorem 6.1 give sup t ∥ E α ( t α L ) ∥ = M α ∗ \sup_t\lVert E_{\alpha}(t^{\alpha}L)\rVert=M_{\alpha}^{*} sup t ∥ E α ( t α L )∥ = M α ∗ and K α ( L ) = 1 α \mathcal{K}_{\alpha}(L)=\frac{1}{\alpha} K α ( L ) = α 1 exactly, so the two sides of Theorem A differ by the factor α M α ∗ \alpha M_{\alpha}^{*} α M α ∗ , while L L L generates no C 0 C_0 C 0 -semigroup. The numerical range of L L L lies on the line through 0 that contains the lower boundary ray of Σ α \Sigma_{\alpha} Σ α , so Theorem 6.4 applies with the half-plane bounded by this line and returns the exact value M α ∗ M_{\alpha}^{*} M α ∗ , whereas for bounded discretisations the bound of Crouzeix and Palencia [8 ] gives ( 1 + 2 ) M α ∗ (1+\sqrt{2})M_{\alpha}^{*} ( 1 + 2 ) M α ∗ and the half-plane estimate of Crouzeix [7 ] does not apply. The memory of the time-fractional Schrödinger equation therefore costs an amplification of exactly M α ∗ M_{\alpha}^{*} M α ∗ , which is 2.34 for α = 1 2 \alpha=\frac{1}{2} α = 2 1 , in a model for which no semigroup bound exists. Finite-difference discretisations reproduce these values on every grid, while the norms ∥ e L ∥ \lVert e^L\rVert ∥ e L ∥ of the discretised semigroups grow without bound with the resolution (Table S9).
Prototype (c): time-changed Markov dynamics
Compartment chains. Let x s \mathbf{x}_s x s be an Ornstein–Uhlenbeck process d x s = − B x s d s + Ξ d W s \mathrm{d}\mathbf{x}_s=-B\mathbf{x}_s\,\mathrm{d}s+\Xi\,\mathrm{d}W_s d x s = − B x s d s + Ξ d W s in R d \mathbb{R}^d R d and let T t T_t T t be the inverse of an independent standard α \alpha α -stable subordinator, the clock of a continuous-time random walk with heavy-tailed waiting times. Since T t T_t T t has the law of t α Y α t^\alpha Y_\alpha t α Y α [5 , 33 ] and E x s = e − s B x 0 \mathbb{E}\mathbf{x}_s=e^{-sB}\mathbf{x}_0 E x s = e − s B x 0 , Theorem 3.3(b) gives
E [ x T t ] = E α ( − t α B ) x 0 , \mathbb{E}\left[\mathbf{x}_{T_t}\right]=E_\alpha(-t^\alpha B)\mathbf{x}_0, E [ x T t ] = E α ( − t α B ) x 0 ,
so the mean of the time-changed process is governed by Theorem A. We take the linear chain B = I − μ Z d T B=I-\mu Z_d^{\mathsf{T}} B = I − μ Z d T , with Z d Z_d Z d the nilpotent shift of § 10 , in which each compartment decays at unit rate and feeds the next one with strength μ \mu μ ; then − B = J d , μ -B=J_{d,\mu} − B = J d , μ is a Jordan block, and the classical mean exhibits the large transient growth that is typical of feed-forward networks. In this example the two Kreiss constants coincide: since μ ≥ 0 \mu\ge0 μ ≥ 0 and
( z + B ) − 1 = ∑ k μ k ( Z d T ) k ( z + 1 ) − k − 1 , (z+B)^{-1}=\sum_k\mu^k(Z_d^{\mathsf{T}})^k(z+1)^{-k-1}, ( z + B ) − 1 = k ∑ μ k ( Z d T ) k ( z + 1 ) − k − 1 ,
its entries are dominated in modulus by those of the nonnegative matrix ( Re z + B ) − 1 (\operatorname{Re}z+B)^{-1} ( Re z + B ) − 1 , and together with a α ( z ) ≤ Re z a_\alpha(z)\le\operatorname{Re}z a α ( z ) ≤ Re z this gives K α ( − B ) = sup x > 0 x ∥ ( x + B ) − 1 ∥ = K 1 ( − B ) \mathcal{K}_\alpha(-B)=\sup_{x>0}x\lVert(x+B)^{-1}\rVert=\mathcal{K}_1(-B) K α ( − B ) = sup x > 0 x ∥( x + B ) − 1 ∥ = K 1 ( − B ) for every α \alpha α . The difference between the classical and the fractional transient is therefore entirely in the propagators. For μ = 2 \mu=2 μ = 2 the classical ratio
sup s ∥ e − s B ∥ K 1 ( − B ) \frac{\sup_s\lVert e^{-sB}\rVert}{\mathcal{K}_1(-B)} K 1 ( − B ) sup s ∥ e − s B ∥
grows like e d 2 π e\sqrt{\frac{d}{2\pi}} e 2 π d , from 1.88 1.88 1.88 at d = 5 d=5 d = 5 to 15.3 15.3 15.3 at d = 200 d=200 d = 200 , while the fractional ratio at d = 200 d=200 d = 200 equals 1.09 1.09 1.09 , 1.31 1.31 1.31 and 2.66 2.66 2.66 for α = 0.3 \alpha=0.3 α = 0.3 , 0.5 0.5 0.5 and 0.8 0.8 0.8 , just below the bounds ϑ α ( d d − 1 ) d \vartheta_\alpha\left(\frac{d}{d-1}\right)^d ϑ α ( d − 1 d ) d that Theorem 9.1(c) gives in the limit of strong coupling (Table S10). The rank is d d d here, so the example illustrates the mechanism of § 10 rather than the slow-component bound: the heavy-tailed clock averages out the nilpotent amplification, and up to d = 200 d=200 d = 200 the ratio stays below 2.7 2.7 2.7 .
Learned low-rank generators. In data-driven practice the generator of a diffusion is replaced by a finite-rank estimate on an infinite-dimensional function space, and the slow-component formulation then yields certificates whose constants depend on the rank of the estimate and not on the number of features. For the transfer operator of a non-reversible Langevin diffusion in R 2 \mathbb{R}^2 R 2 , estimated by reduced-rank regression [23 ] of rank r ≤ 16 r \le16 r ≤ 16 on up to 4000 random Fourier features, every quantity of Theorem 7.1 is evaluated at a cost governed by r r r . The estimated slow dynamics shows no transient growth beyond the obliqueness of the estimator, K α ( L r ; Π ) = sup t ≥ 0 ∥ E α ( t α L r ) Π ∥ = ∥ Π ∥ \mathcal{K}_\alpha(L_r;\Pi)=\sup_{t\ge0}\|E_\alpha(t^\alpha L_r)\Pi\|=\|\Pi\| K α ( L r ; Π ) = sup t ≥ 0 ∥ E α ( t α L r ) Π∥ = ∥Π∥ , so the lower bound of Theorem A is attained, and in one configuration a spurious eigenvalue of size 10 − 6 10^{-6} 1 0 − 6 in Σ α \Sigma_\alpha Σ α makes K α ( L r ; Π ) \mathcal{K}_\alpha(L_r;\Pi) K α ( L r ; Π ) infinite and flags the model as unusable for long-time forecasting under the fractional clock (Section S7.2 of Supplementary material).
Summary
In the three prototypes new theory gives information where the classical tools give little or none. The deflated Kreiss constant follows the propagator within 13% in a model where every semigroup and numerical-range bound is infinite, the amplification caused by memory in the time-fractional Schrödinger equation is exactly M α ∗ M_\alpha^* M α ∗ although no semigroup exists, and nilpotent growth that is unbounded in physical time is tamed by the fractional clock. Among the operators considered here, the ratio of the two sides of Theorem A is large only in the oscillatory construction of § 9 , where it reaches 13.4 at rank 32, at which random operators stay below 5.6, and 138 at rank 512, in agreement with the linear upper bound of Theorem A and the lower bound of Theorem C .
Conclusions and open problems
The fractional Kreiss constant is the resolvent quantity that controls the transient growth of Mittag–Leffler families. It bounds the propagator from below for every generator of a bounded family, and from above, up to a factor linear in the rank, for every generator with a finite slow component, on arbitrary Banach spaces and uniformly in the order α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] . We show that this rank dependence is genuine, and that in infinite dimensions a finite Kreiss constant does not guaranty neither boundedness nor existence of the family. The upper bounds rest on the Bernstein-type inequality for a class of non-rational functions, which can be of independent interest.
The bigger picture is that the fractional transient also behaves differently from the classical one. Normal operators amplify, modes that are unstable in physical time are stable under the fractional clock, and for each fixed α < 1 \alpha< 1 α < 1 the nilpotent amplification that drives the classical Kreiss matrix theorem is averaged out by the Mittag–Leffler distribution, while the growth in our optimality construction is oscillatory. The envelope has distinct short-time, intermediate and long-time regimes, governed by the numerical abscissa, by the Kreiss constant and by the resolvent at the origin.
These results bear directly on the models that motivated them. For subdiffusion driven by a self-excited oscillator, the deflated Kreiss constant tracked the transient within 13% in all computed cases, although the semigroup grows exponentially and the numerical range meets the sector of instability, which makes it a natural candidate for assessing how strongly an unstable local mode can amplify a subdiffusive signal. For the time-fractional Schrödinger equation, the solution operators are bounded by exactly M α ∗ M_{\alpha}^{*} M α ∗ although no semigroup exists, which quantifies the price of memory in this quantum model. For time-changed Markov dynamics, the means of compartment chains are far less amplified under the fractional clock than in physical time, and for finite-rank models learned from data the slow-component formulation yields certificates and diagnostics whose constants are set by the number of retained modes.
We conclude with an open question that remains for future work. It concerns the sharpness of constants J F J_{F} J F in Theorem B and the order of κ α opt ( n ) \kappa_{\alpha}^{\operatorname{opt}}(n) κ α opt ( n ) . For α = 1 \alpha= 1 α = 1 the linear order in the dimension is sharp [28 ] ; for α < 1 \alpha< 1 α < 1 we know n γ ≲ κ α opt ( n ) ≲ n n^{\gamma} \lesssim\kappa_{\alpha}^{\operatorname{opt}}(n) \lesssim n n γ ≲ κ α opt ( n ) ≲ n for every γ < 1 \gamma< 1 γ < 1 . Finally, we note that the proof of Theorem B uses change ω ↦ ω α \omega\mapsto\omega^{\alpha} ω ↦ ω α only when it maps vertical lines to curves whose tangent turns monotonically through an angle less than π \pi π . Hence, we believe that Kreiss-type bounds for resolvent families of Volterra equations with other completely positive kernels, as well as for the families t β − 1 E α , β ( t α L ) t^{\beta-1}E_{\alpha,\beta}(t^{\alpha}L) t β − 1 E α , β ( t α L ) that appear in inhomogeneous problems, could follow as a natural extension of this work.
Data availability
No external data were used. The numerical methods and the complete results are described in the Supplementary Material. The code that produces them, together with the stored results, is available to the referees on request and will be released in a public repository upon publication; every table and figure of the paper and of the Supplementary Material is regenerated by a single script with fixed random seeds.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Funding
This work was funded by the Provincial Secretariat for Higher Education and Scientific Research of Vojvodina (Research Grant 003870560 2025 09418 003 000 000 001 04 004).
Declaration of generative AI and AI-assisted technologies in the writing process
The problem of fractional Kreiss-type theorem for finite rank operators, conceptualization of novel Kreiss constant and developed mathematical approach to obtain the bound is human made, Claude Opus 5.5 has been used to improve presentation, polish writing and identify issues in consistent notations and mathematical formalism. Additionally, Claude Code has been used to produce simple numerical demonstrations provided in the Supplementary material for reader’s convenience.
Appendix
The appendix complements the main text in three self-contained parts. Section Appendix A collects the notation, grouped by topic. Section Appendix B contains the proofs of the two results on sectorial generators, the generation result Theorem 3.4 and the long-time expansion of Theorem Appendix B.3(d); both follow the semigroup case closely, with the Hankel contour adapted to the fractional order. Section Appendix C contains the harmonic-analysis ingredients of the optimality construction of § 9 : the A 2 A_2 A 2 property of the weight, the ray integral (22) , the boundary behaviour (23) of E α E_\alpha E α and the Dirichlet-kernel estimate (24) . Each part restates what it proves, so that it can be read independently of the main text. The complete tables of numerical results, as well as the code used to produce them, are in the separate Supplementary Material.
Notation
Table 2.
Table 3 (PDF p. 46).
Proofs of preliminary results
Mittag–Leffler functions
Here we provide the collection of useful facts on Mittag–Leffler functions, together with their proofs for completeness.
Lemma Appendix B.1. Let 0 < α < 1 0<\alpha<1 0 < α < 1 and fix θ 0 \theta_0 θ 0 with α π 2 < θ 0 < α π \frac{\alpha\pi}{2}<\theta_0<\alpha\pi 2 α π < θ 0 < α π . There are constants C , C ′ > 0 C,C'>0 C , C ′ > 0 such that the following hold.
(a) If ∣ z ∣ ≥ 1 |z|\geq1 ∣ z ∣ ≥ 1 and ∣ Arg z ∣ ≤ θ 0 |\operatorname{Arg}z|\leq\theta_0 ∣ Arg z ∣ ≤ θ 0 , then E α ( z ) = α − 1 exp ( z 1 α ) − z − 1 Γ ( 1 − α ) + r ( z ) E_\alpha(z)=\alpha^{-1}\exp\left(z^{\frac{1}{\alpha}}\right)-\frac{z^{-1}}{\Gamma(1-\alpha)}+r(z) E α ( z ) = α − 1 exp ( z α 1 ) − Γ ( 1 − α ) z − 1 + r ( z ) with ∣ r ( z ) ∣ ≤ C ∣ z ∣ − 2 |r(z)|\leq C|z|^{-2} ∣ r ( z ) ∣ ≤ C ∣ z ∣ − 2 .
(b) If ∣ z ∣ ≥ 1 |z|\geq1 ∣ z ∣ ≥ 1 and θ 0 ≤ ∣ Arg z ∣ ≤ π \theta_0\leq|\operatorname{Arg}z|\leq\pi θ 0 ≤ ∣ Arg z ∣ ≤ π , then ∣ E α ( z ) + z − 1 Γ ( 1 − α ) ∣ ≤ C ∣ z ∣ − 2 \left|E_\alpha(z)+\frac{z^{-1}}{\Gamma(1-\alpha)}\right|\leq C|z|^{-2} E α ( z ) + Γ ( 1 − α ) z − 1 ≤ C ∣ z ∣ − 2 .
(c) E α E_\alpha E α is bounded on Λ α \Lambda_\alpha Λ α ; on every closed sector { θ 1 ≤ ∣ Arg z ∣ ≤ π } \{\theta_1\leq|\operatorname{Arg}z|\leq\pi\} { θ 1 ≤ ∣ Arg z ∣ ≤ π } with θ 1 > α π 2 \theta_1>\frac{\alpha\pi}{2} θ 1 > 2 α π , E α ( j ) ( z ) = O ( ∣ z ∣ − 1 − j ) E_\alpha^{(j)}(z)=O(|z|^{-1-j}) E α ( j ) ( z ) = O ( ∣ z ∣ − 1 − j ) as ∣ z ∣ → ∞ |z|\to\infty ∣ z ∣ → ∞ for every j ∈ N 0 j\in\mathbb{N}_0 j ∈ N 0 ; and q α ( z ) : = E α ( z ) − α − 1 exp ( z 1 α ) q_\alpha(z):=E_\alpha(z)-\alpha^{-1}\exp\left(z^{\frac{1}{\alpha}}\right) q α ( z ) := E α ( z ) − α − 1 exp ( z α 1 ) satisfies ∣ q α ( z ) ∣ ≤ C ′ ∣ z ∣ − 1 |q_\alpha(z)|\leq C'|z|^{-1} ∣ q α ( z ) ∣ ≤ C ′ ∣ z ∣ − 1 and ∣ q α ′ ( z ) ∣ ≤ C ′ ∣ z ∣ − 2 |q'_\alpha(z)|\leq C'|z|^{-2} ∣ q α ′ ( z ) ∣ ≤ C ′ ∣ z ∣ − 2 for ∣ Arg z ∣ = α π 2 |\operatorname{Arg}z|=\frac{\alpha\pi}{2} ∣ Arg z ∣ = 2 α π and ∣ z ∣ ≥ 2 |z|\geq2 ∣ z ∣ ≥ 2 .
(d) ∫ 0 ∞ u k m α ( u ) d u = k ! Γ ( 1 + α k ) \int_0^\infty u^k m_\alpha(u)\,du=\frac{k!}{\Gamma(1+\alpha k)} ∫ 0 ∞ u k m α ( u ) d u = Γ ( 1 + α k ) k ! for k ∈ N 0 k\in\mathbb{N}_0 k ∈ N 0 , and ∫ 0 ∞ e x u m α ( u ) d u = E α ( x ) < ∞ \int_0^\infty e^{xu}m_\alpha(u)\,du=E_\alpha(x)<\infty ∫ 0 ∞ e xu m α ( u ) d u = E α ( x ) < ∞ for every x ∈ R x\in\mathbb{R} x ∈ R .
(e) ϑ α : = sup u > 0 u m α ( u ) < ∞ \vartheta_\alpha:=\sup_{u>0}u\,m_\alpha(u)<\infty ϑ α := sup u > 0 u m α ( u ) < ∞ , and m 1 2 ( u ) = π − 1 2 e − u 2 4 m_{\frac12}(u)=\pi^{-\frac12}e^{-\frac{u^2}{4}} m 2 1 ( u ) = π − 2 1 e − 4 u 2 , so that ϑ 1 2 = 2 e π \vartheta_{\frac12}=\sqrt{\frac{2}{e\pi}} ϑ 2 1 = e π 2 .
Proof. (a) and (b) are the cases β = 1 \beta=1 β = 1 , p = 1 p=1 p = 1 of the classical expansions of the two-parameter function E α , β E_{\alpha,\beta} E α , β , [38 Section 1.2] , see also Gorenflo et al. [17 Chapter 4] .
(c) For z z z with α π 2 ≤ ∣ Arg z ∣ ≤ θ 0 \frac{\alpha\pi}{2}\leq|\operatorname{Arg}z|\leq\theta_0 2 α π ≤ ∣ Arg z ∣ ≤ θ 0 we have ∣ Arg z ∣ α ∈ [ π 2 , π ) \frac{|\operatorname{Arg}z|}{\alpha}\in\left[\frac{\pi}{2},\pi\right) α ∣ Arg z ∣ ∈ [ 2 π , π ) , and therefore ∣ exp ( z 1 α ) ∣ = exp ( ∣ z ∣ 1 α cos ( Arg z α ) ) ≤ 1 \left|\exp\left(z^{\frac{1}{\alpha}}\right)\right|=\exp\left(|z|^{\frac{1}{\alpha}}\cos\left(\frac{\operatorname{Arg}z}{\alpha}\right)\right)\leq1 exp ( z α 1 ) = exp ( ∣ z ∣ α 1 cos ( α Arg z ) ) ≤ 1 ; together with (a), (b) and continuity on { ∣ z ∣ ≤ 1 } \{|z|\leq1\} { ∣ z ∣ ≤ 1 } this gives boundedness on Λ α \Lambda_\alpha Λ α . On { θ 1 ≤ ∣ Arg z ∣ ≤ π } \{\theta_1\leq|\operatorname{Arg}z|\leq\pi\} { θ 1 ≤ ∣ Arg z ∣ ≤ π } the exponential term decays faster than any power, so E α ( z ) = O ( ∣ z ∣ − 1 ) E_\alpha(z)=O(|z|^{-1}) E α ( z ) = O ( ∣ z ∣ − 1 ) there, and Cauchy’s estimates on discs of radius proportional to ∣ z ∣ |z| ∣ z ∣ inside a slightly larger sector give the bounds on the derivatives. By (a), ∣ q α ( z ) ∣ ≤ C ′ ∣ z ∣ − 1 |q_\alpha(z)|\leq C'|z|^{-1} ∣ q α ( z ) ∣ ≤ C ′ ∣ z ∣ − 1 for ∣ z ∣ ≥ 1 |z|\geq1 ∣ z ∣ ≥ 1 and ∣ Arg z ∣ ≤ θ 0 |\operatorname{Arg}z|\leq\theta_0 ∣ Arg z ∣ ≤ θ 0 , where q α q_\alpha q α is holomorphic since θ 0 < π \theta_0<\pi θ 0 < π . Let ∣ Arg z ∣ = α π 2 |\operatorname{Arg}z|=\frac{\alpha\pi}{2} ∣ Arg z ∣ = 2 α π , ∣ z ∣ ≥ 2 |z|\geq2 ∣ z ∣ ≥ 2 and ϱ : = ∣ z ∣ sin ( θ 0 − α π 2 ) 2 \varrho:=\frac{|z|\sin\left(\theta_0-\frac{\alpha\pi}{2}\right)}{2} ϱ := 2 ∣ z ∣ s i n ( θ 0 − 2 α π ) . Every ζ \zeta ζ with ∣ ζ − z ∣ ≤ ϱ |\zeta-z|\leq\varrho ∣ ζ − z ∣ ≤ ϱ satisfies ∣ Arg ζ ∣ ≤ θ 0 |\operatorname{Arg}\zeta|\leq\theta_0 ∣ Arg ζ ∣ ≤ θ 0 and ∣ ζ ∣ ≥ ∣ z ∣ 2 ≥ 1 |\zeta|\geq\frac{|z|}{2}\geq1 ∣ ζ ∣ ≥ 2 ∣ z ∣ ≥ 1 , so Cauchy’s estimate gives
∣ q α ′ ( z ) ∣ ≤ ϱ − 1 max ∣ ζ − z ∣ = ϱ ∣ q α ( ζ ) ∣ = O ( ∣ z ∣ − 2 ) , |q'_\alpha(z)|\leq\varrho^{-1}\max_{|\zeta-z|=\varrho}|q_\alpha(\zeta)|=O(|z|^{-2}), ∣ q α ′ ( z ) ∣ ≤ ϱ − 1 ∣ ζ − z ∣ = ϱ max ∣ q α ( ζ ) ∣ = O ( ∣ z ∣ − 2 ) ,
and it remains to enlarge C ′ C' C ′ .
(d) The first identity is the moment formula E Θ − α k = Γ ( 1 + k ) Γ ( 1 + α k ) \mathbb{E}\Theta^{-\alpha k} = \frac{\Gamma(1+k)}{\Gamma(1+\alpha k)} E Θ − α k = Γ ( 1 + α k ) Γ ( 1 + k ) [15 Section XIII.8] , see also Pollard [39 ] . Since ∑ k x k E Y k k ! = E α ( x ) \sum_k \frac{x^k\mathbb{E}Y_k}{k!} = E_\alpha(x) ∑ k k ! x k E Y k = E α ( x ) , the second follows for x ≥ 0 x \ge0 x ≥ 0 by monotone convergence and for x < 0 x < 0 x < 0 by dominated convergence.
(e) With s = u − 1 α s = u^{-\frac{1}{\alpha}} s = u − α 1 we have u m α ( u ) = α − 1 s g α ( s ) u m_\alpha(u) = \alpha^{-1}s g_\alpha(s) u m α ( u ) = α − 1 s g α ( s ) , which is bounded because g α g_\alpha g α is bounded and continuous with g α ( s ) = O ( s − 1 − α ) g_\alpha(s) = O(s^{-1-\alpha}) g α ( s ) = O ( s − 1 − α ) as s → ∞ s \to\infty s → ∞ [15 Section XVII.6] . For x ≥ 0 x \ge0 x ≥ 0 one checks after some algebra that π − 1 2 ∫ 0 ∞ e − x u − u 2 4 d u = e x 2 erfc ( x ) = E 1 2 ( − x ) \pi^{-\frac{1}{2}}\int_0^\infty e^{-xu-\frac{u^2}{4}}\,du = e^{x^2}\operatorname{erfc}(x) = E_{\frac{1}{2}}(-x) π − 2 1 ∫ 0 ∞ e − xu − 4 u 2 d u = e x 2 erfc ( x ) = E 2 1 ( − x ) , and a distribution on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) is determined by its Laplace transform. □
Sectorial generators
Here we provide the proof of Theorem 3.3.
Proof. (a) Norm continuity and the bound are immediate from the power series. For Re ω > 0 \operatorname{Re}\omega> 0 Re ω > 0 we have ∫ 0 ∞ e − ω t t α k d t = Γ ( 1 + α k ) ω − 1 − α k \int_0^\infty e^{-\omega t}t^{\alpha k}\,dt = \Gamma(1+\alpha k)\omega^{-1-\alpha k} ∫ 0 ∞ e − ω t t α k d t = Γ ( 1 + α k ) ω − 1 − α k , and ∑ k ∥ L ∥ k ( Re ω ) − 1 − α k < ∞ \sum_k \lVert L\rVert^k(\operatorname{Re}\omega)^{-1-\alpha k} < \infty ∑ k ∥ L ∥ k ( Re ω ) − 1 − α k < ∞ when ∥ L ∥ < ( Re ω ) α \lVert L\rVert< (\operatorname{Re}\omega)^\alpha ∥ L ∥ < ( Re ω ) α . Termwise integration gives ∑ k L k ω − 1 − α k = ω − 1 ( I − ω − α L ) − 1 = ω α − 1 ( ω α − L ) − 1 \sum_k L^k\omega^{-1-\alpha k} = \omega^{-1}(I-\omega^{-\alpha}L)^{-1} = \omega^{\alpha-1}(\omega^\alpha-L)^{-1} ∑ k L k ω − 1 − α k = ω − 1 ( I − ω − α L ) − 1 = ω α − 1 ( ω α − L ) − 1 .
(b) By Theorem Appendix B.1(d),
∫ 0 ∞ m α ( u ) ∥ e t α u L ∥ d u ≤ M E α ( ϖ t α ) < ∞ , \int_0^\infty m_\alpha(u)\lVert e^{t^\alpha uL}\rVert\,du \le M E_\alpha(\varpi t^\alpha) < \infty, ∫ 0 ∞ m α ( u ) ∥ e t α uL ∥ d u ≤ M E α ( ϖ t α ) < ∞ ,
and E α ( ϖ t α ) = O ( e ϖ 1 / α t ) E_\alpha(\varpi t^\alpha) = O(e^{\varpi^{1/\alpha}t}) E α ( ϖ t α ) = O ( e ϖ 1/ α t ) by Theorem Appendix B.1(a), so the integrability condition of Theorem 3.2 holds for every w > ϖ 1 / α w > \varpi^{1/\alpha} w > ϖ 1/ α . Hence S ( t ) x : = ∫ 0 ∞ m α ( u ) e t α u L x d u S(t)x := \int_0^\infty m_\alpha(u)e^{t^\alpha uL}x\,du S ( t ) x := ∫ 0 ∞ m α ( u ) e t α uL x d u converges absolutely, S ( 0 ) = I S(0) = I S ( 0 ) = I , ∥ S ( t ) ∥ ≤ sup s ∥ e s L ∥ \lVert S(t)\rVert\le\sup_s\lVert e^{sL}\rVert ∥ S ( t )∥ ≤ sup s ∥ e s L ∥ , and S S S is strongly continuous by dominated convergence. For λ > 0 \lambda> 0 λ > 0 and s > 0 s > 0 s > 0 put ψ λ ( s ) : = ∫ 0 ∞ e − λ t t − α m α ( s t − α ) d t \psi_\lambda(s) := \int_0^\infty e^{-\lambda t}t^{-\alpha}m_\alpha(st^{-\alpha})\,dt ψ λ ( s ) := ∫ 0 ∞ e − λ t t − α m α ( s t − α ) d t . For μ ≥ 0 \mu\ge0 μ ≥ 0 , Tonelli’s theorem, the substitution s = t α u s=t^\alpha u s = t α u and Theorem Appendix B.1(d) give
∫ 0 ∞ e − μ s ψ λ ( s ) d s = ∫ 0 ∞ e − λ t E α ( − μ t α ) d t = λ α − 1 λ α + μ = ∫ 0 ∞ e − μ s λ α − 1 e − λ α s d s . \int_0^\infty e^{-\mu s}\psi_\lambda(s)\,ds = \int_0^\infty e^{-\lambda t}E_\alpha(-\mu t^\alpha)\,dt = \frac{\lambda^{\alpha-1}}{\lambda^\alpha+\mu} = \int_0^\infty e^{-\mu s}\lambda^{\alpha-1}e^{-\lambda^\alpha s}\,ds. ∫ 0 ∞ e − μ s ψ λ ( s ) d s = ∫ 0 ∞ e − λ t E α ( − μ t α ) d t = λ α + μ λ α − 1 = ∫ 0 ∞ e − μ s λ α − 1 e − λ α s d s .
The second equality holds for λ > μ 1 / α \lambda> \mu^{1/\alpha} λ > μ 1/ α by (a) on X = C X = \mathbb{C} X = C , and then for every λ > 0 \lambda> 0 λ > 0 by analytic continuation, since both sides are holomorphic in λ ∈ H \lambda\in\mathbb{H} λ ∈ H . By uniqueness of the Laplace transform, ψ λ ( s ) = λ α − 1 e − λ α s \psi_\lambda(s) = \lambda^{\alpha-1}e^{-\lambda^\alpha s} ψ λ ( s ) = λ α − 1 e − λ α s for almost every s > 0 s > 0 s > 0 . Consequently, for λ α > ϖ \lambda^\alpha> \varpi λ α > ϖ , Fubini’s theorem and the same substitution give
∫ 0 ∞ e − λ t S ( t ) x d t = ∫ 0 ∞ ψ λ ( s ) e s L x d s = λ α − 1 ∫ 0 ∞ e − λ α s e s L x d s = λ α − 1 ( λ α − L ) − 1 x . \int_0^\infty e^{-\lambda t}S(t)x\,dt = \int_0^\infty\psi_\lambda(s)e^{sL}x\,ds = \lambda^{\alpha-1}\int_0^\infty e^{-\lambda^\alpha s}e^{sL}x\,ds = \lambda^{\alpha-1}(\lambda^\alpha-L)^{-1}x. ∫ 0 ∞ e − λ t S ( t ) x d t = ∫ 0 ∞ ψ λ ( s ) e s L x d s = λ α − 1 ∫ 0 ∞ e − λ α s e s L x d s = λ α − 1 ( λ α − L ) − 1 x .
(c) By the Lumer–Phillips theorem ∥ e s L ∥ ≤ 1 \lVert e^{sL}\rVert\le1 ∥ e s L ∥ ≤ 1 [14 ] ; for α < 1 \alpha< 1 α < 1 apply (b). □
Proof of Theorem 3.4. If ∣ Arg ω ∣ ≤ φ \lvert\operatorname{Arg}\omega\rvert\le\varphi ∣ Arg ω ∣ ≤ φ and ω ≠ 0 \omega\ne0 ω = 0 , then ω α \omega^\alpha ω α lies in the closure of Σ ( ψ ′ ) \Sigma(\psi') Σ ( ψ ′ ) , which is contained in Σ ( ψ ) ∪ { 0 } \Sigma(\psi) \cup\{0\} Σ ( ψ ) ∪ { 0 } . By continuity of the resolvent,
∥ ω α − 1 ( ω α − L ) − 1 ∥ ≤ s ψ ′ ( L ) ∣ ω ∣ − 1 ( 0 < ∣ ω ∣ , ∣ Arg ω ∣ ≤ φ ) . (26) \left\lVert\omega^{\alpha-1}(\omega^\alpha-L)^{-1}\right\rVert\le s_{\psi'}(L)\lvert\omega\rvert^{-1} \qquad(0 < \lvert\omega\rvert,\ \lvert\operatorname{Arg}\omega\rvert\le\varphi). \tag*{(26)} ω α − 1 ( ω α − L ) − 1 ≤ s ψ ′ ( L ) ∣ ω ∣ − 1 ( 0 < ∣ ω ∣ , ∣ Arg ω ∣ ≤ φ ) . (26)
Let S ( t ) S(t) S ( t ) denote the right-hand side of (6) . On the rays ∣ e ω t ∣ = e − t ϱ ∣ cos φ ∣ \lvert e^{\omega t}\rvert= e^{-t\varrho\lvert\cos\varphi\rvert} ∣ e ω t ∣ = e − tϱ ∣ c o s φ ∣ , so the integral converges absolutely for t > 0 t > 0 t > 0 and is norm continuous in t > 0 t > 0 t > 0 . The integrand is holomorphic on { ω ≠ 0 : ∣ Arg ω ∣ < φ + ε } \{\omega\ne0 : \lvert\operatorname{Arg}\omega\rvert< \varphi+\varepsilon\} { ω = 0 : ∣ Arg ω ∣ < φ + ε } for small ε > 0 \varepsilon> 0 ε > 0 , because α ( φ + ε ) < ψ \alpha(\varphi+\varepsilon) < \psi α ( φ + ε ) < ψ . By Cauchy’s theorem the unit arc may be replaced by the arc of radius 1 t \frac{1}{t} t 1 , and the substitution ω = s t \omega=\frac{s}{t} ω = t s together with (B.1) gives ∥ S ( t ) ∥ ≤ c φ s ψ ′ ( L ) \|S(t)\|\le c_{\varphi}s_{\psi'}(L) ∥ S ( t ) ∥ ≤ c φ s ψ ′ ( L ) for t > 0 t>0 t > 0 .
Strong continuity at 0. Let x ∈ D ( L ) x\in D(L) x ∈ D ( L ) . Since ω α − 1 ( ω α − L ) − 1 x = ω − 1 x + ω − 1 ( ω α − L ) − 1 L x \omega^{\alpha-1}(\omega^\alpha-L)^{-1}x=\omega^{-1}x+\omega^{-1}(\omega^\alpha-L)^{-1}Lx ω α − 1 ( ω α − L ) − 1 x = ω − 1 x + ω − 1 ( ω α − L ) − 1 Lx and ( 2 π i ) − 1 ∫ H a φ e ω t ω − 1 d ω = 1 (2\pi i)^{-1}\int_{\mathrm{Ha}_{\varphi}}e^{\omega t}\omega^{-1}\,d\omega=1 ( 2 π i ) − 1 ∫ Ha φ e ω t ω − 1 d ω = 1 by Hankel’s formula, we have
S ( t ) x − x = 1 2 π i ∫ H a φ e ω t ω − 1 ( ω α − L ) − 1 L x d ω . S(t)x-x=\frac{1}{2\pi i}\int_{\mathrm{Ha}_{\varphi}}e^{\omega t}\omega^{-1}(\omega^\alpha-L)^{-1}Lx\,d\omega. S ( t ) x − x = 2 π i 1 ∫ Ha φ e ω t ω − 1 ( ω α − L ) − 1 Lx d ω .
For 0 < t ≤ 1 0<t\le1 0 < t ≤ 1 the integrand is bounded by e s ψ ′ ( L ) ∥ L x ∥ ∣ ω ∣ − 1 − α e\,s_{\psi'}(L)\|Lx\||\omega|^{-1-\alpha} e s ψ ′ ( L ) ∥ Lx ∥∣ ω ∣ − 1 − α , which is integrable on H a φ \mathrm{Ha}_{\varphi} Ha φ , so by dominated convergence S ( t ) x − x S(t)x-x S ( t ) x − x converges as t ↓ 0 t\downarrow0 t ↓ 0 to ( 2 π i ) − 1 ∫ H a φ ω − 1 ( ω α − L ) − 1 L x d ω (2\pi i)^{-1}\int_{\mathrm{Ha}_{\varphi}}\omega^{-1}(\omega^\alpha-L)^{-1}Lx\,d\omega ( 2 π i ) − 1 ∫ Ha φ ω − 1 ( ω α − L ) − 1 Lx d ω . This integral vanishes: close the contour by the arcs { R e i ϑ : ∣ ϑ ∣ ≤ φ } \{Re^{i\vartheta}:|\vartheta|\le\varphi\} { R e i ϑ : ∣ ϑ ∣ ≤ φ } , on which the integrand is O ( R − 1 − α ) O(R^{-1-\alpha}) O ( R − 1 − α ) , and apply Cauchy’s theorem in { 1 < ∣ ω ∣ < R , ∣ Arg ω ∣ < φ } \{1<|\omega|<R,\ |\operatorname{Arg}\omega|<\varphi\} { 1 < ∣ ω ∣ < R , ∣ Arg ω ∣ < φ } . Since D ( L ) D(L) D ( L ) is dense and S S S is bounded on ( 0 , 1 ] (0,1] ( 0 , 1 ] , S ( t ) x → x S(t)x\to x S ( t ) x → x for every x x x ; we set S ( 0 ) : = I S(0):=I S ( 0 ) := I .
Laplace transform. Let λ > 1 \lambda>1 λ > 1 . On H a φ \mathrm{Ha}_{\varphi} Ha φ we have Re ω ≤ 1 < λ \operatorname{Re}\omega\le1<\lambda Re ω ≤ 1 < λ and λ − Re ω ≥ λ + ϱ ∣ cos φ ∣ \lambda-\operatorname{Re}\omega\ge\lambda+\varrho|\cos\varphi| λ − Re ω ≥ λ + ϱ ∣ cos φ ∣ on the rays, so Fubini’s theorem gives
∫ 0 ∞ e − λ t S ( t ) d t = ( 2 π i ) − 1 ∫ H a φ ( λ − ω ) − 1 ω α − 1 ( ω α − L ) − 1 d ω . \int_0^\infty e^{-\lambda t}S(t)\,dt=(2\pi i)^{-1}\int_{\mathrm{Ha}_{\varphi}}(\lambda-\omega)^{-1}\omega^{\alpha-1}(\omega^\alpha-L)^{-1}\,d\omega. ∫ 0 ∞ e − λ t S ( t ) d t = ( 2 π i ) − 1 ∫ Ha φ ( λ − ω ) − 1 ω α − 1 ( ω α − L ) − 1 d ω .
Truncate H a φ \mathrm{Ha}_{\varphi} Ha φ at ∣ ω ∣ = R > λ |\omega|=R>\lambda ∣ ω ∣ = R > λ and close it by the arc { R e i ϑ : ∣ ϑ ∣ ≤ φ } \{Re^{i\vartheta}:|\vartheta|\le\varphi\} { R e i ϑ : ∣ ϑ ∣ ≤ φ } , traversed clockwise, on which the integrand is O ( R − 2 ) O(R^{-2}) O ( R − 2 ) . The resulting closed curve bounds { 1 < ∣ ω ∣ < R , ∣ Arg ω ∣ < φ } \{1<|\omega|<R,\ |\operatorname{Arg}\omega|<\varphi\} { 1 < ∣ ω ∣ < R , ∣ Arg ω ∣ < φ } with negative orientation, and the only singularity inside is the simple pole at ω = λ \omega=\lambda ω = λ , with residue − λ α − 1 ( λ α − L ) − 1 -\lambda^{\alpha-1}(\lambda^\alpha-L)^{-1} − λ α − 1 ( λ α − L ) − 1 . Letting R → ∞ R\to\infty R → ∞ gives (4) .
Time scales in Hilbert spaces
The envelope t ↦ ∥ E α ( t α L ) ∥ t\mapsto\|E_\alpha(t^\alpha L)\| t ↦ ∥ E α ( t α L ) ∥ has a short-time, an intermediate and a long-time regime, each governed by a different resolvent quantity. Let ν ( L ) : = sup Re W ( L ) ∈ ( − ∞ , ∞ ] \nu(L):=\sup\operatorname{Re}W(L)\in(-\infty,\infty] ν ( L ) := sup Re W ( L ) ∈ ( − ∞ , ∞ ] be the numerical abscissa of an operator on H H H .
Proposition Appendix B.3 (Time scales). Let 0 < α ≤ 1 0<\alpha\le1 0 < α ≤ 1 and let L L L be closed and densely defined on H H H , generating a Mittag–Leffler family.
(a) For every v ∈ D ( L ) v\in D(L) v ∈ D ( L ) with ∥ v ∥ = 1 \|v\|=1 ∥ v ∥ = 1 , as t ↓ 0 t\downarrow0 t ↓ 0 , ∥ E α ( t α L ) v ∥ = 1 + Re ⟨ L v , v ⟩ t α Γ ( 1 + α ) + o ( t α ) \|E_\alpha(t^\alpha L)v\|=1+\frac{\operatorname{Re}\langle Lv,v\rangle t^\alpha}{\Gamma(1+\alpha)}+o(t^\alpha) ∥ E α ( t α L ) v ∥ = 1 + Γ ( 1 + α ) Re ⟨ Lv , v ⟩ t α + o ( t α ) .
(b) If L ∈ B ( H ) L\in\mathcal{B}(H) L ∈ B ( H ) , then ∥ E α ( t α L ) ∥ = 1 + ν ( L ) t α Γ ( 1 + α ) + O ( t 2 α ) \|E_\alpha(t^\alpha L)\|=1+\frac{\nu(L)t^\alpha}{\Gamma(1+\alpha)}+O(t^{2\alpha}) ∥ E α ( t α L ) ∥ = 1 + Γ ( 1 + α ) ν ( L ) t α + O ( t 2 α ) .
(c) ∥ E α ( t α L ) ∥ ≤ 1 \|E_\alpha(t^\alpha L)\|\le1 ∥ E α ( t α L ) ∥ ≤ 1 for all t ≥ 0 t\ge0 t ≥ 0 if and only if L L L is m m m -dissipative.
(d) Let 0 < α < 1 0<\alpha<1 0 < α < 1 , and suppose that L L L satisfies the hypothesis of Theorem 3.4 for some ψ ∈ ( α π 2 , π ] \psi\in\left(\frac{\alpha\pi}{2},\pi\right] ψ ∈ ( 2 α π , π ] and that 0 ∈ ρ ( L ) 0\in\rho(L) 0 ∈ ρ ( L ) . Then for every q ∈ N q\in\mathbb{N} q ∈ N , in operator norm,
E α ( t α L ) = − ∑ k = 1 q t − α k L − k Γ ( 1 − α k ) + O ( t − α ( q + 1 ) ) ( t → ∞ ) , E_\alpha(t^\alpha L)=-\sum_{k=1}^{q}\frac{t^{-\alpha k}L^{-k}}{\Gamma(1-\alpha k)}+O\left(t^{-\alpha(q+1)}\right)\qquad(t\to\infty), E α ( t α L ) = − k = 1 ∑ q Γ ( 1 − α k ) t − α k L − k + O ( t − α ( q + 1 ) ) ( t → ∞ ) , and in particular ∥ E α ( t α L ) ∥ = ∥ L − 1 ∥ t − α Γ ( 1 − α ) + O ( t − 2 α ) \|E_\alpha(t^\alpha L)\|=\frac{\|L^{-1}\|t^{-\alpha}}{\Gamma(1-\alpha)}+O(t^{-2\alpha}) ∥ E α ( t α L ) ∥ = Γ ( 1 − α ) ∥ L − 1 ∥ t − α + O ( t − 2 α ) .
Proof. (a) Let c t : = t α Γ ( 1 + α ) = ∫ 0 t k α ( t − s ) d s c_t:=\frac{t^\alpha}{\Gamma(1+\alpha)}=\int_0^t k_\alpha(t-s)\,ds c t := Γ ( 1 + α ) t α = ∫ 0 t k α ( t − s ) d s . By (5) , ∥ E α ( t α L ) v − v − c t L v ∥ \|E_\alpha(t^\alpha L)v-v-c_tLv\| ∥ E α ( t α L ) v − v − c t Lv ∥ is bounded by c t sup 0 ≤ s ≤ t ∥ E α ( s α L ) L v − L v ∥ = o ( c t ) c_t\sup_{0\le s\le t}\|E_\alpha(s^\alpha L)Lv-Lv\|=o(c_t) c t sup 0 ≤ s ≤ t ∥ E α ( s α L ) Lv − Lv ∥ = o ( c t ) , and ∥ v + c t L v ∥ 2 = 1 + 2 c t Re ⟨ L v , v ⟩ + c t 2 ∥ L v ∥ 2 \|v+c_tLv\|^2=1+2c_t\operatorname{Re}\langle Lv,v\rangle+c_t^2\|Lv\|^2 ∥ v + c t Lv ∥ 2 = 1 + 2 c t Re ⟨ Lv , v ⟩ + c t 2 ∥ Lv ∥ 2 .
(b) E α ( t α L ) = I + c t L + O ( t 2 α ) E_\alpha(t^\alpha L)=I+c_tL+O(t^{2\alpha}) E α ( t α L ) = I + c t L + O ( t 2 α ) , and taking the supremum over unit vectors in the identity of (a) gives ∥ I + c t L ∥ = 1 + c t ν ( L ) + O ( c t 2 ) \|I+c_tL\|=1+c_t\nu(L)+O(c_t^2) ∥ I + c t L ∥ = 1 + c t ν ( L ) + O ( c t 2 ) . (c) If L L L is m-dissipative, apply Theorem 3.3(c). Conversely, (a) gives Re ⟨ L v , v ⟩ ≤ 0 \operatorname{Re}\langle Lv,v\rangle\le0 Re ⟨ Lv , v ⟩ ≤ 0 , and Theorem 5.1 gives ( 0 , ∞ ) ⊂ Σ α ⊂ ρ ( L ) (0,\infty)\subset\Sigma_\alpha\subset\rho(L) ( 0 , ∞ ) ⊂ Σ α ⊂ ρ ( L ) .
(d) Fix ψ ′ \psi' ψ ′ and φ = ψ ′ α \varphi=\frac{\psi'}{\alpha} φ = α ψ ′ as in Theorem 3.4, so that (6) holds. The closed set Σ ( ψ ′ ) ‾ = { z : z = 0 or ∣ Arg z ∣ ≤ ψ ′ } \overline{\Sigma(\psi')}=\{z:z=0\text{ or }|\operatorname{Arg}z|\le\psi'\} Σ ( ψ ′ ) = { z : z = 0 or ∣ Arg z ∣ ≤ ψ ′ } lies in ρ ( L ) \rho(L) ρ ( L ) , since 0 ∈ ρ ( L ) 0\in\rho(L) 0 ∈ ρ ( L ) and Σ ( ψ ′ ) ‾ ∖ { 0 } ⊂ Σ ( ψ ) \overline{\Sigma(\psi')}\setminus\{0\}\subset\Sigma(\psi) Σ ( ψ ′ ) ∖ { 0 } ⊂ Σ ( ψ ) . Hence the quantity c L : = sup { ∥ ( z − L ) − 1 ∥ : z ∈ Σ ( ψ ′ ) ‾ } c_L:=\sup\{\|(z-L)^{-1}\|:z\in\overline{\Sigma(\psi')}\} c L := sup { ∥ ( z − L ) − 1 ∥ : z ∈ Σ ( ψ ′ ) } is finite: for ∣ z ∣ ≥ 1 |z|\ge1 ∣ z ∣ ≥ 1 the resolvent is bounded by s ψ ′ ′ ( L ) ∣ z ∣ \frac{s_{\psi''}(L)}{|z|} ∣ z ∣ s ψ ′′ ( L ) with ψ ′ < ψ ′ ′ < ψ \psi'<\psi''<\psi ψ ′ < ψ ′′ < ψ , and on the compact remainder it is continuous. For z ∈ ρ ( L ) z\in\rho(L) z ∈ ρ ( L ) we have ( z − L ) − 1 = − L − 1 + z L − 1 ( z − L ) − 1 (z-L)^{-1}=-L^{-1}+zL^{-1}(z-L)^{-1} ( z − L ) − 1 = − L − 1 + z L − 1 ( z − L ) − 1 , and by induction ( z − L ) − 1 = − ∑ k = 0 q − 1 z k L − k − 1 + z q L − q ( z − L ) − 1 (z-L)^{-1}=-\sum_{k=0}^{q-1}z^kL^{-k-1}+z^qL^{-q}(z-L)^{-1} ( z − L ) − 1 = − ∑ k = 0 q − 1 z k L − k − 1 + z q L − q ( z − L ) − 1 . Insert this with z = ω α z=\omega^\alpha z = ω α into (6) . Hankel’s formula
( 2 π i ) − 1 ∫ H a φ e ω t ω − β d ω = t β − 1 Γ ( β ) (2\pi i)^{-1}\int_{\mathrm{Ha}_\varphi}e^{\omega t}\omega^{-\beta}\,d\omega=\frac{t^{\beta-1}}{\Gamma(\beta)} ( 2 π i ) − 1 ∫ Ha φ e ω t ω − β d ω = Γ ( β ) t β − 1
with β = 1 − α ( k + 1 ) \beta=1-\alpha(k+1) β = 1 − α ( k + 1 ) gives the terms of the expansion. In the remainder substitute ω = s t \omega=\frac{s}{t} ω = t s and return to H a φ \mathrm{Ha}_\varphi Ha φ by Cauchy’s theorem, as in the proof of Theorem 3.4; this bounds its norm by t − α ( q + 1 ) ( 2 π ) − 1 c L ∥ L − q ∥ ∫ H a φ ∣ e s ∣ ∣ s ∣ α ( q + 1 ) − 1 ∣ d s ∣ t^{-\alpha(q+1)}(2\pi)^{-1}c_L\|L^{-q}\|\int_{\mathrm{Ha}_\varphi}|e^s||s|^{\alpha(q+1)-1}|ds| t − α ( q + 1 ) ( 2 π ) − 1 c L ∥ L − q ∥ ∫ Ha φ ∣ e s ∣∣ s ∣ α ( q + 1 ) − 1 ∣ d s ∣ .
Thus, as for semigroups, a Mittag–Leffler family is contractive exactly when its generator is m-dissipative, independently of α \alpha α , and the initial slope in the variable t α t^\alpha t α is ν ( L ) Γ ( 1 + α ) \frac{\nu(L)}{\Gamma(1+\alpha)} Γ ( 1 + α ) ν ( L ) . In physical time the envelope has an infinite slope at t = 0 t=0 t = 0 whenever α < 1 \alpha<1 α < 1 and ν ( L ) > 0 \nu(L)>0 ν ( L ) > 0 . The hypothesis of (d) holds for bounded L L L with σ ( L ) ∩ Σ α ‾ = ∅ \sigma(L)\cap\overline{\Sigma_\alpha}=\varnothing σ ( L ) ∩ Σ α = ∅ , for normal L L L with σ ( L ) ⊂ { ∣ Arg z ∣ ≥ ψ } ∖ { 0 } \sigma(L)\subset\{|\operatorname{Arg}z|\ge\psi\}\setminus\{0\} σ ( L ) ⊂ { ∣ Arg z ∣ ≥ ψ } ∖ { 0 } , and for L L L on H H H with 0 ∈ ρ ( L ) 0\in\rho(L) 0 ∈ ρ ( L ) , W ( L ) ⊂ { ∣ Arg z ∣ ≥ ψ } W(L)\subset\{|\operatorname{Arg}z|\ge\psi\} W ( L ) ⊂ { ∣ Arg z ∣ ≥ ψ } and ρ ( L ) ∩ Σ α ‾ ≠ ∅ \rho(L)\cap\overline{\Sigma_\alpha}\ne\varnothing ρ ( L ) ∩ Σ α = ∅ , by the spectral theorem and the argument of Theorem 6.4 (a).
Together, Theorem Appendix B.3 and Theorem A describe the regimes of the envelope: the initial slope is governed by the numerical abscissa, the transient maximum by the fractional Kreiss constant, and the algebraic decay by the resolvent at the origin, ∥ L − 1 ∥ \|L^{-1}\| ∥ L − 1 ∥ , rather than by the spectral abscissa. For α = 1 \alpha=1 α = 1 all coefficients 1 Γ ( 1 − k ) \frac{1}{\Gamma(1-k)} Γ ( 1 − k ) 1 vanish, consistently with exponential decay. The envelopes in Figure 3 (a) display all of them.
Ingredients of the optimality construction
Throughout this appendix 0 < γ < 1 0<\gamma<1 0 < γ < 1 , w : = w γ w:=w_\gamma w := w γ and ∥ ⋅ ∥ : = ∥ ⋅ ∥ γ \|\cdot\|:=\|\cdot\|_\gamma ∥ ⋅ ∥ := ∥ ⋅ ∥ γ . Recall that w ∈ A 2 ( T ) w\in A_2(\mathbb{T}) w ∈ A 2 ( T ) if sup I ( ∣ I ∣ − 1 ∫ I w ) ( ∣ I ∣ − 1 ∫ I w − 1 ) < ∞ \sup_I\left(|I|^{-1}\int_Iw\right)\left(|I|^{-1}\int_Iw^{-1}\right)<\infty sup I ( ∣ I ∣ − 1 ∫ I w ) ( ∣ I ∣ − 1 ∫ I w − 1 ) < ∞ , the supremum being taken over all arcs I ⊂ T I\subset\mathbb{T} I ⊂ T .
We recall Theorem 8.1 (a): w γ ∈ A 2 ( T ) w_\gamma\in A_2(\mathbb{T}) w γ ∈ A 2 ( T ) , the Riesz projections R j R_j R j are uniformly bounded on L γ 2 L^2_\gamma L γ 2 with bound r γ < ∞ r_\gamma<\infty r γ < ∞ , and ∥ S N ∥ ≤ 2 r γ \|S_N\|\le2r_\gamma ∥ S N ∥ ≤ 2 r γ .
Proof of Theorem 8.1 (a). Since w w w and w − 1 w^{-1} w − 1 are integrable, the A 2 A_2 A 2 quotient is bounded over arcs of length at least π 2 \frac{\pi}{2} 2 π . An arc of length less than π 2 \frac{\pi}{2} 2 π is contained in { ∣ x ∣ ≤ 3 π 4 } \{|x|\le\frac{3\pi}{4}\} { ∣ x ∣ ≤ 4 3 π } or in { ∣ x ∣ ≥ π 4 } \{|x|\ge\frac{\pi}{4}\} { ∣ x ∣ ≥ 4 π } . As u ≤ tan u ≤ 3 u u\le\tan u\le3u u ≤ tan u ≤ 3 u on [ 0 , 3 π 8 ] [0,\frac{3\pi}{8}] [ 0 , 8 3 π ] , on the first set w ( x ) w(x) w ( x ) is comparable to ∣ x ∣ γ |x|^\gamma ∣ x ∣ γ , and on the second, with y : = π − ∣ x ∣ y:=\pi-|x| y := π − ∣ x ∣ , to y γ y^\gamma y γ . Power weights ∣ t ∣ β |t|^\beta ∣ t ∣ β with ∣ β ∣ < 1 |\beta|<1 ∣ β ∣ < 1 belong to A 2 ( R ) A_2(\mathbb{R}) A 2 ( R ) [18 Example 7.1.7] , and A 2 A_2 A 2 bounds are stable under multiplication by functions bounded above and below. Hence w ∈ A 2 ( T ) w\in A_2(\mathbb{T}) w ∈ A 2 ( T ) . By the theorem of Hunt et al. [19 ] , the conjugate-function operator f ↦ f ~ f\mapsto\widetilde f f ↦ f , e ~ k = − i sgn ( k ) e k \widetilde e_k=-i\operatorname{sgn}(k)e_k e k = − i sgn ( k ) e k , is bounded on L γ 2 L^2_\gamma L γ 2 . For f ∈ P f\in\mathcal{P} f ∈ P ,
R 0 f = 1 2 ( f + i f ~ ) + 1 2 f ^ ( 0 ) e 0 , ∣ f ^ ( 0 ) ∣ ∥ e 0 ∥ ≤ 1 2 π ( ∫ − π π w − 1 ) 1 2 ( ∫ − π π w ) 1 2 ∥ f ∥ . R_0f=\frac{1}{2}(f+i\widetilde f)+\frac{1}{2}\widehat f(0)e_0,\qquad
|\widehat f(0)|\|e_0\|\le\frac{1}{2\pi}\left(\int_{-\pi}^{\pi}w^{-1}\right)^{\frac{1}{2}}\left(\int_{-\pi}^{\pi}w\right)^{\frac{1}{2}}\|f\|. R 0 f = 2 1 ( f + i f ) + 2 1 f ( 0 ) e 0 , ∣ f ( 0 ) ∣∥ e 0 ∥ ≤ 2 π 1 ( ∫ − π π w − 1 ) 2 1 ( ∫ − π π w ) 2 1 ∥ f ∥.
the latter by the Cauchy–Schwarz inequality, so R 0 R_0 R 0 is bounded. Finally, R j = e j R 0 e − j R_j=e_jR_0e_{-j} R j = e j R 0 e − j and multiplication by e ± j e_{\pm j} e ± j is an isometry of L γ 2 L^2_\gamma L γ 2 , so r γ < ∞ r_\gamma<\infty r γ < ∞ , and ∥ S N ∥ ≤ ∥ R − N ∥ + ∥ R N + 1 ∥ ≤ 2 r γ \|S_N\|\le\|R_{-N}\|+\|R_{N+1}\|\le2r_\gamma ∥ S N ∥ ≤ ∥ R − N ∥ + ∥ R N + 1 ∥ ≤ 2 r γ .
We recall (22) : for a ray R θ = { s e i θ : s ≥ 0 } R_\theta= \{se^{i\theta}:s\ge0\} R θ = { s e i θ : s ≥ 0 } and z ∉ R θ z\notin R_\theta z ∈ / R θ , ∫ 0 ∞ ∣ z − s e i θ ∣ − 2 d s ≤ π dist ( z , R θ ) \int_0^\infty|z-se^{i\theta}|^{-2}\,ds \le\frac{\pi}{\operatorname{dist}(z,R_\theta)} ∫ 0 ∞ ∣ z − s e i θ ∣ − 2 d s ≤ dist ( z , R θ ) π .
Proof of (21). Write z e − i θ = u + i v ze^{-i\theta}=u+iv z e − i θ = u + i v ; the integral equals ∫ 0 ∞ ( ( s − u ) 2 + v 2 ) − 1 d s \int_0^\infty\left((s-u)^2+v^2\right)^{-1}\,ds ∫ 0 ∞ ( ( s − u ) 2 + v 2 ) − 1 d s . If v = 0 v=0 v = 0 , then u < 0 u<0 u < 0 and the integral is 1 ∣ u ∣ = 1 dist ( z , R θ ) \frac{1}{|u|}=\frac{1}{\operatorname{dist}(z,R_\theta)} ∣ u ∣ 1 = dist ( z , R θ ) 1 . If v ≠ 0 v\ne0 v = 0 , it equals ∣ v ∣ − 1 ( π 2 + arctan ( u ∣ v ∣ ) ) |v|^{-1}\left(\frac{\pi}{2}+\arctan\left(\frac{u}{|v|}\right)\right) ∣ v ∣ − 1 ( 2 π + arctan ( ∣ v ∣ u ) ) . For u ≥ 0 u\ge0 u ≥ 0 we have dist ( z , R θ ) = ∣ v ∣ \operatorname{dist}(z,R_\theta)=|v| dist ( z , R θ ) = ∣ v ∣ , and the bound is clear. For u < 0 u<0 u < 0 we have dist ( z , R θ ) = ∣ z ∣ \operatorname{dist}(z,R_\theta)=|z| dist ( z , R θ ) = ∣ z ∣ , and with ϕ : = arctan ( ∣ v ∣ ∣ u ∣ ) ∈ ( 0 , π 2 ) \phi:=\arctan\left(\frac{|v|}{|u|}\right)\in\left(0,\frac{\pi}{2}\right) ϕ := arctan ( ∣ u ∣ ∣ v ∣ ) ∈ ( 0 , 2 π ) the integral equals ϕ ∣ z ∣ sin ϕ ≤ π 2 ∣ z ∣ \frac{\phi}{|z|\sin\phi}\le\frac{\pi}{2|z|} ∣ z ∣ s i n ϕ ϕ ≤ 2∣ z ∣ π .
We recall (23) : b α ( s ) = E α ( ( i s ) α ) − α − 1 e i s b_\alpha(s)=E_\alpha((is)^\alpha)-\alpha^{-1}e^{is} b α ( s ) = E α (( i s ) α ) − α − 1 e i s is continuous on R \mathbb{R} R , bounded and of bounded variation, and b 1 ≡ 0 b_1\equiv0 b 1 ≡ 0 .
Proof of (22). For α = 1 \alpha=1 α = 1 we have Φ 1 ( s ) = e i s \Phi_1(s)=e^{is} Φ 1 ( s ) = e i s . Let 0 < α < 1 0<\alpha<1 0 < α < 1 . Since E α E_\alpha E α has real Taylor coefficients and ( − i s ) α = ( i s ) α ‾ (-is)^\alpha=\overline{(is)^\alpha} ( − i s ) α = ( i s ) α for s > 0 s>0 s > 0 , we have b α ( − s ) = b α ( s ) ‾ b_\alpha(-s)=\overline{b_\alpha(s)} b α ( − s ) = b α ( s ) , so it suffices to work on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . Put σ α : = 2 1 / α \sigma_\alpha:=2^{1/\alpha} σ α := 2 1/ α . For s ≥ σ α s\ge\sigma_\alpha s ≥ σ α the point z = ( i s ) α = s α e i α π / 2 z=(is)^\alpha=s^\alpha e^{i\alpha\pi/2} z = ( i s ) α = s α e i α π /2 satisfies ∣ z ∣ ≥ 2 |z|\ge2 ∣ z ∣ ≥ 2 , ∣ Arg z ∣ = α π 2 |\operatorname{Arg}z|=\frac{\alpha\pi}{2} ∣ Arg z ∣ = 2 α π and z 1 / α = i s z^{1/\alpha}=is z 1/ α = i s . Hence b α ( s ) = q α ( z ) b_\alpha(s)=q_\alpha(z) b α ( s ) = q α ( z ) , and Theorem Appendix B.1(c) gives
∣ b α ( s ) ∣ ≤ C ′ s − α , ∣ b α ′ ( s ) ∣ = ∣ q α ′ ( z ) ∣ α s α − 1 ≤ α C ′ s − 1 − α ( s ≥ σ α ) . |b_\alpha(s)|\le C's^{-\alpha},\qquad|b'_\alpha(s)|=|q'_\alpha(z)|\,\alpha s^{\alpha-1}\le\alpha C's^{-1-\alpha}\qquad(s\ge\sigma_\alpha). ∣ b α ( s ) ∣ ≤ C ′ s − α , ∣ b α ′ ( s ) ∣ = ∣ q α ′ ( z ) ∣ α s α − 1 ≤ α C ′ s − 1 − α ( s ≥ σ α ) .
On [ 0 , σ α ] [0,\sigma_\alpha] [ 0 , σ α ] we have Φ α ( s ) = ∑ k ≥ 0 e i α k π / 2 Γ ( 1 + α k ) s α k \Phi_\alpha(s)=\sum_{k\ge0}\frac{e^{i\alpha k\pi/2}}{\Gamma(1+\alpha k)}s^{\alpha k} Φ α ( s ) = ∑ k ≥ 0 Γ ( 1 + α k ) e i α k π /2 s α k , so ∫ 0 σ α ∣ Φ α ′ ∣ ≤ ∑ k ≥ 1 α k Γ ( 1 + α k ) σ α α k = E α ( 2 ) − 1 \int_0^{\sigma_\alpha}|\Phi'_\alpha|\le\sum_{k\ge1}\frac{\alpha k}{\Gamma(1+\alpha k)}\sigma_\alpha^{\alpha k} = E_\alpha(2)-1 ∫ 0 σ α ∣ Φ α ′ ∣ ≤ ∑ k ≥ 1 Γ ( 1 + α k ) α k σ α α k = E α ( 2 ) − 1 , while s ↦ α − 1 e i s s\mapsto\alpha^{-1}e^{is} s ↦ α − 1 e i s has variation σ α α \frac{\sigma_\alpha}{\alpha} α σ α there. Thus b α b_\alpha b α is continuous and bounded on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , and Var [ 0 , ∞ ) ( b α ) ≤ E α ( 2 ) − 1 + σ α α + α C ′ ∫ σ α ∞ s − 1 − α d s < ∞ \operatorname{Var}_{[0,\infty)}(b_\alpha)\le E_\alpha(2)-1+\frac{\sigma_\alpha}{\alpha}+\alpha C'\int_{\sigma_\alpha}^\infty s^{-1-\alpha}\,ds<\infty Var [ 0 , ∞ ) ( b α ) ≤ E α ( 2 ) − 1 + α σ α + α C ′ ∫ σ α ∞ s − 1 − α d s < ∞ .
We recall (24) : ∥ D N ( ⋅ + π ) ∥ γ ≥ c γ N γ ∥ D N ∥ γ \|D_N(\cdot+\pi)\|_\gamma\ge c_\gamma N^\gamma\|D_N\|_\gamma ∥ D N ( ⋅ + π ) ∥ γ ≥ c γ N γ ∥ D N ∥ γ for all N ∈ N N\in\mathbb{N} N ∈ N , with c γ c_\gamma c γ and Q γ Q_\gamma Q γ as stated there.
Proof of (23). Since sin ( y / 2 ) ≥ y / π \sin(y/2)\ge y/\pi sin ( y /2 ) ≥ y / π on [ 0 , π ] [0,\pi] [ 0 , π ] , we have ∣ D N ( y ) ∣ ≤ min ( 2 N + 1 , π ∣ y ∣ ) |D_N(y)|\le\min\left(2N+1,\frac{\pi}{|y|}\right) ∣ D N ( y ) ∣ ≤ min ( 2 N + 1 , ∣ y ∣ π ) for 0 < ∣ y ∣ ≤ π 0<|y|\le\pi 0 < ∣ y ∣ ≤ π , and ∣ D N ( y ) ∣ ≤ 2 |D_N(y)|\le\sqrt{2} ∣ D N ( y ) ∣ ≤ 2 for π 2 ≤ ∣ y ∣ ≤ π \frac{\pi}{2}\le|y|\le\pi 2 π ≤ ∣ y ∣ ≤ π . Since tan u ≤ 2 u \tan u\le2u tan u ≤ 2 u on [ 0 , π 4 ] \left[0,\frac{\pi}{4}\right] [ 0 , 4 π ] , we have w ( y ) ≤ ∣ y ∣ γ w(y)\le|y|^\gamma w ( y ) ≤ ∣ y ∣ γ for ∣ y ∣ ≤ π 2 |y|\le\frac{\pi}{2} ∣ y ∣ ≤ 2 π . Splitting at ∣ y ∣ = π 2 |y|=\frac{\pi}{2} ∣ y ∣ = 2 π , and the integral over [ 0 , π 2 ] \left[0,\frac{\pi}{2}\right] [ 0 , 2 π ] at y N : = 1 2 N + 1 y_N:=\frac{1}{2N+1} y N := 2 N + 1 1 ,
∥ D N ∥ 2 2 ≤ 2 ∫ 0 y N ( 2 N + 1 ) 2 y γ d y + 2 ∫ y N π / 2 π 2 y γ − 2 d y + 4 ∫ π / 2 π tan γ y 2 d y ≤ Q γ ( 2 N + 1 ) 1 − γ . \|D_N\|_2^2\le2\int_0^{y_N}(2N+1)^2y^\gamma\,dy+2\int_{y_N}^{\pi/2}\pi^2y^{\gamma-2}\,dy+4\int_{\pi/2}^{\pi}\tan^\gamma\frac{y}{2}\,dy\le Q_\gamma(2N+1)^{1-\gamma}. ∥ D N ∥ 2 2 ≤ 2 ∫ 0 y N ( 2 N + 1 ) 2 y γ d y + 2 ∫ y N π /2 π 2 y γ − 2 d y + 4 ∫ π /2 π tan γ 2 y d y ≤ Q γ ( 2 N + 1 ) 1 − γ .
Next, the substitution y = x + π y=x+\pi y = x + π and ∣ tan ( ( y − π ) / 2 ) ∣ = ∣ cot ( y / 2 ) ∣ |\tan((y-\pi)/2)|=|\cot(y/2)| ∣ tan (( y − π ) /2 ) ∣ = ∣ cot ( y /2 ) ∣ give
∥ D N ( ⋅ + π ) ∥ 2 2 = ∫ − π π ∣ D N ( y ) ∣ 2 ∣ cot y 2 ∣ γ d y . \|D_N(\cdot+\pi)\|_2^2=\int_{-\pi}^{\pi}|D_N(y)|^2\left|\cot\frac{y}{2}\right|^\gamma\,dy. ∥ D N ( ⋅ + π ) ∥ 2 2 = ∫ − π π ∣ D N ( y ) ∣ 2 cot 2 y γ d y .
For 0 < ∣ y ∣ ≤ y N 0<|y|\le y_N 0 < ∣ y ∣ ≤ y N we have ( N + 1 2 ) ∣ y ∣ ≤ 1 2 (N+\frac12)|y|\le\frac12 ( N + 2 1 ) ∣ y ∣ ≤ 2 1 , hence D N ( y ) ≥ 2 π ( 2 N + 1 ) ≥ N D_N(y)\ge\frac{2}{\pi}(2N+1)\ge N D N ( y ) ≥ π 2 ( 2 N + 1 ) ≥ N , using sin u ≥ 2 u π \sin u\ge\frac{2u}{\pi} sin u ≥ π 2 u on [ 0 , π 2 ] \left[0,\frac{\pi}{2}\right] [ 0 , 2 π ] and sin u ≤ u \sin u\le u sin u ≤ u , and moreover ∣ cot y 2 ∣ ≥ 1 ∣ y ∣ \left|\cot\frac{y}{2}\right|\ge\frac{1}{|y|} cot 2 y ≥ ∣ y ∣ 1 . Therefore ∥ D N ( ⋅ + π ) ∥ 2 2 ≥ 2 N 2 ∫ 0 y N y − γ d y = 2 N 2 ( 2 N + 1 ) γ − 1 1 − γ \|D_N(\cdot+\pi)\|_2^2\ge2N^2\int_0^{y_N}y^{-\gamma}\,dy=\frac{2N^2(2N+1)^{\gamma-1}}{1-\gamma} ∥ D N ( ⋅ + π ) ∥ 2 2 ≥ 2 N 2 ∫ 0 y N y − γ d y = 1 − γ 2 N 2 ( 2 N + 1 ) γ − 1 . Dividing the two estimates and using 2 N + 1 ≤ 3 N 2N+1\le3N 2 N + 1 ≤ 3 N and 2 γ − 2 < 0 2\gamma-2<0 2 γ − 2 < 0 ,
∥ D N ( ⋅ + π ) ∥ 2 2 ∥ D N ∥ 2 2 ≥ 2 N 2 ( 2 N + 1 ) 2 γ − 2 ( 1 − γ ) Q γ ≥ 2 ⋅ 3 2 γ − 2 ( 1 − γ ) Q γ N 2 γ = c γ 2 N 2 γ . \frac{\|D_N(\cdot+\pi)\|_2^2}{\|D_N\|_2^2}\ge\frac{2N^2(2N+1)^{2\gamma-2}}{(1-\gamma)Q_\gamma}\ge\frac{2\cdot3^{2\gamma-2}}{(1-\gamma)Q_\gamma}N^{2\gamma}=c_\gamma^2N^{2\gamma}. ∥ D N ∥ 2 2 ∥ D N ( ⋅ + π ) ∥ 2 2 ≥ ( 1 − γ ) Q γ 2 N 2 ( 2 N + 1 ) 2 γ − 2 ≥ ( 1 − γ ) Q γ 2 ⋅ 3 2 γ − 2 N 2 γ = c γ 2 N 2 γ .