MSC (2020): 46E35, 42B35 (primary); 26A33, 47G20, 47G40, 42B15
Introduction
The Riesz fractional gradient was probably first studied in its own right in [20 ] by Horváth and then very prominently in the more recent works [33 ] and [34 ] by Shieh and Spector (see their papers also for further historical comments and mentions of earlier appearances of the same object in the literature), even though it is closely related and owes its name to the Riesz potentials introduced by Riesz already in [29 ] (see also [30 ] ). As established in [33 ] , the function spaces associated to the Riesz fractional gradient coincide with the fractional Sobolev spaces known as Bessel potential spaces, going back to the work [2 ] by Aronszajn and Smith. These spaces have been the object of intensive study since the 1960s and play a key role in the modern fields of potential theory, harmonic analysis and interpolation theory (see, e.g., the classical references [2 , 37 , 9 , 38 , 1 ] , as well as the recent works [6 , 3 ] ), not mentioning many others. However, Riesz fractional gradients are only a model example of a wider class of nonlocal gradients, where the common local gradient is replaced by an integral over the first order difference quotient weighted with a general kernel ([25 , 4 , 13 , 8 ] ). Nonlocal gradients have gained popularity in applied mathematics over the course of the last decades as an alternative approach to models involving conventional differential operators, and important applications comprise peridynamics and nonlocal hyperelasticity (see, e.g., [36 , 24 , 5 ] ). Mathematically speaking, nonlocal gradients provide a means to enforce relaxed regularity requirements, in contrast with classical Sobolev spaces, upon the functions in question, while a lot of elements in the standard functional analytic toolbox, as a generalized fundamental theorem of calculus, Poincaré-type inequalities, compactness properties, and embedding relations, are usually still available in this broader setting ([12 , 13 , 8 ] ).
In this work, we address the question whether the nonlocal Sobolev spaces associated to a certain class of generalized nonlocal gradients can be viewed as generalized Bessel potential spaces. In particular, we are interested in whether the well-known equivalent characterizations of Bessel potential spaces carry over to the setting of generalized nonlocal gradients studied in [8 ] . Our main findings are summarized in the following subsections.
Throughout the paper, we put an emphasis on a thorough derivation of the results and keep the exposition as self-contained as possible, striving, in particular, to make the paper accessible also for the interested reader not already equipped with a deep understanding of the Fourier theoretic and/or functional theoretic aspects of the matter.
1.1. A brief overview over the main results. We are concerned subsequently with ρ \rho ρ -nonlocal Sobolev spaces (just nonlocal Sobolev spaces for short) H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^{n}) H ρ , p ( R n ) according to the following framework: for a radial kernel ρ : R n → [ 0 , ∞ ] \rho:\mathbb{R}^{n}\to[0,\infty] ρ : R n → [ 0 , ∞ ] subject to suitable hypothesis, H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^{n}) H ρ , p ( R n ) is given as
H ρ , p ( R n ) : = { u ∈ L p ( R n ) : D ρ u ∈ L p ( R n ; R n ) } , (1) H^{\rho,p}(\mathbb{R}^{n}):=\{u\in L^{p}(\mathbb{R}^{n}):D^{\rho}u\in L^{p}(\mathbb{R}^{n};\mathbb{R}^{n})\}, \tag*{(1)} H ρ , p ( R n ) := { u ∈ L p ( R n ) : D ρ u ∈ L p ( R n ; R n )} , (1)
where the nonlocal derivative is defined for φ ∈ C c ∞ ( R n ) \varphi\in C^{\infty}_{c}(\mathbb{R}^{n}) φ ∈ C c ∞ ( R n ) by
D ρ φ ( x ) = ∫ R n φ ( x ) − φ ( y ) ∣ x − y ∣ x − y ∣ x − y ∣ ρ ( x − y ) d y , x ∈ R n (2) D^{\rho}\varphi(x)=\int_{\mathbb{R}^{n}}\frac{\varphi(x)-\varphi(y)}{|x-y|}\frac{x-y}{|x-y|}\rho(x-y)\,dy,\qquad x\in\mathbb{R}^{n} \tag*{(2)} D ρ φ ( x ) = ∫ R n ∣ x − y ∣ φ ( x ) − φ ( y ) ∣ x − y ∣ x − y ρ ( x − y ) d y , x ∈ R n (2)
and is extended to L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) by duality (see § 2.2 for more details). This setting agrees with the general approach developed in [8 ] , which comprises as special cases the classical Riesz fractional gradient, as studied, for example, in [33 , 12 , 39 ] , and finite horizon fractional gradients appearing in [4 , 5 , 13 , 22 ] . In particular, in the notation of (1) and (2) , the Riesz fractional gradients correspond to the kernel function ρ s ( x ) = c ( n , s ) ∣ x ∣ − ( n + s − 1 ) \rho_s(x)=c(n,s)|x|^{-(n+s-1)} ρ s ( x ) = c ( n , s ) ∣ x ∣ − ( n + s − 1 ) for s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) and suitable constants c ( n , s ) c(n,s) c ( n , s ) , and the respective nonlocal Sobolev spaces are H ρ s , p ( R n ) H^{\rho_s,p}(\mathbb{R}^{n}) H ρ s , p ( R n ) (see also (14) ). For further works on related general non-local gradients we refer to [16 , 25 , 15 , 14 ] .
We establish in this paper several characterizations of the spaces H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^{n}) H ρ , p ( R n ) and their topological duals, generalizing the corresponding identities for Bessel potential spaces (in § 1.1 we recall classical Bessel potential spaces and elaborate on the precise connections to our results). As a disclaimer, we note upfront that the characterizations below may differ in their assumptions on the kernel function ρ \rho ρ , in particular, in the requirements regarding the behavior of ρ \rho ρ in the origin manifesting in the conditions (H3) and (H3’) (the precise hypotheses can be found in § 2.5 ). First, we demonstrate that, under suitable premises already devised in [8 ] , it holds
H ρ , p ( R n ) = { u ∈ S ′ ( R n ) : F − 1 ( ( 1 + ∣ λ ρ ∣ 2 ) 1 / 2 u ^ ) ∈ L p ( R n ) } (3) H^{\rho,p}(\mathbb{R}^{n}) = \{u \in\mathcal{S}'(\mathbb{R}^{n}) : \mathcal{F}^{-1}((1 + |\lambda_{\rho}|^{2})^{1/2}\widehat{u}) \in L^{p}(\mathbb{R}^{n})\} \tag*{(3)} H ρ , p ( R n ) = { u ∈ S ′ ( R n ) : F − 1 (( 1 + ∣ λ ρ ∣ 2 ) 1/2 u ) ∈ L p ( R n )} (3)
for λ ρ \lambda_{\rho} λ ρ the Fourier symbol of the nonlocal gradient. This result can be found in Theorem 3.1 . A similar identity is true for sufficiently regular bounded domains, as demonstrated in Corollary 3.4 . We also derive a related representation of the topological dual space
H − ρ , p ′ ( R n ) : = ( H ρ , p ( R n ) ) ′ . H^{-\rho,p'}(\mathbb{R}^{n}) := (H^{\rho,p}(\mathbb{R}^{n}))'. H − ρ , p ′ ( R n ) := ( H ρ , p ( R n ) ) ′ .
In particular, one has that
H − ρ , p ′ ( R n ) = { u ∈ S ′ ( R n ) : F − 1 ( ( 1 + ∣ λ ρ ∣ 2 ) − 1 / 2 u ^ ) ∈ L p ′ ( R n ) } . (4) H^{-\rho,p'}(\mathbb{R}^{n}) = \{u \in\mathcal{S}'(\mathbb{R}^{n}) : \mathcal{F}^{-1}((1 + |\lambda_{\rho}|^{2})^{-1/2}\widehat{u}) \in L^{p'}(\mathbb{R}^{n})\}. \tag*{(4)} H − ρ , p ′ ( R n ) = { u ∈ S ′ ( R n ) : F − 1 (( 1 + ∣ λ ρ ∣ 2 ) − 1/2 u ) ∈ L p ′ ( R n )} . (4)
It also holds a characterization similar to the classical result stating that, for any open Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n , any element in H − ρ , p ′ ( Ω ) H^{-\rho,p'}(\Omega) H − ρ , p ′ ( Ω ) can be identified with some u 0 + div ρ U u_{0} + \operatorname{div}^{\rho} U u 0 + div ρ U for u 0 ∈ L p ( Ω ) u_{0} \in L^{p}(\Omega) u 0 ∈ L p ( Ω ) and U ∈ L p ( Ω ; R n ) U \in L^{p}(\Omega;\mathbb{R}^{n}) U ∈ L p ( Ω ; R n ) . These two findings are presented in Theorem 3.6 and Proposition 3.8 , respectively. Moreover, we deduce that
H ρ , p ( R n ) = { u ∈ L p ( R n ) : ( − Δ ρ ) 1 / 2 u ∈ L p ( R n ) } (5) H^{\rho,p}(\mathbb{R}^{n}) = \{u \in L^{p}(\mathbb{R}^{n}) : (-\Delta^{\rho})^{1/2}u \in L^{p}(\mathbb{R}^{n})\} \tag*{(5)} H ρ , p ( R n ) = { u ∈ L p ( R n ) : ( − Δ ρ ) 1/2 u ∈ L p ( R n )} (5)
= { u ∈ S ′ ( R n ) : ( 1 − Δ ρ ) 1 / 2 u ∈ L p ( R n ) } , (6) = \{u \in\mathcal{S}'(\mathbb{R}^{n}) : (1-\Delta^{\rho})^{1/2}u \in L^{p}(\mathbb{R}^{n})\}, \tag*{(6)} = { u ∈ S ′ ( R n ) : ( 1 − Δ ρ ) 1/2 u ∈ L p ( R n )} , (6)
where Δ ρ \Delta^{\rho} Δ ρ is the distributional nonlocal generalized ρ \rho ρ -Laplacian (see Definition 4.1 ) and ( − Δ ρ ) 1 / 2 (-\Delta^{\rho})^{1/2} ( − Δ ρ ) 1/2 , ( 1 − Δ ρ ) 1 / 2 (1-\Delta^{\rho})^{1/2} ( 1 − Δ ρ ) 1/2 are square roots of the respective squared operators. The identity (1.5) is proved in Proposition 4.12 and Corollary 4.19 , while (1.6) is the subject of Corollary 4.20 . We note that while ( 1 − Δ ρ ) 1 / 2 (1-\Delta^{\rho})^{1/2} ( 1 − Δ ρ ) 1/2 above is understood as an element of S ′ ( R n ) \mathcal{S}'(\mathbb{R}^{n}) S ′ ( R n ) by duality, more work is required to properly introduce and study ( − Δ ρ ) 1 / 2 (-\Delta^{\rho})^{1/2} ( − Δ ρ ) 1/2 as an unbounded operator in L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) .
Finally, we prove that similar to the case of the Riesz fractional gradient and the Bessel potential spaces, we can characterize H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^{n}) H ρ , p ( R n ) as the image of L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) under a suitable convolution operator. Precisely, we establish that there exists g ρ ∈ L 1 ( R n ) g_{\rho} \in L^{1}(\mathbb{R}^{n}) g ρ ∈ L 1 ( R n ) such that
H ρ , p ( R n ) = { g ρ ∗ f : f ∈ L p ( R n ) } . (7) H^{\rho,p}(\mathbb{R}^{n}) = \{g_{\rho} * f : f \in L^{p}(\mathbb{R}^{n})\}. \tag*{(7)} H ρ , p ( R n ) = { g ρ ∗ f : f ∈ L p ( R n )} . (7)
For this result we refer to Theorem 5.3 .
Let us point out that throughout the paper we confine our analysis to the range p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) , since it is the natural setting for the Fourier methods employed in our approach.
We shall conclude this section by sharing a short observation about the naming of the nonlocal Sobolev spaces treated in this work. When introduced in [8 ] , the authors also discuss possible alternative names for the objects H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^{n}) H ρ , p ( R n ) and conclude to give them the name of ρ \rho ρ -nonlocal Sobolev spaces. Although they acknowledge that ρ \rho ρ -Bessel spaces would be another sensible name, this option was ruled out at the time because there was no immediate connection with the Bessel potential. However, the main result of this work (Theorem 5.3 ) bridges this exact gap and shows that they indeed define a natural notion of generalized Bessel potential spaces.
Finally, we also stress that we are not the first ones coining the term generalized Bessel potential spaces. For example, in [35 ] the same name has been used in a different context for other objects, namely, the spaces of functions that are obtained by replacing the Fourier with the Hankel transform and adding a certain weight in the characterization of Bessel spaces of the type (1.3).
Discussion of results and methodology
We start our discussion with a few comments about the Bessel potential spaces. For the following facts, we refer to the classical treatise [37 Chapter V.3] by Stein, the book [26 Section 7] , and also the recent work [6 ] , as well as the references therein.
To set the stage, we define for s ∈ R s \in\mathbb{R} s ∈ R the operators
Λ s : S ′ ( R n ) → S ′ ( R n ) , f ↦ F − 1 ( ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) − s / 2 f ^ ) \Lambda_s : \mathcal{S}'(\mathbb{R}^n) \to\mathcal{S}'(\mathbb{R}^n), \qquad f \mapsto\mathcal{F}^{-1}\left((1+4\pi^2|\cdot|^2)^{-s/2}\widehat{f}\right) Λ s : S ′ ( R n ) → S ′ ( R n ) , f ↦ F − 1 ( ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) − s /2 f )
and the L 1 L^1 L 1 -function
g s ( x ) : = 1 ( 4 π ) s / 2 Γ ( s 2 ) ∫ 0 ∞ e − π ∣ x ∣ 2 / δ e − δ / 4 π δ ( s − n ) / 2 d δ δ , x ∈ R n . (8) g_s(x) := \frac{1}{(4\pi)^{s/2}\Gamma\left(\frac{s}{2}\right)}\int_0^\infty e^{-\pi|x|^2/\delta}e^{-\delta/4\pi}\delta^{(s-n)/2}\frac{d\delta}{\delta}, \qquad x \in\mathbb{R}^n. \tag*{(8)} g s ( x ) := ( 4 π ) s /2 Γ ( 2 s ) 1 ∫ 0 ∞ e − π ∣ x ∣ 2 / δ e − δ /4 π δ ( s − n ) /2 δ d δ , x ∈ R n . (8)
For s > 0 s>0 s > 0 , it can be shown that
g s ^ = ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) − s / 2 , \widehat{g_s}=(1+4\pi^2|\cdot|^2)^{-s/2}, g s = ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) − s /2 ,
which implies that
Λ s f = g s ∗ f \Lambda_s f=g_s*f Λ s f = g s ∗ f
and proves, together with Young’s convolution inequality, that Λ s \Lambda_s Λ s is a well defined operator from L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) into itself for any p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] . Moreover, since g s > 0 g_s>0 g s > 0 and F ( g s ) ( 0 ) = 1 \mathcal{F}(g_s)(0)=1 F ( g s ) ( 0 ) = 1 , it follows that ∥ g s ∥ L 1 ( R n ) = F ( g s ) ( 0 ) = 1 \|g_s\|_{L^1(\mathbb{R}^n)}=\mathcal{F}(g_s)(0)=1 ∥ g s ∥ L 1 ( R n ) = F ( g s ) ( 0 ) = 1 . In particular, g s g_s g s defines a probability measure on R n \mathbb{R}^n R n and, in the terminology of probability theory, the symbol ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) − s / 2 (1+4\pi^2|\cdot|^2)^{-s/2} ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) − s /2 defines its characteristic function .
We are now ready to formally introduce the Bessel potential spaces, which can be done in two equivalent ways, either following Lions and defining them as
{ u ∈ S ′ ( R n ) : Λ − s f ∈ L p ( R n ) } = { u ∈ S ′ ( R n ) : F − 1 ( ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) s / 2 u ^ ) ∈ L p ( R n ) } , (9) \{u\in\mathcal{S}'(\mathbb{R}^n):\Lambda_{-s}f\in L^p(\mathbb{R}^n)\}=\{u\in\mathcal{S}'(\mathbb{R}^n):\mathcal{F}^{-1}((1+4\pi^2|\cdot|^2)^{s/2}\widehat{u})\in L^p(\mathbb{R}^n)\}, \tag*{(9)} { u ∈ S ′ ( R n ) : Λ − s f ∈ L p ( R n )} = { u ∈ S ′ ( R n ) : F − 1 (( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) s /2 u ) ∈ L p ( R n )} , (9)
or in the spirit of Calderón as potential spaces in the sense
Λ s ( L p ( R n ) ) = { Λ s f : f ∈ L p ( R n ) } = { g s ∗ f : f ∈ L p ( R n ) } . (10) \Lambda_s(L^p(\mathbb{R}^n))=\{\Lambda_s f:f\in L^p(\mathbb{R}^n)\}=\{g_s*f:f\in L^p(\mathbb{R}^n)\}. \tag*{(10)} Λ s ( L p ( R n )) = { Λ s f : f ∈ L p ( R n )} = { g s ∗ f : f ∈ L p ( R n )} . (10)
Indeed, it is easily verified that both definitions agree, using the semigroup property
Λ s 0 Λ s 1 = Λ s 0 + s 1 , s 0 , s 1 ∈ R . (11) \Lambda_{s_0}\Lambda_{s_1}=\Lambda_{s_0+s_1}, \qquad s_0,s_1\in\mathbb{R}. \tag*{(11)} Λ s 0 Λ s 1 = Λ s 0 + s 1 , s 0 , s 1 ∈ R . (11)
We will denote by H s , p ( R n ) H^{s,p}(\mathbb{R}^n) H s , p ( R n ) the Bessel potential space of order s ∈ R s\in\mathbb{R} s ∈ R and exponent p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] , according to (1.9) and (1.10).
The Bessel potential spaces owe their name to the Bessel potentials associated to the operators ( 1 − Δ ) − s / 2 (1-\Delta)^{-s/2} ( 1 − Δ ) − s /2 , whose Fourier symbol is ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) − s / 2 (1+4\pi^2|\cdot|^2)^{-s/2} ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) − s /2 . In particular, H s , p ( R n ) H^{s,p}(\mathbb{R}^n) H s , p ( R n ) is commonly also expressed in the form
H s , p ( R n ) = { u ∈ S ′ ( R n ) : ( 1 − Δ ) s / 2 u ∈ L p ( R n ) } . (12) H^{s,p}(\mathbb{R}^n)=\{u\in\mathcal{S}'(\mathbb{R}^n):(1-\Delta)^{s/2}u\in L^p(\mathbb{R}^n)\}. \tag*{(12)} H s , p ( R n ) = { u ∈ S ′ ( R n ) : ( 1 − Δ ) s /2 u ∈ L p ( R n )} . (12)
The Bessel potentials are themselves closely related to the Riesz potentials based on the operators ( − Δ ) − s / 2 (-\Delta)^{-s/2} ( − Δ ) − s /2 . In particular, it can be shown that for s > 0 s>0 s > 0 the formal quotients
( − Δ ) s / 2 ( 1 − Δ ) s / 2 and ( 1 − Δ ) s / 2 1 + ( − Δ ) s / 2 , \frac{(-\Delta)^{s/2}}{(1-\Delta)^{s/2}} \qquad\text{and} \qquad\frac{(1-\Delta)^{s/2}}{1+(-\Delta)^{s/2}}, ( 1 − Δ ) s /2 ( − Δ ) s /2 and 1 + ( − Δ ) s /2 ( 1 − Δ ) s /2 ,
or in their Fourier representation
( 2 π ∣ ⋅ ∣ ) s ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) s / 2 and ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) s / 2 1 + ( 2 π ∣ ⋅ ∣ ) s , \frac{(2\pi|\cdot|)^s}{(1+4\pi^2|\cdot|^2)^{s/2}} \qquad\text{and} \qquad\frac{(1+4\pi^2|\cdot|^2)^{s/2}}{1+(2\pi|\cdot|)^s}, ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) s /2 ( 2 π ∣ ⋅ ∣ ) s and 1 + ( 2 π ∣ ⋅ ∣ ) s ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) s /2 ,
admit Fourier inverses, i.e., there exist finite signed measures μ 1 s , μ 2 s \mu_1^s,\mu_2^s μ 1 s , μ 2 s on R n \mathbb{R}^n R n such that
μ 1 s ∗ f = ( − Δ ) s / 2 ( 1 − Δ ) s / 2 f and μ 2 s ∗ f = ( 1 − Δ ) s / 2 1 + ( − Δ ) s / 2 f \mu_1^s*f=\frac{(-\Delta)^{s/2}}{(1-\Delta)^{s/2}}f \qquad\text{and} \qquad\mu_2^s*f=\frac{(1-\Delta)^{s/2}}{1+(-\Delta)^{s/2}}f μ 1 s ∗ f = ( 1 − Δ ) s /2 ( − Δ ) s /2 f and μ 2 s ∗ f = 1 + ( − Δ ) s /2 ( 1 − Δ ) s /2 f
for all f ∈ L p ( R n ) f\in L^p(\mathbb{R}^n) f ∈ L p ( R n ) , and
μ 1 s ^ = ( 2 π ∣ ⋅ ∣ ) s ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) s / 2 and μ 2 s ^ = ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) s / 2 1 + ( 2 π ∣ ⋅ ∣ ) s . \widehat{\mu_1^s}=\frac{(2\pi|\cdot|)^s}{(1+4\pi^2|\cdot|^2)^{s/2}} \qquad\text{and} \qquad\widehat{\mu_2^s}=\frac{(1+4\pi^2|\cdot|^2)^{s/2}}{1+(2\pi|\cdot|)^s}. μ 1 s = ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) s /2 ( 2 π ∣ ⋅ ∣ ) s and μ 2 s = 1 + ( 2 π ∣ ⋅ ∣ ) s ( 1 + 4 π 2 ∣ ⋅ ∣ 2 ) s /2 .
Leveraging these relations, the Bessel potential spaces can be represented in terms of the Riesz potential as
H s , p ( R n ) = { u ∈ L p ( R n ) : ( − Δ ) s / 2 u ∈ L p ( R n ) } . (13) H^{s,p}(\mathbb{R}^n)=\{u\in L^p(\mathbb{R}^n):(-\Delta)^{s/2}u\in L^p(\mathbb{R}^n)\}. \tag*{(13)} H s , p ( R n ) = { u ∈ L p ( R n ) : ( − Δ ) s /2 u ∈ L p ( R n )} . (13)
This finishes our recapitulation of the different characterizations of the Bessel potential spaces. We transition now to the question why they gained such an importance over the last years in the context of nonlocal gradients. The answer is contained in the observation that there exists yet another way to describe Bessel potential spaces: as established in [33 ] , they coincide with the nonlocal Sobolev spaces (see § 2.2 ) based on the kernels
ρ s ( x ) : = 1 γ ( n , s ) ∣ x ∣ n − 1 + s for γ ( n , s ) : = π n / 2 Γ ( 1 − s 2 ) 2 s Γ ( n + 1 + s 2 ) , x ∈ R n ∖ { 0 } , (14) \rho_s(x) := \frac{1}{\gamma(n,s)|x|^{n-1+s}} \quad\text{for} \quad\gamma(n,s) := \frac{\pi^{n/2}\Gamma\left(\frac{1-s}{2}\right)}{2^s\Gamma\left(\frac{n+1+s}{2}\right)}, \quad x \in\mathbb{R}^n \setminus\{0\}, \tag*{(14)} ρ s ( x ) := γ ( n , s ) ∣ x ∣ n − 1 + s 1 for γ ( n , s ) := 2 s Γ ( 2 n + 1 + s ) π n /2 Γ ( 2 1 − s ) , x ∈ R n ∖ { 0 } , (14)
this is,
H s , p ( R n ) = H ρ s , p ( R n ) . H^{s,p}(\mathbb{R}^n) = H^{\rho_s,p}(\mathbb{R}^n). H s , p ( R n ) = H ρ s , p ( R n ) .
The path taken in this paper is now opposite to the historical development of the theory of Bessel potential spaces: starting from generalized nonlocal Sobolev spaces, we strive to prove characterizations similar to (1.9), (1.10), (1.12) and (1.13) for very general kernels. In our general theory, the quantity 2 π ∣ ⋅ ∣ 2\pi|\cdot| 2 π ∣ ⋅ ∣ is replaced by the expression ∣ λ ρ ∣ = 2 π ∣ ⋅ ∣ Q ^ ρ |\lambda_\rho| = 2\pi|\cdot|\widehat{\mathcal{Q}}_\rho ∣ λ ρ ∣ = 2 π ∣ ⋅ ∣ Q ρ (see § 2.3 ). We demonstrate that the analogues of (1.9) and (1.12), i.e.,
H ρ , p ( R n ) = { u ∈ S ′ ( R n ) : F − 1 ( ( 1 + ∣ λ ρ ∣ 2 ) 1 / 2 u ^ ) ∈ L p ( R n ) } H^{\rho,p}(\mathbb{R}^n) = \left\{u \in\mathcal{S}'(\mathbb{R}^n) : \mathcal{F}^{-1}\left((1+|\lambda_\rho|^2)^{1/2}\widehat{u}\right) \in L^p(\mathbb{R}^n)\right\} H ρ , p ( R n ) = { u ∈ S ′ ( R n ) : F − 1 ( ( 1 + ∣ λ ρ ∣ 2 ) 1/2 u ) ∈ L p ( R n ) }
and
H ρ , p ( R n ) = { u ∈ S ′ ( R n ) : ( 1 − Δ ρ ) 1 / 2 u ∈ L p ( R n ) } , H^{\rho,p}(\mathbb{R}^n) = \left\{u \in\mathcal{S}'(\mathbb{R}^n) : (1-\Delta^\rho)^{1/2}u \in L^p(\mathbb{R}^n)\right\}, H ρ , p ( R n ) = { u ∈ S ′ ( R n ) : ( 1 − Δ ρ ) 1/2 u ∈ L p ( R n ) } ,
where ( 1 − Δ ρ ) 1 / 2 (1-\Delta^\rho)^{1/2} ( 1 − Δ ρ ) 1/2 is extended by duality to the space of tempered distributions, merely rest on the assumption that the generalized fractions
∣ λ ρ ∣ ( 1 + ∣ λ ρ ∣ 2 ) 1 / 2 and ( 1 + ∣ λ ρ ∣ 2 ) 1 / 2 1 + ∣ λ ρ ∣ (15) \frac{|\lambda_\rho|}{(1+|\lambda_\rho|^2)^{1/2}} \quad\text{and} \quad\frac{(1+|\lambda_\rho|^2)^{1/2}}{1+|\lambda_\rho|} \tag*{(15)} ( 1 + ∣ λ ρ ∣ 2 ) 1/2 ∣ λ ρ ∣ and 1 + ∣ λ ρ ∣ ( 1 + ∣ λ ρ ∣ 2 ) 1/2 (15)
constitute Fourier multipliers; in particular, it is not needed that they admit Fourier inverses at this point. As it turns out, under stronger assumptions (the additional hypotheses being the lower bound of ρ \rho ρ by a Riesz-type fractional gradient in the sense of (H3) in § 2.5 ) they actually do (see Lemma 5.4), and the same ideas can be used to show the potential-type characterization of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) . More concretely, we construct a Fourier inverse in L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) of the expression ( 1 + ∣ λ ρ ∣ 2 ) − 1 / 2 (1+|\lambda_\rho|^2)^{-1/2} ( 1 + ∣ λ ρ ∣ 2 ) − 1/2 , given by
g ρ ( x ) : = 1 ( 4 π 2 ) 1 / 2 Γ ( 1 2 ) ∫ 0 ∞ f ρ δ ( x ) e − δ / ( 4 π 2 ) δ ( 1 − n ) / 2 d δ δ , x ∈ R n , g_\rho(x) := \frac{1}{(4\pi^2)^{1/2}\Gamma\left(\frac{1}{2}\right)} \int_0^\infty f^\delta_\rho(x)e^{-\delta/(4\pi^2)}\delta^{(1-n)/2}\frac{d\delta}{\delta}, \quad x \in\mathbb{R}^n, g ρ ( x ) := ( 4 π 2 ) 1/2 Γ ( 2 1 ) 1 ∫ 0 ∞ f ρ δ ( x ) e − δ / ( 4 π 2 ) δ ( 1 − n ) /2 δ d δ , x ∈ R n ,
for
f ρ δ ( x ) : = δ n / 2 ∫ R n e − ∣ λ ρ ( ξ ) ∣ 2 δ / ( 4 π 2 ) e 2 π i x ⋅ ξ d ξ , x ∈ R n . f^\delta_\rho(x) := \delta^{n/2}\int_{\mathbb{R}^n}e^{-|\lambda_\rho(\xi)|^2\delta/(4\pi^2)}e^{2\pi i x\cdot\xi}\,d\xi, \quad x \in\mathbb{R}^n. f ρ δ ( x ) := δ n /2 ∫ R n e − ∣ λ ρ ( ξ ) ∣ 2 δ / ( 4 π 2 ) e 2 π i x ⋅ ξ d ξ , x ∈ R n .
This is enough to deduce that also Calderon’s definition in (1.10) of the Bessel spaces finds an analogue in the general sense of nonlocal Sobolev spaces. In particular, it holds
H ρ , p ( R n ) = { g ρ ∗ f : f ∈ L p ( R n ) } H^{\rho,p}(\mathbb{R}^n) = \{g_\rho* f : f \in L^p(\mathbb{R}^n)\} H ρ , p ( R n ) = { g ρ ∗ f : f ∈ L p ( R n )}
and the spaces H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) can indeed be viewed as generalized potential spaces. However, for the moment, we were not able to prove that g ρ g_\rho g ρ is positive and ∥ g ρ ∥ L 1 ( R n ) = 1 \|g_\rho\|_{L^1(\mathbb{R}^n)} = 1 ∥ g ρ ∥ L 1 ( R n ) = 1 , which would mirror the situation of the Bessel potential spaces. Consequently, it is not known whether ( 1 + ∣ λ ρ ∣ 2 ) − 1 / 2 (1+|\lambda_\rho|^2)^{-1/2} ( 1 + ∣ λ ρ ∣ 2 ) − 1/2 is again the characteristic function of a probability measure.
Starting from the same stronger assumptions on ρ \rho ρ mentioned above, we further show that Δ ρ \Delta^\rho Δ ρ is well defined as an unbounded operator in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , and that F ∣ λ ρ ∣ F_{|\lambda_\rho|} F ∣ λ ρ ∣ , the closure in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) of the unbounded operator
F ∣ λ ρ ∣ u : = F − 1 ( ∣ λ ρ ∣ u ^ ) , u ∈ S ( R n ) , F_{|\lambda_\rho|}u := \mathcal{F}^{-1}(|\lambda_\rho|\widehat{u}), \quad u \in\mathcal{S}(\mathbb{R}^n), F ∣ λ ρ ∣ u := F − 1 ( ∣ λ ρ ∣ u ) , u ∈ S ( R n ) ,
@/body is a square root of − Δ ρ -\Delta^\rho − Δ ρ . In addition, in the Hilbert space case p = 2 p=2 p = 2 , the operator ( − Δ ρ ) 1 / 2 (-\Delta^\rho)^{1/2} ( − Δ ρ ) 1/2 is uniquely defined as the square root of the positive semidefinite and self-adjoint operator − Δ ρ -\Delta^\rho − Δ ρ . As a consequence of these results, we find that a nonlocal version of (13) holds true, this is
H ρ , p ( R n ) = { u ∈ L p ( R n ) : ( − Δ ρ ) 1 / 2 u ∈ L p ( R n ) } = D ( F ‾ ∣ λ ρ ∣ ) , H^{\rho,p}(\mathbb{R}^n)=\{u\in L^p(\mathbb{R}^n):(-\Delta^\rho)^{1/2}u\in L^p(\mathbb{R}^n)\}=D(\overline{F}_{|\lambda_\rho|}), H ρ , p ( R n ) = { u ∈ L p ( R n ) : ( − Δ ρ ) 1/2 u ∈ L p ( R n )} = D ( F ∣ λ ρ ∣ ) ,
with D ( F ‾ ∣ λ ρ ∣ ) D(\overline{F}_{|\lambda_\rho|}) D ( F ∣ λ ρ ∣ ) the domain in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) of F ‾ ∣ λ ρ ∣ \overline{F}_{|\lambda_\rho|} F ∣ λ ρ ∣ .
In summary, we derive analogous expressions to (9) , (10) , (12) and (13) for general nonlocal Sobolev spaces. As a consequence, in light of the above remarks, the operator ( 1 − Δ ρ ) − 1 / 2 (1-\Delta^\rho)^{-1/2} ( 1 − Δ ρ ) − 1/2 might be interpreted as a generalized Bessel potential operator. Nevertheless, our findings deviate in their generality in some ways from the specific case of the Bessel potential kernels, and we summarize in the following the main differences:
(i) the operators on L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) defined through f ↦ g ρ ∗ f f\mapsto g_\rho* f f ↦ g ρ ∗ f do not form a semigroup in a similar sense to the identity (following from (11) )
g ρ s 0 ∗ ( g ρ s 1 ∗ f ) = g ρ s 0 + s 1 ∗ f , f ∈ L p ( R n ) g_{\rho s_0}*(g_{\rho s_1}*f)=g_{\rho s_0+s_1}*f,\qquad f\in L^p(\mathbb{R}^n) g ρ s 0 ∗ ( g ρ s 1 ∗ f ) = g ρ s 0 + s 1 ∗ f , f ∈ L p ( R n )
holding for the Bessel potential spaces;
(ii) for the moment, the conditions on ρ \rho ρ to prove the characterization of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) similar to (9) is milder than those required to show that the generalized Bessel potential ( 1 − Δ ρ ) − 1 / 2 (1-\Delta^\rho)^{-1/2} ( 1 − Δ ρ ) − 1/2 (and also the fractions in (15) ) admits a Fourier inverse g ρ ∈ L 1 ( R n ) g_\rho\in L^1(\mathbb{R}^n) g ρ ∈ L 1 ( R n ) , and that Δ ρ \Delta^\rho Δ ρ is a closable operator over L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) with well defined square root;
(iii) we were not able to prove the positivity of g ρ g_\rho g ρ , which would immediately imply that g ρ g_\rho g ρ defines a probability distribution with the corresponding characteristic function given by ( 1 + ∣ λ ρ ∣ 2 ) − 1 / 2 (1+|\lambda_\rho|^2)^{-1/2} ( 1 + ∣ λ ρ ∣ 2 ) − 1/2 ;
(iv) we suppose that the kernel ρ \rho ρ has compact support (which is not the case for the Riesz fractional gradient), but this can be assumed without loss of generality, as explained in § 2.5 .
Outline
After fixing the notation used throughout the paper in Section 2.1, we recall the definition and main properties of nonlocal Sobolev spaces in Section 2.2. Section 2.3 then collects the Fourier-analytic aspects of nonlocal gradients needed later on. In Section 2.4, we set up the framework of unbounded operators defined via multiplication in Fourier space, before stating in Section 2.5 the precise hypotheses on the kernel functions that will be assumed in different parts of the paper. The preliminaries conclude with Section 2.6, where we derive key estimates and establish the central Fourier multiplier results on which the subsequent analysis relies.
The presentation of the different characterizations of the nonlocal Sobolev spaces is split into three main parts. First, Section 3 covers what we call the Fourier characterization, where the space is given as the subspace of tempered distributions in terms of a suitable multiplier. Such a description holds for functions defined on the whole space, as well as on bounded domains, which we show in Section 3.1 and Section 3.2, respectively. Afterwards, we also derive the Fourier characterization of the dual spaces in Section 3.3, alongside their representation through derivatives of L p L^p L p -functions.
Second, in Section 4 we establish the characterization of the nonlocal Sobolev spaces through the ρ \rho ρ -Laplacian as subspaces of the L p L^p L p -functions. More precisely, after thoroughly introducing this object in Section 4.1 and spending some time on deriving first essential properties of it, we focus in Section 4.2 on the Hilbert space case p = 2 p=2 p = 2 and then in Section 4.3 on the general case p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) . Later on, in Section 4.4, we also directly recast the Fourier characterization from Section 3 in terms of the operator ( 1 − Δ ρ ) 1 / 2 (1-\Delta^\rho)^{1/2} ( 1 − Δ ρ ) 1/2 .
Third, we conclude the cascade of equivalent characterizations in Section 5, and justify the name generalized Bessel potential spaces, by constructively proving the existence of a suitable potential convolution kernel.
Preliminaries
Notation and terminology
We denote by N \mathbb{N} N the natural numbers and write N 0 \mathbb{N}_0 N 0 if we include zero; R n \mathbb{R}^n R n is the Euclidean space with dimension n ∈ N n \in\mathbb{N} n ∈ N . We write B r ( x ) B_r(x) B r ( x ) for the ball centered at x ∈ R n x \in\mathbb{R}^n x ∈ R n and with radius r > 0 r > 0 r > 0 . We set B r : = B r ( 0 ) B_r := B_r(0) B r := B r ( 0 ) and write S n − 1 : = ∂ B 1 S^{n-1} := \partial B_1 S n − 1 := ∂ B 1 and S + n − 1 : = { z ∈ S n − 1 : z 1 > 0 } S_+^{n-1} := \{z \in S^{n-1}: z_1 > 0\} S + n − 1 := { z ∈ S n − 1 : z 1 > 0 } . The surface area integral over ∂ B r \partial B_r ∂ B r will be denoted by d S \mathrm{d}S d S . By Γ : ( 0 , ∞ ) → R \Gamma: (0,\infty) \to\mathbb{R} Γ : ( 0 , ∞ ) → R , Γ ( x ) : = ∫ 0 ∞ t x − 1 e − t d t \Gamma(x) := \int_0^\infty t^{x-1}e^{-t}\,\mathrm{d}t Γ ( x ) := ∫ 0 ∞ t x − 1 e − t d t we refer to the respective restriction of the common Γ \Gamma Γ -function.
If not mentioned otherwise, we assume throughout the paper that p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) . Moreover, p ′ ∈ ( 1 , ∞ ) p' \in(1,\infty) p ′ ∈ ( 1 , ∞ ) denotes the dual exponent of p p p defined through the relation 1 / p + 1 / p ′ = 1 1/p + 1/p' = 1 1/ p + 1/ p ′ = 1 . With L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) we refer to the usual Lebesgue spaces. We denote the space of compactly supported and smooth functions by C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) , and its topological dual by D ′ ( R n ) \mathcal{D}'(\mathbb{R}^n) D ′ ( R n ) . The space of Schwartz functions and its topological dual, the tempered distributions, are abbreviated as S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) and S ′ ( R n ) \mathcal{S}'(\mathbb{R}^n) S ′ ( R n ) , respectively. We denote by ⟨ ⋅ , ⋅ ⟩ \langle\cdot,\cdot\rangle ⟨ ⋅ , ⋅ ⟩ the duality pairing between L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) and L p ′ ( R n ) L^{p'}(\mathbb{R}^n) L p ′ ( R n ) . Moreover, we write ⟨ ⋅ , ⋅ ⟩ D ′ × D \langle\cdot,\cdot\rangle_{\mathcal{D}'\times\mathcal{D}} ⟨ ⋅ , ⋅ ⟩ D ′ × D , where, for consistency, we use the notation D ( R n ) = C c ∞ ( R n ) \mathcal{D}(\mathbb{R}^n)=C^\infty_c(\mathbb{R}^n) D ( R n ) = C c ∞ ( R n ) , and ⟨ ⋅ , ⋅ ⟩ S ′ × S \langle\cdot,\cdot\rangle_{\mathcal{S}'\times\mathcal{S}} ⟨ ⋅ , ⋅ ⟩ S ′ × S for the respective duality pairings. We say that an element u u u of D ′ ( R n ) \mathcal{D}'(\mathbb{R}^n) D ′ ( R n ) or S ′ ( R n ) \mathcal{S}'(\mathbb{R}^n) S ′ ( R n ) has a representative v v v in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) if ⟨ u , φ ⟩ D ′ × D = ⟨ v , φ ⟩ \langle u,\varphi\rangle_{\mathcal{D}'\times\mathcal{D}}=\langle v,\varphi\rangle ⟨ u , φ ⟩ D ′ × D = ⟨ v , φ ⟩ for every φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) or ⟨ u , φ ⟩ S ′ × S = ⟨ v , φ ⟩ \langle u,\varphi\rangle_{\mathcal{S}'\times\mathcal{S}}=\langle v,\varphi\rangle ⟨ u , φ ⟩ S ′ × S = ⟨ v , φ ⟩ for every φ ∈ S ( R n ) \varphi\in\mathcal{S}(\mathbb{R}^n) φ ∈ S ( R n ) , respectively.
To avoid confusion with the duality pairing, we use the notation ⟦ ⋅ ⟧ : = ( 1 + ∣ ⋅ ∣ 2 ) 1 / 2 \llbracket\cdot\rrbracket:= (1+|\cdot|^2)^{1/2} [ [ ⋅ ] ] := ( 1 + ∣ ⋅ ∣ 2 ) 1/2 instead of the maybe more common expression ⟨ ⋅ ⟩ \langle\cdot\rangle ⟨ ⋅ ⟩ for this quantity.
With α ∈ N 0 n \alpha\in\mathbb{N}_0^n α ∈ N 0 n we always refer to multi-indices, with the convention that ∣ α ∣ = α 1 + α 2 + ⋯ + α n |\alpha|=\alpha_1+\alpha_2+\cdots+\alpha_n ∣ α ∣ = α 1 + α 2 + ⋯ + α n . For another multi-index β ∈ N 0 n \beta\in\mathbb{N}_0^n β ∈ N 0 n , the inequality β ≤ α \beta\leq\alpha β ≤ α is understood componentwise, i.e., β i ≤ α i \beta_i\leq\alpha_i β i ≤ α i for all i ∈ { 1 , … , n } i\in\{1,\ldots,n\} i ∈ { 1 , … , n } ; in particular, ∣ α − β ∣ = ∣ α ∣ − ∣ β ∣ |\alpha-\beta|=|\alpha|-|\beta| ∣ α − β ∣ = ∣ α ∣ − ∣ β ∣ for β ≤ α \beta\leq\alpha β ≤ α . Finally, addition of multi-indices is understood componentwise. If we say that all derivatives of a function u ∈ C ∞ ( R n ) u\in C^\infty(\mathbb{R}^n) u ∈ C ∞ ( R n ) have (at most) polynomial growth, we mean, as this is usually understood (see, e.g., [32 ] ), that for each α ∈ N 0 n \alpha\in\mathbb{N}_0^n α ∈ N 0 n there exist C α > 0 C_\alpha>0 C α > 0 and N α ≥ 0 N_\alpha\geq0 N α ≥ 0 such that
∣ ∂ α u ( x ) ∣ ≤ C α ( 1 + ∣ x ∣ ) N α , x ∈ R n . |\partial^\alpha u(x)|\leq C_\alpha(1+|x|)^{N_\alpha}, \qquad x\in\mathbb{R}^n. ∣ ∂ α u ( x ) ∣ ≤ C α ( 1 + ∣ x ∣ ) N α , x ∈ R n .
Accordingly, we also say that a function has polynomial growth as soon as the estimate holds for α = 0 \alpha=0 α = 0 . Both notions naturally extend to vector-valued functions.
With D ( A ) D(A) D ( A ) we denote the domain of definition of an unbounded linear operator A : D ( A ) ⊂ X → Y A:D(A)\subset X\to Y A : D ( A ) ⊂ X → Y between two Banach spaces X X X and Y Y Y . In the case p = 2 p=2 p = 2 and for some densely defined linear operator T : D ( T ) ⊂ L 2 ( R n ) → L 2 ( R n ) T:D(T)\subset L^2(\mathbb{R}^n)\to L^2(\mathbb{R}^n) T : D ( T ) ⊂ L 2 ( R n ) → L 2 ( R n ) , we write T ∗ T^* T ∗ for the adjoint operator T ∗ : D ( T ∗ ) ⊂ L 2 ( R n ) → L 2 ( R n ) T^*:D(T^*)\subset L^2(\mathbb{R}^n)\to L^2(\mathbb{R}^n) T ∗ : D ( T ∗ ) ⊂ L 2 ( R n ) → L 2 ( R n ) . Standard references for unbounded operators in Banach spaces are [21 Chapter 3, Sections 2.1 and 5.1] and [11 Section 2.1] . We will also use the terminology of closable operators (see, e.g., [21 Chapter 3, Section 5.3] ). For our purposes, it will be enough to know that an unbounded linear operator T T T in the Banach space X X X is closable if and only if the following condition holds: if a sequence ( u j ) j (u_j)_j ( u j ) j in D ( T ) D(T) D ( T ) tends to 0 0 0 in X X X and there exists v ∈ X v\in X v ∈ X such that T u j → v Tu_j\to v T u j → v in X X X , then v = 0 v=0 v = 0 . In that case, the closure T ‾ \overline{T} T of T T T is the unbounded operator in X X X with domain the set of u ∈ X u\in X u ∈ X for which there exist a sequence ( u j ) j ⊂ D ( T ) (u_j)_j\subset D(T) ( u j ) j ⊂ D ( T ) and v ∈ X v\in X v ∈ X such that u j → u u_j\to u u j → u and T u j → v Tu_j\to v T u j → v in X X X . The element v v v is uniquely determined and one sets T ‾ u = v \overline{T}u=v T u = v . Recall that if T 1 , T 2 T_1,T_2 T 1 , T 2 are unbounded operators, the domain of its composition T 2 T 1 T_2T_1 T 2 T 1 is the set
D ( T 2 T 1 ) = { u ∈ D ( T 1 ) : T 1 u ∈ D ( T 2 ) } , (16) D(T_2T_1)=\{u\in D(T_1):T_1u\in D(T_2)\}, \tag*{(16)} D ( T 2 T 1 ) = { u ∈ D ( T 1 ) : T 1 u ∈ D ( T 2 )} , (16)
see, e.g., [21 Chapter 3, Section 5.1] .
Recap on nonlocal Sobolev spaces
Throughout this paper ρ : R n → [ 0 , ∞ ] \rho:\mathbb{R}^n\to[0,\infty] ρ : R n → [ 0 , ∞ ] will always denote an even (radially symmetric) kernel, whose radial profile is given by ρ ~ : R → [ 0 , ∞ ] \widetilde{\rho}:\mathbb{R}\to[0,\infty] ρ : R → [ 0 , ∞ ] . As minimum assumptions on the kernel, we adopt from [8 ] the hypotheses
{ ρ : R n → [ 0 , ∞ ] is radial with ρ ( x ) ∈ R for all x ∈ R n ∖ { 0 } , ρ ∈ L l o c 1 ( R n ) with ∫ R n min { 1 , ∣ x ∣ − 1 } ρ ( x ) d x < ∞ , inf B δ 0 ρ > 0 for some δ 0 > 0. (17) \begin{cases}
\rho:\mathbb{R}^n\to[0,\infty] \text{ is radial with }\rho(x)\in\mathbb{R}\text{ for all }x\in\mathbb{R}^n\setminus\{0\},\\
\rho\in L^1_{\mathrm{loc}}(\mathbb{R}^n)\text{ with }\int_{\mathbb{R}^n}\min\{1,|x|^{-1}\}\rho(x)\,\mathrm{d}x<\infty,\\
\inf_{B_{\delta_0}}\rho>0\text{ for some }\delta_0>0. \tag*{(17)}
\end{cases} ⎩ ⎨ ⎧ ρ : R n → [ 0 , ∞ ] is radial with ρ ( x ) ∈ R for all x ∈ R n ∖ { 0 } , ρ ∈ L loc 1 ( R n ) with ∫ R n min { 1 , ∣ x ∣ − 1 } ρ ( x ) d x < ∞ , inf B δ 0 ρ > 0 for some δ 0 > 0. (17)
In particular, (H0) ensures that for any u ∈ C c ∞ ( R n ) u \in C^\infty_c(\mathbb{R}^n) u ∈ C c ∞ ( R n ) and U ∈ C c ∞ ( R n ; R n ) U \in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n) U ∈ C c ∞ ( R n ; R n ) the nonlocal gradient D 0 ρ u D^\rho_0u D 0 ρ u defined as
D 0 ρ u ( x ) : = ∫ R n u ( x ) − u ( y ) ∣ x − y ∣ x − y ∣ x − y ∣ ρ ( x − y ) d y , x ∈ R n (18) D^\rho_0u(x) := \int_{\mathbb{R}^n} \frac{u(x)-u(y)}{|x-y|}\frac{x-y}{|x-y|}\rho(x-y)\,\mathrm{d}y,\qquad x \in\mathbb{R}^n \tag*{(18)} D 0 ρ u ( x ) := ∫ R n ∣ x − y ∣ u ( x ) − u ( y ) ∣ x − y ∣ x − y ρ ( x − y ) d y , x ∈ R n (18)
lies in L 1 ( R n ; R n ) ∩ L ∞ ( R n ; R n ) L^1(\mathbb{R}^n;\mathbb{R}^n) \cap L^\infty(\mathbb{R}^n;\mathbb{R}^n) L 1 ( R n ; R n ) ∩ L ∞ ( R n ; R n ) , and the nonlocal divergence
div 0 ρ U ( x ) : = ∫ R n U ( x ) − U ( y ) ∣ x − y ∣ ⋅ x − y ∣ x − y ∣ ρ ( x − y ) d y , x ∈ R n (19) \operatorname{div}^\rho_0 U(x) := \int_{\mathbb{R}^n} \frac{U(x)-U(y)}{|x-y|}\cdot\frac{x-y}{|x-y|}\rho(x-y)\,\mathrm{d}y,\qquad x \in\mathbb{R}^n \tag*{(19)} div 0 ρ U ( x ) := ∫ R n ∣ x − y ∣ U ( x ) − U ( y ) ⋅ ∣ x − y ∣ x − y ρ ( x − y ) d y , x ∈ R n (19)
lies in L 1 ( R n ) ∩ L ∞ ( R n ) L^1(\mathbb{R}^n) \cap L^\infty(\mathbb{R}^n) L 1 ( R n ) ∩ L ∞ ( R n ) (see [8 Lemma 2.3] ). Then, for u ∈ L 1 ( R n ) + L ∞ ( R n ) u \in L^1(\mathbb{R}^n) + L^\infty(\mathbb{R}^n) u ∈ L 1 ( R n ) + L ∞ ( R n ) the distributional, or weak, nonlocal gradient is the distribution defined through the relation
⟨ D ρ u , Φ ⟩ D ′ × D = − ∫ R n u ( x ) div 0 ρ Φ ( x ) d x , Φ ∈ C c ∞ ( R n ; R n ) . (20) \langle D^\rho u,\Phi\rangle_{\mathcal{D}'\times\mathcal{D}} = -\int_{\mathbb{R}^n} u(x)\operatorname{div}^\rho_0\Phi(x)\,\mathrm{d}x,\qquad\Phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n). \tag*{(20)} ⟨ D ρ u , Φ ⟩ D ′ × D = − ∫ R n u ( x ) div 0 ρ Φ ( x ) d x , Φ ∈ C c ∞ ( R n ; R n ) . (20)
In particular, thanks to [8 Proposition 3.2] , for u ∈ C c ∞ ( R n ) u \in C^\infty_c(\mathbb{R}^n) u ∈ C c ∞ ( R n ) its pointwise and distributional nonlocal gradient agree, that is,
D 0 ρ u = D ρ u a.e. in Ω , u ∈ C c ∞ ( R n ) . (21) D^\rho_0u = D^\rho u \quad\text{a.e. in } \Omega,\qquad u \in C^\infty_c(\mathbb{R}^n). \tag*{(21)} D 0 ρ u = D ρ u a.e. in Ω , u ∈ C c ∞ ( R n ) . (21)
Moreover, as the proof of [8 Lemma 2.3] shows, D 0 ρ u D^\rho_0u D 0 ρ u and div 0 ρ u \operatorname{div}^\rho_0 u div 0 ρ u are actually well defined for any bounded Lipschitz function u ∈ C 0 , 1 ( R n ) ∩ L ∞ ( R n ) u \in C^{0,1}(\mathbb{R}^n) \cap L^\infty(\mathbb{R}^n) u ∈ C 0 , 1 ( R n ) ∩ L ∞ ( R n ) . In particular, one can quickly check that equality (2.5) extends to such functions (see also the comments after [8 Definition 3.3] ). If ρ \rho ρ has compact support, then the same remains true without the L ∞ L^\infty L ∞ -requirement, i.e., D 0 ρ u D^\rho_0u D 0 ρ u and div 0 ρ u \operatorname{div}^\rho_0 u div 0 ρ u are well defined and identity (2.5) is valid for all u ∈ C 0 , 1 ( R n ) u \in C^{0,1}(\mathbb{R}^n) u ∈ C 0 , 1 ( R n ) .
We recall now from [8 Definition 3.4] the definitions of the ρ \rho ρ -nonlocal Sobolev spaces H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) and H 0 ρ , p ( Ω ) H^{\rho,p}_0(\Omega) H 0 ρ , p ( Ω ) for p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) and Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n open:
H ρ , p ( R n ) : = { u ∈ L p ( R n ) : D ρ u ∈ L p ( R n ; R n ) } , H^{\rho,p}(\mathbb{R}^n) := \{u \in L^p(\mathbb{R}^n): D^\rho u \in L^p(\mathbb{R}^n;\mathbb{R}^n)\}, H ρ , p ( R n ) := { u ∈ L p ( R n ) : D ρ u ∈ L p ( R n ; R n )} ,
H 0 ρ , p ( Ω ) : = { u ∈ H ρ , p ( R n ) : u = 0 a.e. in Ω } ; (22) H^{\rho,p}_0(\Omega) := \{u \in H^{\rho,p}(\mathbb{R}^n): u = 0\ \text{a.e. in }\Omega\}; \tag*{(22)} H 0 ρ , p ( Ω ) := { u ∈ H ρ , p ( R n ) : u = 0 a.e. in Ω } ; (22)
the latter complementary-value space is the nonlocal analogue of classical Sobolev spaces with vanishing trace. Equipped with the norm
∥ u ∥ H ρ , p ( R n ) : = ( ∥ u ∥ L p ( R n ) p + ∥ D ρ u ∥ L p ( R n ; R n ) p ) 1 / p , (23) \|u\|_{H^{\rho,p}(\mathbb{R}^n)} := \left(\|u\|_{L^p(\mathbb{R}^n)}^p + \|D^\rho u\|_{L^p(\mathbb{R}^n;\mathbb{R}^n)}^p\right)^{1/p}, \tag*{(23)} ∥ u ∥ H ρ , p ( R n ) := ( ∥ u ∥ L p ( R n ) p + ∥ D ρ u ∥ L p ( R n ; R n ) p ) 1/ p , (23)
they form separable reflexive Banach spaces, and Hilbert spaces for p = 2 p=2 p = 2 . Moreover, as shown in [8 Theorem 3.9] , the space C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) lies dense in H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) and also C c ∞ ( Ω ) C^\infty_c(\Omega) C c ∞ ( Ω ) in H 0 ρ , p ( Ω ) H^{\rho,p}_0(\Omega) H 0 ρ , p ( Ω ) for open and bounded Ω \Omega Ω with Lipschitz boundary.
An important insight related to the theory above is that for functions in C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) the nonlocal gradient can be written as the convolution of the classical gradient with the function Q ρ Q_\rho Q ρ defined as
Q ρ ( x ) : = ∫ ∣ x ∣ ∞ ρ ( t ) t d t , x ∈ R n ∖ { 0 } . Q_\rho(x) := \int_{|x|}^{\infty} \frac{\rho(t)}{t}\,\mathrm{d}t,\qquad x \in\mathbb{R}^n\setminus\{0\}. Q ρ ( x ) := ∫ ∣ x ∣ ∞ t ρ ( t ) d t , x ∈ R n ∖ { 0 } .
Precisely, for u ∈ C c ∞ ( R n ) u \in C^\infty_c(\mathbb{R}^n) u ∈ C c ∞ ( R n ) it holds (see [8 Proposition 2.6 (i)] )
D 0 ρ u = Q ρ ∗ ∇ u = ∇ ( Q ρ ∗ u ) . (24) D^\rho_0u = Q_\rho* \nabla u = \nabla(Q_\rho* u). \tag*{(24)} D 0 ρ u = Q ρ ∗ ∇ u = ∇ ( Q ρ ∗ u ) . (24)
If ρ ∈ L 1 ( R n ) \rho\in L^1(\mathbb{R}^n) ρ ∈ L 1 ( R n ) , then also Q ρ ∈ L 1 ( R n ) Q_\rho\in L^1(\mathbb{R}^n) Q ρ ∈ L 1 ( R n ) . For further properties of Q ρ Q_\rho Q ρ , we refer to [8 Lemma 2.5] .
Fourier transform
We write S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) for the space of Schwartz functions over R n \mathbb{R}^n R n and denote for f ∈ S ( R n ) f \in\mathcal{S}(\mathbb{R}^n) f ∈ S ( R n ) by F f \mathcal{F}f F f (if sensible, also expressed as f ^ \hat{f} f ^ for brevity) the Fourier transform defined as
F f ( ξ ) : = ∫ R n f ( x ) e − 2 π i x ⋅ ξ d x . \mathcal{F}f(\xi) := \int_{\mathbb{R}^n} f(x)e^{-2\pi i x\cdot\xi}\,\mathrm{d}x. F f ( ξ ) := ∫ R n f ( x ) e − 2 π i x ⋅ ξ d x .
The inverse Fourier transform will be denoted by F − 1 \mathcal{F}^{-1} F − 1 and is defined through the identity F − 1 f ( ξ ) = F f ( − ξ ) \mathcal{F}^{-1}f(\xi) = \mathcal{F}f(-\xi) F − 1 f ( ξ ) = F f ( − ξ ) for ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n . The Fourier transform and its inverse are extended to L 1 ( R n ) + L 2 ( R n ) L^1(\mathbb{R}^n)+L^2(\mathbb{R}^n) L 1 ( R n ) + L 2 ( R n ) and to the tempered distributions S ′ ( R n ) \mathcal{S}'(\mathbb{R}^n) S ′ ( R n ) as usual (see, e.g., [17 Chapter 2] ). Observe that if we denote by f ~ : = f ( − ⋅ ) \widetilde{f}:=f(-\cdot) f := f ( − ⋅ ) the reflection of a function f f f , then
F − 1 f = F F f ~ = F f ~ (25) \mathcal{F}^{-1}f=\mathcal{F}\widetilde{\mathcal{F}f}=\widetilde{\mathcal{F}f} \tag*{(25)} F − 1 f = F F f = F f (25)
for any function f f f for which F − 1 f \mathcal{F}^{-1}f F − 1 f is well defined.
A central role in the subsequent discussion will be played by the function
λ ρ ( ξ ) : = ∫ R d ρ ( x ) x ∣ x ∣ 2 ( e 2 π i ξ ⋅ x − 1 ) d x , ξ ∈ R n , \lambda_\rho(\xi):=\int_{\mathbb{R}^d}\frac{\rho(x)x}{\lvert x\rvert^2}\left(e^{2\pi i\xi\cdot x}-1\right)\,\mathrm{d}x,\qquad\xi\in\mathbb{R}^n, λ ρ ( ξ ) := ∫ R d ∣ x ∣ 2 ρ ( x ) x ( e 2 π i ξ ⋅ x − 1 ) d x , ξ ∈ R n ,
which also admits the representation (see [8 Lemma 2.4] )
λ ρ ( ξ ) = i ξ ∣ ξ ∣ ∫ R n ρ ( x ) x 1 ∣ x ∣ 2 sin ( 2 π ∣ ξ ∣ x 1 ) d x , ξ ∈ R n ∖ { 0 } . \lambda_\rho(\xi)=\frac{i\xi}{\lvert\xi\rvert}\int_{\mathbb{R}^n}\frac{\rho(x)x_1}{\lvert x\rvert^2}\sin(2\pi\lvert\xi\rvert x_1)\,\mathrm{d}x,\qquad\xi\in\mathbb{R}^n\setminus\{0\}. λ ρ ( ξ ) = ∣ ξ ∣ i ξ ∫ R n ∣ x ∣ 2 ρ ( x ) x 1 sin ( 2 π ∣ ξ ∣ x 1 ) d x , ξ ∈ R n ∖ { 0 } .
We point out that the x 1 x_1 x 1 in the integrand above can be replaced by any x i x_i x i for i = 1 , … , n i=1,\ldots,n i = 1 , … , n due to the radial symmetry of ρ \rho ρ . The importance of λ ρ \lambda_\rho λ ρ roots in the fact that it is the Fourier symbol associated to the nonlocal gradient and divergence, i.e., for u ∈ C c ∞ ( R n ) u\in C^\infty_c(\mathbb{R}^n) u ∈ C c ∞ ( R n ) and Φ ∈ C c ∞ ( R n ; R n ) \Phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n) Φ ∈ C c ∞ ( R n ; R n ) it holds (see [8 Lemma 2.4] )
F ( D 0 ρ u ) = λ ρ u ^ , F ( div 0 ρ Φ ) = λ ρ ⋅ Φ ^ . (26) \mathcal{F}(D^ \rho_0u)=\lambda_\rho\widehat{u},\qquad\mathcal{F}(\operatorname{div}^\rho_0\Phi)=\lambda_\rho\cdot\widehat{\Phi}. \tag*{(26)} F ( D 0 ρ u ) = λ ρ u , F ( div 0 ρ Φ ) = λ ρ ⋅ Φ . (26)
This extends readily to u ∈ S ( R n ) u\in\mathcal{S}(\mathbb{R}^n) u ∈ S ( R n ) and Φ ∈ S ( R n ; R n ) \Phi\in\mathcal{S}(\mathbb{R}^n;\mathbb{R}^n) Φ ∈ S ( R n ; R n ) , respectively; we note that the Schwartz space is included in H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) for all p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] (see [8 Proposition 3.5] ) and thus the Fourier transforms above are well defined. For the Fourier inverse of the nonlocal divergence of Φ ∈ S ( R n ; R n ) \Phi\in\mathcal{S}(\mathbb{R}^n;\mathbb{R}^n) Φ ∈ S ( R n ; R n ) , one has
F − 1 ( − div 0 ρ Φ ) = λ ρ ⋅ F − 1 Φ ; (27) \mathcal{F}^{-1}(-\operatorname{div}^\rho_0\Phi)=\lambda_\rho\cdot\mathcal{F}^{-1}\Phi; \tag*{(27)} F − 1 ( − div 0 ρ Φ ) = λ ρ ⋅ F − 1 Φ ; (27)
indeed, the fact that λ ρ \lambda_\rho λ ρ is odd, (2.9) and (2.10) imply
F − 1 ( − div 0 ρ Φ ) = F ( − div 0 ρ Φ ) = − λ ρ ⋅ F Φ = λ ρ ⋅ F − 1 Φ . \mathcal{F}^{-1}(-\operatorname{div}^\rho_0\Phi)=\mathcal{F}(-\operatorname{div}^\rho_0\Phi)=-\lambda_\rho\cdot\mathcal{F}\Phi=\lambda_\rho\cdot\mathcal{F}^{-1}\Phi. F − 1 ( − div 0 ρ Φ ) = F ( − div 0 ρ Φ ) = − λ ρ ⋅ F Φ = λ ρ ⋅ F − 1 Φ.
Furthermore, if ρ ∈ L 1 ( R n ) \rho\in L^1(\mathbb{R}^n) ρ ∈ L 1 ( R n ) , then Q ρ ∈ L 1 ( R n ) Q_\rho\in L^1(\mathbb{R}^n) Q ρ ∈ L 1 ( R n ) and its Fourier transform is well defined. In particular, combining (2.8) and (2.10), one finds that (see [8 Proposition 2.6 (ii)] )
λ ρ ( ξ ) = 2 π i ξ Q ρ ^ ( ξ ) , ξ ∈ R n \lambda_\rho(\xi)=2\pi i\xi\widehat{Q_\rho}(\xi),\qquad\xi\in\mathbb{R}^n λ ρ ( ξ ) = 2 π i ξ Q ρ ( ξ ) , ξ ∈ R n
and
Q ρ ^ ( ξ ) = 1 2 π ∣ ξ ∣ ∫ R n ρ ( x ) x 1 ∣ x ∣ 2 sin ( 2 π ∣ ξ ∣ x 1 ) d x , ξ ∈ R n ∖ { 0 } . (28) \widehat{Q_\rho}(\xi)=\frac{1}{2\pi\lvert\xi\rvert}\int_{\mathbb{R}^n}\frac{\rho(x)x_1}{\lvert x\rvert^2}\sin(2\pi\lvert\xi\rvert x_1)\,\mathrm{d}x,\qquad\xi\in\mathbb{R}^n\setminus\{0\}. \tag*{(28)} Q ρ ( ξ ) = 2 π ∣ ξ ∣ 1 ∫ R n ∣ x ∣ 2 ρ ( x ) x 1 sin ( 2 π ∣ ξ ∣ x 1 ) d x , ξ ∈ R n ∖ { 0 } . (28)
Starting from the identity above, the authors of [8 ] studied the properties of Q ρ ^ \widehat{Q_\rho} Q ρ and devised sufficient conditions ensuring, in particular, that Q ρ ^ \widehat{Q_\rho} Q ρ is positive and satisfies certain lower bounds. We will recall the corresponding results in Section 2.5. Later on in Lemma 2.9 it is derived that, as a consequence of these conditions, λ ρ \lambda_\rho λ ρ , understood as a multiplier in Fourier space, maps S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) into S ( R n ; R n ) \mathcal{S}(\mathbb{R}^n;\mathbb{R}^n) S ( R n ; R n ) .
It is also worth mentioning that the Fourier transform preserves radiality (see [17 Corollary 2.2.12] ); thus, since ρ \rho ρ is radial by hypothesis (H0), Q ρ ^ \widehat{Q_\rho} Q ρ is again radial, while λ ρ \lambda_\rho λ ρ is vector radial.
Fourier multiplier theory
Let us now turn to L p L^p L p -Fourier multipliers and, more generally, to unbounded linear operators on L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) defined via multiplication in Fourier space. For a tempered distribution w ∈ S ′ ( R n ) w\in\mathcal{S}'(\mathbb{R}^n) w ∈ S ′ ( R n ) , we write F − 1 w ∈ L p ( R n ) \mathcal{F}^{-1}w\in L^p(\mathbb{R}^n) F − 1 w ∈ L p ( R n ) if there exists a (necessarily unique) v ∈ L p ( R n ) v\in L^p(\mathbb{R}^n) v ∈ L p ( R n ) such that
⟨ v , φ ⟩ = ⟨ w , F − 1 φ ⟩ S ′ × S , φ ∈ S ( R n ) , (29) \langle v,\varphi\rangle=\langle w,\mathcal{F}^{-1}\varphi\rangle_{\mathcal{S}'\times\mathcal{S}},\qquad\varphi\in\mathcal{S}(\mathbb{R}^n), \tag*{(29)} ⟨ v , φ ⟩ = ⟨ w , F − 1 φ ⟩ S ′ × S , φ ∈ S ( R n ) , (29)
and we identify F − 1 w \mathcal{F}^{-1}w F − 1 w with v v v . Since this convention appears in two contexts below (see Sections 3 and 4), we briefly discuss it here.
If m ∈ C ∞ ( R n ) m \in C^{\infty}(\mathbb{R}^{n}) m ∈ C ∞ ( R n ) is such that all its derivatives have polynomial growth and u ∈ S ′ ( R n ) u \in\mathcal{S}'(\mathbb{R}^{n}) u ∈ S ′ ( R n ) , then m u ^ m\widehat{u} m u and F − 1 ( m u ^ ) \mathcal{F}^{-1}(m\widehat{u}) F − 1 ( m u ) are again tempered distributions, defined as usual by duality, i.e.,
⟨ F − 1 ( m u ^ ) , φ ⟩ S ′ × S = ⟨ u , F ( m F − 1 φ ) ⟩ S ′ × S , φ ∈ S ( R n ) , \left\langle\mathcal{F}^{-1}(m\widehat{u}),\varphi\right\rangle_{\mathcal{S}'\times\mathcal{S}}=\left\langle u,\mathcal{F}(m\mathcal{F}^{-1}\varphi)\right\rangle_{\mathcal{S}'\times\mathcal{S}},\qquad\varphi\in\mathcal{S}(\mathbb{R}^{n}), ⟨ F − 1 ( m u ) , φ ⟩ S ′ × S = ⟨ u , F ( m F − 1 φ ) ⟩ S ′ × S , φ ∈ S ( R n ) ,
where we note that since m F − 1 φ ∈ S ( R n ) m\mathcal{F}^{-1}\varphi\in\mathcal{S}(\mathbb{R}^{n}) m F − 1 φ ∈ S ( R n ) , its Fourier transform is guaranteed to be well defined. Choosing w = m u ^ w=m\widehat{u} w = m u , we understand F − 1 ( m u ^ ) ∈ L p ( R n ) \mathcal{F}^{-1}(m\widehat{u})\in L^{p}(\mathbb{R}^{n}) F − 1 ( m u ) ∈ L p ( R n ) in the sense of (2.13).
If, on the other hand, m : R n → R m:\mathbb{R}^{n}\to\mathbb{R} m : R n → R is merely measurable and of polynomial growth, but u ∈ S ( R n ) u\in\mathcal{S}(\mathbb{R}^{n}) u ∈ S ( R n ) , then m u ^ m\widehat{u} m u is integrable and m u ^ m\widehat{u} m u defines a tempered distribution via
⟨ m u ^ , φ ⟩ S ′ × S = ⟨ m u ^ , φ ⟩ = ∫ R n m u ^ φ d x , φ ∈ S ( R n ) . \left\langle m\widehat{u},\varphi\right\rangle_{\mathcal{S}'\times\mathcal{S}}=\langle m\widehat{u},\varphi\rangle=\int_{\mathbb{R}^{n}}m\widehat{u}\varphi\,\mathrm{d}x,\qquad\varphi\in\mathcal{S}(\mathbb{R}^{n}). ⟨ m u , φ ⟩ S ′ × S = ⟨ m u , φ ⟩ = ∫ R n m u φ d x , φ ∈ S ( R n ) .
By taking w = m u ^ w=m\widehat{u} w = m u in (2.13), we can therefore again give meaning to the expression F − 1 ( m u ^ ) \mathcal{F}^{-1}(m\widehat{u}) F − 1 ( m u ) . Moreover, since m u ^ ∈ L 1 ( R n ) ∩ L 2 ( R n ) m\widehat{u}\in L^{1}(\mathbb{R}^{n})\cap L^{2}(\mathbb{R}^{n}) m u ∈ L 1 ( R n ) ∩ L 2 ( R n ) , this distributional inverse Fourier transform coincides with the pointwise one and, in particular, F − 1 ( m u ^ ) ∈ L 2 ( R n ) ∩ L ∞ ( R n ) \mathcal{F}^{-1}(m\widehat{u})\in L^{2}(\mathbb{R}^{n})\cap L^{\infty}(\mathbb{R}^{n}) F − 1 ( m u ) ∈ L 2 ( R n ) ∩ L ∞ ( R n ) .
For later use, we state an auxiliary result on symmetry relations for multiplication in Fourier space by even or odd functions.
Lemma 2.1. Assume that m 1 , m 2 : R n → R m_{1},m_{2}:\mathbb{R}^{n}\to\mathbb{R} m 1 , m 2 : R n → R are measurable functions with polynomial growth such that m 1 m_{1} m 1 is even and m 2 m_{2} m 2 is odd. Then, for any u ∈ S ( R n ) u\in\mathcal{S}(\mathbb{R}^{n}) u ∈ S ( R n ) ,
(i) F − 1 ( m 1 u ^ ) = F ( m 1 F − 1 u ) \mathcal{F}^{-1}(m_{1}\widehat{u})=\mathcal{F}(m_{1}\mathcal{F}^{-1}u) F − 1 ( m 1 u ) = F ( m 1 F − 1 u ) and F − 1 ( F − 1 ( m 1 u ^ ) ) = m 1 F − 1 u \mathcal{F}^{-1}(\mathcal{F}^{-1}(m_{1}\widehat{u}))=m_{1}\mathcal{F}^{-1}u F − 1 ( F − 1 ( m 1 u )) = m 1 F − 1 u ;
(ii) F − 1 ( m 2 u ^ ) = − F ( m 2 F − 1 u ) \mathcal{F}^{-1}(m_{2}\widehat{u})=-\mathcal{F}(m_{2}\mathcal{F}^{-1}u) F − 1 ( m 2 u ) = − F ( m 2 F − 1 u ) and F − 1 ( F − 1 ( m 2 u ^ ) ) = − m 2 F − 1 u \mathcal{F}^{-1}(\mathcal{F}^{-1}(m_{2}\widehat{u}))=-m_{2}\mathcal{F}^{-1}u F − 1 ( F − 1 ( m 2 u )) = − m 2 F − 1 u .
Proof. Let m ∈ { m 1 , m 2 } m\in\{m_{1},m_{2}\} m ∈ { m 1 , m 2 } and denote by m ~ \widetilde{m} m its reflection. Note that, as m m m has polynomial growth, for u ∈ S ( R n ) u\in\mathcal{S}(\mathbb{R}^{n}) u ∈ S ( R n ) , the expression F − 1 ( m u ^ ) \mathcal{F}^{-1}(m\widehat{u}) F − 1 ( m u ) is always well defined, since m u ^ ∈ L 1 ( R n ) m\widehat{u}\in L^{1}(\mathbb{R}^{n}) m u ∈ L 1 ( R n ) . Using (2.9), we have for x ∈ R n x\in\mathbb{R}^{n} x ∈ R n ,
F − 1 ( m u ^ ) ( x ) = ∫ R n m ( ξ ) u ^ ( ξ ) e 2 π i x ⋅ ξ d ξ = ∫ R n m ( ξ ) u ^ ( ξ ) e − 2 π i x ⋅ ( − ξ ) d ξ = ∫ R n m ~ ( ξ ) F u ( ξ ) e − 2 π i x ⋅ ξ d ξ = F ( m ~ F − 1 u ) ( x ) . \begin{aligned}
\mathcal{F}^{-1}(m\widehat{u})(x)&=\int_{\mathbb{R}^{n}}m(\xi)\widehat{u}(\xi)e^{2\pi ix\cdot\xi}\,\mathrm{d}\xi\\
&=\int_{\mathbb{R}^{n}}m(\xi)\widehat{u}(\xi)e^{-2\pi ix\cdot(-\xi)}\,\mathrm{d}\xi\\
&=\int_{\mathbb{R}^{n}}\widetilde{m}(\xi)\mathcal{F}u(\xi)e^{-2\pi ix\cdot\xi}\,\mathrm{d}\xi\\
&=\mathcal{F}(\widetilde{m}\mathcal{F}^{-1}u)(x).
\end{aligned} F − 1 ( m u ) ( x ) = ∫ R n m ( ξ ) u ( ξ ) e 2 π i x ⋅ ξ d ξ = ∫ R n m ( ξ ) u ( ξ ) e − 2 π i x ⋅ ( − ξ ) d ξ = ∫ R n m ( ξ ) F u ( ξ ) e − 2 π i x ⋅ ξ d ξ = F ( m F − 1 u ) ( x ) .
With this, we obtain the first equalities of (i) and (ii). The second ones follow by taking the inverse Fourier transform of both sides of the first one. The lemma is proved. □
We next define the following unbounded multiplier operators, which will play a prominent role throughout Section 4.
Definition 2.2. Let p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) . Let m : R n → R m:\mathbb{R}^{n}\to\mathbb{R} m : R n → R be a measurable function with polynomial growth, and let u ∈ S ( R n ) u\in\mathcal{S}(\mathbb{R}^{n}) u ∈ S ( R n ) . The unbounded operator F m \mathcal{F}_{m} F m in L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) with domain
D ( F m ) : = { u ∈ S ( R n ) : F − 1 ( m u ^ ) ∈ L p ( R n ) } D(\mathcal{F}_{m}):=\{u\in\mathcal{S}(\mathbb{R}^{n}):\mathcal{F}^{-1}(m\widehat{u})\in L^{p}(\mathbb{R}^{n})\} D ( F m ) := { u ∈ S ( R n ) : F − 1 ( m u ) ∈ L p ( R n )} is defined as
F m u : = F − 1 ( m u ^ ) , u ∈ D ( F m ) . \mathcal{F}_{m}u:=\mathcal{F}^{-1}(m\widehat{u}),\qquad u\in D(\mathcal{F}_{m}). F m u := F − 1 ( m u ) , u ∈ D ( F m ) .
Note that in the case of p ∈ [ 2 , ∞ ] p\in[2,\infty] p ∈ [ 2 , ∞ ] , one automatically has D ( F m ) = S ( R n ) D(\mathcal{F}_{m})=\mathcal{S}(\mathbb{R}^{n}) D ( F m ) = S ( R n ) as F − 1 ( m u ^ ) ∈ L p ( R n ) \mathcal{F}^{-1}(m\widehat{u})\in L^{p}(\mathbb{R}^{n}) F − 1 ( m u ) ∈ L p ( R n ) for all u ∈ S ( R n ) u\in\mathcal{S}(\mathbb{R}^{n}) u ∈ S ( R n ) .
Notation 2.3. If the operator F m \mathcal{F}_{m} F m in Definition 2.2 is closable (see Section 2.1), its closure in L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) is denoted by F ‾ m \overline{\mathcal{F}}_{m} F m .
If the domain of F ‾ m \overline{\mathcal{F}}_{m} F m is L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) , then m m m is called an L p L^{p} L p -Fourier multiplier, while F ‾ m \overline{\mathcal{F}}_{m} F m is called an L p L^{p} L p -Fourier multiplier operator. It is well known that for p = 2 p=2 p = 2 the condition m ∈ L ∞ ( R n ) m\in L^{\infty}(\mathbb{R}^{n}) m ∈ L ∞ ( R n ) is necessary and sufficient for m m m to be an L 2 L^{2} L 2 -Fourier multiplier (see [17 Theorem 2.5.10] ). For general p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) , one needs to require stronger properties for m m m to extend F m \mathcal{F}_{m} F m to a bounded operator in all of L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Sufficient conditions are, for example, devised by the Mihlin-Hörmander multiplier theorem, which will be used repeatedly throughout the paper (see, e.g., [17 Theorem 6.2.7] ).
The next lemma shows that for smooth multipliers m m m with derivatives of polynomial growth, the operator F m \mathcal{F}_m F m has the following basic properties, which, in particular, allow us to define its closure F m ‾ \overline{\mathcal{F}_m} F m .
Lemma 2.4. Let m : R n → R m:\mathbb{R}^n \to\mathbb{R} m : R n → R be a smooth function such that all its derivatives have polynomial growth. Then the unbounded operator F m \mathcal{F}_m F m in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) maps from D ( F m ) = S ( R n ) D(\mathcal{F}_m)=\mathcal{S}(\mathbb{R}^n) D ( F m ) = S ( R n ) into S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) and is closable. Moreover, for all u ∈ D ( F m ‾ ) u\in D(\overline{\mathcal{F}_m}) u ∈ D ( F m ) the expression F − 1 ( m u ^ ) \mathcal{F}^{-1}(m\hat{u}) F − 1 ( m u ^ ) exists as tempered distribution in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , and it holds
F − 1 ( m u ^ ) = F m ‾ u . \mathcal{F}^{-1}(m\hat{u})=\overline{\mathcal{F}_m}u. F − 1 ( m u ^ ) = F m u .
Proof. Let u ∈ S ( R n ) u\in\mathcal{S}(\mathbb{R}^n) u ∈ S ( R n ) . Then u ^ ∈ S ( R n ) \hat{u}\in\mathcal{S}(\mathbb{R}^n) u ^ ∈ S ( R n ) , and since all derivatives of m m m have polynomial growth, m u ^ ∈ S ( R n ) m\hat{u}\in\mathcal{S}(\mathbb{R}^n) m u ^ ∈ S ( R n ) . Consequently, F − 1 ( m u ^ ) \mathcal{F}^{-1}(m\hat{u}) F − 1 ( m u ^ ) is well defined and lies in S ( R n ) ⊂ L p ( R n ) \mathcal{S}(\mathbb{R}^n)\subset L^p(\mathbb{R}^n) S ( R n ) ⊂ L p ( R n ) .
Now we prove that F m \mathcal{F}_m F m is closable. Let ( u j ) j (u_j)_j ( u j ) j be a sequence in D ( F m ) D(\mathcal{F}_m) D ( F m ) such that u j → 0 u_j\to0 u j → 0 in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) and F m u j → v \mathcal{F}_m u_j\to v F m u j → v in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) for some v ∈ L p ( R n ) v\in L^p(\mathbb{R}^n) v ∈ L p ( R n ) . Using that F m u j → v \mathcal{F}_m u_j\to v F m u j → v in S ′ ( R n ) \mathcal{S}'(\mathbb{R}^n) S ′ ( R n ) , we obtain that for all φ ∈ S ( R n ) \varphi\in\mathcal{S}(\mathbb{R}^n) φ ∈ S ( R n ) ,
lim j → ∞ ⟨ F m u j , φ ⟩ = ⟨ v , φ ⟩ . \lim_{j\to\infty}\langle\mathcal{F}_m u_j,\varphi\rangle=\langle v,\varphi\rangle. j → ∞ lim ⟨ F m u j , φ ⟩ = ⟨ v , φ ⟩ .
But, as F ( m F − 1 φ ) ∈ S ( R n ) \mathcal{F}(m\mathcal{F}^{-1}\varphi)\in\mathcal{S}(\mathbb{R}^n) F ( m F − 1 φ ) ∈ S ( R n ) by the assumption on m m m ,
lim j → ∞ ⟨ F m u j , φ ⟩ = lim j → ∞ ⟨ m u j ^ , F − 1 φ ⟩ = lim j → ∞ ⟨ u j , F ( m F − 1 φ ) ⟩ = 0. \lim_{j\to\infty}\langle\mathcal{F}_m u_j,\varphi\rangle
=\lim_{j\to\infty}\langle m\widehat{u_j},\mathcal{F}^{-1}\varphi\rangle
=\lim_{j\to\infty}\langle u_j,\mathcal{F}(m\mathcal{F}^{-1}\varphi)\rangle=0. j → ∞ lim ⟨ F m u j , φ ⟩ = j → ∞ lim ⟨ m u j , F − 1 φ ⟩ = j → ∞ lim ⟨ u j , F ( m F − 1 φ )⟩ = 0.
Thus, ⟨ v , φ ⟩ = 0 \langle v,\varphi\rangle=0 ⟨ v , φ ⟩ = 0 for all φ ∈ S ( R n ) \varphi\in\mathcal{S}(\mathbb{R}^n) φ ∈ S ( R n ) , yielding v = 0 v=0 v = 0 .
Finally, in order to see the last claim, let u ∈ D ( F m ‾ ) u\in D(\overline{\mathcal{F}_m}) u ∈ D ( F m ) and ( w j ) j ⊂ D ( F m ) (w_j)_j\subset D(\mathcal{F}_m) ( w j ) j ⊂ D ( F m ) with w j → u w_j\to u w j → u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) and such that F m w j → F m ‾ u \mathcal{F}_m w_j\to\overline{\mathcal{F}_m}u F m w j → F m u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Then, one finds
⟨ F m ‾ u , φ ⟩ = lim j → ∞ ⟨ F m w j , φ ⟩ = lim j → ∞ ⟨ w j , F ( m F − 1 φ ) ⟩ = ⟨ u , F ( m F − 1 φ ) ⟩ , φ ∈ S ( R n ) , \langle\overline{\mathcal{F}_m}u,\varphi\rangle
=\lim_{j\to\infty}\langle\mathcal{F}_m w_j,\varphi\rangle
=\lim_{j\to\infty}\langle w_j,\mathcal{F}(m\mathcal{F}^{-1}\varphi)\rangle
=\langle u,\mathcal{F}(m\mathcal{F}^{-1}\varphi)\rangle,\qquad\varphi\in\mathcal{S}(\mathbb{R}^n), ⟨ F m u , φ ⟩ = j → ∞ lim ⟨ F m w j , φ ⟩ = j → ∞ lim ⟨ w j , F ( m F − 1 φ )⟩ = ⟨ u , F ( m F − 1 φ )⟩ , φ ∈ S ( R n ) ,
so that F − 1 ( m u ^ ) \mathcal{F}^{-1}(m\hat{u}) F − 1 ( m u ^ ) as a tempered distribution is equal to F m ‾ u \overline{\mathcal{F}_m}u F m u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , concluding the proof. □
Naturally, and for our later purposes in Section 4 very important, the question arises, how concatenations of unbounded operators as introduced in Definition 2.2 behave. In that respect, one always has at least that F m 2 m 1 \mathcal{F}_{m_2m_1} F m 2 m 1 extends F m 2 ‾ F m 1 \overline{\mathcal{F}_{m_2}}\mathcal{F}_{m_1} F m 2 F m 1 as unbounded operators. For the reader’s convenience, we give below a detailed proof of this property. We add that this behavior is well known in the case of L p L^p L p -Fourier multipliers (see [17 Proposition 2.5.13] ), where even the equality F m 2 ‾ F m 1 = F m 2 m 1 \overline{\mathcal{F}_{m_2}}\mathcal{F}_{m_1}=\mathcal{F}_{m_2m_1} F m 2 F m 1 = F m 2 m 1 holds.
Lemma 2.5. Let m 1 , m 2 : R n → R m_1,m_2:\mathbb{R}^n\to\mathbb{R} m 1 , m 2 : R n → R be measurable functions with polynomial growth, consider the operators F m 1 , F m 2 \mathcal{F}_{m_1},\mathcal{F}_{m_2} F m 1 , F m 2 in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , and further assume that F m 2 \mathcal{F}_{m_2} F m 2 is closable. Then
D ( F m 2 ‾ F m 1 ) ⊂ D ( F m 2 m 1 ) D(\overline{\mathcal{F}_{m_2}}\mathcal{F}_{m_1})\subset D(\mathcal{F}_{m_2m_1}) D ( F m 2 F m 1 ) ⊂ D ( F m 2 m 1 ) and
F m 2 ‾ F m 1 u = F m 2 m 1 u , u ∈ D ( F m 2 ‾ F m 1 ) . (30) \overline{\mathcal{F}_{m_2}}\mathcal{F}_{m_1}u=\mathcal{F}_{m_2m_1}u,\qquad u\in D(\overline{\mathcal{F}_{m_2}}\mathcal{F}_{m_1}). \tag*{(30)} F m 2 F m 1 u = F m 2 m 1 u , u ∈ D ( F m 2 F m 1 ) . (30)
Proof. Let u ∈ D ( F m 2 ‾ F m 1 ) u\in D(\overline{\mathcal{F}_{m_2}}\mathcal{F}_{m_1}) u ∈ D ( F m 2 F m 1 ) , recalling (2.1). As F m 1 u ∈ D ( F m 2 ‾ ) \mathcal{F}_{m_1}u\in D(\overline{\mathcal{F}_{m_2}}) F m 1 u ∈ D ( F m 2 ) , there exists a sequence ( v j ) j ⊂ D ( F m 2 ) (v_j)_j\subset D(\mathcal{F}_{m_2}) ( v j ) j ⊂ D ( F m 2 ) such that v j → F m 1 u v_j\to\mathcal{F}_{m_1}u v j → F m 1 u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) and
F m 2 v j → F m 2 ‾ F m 1 u in L p ( R n ) . \mathcal{F}_{m_2}v_j\to\overline{\mathcal{F}_{m_2}}\mathcal{F}_{m_1}u\quad\text{in }L^p(\mathbb{R}^n). F m 2 v j → F m 2 F m 1 u in L p ( R n ) .
At the same time, by Definition 2.2, for all φ ∈ C ∞ ( R n ) \varphi\in C^\infty(\mathbb{R}^n) φ ∈ C ∞ ( R n ) ,
⟨ F m 2 v j , φ ⟩ = ⟨ m 2 v j ^ , F − 1 φ ⟩ = ⟨ v j ^ , m 2 F − 1 φ ⟩ . \langle\mathcal{F}_{m_2}v_j,\varphi\rangle
=\langle m_2\widehat{v_j},\mathcal{F}^{-1}\varphi\rangle
=\langle\widehat{v_j},m_2\mathcal{F}^{-1}\varphi\rangle. ⟨ F m 2 v j , φ ⟩ = ⟨ m 2 v j , F − 1 φ ⟩ = ⟨ v j , m 2 F − 1 φ ⟩ .
Note that m 2 F − 1 φ ∈ L p ′ ( R n ) m_2\mathcal{F}^{-1}\varphi\in L^{p'}(\mathbb{R}^n) m 2 F − 1 φ ∈ L p ′ ( R n ) , since F − 1 φ ∈ S ( R n ) \mathcal{F}^{-1}\varphi\in\mathcal{S}(\mathbb{R}^n) F − 1 φ ∈ S ( R n ) and m 2 m_2 m 2 has polynomial growth. Hence, we can choose a second sequence ( w k ) k ⊂ S ( R n ) (w_k)_k\subset\mathcal{S}(\mathbb{R}^n) ( w k ) k ⊂ S ( R n ) with w k → m 2 F − 1 φ w_k\to m_2\mathcal{F}^{-1}\varphi w k → m 2 F − 1 φ in L p ′ ( R n ) L^{p'}(\mathbb{R}^n) L p ′ ( R n ) . On the one hand, this entails that
lim j → ∞ lim k → ∞ ⟨ v j ^ , w k ⟩ = lim j → ∞ ⟨ v j ^ , m 2 F − 1 φ ⟩ = lim j → ∞ ⟨ F m 2 v j , φ ⟩ = ⟨ F m 2 ‾ F m 1 u , φ ⟩ . (31) \lim_{j\to\infty}\lim_{k\to\infty}\langle\widehat{v_j},w_k\rangle
=\lim_{j\to\infty}\langle\widehat{v_j},m_2\mathcal{F}^{-1}\varphi\rangle
=\lim_{j\to\infty}\langle\mathcal{F}_{m_2}v_j,\varphi\rangle
=\langle\overline{\mathcal{F}_{m_2}}\mathcal{F}_{m_1}u,\varphi\rangle. \tag*{(31)} j → ∞ lim k → ∞ lim ⟨ v j , w k ⟩ = j → ∞ lim ⟨ v j , m 2 F − 1 φ ⟩ = j → ∞ lim ⟨ F m 2 v j , φ ⟩ = ⟨ F m 2 F m 1 u , φ ⟩ . (31)
while, on the other hand,
lim j → ∞ ⟨ v j ^ , w k ⟩ = lim j → ∞ ⟨ v j , w k ^ ⟩ = ⟨ F m 1 u , w k ^ ⟩ = ⟨ m 1 u ^ , w k ⟩ , k ∈ N (32) \lim_{j\to\infty}\langle\widehat{v_j},w_k\rangle=\lim_{j\to\infty}\langle v_j,\widehat{w_k}\rangle=\langle F_{m_1}u,\widehat{w_k}\rangle=\langle\widehat{m_1u},w_k\rangle,\qquad k\in\mathbb{N} \tag*{(32)} j → ∞ lim ⟨ v j , w k ⟩ = j → ∞ lim ⟨ v j , w k ⟩ = ⟨ F m 1 u , w k ⟩ = ⟨ m 1 u , w k ⟩ , k ∈ N (32)
and
lim k → ∞ ⟨ m 1 u ^ , w k ⟩ = ⟨ m 1 u ^ , m 2 F − 1 φ ⟩ = ⟨ m 2 m 1 u ^ , F − 1 φ ⟩ , \lim_{k\to\infty}\langle\widehat{m_1u},w_k\rangle=\langle\widehat{m_1u},m_2\mathcal{F}^{-1}\varphi\rangle=\langle\widehat{m_2m_1u},\mathcal{F}^{-1}\varphi\rangle, k → ∞ lim ⟨ m 1 u , w k ⟩ = ⟨ m 1 u , m 2 F − 1 φ ⟩ = ⟨ m 2 m 1 u , F − 1 φ ⟩ ,
and together
lim k → ∞ lim j → ∞ ⟨ v j ^ , w k ⟩ = ⟨ m 2 m 1 u ^ , F − 1 φ ⟩ . (33) \lim_{k\to\infty}\lim_{j\to\infty}\langle\widehat{v_j},w_k\rangle=\langle\widehat{m_2m_1u},\mathcal{F}^{-1}\varphi\rangle. \tag*{(33)} k → ∞ lim j → ∞ lim ⟨ v j , w k ⟩ = ⟨ m 2 m 1 u , F − 1 φ ⟩ . (33)
As m 1 m_1 m 1 has polynomial growth and u ∈ S ( R n ) u\in\mathcal{S}(\mathbb{R}^n) u ∈ S ( R n ) , it holds m 1 u ^ ∈ L p ( R n ) \widehat{m_1u}\in L^p(\mathbb{R}^n) m 1 u ∈ L p ( R n ) . Moreover, similarly to (32) one infers that ⟨ v j ^ , ψ ⟩ → ⟨ m 1 u ^ , ψ ⟩ \langle\widehat{v_j},\psi\rangle\to\langle\widehat{m_1u},\psi\rangle ⟨ v j , ψ ⟩ → ⟨ m 1 u , ψ ⟩ for all ψ ∈ S ( R n ) \psi\in\mathcal{S}(\mathbb{R}^n) ψ ∈ S ( R n ) , and hence, ( v j ^ ) j (\widehat{v_j})_j ( v j ) j converges weakly to m 1 u ^ \widehat{m_1u} m 1 u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . In particular, ( v j ^ ) j (\widehat{v_j})_j ( v j ) j is bounded in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . We conclude that the limit lim k → ∞ ⟨ v j ^ , w k ⟩ \lim_{k\to\infty}\langle\widehat{v_j},w_k\rangle lim k → ∞ ⟨ v j , w k ⟩ is attained uniformly in j ∈ N j\in\mathbb{N} j ∈ N , which follows from
lim k → ∞ sup j ∈ N ∣ ⟨ v j ^ , w k ⟩ − ⟨ v j ^ , m 2 F − 1 φ ⟩ ∣ ≤ lim k → ∞ ( sup j ∈ N ∥ v j ^ ∥ L p ( R n ) ) ∥ w k − m 2 F − 1 φ ∥ L p ′ ( R n ) = 0. \lim_{k\to\infty}\sup_{j\in\mathbb{N}}\left|\langle\widehat{v_j},w_k\rangle-\langle\widehat{v_j},m_2\mathcal{F}^{-1}\varphi\rangle\right|\leq\lim_{k\to\infty}\left(\sup_{j\in\mathbb{N}}\|\widehat{v_j}\|_{L^p(\mathbb{R}^n)}\right)\|w_k-m_2\mathcal{F}^{-1}\varphi\|_{L^{p'}(\mathbb{R}^n)}=0. k → ∞ lim j ∈ N sup ⟨ v j , w k ⟩ − ⟨ v j , m 2 F − 1 φ ⟩ ≤ k → ∞ lim ( j ∈ N sup ∥ v j ∥ L p ( R n ) ) ∥ w k − m 2 F − 1 φ ∥ L p ′ ( R n ) = 0.
By a classical argument,the Moore–Osgood theorem on the interchange of limits, this implies that the two double-limits in (31) and (33) agree. By Definition 2.2, u ∈ D ( F m 2 m 1 ‾ ) u\in D(\overline{F_{m_2m_1}}) u ∈ D ( F m 2 m 1 ) with
F m 2 m 1 ‾ u = F m 2 ‾ F m 1 ‾ u , \overline{F_{m_2m_1}}u=\overline{F_{m_2}}\overline{F_{m_1}}u, F m 2 m 1 u = F m 2 F m 1 u ,
which yields the statement. □
Moreover, one has the inclusion of domains D ( F m 1 ‾ ) ⊂ D ( F m 2 m 1 ‾ ) D(\overline{F_{m_1}})\subset D(\overline{F_{m_2m_1}}) D ( F m 1 ) ⊂ D ( F m 2 m 1 ) of the respective closed operators, if at least m 2 m_2 m 2 is an L p L^p L p -Fourier multiplier.
Lemma 2.6. Let m 1 , m 2 : R n → R m_1,m_2:\mathbb{R}^n\to\mathbb{R} m 1 , m 2 : R n → R be measurable functions with polynomial growth, consider the operators F m 1 , F m 2 F_{m_1},F_{m_2} F m 1 , F m 2 in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , and further assume that F m 1 F_{m_1} F m 1 is closable. If m 2 m_2 m 2 is an L p L^p L p -Fourier multiplier and also F m 1 m 2 F_{m_1m_2} F m 1 m 2 is closable, then
D ( F m 1 ‾ ) ⊂ D ( F m 2 m 1 ‾ ) , D(\overline{F_{m_1}})\subset D(\overline{F_{m_2m_1}}), D ( F m 1 ) ⊂ D ( F m 2 m 1 ) , and for any u ∈ D ( F m 1 ‾ ) u\in D(\overline{F_{m_1}}) u ∈ D ( F m 1 ) and ( u j ) j ⊂ D ( F m 1 ‾ ) (u_j)_j\subset D(\overline{F_{m_1}}) ( u j ) j ⊂ D ( F m 1 ) with u j → u u_j\to u u j → u and F m 1 u j → F m 1 ‾ u F_{m_1}u_j\to\overline{F_{m_1}}u F m 1 u j → F m 1 u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , it holds
F m 2 m 1 ‾ u j → F m 2 m 1 ‾ u in L p ( R n ) . \overline{F_{m_2m_1}}u_j\to\overline{F_{m_2m_1}}u\quad\text{in }L^p(\mathbb{R}^n). F m 2 m 1 u j → F m 2 m 1 u in L p ( R n ) .
Proof. First observe that since F m 2 ‾ \overline{F_{m_2}} F m 2 is an L p L^p L p -multiplier operator, and thus, D ( F m 2 ‾ ) = L p ( R n ) D(\overline{F_{m_2}})=L^p(\mathbb{R}^n) D ( F m 2 ) = L p ( R n ) , it holds in view of Definition 2.2 that D ( F m 1 ‾ ) = { u ∈ S ( R n ) : F m 1 u ∈ L p ( R n ) } = D ( F m 2 F m 1 ‾ ) D(\overline{F_{m_1}})=\{u\in\mathcal{S}(\mathbb{R}^n):F_{m_1}u\in L^p(\mathbb{R}^n)\}=D(\overline{F_{m_2}F_{m_1}}) D ( F m 1 ) = { u ∈ S ( R n ) : F m 1 u ∈ L p ( R n )} = D ( F m 2 F m 1 ) . Consequently, Lemma 2.5 implies in particular that D ( F m 1 ‾ ) ⊂ D ( F m 2 m 1 ‾ ) D(\overline{F_{m_1}})\subset D(\overline{F_{m_2m_1}}) D ( F m 1 ) ⊂ D ( F m 2 m 1 ) .
Let u ∈ D ( F m 1 ‾ ) u\in D(\overline{F_{m_1}}) u ∈ D ( F m 1 ) and choose a sequence ( u j ) j ⊂ D ( F m 1 ‾ ) (u_j)_j\subset D(\overline{F_{m_1}}) ( u j ) j ⊂ D ( F m 1 ) with u j → u u_j\to u u j → u and F m 1 u j → F m 1 ‾ u F_{m_1}u_j\to\overline{F_{m_1}}u F m 1 u j → F m 1 u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Since F m 2 ‾ F m 1 u j = F m 1 m 2 u j \overline{F_{m_2}}F_{m_1}u_j=F_{m_1m_2}u_j F m 2 F m 1 u j = F m 1 m 2 u j for all j ∈ N j\in\mathbb{N} j ∈ N by (30) , and because F m 2 ‾ \overline{F_{m_2}} F m 2 is an L p L^p L p -Fourier multiplier operator, one deduces that
F m 2 m 1 ‾ u j = F m 2 ‾ F m 1 u j → F m 2 m 1 ‾ u in L p ( R n ) , \overline{F_{m_2m_1}}u_j=\overline{F_{m_2}}F_{m_1}u_j\to\overline{F_{m_2m_1}}u\quad\text{in }L^p(\mathbb{R}^n), F m 2 m 1 u j = F m 2 F m 1 u j → F m 2 m 1 u in L p ( R n ) ,
which yields u ∈ D ( F m 2 m 1 ‾ ) . □ u\in D(\overline{F_{m_2m_1}}). \square u ∈ D ( F m 2 m 1 ) . □
We conclude this section with a result that, in the cases we will be interested in, reduces proving an identity of the type F m 2 ‾ F m 1 ‾ = F m 2 m 1 ‾ \overline{F_{m_2}}\overline{F_{m_1}}=\overline{F_{m_2m_1}} F m 2 F m 1 = F m 2 m 1 of operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) to showing the equality D ( F m 2 ‾ F m 1 ‾ ) = D ( F m 2 m 1 ‾ ) D(\overline{F_{m_2}}\overline{F_{m_1}})=D(\overline{F_{m_2m_1}}) D ( F m 2 F m 1 ) = D ( F m 2 m 1 ) of their domains.
Lemma 2.7. Let m 1 , m 2 : R n → R m_1,m_2:\mathbb{R}^n\to\mathbb{R} m 1 , m 2 : R n → R be measurable functions with polynomial growth, consider the operators F m 1 , F m 2 F_{m_1},F_{m_2} F m 1 , F m 2 in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , and further assume that F m 1 , F m 2 F_{m_1},F_{m_2} F m 1 , F m 2 are closable and that F ( m 2 F − 1 φ ) , F ( m 2 m 1 F − 1 φ ) \mathcal{F}(m_2\mathcal{F}^{-1}\varphi),\mathcal{F}(m_2m_1\mathcal{F}^{-1}\varphi) F ( m 2 F − 1 φ ) , F ( m 2 m 1 F − 1 φ ) lie in L p ′ ( R n ) L^{p'}(\mathbb{R}^n) L p ′ ( R n ) for any φ ∈ S ( R n ) \varphi\in\mathcal{S}(\mathbb{R}^n) φ ∈ S ( R n ) . Then, if
D ( F m 2 ‾ F m 1 ‾ ) = D ( F m 2 m 1 ‾ ) , D(\overline{F_{m_2}}\overline{F_{m_1}})=D(\overline{F_{m_2m_1}}), D ( F m 2 F m 1 ) = D ( F m 2 m 1 ) , it follows that
F m 2 ‾ F m 1 ‾ = F m 2 m 1 ‾ \overline{F_{m_2}}\overline{F_{m_1}}=\overline{F_{m_2m_1}} F m 2 F m 1 = F m 2 m 1 as operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) .
Proof. Let u ∈ D ( F ‾ m 2 F ‾ m 1 ) = D ( F ‾ m 2 m 1 ) u \in D(\overline{F}_{m_2}\overline{F}_{m_1}) = D(\overline{F}_{m_2m_1}) u ∈ D ( F m 2 F m 1 ) = D ( F m 2 m 1 ) , and take ( u j ) j , ( v j ) j ⊂ S ( R n ) (u_j)_j,(v_j)_j \subset\mathcal{S}(\mathbb{R}^n) ( u j ) j , ( v j ) j ⊂ S ( R n ) such that u j → u u_j \to u u j → u , F ‾ m 1 u j → F ‾ m 1 u \overline{F}_{m_1}u_j \to\overline{F}_{m_1}u F m 1 u j → F m 1 u , v j → F ‾ m 1 u v_j \to\overline{F}_{m_1}u v j → F m 1 u , F ‾ m 2 v j → F ‾ m 2 F ‾ m 1 u \overline{F}_{m_2}v_j \to\overline{F}_{m_2}\overline{F}_{m_1}u F m 2 v j → F m 2 F m 1 u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) .
We compute
⟨ F ‾ m 2 F ‾ m 1 u , φ ⟩ = lim j → ∞ ⟨ F ‾ m 2 v j , φ ⟩ = lim j → ∞ ⟨ v j , F ( m 2 F − 1 φ ) ⟩ = lim j → ∞ ⟨ F ‾ m 1 u j , F ( m 2 F − 1 φ ) ⟩ = lim j → ∞ ⟨ u j , F ( m 2 m 1 F − 1 φ ) ⟩ = ⟨ u , F ( m 2 m 1 F − 1 φ ) ⟩ = ⟨ F ‾ m 2 m 1 u , φ ⟩ , φ ∈ S ( R n ) , \begin{aligned}
\langle\overline{F}_{m_2}\overline{F}_{m_1}u,\varphi\rangle
&= \lim_{j\to\infty}\langle\overline{F}_{m_2}v_j,\varphi\rangle
= \lim_{j\to\infty}\langle v_j,\mathcal{F}(m_2\mathcal{F}^{-1}\varphi)\rangle\\
&= \lim_{j\to\infty}\langle\overline{F}_{m_1}u_j,\mathcal{F}(m_2\mathcal{F}^{-1}\varphi)\rangle
= \lim_{j\to\infty}\langle u_j,\mathcal{F}(m_2m_1\mathcal{F}^{-1}\varphi)\rangle\\
&= \langle u,\mathcal{F}(m_2m_1\mathcal{F}^{-1}\varphi)\rangle
= \langle\overline{F}_{m_2m_1}u,\varphi\rangle,\qquad\varphi\in\mathcal{S}(\mathbb{R}^n),
\end{aligned} ⟨ F m 2 F m 1 u , φ ⟩ = j → ∞ lim ⟨ F m 2 v j , φ ⟩ = j → ∞ lim ⟨ v j , F ( m 2 F − 1 φ )⟩ = j → ∞ lim ⟨ F m 1 u j , F ( m 2 F − 1 φ )⟩ = j → ∞ lim ⟨ u j , F ( m 2 m 1 F − 1 φ )⟩ = ⟨ u , F ( m 2 m 1 F − 1 φ )⟩ = ⟨ F m 2 m 1 u , φ ⟩ , φ ∈ S ( R n ) ,
where we have used in particular that F ( m 2 F − 1 φ ) , F ( m 2 m 1 F − 1 φ ) ∈ L p ( R n ) \mathcal{F}(m_2\mathcal{F}^{-1}\varphi),\mathcal{F}(m_2m_1\mathcal{F}^{-1}\varphi) \in L^p(\mathbb{R}^n) F ( m 2 F − 1 φ ) , F ( m 2 m 1 F − 1 φ ) ∈ L p ( R n ) by assumption. This shows that F ‾ m 2 F ‾ m 1 = F ‾ m 2 m 1 \overline{F}_{m_2}\overline{F}_{m_1} = \overline{F}_{m_2m_1} F m 2 F m 1 = F m 2 m 1 as operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) .
Further hypotheses on the kernel function
The properties of the function Q ^ ρ \widehat{Q}_\rho Q ρ depend strongly on those of the kernel ρ \rho ρ . We assume in the following that ρ ∈ L 1 ( R n ) \rho\in L^1(\mathbb{R}^n) ρ ∈ L 1 ( R n ) .
It is straightforward, using the coarea formula, to rewrite the expression of Q ρ Q_\rho Q ρ in (28) as (see [8 Lemma 4.3] )
Q ^ ρ ( ξ ) = 1 π ∣ ξ ∣ ∫ S + n − 1 ∫ 0 ∞ r n − 2 ρ ‾ ( r ) sin ( 2 π ∣ ξ ∣ r z 1 ) d r d H n − 1 ( z ) , ξ ∈ R n ∖ { 0 } . \widehat{Q}_\rho(\xi)=\frac{1}{\pi|\xi|}\int_{\mathbb{S}^{n-1}_+}\int_0^\infty r^{n-2}\overline{\rho}(r)\sin(2\pi|\xi|rz_1)\,dr\,d\mathcal{H}^{n-1}(z),\qquad\xi\in\mathbb{R}^n\setminus\{0\}. Q ρ ( ξ ) = π ∣ ξ ∣ 1 ∫ S + n − 1 ∫ 0 ∞ r n − 2 ρ ( r ) sin ( 2 π ∣ ξ ∣ r z 1 ) d r d H n − 1 ( z ) , ξ ∈ R n ∖ { 0 } .
Basically by a variant of Pólya’s criterion for the non-negativity of oscillatory integrals of sin-type (see [27 ] ), it was shown in [8 Lemma 4.3] that Q ^ ρ > 0 \widehat{Q}_\rho>0 Q ρ > 0 if r ↦ r n − 2 ρ ‾ ( r ) r\mapsto r^{n-2}\overline{\rho}(r) r ↦ r n − 2 ρ ( r ) is strictly decreasing (actually, in [8 Lemma 4.3] it is only assumed that the function r ↦ r n − 2 ρ ‾ ( r ) r\mapsto r^{n-2}\overline{\rho}(r) r ↦ r n − 2 ρ ( r ) is decreasing, not strictly decreasing; but this is compensated there by the additional doubling property found below in assumption (H1)). In [8 ] , the positivity of Q ^ ρ \widehat{Q}_\rho Q ρ is then used to formally invert the nonlocal derivative in Fourier space by introducing the symbol Q ^ ρ − 1 \widehat{Q}_\rho^{-1} Q ρ − 1 . Additional conditions on ρ \rho ρ are then imposed to derive Poincaré inequalities, compact embeddings and a nonlocal version of the fundamental theorem of calculus. Most of these conditions prove equally relevant to our purposes. We therefore recall those that we shall use, and comment on their implications afterwards. Let δ 0 > 0 \delta_0>0 δ 0 > 0 such that inf B δ 0 ρ > 0 \inf_{B_{\delta_0}}\rho>0 inf B δ 0 ρ > 0 as in (H0).
(H1) The function f ρ : ( 0 , ∞ ) → R f_\rho:(0,\infty)\to\mathbb{R} f ρ : ( 0 , ∞ ) → R , r ↦ r n − 2 ρ ‾ ( r ) r\mapsto r^{n-2}\overline{\rho}(r) r ↦ r n − 2 ρ ( r ) is decreasing, and there exists μ ∈ ( 0 , 1 ) \mu\in(0,1) μ ∈ ( 0 , 1 ) such that μ f ρ ( r / 2 ) ≥ f ρ ( r ) \mu f_\rho(r/2)\ge f_\rho(r) μ f ρ ( r /2 ) ≥ f ρ ( r ) for r ∈ ( 0 , δ 0 ) r\in(0,\delta_0) r ∈ ( 0 , δ 0 ) .
(H2) The function f ρ f_\rho f ρ is smooth in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , and for r ∈ ( 0 , δ 0 ) r\in(0,\delta_0) r ∈ ( 0 , δ 0 ) it holds
− C d d r f ρ ( r ) ≥ f ρ ( r ) r and ∣ d k d r k f ρ ( r ) ∣ ≤ C k f ρ ( r ) r k , k ∈ N . -C\frac{d}{dr}f_\rho(r)\ge\frac{f_\rho(r)}{r}
\qquad\text{and}\qquad
\left|\frac{d^k}{dr^k}f_\rho(r)\right|\le C_k\frac{f_\rho(r)}{r^k},\qquad k\in\mathbb{N}. − C d r d f ρ ( r ) ≥ r f ρ ( r ) and d r k d k f ρ ( r ) ≤ C k r k f ρ ( r ) , k ∈ N .
(H3) There exists σ ∈ ( 0 , 1 ) \sigma\in(0,1) σ ∈ ( 0 , 1 ) such that the function g ρ : ( 0 , ∞ ) → R g_\rho:(0,\infty)\to\mathbb{R} g ρ : ( 0 , ∞ ) → R , r ↦ r n + σ − 1 ρ ‾ ( r ) r\mapsto r^{n+\sigma-1}\overline{\rho}(r) r ↦ r n + σ − 1 ρ ( r ) is almost decreasing on ( 0 , δ 0 ) (0,\delta_0) ( 0 , δ 0 ) .
The expression almost decreasing in (H3) means that g ρ ( r 1 ) ≥ C g ρ ( r 2 ) g_\rho(r_1)\ge Cg_\rho(r_2) g ρ ( r 1 ) ≥ C g ρ ( r 2 ) for all r 1 ≤ r 2 ∈ ( 0 , δ 0 ) r_1\le r_2\in(0,\delta_0) r 1 ≤ r 2 ∈ ( 0 , δ 0 ) and some fixed constant C > 0 C>0 C > 0 . Qualitatively speaking, (H3) encodes that ρ \rho ρ lies around the origin above the kernel ∣ ⋅ ∣ − ( n + σ − 1 ) |\cdot|^{-(n+\sigma-1)} ∣ ⋅ ∣ − ( n + σ − 1 ) (which gives rise to the Riesz fractional gradient of order σ \sigma σ ), and thus, has at least a singularity of the corresponding order in the origin. We also introduce as a weaker replacement of condition (H3)
lim inf t ↓ 0 t n − 1 ρ ‾ ( t ) > 0. (34) \liminf_{t\downarrow0}t^{n-1}\overline{\rho}(t)>0. \tag*{(34)} t ↓ 0 lim inf t n − 1 ρ ( t ) > 0. (34)
For more details on the different hypotheses we refer the reader to [8 ] .
Hypothesis (H1) ensures not only that Q ^ ρ \widehat{Q}_\rho Q ρ is positive, as remarked above, but it also guarantees the validity of the estimate (see [8 Lemma 4.3] )
Q ^ ρ ( ξ ) ≥ C ρ ‾ ( 1 / ∣ ξ ∣ ) ∣ ξ ∣ n , ξ ∈ B 2 / δ 0 c . \widehat{Q}_\rho(\xi)\ge C\frac{\overline{\rho}(1/|\xi|)}{|\xi|^n},\qquad\xi\in B^c_{2/\delta_0}. Q ρ ( ξ ) ≥ C ∣ ξ ∣ n ρ ( 1/∣ ξ ∣ ) , ξ ∈ B 2/ δ 0 c .
Building on this estimate, as outlined at the start of the proof of [8 Lemma 4.11] , under the additional hypothesis (H3’) it holds
Q ^ ρ ( ξ ) ≥ C ∣ ξ ∣ − 1 , ξ ∈ B 2 / δ 0 c . (35) \widehat{Q}_{\rho}(\xi) \ge C|\xi|^{-1}, \qquad\xi\in B^{c}_{2/\delta_{0}}. \tag*{(35)} Q ρ ( ξ ) ≥ C ∣ ξ ∣ − 1 , ξ ∈ B 2/ δ 0 c . (35)
Moreover, with hypothesis (H3) one can strengthen this lower bound. In fact, under conditions (H0), (H1) and (H3) one has (see in [8 Eq. (5.1)] )
Q ^ ρ ( ξ ) ≥ C ∣ ξ ∣ σ − 1 ≥ C ∣ ξ ∣ − 1 , ξ ∈ B 2 / δ 0 c . (36) \widehat{Q}_{\rho}(\xi) \ge C|\xi|^{\sigma-1} \ge C|\xi|^{-1}, \qquad\xi\in B^{c}_{2/\delta_{0}}. \tag*{(36)} Q ρ ( ξ ) ≥ C ∣ ξ ∣ σ − 1 ≥ C ∣ ξ ∣ − 1 , ξ ∈ B 2/ δ 0 c . (36)
Finally, assuming additionally that ρ \rho ρ has compact support, so that Q ^ ρ \widehat{Q}_{\rho} Q ρ is smooth by the Paley–Wiener theorem, and imposing hypothesis (H2), allows to deduce upper bounds for the derivatives ∂ α Q ^ ρ \partial^{\alpha}\widehat{Q}_{\rho} ∂ α Q ρ (see [8 Lemma 4.10] ). In particular, if (H0), (H1), (H2) and (H3’) are satisfied, then
∣ ∂ α Q ^ ρ ( ξ ) ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ Q ^ ρ ( ξ ) , ξ ∈ B 2 / δ 0 c . (37) |\partial^{\alpha}\widehat{Q}_{\rho}(\xi)| \le C_{\alpha}|\xi|^{-|\alpha|}\widehat{Q}_{\rho}(\xi), \qquad\xi\in B^{c}_{2/\delta_{0}}. \tag*{(37)} ∣ ∂ α Q ρ ( ξ ) ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ Q ρ ( ξ ) , ξ ∈ B 2/ δ 0 c . (37)
If ρ \rho ρ has compact support, then, by the smoothness of Q ^ ρ \widehat{Q}_{\rho} Q ρ , all estimates above that are valid on B 2 / δ 0 c B^{c}_{2/\delta_{0}} B 2/ δ 0 c also hold true on B 1 c B^{c}_{1} B 1 c (up to adapting the constants in the inequalities). For this reason, we will drop the specific δ 0 > 0 \delta_{0}>0 δ 0 > 0 in the estimates from here on and work throughout the paper exclusively on B 1 B_{1} B 1 and its complement B 1 c B^{c}_{1} B 1 c .
As we have seen above, assuming that ρ \rho ρ has compact support simplifies the analysis in some aspects. Not only does this imply that ρ ∈ L 1 ( R n ) \rho\in L^{1}(\mathbb{R}^{n}) ρ ∈ L 1 ( R n ) and Q ρ ∈ L 1 ( R n ) Q_{\rho} \in L^{1}(\mathbb{R}^{n}) Q ρ ∈ L 1 ( R n ) , since ρ ∈ L l o c 1 ( R n ) \rho\in L^{1}_{\mathrm{loc}}(\mathbb{R}^{n}) ρ ∈ L loc 1 ( R n ) according to hypothesis (H0), but more importantly, it also guarantees that Q ρ Q_{\rho} Q ρ again has compact support and Q ^ ρ \widehat{Q}_{\rho} Q ρ is smooth. As it turns out, the assumption of compact support is not too restrictive, since for any kernel ρ \rho ρ satisfying (H0) and any radially decreasing χ ∈ L ∞ ( R n ) \chi\in L^{\infty}(\mathbb{R}^{n}) χ ∈ L ∞ ( R n ) such that χ ∣ B r ≡ 1 \chi|_{B_{r}}\equiv1 χ ∣ B r ≡ 1 for some r > 0 r>0 r > 0 , we have H 0 ρ , p ( Ω ) = H 0 χ ρ , p ( Ω ) H_{0}^{\rho,p}(\Omega)=H_{0}^{\chi\rho,p}(\Omega) H 0 ρ , p ( Ω ) = H 0 χ ρ , p ( Ω ) (see [8 Example 3.11] ). In addition, if ρ \rho ρ satisfies any of the conditions (H1)–(H3) and (H3’), then so does χ ρ \chi\rho χ ρ . For completeness, we also mention that [8 Proposition 3.10] proves that if H 0 ρ , p ( Ω ) H_{0}^{\rho,p}(\Omega) H 0 ρ , p ( Ω ) , admits a Poincaré inequality, then so does H 0 χ ρ , p ( Ω ) H_{0}^{\chi\rho,p}(\Omega) H 0 χ ρ , p ( Ω ) .
Basic Fourier multiplier results involving the symbol λ ρ \lambda_{\rho} λ ρ
We will start this section by presenting important estimates involving the Fourier symbol λ ρ \lambda_{\rho} λ ρ (see § 2.3 ) of the nonlocal derivative D ρ p D_{\rho}^{p} D ρ p . These estimates build upon the inequalities in § 2.5 , and their proof, only consisting of careful computations, is deferred to Appendix A.
Lemma 2.8. Assume that ρ \rho ρ has compact support and satisfies assumptions (H0)–(H2) and (H3’). Then, for all ξ ∈ B 1 c \xi\in B^{c}_{1} ξ ∈ B 1 c and all multi-indices α ∈ N 0 n \alpha\in\mathbb{N}_{0}^{n} α ∈ N 0 n ,
∣ ∂ α λ ρ ( ξ ) ∣ , ∣ ∂ α ∣ λ ρ ( ξ ) ∣ ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ ∣ ξ ∣ − ∣ α ∣ , (38) |\partial^{\alpha}\lambda_{\rho}(\xi)|, |\partial^{\alpha}|\lambda_{\rho}(\xi)|| \le C_{\alpha}|\lambda_{\rho}(\xi)||\xi|^{-|\alpha|}, \tag*{(38)} ∣ ∂ α λ ρ ( ξ ) ∣ , ∣ ∂ α ∣ λ ρ ( ξ ) ∣∣ ≤ C α ∣ λ ρ ( ξ ) ∣∣ ξ ∣ − ∣ α ∣ , (38) ∣ ∂ α ( ∣ λ ρ ( ξ ) ∣ γ ) ∣ , ∣ ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ ) γ ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ γ ∣ ξ ∣ − ∣ α ∣ , (39) |\partial^{\alpha}(|\lambda_{\rho}(\xi)|^{\gamma})|, |\partial^{\alpha}(1+|\lambda_{\rho}(\xi)|)^{\gamma}| \le C_{\alpha}|\lambda_{\rho}(\xi)|^{\gamma}|\xi|^{-|\alpha|}, \tag*{(39)} ∣ ∂ α ( ∣ λ ρ ( ξ ) ∣ γ ) ∣ , ∣ ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ ) γ ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ γ ∣ ξ ∣ − ∣ α ∣ , (39) ∣ ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ 2 ∣ ξ ∣ − ∣ α ∣ , (40) |\partial^{\alpha}(1+|\lambda_{\rho}(\xi)|^{2})| \le C_{\alpha}|\lambda_{\rho}(\xi)|^{2}|\xi|^{-|\alpha|}, \tag*{(40)} ∣ ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ 2 ∣ ξ ∣ − ∣ α ∣ , (40) ∣ ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) γ ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ 2 γ ∣ ξ ∣ − ∣ α ∣ , γ ∈ R , (41) |\partial^{\alpha}(1+|\lambda_{\rho}(\xi)|^{2})^{\gamma}| \le C_{\alpha}|\lambda_{\rho}(\xi)|^{2\gamma}|\xi|^{-|\alpha|}, \qquad\gamma\in\mathbb{R}, \tag*{(41)} ∣ ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) γ ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ 2 γ ∣ ξ ∣ − ∣ α ∣ , γ ∈ R , (41) with constants C α > 0 C_{\alpha}>0 C α > 0 .
It is important for our purposes to note that the operators D ρ p D_{\rho}^{p} D ρ p and div ρ p \operatorname{div}_{\rho}^{p} div ρ p map Schwartz functions again to Schwartz functions. Indeed, in view of (26) , it suffices to prove that λ ρ \lambda_{\rho} λ ρ is smooth and all its derivatives have polynomial growth, as Lemma 2.9 below shows.
Lemma 2.9. Assume that ρ \rho ρ has compact support. If ρ \rho ρ satisfies assumptions (H0), then λ ρ ∈ C ∞ ( R n ) \lambda_{\rho} \in C^{\infty}(\mathbb{R}^{n}) λ ρ ∈ C ∞ ( R n ) has linear growth, and if it additionally fulfills (H1)–(H2) and (H3’), then
∣ ∂ α λ ρ ( ξ ) ∣ ≤ C α ∣ ξ ∣ 1 − ∣ α ∣ |\partial^{\alpha}\lambda_{\rho}(\xi)| \le C_{\alpha}|\xi|^{1-|\alpha|} ∣ ∂ α λ ρ ( ξ ) ∣ ≤ C α ∣ ξ ∣ 1 − ∣ α ∣ for all ξ ∈ R n \xi\in\mathbb{R}^{n} ξ ∈ R n and multi-indices α ∈ N 0 n \alpha\in\mathbb{N}_{0}^{n} α ∈ N 0 n , for some suitable constants C α > 0 C_{\alpha}>0 C α > 0 .
Proof. The property that λ ρ ∈ C ∞ ( R n ) \lambda_\rho\in C^\infty(\mathbb{R}^n) λ ρ ∈ C ∞ ( R n ) is a direct consequence of the fact that ρ \rho ρ has compact support. Let ξ ∈ B 1 c \xi\in B_1^c ξ ∈ B 1 c . According to [8 Lemma 2.4] it holds
∣ λ ρ ( ξ ) ∣ ≤ ∫ R n ρ ( x ) min { ∣ x ∣ − 1 , 2 π ∣ ξ ∣ } d x , |\lambda_\rho(\xi)| \le\int_{\mathbb{R}^n} \rho(x)\min\{|x|^{-1},2\pi|\xi|\}\,dx, ∣ λ ρ ( ξ ) ∣ ≤ ∫ R n ρ ( x ) min { ∣ x ∣ − 1 , 2 π ∣ ξ ∣ } d x ,
which, using that ρ ∈ L 1 ( R n ) \rho\in L^1(\mathbb{R}^n) ρ ∈ L 1 ( R n ) , can be further estimated by
∣ λ ρ ( ξ ) ∣ ≤ 2 π ∥ ρ ∥ L 1 ( R n ) ∣ ξ ∣ , ξ ∈ B 1 c . |\lambda_\rho(\xi)| \le2\pi\|\rho\|_{L^1(\mathbb{R}^n)}|\xi|,\qquad\xi\in B_1^c. ∣ λ ρ ( ξ ) ∣ ≤ 2 π ∥ ρ ∥ L 1 ( R n ) ∣ ξ ∣ , ξ ∈ B 1 c .
Exploiting also (H1)–(H2) and (H3’), it follows then immediately from Lemma 2.8 that ∣ ∂ α λ ρ ( ξ ) ∣ ≤ C ∣ ξ ∣ 1 − ∣ α ∣ |\partial^\alpha\lambda_\rho(\xi)| \le C|\xi|^{1-|\alpha|} ∣ ∂ α λ ρ ( ξ ) ∣ ≤ C ∣ ξ ∣ 1 − ∣ α ∣ for all multi-indices α ∈ N 0 n \alpha\in\mathbb{N}_0^n α ∈ N 0 n , and for some constants C α > 0 C_\alpha> 0 C α > 0 depending on ∥ ρ ∥ L 1 ( R n ) \|\rho\|_{L^1(\mathbb{R}^n)} ∥ ρ ∥ L 1 ( R n ) . We conclude by noting that the inequality extends to all ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n , since λ ρ ∈ C ∞ ( R n ) \lambda_\rho\in C^\infty(\mathbb{R}^n) λ ρ ∈ C ∞ ( R n ) .
Lemma 2.8 allows us to deduce the following Fourier multiplier result by means of the Mihlin-Hörmander multiplier theorem. We set from now on
⟦ λ ρ ⟧ : = ( 1 + ∣ λ ρ ∣ 2 ) 1 / 2 . \llbracket\lambda_\rho\rrbracket:= (1+|\lambda_\rho|^2)^{1/2}. [ [ λ ρ ] ] := ( 1 + ∣ λ ρ ∣ 2 ) 1/2 .
Proposition 2.10 (Central Fourier multipliers). Let p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) . Assume that ρ \rho ρ has compact support and satisfies assumptions (H0)–(H2) and (H3’). Then, the symbols
( i ) ∣ λ ρ ∣ ⟦ λ ρ ⟧ , ( i i ) ⟦ λ ρ ⟧ 1 + ∣ λ ρ ∣ , ( i i i ) 1 ⟦ λ ρ ⟧ , ( i v ) 1 ⟦ λ ρ ⟧ 2 (i)\quad\frac{|\lambda_\rho|}{\llbracket\lambda_\rho\rrbracket},\qquad
(ii)\quad\frac{\llbracket\lambda_\rho\rrbracket}{1+|\lambda_\rho|},\qquad
(iii)\quad\frac{1}{\llbracket\lambda_\rho\rrbracket},\qquad
(iv)\quad\frac{1}{\llbracket\lambda_\rho\rrbracket^2} ( i ) [ [ λ ρ ] ] ∣ λ ρ ∣ , ( ii ) 1 + ∣ λ ρ ∣ [ [ λ ρ ] ] , ( iii ) [ [ λ ρ ] ] 1 , ( i v ) [ [ λ ρ ] ] 2 1 are L p L^p L p -Fourier multipliers with corresponding L p L^p L p -Fourier multiplier operators
F ∣ λ ρ ∣ / ⟦ λ ρ ⟧ ‾ , F ⟦ λ ρ ⟧ / ( 1 + ∣ λ ρ ∣ ) ‾ , F 1 / ⟦ λ ρ ⟧ ‾ and F 1 / ⟦ λ ρ ⟧ 2 ‾ . \overline{F_{|\lambda_\rho|/\llbracket\lambda_\rho\rrbracket}},\quad
\overline{F_{\llbracket\lambda_\rho\rrbracket/(1+|\lambda_\rho|)}},\quad
\overline{F_{1/\llbracket\lambda_\rho\rrbracket}}
\quad\text{and}\quad
\overline{F_{1/\llbracket\lambda_\rho\rrbracket^2}}. F ∣ λ ρ ∣/ [ [ λ ρ ] ] , F [ [ λ ρ ] ] / ( 1 + ∣ λ ρ ∣ ) , F 1/ [ [ λ ρ ] ] and F 1/ [ [ λ ρ ] ] 2 .
Proof. (i). We define m ρ : R n → R m_\rho:\mathbb{R}^n \to\mathbb{R} m ρ : R n → R by
m ρ : = ∣ λ ρ ∣ ⟦ λ ρ ⟧ . m_\rho:= \frac{|\lambda_\rho|}{\llbracket\lambda_\rho\rrbracket}. m ρ := [ [ λ ρ ] ] ∣ λ ρ ∣ .
In order to apply the Mihlin-Hörmander multiplier theorem, we verify subsequently that
∣ ∂ α m ρ ( ξ ) ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ , ξ ∈ R n ∖ { 0 } (42) |\partial^\alpha m_\rho(\xi)| \le C_\alpha|\xi|^{-|\alpha|},\qquad\xi\in\mathbb{R}^n\setminus\{0\} \tag*{(42)} ∣ ∂ α m ρ ( ξ ) ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ , ξ ∈ R n ∖ { 0 } (42)
for all multi-indices α ∈ N 0 n \alpha\in\mathbb{N}_0^n α ∈ N 0 n . We note that by their smoothness, the functions ∂ α f i \partial^\alpha f_i ∂ α f i , i = 1 , 2 i=1,2 i = 1 , 2 , are well defined on B 1 ∖ { 0 } B_1\setminus\{0\} B 1 ∖ { 0 } , and it is readily seen that m ρ m_\rho m ρ and all its derivatives are bounded there. So, it will be sufficient to establish inequality (2.25) for ξ ∈ B 1 c \xi\in B_1^c ξ ∈ B 1 c . To that end we first note that, actually for all ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n ,
∣ m ρ ( ξ ) ∣ ≤ 1. (43) |m_\rho(\xi)|\le1. \tag*{(43)} ∣ m ρ ( ξ ) ∣ ≤ 1. (43)
Choose now some α ∈ N 0 n ∖ { 0 } \alpha\in\mathbb{N}_0^n\setminus\{0\} α ∈ N 0 n ∖ { 0 } . Then, with Leibniz’ rule, it holds
∣ ∂ α m ρ ( ξ ) ∣ ≤ C α ∑ β ≤ α ∣ ∂ β λ ρ ( ξ ) ∣ ∥ ∂ α − β ⟦ λ ρ ( ξ ) ⟧ ∥ , ξ ∈ B 1 c . (44) |\partial^\alpha m_\rho(\xi)|\le C_\alpha\sum_{\beta\le\alpha}|\partial^\beta\lambda_\rho(\xi)|\|\partial^{\alpha-\beta}\llbracket\lambda_\rho(\xi)\rrbracket\|,
\qquad\xi\in B_1^c. \tag*{(44)} ∣ ∂ α m ρ ( ξ ) ∣ ≤ C α β ≤ α ∑ ∣ ∂ β λ ρ ( ξ ) ∣∥ ∂ α − β [ [ λ ρ ( ξ ) ] ] ∥ , ξ ∈ B 1 c . (44)
Resorting to (2.21) and (2.24) in Lemma 2.8 , we conclude that the right-hand side is bounded by
C α ∑ β ≤ α ∣ λ ρ ( ξ ) ∣ ∣ ξ ∣ − ∣ β ∣ ∣ λ ρ ( ξ ) ∣ − 1 ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ = C α ∣ ξ ∣ − ∣ α ∣ , ξ ∈ B 1 c . (45) C_\alpha\sum_{\beta\le\alpha}|\lambda_\rho(\xi)||\xi|^{-|\beta|}|\lambda_\rho(\xi)|^{-1}|\xi|^{-|\alpha|+|\beta|}
=C_\alpha|\xi|^{-|\alpha|},\qquad\xi\in B_1^c. \tag*{(45)} C α β ≤ α ∑ ∣ λ ρ ( ξ ) ∣∣ ξ ∣ − ∣ β ∣ ∣ λ ρ ( ξ ) ∣ − 1 ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ = C α ∣ ξ ∣ − ∣ α ∣ , ξ ∈ B 1 c . (45)
Combining estimate (2.26) with (2.27) and (2.28), proves (2.25). Consequently, thanks to the Mihlin-Hörmander multiplier theorem, m m m is an L p L^p L p -Fourier multiplier, for any p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) .
(ii). The proof is obtained similarly to case (i). In fact, once more with Lemma 2.8 , now estimates (2.22) and (2.23), we can compute again
∑ β ≤ α ∣ ∂ β ⟦ λ ρ ⟧ ∣ ∣ ∂ α − β ( 1 + ∣ λ ρ ( ξ ) ∣ ) − 1 ∣ ≤ C α ∑ β ≤ α ∣ λ ρ ( ξ ) ∣ ∣ ξ ∣ − ∣ β ∣ ∣ λ ρ ( ξ ) ∣ − 1 ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ = C α ∣ ξ ∣ − ∣ α ∣ , ξ ∈ B 1 c . \sum_{\beta\le\alpha}|\partial^\beta\llbracket\lambda_\rho\rrbracket||\partial^{\alpha-\beta}(1+|\lambda_\rho(\xi)|)^{-1}|
\le C_\alpha\sum_{\beta\le\alpha}|\lambda_\rho(\xi)||\xi|^{-|\beta|}|\lambda_\rho(\xi)|^{-1}|\xi|^{-|\alpha|+|\beta|}
= C_\alpha|\xi|^{-|\alpha|},\qquad\xi\in B_1^c. β ≤ α ∑ ∣ ∂ β [ [ λ ρ ] ] ∣∣ ∂ α − β ( 1 + ∣ λ ρ ( ξ ) ∣ ) − 1 ∣ ≤ C α β ≤ α ∑ ∣ λ ρ ( ξ ) ∣∣ ξ ∣ − ∣ β ∣ ∣ λ ρ ( ξ ) ∣ − 1 ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ = C α ∣ ξ ∣ − ∣ α ∣ , ξ ∈ B 1 c .
(iii) and (iv). It follows from (2.18) that ∣ λ ρ ∣ − 1 |\lambda_\rho|^{-1} ∣ λ ρ ∣ − 1 is bounded, and we get from Lemma 2.8 immediately that
∣ ∂ α [ λ ρ ] − 1 ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ − 1 ∣ ξ ∣ − ∣ α ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ , |\partial^\alpha[\lambda_\rho]^{-1}| \le C_\alpha|\lambda_\rho(\xi)|^{-1}|\xi|^{-|\alpha|} \le C_\alpha|\xi|^{-|\alpha|}, ∣ ∂ α [ λ ρ ] − 1 ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ − 1 ∣ ξ ∣ − ∣ α ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ ,
∣ ∂ α [ λ ρ ] − 2 ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ − 2 ∣ ξ ∣ − ∣ α ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ , ξ ∈ B 1 c |\partial^\alpha[\lambda_\rho]^{-2}| \le C_\alpha|\lambda_\rho(\xi)|^{-2}|\xi|^{-|\alpha|} \le C_\alpha|\xi|^{-|\alpha|}, \qquad\xi\in B_1^c ∣ ∂ α [ λ ρ ] − 2 ∣ ≤ C α ∣ λ ρ ( ξ ) ∣ − 2 ∣ ξ ∣ − ∣ α ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ , ξ ∈ B 1 c
for all α ∈ N 0 n \alpha\in\mathbb{N}_0^n α ∈ N 0 n . Thus, one can again apply the Mihlin-Hörmander multiplier theorem in both cases. □
Our final result in this section collects four specific multiplier operators involving the symbol λ ρ \lambda_\rho λ ρ . Their closures will turn out to be central for the later analysis, since they correspond to the ρ \rho ρ -Laplacian and its square roots.
Lemma 2.11. Let p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) . Assume that ρ \rho ρ has compact support and satisfies ( H 0 ) (H0) ( H 0 ) –( H 2 ) (H2) ( H 2 ) and ( H 3 ′ ) (H3') ( H 3 ′ ) . Then the following holds:
(i) The unbounded operators F [ λ ρ ] \mathcal{F}_{[\lambda_\rho]} F [ λ ρ ] , F [ λ ρ ] 2 \mathcal{F}_{[\lambda_\rho]^2} F [ λ ρ ] 2 , F 1 / [ λ ρ ] 2 \mathcal{F}_{1/[\lambda_\rho]^2} F 1/ [ λ ρ ] 2 and F ∣ λ ρ ∣ 2 \mathcal{F}_{|\lambda_\rho|^2} F ∣ λ ρ ∣ 2 in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) map their domain S ( R n ) S(\mathbb{R}^n) S ( R n ) into itself and are closable.
(ii) The unbounded operators F ∣ λ ρ ∣ \mathcal{F}_{|\lambda_\rho|} F ∣ λ ρ ∣ and F 1 + ∣ λ ρ ∣ \mathcal{F}_{1+|\lambda_\rho|} F 1 + ∣ λ ρ ∣ in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) are closable with domains S ( R n ) S(\mathbb{R}^n) S ( R n ) .
Moreover, it holds
∥ F [ λ ρ ] ‾ u ∥ L p ( R n ) ≤ 1 2 ∥ u ∥ L p ( R n ) + 1 2 ∥ F [ λ ρ ] 2 ‾ u ∥ L p ( R n ) , u ∈ D ( F [ λ ρ ] ‾ ) ∩ D ( F [ λ ρ ] 2 ‾ ) , (46) \|\overline{\mathcal{F}_{[\lambda_\rho]}}u\|_{L^p(\mathbb{R}^n)} \le\frac{1}{2}\|u\|_{L^p(\mathbb{R}^n)}+\frac{1}{2}\|\overline{\mathcal{F}_{[\lambda_\rho]^2}}u\|_{L^p(\mathbb{R}^n)}, \qquad u \in D(\overline{\mathcal{F}_{[\lambda_\rho]}})\cap D(\overline{\mathcal{F}_{[\lambda_\rho]^2}}), \tag*{(46)} ∥ F [ λ ρ ] u ∥ L p ( R n ) ≤ 2 1 ∥ u ∥ L p ( R n ) + 2 1 ∥ F [ λ ρ ] 2 u ∥ L p ( R n ) , u ∈ D ( F [ λ ρ ] ) ∩ D ( F [ λ ρ ] 2 ) , (46) ∥ F ∣ λ ρ ∣ u ∥ L p ( R n ) ≤ 1 2 ∥ u ∥ L p ( R n ) + 1 2 ∥ F ∣ λ ρ ∣ 2 u ∥ L p ( R n ) , u ∈ S ( R n ) . (47) \|\mathcal{F}_{|\lambda_\rho|}u\|_{L^p(\mathbb{R}^n)} \le\frac{1}{2}\|u\|_{L^p(\mathbb{R}^n)}+\frac{1}{2}\|\mathcal{F}_{|\lambda_\rho|^2}u\|_{L^p(\mathbb{R}^n)}, \qquad u \in S(\mathbb{R}^n). \tag*{(47)} ∥ F ∣ λ ρ ∣ u ∥ L p ( R n ) ≤ 2 1 ∥ u ∥ L p ( R n ) + 2 1 ∥ F ∣ λ ρ ∣ 2 u ∥ L p ( R n ) , u ∈ S ( R n ) . (47)
Proof. Step 1: proof of (i) and (2.29). Part (i) is an immediate consequence of Lemma 2.4 in combination with Lemma 2.9. Furthermore, with the aim of proving (2.29), let u ∈ D ( F [ λ ρ ] ‾ ) ∩ D ( F [ λ ρ ] 2 ‾ ) u \in D(\overline{\mathcal{F}_{[\lambda_\rho]}})\cap D(\overline{\mathcal{F}_{[\lambda_\rho]^2}}) u ∈ D ( F [ λ ρ ] ) ∩ D ( F [ λ ρ ] 2 ) and choose a sequence, existing by definition of the closure, ( u j ) j ⊂ S ( R n ) (u_j)_j \subset S(\mathbb{R}^n) ( u j ) j ⊂ S ( R n ) such that u j → u u_j \to u u j → u and F [ λ ρ ] 2 u j → F [ λ ρ ] 2 ‾ u \mathcal{F}_{[\lambda_\rho]^2}u_j \to\overline{\mathcal{F}_{[\lambda_\rho]^2}}u F [ λ ρ ] 2 u j → F [ λ ρ ] 2 u . Since F 1 / [ λ ρ ] \mathcal{F}_{1/[\lambda_\rho]} F 1/ [ λ ρ ] is an L p L^p L p -Fourier multiplier operator according to Proposition 2.10, we infer with Lemma 2.6 that F [ λ ρ ] u j → F [ λ ρ ] ‾ u \mathcal{F}_{[\lambda_\rho]}u_j \to\overline{\mathcal{F}_{[\lambda_\rho]}}u F [ λ ρ ] u j → F [ λ ρ ] u . Then, as a consequence of Young’s inequality, one has
∣ ⟨ F [ λ ρ ] ‾ u , φ ⟩ ∣ = lim j → ∞ ∣ ⟨ F [ λ ρ ] u j , φ ⟩ ∣ = lim j → ∞ ∣ ⟨ [ λ ρ ] u j ^ , F − 1 φ ⟩ ∣ ≤ lim j → ∞ ( 1 2 ∣ ⟨ u j , F − 1 φ ⟩ ∣ + 1 2 ∣ ⟨ [ λ ρ ] 2 u j ^ , F − 1 φ ⟩ ∣ ) = 1 2 ∣ ⟨ u , φ ⟩ ∣ + 1 2 ∣ ⟨ F [ λ ρ ] 2 ‾ u , φ ⟩ ∣ , φ ∈ C c ∞ ( R n ) . \begin{aligned}
|\langle\overline{\mathcal{F}_{[\lambda_\rho]}}u,\varphi\rangle|=\lim_{j\to\infty}|\langle\mathcal{F}_{[\lambda_\rho]}u_j,\varphi\rangle|=\lim_{j\to\infty}|\langle[\lambda_\rho]\widehat{u_j},\mathcal{F}^{-1}\varphi\rangle| \\
&\le\lim_{j\to\infty}\left(\frac{1}{2}|\langle u_j,\mathcal{F}^{-1}\varphi\rangle|+\frac{1}{2}|\langle[\lambda_\rho]^2\widehat{u_j},\mathcal{F}^{-1}\varphi\rangle|\right)=\frac{1}{2}|\langle u,\varphi\rangle|+\frac{1}{2}|\langle\overline{\mathcal{F}_{[\lambda_\rho]^2}}u,\varphi\rangle|,\qquad\varphi\in C^\infty_c(\mathbb{R}^n).
\end{aligned} ∣ ⟨ F [ λ ρ ] u , φ ⟩ ∣ = j → ∞ lim ∣ ⟨ F [ λ ρ ] u j , φ ⟩ ∣ = j → ∞ lim ∣ ⟨[ λ ρ ] u j , F − 1 φ ⟩ ∣ ≤ j → ∞ lim ( 2 1 ∣ ⟨ u j , F − 1 φ ⟩ ∣ + 2 1 ∣ ⟨[ λ ρ ] 2 u j , F − 1 φ ⟩ ∣ ) = 2 1 ∣ ⟨ u , φ ⟩ ∣ + 2 1 ∣ ⟨ F [ λ ρ ] 2 u , φ ⟩ ∣ , φ ∈ C c ∞ ( R n ) .
Taking the supremum over all φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) such that ∥ φ ∥ L p ′ ( R n ) ≤ 1 \|\varphi\|_{L^{p'}(\mathbb{R}^n)}\le1 ∥ φ ∥ L p ′ ( R n ) ≤ 1 , we obtain estimate (2.29).
Step 2: proof of (ii) and (2.30). We observe upfront that
D ( F ∣ λ ρ ∣ ) = D ( F 1 + ∣ λ ρ ∣ ) and F 1 + ∣ λ ρ ∣ u = u + F ∣ λ ρ ∣ u , u ∈ D ( F ∣ λ ρ ∣ ) , (48) D(\mathcal{F}_{|\lambda_\rho|})=D(\mathcal{F}_{1+|\lambda_\rho|}) \quad\text{and} \quad\mathcal{F}_{1+|\lambda_\rho|}u=u+\mathcal{F}_{|\lambda_\rho|}u,\qquad u\in D(\mathcal{F}_{|\lambda_\rho|}), \tag*{(48)} D ( F ∣ λ ρ ∣ ) = D ( F 1 + ∣ λ ρ ∣ ) and F 1 + ∣ λ ρ ∣ u = u + F ∣ λ ρ ∣ u , u ∈ D ( F ∣ λ ρ ∣ ) , (48)
following the linearity of the Fourier transform. To establish the claim that F ∣ λ ρ ∣ \mathcal{F}_{|\lambda_\rho|} F ∣ λ ρ ∣ and F 1 + ∣ λ ρ ∣ \mathcal{F}_{1+|\lambda_\rho|} F 1 + ∣ λ ρ ∣ have domain equal to S ( R n ) S(\mathbb{R}^n) S ( R n ) , it thus suffices to show that D ( F ∣ λ ρ ∣ ) = S ( R n ) D(\mathcal{F}_{|\lambda_\rho|})=S(\mathbb{R}^n) D ( F ∣ λ ρ ∣ ) = S ( R n ) , which reduces to proving that S ( R n ) ⊂ D ( F ∣ λ ρ ∣ ) S(\mathbb{R}^n)\subset D(\mathcal{F}_{|\lambda_\rho|}) S ( R n ) ⊂ D ( F ∣ λ ρ ∣ ) , since the inverse inclusion holds by definition.
To that end, we pick u ∈ S ( R n ) u\in S(\mathbb{R}^n) u ∈ S ( R n ) and point out that the product ∣ λ ρ ∣ u ^ |\lambda_\rho|\widehat{u} ∣ λ ρ ∣ u lies in L 2 ( R n ) L^2(\mathbb{R}^n) L 2 ( R n ) , since λ ρ \lambda_\rho λ ρ has polynomial growth thanks to Lemma 2.9, and, hence, the Fourier inverse is well defined and F − 1 ( ∣ λ ρ ∣ u ^ ) ∈ L 2 ( R n ) \mathcal{F}^{-1}(|\lambda_\rho|\widehat{u})\in L^2(\mathbb{R}^n) F − 1 ( ∣ λ ρ ∣ u ) ∈ L 2 ( R n ) . At the same time, it holds, similar to Step 1, with Young’s inequality and part (i) that
∣ ⟨ F − 1 ( ∣ λ ρ ∣ u ^ ) , φ ⟩ ∣ = ∣ ⟨ ∣ λ ρ ∣ u ^ , F − 1 φ ⟩ ∣ ≤ 1 2 ∣ ⟨ u , φ ⟩ ∣ + 1 2 ∣ ⟨ F ∣ λ ρ ∣ 2 u , φ ⟩ ∣ , φ ∈ C c ∞ ( R n ) , |\langle\mathcal{F}^{-1}(|\lambda_\rho|\widehat{u}),\varphi\rangle|=|\langle|\lambda_\rho|\widehat{u},\mathcal{F}^{-1}\varphi\rangle|\le\frac{1}{2}|\langle u,\varphi\rangle|+\frac{1}{2}|\langle\mathcal{F}_{|\lambda_\rho|^2}u,\varphi\rangle|,\qquad\varphi\in C^\infty_c(\mathbb{R}^n), ∣ ⟨ F − 1 ( ∣ λ ρ ∣ u ) , φ ⟩ ∣ = ∣ ⟨ ∣ λ ρ ∣ u , F − 1 φ ⟩ ∣ ≤ 2 1 ∣ ⟨ u , φ ⟩ ∣ + 2 1 ∣ ⟨ F ∣ λ ρ ∣ 2 u , φ ⟩ ∣ , φ ∈ C c ∞ ( R n ) ,
and taking again the supremum over all φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) such that ∥ φ ∥ L p ′ ( R n ) ≤ 1 \|\varphi\|_{L^{p'}(\mathbb{R}^n)}\le1 ∥ φ ∥ L p ′ ( R n ) ≤ 1 , yields
∥ F − 1 ( ∣ λ ρ ∣ u ^ ) ∥ L p ( R n ) ≤ 1 2 ∥ u ∥ L p ( R n ) + 1 2 ∥ F ∣ λ ρ ∣ 2 u ∥ L p ( R n ) . (49) \|\mathcal{F}^{-1}(|\lambda_\rho|\widehat{u})\|_{L^p(\mathbb{R}^n)}\le\frac{1}{2}\|u\|_{L^p(\mathbb{R}^n)}+\frac{1}{2}\|\mathcal{F}_{|\lambda_\rho|^2}u\|_{L^p(\mathbb{R}^n)}. \tag*{(49)} ∥ F − 1 ( ∣ λ ρ ∣ u ) ∥ L p ( R n ) ≤ 2 1 ∥ u ∥ L p ( R n ) + 2 1 ∥ F ∣ λ ρ ∣ 2 u ∥ L p ( R n ) . (49)
This shows that F − 1 ( ∣ λ ρ ∣ u ^ ) ∈ L p ( R n ) \mathcal{F}^{-1}(|\lambda_\rho|\widehat{u})\in L^p(\mathbb{R}^n) F − 1 ( ∣ λ ρ ∣ u ) ∈ L p ( R n ) , and hence, u ∈ D ( F ∣ λ ρ ∣ ) u\in D(\mathcal{F}_{|\lambda_\rho|}) u ∈ D ( F ∣ λ ρ ∣ ) . Consequently, D ( F ∣ λ ρ ∣ ) = D ( F 1 + ∣ λ ρ ∣ ) = S ( R n ) D(\mathcal{F}_{|\lambda_\rho|})=D(\mathcal{F}_{1+|\lambda_\rho|})=S(\mathbb{R}^n) D ( F ∣ λ ρ ∣ ) = D ( F 1 + ∣ λ ρ ∣ ) = S ( R n ) , and (2.30) is a direct consequence of (2.32).
It remains to prove that F ∣ λ ρ ∣ \mathcal{F}_{\lvert\lambda_{\rho}\rvert} F ∣ λ ρ ∣ and F 1 + ∣ λ ρ ∣ \mathcal{F}_{1+\lvert\lambda_{\rho}\rvert} F 1 + ∣ λ ρ ∣ are both closable. With (2.31) in mind, choosing now a sequence ( u j ) j ⊂ D ( F ∣ λ ρ ∣ ) = S ( R n ) (u_j)_j \subset D(\mathcal{F}_{\lvert\lambda_{\rho}\rvert})=\mathcal{S}(\mathbb{R}^n) ( u j ) j ⊂ D ( F ∣ λ ρ ∣ ) = S ( R n ) such that u j → 0 u_j \to0 u j → 0 in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , there exists v ∈ L p ( R n ) v\in L^p(\mathbb{R}^n) v ∈ L p ( R n ) such that F ∣ λ ρ ∣ u j \mathcal{F}_{\lvert\lambda_{\rho}\rvert}u_j F ∣ λ ρ ∣ u j and F 1 + ∣ λ ρ ∣ u j → v \mathcal{F}_{1+\lvert\lambda_{\rho}\rvert}u_j \to v F 1 + ∣ λ ρ ∣ u j → v in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Proposition 2.10 (ii) implies the continuity of the L p L^p L p -Fourier multiplier operator F ∣ λ ρ ∣ ( 1 + ∣ λ ρ ∣ ) \frac{\mathcal{F}_{\lvert\lambda_{\rho}\rvert}}{(1+\lvert\lambda_{\rho}\rvert)} ( 1 + ∣ λ ρ ∣) F ∣ λ ρ ∣ , and so we find with Lemma 2.5 that
F ∣ λ ρ ∣ u j = F ∣ λ ρ ∣ ( 1 + ∣ λ ρ ∣ ) F 1 + ∣ λ ρ ∣ u j → F ∣ λ ρ ∣ ( 1 + ∣ λ ρ ∣ ) v in L p ( R n ) . \mathcal{F}_{\lvert\lambda_{\rho}\rvert}u_j=\frac{\mathcal{F}_{\lvert\lambda_{\rho}\rvert}}{(1+\lvert\lambda_{\rho}\rvert)}\mathcal{F}_{1+\lvert\lambda_{\rho}\rvert}u_j\to\frac{\mathcal{F}_{\lvert\lambda_{\rho}\rvert}}{(1+\lvert\lambda_{\rho}\rvert)}v \quad\text{in }L^p(\mathbb{R}^n). F ∣ λ ρ ∣ u j = ( 1 + ∣ λ ρ ∣) F ∣ λ ρ ∣ F 1 + ∣ λ ρ ∣ u j → ( 1 + ∣ λ ρ ∣) F ∣ λ ρ ∣ v in L p ( R n ) .
As the operator F ∣ λ ρ ∣ \mathcal{F}_{\lvert\lambda_{\rho}\rvert} F ∣ λ ρ ∣ is closable in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) according to part (i), we conclude that the limit above is zero, thus, F ∣ λ ρ ∣ u j → 0 \mathcal{F}_{\lvert\lambda_{\rho}\rvert}u_j\to0 F ∣ λ ρ ∣ u j → 0 in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Finally, applying Proposition 2.10 (i) and again Lemma 2.5 yields
F ∣ λ ρ ∣ u j = F ∣ λ ρ ∣ [ λ ρ ] F [ λ ρ ] u j → 0 in L p ( R n ) . \mathcal{F}_{\lvert\lambda_{\rho}\rvert}u_j=\frac{\mathcal{F}_{\lvert\lambda_{\rho}\rvert}}{\lbrack\lambda_{\rho}\rbrack}\mathcal{F}_{\lbrack\lambda_{\rho}\rbrack}u_j\to0 \quad\text{in }L^p(\mathbb{R}^n). F ∣ λ ρ ∣ u j = [ λ ρ ] F ∣ λ ρ ∣ F [ λ ρ ] u j → 0 in L p ( R n ) .
Hence, v = 0 v=0 v = 0 and, consequently, F ∣ λ ρ ∣ \mathcal{F}_{\lvert\lambda_{\rho}\rvert} F ∣ λ ρ ∣ is closable. Similarly, because F 1 + ∣ λ ρ ∣ u j → 0 \mathcal{F}_{1+\lvert\lambda_{\rho}\rvert}u_j\to0 F 1 + ∣ λ ρ ∣ u j → 0 in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) as well, we have also proved that F 1 + ∣ λ ρ ∣ \mathcal{F}_{1+\lvert\lambda_{\rho}\rvert} F 1 + ∣ λ ρ ∣ is closable. □
Fourier characterization
In this section, we establish the characterization of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) via Fourier transform, and then conclude a similar result for H 0 ρ , p ( Ω ) H^{\rho,p}_0(\Omega) H 0 ρ , p ( Ω ) with Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n open. Finally, the dual of H 0 ρ , p ( Ω ) H^{\rho,p}_0(\Omega) H 0 ρ , p ( Ω ) is also characterized.
Fourier characterization of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n )
In the following, we will devise a representation of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) in terms of a Fourier characterization over the space S ′ ( R n ) \mathcal{S}'(\mathbb{R}^n) S ′ ( R n ) . In order for this characterization to make sense, we rely, as outlined in Section 2.3, on the fact that [ λ ρ ] \lbrack\lambda_{\rho}\rbrack [ λ ρ ] lies in C ∞ ( R n ) C^\infty(\mathbb{R}^n) C ∞ ( R n ) and all its derivatives have polynomial growth thanks to Lemma 2.9.
Theorem 3.1 (Fourier characterization). Let p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and assume that ρ \rho ρ has compact support and fulfills hypotheses (H0)–(H2) and (H3’). It holds
H ρ , p ( R n ) = { u ∈ S ′ ( R n ) : F − 1 ( [ λ ρ ] u ^ ) ∈ L p ( R n ) } = { u ∈ L p ( R n ) : F − 1 ( [ λ ρ ] u ^ ) ∈ L p ( R n ) } (50) H^{\rho,p}(\mathbb{R}^n)=\{u\in\mathcal{S}'(\mathbb{R}^n):\mathcal{F}^{-1}(\lbrack\lambda_{\rho}\rbrack\hat{u})\in L^p(\mathbb{R}^n)\}
=\{u\in L^p(\mathbb{R}^n):\mathcal{F}^{-1}(\lbrack\lambda_{\rho}\rbrack\hat{u})\in L^p(\mathbb{R}^n)\} \tag*{(50)} H ρ , p ( R n ) = { u ∈ S ′ ( R n ) : F − 1 ([ λ ρ ] u ^ ) ∈ L p ( R n )} = { u ∈ L p ( R n ) : F − 1 ([ λ ρ ] u ^ ) ∈ L p ( R n )} (50) with
c p ∥ F − 1 ( [ λ ρ ] u ^ ) ∥ L p ( R n ) ≤ ∥ u ∥ H ρ , p ( R n ) ≤ C p ∥ F − 1 ( [ λ ρ ] u ^ ) ∥ L p ( R n ) (51) c_p\|\mathcal{F}^{-1}(\lbrack\lambda_{\rho}\rbrack\hat{u})\|_{L^p(\mathbb{R}^n)}
\leq\|u\|_{H^{\rho,p}(\mathbb{R}^n)}
\leq C_p\|\mathcal{F}^{-1}(\lbrack\lambda_{\rho}\rbrack\hat{u})\|_{L^p(\mathbb{R}^n)} \tag*{(51)} c p ∥ F − 1 ([ λ ρ ] u ^ ) ∥ L p ( R n ) ≤ ∥ u ∥ H ρ , p ( R n ) ≤ C p ∥ F − 1 ([ λ ρ ] u ^ ) ∥ L p ( R n ) (51) for all u ∈ H ρ , p ( R n ) u\in H^{\rho,p}(\mathbb{R}^n) u ∈ H ρ , p ( R n ) and some constants c p , C p > 0 c_p,C_p>0 c p , C p > 0 .
Moreover,
D ρ u = F − 1 ( λ ρ u ^ ) , u ∈ H ρ , p ( R n ) , (52) D^\rho u=\mathcal{F}^{-1}(\lambda_\rho\hat{u}),\qquad u\in H^{\rho,p}(\mathbb{R}^n), \tag*{(52)} D ρ u = F − 1 ( λ ρ u ^ ) , u ∈ H ρ , p ( R n ) , (52) where the right-hand side is understood in the sense of tempered distributions.
Proof. We define
X ρ , p ( R n ) : = { u ∈ S ′ ( R n ) : F − 1 ( [ λ ρ ] u ^ ) ∈ L p ( R n ) } . X^{\rho,p}(\mathbb{R}^n):=\{u\in\mathcal{S}'(\mathbb{R}^n):\mathcal{F}^{-1}(\lbrack\lambda_{\rho}\rbrack\hat{u})\in L^p(\mathbb{R}^n)\}. X ρ , p ( R n ) := { u ∈ S ′ ( R n ) : F − 1 ([ λ ρ ] u ^ ) ∈ L p ( R n )} .
We point out that the inclusions L p ( R n ) ⊂ S ′ ( R n ) L^p(\mathbb{R}^n)\subset\mathcal{S}'(\mathbb{R}^n) L p ( R n ) ⊂ S ′ ( R n ) and H ρ , p ( R n ) ⊂ S ′ ( R n ) H^{\rho,p}(\mathbb{R}^n)\subset\mathcal{S}'(\mathbb{R}^n) H ρ , p ( R n ) ⊂ S ′ ( R n ) are fulfilled automatically. Moreover, as soon as we know that H ρ , p ( R n ) = X ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n)=X^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) = X ρ , p ( R n ) , then the second equality in (3.1) follows directly from H ρ , p ( R n ) ⊂ L p ( R n ) H^{\rho,p}(\mathbb{R}^n)\subset L^p(\mathbb{R}^n) H ρ , p ( R n ) ⊂ L p ( R n ) . Henceforth, we will use C > 0 C>0 C > 0 to denote a generic positive constant that may change throughout the proof.
First, let us show that H ρ , p ( R n ) ⊂ X ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n)\subset X^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) ⊂ X ρ , p ( R n ) continuously, i.e., that the symbol [ λ ρ ] \lbrack\lambda_{\rho}\rbrack [ λ ρ ] defines a bounded operator from H ρ , p ( R n ) ⊂ L p ( R n ) H^{\rho,p}(\mathbb{R}^n)\subset L^p(\mathbb{R}^n) H ρ , p ( R n ) ⊂ L p ( R n ) into L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . We pick some arbitrary u ∈ C c ∞ ( R n ) ⊂ H ρ , p ( R n ) u\in C^\infty_c(\mathbb{R}^n)\subset H^{\rho,p}(\mathbb{R}^n) u ∈ C c ∞ ( R n ) ⊂ H ρ , p ( R n ) and observe that F − 1 ( ( 1 + ∣ λ ρ ∣ ) u ^ ) \mathcal{F}^{-1}((1+\lvert\lambda_{\rho}\rvert)\hat{u}) F − 1 (( 1 + ∣ λ ρ ∣) u ^ ) is well defined as an element in S ′ ( R n ) \mathcal{S}'(\mathbb{R}^n) S ′ ( R n ) due to the polynomial growth of λ ρ \lambda_\rho λ ρ (see Lemma 2.9). Our goal is to demonstrate that F − 1 ( ( 1 + ∣ λ ρ ∣ ) u ^ ) ∈ L p ( R n ) \mathcal{F}^{-1}((1+\lvert\lambda_{\rho}\rvert)\hat{u})\in L^p(\mathbb{R}^n) F − 1 (( 1 + ∣ λ ρ ∣) u ^ ) ∈ L p ( R n ) because Proposition 2.10 (ii) then entails
∥ F − 1 ( [ λ ρ ] u ^ ) ∥ L p ( R n ) ≤ C ∥ F − 1 ( ( 1 + ∣ λ ρ ∣ ) u ^ ) ∥ L p ( R n ) . (53) \|\mathcal{F}^{-1}(\lbrack\lambda_{\rho}\rbrack\hat{u})\|_{L^p(\mathbb{R}^n)}
\leq C\|\mathcal{F}^{-1}((1+\lvert\lambda_{\rho}\rvert)\hat{u})\|_{L^p(\mathbb{R}^n)}. \tag*{(53)} ∥ F − 1 ([ λ ρ ] u ^ ) ∥ L p ( R n ) ≤ C ∥ F − 1 (( 1 + ∣ λ ρ ∣) u ^ ) ∥ L p ( R n ) . (53)
By definition, we obtain for any φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n )
⟨ F − 1 ( ( 1 + ∣ λ ρ ∣ ) u ^ ) , φ ⟩ S ′ × S = ⟨ u , F ( ( 1 + ∣ λ ρ ∣ ) F − 1 φ ) ⟩ = ⟨ u ^ , F − 1 φ ⟩ + ⟨ u ^ , ∣ λ ρ ∣ F − 1 φ ⟩ . \langle\mathcal{F}^{-1}((1+\lvert\lambda_{\rho}\rvert)\hat{u}),\varphi\rangle_{\mathcal{S}'\times\mathcal{S}}
=\langle u,\mathcal{F}((1+\lvert\lambda_{\rho}\rvert)\mathcal{F}^{-1}\varphi)\rangle
=\langle\hat{u},\mathcal{F}^{-1}\varphi\rangle+\langle\hat{u},\lvert\lambda_{\rho}\rvert\mathcal{F}^{-1}\varphi\rangle. ⟨ F − 1 (( 1 + ∣ λ ρ ∣) u ^ ) , φ ⟩ S ′ × S = ⟨ u , F (( 1 + ∣ λ ρ ∣) F − 1 φ )⟩ = ⟨ u ^ , F − 1 φ ⟩ + ⟨ u ^ , ∣ λ ρ ∣ F − 1 φ ⟩ .
The first contribution on the right-hand side is bounded with the help of Hölder’s inequality. In order to estimate the remaining summand on the right-hand side, we observe that that, using the identity F ( D 0 ρ u ) = λ ρ u ^ = 2 π i Q ρ u ^ \mathcal{F}(D_{0}^{\rho}u)=\lambda_{\rho}\hat{u}=2\pi iQ_{\rho}\hat{u} F ( D 0 ρ u ) = λ ρ u ^ = 2 π i Q ρ u ^ (see (2.10)), it can be rewritten as
⟨ − i ξ 2 π ∣ ξ ∣ 2 Q ρ ( ξ ) ⋅ F ( D 0 ρ u ) , ∣ λ ρ ∣ F − 1 φ ⟩ = ⟨ ∑ j = 1 n − i ξ j ∣ ξ ∣ ( F ( D 0 ρ u ) ) j , F − 1 φ ⟩ = ⟨ F ( R D 0 ρ u ) , F − 1 φ ⟩ , \left\langle\frac{-i\xi}{2\pi|\xi|^{2}Q_{\rho}(\xi)}\cdot\mathcal{F}(D_{0}^{\rho}u),|\lambda_{\rho}|\mathcal{F}^{-1}\varphi\right\rangle=\left\langle\sum_{j=1}^{n}\frac{-i\xi_{j}}{|\xi|}(\mathcal{F}(D_{0}^{\rho}u))_{j},\mathcal{F}^{-1}\varphi\right\rangle=\left\langle\mathcal{F}(RD_{0}^{\rho}u),\mathcal{F}^{-1}\varphi\right\rangle, ⟨ 2 π ∣ ξ ∣ 2 Q ρ ( ξ ) − i ξ ⋅ F ( D 0 ρ u ) , ∣ λ ρ ∣ F − 1 φ ⟩ = ⟨ j = 1 ∑ n ∣ ξ ∣ − i ξ j ( F ( D 0 ρ u ) ) j , F − 1 φ ⟩ = ⟨ F ( R D 0 ρ u ) , F − 1 φ ⟩ ,
where
R U : = ∑ j = 1 n R j U j , U ∈ L p ( R n ; R n ) (54) RU:=\sum_{j=1}^{n}R_{j}U_{j},\qquad U\in L^{p}(\mathbb{R}^{n};\mathbb{R}^{n}) \tag*{(54)} R U := j = 1 ∑ n R j U j , U ∈ L p ( R n ; R n ) (54)
for R j R_{j} R j the j j j -th Riesz transforms with the Fourier symbol ξ ↦ − i ξ j / ∣ ξ ∣ \xi\mapsto-i\xi_{j}/|\xi| ξ ↦ − i ξ j /∣ ξ ∣ . Importantly, the Riesz transforms are bounded operators on L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) for p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and j ∈ { 1 , … , n } j\in\{1,\ldots,n\} j ∈ { 1 , … , n } (cf. [17 Corollary 5.2.8] ). Since ρ \rho ρ has compact support, we know that D 0 ρ u ∈ C c ∞ ( R n ) D_{0}^{\rho}u\in C_{c}^{\infty}(\mathbb{R}^{n}) D 0 ρ u ∈ C c ∞ ( R n ) and, in particular, R D 0 ρ u ∈ L 2 ( R n ) RD_{0}^{\rho}u\in L^{2}(\mathbb{R}^{n}) R D 0 ρ u ∈ L 2 ( R n ) . Thus, Hölder’s inequality yields
∣ ⟨ F ( R D 0 ρ u ) , F − 1 φ ⟩ ∣ = ∣ ⟨ R D 0 ρ u , φ ⟩ ∣ ≤ ∥ R D 0 ρ u ∥ L p ( R n ) ∥ φ ∥ L p ′ ( R n ) ≤ C ∥ D 0 ρ u ∥ L p ( R n ; R n ) ∥ φ ∥ L p ′ ( R n ) . |\langle\mathcal{F}(RD_{0}^{\rho}u),\mathcal{F}^{-1}\varphi\rangle|=|\langle RD_{0}^{\rho}u,\varphi\rangle|\leq\|RD_{0}^{\rho}u\|_{L^{p}(\mathbb{R}^{n})}\|\varphi\|_{L^{p'}(\mathbb{R}^{n})}\leq C\|D_{0}^{\rho}u\|_{L^{p}(\mathbb{R}^{n};\mathbb{R}^{n})}\|\varphi\|_{L^{p'}(\mathbb{R}^{n})}. ∣ ⟨ F ( R D 0 ρ u ) , F − 1 φ ⟩ ∣ = ∣ ⟨ R D 0 ρ u , φ ⟩ ∣ ≤ ∥ R D 0 ρ u ∥ L p ( R n ) ∥ φ ∥ L p ′ ( R n ) ≤ C ∥ D 0 ρ u ∥ L p ( R n ; R n ) ∥ φ ∥ L p ′ ( R n ) .
Taking now the supremum of the left-hand side over all φ ∈ C c ∞ ( R n ) \varphi\in C_{c}^{\infty}(\mathbb{R}^{n}) φ ∈ C c ∞ ( R n ) with ∥ φ ∥ L p ′ ( R n ) ≤ 1 \|\varphi\|_{L^{p'}(\mathbb{R}^{n})}\leq1 ∥ φ ∥ L p ′ ( R n ) ≤ 1 , we arrive at
∥ F − 1 ( ( 1 + ∣ λ ρ ∣ ) u ^ ) ∥ L p ( R n ) ≤ C ( ∥ u ∥ L p ( R n ) + ∥ D 0 ρ u ∥ L p ( R n ; R n ) ) , \|\mathcal{F}^{-1}((1+|\lambda_{\rho}|)\hat{u})\|_{L^{p}(\mathbb{R}^{n})}\leq C(\|u\|_{L^{p}(\mathbb{R}^{n})}+\|D_{0}^{\rho}u\|_{L^{p}(\mathbb{R}^{n};\mathbb{R}^{n})}), ∥ F − 1 (( 1 + ∣ λ ρ ∣ ) u ^ ) ∥ L p ( R n ) ≤ C ( ∥ u ∥ L p ( R n ) + ∥ D 0 ρ u ∥ L p ( R n ; R n ) ) ,
where F − 1 ( ( 1 + ∣ λ ρ ∣ ) u ^ ) \mathcal{F}^{-1}((1+|\lambda_{\rho}|)\hat{u}) F − 1 (( 1 + ∣ λ ρ ∣ ) u ^ ) has a representative in L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) due to Riesz’ representation theorem (see, e.g., [10 Theorem 4.11] ). In view of (3.4), this proves the claim for functions in C c ∞ ( R n ) C_{c}^{\infty}(\mathbb{R}^{n}) C c ∞ ( R n ) . For arbitrary u ∈ H ρ , p ( R n ) u\in H^{\rho,p}(\mathbb{R}^{n}) u ∈ H ρ , p ( R n ) , the statement can be deduced by the density of C c ∞ ( R n ) C_{c}^{\infty}(\mathbb{R}^{n}) C c ∞ ( R n ) in H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^{n}) H ρ , p ( R n ) (see [8 Theorem 3.9] ).
For the other inclusion, X ρ , p ( R n ) ⊂ H ρ , p ( R n ) X^{\rho,p}(\mathbb{R}^{n})\subset H^{\rho,p}(\mathbb{R}^{n}) X ρ , p ( R n ) ⊂ H ρ , p ( R n ) , let now u ∈ X ρ , p ( R n ) u\in X^{\rho,p}(\mathbb{R}^{n}) u ∈ X ρ , p ( R n ) . We demonstrate first that u u u can be regarded as an element in L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) , arguing by means of Proposition 2.10 (iii). For any φ ∈ S ( R n ) \varphi\in\mathcal{S}(\mathbb{R}^{n}) φ ∈ S ( R n ) , we define the auxiliary function
f 1 : = F − 1 ( [ λ ρ ] − 1 φ ^ ) , f_{1}:=\mathcal{F}^{-1}([\lambda_{\rho}]^{-1}\hat{\varphi}), f 1 := F − 1 ([ λ ρ ] − 1 φ ^ ) ,
and observe that, by virtue of the second identity of Lemma 2.1 (i) for the even multiplier [ λ ρ ] − 1 [\lambda_{\rho}]^{-1} [ λ ρ ] − 1 , it holds [ λ ρ ] − 1 F − 1 φ = F − 1 f 1 [\lambda_{\rho}]^{-1}\mathcal{F}^{-1}\varphi=\mathcal{F}^{-1}f_{1} [ λ ρ ] − 1 F − 1 φ = F − 1 f 1 . Then we compute
⟨ u , φ ⟩ S ′ × S = ⟨ u , F ( F − 1 φ ) ⟩ S ′ × S = ⟨ u , F ( [ λ ρ ] ( [ λ ρ ] − 1 F − 1 φ ) ) ⟩ S ′ × S = ⟨ u , F ( [ λ ρ ] F − 1 f 1 ) ⟩ S ′ × S = ⟨ F − 1 ( [ λ ρ ] u ^ ) , f 1 ⟩ . \begin{aligned}
\langle u,\varphi\rangle_{\mathcal{S}'\times\mathcal{S}}&=\langle u,\mathcal{F}(\mathcal{F}^{-1}\varphi)\rangle_{\mathcal{S}'\times\mathcal{S}}\\
&=\langle u,\mathcal{F}([\lambda_{\rho}]([\lambda_{\rho}]^{-1}\mathcal{F}^{-1}\varphi))\rangle_{\mathcal{S}'\times\mathcal{S}}\\
&=\langle u,\mathcal{F}([\lambda_{\rho}]\mathcal{F}^{-1}f_{1})\rangle_{\mathcal{S}'\times\mathcal{S}}\\
&=\langle\mathcal{F}^{-1}([\lambda_{\rho}]\hat{u}),f_{1}\rangle.
\end{aligned} ⟨ u , φ ⟩ S ′ × S = ⟨ u , F ( F − 1 φ ) ⟩ S ′ × S = ⟨ u , F ([ λ ρ ] ([ λ ρ ] − 1 F − 1 φ )) ⟩ S ′ × S = ⟨ u , F ([ λ ρ ] F − 1 f 1 ) ⟩ S ′ × S = ⟨ F − 1 ([ λ ρ ] u ^ ) , f 1 ⟩ .
Combining the assumption F − 1 ( [ λ ρ ] u ^ ) ∈ L p ( R n ) \mathcal{F}^{-1}([\lambda_{\rho}]\hat{u})\in L^{p}(\mathbb{R}^{n}) F − 1 ([ λ ρ ] u ^ ) ∈ L p ( R n ) with Hölder’s inequality and Proposition 2.10 (iii), we conclude that
∣ ⟨ u , φ ⟩ S ′ × S ∣ ≤ ∥ F − 1 ( [ λ ρ ] u ^ ) ∥ L p ( R n ) ∥ f 1 ∥ L p ′ ( R n ) ≤ C ∥ F − 1 ( [ λ ρ ] u ^ ) ∥ L p ( R n ) ∥ φ ∥ L p ′ ( R n ) . (55) |\langle u,\varphi\rangle_{\mathcal{S}'\times\mathcal{S}}|\leq\|\mathcal{F}^{-1}([\lambda_{\rho}]\hat{u})\|_{L^{p}(\mathbb{R}^{n})}\|f_{1}\|_{L^{p'}(\mathbb{R}^{n})}\leq C\|\mathcal{F}^{-1}([\lambda_{\rho}]\hat{u})\|_{L^{p}(\mathbb{R}^{n})}\|\varphi\|_{L^{p'}(\mathbb{R}^{n})}. \tag*{(55)} ∣ ⟨ u , φ ⟩ S ′ × S ∣ ≤ ∥ F − 1 ([ λ ρ ] u ^ ) ∥ L p ( R n ) ∥ f 1 ∥ L p ′ ( R n ) ≤ C ∥ F − 1 ([ λ ρ ] u ^ ) ∥ L p ( R n ) ∥ φ ∥ L p ′ ( R n ) . (55)
Choose now some φ ∈ L p ′ ( R n ) \varphi\in L^{p'}(\mathbb{R}^{n}) φ ∈ L p ′ ( R n ) and ( φ k ) k ⊂ C c ∞ ( R n ) (\varphi_{k})_{k}\subset C_{c}^{\infty}(\mathbb{R}^{n}) ( φ k ) k ⊂ C c ∞ ( R n ) such that φ k → φ \varphi_{k}\to\varphi φ k → φ in L p ′ ( R n ) L^{p'}(\mathbb{R}^{n}) L p ′ ( R n ) . It follows by linearity from (3.6) that ( ⟨ u , φ k ⟩ S ′ × S ) k (\langle u,\varphi_{k}\rangle_{\mathcal{S}'\times\mathcal{S}})_{k} (⟨ u , φ k ⟩ S ′ × S ) k constitutes a Cauchy sequence in R \mathbb{R} R , which attains a well defined limit. In particular, setting ⟨ u , φ ⟩ : = lim k → ∞ ⟨ u , φ k ⟩ S ′ × S \langle u,\varphi\rangle:=\lim_{k\to\infty}\langle u,\varphi_{k}\rangle_{\mathcal{S}'\times\mathcal{S}} ⟨ u , φ ⟩ := lim k → ∞ ⟨ u , φ k ⟩ S ′ × S , the expression ⟨ u , φ ⟩ \langle u,\varphi\rangle ⟨ u , φ ⟩ is independent of the choice of the approximating sequence. At the same time, we obtain along with (3.6) that
∣ ⟨ u , φ ⟩ S ′ × S ∣ ≤ C ∥ F − 1 ( [ λ ρ ] u ^ ) ∥ L p ( R n ) ∥ φ ∥ L p ′ ( R n ) . |\langle u,\varphi\rangle_{\mathcal{S}'\times\mathcal{S}}|\leq C\|\mathcal{F}^{-1}([\lambda_{\rho}]\hat{u})\|_{L^{p}(\mathbb{R}^{n})}\|\varphi\|_{L^{p'}(\mathbb{R}^{n})}. ∣ ⟨ u , φ ⟩ S ′ × S ∣ ≤ C ∥ F − 1 ([ λ ρ ] u ^ ) ∥ L p ( R n ) ∥ φ ∥ L p ′ ( R n ) .
The Riesz representation theorem yields that u u u indeed coincides with an element in L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) . It remains to prove that D ρ u ∈ L p ( R n ; R n ) D^{\rho}u\in L^{p}(\mathbb{R}^{n};\mathbb{R}^{n}) D ρ u ∈ L p ( R n ; R n ) . Precisely, we claim that F − 1 ( λ ρ u ^ ) \mathcal{F}^{-1}(\lambda_{\rho}\hat{u}) F − 1 ( λ ρ u ^ ) is well defined as an element in L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) and that D ρ u = F − 1 ( λ ρ u ^ ) D^{\rho}u=\mathcal{F}^{-1}(\lambda_{\rho}\hat{u}) D ρ u = F − 1 ( λ ρ u ^ ) . Indeed, similar to the steps taken above to establish that u u u coincides with an L p L^p L p -function, one can also show that F − 1 ( λ ρ u ^ ) ∈ L p ( R n ) \mathcal{F}^{-1}(\lambda_\rho\hat{u}) \in L^p(\mathbb{R}^n) F − 1 ( λ ρ u ^ ) ∈ L p ( R n ) . To see this, we note that, for any Φ ∈ C c ∞ ( R n ; R n ) \Phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n) Φ ∈ C c ∞ ( R n ; R n ) and recalling (54) , the function
f 2 : = − F − 1 ( ∣ λ ρ ∣ [ λ ρ ] F ( R Φ ) ) f_2 := -\mathcal{F}^{-1}\left(\frac{|\lambda_\rho|}{[\lambda_\rho]}\mathcal{F}(\mathcal{R}\Phi)\right) f 2 := − F − 1 ( [ λ ρ ] ∣ λ ρ ∣ F ( R Φ ) )
is an element of L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) with
∥ f 2 ∥ L p ( R n ) ≤ C ∥ Φ ∥ L p ( R n ; R n ) , \|f_2\|_{L^p(\mathbb{R}^n)} \le C\|\Phi\|_{L^p(\mathbb{R}^n;\mathbb{R}^n)}, ∥ f 2 ∥ L p ( R n ) ≤ C ∥Φ ∥ L p ( R n ; R n ) ,
by virtue of Proposition 2.10 ( i ) (i) ( i ) and the boundedness of the Riesz transforms. Then it holds, with the second identity of Lemma 2.1 ( i i ) (ii) ( ii ) applied componentwise,
⟨ F − 1 ( λ ρ u ^ ) , Φ ⟩ S ′ × S = ⟨ u , F ( λ ρ ⋅ F − 1 Φ ) ⟩ S ′ × S = ⟨ u , F ( [ λ ρ ] i ξ ∣ ξ ∣ ∣ λ ρ ∣ [ λ ρ ] ⋅ F − 1 Φ ) ⟩ S ′ × S = ⟨ u , F ( [ λ ρ ] F − 1 ( f 2 ) ) ⟩ S ′ × S = ⟨ F − 1 ( [ λ ρ ] u ^ ) , f 2 ⟩ . \begin{aligned}
\langle\mathcal{F}^{-1}(\lambda_\rho\hat{u}),\Phi\rangle_{\mathcal{S}'\times\mathcal{S}} &= \langle u,\mathcal{F}(\lambda_\rho\cdot\mathcal{F}^{-1}\Phi)\rangle_{\mathcal{S}'\times\mathcal{S}}\\
&= \left\langle u,\mathcal{F}\left([\lambda_\rho]\frac{i\xi}{|\xi|}\frac{|\lambda_\rho|}{[\lambda_\rho]}\cdot\mathcal{F}^{-1}\Phi\right)\right\rangle_{\mathcal{S}'\times\mathcal{S}}\\
&= \langle u,\mathcal{F}([\lambda_\rho]\mathcal{F}^{-1}(f_2))\rangle_{\mathcal{S}'\times\mathcal{S}}\\
&= \langle\mathcal{F}^{-1}([\lambda_\rho]\hat{u}),f_2\rangle.
\end{aligned} ⟨ F − 1 ( λ ρ u ^ ) , Φ ⟩ S ′ × S = ⟨ u , F ( λ ρ ⋅ F − 1 Φ ) ⟩ S ′ × S = ⟨ u , F ( [ λ ρ ] ∣ ξ ∣ i ξ [ λ ρ ] ∣ λ ρ ∣ ⋅ F − 1 Φ ) ⟩ S ′ × S = ⟨ u , F ([ λ ρ ] F − 1 ( f 2 )) ⟩ S ′ × S = ⟨ F − 1 ([ λ ρ ] u ^ ) , f 2 ⟩ .
Following, mutatis mutandis, the computations in (55) and the steps thereafter, and applying Riesz’ representation theorem, it follows as claimed that F − 1 ( λ ρ u ^ ) \mathcal{F}^{-1}(\lambda_\rho\hat{u}) F − 1 ( λ ρ u ^ ) has a representative in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Finally, we verify that D ρ u = F − 1 ( λ ρ u ^ ) D_\rho u = \mathcal{F}^{-1}(\lambda_\rho\hat{u}) D ρ u = F − 1 ( λ ρ u ^ ) . We recall that since ρ \rho ρ has compact support, it holds div 0 ρ Φ ∈ C c ∞ ( R n ) \operatorname{div}_0^\rho\Phi\in C^\infty_c(\mathbb{R}^n) div 0 ρ Φ ∈ C c ∞ ( R n ) for any Φ ∈ C c ∞ ( R n ; R n ) \Phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n) Φ ∈ C c ∞ ( R n ; R n ) and the integral
∫ R n u ( x ) div 0 ρ Φ ( x ) d x \int_{\mathbb{R}^n} u(x)\operatorname{div}_0^\rho\Phi(x)\,\mathrm{d}x ∫ R n u ( x ) div 0 ρ Φ ( x ) d x
is well defined. With (27) we find
− ⟨ u , div 0 ρ Φ ⟩ = ⟨ u ^ , F − 1 ( − div 0 ρ Φ ) ⟩ S ′ × S = ⟨ u ^ , λ ρ ⋅ F − 1 Φ ⟩ S ′ × S = ⟨ F − 1 ( λ ρ u ^ ) , Φ ⟩ -\langle u,\operatorname{div}_0^\rho\Phi\rangle= \langle\hat{u},\mathcal{F}^{-1}(-\operatorname{div}_0^\rho\Phi)\rangle_{\mathcal{S}'\times\mathcal{S}} = \langle\hat{u},\lambda_\rho\cdot\mathcal{F}^{-1}\Phi\rangle_{\mathcal{S}'\times\mathcal{S}} = \langle\mathcal{F}^{-1}(\lambda_\rho\hat{u}),\Phi\rangle − ⟨ u , div 0 ρ Φ ⟩ = ⟨ u ^ , F − 1 ( − div 0 ρ Φ ) ⟩ S ′ × S = ⟨ u ^ , λ ρ ⋅ F − 1 Φ ⟩ S ′ × S = ⟨ F − 1 ( λ ρ u ^ ) , Φ ⟩
for all Φ ∈ C c ∞ ( R n ; R n ) \Phi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n) Φ ∈ C c ∞ ( R n ; R n ) . By definition, this establishes the sought identity D ρ u = F − 1 ( λ ρ u ^ ) D_\rho u = \mathcal{F}^{-1}(\lambda_\rho\hat{u}) D ρ u = F − 1 ( λ ρ u ^ ) , which is exactly (52) . □
Remark 3.2 (Hilbert space case). In the case p = 2 p=2 p = 2 , by Plancherel’s identity, the alternative characterization
H ρ ( R n ) = { u ∈ S ′ ( R n ) : [ λ ρ ] u ^ ∈ L 2 ( R n ) } H^\rho(\mathbb{R}^n)=\{u\in\mathcal{S}'(\mathbb{R}^n):[\lambda_\rho]\hat{u}\in L^2(\mathbb{R}^n)\} H ρ ( R n ) = { u ∈ S ′ ( R n ) : [ λ ρ ] u ^ ∈ L 2 ( R n )}
holds true (see also [7 Proposition 4.7] for a recent related result). Moreover, the statement of Proposition 2.10 can be easily proved in this case, since the space of L 2 L^2 L 2 -Fourier multipliers coincides with L ∞ L^\infty L ∞ (see [17 Theorem 2.5.10] ), noting that the assumptions on ρ \rho ρ can be weakened, as done, for example, in [8 Section 4.2 and Theorem 7.2 (ii)] .
Remark 3.3 (Kernels without compact support). We recall from Section 2.5 that for any kernel ρ \rho ρ satisfying conditions (H0)–(H2) and (H3’), the space H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) is equally described as H ρ ~ , p ( R n ) H^{\tilde{\rho},p}(\mathbb{R}^n) H ρ ~ , p ( R n ) for some suitable kernel ρ ~ \tilde{\rho} ρ ~ of compact support still fulfilling conditions (H0)–(H2) and (H3’). Then, Theorem 3.1 entails that
c ∥ F − 1 ( [ λ ρ ~ ] u ^ ) ∥ L p ( R n ) ≤ ∥ u ∥ H ρ , p ( R n ) ≤ C ∥ F − 1 ( [ λ ρ ~ ] u ^ ) ∥ L p ( R n ) c\|\mathcal{F}^{-1}([\lambda_{\tilde{\rho}}]\hat{u})\|_{L^p(\mathbb{R}^n)} \le\|u\|_{H^{\rho,p}(\mathbb{R}^n)} \le C\|\mathcal{F}^{-1}([\lambda_{\tilde{\rho}}]\hat{u})\|_{L^p(\mathbb{R}^n)} c ∥ F − 1 ([ λ ρ ~ ] u ^ ) ∥ L p ( R n ) ≤ ∥ u ∥ H ρ , p ( R n ) ≤ C ∥ F − 1 ([ λ ρ ~ ] u ^ ) ∥ L p ( R n )
for all u ∈ H ρ , p ( R n ) u\in H^{\rho,p}(\mathbb{R}^n) u ∈ H ρ , p ( R n ) and some constants c , C > 0 c,C>0 c , C > 0 .
Fourier characterization of H 0 ρ , p ( Ω ) H^{\rho,p}_0(\Omega) H 0 ρ , p ( Ω )
Let for the rest of the section Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be open. Recalling (22) , the space H 0 ρ , p ( Ω ) H^{\rho,p}_0(\Omega) H 0 ρ , p ( Ω ) is defined as the set of u ∈ H ρ , p ( R n ) u\in H^{\rho,p}(\mathbb{R}^n) u ∈ H ρ , p ( R n ) such that u = 0 u=0 u = 0 a.e. in Ω c \Omega^c Ω c . If Ω \Omega Ω has a Lipschitz boundary, this set can be described as the closure of C c ∞ ( Ω ) C^\infty_c(\Omega) C c ∞ ( Ω ) in H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) ; see [8 Theorem 3.9] . The Fourier characterization of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) given by Theorem 3.1, translates directly to the space H 0 ρ , p ( Ω ) H^{\rho,p}_0(\Omega) H 0 ρ , p ( Ω ) .
Corollary 3.4 (Fourier characterization for complementary-value spaces). Let p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) , let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be open such that ∣ ∂ Ω ∣ = 0 |\partial\Omega| = 0 ∣ ∂ Ω∣ = 0 and assume that ρ \rho ρ has compact support and satisfies hypotheses (H0)–(H2) and (H3′). Then
H 0 ρ , p ( Ω ) = { u ∈ S ′ ( R n ) : supp u ⊂ Ω ‾ , F − 1 ( [ λ ρ ] u ^ ) ∈ L p ( R n ) } H_0^{\rho,p}(\Omega) = \{u \in\mathcal{S}'(\mathbb{R}^n) : \operatorname{supp} u \subset\overline{\Omega},\ \mathcal{F}^{-1}([\lambda_\rho]\widehat{u}) \in L^p(\mathbb{R}^n)\} H 0 ρ , p ( Ω ) = { u ∈ S ′ ( R n ) : supp u ⊂ Ω , F − 1 ([ λ ρ ] u ) ∈ L p ( R n )} and, for all u ∈ H 0 ρ , p ( Ω ) u \in H_0^{\rho,p}(\Omega) u ∈ H 0 ρ , p ( Ω ) , inequalities (3.2) hold.
Proof. As ∣ ∂ Ω ∣ = 0 |\partial\Omega| = 0 ∣ ∂ Ω∣ = 0 , the subspace H 0 ρ , p ( Ω ) H_0^{\rho,p}(\Omega) H 0 ρ , p ( Ω ) can be equivalently described as the set of u ∈ H ρ , p ( R n ) u \in H^{\rho,p}(\mathbb{R}^n) u ∈ H ρ , p ( R n ) such that supp u ⊂ Ω ‾ \operatorname{supp} u \subset\overline{\Omega} supp u ⊂ Ω . Since, for a locally integrable function, its support as a function coincides with its support as a distribution (see, e.g., [32 Sections I.1 and I.3] ), we conclude by Theorem 3.1 the desired characterization. □
Remark 3.5. Instead of [8 Definition 3.4] , an equally natural definition of H 0 ρ , p ( Ω ) H_0^{\rho,p}(\Omega) H 0 ρ , p ( Ω ) would have been the set of u ∈ H ρ , p ( R n ) u \in H^{\rho,p}(\mathbb{R}^n) u ∈ H ρ , p ( R n ) such that u = 0 u = 0 u = 0 a.e. in Ω ‾ c \overline{\Omega}^{\,c} Ω c ; in other words, the set of u ∈ H ρ , p ( R n ) u \in H^{\rho,p}(\mathbb{R}^n) u ∈ H ρ , p ( R n ) such that supp u ⊂ Ω ‾ \operatorname{supp} u \subset\overline{\Omega} supp u ⊂ Ω . In that case, the analogue of the characterization given in Corollary 3.4 holds without the assumption ∣ ∂ Ω ∣ = 0 |\partial\Omega| = 0 ∣ ∂ Ω∣ = 0 .
Fourier characterization of the dual space
For p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) we denote by p ′ ∈ ( 1 , ∞ ) p' \in(1,\infty) p ′ ∈ ( 1 , ∞ ) its dual exponent fulfilling 1 / p + 1 / p ′ = 1 1/p + 1/p' = 1 1/ p + 1/ p ′ = 1 , and define the topological dual space of H 0 ρ , p ( Ω ) H_0^{\rho,p}(\Omega) H 0 ρ , p ( Ω ) by
H − ρ , p ′ ( Ω ) : = ( H 0 ρ , p ( Ω ) ) ′ . H^{-\rho,p'}(\Omega) := (H_0^{\rho,p}(\Omega))'. H − ρ , p ′ ( Ω ) := ( H 0 ρ , p ( Ω ) ) ′ .
As C c ∞ ( Ω ) C^\infty_c(\Omega) C c ∞ ( Ω ) is dense in H 0 ρ , p ( Ω ) H_0^{\rho,p}(\Omega) H 0 ρ , p ( Ω ) due to [8 Theorem 3.9] (under assumption (H0)), we have H − ρ , p ′ ( Ω ) ⊂ D ′ ( Ω ) H^{-\rho,p'}(\Omega) \subset\mathcal{D}'(\Omega) H − ρ , p ′ ( Ω ) ⊂ D ′ ( Ω ) . We point out that, if Ω \Omega Ω is bounded, then H − ρ , p ′ ( Ω ) H^{-\rho,p'}(\Omega) H − ρ , p ′ ( Ω ) is in general not a subset of S ′ ( R n ) \mathcal{S}'(\mathbb{R}^n) S ′ ( R n ) since H − ρ , p ′ ( Ω ) H^{-\rho,p'}(\Omega) H − ρ , p ′ ( Ω ) only restricts the behavior of the functionals acting on functions defined over Ω \Omega Ω .
The embedding H 0 ρ , p ( Ω ) ↪ L p ( Ω ) H_0^{\rho,p}(\Omega) \hookrightarrow L^p(\Omega) H 0 ρ , p ( Ω ) ↪ L p ( Ω ) is dense, hence the dual embedding L p ′ ( Ω ) ↪ H − ρ , p ′ ( Ω ) L^{p'}(\Omega) \hookrightarrow H^{-\rho,p'}(\Omega) L p ′ ( Ω ) ↪ H − ρ , p ′ ( Ω ) holds. Moreover, if Ω \Omega Ω is bounded, by [8 Theorem 4.11] (provided that (H0)–(H2) and (H3′) are fulfilled), the embedding H 0 ρ , p ( Ω ) ↪ L p ( Ω ) H_0^{\rho,p}(\Omega) \hookrightarrow L^p(\Omega) H 0 ρ , p ( Ω ) ↪ L p ( Ω ) is compact, and, consequently, so is the embedding L p ′ ( Ω ) ↪ H − ρ , p ′ ( Ω ) L^{p'}(\Omega) \hookrightarrow H^{-\rho,p'}(\Omega) L p ′ ( Ω ) ↪ H − ρ , p ′ ( Ω ) (see, e.g., [10 Theorem 6.4] ).
The Fourier characterization of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) given in Theorem 3.1 has a natural counterpart for its dual.
Theorem 3.6 (Dual Fourier characterization). Let p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) , and assume that ρ \rho ρ has compact support and fulfills hypotheses (H0)–(H2) and (H3′). Then
H − ρ , p ′ ( R n ) = { u ∈ S ′ ( R n ) : F − 1 ( [ λ ρ ] − 1 u ^ ) ∈ L p ′ ( R n ) } H^{-\rho,p'}(\mathbb{R}^n) = \{u \in\mathcal{S}'(\mathbb{R}^n) : \mathcal{F}^{-1}([\lambda_\rho]^{-1}\widehat{u}) \in L^{p'}(\mathbb{R}^n)\} H − ρ , p ′ ( R n ) = { u ∈ S ′ ( R n ) : F − 1 ([ λ ρ ] − 1 u ) ∈ L p ′ ( R n )} and, for all u ∈ H − ρ , p ′ ( R n ) u \in H^{-\rho,p'}(\mathbb{R}^n) u ∈ H − ρ , p ′ ( R n ) ,
C p − 1 ∥ F − 1 ( [ λ ρ ] − 1 u ^ ) ∥ L p ′ ( R n ) ≤ ∥ u ∥ H − ρ , p ′ ( R n ) ≤ c p − 1 ∥ F − 1 ( [ λ ρ ] − 1 u ^ ) ∥ L p ′ ( R n ) C_p^{-1}\lVert\mathcal{F}^{-1}([\lambda_\rho]^{-1}\widehat{u})\rVert_{L^{p'}(\mathbb{R}^n)} \le\lVert u\rVert_{H^{-\rho,p'}(\mathbb{R}^n)} \le c_p^{-1}\lVert\mathcal{F}^{-1}([\lambda_\rho]^{-1}\widehat{u})\rVert_{L^{p'}(\mathbb{R}^n)} C p − 1 ∥ F − 1 ([ λ ρ ] − 1 u ) ∥ L p ′ ( R n ) ≤ ∥ u ∥ H − ρ , p ′ ( R n ) ≤ c p − 1 ∥ F − 1 ([ λ ρ ] − 1 u ) ∥ L p ′ ( R n ) where c p c_p c p , C p > 0 C_p > 0 C p > 0 are the constants from Theorem 3.1 .
We will deduce this statement as a special case of the following general result, which characterizes the duals of spaces defined via suitable multiplication in Fourier space. The proof relies on standard techniques; nevertheless, since we were unable to locate this precise statement in the literature, we include it here for the sake of completeness.
Henceforth, let m ∈ C ∞ ( R n ) m \in C^\infty(\mathbb{R}^n) m ∈ C ∞ ( R n ) be positive, even, and such that all its derivatives have polynomial growth. Further assume that also m − 1 = 1 / m m^{-1} = 1/m m − 1 = 1/ m has polynomial growth. Set
H m p ( R n ) : = { u ∈ S ′ ( R n ) : F − 1 ( m u ^ ) ∈ L p ( R n ) } H_m^p(\mathbb{R}^n) := \{u \in\mathcal{S}'(\mathbb{R}^n) : \mathcal{F}^{-1}(m\widehat{u}) \in L^p(\mathbb{R}^n)\} H m p ( R n ) := { u ∈ S ′ ( R n ) : F − 1 ( m u ) ∈ L p ( R n )}
and assume that ∥ ⋅ ∥ H m p ( R n ) \lVert\cdot\rVert_{H_m^p(\mathbb{R}^n)} ∥ ⋅ ∥ H m p ( R n ) is a norm equivalent to ∥ F − 1 ( m F ( ⋅ ) ) ∥ L p ( R n ) \lVert\mathcal{F}^{-1}(m\mathcal{F}(\cdot))\rVert_{L^p(\mathbb{R}^n)} ∥ F − 1 ( m F ( ⋅ )) ∥ L p ( R n ) , i.e.,
c H m p ∥ F − 1 ( m u ^ ) ∥ L p ( R n ) ≤ ∥ u ∥ H m p ( R n ) ≤ C H m p ∥ F − 1 ( m u ^ ) ∥ L p ( R n ) (56) c_{H_m^p}\lVert\mathcal{F}^{-1}(m\widehat{u})\rVert_{L^p(\mathbb{R}^n)} \le\lVert u\rVert_{H_m^p(\mathbb{R}^n)} \le C_{H_m^p}\lVert\mathcal{F}^{-1}(m\widehat{u})\rVert_{L^p(\mathbb{R}^n)} \tag*{(56)} c H m p ∥ F − 1 ( m u ) ∥ L p ( R n ) ≤ ∥ u ∥ H m p ( R n ) ≤ C H m p ∥ F − 1 ( m u ) ∥ L p ( R n ) (56)
for some c H m p , C H m p > 0 c_{H_m^p}, C_{H_m^p} > 0 c H m p , C H m p > 0 . We point out that H m p ( R n ) H_m^p(\mathbb{R}^n) H m p ( R n ) is well defined following the assumptions on m m m . Finally, let ( H m p ( R n ) ) ′ (H_m^p(\mathbb{R}^n))' ( H m p ( R n ) ) ′ be the topological dual space of H m p ( R n ) H_m^p(\mathbb{R}^n) H m p ( R n ) .
Lemma 3.7. Let p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) . If C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) is dense in H m p ( R n ) H^p_m(\mathbb{R}^n) H m p ( R n ) , then
( H m p ( R n ) ) ′ = { u ∈ S ′ ( R n ) : F − 1 ( m − 1 u ^ ) ∈ L p ′ ( R n ) } , (H^p_m(\mathbb{R}^n))'=\{u\in\mathcal{S}'(\mathbb{R}^n):\mathcal{F}^{-1}(m^{-1}\widehat{u})\in L^{p'}(\mathbb{R}^n)\}, ( H m p ( R n ) ) ′ = { u ∈ S ′ ( R n ) : F − 1 ( m − 1 u ) ∈ L p ′ ( R n )} , and, for all u ∈ H m p ( R n ) u\in H^p_m(\mathbb{R}^n) u ∈ H m p ( R n ) ,
C H m p − 1 ∥ F − 1 ( m − 1 u ^ ) ∥ L p ′ ( R n ) ≤ ∥ u ∥ ( H m p ( R n ) ) ′ ≤ c H m p − 1 ∥ F − 1 ( m − 1 u ^ ) ∥ L p ′ ( R n ) . C_{H^p_m}^{-1}\lVert\mathcal{F}^{-1}(m^{-1}\widehat{u})\rVert_{L^{p'}(\mathbb{R}^n)}
\leq\lVert u\rVert_{(H^p_m(\mathbb{R}^n))'}
\leq c_{H^p_m}^{-1}\lVert\mathcal{F}^{-1}(m^{-1}\widehat{u})\rVert_{L^{p'}(\mathbb{R}^n)}. C H m p − 1 ∥ F − 1 ( m − 1 u ) ∥ L p ′ ( R n ) ≤ ∥ u ∥ ( H m p ( R n ) ) ′ ≤ c H m p − 1 ∥ F − 1 ( m − 1 u ) ∥ L p ′ ( R n ) . Proof. We define
X ′ : = { u ∈ S ′ ( R n ) : F − 1 ( m − 1 u ^ ) ∈ L p ′ ( R n ) } X':=\{u\in\mathcal{S}'(\mathbb{R}^n):\mathcal{F}^{-1}(m^{-1}\widehat{u})\in L^{p'}(\mathbb{R}^n)\} X ′ := { u ∈ S ′ ( R n ) : F − 1 ( m − 1 u ) ∈ L p ′ ( R n )} and prove first that ( H m p ( R n ) ) ′ ⊂ X ′ (H^p_m(\mathbb{R}^n))'\subset X' ( H m p ( R n ) ) ′ ⊂ X ′ . Applying the definition of the Fourier transform and its inverse for tempered distributions, it follows with the first identity of Lemma 2.1 ( i ) (i) ( i ) that for any u ∈ ( H m p ( R n ) ) ′ u\in(H^p_m(\mathbb{R}^n))' u ∈ ( H m p ( R n ) ) ′ and any φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) ,
⟨ F − 1 ( m − 1 u ^ ) , φ ⟩ S ′ × S = ⟨ u , F ( m − 1 F − 1 φ ) ⟩ = ⟨ u , F − 1 ( m − 1 φ ^ ) ⟩ . (57) \langle\mathcal{F}^{-1}(m^{-1}\widehat{u}),\varphi\rangle_{\mathcal{S}'\times\mathcal{S}}
=\langle u,\mathcal{F}(m^{-1}\mathcal{F}^{-1}\varphi)\rangle
=\langle u,\mathcal{F}^{-1}(m^{-1}\widehat{\varphi})\rangle. \tag*{(57)} ⟨ F − 1 ( m − 1 u ) , φ ⟩ S ′ × S = ⟨ u , F ( m − 1 F − 1 φ )⟩ = ⟨ u , F − 1 ( m − 1 φ )⟩ . (57) Moreover, additionally assuming that ∥ φ ∥ L p ≤ 1 \lVert\varphi\rVert_{L^p}\leq1 ∥ φ ∥ L p ≤ 1 , we have thanks to (3.7)
∥ F − 1 ( m − 1 φ ^ ) ∥ H m p ( R n ) ≤ C H m p ∥ F − 1 ( m ( m − 1 φ ^ ) ) ∥ L p ( R n ) ≤ C H m p \lVert\mathcal{F}^{-1}(m^{-1}\widehat{\varphi})\rVert_{H^p_m(\mathbb{R}^n)}
\leq C_{H^p_m}\lVert\mathcal{F}^{-1}(m(m^{-1}\widehat{\varphi}))\rVert_{L^p(\mathbb{R}^n)}
\leq C_{H^p_m} ∥ F − 1 ( m − 1 φ ) ∥ H m p ( R n ) ≤ C H m p ∥ F − 1 ( m ( m − 1 φ )) ∥ L p ( R n ) ≤ C H m p and, hence,
∣ ⟨ u , F − 1 ( m − 1 φ ^ ) ⟩ ∣ ≤ C H m p ∥ u ∥ ( H m p ( R n ) ) ′ . |\langle u,\mathcal{F}^{-1}(m^{-1}\widehat{\varphi})\rangle|
\leq C_{H^p_m}\lVert u\rVert_{(H^p_m(\mathbb{R}^n))'}. ∣ ⟨ u , F − 1 ( m − 1 φ )⟩ ∣ ≤ C H m p ∥ u ∥ ( H m p ( R n ) ) ′ . Taking the supremum over all φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) with ∥ φ ∥ L p ≤ 1 \lVert\varphi\rVert_{L^p}\leq1 ∥ φ ∥ L p ≤ 1 , it follows from (3.8), together with the Riesz representation theorem, that
C H m p − 1 ∥ F − 1 ( m u ^ ) ∥ L p ′ ( R n ) ≤ ∥ u ∥ ( H m p ( R n ) ) ′ , C_{H^p_m}^{-1}\lVert\mathcal{F}^{-1}(m\widehat{u})\rVert_{L^{p'}(\mathbb{R}^n)}
\leq\lVert u\rVert_{(H^p_m(\mathbb{R}^n))'}, C H m p − 1 ∥ F − 1 ( m u ) ∥ L p ′ ( R n ) ≤ ∥ u ∥ ( H m p ( R n ) ) ′ , and we obtain that ( H m p ( R n ) ) ′ ⊂ X ′ (H^p_m(\mathbb{R}^n))'\subset X' ( H m p ( R n ) ) ′ ⊂ X ′ .
For the reverse inclusion X ′ ⊂ ( H m p ( R n ) ) ′ X'\subset(H^p_m(\mathbb{R}^n))' X ′ ⊂ ( H m p ( R n ) ) ′ , we pick u ∈ X ′ u\in X' u ∈ X ′ and again φ ∈ C c ∞ ( R n ) ⊂ H m p ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n)\subset H^p_m(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) ⊂ H m p ( R n ) , where the inclusion holds by assumption. By virtue of (3.7), the function
f : = F − 1 ( m φ ^ ) f:=\mathcal{F}^{-1}(m\widehat{\varphi}) f := F − 1 ( m φ ) is an element of S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) by the assumptions on m m m and
∥ f ∥ L p ( R n ) ≤ c H m p − 1 ∥ φ ∥ H m p ( R n ) . (58) \lVert f\rVert_{L^p(\mathbb{R}^n)}
\leq c_{H^p_m}^{-1}\lVert\varphi\rVert_{H^p_m(\mathbb{R}^n)}. \tag*{(58)} ∥ f ∥ L p ( R n ) ≤ c H m p − 1 ∥ φ ∥ H m p ( R n ) . (58) Now, again with the first identity of Lemma 2.1 ( i ) (i) ( i ) ,
⟨ u , φ ⟩ = ⟨ u , F − 1 ( m − 1 F ( f ) ) ⟩ = ⟨ u , F ( m − 1 F − 1 ( f ) ) ⟩ = ⟨ F − 1 ( m − 1 u ^ ) , f ⟩ . \langle u,\varphi\rangle
=\langle u,\mathcal{F}^{-1}(m^{-1}\mathcal{F}(f))\rangle
=\langle u,\mathcal{F}(m^{-1}\mathcal{F}^{-1}(f))\rangle
=\langle\mathcal{F}^{-1}(m^{-1}\widehat{u}),f\rangle. ⟨ u , φ ⟩ = ⟨ u , F − 1 ( m − 1 F ( f ))⟩ = ⟨ u , F ( m − 1 F − 1 ( f ))⟩ = ⟨ F − 1 ( m − 1 u ) , f ⟩ . If φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) is chosen such that ∥ φ ∥ H m p ( R n ) ≤ 1 \lVert\varphi\rVert_{H^p_m(\mathbb{R}^n)}\leq1 ∥ φ ∥ H m p ( R n ) ≤ 1 , it follows from the above computations, Hölder’s inequality and the norm estimate in (3.9) that
∣ ⟨ u , φ ⟩ ∣ ≤ c H m p − 1 ∥ F − 1 ( m − 1 u ^ ) ∥ L p ′ ( R n ) . |\langle u,\varphi\rangle|
\leq c_{H^p_m}^{-1}\lVert\mathcal{F}^{-1}(m^{-1}\widehat{u})\rVert_{L^{p'}(\mathbb{R}^n)}. ∣ ⟨ u , φ ⟩ ∣ ≤ c H m p − 1 ∥ F − 1 ( m − 1 u ) ∥ L p ′ ( R n ) . Taking in the left-hand side the supremum over all φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) , and using the density of C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) in H m p ( R n ) H^p_m(\mathbb{R}^n) H m p ( R n ) , yields
∥ u ∥ ( H m p ( R n ) ) ′ ≤ c H m p − 1 ∥ F − 1 ( m − 1 u ^ ) ∥ L p ′ ( R n ) . \lVert u\rVert_{(H^p_m(\mathbb{R}^n))'}
\leq c_{H^p_m}^{-1}\lVert\mathcal{F}^{-1}(m^{-1}\widehat{u})\rVert_{L^{p'}(\mathbb{R}^n)}. ∥ u ∥ ( H m p ( R n ) ) ′ ≤ c H m p − 1 ∥ F − 1 ( m − 1 u ) ∥ L p ′ ( R n ) . We conclude that ( H m p ( R n ) ) ′ = X ′ (H^p_m(\mathbb{R}^n))'=X' ( H m p ( R n ) ) ′ = X ′ , as claimed. □
The proof of Theorem 3.6 is now obtained as a simple corollary.
Proof of Theorem 3.6. The result follows from Lemma 3.7 with m : = ⌈ λ ρ ⌉ m:=\lceil\lambda_\rho\rceil m := ⌈ λ ρ ⌉ so that H m p ( R n ) = H ρ , p ( R n ) H^p_m(\mathbb{R}^n)=H^{\rho,p}(\mathbb{R}^n) H m p ( R n ) = H ρ , p ( R n ) . We note that the validity of the estimates in (3.7) follows from Theorem 3.1, while the fact that C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) is dense in H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) was shown in [8 Theorem 3.9] and only relies on hypothesis (H0). □
Although not of Fourier-type, we end this section with an alternative characterization of the dual through derivatives of L p L^p L p -functions.
Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be open. Given u 0 ∈ L p ( Ω ) u_0 \in L^p(\Omega) u 0 ∈ L p ( Ω ) and U ∈ L p ( Ω ; R n ) U \in L^p(\Omega;\mathbb{R}^n) U ∈ L p ( Ω ; R n ) , then u 0 + div ρ U ∈ H − ρ , p ( Ω ) u_0+\operatorname{div}^{\rho} U \in H^{-\rho,p}(\Omega) u 0 + div ρ U ∈ H − ρ , p ( Ω ) , since for all φ ∈ H ρ , p ′ ( Ω ) \varphi\in H^{\rho,p'}(\Omega) φ ∈ H ρ , p ′ ( Ω ) ,
⟨ u 0 + div ρ U , φ ⟩ = ∫ Ω ( u 0 φ − U ⋅ D ρ φ ) d x ≤ ∥ u 0 ∥ L p ( Ω ) ∥ φ ∥ L p ′ ( Ω ) + ∥ U ∥ L p ( Ω ; R n ) ∥ D ρ φ ∥ L p ′ ( Ω ; R n ) , \langle u_0+\operatorname{div}^{\rho} U,\varphi\rangle=\int_{\Omega}(u_0\varphi-U\cdot D^{\rho}\varphi)\,\mathrm{d}x\leq\|u_0\|_{L^p(\Omega)}\|\varphi\|_{L^{p'}(\Omega)}+\|U\|_{L^p(\Omega;\mathbb{R}^n)}\|D^{\rho}\varphi\|_{L^{p'}(\Omega;\mathbb{R}^n)}, ⟨ u 0 + div ρ U , φ ⟩ = ∫ Ω ( u 0 φ − U ⋅ D ρ φ ) d x ≤ ∥ u 0 ∥ L p ( Ω ) ∥ φ ∥ L p ′ ( Ω ) + ∥ U ∥ L p ( Ω ; R n ) ∥ D ρ φ ∥ L p ′ ( Ω ; R n ) ,
showing that
∥ u 0 + div ρ U ∥ H − ρ , p ( Ω ) ≤ ( ∥ u 0 ∥ L p ( Ω ) p + ∥ U ∥ L p ( Ω ; R n ) p ) 1 / p . \|u_0+\operatorname{div}^{\rho} U\|_{H^{-\rho,p}(\Omega)}\leq\left(\|u_0\|_{L^p(\Omega)}^p+\|U\|_{L^p(\Omega;\mathbb{R}^n)}^p\right)^{1/p}. ∥ u 0 + div ρ U ∥ H − ρ , p ( Ω ) ≤ ( ∥ u 0 ∥ L p ( Ω ) p + ∥ U ∥ L p ( Ω ; R n ) p ) 1/ p .
On the contrary, any element in H − ρ , p ( Ω ) H^{-\rho,p}(\Omega) H − ρ , p ( Ω ) can be identified with an element of L p ( Ω ) + L p ( Ω ; R n ) L^p(\Omega)+L^p(\Omega;\mathbb{R}^n) L p ( Ω ) + L p ( Ω ; R n ) , and the proof of this property is the same as the classical one (see, e.g., Proposition 8.14 and Proposition 9.20 in [10 ] ).
Proposition 3.8 (Dual characterization through derivatives). Let Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n be open, p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) , and assume that ρ \rho ρ satisfies (H0). For every F ∈ H − ρ , p ( Ω ) F\in H^{-\rho,p}(\Omega) F ∈ H − ρ , p ( Ω ) there exist u 0 ∈ L p ( Ω ) u_0\in L^p(\Omega) u 0 ∈ L p ( Ω ) and U ∈ L p ( Ω ; R n ) U\in L^p(\Omega;\mathbb{R}^n) U ∈ L p ( Ω ; R n ) such that F = u 0 + div ρ U F=u_0+\operatorname{div}^{\rho}U F = u 0 + div ρ U and
c ~ p ∥ F ∥ H − ρ , p ( Ω ) ≤ max { ∥ u 0 ∥ L p ( Ω ) , ∥ U ∥ L p ( Ω ; R n ) } ≤ C ~ p ∥ F ∥ H − ρ , p ( Ω ) \widetilde{c}_p\|F\|_{H^{-\rho,p}(\Omega)}\leq\max\{\|u_0\|_{L^p(\Omega)},\|U\|_{L^p(\Omega;\mathbb{R}^n)}\}\leq\widetilde{C}_p\|F\|_{H^{-\rho,p}(\Omega)} c p ∥ F ∥ H − ρ , p ( Ω ) ≤ max { ∥ u 0 ∥ L p ( Ω ) , ∥ U ∥ L p ( Ω ; R n ) } ≤ C p ∥ F ∥ H − ρ , p ( Ω ) for some constants c ~ p , C ~ p > 0 \widetilde{c}_p,\widetilde{C}_p>0 c p , C p > 0 . When Ω \Omega Ω is bounded and ρ \rho ρ satisfies additionally (H1), (H2) and (H3’), one can take u 0 = 0 u_0=0 u 0 = 0 .
Proof. The proof can be obtained following the steps in [10 Proposition 8.14] . Note that the norm defined in (23) is equivalent to the one used in [10 Proposition 8.14] , so we find the equivalent dual norm characterization claimed in the statement; the constants c ~ p , C ~ p > 0 \widetilde{c}_p,\widetilde{C}_p>0 c p , C p > 0 only depend on p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) . To derive the last claim we use that for any ρ \rho ρ subject to hypotheses (H0)–(H2) and (H3’) it holds with [8 Theorem 4.11 (i)] a Poincaré-Wirtinger-type inequality in H 0 ρ , p ′ ( Ω ) H_0^{\rho,p'}(\Omega) H 0 ρ , p ′ ( Ω ) . With this result one can again adapt the respective steps in the proof of [10 Proposition 8.14] accordingly. □
We conclude with the comment that unlike in the classical case, if u ∈ C c ∞ ( R n ) u\in C^\infty_c(\mathbb{R}^n) u ∈ C c ∞ ( R n ) , then D ρ u D^\rho u D ρ u is not, in general, a distribution. Indeed, if T T T is a distribution then D ρ T D^\rho T D ρ T is defined, for all φ ∈ C c ∞ ( R n ; R n ) \varphi\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n) φ ∈ C c ∞ ( R n ; R n ) , as
⟨ D ρ T , φ ⟩ = − ⟨ T , D ρ φ ⟩ , \langle D^\rho T,\varphi\rangle=-\langle T,D^\rho\varphi\rangle, ⟨ D ρ T , φ ⟩ = − ⟨ T , D ρ φ ⟩ ,
but D ρ φ D^\rho\varphi D ρ φ is not, in general, in C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) , since it does not need to be compactly supported if the horizon of ρ \rho ρ is not finite. This may be solved by working in the space T ( R n ) \mathcal{T}(\mathbb{R}^n) T ( R n ) of C ∞ ( R n ) C^\infty(\mathbb{R}^n) C ∞ ( R n ) functions u u u such that ∂ α u ∈ L 1 ( R n ) ∩ L ∞ ( R n ) \partial^\alpha u\in L^1(\mathbb{R}^n)\cap L^\infty(\mathbb{R}^n) ∂ α u ∈ L 1 ( R n ) ∩ L ∞ ( R n ) for all α ∈ N n \alpha\in\mathbb{N}^n α ∈ N n . In this case, D ρ φ ∈ T ( R n ) D^\rho\varphi\in\mathcal{T}(\mathbb{R}^n) D ρ φ ∈ T ( R n ) when φ ∈ T ( R n ) \varphi\in\mathcal{T}(\mathbb{R}^n) φ ∈ T ( R n ) , so D ρ T ∈ T ( R n ) ′ D^\rho T\in\mathcal{T}(\mathbb{R}^n)' D ρ T ∈ T ( R n ) ′ whenever T ∈ T ( R n ) ′ T\in\mathcal{T}(\mathbb{R}^n)' T ∈ T ( R n ) ′ (see, e.g., [23 , 39 ] ).
Characterization in terms of the ρ \rho ρ -Laplacian
In this section, we prove a characterization of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) using the square root of the negative ρ \rho ρ -Laplacian, in the spirit of (12) for the Bessel space. After introducing the ρ \rho ρ -Laplacian and deriving its main properties that are important for our purposes, the presentation is split into the cases p = 2 p=2 p = 2 and p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) . For the latter, we shall use the language of unbounded operators in Banach spaces, and for p = 2 p=2 p = 2 , the special theory of unbounded operators in Hilbert spaces (see, e.g., [21 , 28 , 31 , 11 ] ).
Definition and first properties of the ρ \rho ρ -Laplacian
In analogy to the definition of the pointwise and distributional nonlocal gradient and divergence in Section 2.2, we introduce in the following the pointwise and distributional nonlocal ρ \rho ρ -Laplacian. For u ∈ C c ∞ ( R n ) u\in C^\infty_c(\mathbb{R}^n) u ∈ C c ∞ ( R n ) , we define its pointwise ρ \rho ρ -Laplacian Δ 0 ρ \Delta^\rho_0 Δ 0 ρ in the natural way as
Δ 0 ρ u : = div 0 ρ D 0 ρ u , (59) \Delta^\rho_0u:=\operatorname{div}^{\rho}_0D^\rho_0u, \tag*{(59)} Δ 0 ρ u := div 0 ρ D 0 ρ u , (59)
with D 0 ρ D_{0}^{\rho} D 0 ρ and div 0 ρ \operatorname{div}_{0}^{\rho} div 0 ρ given by (2.2) and (2.3). Note that right-hand side in (4.1) is well defined for any ρ \rho ρ subject to (H0), since D 0 ρ D_{0}^{\rho} D 0 ρ is a bounded Lipschitz function (see comments after (2.5)). We will assume from here on that ρ \rho ρ has compact support, and recall that if this is the case, then we have D 0 ρ u , Δ 0 ρ u ∈ C c ∞ ( R n ) D_{0}^{\rho}u,\Delta_{0}^{\rho}u\in C_{c}^{\infty}(\mathbb{R}^{n}) D 0 ρ u , Δ 0 ρ u ∈ C c ∞ ( R n ) for any u ∈ C c ∞ ( R n ) u\in C_{c}^{\infty}(\mathbb{R}^{n}) u ∈ C c ∞ ( R n ) .
We shall now first collect some preliminary relations. By (2.10), it holds
F ( Δ 0 ρ u ) = λ ρ ⋅ F ( D 0 ρ u ) = λ ρ ⋅ λ ρ u ^ = − ∣ λ ρ ∣ 2 u ^ , u ∈ C c ∞ ( R n ) , (60) \mathcal{F}(\Delta_{0}^{\rho}u)=\lambda_{\rho}\cdot\mathcal{F}(D_{0}^{\rho}u)=\lambda_{\rho}\cdot\lambda_{\rho}\widehat{u}=-\lvert\lambda_{\rho}\rvert^{2}\widehat{u},\qquad u\in C_{c}^{\infty}(\mathbb{R}^{n}), \tag*{(60)} F ( Δ 0 ρ u ) = λ ρ ⋅ F ( D 0 ρ u ) = λ ρ ⋅ λ ρ u = − ∣ λ ρ ∣ 2 u , u ∈ C c ∞ ( R n ) , (60)
since λ ρ \lambda_{\rho} λ ρ is purely imaginary on R n \mathbb{R}^{n} R n .
In addition, as demonstrated in the proof of [8 Lemma 3.6] , if ρ \rho ρ fulfills (H0), then for any k ∈ C c ∞ ( R n ) k\in C_{c}^{\infty}(\mathbb{R}^{n}) k ∈ C c ∞ ( R n ) , U ∈ C c ∞ ( R n ; R n ) U\in C_{c}^{\infty}(\mathbb{R}^{n};\mathbb{R}^{n}) U ∈ C c ∞ ( R n ; R n ) it holds k ∗ div 0 ρ U = div 0 ρ ( k ∗ U ) k\ast\operatorname{div}_{0}^{\rho}U=\operatorname{div}_{0}^{\rho}(k\ast U) k ∗ div 0 ρ U = div 0 ρ ( k ∗ U ) . This remains true for kernels k ∈ L 1 ( R n ) k\in L^{1}(\mathbb{R}^{n}) k ∈ L 1 ( R n ) and, by a similar calculation, one gets an analogous identity for the gradient operator D 0 ρ D_{0}^{\rho} D 0 ρ . Hence, for ρ \rho ρ with compact support, it follows from Δ 0 ρ = div 0 ρ D 0 ρ \Delta_{0}^{\rho}=\operatorname{div}_{0}^{\rho}D_{0}^{\rho} Δ 0 ρ = div 0 ρ D 0 ρ that
k ∗ D 0 ρ u = D 0 ρ ( k ∗ u ) , k ∗ Δ 0 ρ u = Δ 0 ρ ( k ∗ u ) , k ∈ L 1 ( R n ) , u ∈ C c ∞ ( R n ) . (61) k\ast D_{0}^{\rho}u=D_{0}^{\rho}(k\ast u),\qquad k\ast\Delta_{0}^{\rho}u=\Delta_{0}^{\rho}(k\ast u),\qquad k\in L^{1}(\mathbb{R}^{n}),\ u\in C_{c}^{\infty}(\mathbb{R}^{n}). \tag*{(61)} k ∗ D 0 ρ u = D 0 ρ ( k ∗ u ) , k ∗ Δ 0 ρ u = Δ 0 ρ ( k ∗ u ) , k ∈ L 1 ( R n ) , u ∈ C c ∞ ( R n ) . (61)
For u ∈ L 1 ( R n ) + L ∞ ( R n ) u\in L^{1}(\mathbb{R}^{n})+L^{\infty}(\mathbb{R}^{n}) u ∈ L 1 ( R n ) + L ∞ ( R n ) , we define the distributional, or weak, nonlocal ρ \rho ρ -Laplacian Δ ρ u \Delta^{\rho}u Δ ρ u as the element in D ′ ( R n ) \mathcal{D}'(\mathbb{R}^{n}) D ′ ( R n ) such that
⟨ Δ ρ u , φ ⟩ D ′ × D = ⟨ u , Δ 0 ρ φ ⟩ , φ ∈ C c ∞ ( R n ) . (62) \langle\Delta^{\rho}u,\varphi\rangle_{\mathcal{D}'\times\mathcal{D}}=\langle u,\Delta_{0}^{\rho}\varphi\rangle,\qquad\varphi\in C_{c}^{\infty}(\mathbb{R}^{n}). \tag*{(62)} ⟨ Δ ρ u , φ ⟩ D ′ × D = ⟨ u , Δ 0 ρ φ ⟩ , φ ∈ C c ∞ ( R n ) . (62)
In what follows, we understand the ρ \rho ρ -Laplacian as an unbounded operator in L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) .
Definition 4.1 (ρ \rho ρ -Laplacian). Let p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and let ρ \rho ρ have compact support and satisfy (H0). The ρ \rho ρ -Laplacian Δ ρ \Delta^{\rho} Δ ρ is the unbounded operator in L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) with domain
D ( Δ ρ ) = { u ∈ L p ( R n ) : Δ ρ u ∈ L p ( R n ) } , D(\Delta^{\rho})=\{u\in L^{p}(\mathbb{R}^{n}):\Delta^{\rho}u\in L^{p}(\mathbb{R}^{n})\}, D ( Δ ρ ) = { u ∈ L p ( R n ) : Δ ρ u ∈ L p ( R n )} , where Δ ρ u \Delta^{\rho}u Δ ρ u is defined as in (4.4) for u ∈ L p ( R n ) u\in L^{p}(\mathbb{R}^{n}) u ∈ L p ( R n ) .
Strictly speaking, the ρ \rho ρ -Laplacian in L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) depends on p p p , but, for the sake of simplicity, we will plainly denote it by Δ ρ \Delta^{\rho} Δ ρ .
The following auxiliary estimate and the resulting approximation result will be used repeatedly below. The proof can be found in Appendix B.
Lemma 4.3. Let p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and assume that ρ \rho ρ satisfies (H0) and has compact support. For any g ∈ C ∞ ( R n ) ∩ H p , p ( R n ) ∩ D ( Δ ρ ) g\in C^{\infty}(\mathbb{R}^{n})\cap H^{p,p}(\mathbb{R}^{n})\cap D(\Delta^{\rho}) g ∈ C ∞ ( R n ) ∩ H p , p ( R n ) ∩ D ( Δ ρ ) and f ∈ C c ∞ ( R n ) f\in C_{c}^{\infty}(\mathbb{R}^{n}) f ∈ C c ∞ ( R n ) it holds
∥ Δ 0 ρ ( f g ) − f Δ ρ g ∥ L p ( R n ) ≤ C ∥ ∇ f ∥ W 1 , ∞ ( R n ; R n ) ∥ g ∥ H p , p ( R n ) (65) \lVert\Delta_{0}^{\rho}(fg)-f\Delta^{\rho}g\rVert_{L^{p}(\mathbb{R}^{n})}\leq C\lVert\nabla f\rVert_{W^{1,\infty}(\mathbb{R}^{n};\mathbb{R}^{n})}\lVert g\rVert_{H^{p,p}(\mathbb{R}^{n})} \tag*{(65)} ∥ Δ 0 ρ ( f g ) − f Δ ρ g ∥ L p ( R n ) ≤ C ∥ ∇ f ∥ W 1 , ∞ ( R n ; R n ) ∥ g ∥ H p , p ( R n ) (65) for some constant C = C ( ρ ) > 0 C=C(\rho)>0 C = C ( ρ ) > 0 .
In particular, for σ ∈ C c ∞ ( R n ) \sigma\in C_{c}^{\infty}(\mathbb{R}^{n}) σ ∈ C c ∞ ( R n ) with σ = 1 \sigma=1 σ = 1 in B 1 B_{1} B 1 , and the smooth sequence of cut-off functions ( σ j ) j (\sigma_{j})_{j} ( σ j ) j defined as σ j ( ⋅ ) : = σ ( ⋅ / j ) \sigma_{j}(\cdot):=\sigma(\cdot/j) σ j ( ⋅ ) := σ ( ⋅ / j ) , one gets
σ j g → g , Δ 0 ρ ( σ j g ) → Δ ρ g in L p ( R n ) . (66) \sigma_{j}g\to g,\qquad\Delta_{0}^{\rho}(\sigma_{j}g)\to\Delta^{\rho}g\quad\text{in }L^{p}(\mathbb{R}^{n}). \tag*{(66)} σ j g → g , Δ 0 ρ ( σ j g ) → Δ ρ g in L p ( R n ) . (66)
In order to see this, we observe that σ j g → g \sigma_jg \to g σ j g → g and σ j Δ ρ g → Δ ρ g \sigma_j\Delta^\rho g \to\Delta^\rho g σ j Δ ρ g → Δ ρ g by dominated convergence, and that ∥ ∇ σ j ∥ W 1 , ∞ ≤ C / j \|\nabla\sigma_j\|_{W^{1,\infty}} \le C/j ∥∇ σ j ∥ W 1 , ∞ ≤ C / j for some uniform constant C > 0 C > 0 C > 0 . The claimed convergence of the ρ \rho ρ -Laplacian follows then immediately from the above-proved estimate (65) for the choice f = σ j f=\sigma_j f = σ j , since
∥ Δ 0 ρ ( σ j g ) − σ j Δ ρ g ∥ L p ( R n ) ≤ C ∥ ∇ σ j ∥ W 1 , ∞ ( R n ; R n ) ∥ g ∥ H ρ , p ( R n ) . \|\Delta_0^\rho(\sigma_jg)-\sigma_j\Delta^\rho g\|_{L^p(\mathbb{R}^n)} \le C\|\nabla\sigma_j\|_{W^{1,\infty}(\mathbb{R}^n;\mathbb{R}^n)}\|g\|_{H^{\rho,p}(\mathbb{R}^n)}. ∥ Δ 0 ρ ( σ j g ) − σ j Δ ρ g ∥ L p ( R n ) ≤ C ∥∇ σ j ∥ W 1 , ∞ ( R n ; R n ) ∥ g ∥ H ρ , p ( R n ) .
We demonstrate next that D ( Δ ρ ) ⊂ H ρ , p ( R n ) D(\Delta^\rho) \subset H^{\rho,p}(\mathbb{R}^n) D ( Δ ρ ) ⊂ H ρ , p ( R n ) . The proof is inspired by the techniques of [8 Theorem 5.2] (see also Remark 4.6 below).
Proposition 4.4. Let p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) and assume that ρ \rho ρ has compact support and satisfies ( H 0 ) (H0) ( H 0 ) –( H 3 ) (H3) ( H 3 ) . Then there exists a vector radial function K ρ ∈ L 1 ( R n ; R n ) K_\rho\in L^1(\mathbb{R}^n;\mathbb{R}^n) K ρ ∈ L 1 ( R n ; R n ) such that
K ρ ^ = λ ρ 1 + ∣ λ ρ ∣ 2 . (67) \widehat{K_\rho}=\frac{\lambda_\rho}{1+|\lambda_\rho|^2}. \tag*{(67)} K ρ = 1 + ∣ λ ρ ∣ 2 λ ρ . (67) Moreover,
D ρ u = K ρ ∗ ( u − Δ ρ u ) ∈ L p ( R n ) , u ∈ D ( Δ ρ ) (68) D^\rho u=K_\rho*(u-\Delta^\rho u)\in L^p(\mathbb{R}^n),\qquad u\in D(\Delta^\rho) \tag*{(68)} D ρ u = K ρ ∗ ( u − Δ ρ u ) ∈ L p ( R n ) , u ∈ D ( Δ ρ ) (68) and, in particular,
D ( Δ ρ ) ⊂ H ρ , p ( R n ) . D(\Delta^\rho)\subset H^{\rho,p}(\mathbb{R}^n). D ( Δ ρ ) ⊂ H ρ , p ( R n ) .
Proof. By C > 0 C>0 C > 0 we will denote a constant that may change from line to line and possibly depends on ρ \rho ρ and, occasionally, on some multi-index α ∈ N 0 n \alpha\in\mathbb{N}_0^n α ∈ N 0 n .
Step 1: Preliminary estimates. We recall initially that due to the assumptions and by (36) it holds λ ρ ( ξ ) ≥ C ∣ ξ ∣ σ ≥ C \lambda_\rho(\xi)\ge C|\xi|^\sigma\ge C λ ρ ( ξ ) ≥ C ∣ ξ ∣ σ ≥ C for all ξ ∈ B 1 c \xi\in B_1^c ξ ∈ B 1 c . In particular, since
∣ λ ρ ∣ 1 + ∣ λ ρ ∣ 2 ≤ C 1 ∣ λ ρ ∣ ≤ C , ξ ∈ B 1 c , \frac{|\lambda_\rho|}{1+|\lambda_\rho|^2}\le C\frac{1}{|\lambda_\rho|}\le C,\qquad\xi\in B_1^c, 1 + ∣ λ ρ ∣ 2 ∣ λ ρ ∣ ≤ C ∣ λ ρ ∣ 1 ≤ C , ξ ∈ B 1 c ,
and because λ ρ ∈ C ∞ ( R n ) \lambda_\rho\in C^\infty(\mathbb{R}^n) λ ρ ∈ C ∞ ( R n ) (see Lemma 2.9), the right-hand side of (4.9) is uniformly bounded in R n \mathbb{R}^n R n and defines a tempered distribution, as which its Fourier inverse is well defined. Moreover, with the help of Lemma 2.8 one can compute that
∣ ∂ α λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 ∣ ≤ ∑ β ≤ α ∣ ∂ β λ ρ ( ξ ) ∣ ∣ ∂ α − β ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) − 1 ∣ ≤ C ∑ β ≤ α ∣ λ ρ ( ξ ) ∣ ∣ ξ ∣ − ∣ β ∣ ∣ λ ρ ( ξ ) ∣ − 2 ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ = C ∣ ξ ∣ − ∣ α ∣ ∣ λ ρ ( ξ ) ∣ − 1 ≤ C ∣ ξ ∣ − ∣ α ∣ − σ , ξ ∈ B 1 c (69) \begin{aligned}
\left|\partial^\alpha\frac{\lambda_\rho(\xi)}{1+|\lambda_\rho(\xi)|^2}\right|
\le\sum_{\beta\le\alpha}|\partial^\beta\lambda_\rho(\xi)|\left|\partial^{\alpha-\beta}(1+|\lambda_\rho(\xi)|^2)^{-1}\right| \\
&\le C\sum_{\beta\le\alpha}|\lambda_\rho(\xi)||\xi|^{-|\beta|}|\lambda_\rho(\xi)|^{-2}|\xi|^{-|\alpha|+|\beta|} \\
&= C|\xi|^{-|\alpha|}|\lambda_\rho(\xi)|^{-1} \\
&\le C|\xi|^{-|\alpha|-\sigma},\qquad\xi\in B_1^c \tag*{(69)}
\end{aligned} ∂ α 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) ≤ β ≤ α ∑ ∣ ∂ β λ ρ ( ξ ) ∣ ∂ α − β ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) − 1 ≤ C β ≤ α ∑ ∣ λ ρ ( ξ ) ∣∣ ξ ∣ − ∣ β ∣ ∣ λ ρ ( ξ ) ∣ − 2 ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ = C ∣ ξ ∣ − ∣ α ∣ ∣ λ ρ ( ξ ) ∣ − 1 ≤ C ∣ ξ ∣ − ∣ α ∣ − σ , ξ ∈ B 1 c (69)
for any α ∈ N 0 n \alpha\in\mathbb{N}_0^n α ∈ N 0 n .
Step 2: Validating that K ρ ∈ L 1 ( R n ) K_\rho\in L^1(\mathbb{R}^n) K ρ ∈ L 1 ( R n ) . We show now that
K ρ ( x ) : = F − 1 ( λ ρ 1 + ∣ λ ρ ∣ 2 ) ( x ) = ∫ R n λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 e 2 π i x ⋅ ξ d ξ = ∫ R n λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 e 2 π i ∣ x ∣ ξ 1 d ξ K_\rho(x):=\mathcal{F}^{-1}\left(\frac{\lambda_\rho}{1+|\lambda_\rho|^2}\right)(x)
=\int_{\mathbb{R}^n}\frac{\lambda_\rho(\xi)}{1+|\lambda_\rho(\xi)|^2}e^{2\pi i x\cdot\xi}\,d\xi
=\int_{\mathbb{R}^n}\frac{\lambda_\rho(\xi)}{1+|\lambda_\rho(\xi)|^2}e^{2\pi i|x|\xi_1}\,d\xi K ρ ( x ) := F − 1 ( 1 + ∣ λ ρ ∣ 2 λ ρ ) ( x ) = ∫ R n 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) e 2 π i x ⋅ ξ d ξ = ∫ R n 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ
exists pointwise for a.e. x ∈ R n x\in\mathbb{R}^n x ∈ R n and lies actually in L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) ; note that the last identity above follows from the symmetry of λ ρ \lambda_\rho λ ρ and similar arguments as in [8 Lemma 2.4] . To that end, we will derive the estimates
∣ K ρ ( x ) ∣ ≤ C 1 ∣ x ∣ n − σ , x ∈ B 1 ∖ { 0 } , (70) |K_\rho(x)|\le C\frac{1}{|x|^{n-\sigma}},\qquad x\in B_1\setminus\{0\}, \tag*{(70)} ∣ K ρ ( x ) ∣ ≤ C ∣ x ∣ n − σ 1 , x ∈ B 1 ∖ { 0 } , (70)
∣ K ρ ( x ) ∣ ≤ C 1 ∣ x ∣ n + 1 , x ∈ B 1 c . (71) |K_\rho(x)|\le C\frac{1}{|x|^{n+1}},\qquad x\in B_1^c. \tag*{(71)} ∣ K ρ ( x ) ∣ ≤ C ∣ x ∣ n + 1 1 , x ∈ B 1 c . (71)
Together, those yield
∫ R n ∣ K ρ ( x ) ∣ d x ≤ C ∫ B 1 1 ∣ x ∣ n − σ d x + C ∫ B 1 c 1 ∣ x ∣ n + 1 d x < ∞ , (72) \int_{\mathbb{R}^n}|K_\rho(x)|\,dx\le C\int_{B_1}\frac{1}{|x|^{n-\sigma}}\,dx+C\int_{B_1^c}\frac{1}{|x|^{n+1}}\,dx<\infty, \tag*{(72)} ∫ R n ∣ K ρ ( x ) ∣ d x ≤ C ∫ B 1 ∣ x ∣ n − σ 1 d x + C ∫ B 1 c ∣ x ∣ n + 1 1 d x < ∞ , (72)
which proves indeed that K ρ ∈ L 1 ( R n ) K_\rho\in L^1(\mathbb{R}^n) K ρ ∈ L 1 ( R n ) .
We assume first that x ∈ B 1 c x \in B_1^c x ∈ B 1 c and show (4.13). With repeated integration by parts, we compute
∫ R n λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 e 2 π i ∣ x ∣ ξ 1 d ξ = 1 ( 2 π i ∣ x ∣ ) n + 1 [ ∫ B 1 + ∫ B 1 c ] ∂ ξ 1 n + 1 ( λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 ) e 2 π i ∣ x ∣ ξ 1 d ξ . \int_{\mathbb{R}^n} \frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}\xi
=
\frac{1}{(2\pi i\lvert x\rvert)^{n+1}}
\left[\int_{B_1}+\int_{B_1^c}\right]
\partial_{\xi_1}^{n+1}\left(\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}\right)e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}\xi. ∫ R n 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ = ( 2 π i ∣ x ∣ ) n + 1 1 [ ∫ B 1 + ∫ B 1 c ] ∂ ξ 1 n + 1 ( 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) ) e 2 π i ∣ x ∣ ξ 1 d ξ .
Inserting in the integral above over B 1 c B_1^c B 1 c the estimate (4.11) with the choice α = ( n + 1 , 0 , … , 0 ) \alpha=(n+1,0,\ldots,0) α = ( n + 1 , 0 , … , 0 ) and using on B 1 B_1 B 1 the smoothness of λ ρ \lambda_\rho λ ρ , yields
∫ R n λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 e 2 π i ∣ x ∣ ξ 1 d ξ ≤ C 1 ∣ x ∣ n + 1 [ 1 + ∫ B 1 c ∣ ξ ∣ − ( n + 1 ) d ξ ] ≤ C 1 ∣ x ∣ n + 1 , x ∈ B 1 c . (73) \int_{\mathbb{R}^n}\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}\xi
\leq C\frac{1}{\lvert x\rvert^{n+1}}\left[1+\int_{B_1^c}\lvert\xi\rvert^{-(n+1)}\,\mathrm{d}\xi\right]
\leq C\frac{1}{\lvert x\rvert^{n+1}},\qquad x\in B_1^c. \tag*{(73)} ∫ R n 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ ≤ C ∣ x ∣ n + 1 1 [ 1 + ∫ B 1 c ∣ ξ ∣ − ( n + 1 ) d ξ ] ≤ C ∣ x ∣ n + 1 1 , x ∈ B 1 c . (73)
Now we study the remaining case that x ∈ B 1 ∖ { 0 } x\in B_1\setminus\{0\} x ∈ B 1 ∖ { 0 } and derive (4.12). The first contribution in the sum
∫ R n λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 e 2 π i ∣ x ∣ ξ 1 d ξ = [ ∫ B 1 + ∫ B 1 c ] λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 e 2 π i ∣ x ∣ ξ 1 d ξ \int_{\mathbb{R}^n}\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}\xi
=
\left[\int_{B_1}+\int_{B_1^c}\right]\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}\xi ∫ R n 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ = [ ∫ B 1 + ∫ B 1 c ] 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ
is, again by the smoothness of λ ρ \lambda_\rho λ ρ , bounded by
∣ ∫ B 1 λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 e 2 π i ∣ x ∣ ξ 1 d ξ ∣ ≤ C ≤ C 1 ∣ x ∣ n − σ , x ∈ B 1 ∖ { 0 } , \left\lvert\int_{B_1}\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}\xi\right\rvert
\leq C\leq C\frac{1}{\lvert x\rvert^{n-\sigma}},\qquad x\in B_1\setminus\{0\}, ∫ B 1 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ ≤ C ≤ C ∣ x ∣ n − σ 1 , x ∈ B 1 ∖ { 0 } ,
and we derive subsequently a similar bound also for the second integral. To that end, we partially integrate ( n − 1 ) (n-1) ( n − 1 ) -times to obtain
∫ B 1 c λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 e 2 π i ∣ x ∣ ξ 1 d ξ = ( − 1 ) n − 1 ( 2 π i ∣ x ∣ ) n − 1 ∫ B 1 c ∂ ξ 1 n − 1 ( λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 ) e 2 π i ∣ x ∣ ξ 1 d ξ + R ( x ) , (74) \int_{B_1^c}\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}\xi
=
\frac{(-1)^{n-1}}{(2\pi i\lvert x\rvert)^{n-1}}
\int_{B_1^c}\partial_{\xi_1}^{n-1}\left(\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}\right)e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}\xi+R(x), \tag*{(74)} ∫ B 1 c 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ = ( 2 π i ∣ x ∣ ) n − 1 ( − 1 ) n − 1 ∫ B 1 c ∂ ξ 1 n − 1 ( 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) ) e 2 π i ∣ x ∣ ξ 1 d ξ + R ( x ) , (74)
where the boundary integrals resulting from the partial integrations are collected in the remainder term
R ( x ) = ∑ k = 1 n − 1 ( − 1 ) k − 1 ( 2 π i ∣ x ∣ ) n − 1 − k ∫ ∂ B 1 ∂ ξ 1 k − 1 ( λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ∂ ξ 1 n − 1 − k e 2 π i ∣ x ∣ ξ 1 d S ( ξ ) . R(x)=\sum_{k=1}^{n-1}\frac{(-1)^{k-1}}{(2\pi i\lvert x\rvert)^{n-1-k}}\int_{\partial B_1}\partial_{\xi_1}^{k-1}\left(\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}\right)\partial_{\xi_1}^{n-1-k}e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}S(\xi). R ( x ) = k = 1 ∑ n − 1 ( 2 π i ∣ x ∣ ) n − 1 − k ( − 1 ) k − 1 ∫ ∂ B 1 ∂ ξ 1 k − 1 ( 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) ) ∂ ξ 1 n − 1 − k e 2 π i ∣ x ∣ ξ 1 d S ( ξ ) .
With (4.11) we get
∣ R ( x ) ∣ ≤ C ∑ k = 1 n − 1 1 ∣ x ∣ k ∫ ∂ B 1 ∣ ξ ∣ − ( k − 1 + σ ) d S ( ξ ) ≤ C 1 ∣ x ∣ n − σ , x ∈ B 1 ∖ { 0 } . (75) \lvert R(x)\rvert\leq C\sum_{k=1}^{n-1}\frac{1}{\lvert x\rvert^k}\int_{\partial B_1}\lvert\xi\rvert^{-(k-1+\sigma)}\,\mathrm{d}S(\xi)\leq C\frac{1}{\lvert x\rvert^{n-\sigma}},\qquad x\in B_1\setminus\{0\}. \tag*{(75)} ∣ R ( x )∣ ≤ C k = 1 ∑ n − 1 ∣ x ∣ k 1 ∫ ∂ B 1 ∣ ξ ∣ − ( k − 1 + σ ) d S ( ξ ) ≤ C ∣ x ∣ n − σ 1 , x ∈ B 1 ∖ { 0 } . (75)
To estimate the right-hand side in (4.16) further, we apply in the following the decomposition of B 1 c B_1^c B 1 c into B 1 / ∣ x ∣ ∖ B 1 B_{1/\lvert x\rvert}\setminus B_1 B 1/ ∣ x ∣ ∖ B 1 and B 1 / ∣ x ∣ c B_{1/\lvert x\rvert}^c B 1/ ∣ x ∣ c . Partially integrating once more on the latter domain it holds
( − 1 ) n − 1 ( 2 π i ∣ x ∣ ) n − 1 ∫ B 1 / ∣ x ∣ c ∂ ξ 1 n − 1 ( λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 ) e 2 π i ∣ x ∣ ξ 1 d ξ = ( − 1 ) n ( 2 π i ∣ x ∣ ) n ∫ ∂ B 1 / ∣ x ∣ ∂ ξ 1 n − 1 ( λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 ) e 2 π i ∣ x ∣ ξ 1 d S ( ξ ) + ( − 1 ) n ( 2 π i ∣ x ∣ ) n ∫ B 1 / ∣ x ∣ c ∂ ξ 1 n ( λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 ) e 2 π i ∣ x ∣ ξ 1 d ξ . (76) \begin{aligned}
\frac{(-1)^{n-1}}{(2\pi i\lvert x\rvert)^{n-1}}\int_{B_{1/\lvert x\rvert}^c}\partial_{\xi_1}^{n-1}\left(\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}\right)e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}\xi\\
&=
\frac{(-1)^n}{(2\pi i\lvert x\rvert)^n}\int_{\partial B_{1/\lvert x\rvert}}\partial_{\xi_1}^{n-1}\left(\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}\right)e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}S(\xi)
+
\frac{(-1)^n}{(2\pi i\lvert x\rvert)^n}\int_{B_{1/\lvert x\rvert}^c}\partial_{\xi_1}^{n}\left(\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}\right)e^{2\pi i\lvert x\rvert\xi_1}\,\mathrm{d}\xi. \tag*{(76)}
\end{aligned} ( 2 π i ∣ x ∣ ) n − 1 ( − 1 ) n − 1 ∫ B 1/ ∣ x ∣ c ∂ ξ 1 n − 1 ( 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) ) e 2 π i ∣ x ∣ ξ 1 d ξ = ( 2 π i ∣ x ∣ ) n ( − 1 ) n ∫ ∂ B 1/ ∣ x ∣ ∂ ξ 1 n − 1 ( 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) ) e 2 π i ∣ x ∣ ξ 1 d S ( ξ ) + ( 2 π i ∣ x ∣ ) n ( − 1 ) n ∫ B 1/ ∣ x ∣ c ∂ ξ 1 n ( 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) ) e 2 π i ∣ x ∣ ξ 1 d ξ . (76)
Estimating now the remaining contributions on B 1 c B_1^c B 1 c with the bounds from (4.11) leads in view of (4.16) to
1 ∣ x ∣ n − 1 ∫ B 1 / ∣ x ∣ ∖ B 1 ∣ ∂ ξ 1 n − 1 ( λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ∣ d ξ ≤ C 1 ∣ x ∣ n − 1 ∫ B 1 / ∣ x ∣ 1 ∣ ξ ∣ n − 1 + σ d ξ = C 1 ∣ x ∣ n − σ , (77) \frac{1}{\lvert x\rvert^{n-1}}\int_{B_{1/\lvert x\rvert}\setminus B_1}\left\lvert\partial_{\xi_1}^{n-1}\left(\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}\right)\right\rvert\,\mathrm{d}\xi
\leq C\frac{1}{\lvert x\rvert^{n-1}}\int_{B_{1/\lvert x\rvert}}\frac{1}{\lvert\xi\rvert^{n-1+\sigma}}\,\mathrm{d}\xi
=C\frac{1}{\lvert x\rvert^{n-\sigma}}, \tag*{(77)} ∣ x ∣ n − 1 1 ∫ B 1/ ∣ x ∣ ∖ B 1 ∂ ξ 1 n − 1 ( 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) ) d ξ ≤ C ∣ x ∣ n − 1 1 ∫ B 1/ ∣ x ∣ ∣ ξ ∣ n − 1 + σ 1 d ξ = C ∣ x ∣ n − σ 1 , (77)
and, with the help of (4.18), one gets
1 ∣ x ∣ n ∫ ∂ B 1 / ∣ x ∣ ∣ ∂ ξ 1 n − 1 ( λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ∣ d S ( ξ ) ≤ C 1 ∣ x ∣ n ∫ ∂ B 1 / ∣ x ∣ 1 ∣ ξ ∣ n − 1 + σ d S ( ξ ) = C 1 ∣ x ∣ n − σ . (78) \frac{1}{\lvert x\rvert^n}\int_{\partial B_{1/\lvert x\rvert}}\left\lvert\partial_{\xi_1}^{n-1}\left(\frac{\lambda_\rho(\xi)}{1+\lvert\lambda_\rho(\xi)\rvert^2}\right)\right\rvert\,\mathrm{d}S(\xi)
\leq C\frac{1}{\lvert x\rvert^n}\int_{\partial B_{1/\lvert x\rvert}}\frac{1}{\lvert\xi\rvert^{n-1+\sigma}}\,\mathrm{d}S(\xi)
=C\frac{1}{\lvert x\rvert^{n-\sigma}}. \tag*{(78)} ∣ x ∣ n 1 ∫ ∂ B 1/ ∣ x ∣ ∂ ξ 1 n − 1 ( 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) ) d S ( ξ ) ≤ C ∣ x ∣ n 1 ∫ ∂ B 1/ ∣ x ∣ ∣ ξ ∣ n − 1 + σ 1 d S ( ξ ) = C ∣ x ∣ n − σ 1 . (78)
as well as,
1 ∣ x ∣ n ∫ B 1 / ∣ x ∣ c ∣ ∂ ξ 1 n ( λ ρ ( ξ ) 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ∣ d ξ ≤ C 1 ∣ x ∣ n ∫ B 1 / ∣ x ∣ c 1 ∣ ξ ∣ n + σ d ξ = C 1 ∣ x ∣ n − σ , (79) \frac{1}{|x|^{n}}\int_{B_{1/|x|}^{c}}\left|\partial_{\xi_1}^{n}\left(\frac{\lambda_{\rho}(\xi)}{1+|\lambda_{\rho}(\xi)|^{2}}\right)\right|\,\mathrm{d}\xi\le C\frac{1}{|x|^{n}}\int_{B_{1/|x|}^{c}}\frac{1}{|\xi|^{n+\sigma}}\,\mathrm{d}\xi= C\frac{1}{|x|^{n-\sigma}}, \tag*{(79)} ∣ x ∣ n 1 ∫ B 1/∣ x ∣ c ∂ ξ 1 n ( 1 + ∣ λ ρ ( ξ ) ∣ 2 λ ρ ( ξ ) ) d ξ ≤ C ∣ x ∣ n 1 ∫ B 1/∣ x ∣ c ∣ ξ ∣ n + σ 1 d ξ = C ∣ x ∣ n − σ 1 , (79)
where always x ∈ B 1 ∖ { 0 } x \in B_{1} \setminus\{0\} x ∈ B 1 ∖ { 0 } .
Overall, from (4.15) for x ∈ B 1 c x \in B_{1}^{c} x ∈ B 1 c and (4.16) for x ∈ B 1 ∖ { 0 } x \in B_{1} \setminus\{0\} x ∈ B 1 ∖ { 0 } together with estimates (4.17)–(4.21), we obtain (4.14) and K ρ ∈ L 1 ( R n ; R n ) K_{\rho} \in L^{1}(\mathbb{R}^{n};\mathbb{R}^{n}) K ρ ∈ L 1 ( R n ; R n ) . Note, in addition, that K ρ K_{\rho} K ρ is vector radial, since this property is preserved by Fourier inversion.
Step 3: Proof of (4.10) and D ( Δ ρ ) ⊂ H ρ , p ( R n ) D(\Delta^{\rho}) \subset H^{\rho,p}(\mathbb{R}^{n}) D ( Δ ρ ) ⊂ H ρ , p ( R n ) . Next, the identity K ρ ∗ ( Φ − Δ 0 ρ Φ ) = div 0 ρ Φ K_{\rho} * (\Phi-\Delta_{0}^{\rho}\Phi)=\operatorname{div}_{0}^{\rho}\Phi K ρ ∗ ( Φ − Δ 0 ρ Φ ) = div 0 ρ Φ for Φ ∈ C c ∞ ( R n ; R n ) \Phi\in C_{c}^{\infty}(\mathbb{R}^{n};\mathbb{R}^{n}) Φ ∈ C c ∞ ( R n ; R n ) (where K ρ ∗ Φ K_{\rho} * \Phi K ρ ∗ Φ is defined as the trace of the matrix ( ( K ρ ) i ∗ Φ j ) i j ((K_{\rho})_{i} * \Phi_{j})_{ij} (( K ρ ) i ∗ Φ j ) ij , and similarly for K ρ ∗ Δ 0 ρ Φ K_{\rho} * \Delta_{0}^{\rho}\Phi K ρ ∗ Δ 0 ρ Φ ) follows by definition of K ρ K_{\rho} K ρ and (2.10), (4.2), stating the Fourier symbols of D 0 ρ D_{0}^{\rho} D 0 ρ , div 0 ρ \operatorname{div}_{0}^{\rho} div 0 ρ and Δ 0 ρ \Delta_{0}^{\rho} Δ 0 ρ . Precisely, we have
K ρ ∗ ( Φ − Δ 0 ρ Φ ) = F − 1 ( λ ρ 1 + ∣ λ ρ ∣ 2 ⋅ F ( Φ − Δ 0 ρ Φ ) ) = F − 1 ( λ ρ ⋅ Φ ^ ) = div 0 ρ Φ . (80) K_{\rho} * (\Phi-\Delta_{0}^{\rho}\Phi)=\mathcal{F}^{-1}\left(\frac{\lambda_{\rho}}{1+|\lambda_{\rho}|^{2}}\cdot\mathcal{F}(\Phi-\Delta_{0}^{\rho}\Phi)\right)=\mathcal{F}^{-1}(\lambda_{\rho}\cdot\widehat{\Phi})=\operatorname{div}_{0}^{\rho}\Phi. \tag*{(80)} K ρ ∗ ( Φ − Δ 0 ρ Φ ) = F − 1 ( 1 + ∣ λ ρ ∣ 2 λ ρ ⋅ F ( Φ − Δ 0 ρ Φ ) ) = F − 1 ( λ ρ ⋅ Φ ) = div 0 ρ Φ. (80)
By duality, this also implies the property D ( Δ ρ ) ⊂ H ρ , p ( R n ) D(\Delta^{\rho}) \subset H^{\rho,p}(\mathbb{R}^{n}) D ( Δ ρ ) ⊂ H ρ , p ( R n ) , as we show subsequently. In order to see that, we will need additional properties of K ρ ∗ Φ K_{\rho} * \Phi K ρ ∗ Φ , namely that K ρ ∗ Φ ∈ H ρ , p ( R n ) K_{\rho} * \Phi\in H^{\rho,p}(\mathbb{R}^{n}) K ρ ∗ Φ ∈ H ρ , p ( R n ) and K ρ ∗ Φ ∈ D ( Δ ρ ) K_{\rho} * \Phi\in D(\Delta^{\rho}) K ρ ∗ Φ ∈ D ( Δ ρ ) for any Φ ∈ C c ∞ ( R n ; R n ) \Phi\in C_{c}^{\infty}(\mathbb{R}^{n};\mathbb{R}^{n}) Φ ∈ C c ∞ ( R n ; R n ) . Indeed, D 0 ρ ( K ρ ∗ Φ ) D_{0}^{\rho}(K_{\rho} * \Phi) D 0 ρ ( K ρ ∗ Φ ) is pointwise well defined (see (2.5) and the comments thereafter) and an element of L p ( R n ) L^{p}(\mathbb{R}^{n}) L p ( R n ) by Young’s convolution inequality and K ρ , Q ρ ∈ L 1 ( R n ) K_{\rho},Q_{\rho} \in L^{1}(\mathbb{R}^{n}) K ρ , Q ρ ∈ L 1 ( R n ) , since, with (4.3) and (2.8),
D 0 ρ ( K ρ ∗ Φ ) = K ρ ∗ D 0 ρ Φ = K ρ ∗ ( Q ρ ∗ ∇ Φ ) , j ∈ N . D_{0}^{\rho}(K_{\rho} * \Phi)=K_{\rho} * D_{0}^{\rho}\Phi=K_{\rho} * (Q_{\rho} * \nabla\Phi),\qquad j\in\mathbb{N}. D 0 ρ ( K ρ ∗ Φ ) = K ρ ∗ D 0 ρ Φ = K ρ ∗ ( Q ρ ∗ ∇Φ ) , j ∈ N .
A similar computation shows that Δ ρ K ρ ∗ Φ \Delta^{\rho}K_{\rho} * \Phi Δ ρ K ρ ∗ Φ exists pointwise and that K ρ ∗ Φ ∈ D ( Δ ρ ) K_{\rho} * \Phi\in D(\Delta^{\rho}) K ρ ∗ Φ ∈ D ( Δ ρ ) . This allows us now to apply (4.8) and to conclude that there exists ( σ j ) j ⊂ C c ∞ ( R n ) (\sigma_{j})_{j} \subset C_{c}^{\infty}(\mathbb{R}^{n}) ( σ j ) j ⊂ C c ∞ ( R n ) such that ( σ j ( K ρ ∗ Φ ) ) j ⊂ C c ∞ ( R n ) (\sigma_{j}(K_{\rho} * \Phi))_{j} \subset C_{c}^{\infty}(\mathbb{R}^{n}) ( σ j ( K ρ ∗ Φ ) ) j ⊂ C c ∞ ( R n ) and
σ j ( K ρ ∗ Φ ) → K ρ ∗ Φ , Δ ρ ( σ j ( K ρ ∗ Φ ) ) → Δ ρ ( K ρ ∗ Φ ) in L p ( R n ) . (81) \sigma_{j}(K_{\rho} * \Phi)\to K_{\rho} * \Phi,\qquad\Delta^{\rho}(\sigma_{j}(K_{\rho} * \Phi))\to\Delta^{\rho}(K_{\rho} * \Phi)\quad\text{in }L^{p}(\mathbb{R}^{n}). \tag*{(81)} σ j ( K ρ ∗ Φ ) → K ρ ∗ Φ , Δ ρ ( σ j ( K ρ ∗ Φ )) → Δ ρ ( K ρ ∗ Φ ) in L p ( R n ) . (81)
We have collected now all ingredients for the final argument. Using first the fact that K ρ K_{\rho} K ρ is vector radial together with the convolution identity from [10 Proposition 4.16] and then later the commutation relation in (4.3), it holds
⟨ K ρ ∗ ( u − Δ ρ u ) , Φ ⟩ = − ⟨ u − Δ ρ u , K ρ ∗ Φ ⟩ = − lim j → ∞ ⟨ u − Δ ρ u , σ j ( K ρ ∗ Φ ) ⟩ = − lim j → ∞ ⟨ u , ( Id − Δ 0 ρ ) ( σ j ( K ρ ∗ Φ ) ) ⟩ = − ⟨ u , K ρ ∗ ( Φ − Δ 0 ρ Φ ) ⟩ = − ⟨ u , div 0 ρ Φ ⟩ . \begin{aligned}
\langle K_{\rho} * (u-\Delta^{\rho}u),\Phi\rangle
&=-\langle u-\Delta^{\rho}u,K_{\rho} * \Phi\rangle
=-\lim_{j\to\infty}\langle u-\Delta^{\rho}u,\sigma_{j}(K_{\rho} * \Phi)\rangle\\
&=-\lim_{j\to\infty}\langle u,(\operatorname{Id}-\Delta_{0}^{\rho})(\sigma_{j}(K_{\rho} * \Phi))\rangle
=-\langle u,K_{\rho} * (\Phi-\Delta_{0}^{\rho}\Phi)\rangle
=-\langle u,\operatorname{div}_{0}^{\rho}\Phi\rangle.
\end{aligned} ⟨ K ρ ∗ ( u − Δ ρ u ) , Φ ⟩ = − ⟨ u − Δ ρ u , K ρ ∗ Φ ⟩ = − j → ∞ lim ⟨ u − Δ ρ u , σ j ( K ρ ∗ Φ )⟩ = − j → ∞ lim ⟨ u , ( Id − Δ 0 ρ ) ( σ j ( K ρ ∗ Φ ))⟩ = − ⟨ u , K ρ ∗ ( Φ − Δ 0 ρ Φ )⟩ = − ⟨ u , div 0 ρ Φ ⟩ .
Furthermore, by Young’s convolution inequality and K ρ ∈ L 1 ( R n ) K_{\rho}\in L^{1}(\mathbb{R}^{n}) K ρ ∈ L 1 ( R n ) , we deduce that K ρ ∗ ( u − Δ ρ u ) ∈ L p ( R n ) K_{\rho}*(u-\Delta^{\rho}u)\in L^{p}(\mathbb{R}^{n}) K ρ ∗ ( u − Δ ρ u ) ∈ L p ( R n ) whenever u ∈ D ( Δ ρ ) u\in D(\Delta^{\rho}) u ∈ D ( Δ ρ ) . It follows then by definition that D ρ u = K ρ ∗ ( u − Δ ρ u ) ∈ L p ( R n ) D^{\rho}u=K_{\rho}*(u-\Delta^{\rho}u)\in L^{p}(\mathbb{R}^{n}) D ρ u = K ρ ∗ ( u − Δ ρ u ) ∈ L p ( R n ) exists, implying that u ∈ H ρ , p ( R n ) u\in H^{\rho,p}(\mathbb{R}^{n}) u ∈ H ρ , p ( R n ) .
Remark 4.5 (Hilbert space case). For p = 2 p=2 p = 2 , the inclusion D ( Δ ρ ) ⊂ H ρ , 2 ( R n ) D(\Delta^{\rho})\subset H^{\rho,2}(\mathbb{R}^{n}) D ( Δ ρ ) ⊂ H ρ , 2 ( R n ) holds true under the milder conditions that ρ \rho ρ has compact support and merely satisfies (H0), (H1) and (H3’). In fact, using Fourier multiplier theory in L 2 ( R n ) L^{2}(\mathbb{R}^{n}) L 2 ( R n ) instead, one does not need to rely on the property that λ ρ / ( 1 + ∣ λ ρ ∣ 2 ) \lambda_{\rho}/(1+|\lambda_{\rho}|^{2}) λ ρ / ( 1 + ∣ λ ρ ∣ 2 ) admits a Fourier inverse in L 1 ( R n ) L^{1}(\mathbb{R}^{n}) L 1 ( R n ) . We sketch here the arguments.
Due to (2.18), the assumptions on ρ \rho ρ assure that λ ρ / ( 1 + ∣ λ ρ ∣ 2 ) ∈ L ∞ ( R n ) \lambda_{\rho}/(1+|\lambda_{\rho}|^{2})\in L^{\infty}(\mathbb{R}^{n}) λ ρ / ( 1 + ∣ λ ρ ∣ 2 ) ∈ L ∞ ( R n ) and this implies that this fraction is an L 2 L^{2} L 2 -Fourier multiplier (see [17 Theorem 2.5.10] ). We define
R ρ : L 2 ( R n ) → L 2 ( R n ; R n ) , u ↦ F − 1 ( λ ρ 1 + ∣ λ ρ ∣ 2 u ^ ) R_{\rho}:L^{2}(\mathbb{R}^{n})\to L^{2}(\mathbb{R}^{n};\mathbb{R}^{n}),\qquad u\mapsto\mathcal{F}^{-1}\left(\frac{\lambda_{\rho}}{1+|\lambda_{\rho}|^{2}}\widehat{u}\right) R ρ : L 2 ( R n ) → L 2 ( R n ; R n ) , u ↦ F − 1 ( 1 + ∣ λ ρ ∣ 2 λ ρ u )
and extend it to a map L 2 ( R n ; R n ) → L 2 ( R n ) L^{2}(\mathbb{R}^{n};\mathbb{R}^{n})\to L^{2}(\mathbb{R}^{n}) L 2 ( R n ; R n ) → L 2 ( R n ) by inserting the scalar product in Fourier space. It follows from Parseval’s identity that the adjoint operator of R ρ R_{\rho} R ρ is given by − R ρ -R_{\rho} − R ρ . Now, for any Φ ∈ C c ∞ ( R n ; R n ) \Phi\in C_{c}^{\infty}(\mathbb{R}^{n};\mathbb{R}^{n}) Φ ∈ C c ∞ ( R n ; R n ) the functions
F − 1 ( λ ρ ⊗ λ ρ 1 + ∣ λ ρ ∣ 2 Φ ^ ) , − F − 1 ( ∣ λ ρ ∣ 2 1 + ∣ λ ρ ∣ 2 Φ ^ ) (82) \mathcal{F}^{-1}\left(\frac{\lambda_{\rho}\otimes\lambda_{\rho}}{1+|\lambda_{\rho}|^{2}}\widehat{\Phi}\right),\qquad
-\mathcal{F}^{-1}\left(\frac{|\lambda_{\rho}|^{2}}{1+|\lambda_{\rho}|^{2}}\widehat{\Phi}\right) \tag*{(82)} F − 1 ( 1 + ∣ λ ρ ∣ 2 λ ρ ⊗ λ ρ Φ ) , − F − 1 ( 1 + ∣ λ ρ ∣ 2 ∣ λ ρ ∣ 2 Φ ) (82)
lie in L 2 ( R n ; R n ) L^2(\mathbb{R}^n;\mathbb{R}^n) L 2 ( R n ; R n ) and L 2 ( R n ) L^2(\mathbb{R}^n) L 2 ( R n ) , respectively, because λ ρ ⊗ λ ρ / ( 1 + ∣ λ ρ ∣ 2 ) \lambda_\rho\otimes\lambda_\rho/(1+|\lambda_\rho|^2) λ ρ ⊗ λ ρ / ( 1 + ∣ λ ρ ∣ 2 ) , ∣ λ ρ ∣ λ ρ ∣ 2 / ( 1 + ∣ λ ρ ∣ 2 ) |\lambda_\rho|\lambda_\rho|^2/(1+|\lambda_\rho|^2) ∣ λ ρ ∣ λ ρ ∣ 2 / ( 1 + ∣ λ ρ ∣ 2 ) have polynomial growth (which follows again from (35) and the fact that λ ρ \lambda_\rho λ ρ has linear growth as shown in Lemma 2.9, where the arguments merely use ρ ∈ L 1 ( R n ) \rho\in L^1(\mathbb{R}^n) ρ ∈ L 1 ( R n ) ). Additionally, the expressions in (82) are equal to D ρ R ρ Φ D^\rho R_\rho\Phi D ρ R ρ Φ and Δ ρ R ρ Φ \Delta^\rho R_\rho\Phi Δ ρ R ρ Φ , respectively, as can be quickly verified using Parseval’s identity. We conclude from this that R ρ Φ ∈ H ρ , 2 ( R n ) R_\rho\Phi\in H^{\rho,2}(\mathbb{R}^n) R ρ Φ ∈ H ρ , 2 ( R n ) and, with similar arguments as in (80) , that ( 1 − Δ ρ ) R ρ Φ = div 0 ρ Φ (1-\Delta^\rho)R_\rho\Phi=\operatorname{div}_0^\rho\Phi ( 1 − Δ ρ ) R ρ Φ = div 0 ρ Φ . Hence, again by a suitable approximation and (66) ,
⟨ R ρ ( u − Δ ρ u ) , Φ ⟩ = − ⟨ u − Δ ρ u , R ρ Φ ⟩ = − ⟨ u , ( Id − Δ ρ ) R ρ Φ ⟩ = ⟨ u , div 0 ρ Φ ⟩ . \langle R_\rho(u-\Delta^\rho u),\Phi\rangle=-\langle u-\Delta^\rho u,R_\rho\Phi\rangle=-\langle u,(\operatorname{Id}-\Delta^\rho)R_\rho\Phi\rangle=\langle u,\operatorname{div}_0^\rho\Phi\rangle. ⟨ R ρ ( u − Δ ρ u ) , Φ ⟩ = − ⟨ u − Δ ρ u , R ρ Φ ⟩ = − ⟨ u , ( Id − Δ ρ ) R ρ Φ ⟩ = ⟨ u , div 0 ρ Φ ⟩ .
Thus, we have shown that D ρ u = R ρ ( u − Δ ρ u ) ∈ L 2 ( R n ; R n ) D^\rho u=R_\rho(u-\Delta^\rho u)\in L^2(\mathbb{R}^n;\mathbb{R}^n) D ρ u = R ρ ( u − Δ ρ u ) ∈ L 2 ( R n ; R n ) and u ∈ H ρ , 2 ( R n ) u\in H^{\rho,2}(\mathbb{R}^n) u ∈ H ρ , 2 ( R n ) .
Remark 4.6 (Refinement of a nonlocal fundamental theorem of calculus). We comment briefly on [8 Theorem 5.2] where the authors show a nonlocal analogue of the classical fundamental theorem of calculus. In particular, they establish that for compactly supported ρ \rho ρ subject to (H0)–(H4) (where (H4) is an additional hypothesis that they impose) there exists a function V ρ ∈ C ∞ ( R n ∖ { 0 } ; R n ) ∩ L l o c 1 ( R n ; R n ) V_\rho\in C^\infty(\mathbb{R}^n\setminus\{0\};\mathbb{R}^n)\cap L^1_{\mathrm{loc}}(\mathbb{R}^n;\mathbb{R}^n) V ρ ∈ C ∞ ( R n ∖ { 0 } ; R n ) ∩ L loc 1 ( R n ; R n ) such that
u ( x ) = ∫ R n V ρ ( x − y ) ⋅ D 0 ρ u ( y ) d y , x ∈ R n , u ∈ C c ∞ ( R n ) . u(x)=\int_{\mathbb{R}^n}V_\rho(x-y)\cdot D_0^\rho u(y)\,\mathrm{d}y,\qquad x\in\mathbb{R}^n,\ u\in C_c^\infty(\mathbb{R}^n). u ( x ) = ∫ R n V ρ ( x − y ) ⋅ D 0 ρ u ( y ) d y , x ∈ R n , u ∈ C c ∞ ( R n ) .
Later in [8 Corollary 5.3] , this is extended to u ∈ H 0 ρ , p ( Ω ) u\in H_0^{\rho,p}(\Omega) u ∈ H 0 ρ , p ( Ω ) for open and bounded Ω ⊂ R n \Omega\subset\mathbb{R}^n Ω ⊂ R n . Inspecting the proof of [8 Theorem 5.2] one finds that hypothesis (H4) is only used to derive that an expression similar to
1 ∣ x ∣ n − 1 ∫ B 1 ∖ B 1 / ∣ x ∣ 1 ∣ ξ ∣ Q ρ ^ ( ξ ) d ξ , x ∈ B 1 \frac{1}{|x|^{n-1}}\int_{B_1\setminus B_{1/|x|}}\frac{1}{|\xi|\widehat{Q_\rho}(\xi)}\,\mathrm{d}\xi,\qquad x\in B_1 ∣ x ∣ n − 1 1 ∫ B 1 ∖ B 1/∣ x ∣ ∣ ξ ∣ Q ρ ( ξ ) 1 d ξ , x ∈ B 1
is finite and integrable over B 1 B_1 B 1 . From (36) (which does not require (H4)) and the computations in (77) it follows now that
1 ∣ x ∣ n − 1 ∫ B 1 ∖ B 1 / ∣ x ∣ 1 ∣ ξ ∣ Q ρ ^ ( ξ ) d ξ ≤ C 1 ∣ x ∣ n − 1 ∫ B 1 ∖ B 1 / ∣ x ∣ 1 ∣ ξ ∣ n − 1 + σ d ξ = C 1 ∣ x ∣ n − σ , x ∈ B 1 , \frac{1}{|x|^{n-1}}\int_{B_1\setminus B_{1/|x|}}\frac{1}{|\xi|\widehat{Q_\rho}(\xi)}\,\mathrm{d}\xi\leq C\frac{1}{|x|^{n-1}}\int_{B_1\setminus B_{1/|x|}}\frac{1}{|\xi|^{n-1+\sigma}}\,\mathrm{d}\xi=C\frac{1}{|x|^{n-\sigma}},\qquad x\in B_1, ∣ x ∣ n − 1 1 ∫ B 1 ∖ B 1/∣ x ∣ ∣ ξ ∣ Q ρ ( ξ ) 1 d ξ ≤ C ∣ x ∣ n − 1 1 ∫ B 1 ∖ B 1/∣ x ∣ ∣ ξ ∣ n − 1 + σ 1 d ξ = C ∣ x ∣ n − σ 1 , x ∈ B 1 ,
where, as desired, the right-hand side is integrable. Consequently, the nonlocal fundamental theorem of calculus presented in [8 Theorem 5.2] and [8 Corollary 5.3] is valid for the larger class of kernels being compactly supported and only constrained by hypotheses (H0)–(H3).
The result of Proposition 4.4 allows us to prove that C c ∞ ( R n ) C_c^\infty(\mathbb{R}^n) C c ∞ ( R n ) is a core for Δ ρ \Delta^\rho Δ ρ .
Lemma 4.7 (Core property). Let p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and assume that ρ \rho ρ has compact support and satisfies (H0)–(H3). For each u ∈ D ( Δ ρ ) u\in D(\Delta^\rho) u ∈ D ( Δ ρ ) there exists a sequence ( u j ) j (u_j)_j ( u j ) j in C c ∞ ( R n ) C_c^\infty(\mathbb{R}^n) C c ∞ ( R n ) such that
u j → u , Δ 0 ρ u j → Δ ρ u in L p ( R n ) . u_j\to u,\qquad\Delta_0^\rho u_j\to\Delta^\rho u\qquad\text{in }L^p(\mathbb{R}^n). u j → u , Δ 0 ρ u j → Δ ρ u in L p ( R n ) . Proof. Let u ∈ D ( Δ ρ ) u\in D(\Delta^\rho) u ∈ D ( Δ ρ ) . Let ( η ε ) ε > 0 ⊂ C c ∞ ( R n ) (\eta_\varepsilon)_{\varepsilon>0}\subset C_c^\infty(\mathbb{R}^n) ( η ε ) ε > 0 ⊂ C c ∞ ( R n ) be a standard family of mollifiers. Then η ε ∗ u → u \eta_\varepsilon*u\to u η ε ∗ u → u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) for ε → 0 \varepsilon\to0 ε → 0 . Thanks to (61) , we have η ε ∗ Δ 0 ρ φ = Δ 0 ρ ( η ε ∗ φ ) \eta_\varepsilon*\Delta_0^\rho\varphi=\Delta_0^\rho(\eta_\varepsilon*\varphi) η ε ∗ Δ 0 ρ φ = Δ 0 ρ ( η ε ∗ φ ) for any φ ∈ C c ∞ ( R n ) \varphi\in C_c^\infty(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) . Together with [10 Proposition 4.16] and the fact that η ε \eta_\varepsilon η ε is radial, we obtain
⟨ η ε ∗ u , Δ 0 ρ φ ⟩ = ⟨ u , η ε ∗ Δ 0 ρ φ ⟩ = ⟨ u , Δ 0 ρ ( η ε ∗ φ ) ⟩ = ⟨ Δ ρ u , η ε ∗ φ ⟩ = ⟨ η ε ∗ Δ ρ u , φ ⟩ \langle\eta_\varepsilon*u,\Delta_0^\rho\varphi\rangle=\langle u,\eta_\varepsilon*\Delta_0^\rho\varphi\rangle=\langle u,\Delta_0^\rho(\eta_\varepsilon*\varphi)\rangle=\langle\Delta^\rho u,\eta_\varepsilon*\varphi\rangle=\langle\eta_\varepsilon*\Delta^\rho u,\varphi\rangle ⟨ η ε ∗ u , Δ 0 ρ φ ⟩ = ⟨ u , η ε ∗ Δ 0 ρ φ ⟩ = ⟨ u , Δ 0 ρ ( η ε ∗ φ )⟩ = ⟨ Δ ρ u , η ε ∗ φ ⟩ = ⟨ η ε ∗ Δ ρ u , φ ⟩ for all φ ∈ C c ∞ ( R n ) \varphi\in C_c^\infty(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) . Therefore, η ε ∗ u ∈ D ( Δ ρ ) \eta_\varepsilon*u\in D(\Delta^\rho) η ε ∗ u ∈ D ( Δ ρ ) and Δ ρ ( η ε ∗ u ) = η ε ∗ Δ ρ u \Delta^\rho(\eta_\varepsilon*u)=\eta_\varepsilon*\Delta^\rho u Δ ρ ( η ε ∗ u ) = η ε ∗ Δ ρ u , and consequently, Δ ρ ( η ε ∗ u ) → Δ ρ u \Delta^\rho(\eta_\varepsilon*u)\to\Delta^\rho u Δ ρ ( η ε ∗ u ) → Δ ρ u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) as ε → 0 \varepsilon\to0 ε → 0 . Applying Proposition 4.4 , further yields η ε ∗ u ∈ H ρ , p ( R n ) \eta_\varepsilon*u\in H^{\rho,p}(\mathbb{R}^n) η ε ∗ u ∈ H ρ , p ( R n ) .
Thus, for each fixed ε > 0 \varepsilon>0 ε > 0 , we infer from (66) that there exists ( σ j ) j ⊂ C c ∞ ( R n ) (\sigma_j)_j\subset C_c^\infty(\mathbb{R}^n) ( σ j ) j ⊂ C c ∞ ( R n ) such that ( σ j ( η ε ∗ u ) ) j ⊂ C c ∞ ( R n ) (\sigma_j(\eta_\varepsilon*u))_j\subset C_c^\infty(\mathbb{R}^n) ( σ j ( η ε ∗ u ) ) j ⊂ C c ∞ ( R n ) and
σ j ( η ε ∗ u ) → η ε ∗ u , Δ 0 ρ ( σ j ( η ε ∗ u ) ) → Δ ρ ( η ε ∗ u ) in L p ( R n ) . \sigma_j(\eta_\varepsilon*u)\to\eta_\varepsilon*u,\qquad\Delta_0^\rho(\sigma_j(\eta_\varepsilon*u))\to\Delta^\rho(\eta_\varepsilon*u)\qquad\text{in }L^p(\mathbb{R}^n). σ j ( η ε ∗ u ) → η ε ∗ u , Δ 0 ρ ( σ j ( η ε ∗ u )) → Δ ρ ( η ε ∗ u ) in L p ( R n ) . Finally, by choosing a suitable diagonal sequence ( σ j ( ε ) ( η ε ∗ u ) ) ε (\sigma_{j(\varepsilon)}(\eta_\varepsilon*u))_\varepsilon ( σ j ( ε ) ( η ε ∗ u ) ) ε depending on ε \varepsilon ε , and letting ε → 0 \varepsilon\to0 ε → 0 , we obtain the desired result. □ \square □
Remark 4.8. a) Since in the above proof, hypothesis (H3) is only used to apply Proposition 4.4 to obtain that η ε ∗ u ∈ H ρ , p ( R n ) \eta_{\varepsilon} * u \in H^{\rho,p}(\mathbb{R}^{n}) η ε ∗ u ∈ H ρ , p ( R n ) , it follows from Remark 4.5 that for p = 2 p=2 p = 2 , Lemma 4.7 is true for all kernels ρ \rho ρ with compact support and fulfilling merely (H0), (H1) and (H3’).
b) For u ∈ C ∞ ( R n ) ∩ D ( Δ ρ ) u \in C^{\infty}(\mathbb{R}^{n}) \cap D(\Delta^{\rho}) u ∈ C ∞ ( R n ) ∩ D ( Δ ρ ) the proof of Lemma 4.7 reduces to applying the cut-off argument from (66) which only relies on Lemma 4.3 . Thus, the claimed simultaneous approximation exists as soon as hypothesis (H0) is satisfied and ρ \rho ρ has compact support.
c) It was shown in [8 Theorem 3.9] that by the very same construction used above in Lemma 4.7 , but here only requiring (H0), any function in u ∈ H ρ , p ( R n ) u \in H^{\rho,p}(\mathbb{R}^{n}) u ∈ H ρ , p ( R n ) can be approximated by a sequence in C c ∞ ( R n ) C^{\infty}_{c}(\mathbb{R}^{n}) C c ∞ ( R n ) . Since additionally D ( Δ ρ ) ⊂ H ρ , p ( R n ) D(\Delta^{\rho}) \subset H^{\rho,p}(\mathbb{R}^{n}) D ( Δ ρ ) ⊂ H ρ , p ( R n ) by virtue of Proposition 4.4 , this proves that for each u ∈ D ( Δ ρ ) u \in D(\Delta^{\rho}) u ∈ D ( Δ ρ ) there exists a sequence ( u j ) j ⊂ C c ∞ ( R n ) (u_{j})_{j} \subset C^{\infty}_{c}(\mathbb{R}^{n}) ( u j ) j ⊂ C c ∞ ( R n ) such that u j → u u_{j} \to u u j → u , D 0 ρ u j → D ρ u D^{\rho}_{0}u_{j} \to D^{\rho}u D 0 ρ u j → D ρ u , and Δ 0 ρ u j → Δ ρ u \Delta^{\rho}_{0}u_{j} \to\Delta^{\rho}u Δ 0 ρ u j → Δ ρ u simultaneously in L p L^{p} L p -norm.
The Hilbert space case
In this section, we will write H ρ ( R n ) H^{\rho}(\mathbb{R}^{n}) H ρ ( R n ) for H ρ , 2 ( R n ) H^{\rho,2}(\mathbb{R}^{n}) H ρ , 2 ( R n ) to ease the notation. We have the following essential properties of the ρ \rho ρ -Laplacian Δ ρ \Delta^{\rho} Δ ρ introduced in Definition 4.1 when p = 2 p=2 p = 2 .
Proposition 4.9 (Basic properties of the ρ \rho ρ -Laplacian for p = 2 p=2 p = 2 ). Assume that ρ \rho ρ has compact support and satisfies (H0), (H1) and (H3’). Then Δ ρ \Delta^{\rho} Δ ρ has the following properties:
(i) Δ ρ \Delta^{\rho} Δ ρ is self-adjoint and dissipative.
(ii) For all u ∈ D ( Δ ρ ) u \in D(\Delta^{\rho}) u ∈ D ( Δ ρ ) and v ∈ H ρ ( R n ) v \in H^{\rho}(\mathbb{R}^{n}) v ∈ H ρ ( R n ) , it holds that
⟨ Δ ρ u , v ⟩ = − ⟨ D ρ u , D ρ v ⟩ . (83) \langle\Delta^{\rho}u,v\rangle=-\langle D^{\rho}u,D^{\rho}v\rangle. \tag*{(83)} ⟨ Δ ρ u , v ⟩ = − ⟨ D ρ u , D ρ v ⟩ . (83) (iii) The domain of Δ ρ \Delta^{\rho} Δ ρ fulfills
D ( Δ ρ ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ 2 u ^ ∈ L 2 ( R n ) } , D(\Delta^{\rho})=\{u\in L^{2}(\mathbb{R}^{n}):|\lambda_{\rho}|^{2}\widehat{u}\in L^{2}(\mathbb{R}^{n})\}, D ( Δ ρ ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ 2 u ∈ L 2 ( R n )} , and
Δ ρ u = F − 1 ( − ∣ λ ρ ∣ 2 u ^ ) , u ∈ D ( Δ ρ ) . (84) \Delta^{\rho}u=\mathcal{F}^{-1}(-|\lambda_{\rho}|^{2}\widehat{u}),\qquad u\in D(\Delta^{\rho}). \tag*{(84)} Δ ρ u = F − 1 ( − ∣ λ ρ ∣ 2 u ) , u ∈ D ( Δ ρ ) . (84)
Proof. (i) and (ii). We start by showing that Δ ρ \Delta^{\rho} Δ ρ is self-adjoint. As shown in (63) , we have ⟨ Δ 0 ρ u , v ⟩ = ⟨ u , Δ 0 ρ v ⟩ \langle\Delta^{\rho}_{0}u,v\rangle=\langle u,\Delta^{\rho}_{0}v\rangle ⟨ Δ 0 ρ u , v ⟩ = ⟨ u , Δ 0 ρ v ⟩ for any u , φ ∈ C c ∞ ( R n ) u,\varphi\in C^{\infty}_{c}(\mathbb{R}^{n}) u , φ ∈ C c ∞ ( R n ) . This entails that Δ ρ \Delta^{\rho} Δ ρ is an extension of Δ 0 ρ \Delta^{\rho}_{0} Δ 0 ρ and that Δ 0 ρ \Delta^{\rho}_{0} Δ 0 ρ is symmetric. In addition, invoking the core property of Lemma 4.7 , we find that Δ ρ \Delta^{\rho} Δ ρ is symmetric as well. Let us demonstrate now that ( Δ 0 ρ ) ∗ = Δ ρ (\Delta^{\rho}_{0})^{*}=\Delta^{\rho} ( Δ 0 ρ ) ∗ = Δ ρ , where we understand Δ 0 ρ \Delta^{\rho}_{0} Δ 0 ρ as an unbounded operator in L 2 ( R n ) L^{2}(\mathbb{R}^{n}) L 2 ( R n ) with domain C c ∞ ( R n ) C^{\infty}_{c}(\mathbb{R}^{n}) C c ∞ ( R n ) . The adjoint ( Δ 0 ρ ) ∗ (\Delta^{\rho}_{0})^{*} ( Δ 0 ρ ) ∗ exists because C c ∞ ( R n ) C^{\infty}_{c}(\mathbb{R}^{n}) C c ∞ ( R n ) is dense in L 2 ( R n ) L^{2}(\mathbb{R}^{n}) L 2 ( R n ) . By [11 Section 2.4] and Riesz’ representation theorem, a function u ∈ L 2 ( R n ) u\in L^{2}(\mathbb{R}^{n}) u ∈ L 2 ( R n ) belongs to D ( ( Δ 0 ρ ) ∗ ) D((\Delta^{\rho}_{0})^{*}) D (( Δ 0 ρ ) ∗ ) if and only if there exists v ∈ L 2 ( R n ) v\in L^{2}(\mathbb{R}^{n}) v ∈ L 2 ( R n ) such that
⟨ u , Δ 0 ρ φ ⟩ = ⟨ v , φ ⟩ , φ ∈ C c ∞ ( R n ) , \langle u,\Delta^{\rho}_{0}\varphi\rangle=\langle v,\varphi\rangle,\qquad\varphi\in C^{\infty}_{c}(\mathbb{R}^{n}), ⟨ u , Δ 0 ρ φ ⟩ = ⟨ v , φ ⟩ , φ ∈ C c ∞ ( R n ) ,
which, according to Definition 4.1 , holds if and only if u ∈ D ( Δ ρ ) u\in D(\Delta^{\rho}) u ∈ D ( Δ ρ ) . Moreover, in this case, ( Δ 0 ρ ) ∗ u = v = Δ ρ u (\Delta^{\rho}_{0})^{*}u=v=\Delta^{\rho}u ( Δ 0 ρ ) ∗ u = v = Δ ρ u . Together with the symmetry of Δ ρ \Delta^{\rho} Δ ρ it follows Δ ρ ⊂ ( Δ ρ ) ∗ ⊂ ( Δ 0 ρ ) ∗ = Δ ρ \Delta^{\rho}\subset(\Delta^{\rho})^{*}\subset(\Delta^{\rho}_{0})^{*}=\Delta^{\rho} Δ ρ ⊂ ( Δ ρ ) ∗ ⊂ ( Δ 0 ρ ) ∗ = Δ ρ ; in other words, Δ ρ \Delta^{\rho} Δ ρ is self-adjoint.
We are now in the position to prove (ii). Taking u ∈ D ( Δ ρ ) u\in D(\Delta^{\rho}) u ∈ D ( Δ ρ ) , we find with the help of Lemma 4.7 , Remark 4.8a) and c) a sequence ( u j ) ⊂ C c ∞ ( R n ) (u_{j})\subset C^{\infty}_{c}(\mathbb{R}^{n}) ( u j ) ⊂ C c ∞ ( R n ) with the property that u j → u u_{j}\to u u j → u , D 0 ρ u j → D ρ u D^{\rho}_{0}u_{j}\to D^{\rho}u D 0 ρ u j → D ρ u , and Δ 0 ρ u j → Δ ρ u \Delta^{\rho}_{0}u_{j}\to\Delta^{\rho}u Δ 0 ρ u j → Δ ρ u in L 2 ( R n ) L^{2}(\mathbb{R}^{n}) L 2 ( R n ) . We conclude that
⟨ Δ ρ u , v ⟩ = lim j → ∞ ⟨ Δ 0 ρ u j , v ⟩ = lim j → ∞ ⟨ div 0 ρ D 0 ρ u j , v ⟩ = − lim j → ∞ ⟨ D 0 ρ u j , D ρ v ⟩ = − ⟨ D ρ u , D ρ v ⟩ . \langle\Delta^{\rho}u,v\rangle=\lim_{j\to\infty}\langle\Delta^{\rho}_{0}u_{j},v\rangle=\lim_{j\to\infty}\langle\operatorname{div}^{\rho}_{0}D^{\rho}_{0}u_{j},v\rangle=-\lim_{j\to\infty}\langle D^{\rho}_{0}u_{j},D^{\rho}v\rangle=-\langle D^{\rho}u,D^{\rho}v\rangle. ⟨ Δ ρ u , v ⟩ = j → ∞ lim ⟨ Δ 0 ρ u j , v ⟩ = j → ∞ lim ⟨ div 0 ρ D 0 ρ u j , v ⟩ = − j → ∞ lim ⟨ D 0 ρ u j , D ρ v ⟩ = − ⟨ D ρ u , D ρ v ⟩ .
This implies (83) and (ii). In order to conclude the proof of (i), it remains to show that Δ ρ \Delta^{\rho} Δ ρ is dissipative. But it follows immediately from (83) that
⟨ Δ 0 ρ u j , u j ⟩ = − ∥ D ρ u j ∥ L 2 ( R n ) 2 ≤ 0. (85) \langle\Delta^{\rho}_{0}u_{j},u_{j}\rangle=-\|D^{\rho}u_{j}\|^{2}_{L^{2}(\mathbb{R}^{n})}\leq0. \tag*{(85)} ⟨ Δ 0 ρ u j , u j ⟩ = − ∥ D ρ u j ∥ L 2 ( R n ) 2 ≤ 0. (85)
By Lemma 4.7 this carries over to general u ∈ D ( Δ ρ ) u\in D(\Delta^{\rho}) u ∈ D ( Δ ρ ) , and, hence, Δ ρ \Delta^{\rho} Δ ρ is indeed dissipative.
(iii). We define in L 2 ( R n ) L^{2}(\mathbb{R}^{n}) L 2 ( R n ) the unbounded operator A A A with domain D ( A ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ 2 u ^ ∈ L 2 ( R n ) } D(A)=\{u\in L^{2}(\mathbb{R}^{n}):|\lambda_{\rho}|^{2}\widehat{u}\in L^{2}(\mathbb{R}^{n})\} D ( A ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ 2 u ∈ L 2 ( R n )} as A u : = F − 1 ( − ∣ λ ρ ∣ 2 u ^ ) Au:=\mathcal{F}^{-1}(-|\lambda_{\rho}|^{2}\widehat{u}) A u := F − 1 ( − ∣ λ ρ ∣ 2 u ) , u ∈ D ( A ) u\in D(A) u ∈ D ( A ) . By properties of the Fourier transform, the operator A A A is unitarily equivalent to the Fourier multiplication operator with multiplier − ∣ λ ρ ∣ 2 -|\lambda_{\rho}|^{2} − ∣ λ ρ ∣ 2 . Hence, as − ∣ λ ρ ∣ 2 -|\lambda_\rho|^2 − ∣ λ ρ ∣ 2 is real, the operator A A A is self-adjoint (see, e.g., [28 Section VIII.3, Proposition 1] ).
Now, for each u ∈ C ∞ ( R n ) u \in C^\infty(\mathbb{R}^n) u ∈ C ∞ ( R n ) , we have that u ^ ∈ S ( R n ) \widehat{u} \in\mathcal{S}(\mathbb{R}^n) u ∈ S ( R n ) , so ∣ λ ρ ∣ 2 u ^ ∈ L 2 ( R n ) |\lambda_\rho|^2\widehat{u} \in L^2(\mathbb{R}^n) ∣ λ ρ ∣ 2 u ∈ L 2 ( R n ) because λ ρ \lambda_\rho λ ρ has linear growth according to Lemma 2.9. In addition, F ( Δ 0 ρ u ) = − ∣ λ ρ ∣ 2 u ^ \mathcal{F}(\Delta^\rho_0u)=-|\lambda_\rho|^2\widehat{u} F ( Δ 0 ρ u ) = − ∣ λ ρ ∣ 2 u according to (60) . Therefore, Δ 0 ρ \Delta^\rho_0 Δ 0 ρ and A A A coincide in C ∞ ( R n ) C^\infty(\mathbb{R}^n) C ∞ ( R n ) . As A A A is a self-adjoint extension of the symmetric operator Δ 0 ρ \Delta^\rho_0 Δ 0 ρ , we obtain that ( Δ 0 ρ ) ∗ (\Delta^\rho_0)^* ( Δ 0 ρ ) ∗ is an extension of A A A , so, by Step 1, Δ ρ \Delta^\rho Δ ρ is an extension of A A A .
Finally, we show that D ( Δ ρ ) ⊂ D ( A ) D(\Delta^\rho) \subset D(A) D ( Δ ρ ) ⊂ D ( A ) . Since F ( Δ 0 ρ u ) = − ∣ λ ρ ∣ 2 u ^ \mathcal{F}(\Delta^\rho_0u)=-|\lambda_\rho|^2\widehat{u} F ( Δ 0 ρ u ) = − ∣ λ ρ ∣ 2 u for u ∈ C c ∞ ( R n ) u \in C^\infty_c(\mathbb{R}^n) u ∈ C c ∞ ( R n ) , it follows for u ∈ D ( Δ ρ ) u \in D(\Delta^\rho) u ∈ D ( Δ ρ ) with the approximating sequence provided by Lemma 4.7 that F ( Δ 0 ρ u ) = − ∣ λ ρ ∣ 2 u ^ \mathcal{F}(\Delta^\rho_0u)=-|\lambda_\rho|^2\widehat{u} F ( Δ 0 ρ u ) = − ∣ λ ρ ∣ 2 u a.e. and, in particular, that ∣ λ ρ ∣ 2 u ^ ∈ L 2 ( R n ) |\lambda_\rho|^2\widehat{u} \in L^2(\mathbb{R}^n) ∣ λ ρ ∣ 2 u ∈ L 2 ( R n ) . Thus, u ∈ D ( A ) u \in D(A) u ∈ D ( A ) . We conclude that Δ ρ = A \Delta^\rho=A Δ ρ = A as unbounded operators. This proves (iii). □
Remark 4.10. a) In the language introduced in Section 2.3, the statement of Proposition 4.9 (iii) can be equivalently expressed as Δ ρ = F − 1 ∣ λ ρ ∣ 2 \Delta^\rho=\mathcal{F}^{-1}|\lambda_\rho|^2 Δ ρ = F − 1 ∣ λ ρ ∣ 2 . We show below in Proposition 4.13 that this identity extends to general p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) .
b) An alternative proof of the self-adjointness of Δ ρ \Delta^\rho Δ ρ can also be obtained following the chain of reasoning in [11 Section 2.6.1] and arguing via the Lax-Milgram theorem in H ρ ( R n ) H^\rho(\mathbb{R}^n) H ρ ( R n ) .
In view of Proposition 4.9, it follows from a well-known result (see, e.g., [31 Theorem 13.31] ) that there exists a unique self-adjoint, positive semidefinite, unbounded operator in L 2 ( R n ) L^2(\mathbb{R}^n) L 2 ( R n ) whose square is − Δ ρ -\Delta^\rho − Δ ρ . It will be denoted by ( − Δ ρ ) 1 / 2 (-\Delta^\rho)^{1/2} ( − Δ ρ ) 1/2 . Moreover, ( − Δ ρ ) 1 / 2 (-\Delta^\rho)^{1/2} ( − Δ ρ ) 1/2 can be identified with the unbounded Fourier multiplication operator with symbol ∣ λ ρ ∣ |\lambda_\rho| ∣ λ ρ ∣ . While the latter property can also be derived from the multiplication operator form of the spectral theorem and the associated functional calculus (see, e.g., [28 ] ), we provide, for the reader’s convenience, a brief self-contained proof of this fact.
Lemma 4.11 (Square root of − Δ ρ -\Delta^\rho − Δ ρ ). Assume that ρ \rho ρ has compact support and satisfies (H0), (H1) and (H3′). Then
D ( ( − Δ ρ ) 1 / 2 ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ u ^ ∈ L 2 ( R n ) } D((-\Delta^\rho)^{1/2})=\{u\in L^2(\mathbb{R}^n):|\lambda_\rho|\widehat{u}\in L^2(\mathbb{R}^n)\} D (( − Δ ρ ) 1/2 ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ u ∈ L 2 ( R n )} and
( − Δ ρ ) 1 / 2 u = F − 1 ( ∣ λ ρ ∣ u ^ ) , u ∈ D ( ( − Δ ρ ) 1 / 2 ) . (-\Delta^\rho)^{1/2}u=\mathcal{F}^{-1}(|\lambda_\rho|\widehat{u}),\qquad u\in D((-\Delta^\rho)^{1/2}). ( − Δ ρ ) 1/2 u = F − 1 ( ∣ λ ρ ∣ u ) , u ∈ D (( − Δ ρ ) 1/2 ) . Proof. We define the operator B B B in L 2 ( R n ) L^2(\mathbb{R}^n) L 2 ( R n ) with domain D ( B ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ u ^ ∈ L 2 ( R n ) } D(B)=\{u\in L^2(\mathbb{R}^n):|\lambda_\rho|\widehat{u}\in L^2(\mathbb{R}^n)\} D ( B ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ u ∈ L 2 ( R n )} by B u : = F − 1 ( ∣ λ ρ ∣ u ^ ) Bu:=\mathcal{F}^{-1}(|\lambda_\rho|\widehat{u}) B u := F − 1 ( ∣ λ ρ ∣ u ) , u ∈ L 2 ( R n ) u\in L^2(\mathbb{R}^n) u ∈ L 2 ( R n ) , which is self-adjoint because ∣ λ ρ ∣ |\lambda_\rho| ∣ λ ρ ∣ is real (see, e.g., [28 Section VIII.3, Proposition 1] ). Since
⟨ B u , u ⟩ = ⟨ ∣ λ ρ ∣ u ^ , u ^ ⟩ = ∫ R n ∣ λ ρ ( ξ ) ∣ ∣ u ^ ( ξ ) ∣ 2 d ξ ≥ 0 , u ∈ D ( B ) , \langle Bu,u\rangle=\langle|\lambda_\rho|\widehat{u},\widehat{u}\rangle=\int_{\mathbb{R}^n}|\lambda_\rho(\xi)||\widehat{u}(\xi)|^2d\xi\geq0,\qquad u\in D(B), ⟨ B u , u ⟩ = ⟨ ∣ λ ρ ∣ u , u ⟩ = ∫ R n ∣ λ ρ ( ξ ) ∣∣ u ( ξ ) ∣ 2 d ξ ≥ 0 , u ∈ D ( B ) , B B B is positive semidefinite. Its square B 2 B^2 B 2 is defined in
D ( B 2 ) = { u ∈ D ( B ) : B u ∈ D ( B ) } = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ u ^ ∈ L 2 ( R n ) , ∣ λ ρ ∣ 2 u ^ ∈ L 2 ( R n ) } . D(B^2)=\{u\in D(B):Bu\in D(B)\}=\{u\in L^2(\mathbb{R}^n):|\lambda_\rho|\widehat{u}\in L^2(\mathbb{R}^n),|\lambda_\rho|^2\widehat{u}\in L^2(\mathbb{R}^n)\}. D ( B 2 ) = { u ∈ D ( B ) : B u ∈ D ( B )} = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ u ∈ L 2 ( R n ) , ∣ λ ρ ∣ 2 u ∈ L 2 ( R n )} . Thanks to Hölder’s inequality, for any f ∈ L 2 ( R n ) f\in L^2(\mathbb{R}^n) f ∈ L 2 ( R n ) with ∣ λ ρ ∣ 2 f ∈ L 2 ( R n ) |\lambda_\rho|^2f\in L^2(\mathbb{R}^n) ∣ λ ρ ∣ 2 f ∈ L 2 ( R n ) we have
∥ ∣ λ ρ ∣ f ∥ L 2 ( R n ) ≤ ∥ ∣ f ∣ 1 / 2 ∥ L 4 ( R n ) ∥ ∣ λ ρ ∣ ∣ f ∣ 1 / 2 ∥ L 4 ( R n ) = ∥ f ∥ L 2 ( R n ) 1 / 2 ∥ ∣ λ ρ ∣ 2 f ∥ L 2 ( R n ) 1 / 2 . \bigl\||\lambda_\rho|f\bigr\|_{L^2(\mathbb{R}^n)}\leq\bigl\||f|^{1/2}\bigr\|_{L^4(\mathbb{R}^n)}\bigl\||\lambda_\rho||f|^{1/2}\bigr\|_{L^4(\mathbb{R}^n)}=\|f\|_{L^2(\mathbb{R}^n)}^{1/2}\bigl\||\lambda_\rho|^2f\bigr\|_{L^2(\mathbb{R}^n)}^{1/2}. ∣ λ ρ ∣ f L 2 ( R n ) ≤ ∣ f ∣ 1/2 L 4 ( R n ) ∣ λ ρ ∣∣ f ∣ 1/2 L 4 ( R n ) = ∥ f ∥ L 2 ( R n ) 1/2 ∣ λ ρ ∣ 2 f L 2 ( R n ) 1/2 . Therefore,
D ( B 2 ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ 2 u ^ ∈ L 2 ( R n ) } . D(B^2)=\{u\in L^2(\mathbb{R}^n):|\lambda_\rho|^2\widehat{u}\in L^2(\mathbb{R}^n)\}. D ( B 2 ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ 2 u ∈ L 2 ( R n )} . This operator acts, for each u ∈ D ( B 2 ) u\in D(B^2) u ∈ D ( B 2 ) , as
B 2 u = B ( F − 1 ( ∣ λ ρ ∣ u ^ ) ) = F − 1 ( ∣ λ ρ ∣ 2 u ^ ) . B^2u=B(\mathcal{F}^{-1}(|\lambda_\rho|\widehat{u}))=\mathcal{F}^{-1}(|\lambda_\rho|^2\widehat{u}). B 2 u = B ( F − 1 ( ∣ λ ρ ∣ u )) = F − 1 ( ∣ λ ρ ∣ 2 u ) . By Proposition 4.9 (iii), B 2 = − Δ ρ B^2=-\Delta^\rho B 2 = − Δ ρ . Because B B B is self-adjoint and positive semidefinite according to Proposition 4.9 (i), by uniqueness (see, e.g., [31 Theorem 13.31] ), B = ( − Δ ρ ) 1 / 2 B=(-\Delta^\rho)^{1/2} B = ( − Δ ρ ) 1/2 . □
We have prepared the following characterization of H ρ ( R n ) H^\rho(\mathbb{R}^n) H ρ ( R n ) in terms of the square root of the generalized ρ \rho ρ -Laplace operator. Our proof is based on the Fourier characterization of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) obtained in Theorem 3.1
Proposition 4.12. Assume that ρ \rho ρ has compact support and satisfies ( H 0 ) (H0) ( H 0 ) –( H 2 ) (H2) ( H 2 ) and ( H 3 ′ ) (H3') ( H 3 ′ ) . Then
H ρ ( R n ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ u ^ ∈ L 2 ( R n ) } = D ( ( − Δ ρ ) 1 / 2 ) . H^\rho(\mathbb{R}^n)=\{u\in L^2(\mathbb{R}^n):|\lambda_\rho|\hat{u}\in L^2(\mathbb{R}^n)\}=D((-\Delta^\rho)^{1/2}). H ρ ( R n ) = { u ∈ L 2 ( R n ) : ∣ λ ρ ∣ u ^ ∈ L 2 ( R n )} = D (( − Δ ρ ) 1/2 ) . Proof. Let u ∈ H ρ ( R n ) u\in H^\rho(\mathbb{R}^n) u ∈ H ρ ( R n ) . Then u ∈ L 2 ( R n ) u\in L^2(\mathbb{R}^n) u ∈ L 2 ( R n ) and, by Theorem 3.1, F − 1 ( J λ ρ K u ^ ) ∈ L 2 ( R n ) \mathcal{F}^{-1}(J_{\lambda_\rho K}\hat{u})\in L^2(\mathbb{R}^n) F − 1 ( J λ ρ K u ^ ) ∈ L 2 ( R n ) . By Proposition 2.10 (i), F − 1 ( ∣ λ ρ ∣ u ^ ) ∈ L 2 ( R n ) \mathcal{F}^{-1}(|\lambda_\rho|\hat{u})\in L^2(\mathbb{R}^n) F − 1 ( ∣ λ ρ ∣ u ^ ) ∈ L 2 ( R n ) and, hence, ∣ λ ρ ∣ u ^ ∈ L 2 ( R n ) |\lambda_\rho|\hat{u}\in L^2(\mathbb{R}^n) ∣ λ ρ ∣ u ^ ∈ L 2 ( R n ) . Conversely, let u ∈ L 2 ( R n ) u\in L^2(\mathbb{R}^n) u ∈ L 2 ( R n ) satisfy ∣ λ ρ ∣ u ^ ∈ L 2 ( R n ) |\lambda_\rho|\hat{u}\in L^2(\mathbb{R}^n) ∣ λ ρ ∣ u ^ ∈ L 2 ( R n ) . Then, F − 1 ( ∣ λ ρ ∣ u ^ ) ∈ L 2 ( R n ) \mathcal{F}^{-1}(|\lambda_\rho|\hat{u})\in L^2(\mathbb{R}^n) F − 1 ( ∣ λ ρ ∣ u ^ ) ∈ L 2 ( R n ) . As u ∈ L 2 ( R n ) u\in L^2(\mathbb{R}^n) u ∈ L 2 ( R n ) , then F − 1 ( ( 1 + ∣ λ ρ ∣ ) u ^ ) ∈ L 2 ( R n ) \mathcal{F}^{-1}((1+|\lambda_\rho|)\hat{u})\in L^2(\mathbb{R}^n) F − 1 (( 1 + ∣ λ ρ ∣ ) u ^ ) ∈ L 2 ( R n ) . By Proposition 2.10 (ii), F − 1 ( J λ ρ K u ^ ) ∈ L 2 ( R n ) \mathcal{F}^{-1}(J_{\lambda_\rho K}\hat{u})\in L^2(\mathbb{R}^n) F − 1 ( J λ ρ K u ^ ) ∈ L 2 ( R n ) . The first equality of the statement is proved; the second follows from Lemma 4.11.
The general-p p p case
The theory of square roots of unbounded operators in Banach spaces is considerably more involved than in Hilbert spaces (see, e.g., [19 ] ) and lies beyond the scope of this work. Yet, we can define the ρ \rho ρ -Laplacian and its square root in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) in a rather canonical way. For the rest of the section, we fix p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and, as for p = 2 p=2 p = 2 in the previous section, turn to the problem of representing H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) in terms of the domain of ( − Δ ρ ) 1 / 2 (-\Delta^\rho)^{1/2} ( − Δ ρ ) 1/2 .
We remind the reader of the terminology and basic properties introduced in Section 2.6 of unbounded operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) defined through a multiplier in Fourier space and closable unbounded operators (especially, Definition 2.2 and the results thereafter). Also recall that, as stated in Proposition 2.10, for m m m being any of the four functions ∣ λ ρ ∣ / J λ ρ K |\lambda_\rho|/J_{\lambda_\rho K} ∣ λ ρ ∣/ J λ ρ K , J λ ρ K / ( 1 + ∣ λ ρ ∣ ) J_{\lambda_\rho K}/(1+|\lambda_\rho|) J λ ρ K / ( 1 + ∣ λ ρ ∣ ) , 1 / J λ ρ K 1/J_{\lambda_\rho K} 1/ J λ ρ K or 1 / J λ ρ K 2 1/J_{\lambda_\rho K}^2 1/ J λ ρ K 2 , the corresponding closure F ‾ m \overline{\mathcal{F}}_m F m is a bounded linear operator in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) .
The following proposition is the general counterpart of Proposition 4.9 and collects again the main properties of the ρ \rho ρ -Laplacian. It varies from the result in the Hilbert space case, since there is no concept of self-adjointness in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) .
Proposition 4.13 (Basic properties of the ρ \rho ρ -Laplacian). Assume that ρ \rho ρ has compact support and satisfies ( H 0 ) (H0) ( H 0 ) –( H 3 ) (H3) ( H 3 ) . The operator Δ ρ \Delta^\rho Δ ρ has the following properties:
(i) The closure of Δ 0 ρ \Delta^\rho_0 Δ 0 ρ in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) is Δ ρ \Delta^\rho Δ ρ .
(ii) For all u ∈ D ( Δ ρ ) u\in D(\Delta^\rho) u ∈ D ( Δ ρ ) in the sense of L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) and v ∈ D ( Δ ρ ) v\in D(\Delta^\rho) v ∈ D ( Δ ρ ) in the sense of L p ′ ( R n ) L^{p'}(\mathbb{R}^n) L p ′ ( R n ) it holds
⟨ Δ ρ u , v ⟩ = ⟨ u , Δ ρ v ⟩ . (86) \langle\Delta^\rho u,v\rangle=\langle u,\Delta^\rho v\rangle. \tag*{(86)} ⟨ Δ ρ u , v ⟩ = ⟨ u , Δ ρ v ⟩ . (86) (iii) It holds Δ ρ = F ‾ − ∣ λ ρ ∣ 2 \Delta^\rho=\overline{\mathcal{F}}_{-|\lambda_\rho|^2} Δ ρ = F − ∣ λ ρ ∣ 2 in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , in particular,
Δ ρ u = F − 1 ( − ∣ λ ρ ∣ 2 u ^ ) , u ∈ S ( R n ) . (87) \Delta^\rho u=\mathcal{F}^{-1}(-|\lambda_\rho|^2\hat{u}),\qquad u\in\mathcal{S}(\mathbb{R}^n). \tag*{(87)} Δ ρ u = F − 1 ( − ∣ λ ρ ∣ 2 u ^ ) , u ∈ S ( R n ) . (87)
Proof. (i). We show first that Δ 0 ρ \Delta^\rho_0 Δ 0 ρ is closable as an operator in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Let ( u j ) j (u_j)_j ( u j ) j be a sequence in C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) such that u j → 0 u_j\to0 u j → 0 in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) and Δ 0 ρ u j → v \Delta^\rho_0u_j\to v Δ 0 ρ u j → v in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) for some v ∈ L p ( R n ) v\in L^p(\mathbb{R}^n) v ∈ L p ( R n ) . Then, for all φ ∈ C c ∞ ( R n ) \varphi\in C^\infty_c(\mathbb{R}^n) φ ∈ C c ∞ ( R n ) ,
⟨ Δ 0 ρ u j , φ ⟩ → ⟨ v , φ ⟩ and ⟨ Δ 0 ρ u j , φ ⟩ = ⟨ u j , Δ 0 ρ φ ⟩ → 0 , \langle\Delta^\rho_0u_j,\varphi\rangle\to\langle v,\varphi\rangle\quad\text{and}\quad\langle\Delta^\rho_0u_j,\varphi\rangle=\langle u_j,\Delta^\rho_0\varphi\rangle\to0, ⟨ Δ 0 ρ u j , φ ⟩ → ⟨ v , φ ⟩ and ⟨ Δ 0 ρ u j , φ ⟩ = ⟨ u j , Δ 0 ρ φ ⟩ → 0 ,
so ⟨ v , φ ⟩ = 0 \langle v,\varphi\rangle=0 ⟨ v , φ ⟩ = 0 , which implies that v = 0 v=0 v = 0 .
It remains to verify that the closure of Δ 0 ρ \Delta^\rho_0 Δ 0 ρ in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) is Δ ρ \Delta^\rho Δ ρ . Let therefore Δ 0 ρ ‾ \overline{\Delta^\rho_0} Δ 0 ρ denote the closure of Δ 0 ρ \Delta^\rho_0 Δ 0 ρ in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . We start by proving that D ( Δ 0 ρ ‾ ) ⊂ D ( Δ ρ ) D(\overline{\Delta^\rho_0})\subset D(\Delta^\rho) D ( Δ 0 ρ ) ⊂ D ( Δ ρ ) and that Δ ρ \Delta^\rho Δ ρ extends Δ 0 ρ ‾ \overline{\Delta^\rho_0} Δ 0 ρ . Let ( u j ) j (u_j)_j ( u j ) j be a sequence in C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) such that u j → u u_j\to u u j → u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) and Δ 0 ρ u j → w \Delta^\rho_0u_j\to w Δ 0 ρ u j → w in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) for some u , w ∈ L p ( R n ) u,w\in L^p(\mathbb{R}^n) u , w ∈ L p ( R n ) . Then
⟨ u , Δ 0 ρ φ ⟩ = lim j → ∞ ⟨ u j , Δ 0 ρ φ ⟩ = lim j → ∞ ⟨ Δ 0 ρ u j , φ ⟩ = ⟨ w , φ ⟩ , φ ∈ C c ∞ ( R n ) , \langle u,\Delta^\rho_0\varphi\rangle=\lim_{j\to\infty}\langle u_j,\Delta^\rho_0\varphi\rangle=\lim_{j\to\infty}\langle\Delta^\rho_0u_j,\varphi\rangle=\langle w,\varphi\rangle,\qquad\varphi\in C^\infty_c(\mathbb{R}^n), ⟨ u , Δ 0 ρ φ ⟩ = j → ∞ lim ⟨ u j , Δ 0 ρ φ ⟩ = j → ∞ lim ⟨ Δ 0 ρ u j , φ ⟩ = ⟨ w , φ ⟩ , φ ∈ C c ∞ ( R n ) ,
so u ∈ D ( Δ ρ ) u\in D(\Delta^\rho) u ∈ D ( Δ ρ ) and Δ ρ u = w = Δ 0 ρ ‾ u \Delta^\rho u=w=\overline{\Delta^\rho_0}u Δ ρ u = w = Δ 0 ρ u . Thus, Δ ρ \Delta^\rho Δ ρ is an extension of Δ 0 ρ ‾ \overline{\Delta^\rho_0} Δ 0 ρ .
The inclusion D ( Δ ρ ) ⊂ D ( Δ 0 ρ ‾ ) D(\Delta^\rho)\subset D(\overline{\Delta^\rho_0}) D ( Δ ρ ) ⊂ D ( Δ 0 ρ ) and the fact that Δ 0 ρ ‾ \overline{\Delta^\rho_0} Δ 0 ρ extends Δ ρ \Delta^\rho Δ ρ follow directly from the statement of Lemma 4.7. This proves part ( i ) (i) ( i ) .
(ii). Let u ∈ D ( Δ ρ ) u\in D(\Delta^\rho) u ∈ D ( Δ ρ ) in the sense of L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) and v ∈ D ( Δ ρ ) v\in D(\Delta^\rho) v ∈ D ( Δ ρ ) in the sense of L p ′ ( R n ) L^{p'}(\mathbb{R}^n) L p ′ ( R n ) . By the core property in Lemma 4.7, there exists a sequence ( v j ) j ⊂ C c ∞ ( R n ) (v_j)_j\subset C^\infty_c(\mathbb{R}^n) ( v j ) j ⊂ C c ∞ ( R n ) such that v j → v v_j\to v v j → v , Δ 0 ρ v j → Δ ρ v \Delta^\rho_0v_j\to\Delta^\rho v Δ 0 ρ v j → Δ ρ v in L p ′ ( R n ) L^{p'}(\mathbb{R}^n) L p ′ ( R n ) . It follows that
⟨ Δ ρ u , v ⟩ = lim j → ∞ ⟨ Δ ρ u , v j ⟩ = lim j → ∞ ⟨ u , Δ 0 ρ v j ⟩ = ⟨ u , Δ ρ v ⟩ . \langle\Delta^\rho u,v\rangle=\lim_{j\to\infty}\langle\Delta^\rho u,v_j\rangle=\lim_{j\to\infty}\langle u,\Delta^\rho_0v_j\rangle=\langle u,\Delta^\rho v\rangle. ⟨ Δ ρ u , v ⟩ = j → ∞ lim ⟨ Δ ρ u , v j ⟩ = j → ∞ lim ⟨ u , Δ 0 ρ v j ⟩ = ⟨ u , Δ ρ v ⟩ .
This proves (4.28) and thus, ( i i ) (ii) ( ii ) .
:::
(iii). Note that the closure F − ∣ λ ρ ∣ 2 ‾ = − F ∣ λ ρ ∣ 2 ‾ \overline{F_{-|\lambda_\rho|^2}}=-\overline{F_{|\lambda_\rho|^2}} F − ∣ λ ρ ∣ 2 = − F ∣ λ ρ ∣ 2 is well defined thanks to Lemma 2.11, which also implies that D ( F − ∣ λ ρ ∣ 2 ) = S ( R n ) D(F_{-|\lambda_\rho|^2})=\mathcal{S}(\mathbb{R}^n) D ( F − ∣ λ ρ ∣ 2 ) = S ( R n ) . Comparing this with (4.2), we observe that Δ 0 ρ \Delta_0^\rho Δ 0 ρ and F − ∣ λ ρ ∣ 2 F_{-|\lambda_\rho|^2} F − ∣ λ ρ ∣ 2 coincide in C c ∞ ( R n ) C_c^\infty(\mathbb{R}^n) C c ∞ ( R n ) . Since, as was shown in (i), the closure of Δ 0 ρ \Delta_0^\rho Δ 0 ρ in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) with respect to C c ∞ ( R n ) C_c^\infty(\mathbb{R}^n) C c ∞ ( R n ) is exactly Δ ρ \Delta^\rho Δ ρ , it follows with the inclusion C c ∞ ( R n ) ⊂ S ( R n ) C_c^\infty(\mathbb{R}^n)\subset\mathcal{S}(\mathbb{R}^n) C c ∞ ( R n ) ⊂ S ( R n ) that D ( Δ ρ ) ⊂ D ( F − ∣ λ ρ ∣ 2 ‾ ) D(\Delta^\rho)\subset D(\overline{F_{-|\lambda_\rho|^2}}) D ( Δ ρ ) ⊂ D ( F − ∣ λ ρ ∣ 2 ) and that F − ∣ λ ρ ∣ 2 ‾ \overline{F_{-|\lambda_\rho|^2}} F − ∣ λ ρ ∣ 2 extends Δ ρ \Delta^\rho Δ ρ .
To see the reverse connection, let ( u j ) j ⊂ S ( R n ) (u_j)_j\subset\mathcal{S}(\mathbb{R}^n) ( u j ) j ⊂ S ( R n ) and u ∈ L p ( R n ) u\in L^p(\mathbb{R}^n) u ∈ L p ( R n ) such that u j → u u_j\to u u j → u and F − ∣ λ ρ ∣ 2 ‾ u j → F − ∣ λ ρ ∣ 2 ‾ u \overline{F_{-|\lambda_\rho|^2}}u_j\to\overline{F_{-|\lambda_\rho|^2}}u F − ∣ λ ρ ∣ 2 u j → F − ∣ λ ρ ∣ 2 u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . By identities (2.5) and (2.10) and the comments thereafter, it holds (4.29) and Δ ρ u j = div ρ 0 D ρ 0 u j = F − ∣ λ ρ ∣ 2 u j \Delta^\rho u_j=\operatorname{div}_\rho^0D_\rho^0u_j=F_{-|\lambda_\rho|^2}u_j Δ ρ u j = div ρ 0 D ρ 0 u j = F − ∣ λ ρ ∣ 2 u j , and, similarly to the proof of (i),
⟨ u , Δ 0 ρ φ ⟩ = lim j → ∞ ⟨ u j , Δ 0 ρ φ ⟩ = lim j → ∞ ⟨ Δ ρ u j , φ ⟩ = lim j → ∞ ⟨ F − ∣ λ ρ ∣ 2 u j , φ ⟩ = ⟨ F − ∣ λ ρ ∣ 2 ‾ u , φ ⟩ , φ ∈ C c ∞ ( R n ) . \langle u,\Delta_0^\rho\varphi\rangle=\lim_{j\to\infty}\langle u_j,\Delta_0^\rho\varphi\rangle=\lim_{j\to\infty}\langle\Delta^\rho u_j,\varphi\rangle=\lim_{j\to\infty}\langle F_{-|\lambda_\rho|^2}u_j,\varphi\rangle=\langle\overline{F_{-|\lambda_\rho|^2}}u,\varphi\rangle,\quad\varphi\in C_c^\infty(\mathbb{R}^n). ⟨ u , Δ 0 ρ φ ⟩ = j → ∞ lim ⟨ u j , Δ 0 ρ φ ⟩ = j → ∞ lim ⟨ Δ ρ u j , φ ⟩ = j → ∞ lim ⟨ F − ∣ λ ρ ∣ 2 u j , φ ⟩ = ⟨ F − ∣ λ ρ ∣ 2 u , φ ⟩ , φ ∈ C c ∞ ( R n ) .
This entails that Δ ρ u \Delta^\rho u Δ ρ u exists in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) and simultaneously that D ( F − ∣ λ ρ ∣ 2 ‾ ) ⊂ D ( Δ ρ ) D(\overline{F_{-|\lambda_\rho|^2}})\subset D(\Delta^\rho) D ( F − ∣ λ ρ ∣ 2 ) ⊂ D ( Δ ρ ) with Δ ρ = F − ∣ λ ρ ∣ 2 ‾ \Delta^\rho=\overline{F_{-|\lambda_\rho|^2}} Δ ρ = F − ∣ λ ρ ∣ 2 on D ( F − ∣ λ ρ ∣ 2 ‾ ) D(\overline{F_{-|\lambda_\rho|^2}}) D ( F − ∣ λ ρ ∣ 2 ) . We have thus demonstrated (4.29) and (iii), where we note that (4.29) alone already follows from (4.6).
In what follows, our goal is to show that F ∣ λ ρ ∣ ‾ \overline{F_{|\lambda_\rho|}} F ∣ λ ρ ∣ defines a square root of − Δ ρ -\Delta^\rho − Δ ρ , generalizing the result of Lemma 4.11. The proof is more delicate than in the Hilbert space case, and we will establish several auxiliary results building up to it. With Proposition 4.13 (iii) in mind, it will suffice to establish that F ∣ λ ρ ∣ ‾ 2 = F ∣ λ ρ ∣ 2 \overline{F_{|\lambda_\rho|}}^2=F_{|\lambda_\rho|^2} F ∣ λ ρ ∣ 2 = F ∣ λ ρ ∣ 2 as operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) .
First, we will show the representation of the nonlocal Sobolev space H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) as the domain of the operators F [ λ ρ ] ‾ \overline{F_{[\lambda_\rho]}} F [ λ ρ ] and F ∣ λ ρ ∣ ‾ \overline{F_{|\lambda_\rho|}} F ∣ λ ρ ∣ , respectively. The result is based on the Fourier characterization in Theorem 3.1 and the density of smooth compactly supported functions in H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) .
Lemma 4.14. Assume that ρ \rho ρ has compact support and fulfills hypotheses (H0)–(H2) and (H3'). Then it holds
H ρ , p ( R n ) = D ( F [ λ ρ ] ‾ ) = D ( F ∣ λ ρ ∣ ‾ ) . H^{\rho,p}(\mathbb{R}^n)=D(\overline{F_{[\lambda_\rho]}})=D(\overline{F_{|\lambda_\rho|}}). H ρ , p ( R n ) = D ( F [ λ ρ ] ) = D ( F ∣ λ ρ ∣ ) .
Proof. The proof will be given in two steps, corresponding to the two claimed identities.
Step 1. We start by showing that H ρ , p ( R n ) = D ( F [ λ ρ ] ‾ ) H^{\rho,p}(\mathbb{R}^n)=D(\overline{F_{[\lambda_\rho]}}) H ρ , p ( R n ) = D ( F [ λ ρ ] ) . We use that H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) in the characterization of Theorem 3.1 can be written as
H ρ , p ( R n ) = { u ∈ L p ( R n ) : F − 1 ( [ λ ρ ] u ^ ) ∈ L p ( R n ) } , (88) H^{\rho,p}(\mathbb{R}^n)=\{u\in L^p(\mathbb{R}^n):\mathcal{F}^{-1}([\lambda_\rho]\widehat{u})\in L^p(\mathbb{R}^n)\}, \tag*{(88)} H ρ , p ( R n ) = { u ∈ L p ( R n ) : F − 1 ([ λ ρ ] u ) ∈ L p ( R n )} , (88)
where F − 1 ( [ λ ρ ] u ^ ) \mathcal{F}^{-1}([\lambda_\rho]\widehat{u}) F − 1 ([ λ ρ ] u ) is understood in the sense of S ′ ( R n ) \mathcal{S}'(\mathbb{R}^n) S ′ ( R n ) . In view of (4.30), the inclusion H ρ , p ( R n ) ⊂ D ( F [ λ ρ ] ‾ ) H^{\rho,p}(\mathbb{R}^n)\subset D(\overline{F_{[\lambda_\rho]}}) H ρ , p ( R n ) ⊂ D ( F [ λ ρ ] ) follows from the fact that C c ∞ ( R n ) C_c^\infty(\mathbb{R}^n) C c ∞ ( R n ) is dense in H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) (see [8 Theorem 3.9] ). Indeed, the density result and the equivalent norm characterization of Theorem 3.1 imply that for every u ∈ H ρ , p ( R n ) u\in H^{\rho,p}(\mathbb{R}^n) u ∈ H ρ , p ( R n ) , there exists ( u j ) j ⊂ C c ∞ ( R n ) ⊂ S ( R n ) (u_j)_j\subset C_c^\infty(\mathbb{R}^n)\subset\mathcal{S}(\mathbb{R}^n) ( u j ) j ⊂ C c ∞ ( R n ) ⊂ S ( R n ) such that
F [ λ ρ ] u j = F − 1 ( [ λ ρ ] u j ^ ) → F − 1 ( [ λ ρ ] u ^ ) in L p ( R n ) , F_{[\lambda_\rho]}u_j=\mathcal{F}^{-1}([\lambda_\rho]\widehat{u_j})\to\mathcal{F}^{-1}([\lambda_\rho]\widehat{u})\quad\text{in }L^p(\mathbb{R}^n), F [ λ ρ ] u j = F − 1 ([ λ ρ ] u j ) → F − 1 ([ λ ρ ] u ) in L p ( R n ) ,
showing that u ∈ D ( F [ λ ρ ] ‾ ) u\in D(\overline{F_{[\lambda_\rho]}}) u ∈ D ( F [ λ ρ ] ) . The reverse inclusion is a mere consequence of the property that the tempered distribution F − 1 ( [ λ ρ ] u ^ ) \mathcal{F}^{-1}([\lambda_\rho]\widehat{u}) F − 1 ([ λ ρ ] u ) coincides with the function F [ λ ρ ] ‾ u \overline{F_{[\lambda_\rho]}}u F [ λ ρ ] u , whenever u ∈ D ( F [ λ ρ ] ‾ ) u\in D(\overline{F_{[\lambda_\rho]}}) u ∈ D ( F [ λ ρ ] ) , following from Lemma 2.4 for m = [ λ ρ ] m=[\lambda_\rho] m = [ λ ρ ] .
Step 2. We will show in this step that
D ( F [ λ ρ ] ‾ ) = D ( F ∣ λ ρ ∣ ‾ ) , (89) D(\overline{F_{[\lambda_\rho]}})=D(\overline{F_{|\lambda_\rho|}}), \tag*{(89)} D ( F [ λ ρ ] ) = D ( F ∣ λ ρ ∣ ) , (89)
which will follow essentially from the properties of the L p L^p L p -Fourier multipliers in Proposition 2.10. To be precise, by applying Lemma 2.6 twice, first with m 1 = [ λ ρ ] m_1=[\lambda_\rho] m 1 = [ λ ρ ] and m 2 = ∣ λ ρ ∣ / [ λ ρ ] m_2=|\lambda_\rho|/[\lambda_\rho] m 2 = ∣ λ ρ ∣/ [ λ ρ ] (see Proposition 2.10 (i)), and then with m 1 = 1 + ∣ λ ρ ∣ m_1=1+|\lambda_\rho| m 1 = 1 + ∣ λ ρ ∣ and m 2 = [ λ ρ ] / ( 1 + ∣ λ ρ ∣ ) m_2=[\lambda_\rho]/(1+|\lambda_\rho|) m 2 = [ λ ρ ] / ( 1 + ∣ λ ρ ∣ ) (see Proposition 2.10 (ii)), we conclude that
D ( F [ λ ρ ] ‾ ) ⊂ D ( F ∣ λ ρ ∣ ‾ ) , D ( F 1 + ∣ λ ρ ∣ ‾ ) ⊂ D ( F [ λ ρ ] ‾ ) . (90) D(\overline{F_{[\lambda_\rho]}})\subset D(\overline{F_{|\lambda_\rho|}}),\qquad D(\overline{F_{1+|\lambda_\rho|}})\subset D(\overline{F_{[\lambda_\rho]}}). \tag*{(90)} D ( F [ λ ρ ] ) ⊂ D ( F ∣ λ ρ ∣ ) , D ( F 1 + ∣ λ ρ ∣ ) ⊂ D ( F [ λ ρ ] ) . (90)
All that remains is to prove the inclusion D ( F ∣ λ ρ ∣ ‾ ) ⊂ D ( F 1 + ∣ λ ρ ∣ ‾ ) D(\overline{F_{|\lambda_\rho|}})\subset D(\overline{F_{1+|\lambda_\rho|}}) D ( F ∣ λ ρ ∣ ) ⊂ D ( F 1 + ∣ λ ρ ∣ ) , and we will see that this is a simple consequence of the notion of domain. Indeed, let u ∈ D ( F ∣ λ ρ ∣ ‾ ) u\in D(\overline{F_{|\lambda_\rho|}}) u ∈ D ( F ∣ λ ρ ∣ ) , implying, by definition of F ‾ ∣ λ ρ ∣ \overline{\mathcal{F}}_{|\lambda_\rho|} F ∣ λ ρ ∣ , that there exists ( u j ) j ⊂ S ( R n ) (u_j)_j \subset\mathcal{S}(\mathbb{R}^n) ( u j ) j ⊂ S ( R n ) such that u j → u u_j \to u u j → u and F ‾ ∣ λ ρ ∣ u j → F ‾ ∣ λ ρ ∣ u \overline{\mathcal{F}}_{|\lambda_\rho|}u_j \to\overline{\mathcal{F}}_{|\lambda_\rho|}u F ∣ λ ρ ∣ u j → F ∣ λ ρ ∣ u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Together, it holds
F 1 + ∣ λ ρ ∣ u j = u j + F ‾ ∣ λ ρ ∣ u j → u + F ‾ ∣ λ ρ ∣ u = F ‾ 1 + ∣ λ ρ ∣ u in L p ( R n ) , (91) \mathcal{F}_{1+|\lambda_\rho|}u_j=u_j+\overline{\mathcal{F}}_{|\lambda_\rho|}u_j\to u+\overline{\mathcal{F}}_{|\lambda_\rho|}u=\overline{\mathcal{F}}_{1+|\lambda_\rho|}u \quad\text{in } L^p(\mathbb{R}^n), \tag*{(91)} F 1 + ∣ λ ρ ∣ u j = u j + F ∣ λ ρ ∣ u j → u + F ∣ λ ρ ∣ u = F 1 + ∣ λ ρ ∣ u in L p ( R n ) , (91)
showing u ∈ D ( F ‾ 1 + ∣ λ ρ ∣ ) u\in D(\overline{\mathcal{F}}_{1+|\lambda_\rho|}) u ∈ D ( F 1 + ∣ λ ρ ∣ ) . In combination with (90) , this yields (89) , as desired. □
We prove below that the unbounded operators F ‾ ∣ λ ρ ∣ \overline{\mathcal{F}}_{|\lambda_\rho|} F ∣ λ ρ ∣ and F ‾ ⟦ λ ρ ⟧ \overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket} F [ [ λ ρ ] ] commute with each other.
Lemma 4.15. Assume that ρ \rho ρ has compact support and fulfills hypotheses (H0)–(H2) and (H3’). Then
F ‾ ⟦ λ ρ ⟧ F ‾ ∣ λ ρ ∣ = F ‾ ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ \overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}\overline{\mathcal{F}}_{|\lambda_\rho|}
=
\overline{\mathcal{F}}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket} F [ [ λ ρ ] ] F ∣ λ ρ ∣ = F ∣ λ ρ ∣ F [ [ λ ρ ] ] as operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) .
Proof. Step 1. We first prove the commutation relation on the level of functions in S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) . Precisely, we claim that it holds the identity
F ‾ ⟦ λ ρ ⟧ F ∣ λ ρ ∣ u = F ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ u , u ∈ S ( R n ) . (92) \overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}\mathcal{F}_{|\lambda_\rho|}u
=
\mathcal{F}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u,
\qquad u\in\mathcal{S}(\mathbb{R}^n). \tag*{(92)} F [ [ λ ρ ] ] F ∣ λ ρ ∣ u = F ∣ λ ρ ∣ F [ [ λ ρ ] ] u , u ∈ S ( R n ) . (92)
and, in particular, that the left-hand side is always well defined as an element in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Note that the right-hand side exists due to Lemma 2.11, which shows that F ‾ ⟦ λ ρ ⟧ \overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket} F [ [ λ ρ ] ] maps S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) into itself and that D ( F ∣ λ ρ ∣ ) = S ( R n ) D(\mathcal{F}_{|\lambda_\rho|})=\mathcal{S}(\mathbb{R}^n) D ( F ∣ λ ρ ∣ ) = S ( R n ) . To obtain (92) , we compute, for u ∈ S ( R n ) u\in\mathcal{S}(\mathbb{R}^n) u ∈ S ( R n ) ,
⟨ F ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ u , φ ⟩ = ⟨ ⟦ λ ρ ⟧ u ^ , ∣ λ ρ ∣ F − 1 φ ⟩ = ⟨ ∣ λ ρ ∣ u ^ , ⟦ λ ρ ⟧ F − 1 φ ⟩ , φ ∈ S ( R n ) , \langle\mathcal{F}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u,\varphi\rangle
=
\langle\llbracket\lambda_\rho\rrbracket\widehat{u},|\lambda_\rho|\mathcal{F}^{-1}\varphi\rangle
=
\langle|\lambda_\rho|\widehat{u},\llbracket\lambda_\rho\rrbracket\mathcal{F}^{-1}\varphi\rangle,
\qquad\varphi\in\mathcal{S}(\mathbb{R}^n), ⟨ F ∣ λ ρ ∣ F [ [ λ ρ ] ] u , φ ⟩ = ⟨ [ [ λ ρ ] ] u , ∣ λ ρ ∣ F − 1 φ ⟩ = ⟨ ∣ λ ρ ∣ u , [ [ λ ρ ] ] F − 1 φ ⟩ , φ ∈ S ( R n ) ,
and, by definition, the right-hand side equals F − 1 ( ∣ λ ρ ∣ F ∣ λ ρ ∣ u ) \mathcal{F}^{-1}(|\lambda_\rho|\mathcal{F}_{|\lambda_\rho|}u) F − 1 ( ∣ λ ρ ∣ F ∣ λ ρ ∣ u ) as a tempered distribution, i.e., it lies in H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) by virtue of Theorem 3.1. With Lemma 4.14, we have that F ∣ λ ρ ∣ u ∈ D ( F ‾ ⟦ λ ρ ⟧ ) \mathcal{F}_{|\lambda_\rho|}u\in D(\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}) F ∣ λ ρ ∣ u ∈ D ( F [ [ λ ρ ] ] ) , and equality (92) is then a direct consequence of the fact that, as a result of Lemma 2.4 for m = ⟦ λ ρ ⟧ m=\llbracket\lambda_\rho\rrbracket m = [ [ λ ρ ] ] , F ‾ ⟦ λ ρ ⟧ v \overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}v F [ [ λ ρ ] ] v agrees with F − 1 ( ⟦ λ ρ ⟧ v ^ ) \mathcal{F}^{-1}(\llbracket\lambda_\rho\rrbracket\widehat{v}) F − 1 ( [ [ λ ρ ] ] v ) for any v ∈ D ( F ‾ ⟦ λ ρ ⟧ ) v\in D(\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}) v ∈ D ( F [ [ λ ρ ] ] ) .
Step 2. In this step we are concerned with deriving the inclusion
D ( F ‾ ⟦ λ ρ ⟧ F ∣ λ ρ ∣ ) ⊂ D ( F ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ ) (93) D(\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}\mathcal{F}_{|\lambda_\rho|})
\subset D(\mathcal{F}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}) \tag*{(93)} D ( F [ [ λ ρ ] ] F ∣ λ ρ ∣ ) ⊂ D ( F ∣ λ ρ ∣ F [ [ λ ρ ] ] ) (93)
and the identity
F ‾ ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ u = F ‾ ⟦ λ ρ ⟧ F ∣ λ ρ ∣ u , u ∈ D ( F ‾ ⟦ λ ρ ⟧ F ∣ λ ρ ∣ ) . (94) \overline{\mathcal{F}}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u
=
\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}\mathcal{F}_{|\lambda_\rho|}u,
\qquad u\in D(\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}\mathcal{F}_{|\lambda_\rho|}). \tag*{(94)} F ∣ λ ρ ∣ F [ [ λ ρ ] ] u = F [ [ λ ρ ] ] F ∣ λ ρ ∣ u , u ∈ D ( F [ [ λ ρ ] ] F ∣ λ ρ ∣ ) . (94)
Let therefore u ∈ D ( F ‾ ⟦ λ ρ ⟧ F ∣ λ ρ ∣ ) u\in D(\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}\mathcal{F}_{|\lambda_\rho|}) u ∈ D ( F [ [ λ ρ ] ] F ∣ λ ρ ∣ ) and choose ( u j ) j ⊂ S ( R n ) (u_j)_j\subset\mathcal{S}(\mathbb{R}^n) ( u j ) j ⊂ S ( R n ) such that u j → u u_j\to u u j → u and F ∣ λ ρ ∣ u j → F ‾ ∣ λ ρ ∣ u \mathcal{F}_{|\lambda_\rho|}u_j\to\overline{\mathcal{F}}_{|\lambda_\rho|}u F ∣ λ ρ ∣ u j → F ∣ λ ρ ∣ u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . With Lemma 2.6 applied to m 1 = 1 + ∣ λ ρ ∣ m_1=1+|\lambda_\rho| m 1 = 1 + ∣ λ ρ ∣ and m 2 = ⟦ λ ρ ⟧ / ( 1 + ∣ λ ρ ∣ ) m_2=\llbracket\lambda_\rho\rrbracket/(1+|\lambda_\rho|) m 2 = [ [ λ ρ ] ] / ( 1 + ∣ λ ρ ∣ ) , it also holds u ∈ D ( F ‾ ⟦ λ ρ ⟧ ) u\in D(\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}) u ∈ D ( F [ [ λ ρ ] ] ) with F ‾ ⟦ λ ρ ⟧ u j → F ‾ ⟦ λ ρ ⟧ u \overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u_j\to\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u F [ [ λ ρ ] ] u j → F [ [ λ ρ ] ] u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Invoking Step 1 , yields that
F ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ u j = F ‾ ⟦ λ ρ ⟧ F ∣ λ ρ ∣ u j , j ∈ N . (95) \mathcal{F}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u_j
=
\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}\mathcal{F}_{|\lambda_\rho|}u_j,
\qquad j\in\mathbb{N}. \tag*{(95)} F ∣ λ ρ ∣ F [ [ λ ρ ] ] u j = F [ [ λ ρ ] ] F ∣ λ ρ ∣ u j , j ∈ N . (95)
Thus, since F ∣ λ ρ ∣ u j → F ‾ ∣ λ ρ ∣ u ∈ D ( F ‾ ⟦ λ ρ ⟧ ) \mathcal{F}_{|\lambda_\rho|}u_j\to\overline{\mathcal{F}}_{|\lambda_\rho|}u\in D(\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}) F ∣ λ ρ ∣ u j → F ∣ λ ρ ∣ u ∈ D ( F [ [ λ ρ ] ] ) and F ‾ ⟦ λ ρ ⟧ u j → F ‾ ⟦ λ ρ ⟧ u \overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u_j\to\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u F [ [ λ ρ ] ] u j → F [ [ λ ρ ] ] u , we find in the limit indeed that F ‾ ⟦ λ ρ ⟧ u ∈ D ( F ∣ λ ρ ∣ ) \overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u\in D(\mathcal{F}_{|\lambda_\rho|}) F [ [ λ ρ ] ] u ∈ D ( F ∣ λ ρ ∣ ) with F ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ u = F ‾ ⟦ λ ρ ⟧ F ∣ λ ρ ∣ u \mathcal{F}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u=\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}\mathcal{F}_{|\lambda_\rho|}u F ∣ λ ρ ∣ F [ [ λ ρ ] ] u = F [ [ λ ρ ] ] F ∣ λ ρ ∣ u . We point out that the regularity of the approximating sequence ( F ‾ ⟦ λ ρ ⟧ u j ) j (\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u_j)_j ( F [ [ λ ρ ] ] u j ) j is not important for this argument to hold; instead, the convergence of the left-hand side in (95) rests merely upon the fact that the graph of F ∣ λ ρ ∣ \mathcal{F}_{|\lambda_\rho|} F ∣ λ ρ ∣ is closed by definition of F ∣ λ ρ ∣ \mathcal{F}_{|\lambda_\rho|} F ∣ λ ρ ∣ being a closed operator. We have hence shown (93) and (94) at the same time.
Step 3. At this point, the overall claim follows as soon as we prove that
D ( F ‾ ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ ) ⊂ D ( F ‾ ⟦ λ ρ ⟧ F ∣ λ ρ ∣ ) (96) D(\overline{\mathcal{F}}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket})
\subset D(\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}\mathcal{F}_{|\lambda_\rho|}) \tag*{(96)} D ( F ∣ λ ρ ∣ F [ [ λ ρ ] ] ) ⊂ D ( F [ [ λ ρ ] ] F ∣ λ ρ ∣ ) (96)
and
F ‾ ⟦ λ ρ ⟧ F ∣ λ ρ ∣ u = F ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ u , u ∈ D ( F ‾ ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ ) . (97) \overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}\mathcal{F}_{|\lambda_\rho|}u
=
\mathcal{F}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u,
\qquad u\in D(\overline{\mathcal{F}}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}). \tag*{(97)} F [ [ λ ρ ] ] F ∣ λ ρ ∣ u = F ∣ λ ρ ∣ F [ [ λ ρ ] ] u , u ∈ D ( F ∣ λ ρ ∣ F [ [ λ ρ ] ] ) . (97)
The approach is similar to Step 2 . More precisely, for u ∈ D ( F ‾ ∣ λ ρ ∣ F ‾ ⟦ λ ρ ⟧ ) u\in D(\overline{\mathcal{F}}_{|\lambda_\rho|}\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}) u ∈ D ( F ∣ λ ρ ∣ F [ [ λ ρ ] ] ) pick ( u j ) j ⊂ S ( R n ) (u_j)_j\subset\mathcal{S}(\mathbb{R}^n) ( u j ) j ⊂ S ( R n ) with u j → u u_j\to u u j → u and F ⟦ λ ρ ⟧ u j → F ‾ ⟦ λ ρ ⟧ u \mathcal{F}_{\llbracket\lambda_\rho\rrbracket}u_j\to\overline{\mathcal{F}}_{\llbracket\lambda_\rho\rrbracket}u F [ [ λ ρ ] ] u j → F [ [ λ ρ ] ] u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , infer from Lemma 2.6, now with m 1 = ⟦ λ ρ ⟧ m_1=\llbracket\lambda_\rho\rrbracket m 1 = [ [ λ ρ ] ] and m 2 = ∣ λ ρ ∣ / ⟦ λ ρ ⟧ m_2=|\lambda_\rho|/\llbracket\lambda_\rho\rrbracket m 2 = ∣ λ ρ ∣/ [ [ λ ρ ] ] , that u ∈ D ( F ‾ ∣ λ ρ ∣ ) u\in D(\overline{\mathcal{F}}_{|\lambda_\rho|}) u ∈ D ( F ∣ λ ρ ∣ ) with F ∣ λ ρ ∣ u j → F ‾ ∣ λ ρ ∣ u \mathcal{F}_{|\lambda_\rho|}u_j\to\overline{\mathcal{F}}_{|\lambda_\rho|}u F ∣ λ ρ ∣ u j → F ∣ λ ρ ∣ u , and use Step 1 to conclude that (95) holds true again. In the limit we thus find exactly (96) and (97) . □
Next, we establish the identity F ‾ [ λ ρ ] 2 = F ‾ [ λ ρ ] 2 \overline{F}_{[\lambda_\rho]}^{2}=\overline{F}_{[\lambda_\rho]^2} F [ λ ρ ] 2 = F [ λ ρ ] 2 as operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , which is the analogue of the sought equality for [ λ ρ ] [\lambda_\rho] [ λ ρ ] instead of ∣ λ ρ ∣ |\lambda_\rho| ∣ λ ρ ∣ . We deduce the requisite core property of the operator F ‾ [ λ ρ ] 2 \overline{F}_{[\lambda_\rho]}^{2} F [ λ ρ ] 2 with the help of the L p L^p L p -Fourier multiplier operator F 1 / [ λ ρ ] 2 F_{1/[\lambda_\rho]^2} F 1/ [ λ ρ ] 2 .
Lemma 4.16. Assume that ρ \rho ρ has compact support and fulfills hypotheses ( H 0 ) (H0) ( H 0 ) –( H 2 ) (H2) ( H 2 ) and ( H 3 ′ ) (H3') ( H 3 ′ ) . Then
F ‾ [ λ ρ ] 2 = F ‾ [ λ ρ ] 2 \overline{F}_{[\lambda_\rho]}^{2}=\overline{F}_{[\lambda_\rho]^2} F [ λ ρ ] 2 = F [ λ ρ ] 2 as operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) .
Proof. We define, inspired by the Fourier characterization in Theorem 3.1 , the auxiliary space
X ρ , p ( R n ) : = { u ∈ L p ( R n ) : F − 1 ( [ λ ρ ] 2 u ^ ) ∈ L p ( R n ) } , X^{\rho,p}(\mathbb{R}^n):=\{u\in L^p(\mathbb{R}^n):\mathcal{F}^{-1}([\lambda_\rho]^2\widehat{u})\in L^p(\mathbb{R}^n)\}, X ρ , p ( R n ) := { u ∈ L p ( R n ) : F − 1 ([ λ ρ ] 2 u ) ∈ L p ( R n )} ,
and we proof in Steps 1 and 2 the statement D ( F ‾ [ λ ρ ] 2 ) = D ( F ‾ [ λ ρ ] 2 ) D(\overline{F}_{[\lambda_\rho]}^{2})=D(\overline{F}_{[\lambda_\rho]^2}) D ( F [ λ ρ ] 2 ) = D ( F [ λ ρ ] 2 ) by showing the two identities
D ( F ‾ [ λ ρ ] 2 ) = X ρ , p ( R n ) (98) D(\overline{F}_{[\lambda_\rho]}^{2})=X^{\rho,p}(\mathbb{R}^n) \tag*{(98)} D ( F [ λ ρ ] 2 ) = X ρ , p ( R n ) (98)
and
X ρ , p ( R n ) = D ( F ‾ [ λ ρ ] 2 ) . (99) X^{\rho,p}(\mathbb{R}^n)=D(\overline{F}_{[\lambda_\rho]^2}). \tag*{(99)} X ρ , p ( R n ) = D ( F [ λ ρ ] 2 ) . (99)
Appealing to Lemma 2.7 for m 1 = [ λ ρ ] = m 2 m_1=[\lambda_\rho]=m_2 m 1 = [ λ ρ ] = m 2 , this automatically implies that F ‾ [ λ ρ ] 2 = F ‾ [ λ ρ ] 2 \overline{F}_{[\lambda_\rho]}^{2}=\overline{F}_{[\lambda_\rho]^2} F [ λ ρ ] 2 = F [ λ ρ ] 2 as operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) .
Step 1: proof of (4.40). We recall that the domain of F ‾ [ λ ρ ] 2 \overline{F}_{[\lambda_\rho]}^{2} F [ λ ρ ] 2 is given by
D ( F ‾ [ λ ρ ] 2 ) = D ( F ‾ [ λ ρ ] F ‾ [ λ ρ ] ) = { u ∈ D ( F ‾ [ λ ρ ] ) : F ‾ [ λ ρ ] u ∈ D ( F ‾ [ λ ρ ] ) } , D(\overline{F}_{[\lambda_\rho]}^{2})=D(\overline{F}_{[\lambda_\rho]}\overline{F}_{[\lambda_\rho]})=\{u\in D(\overline{F}_{[\lambda_\rho]}):\overline{F}_{[\lambda_\rho]}u\in D(\overline{F}_{[\lambda_\rho]})\}, D ( F [ λ ρ ] 2 ) = D ( F [ λ ρ ] F [ λ ρ ] ) = { u ∈ D ( F [ λ ρ ] ) : F [ λ ρ ] u ∈ D ( F [ λ ρ ] )} ,
cf. (16) . With the help of Theorem 3.1 and Lemma 4.14, we can reformulate the right-hand side as
{ u ∈ D ( F ‾ [ λ ρ ] ) : F ‾ [ λ ρ ] u ∈ D ( F ‾ [ λ ρ ] ) } = { u ∈ H ρ , p ( R n ) : F − 1 ( [ λ ρ ] u ^ ) ∈ H ρ , p ( R n ) } = { u ∈ L p ( R n ) : F − 1 ( [ λ ρ ] u ^ ) , F − 1 ( [ λ ρ ] F ( F − 1 ( [ λ ρ ] u ^ ) ) ) ∈ L p ( R n ) } = { u ∈ L p ( R n ) : F − 1 ( [ λ ρ ] u ^ ) , F − 1 ( [ λ ρ ] 2 u ^ ) ∈ L p ( R n ) } . \begin{aligned}
\{u\in D(\overline{F}_{[\lambda_\rho]}):\overline{F}_{[\lambda_\rho]}u\in D(\overline{F}_{[\lambda_\rho]})\}
&=\{u\in H^{\rho,p}(\mathbb{R}^n):\mathcal{F}^{-1}([\lambda_\rho]\widehat{u})\in H^{\rho,p}(\mathbb{R}^n)\}\\
&=\{u\in L^p(\mathbb{R}^n):\mathcal{F}^{-1}([\lambda_\rho]\widehat{u}),\mathcal{F}^{-1}([\lambda_\rho]\mathcal{F}(\mathcal{F}^{-1}([\lambda_\rho]\widehat{u})))\in L^p(\mathbb{R}^n)\}\\
&=\{u\in L^p(\mathbb{R}^n):\mathcal{F}^{-1}([\lambda_\rho]\widehat{u}),\mathcal{F}^{-1}([\lambda_\rho]^2\widehat{u})\in L^p(\mathbb{R}^n)\}.
\end{aligned} { u ∈ D ( F [ λ ρ ] ) : F [ λ ρ ] u ∈ D ( F [ λ ρ ] )} = { u ∈ H ρ , p ( R n ) : F − 1 ([ λ ρ ] u ) ∈ H ρ , p ( R n )} = { u ∈ L p ( R n ) : F − 1 ([ λ ρ ] u ) , F − 1 ([ λ ρ ] F ( F − 1 ([ λ ρ ] u ))) ∈ L p ( R n )} = { u ∈ L p ( R n ) : F − 1 ([ λ ρ ] u ) , F − 1 ([ λ ρ ] 2 u ) ∈ L p ( R n )} .
Let u ∈ L p ( R n ) u\in L^p(\mathbb{R}^n) u ∈ L p ( R n ) be an element of the right-hand side above. Then, by (46) in Lemma 2.11 , in combination with Lemma 2.4 for m = [ λ ρ ] m=[\lambda_\rho] m = [ λ ρ ] and m = [ λ ρ ] 2 m=[\lambda_\rho]^2 m = [ λ ρ ] 2 , we have that
∥ F − 1 ( [ λ ρ ] u ^ ) ∥ L p ( R n ) ≤ 1 2 ∥ u ∥ L p ( R n ) + 1 2 ∥ F − 1 ( [ λ ρ ] 2 u ^ ) ∥ L p ( R n ) . \|\mathcal{F}^{-1}([\lambda_\rho]\widehat{u})\|_{L^p(\mathbb{R}^n)}
\leq\frac{1}{2}\|u\|_{L^p(\mathbb{R}^n)}+\frac{1}{2}\|\mathcal{F}^{-1}([\lambda_\rho]^2\widehat{u})\|_{L^p(\mathbb{R}^n)}. ∥ F − 1 ([ λ ρ ] u ) ∥ L p ( R n ) ≤ 2 1 ∥ u ∥ L p ( R n ) + 2 1 ∥ F − 1 ([ λ ρ ] 2 u ) ∥ L p ( R n ) .
Thus, we conclude
D ( F ‾ [ λ ρ ] 2 ) = { u ∈ L p ( R n ) : F − 1 ( [ λ ρ ] 2 u ^ ) ∈ L p ( R n ) } = X ρ , p ( R n ) . D(\overline{F}_{[\lambda_\rho]}^{2})=\{u\in L^p(\mathbb{R}^n):\mathcal{F}^{-1}([\lambda_\rho]^2\widehat{u})\in L^p(\mathbb{R}^n)\}=X^{\rho,p}(\mathbb{R}^n). D ( F [ λ ρ ] 2 ) = { u ∈ L p ( R n ) : F − 1 ([ λ ρ ] 2 u ) ∈ L p ( R n )} = X ρ , p ( R n ) .
Step 2: proof of (4.41). The inclusion D ( F ‾ [ λ ρ ] 2 ) ⊂ X ρ , p ( R n ) D(\overline{F}_{[\lambda_\rho]^2})\subset X^{\rho,p}(\mathbb{R}^n) D ( F [ λ ρ ] 2 ) ⊂ X ρ , p ( R n ) holds true since Lemma 2.4 for m = [ λ ρ ] 2 m=[\lambda_\rho]^2 m = [ λ ρ ] 2 implies that F − 1 ( [ λ ρ ] 2 u ^ ) = F ‾ [ λ ρ ] 2 u \mathcal{F}^{-1}([\lambda_\rho]^2\widehat{u})=\overline{F}_{[\lambda_\rho]^2}u F − 1 ([ λ ρ ] 2 u ) = F [ λ ρ ] 2 u for any u ∈ D ( F ‾ [ λ ρ ] 2 ) u\in D(\overline{F}_{[\lambda_\rho]^2}) u ∈ D ( F [ λ ρ ] 2 ) . We thus turn to the reverse inclusion. Observe that proving D ( F ‾ [ λ ρ ] 2 ) ⊂ X ρ , p ( R n ) D(\overline{F}_{[\lambda_\rho]^2})\subset X^{\rho,p}(\mathbb{R}^n) D ( F [ λ ρ ] 2 ) ⊂ X ρ , p ( R n ) amounts to finding for each u ∈ X ρ , p ( R n ) u\in X^{\rho,p}(\mathbb{R}^n) u ∈ X ρ , p ( R n ) an approximating sequence ( u j ) j ⊂ S ( R n ) (u_j)_j\subset\mathcal{S}(\mathbb{R}^n) ( u j ) j ⊂ S ( R n ) such that
u j → u , F ‾ [ λ ρ ] 2 u j → F ‾ [ λ ρ ] 2 u , in L p ( R n ) . (100) u_j\to u,\qquad\overline{F}_{[\lambda_\rho]^2}u_j\to\overline{F}_{[\lambda_\rho]^2}u,\qquad\text{in }L^p(\mathbb{R}^n). \tag*{(100)} u j → u , F [ λ ρ ] 2 u j → F [ λ ρ ] 2 u , in L p ( R n ) . (100)
To this end, since F − 1 ( [ λ ρ ] 2 u ^ ) ∈ L p ( R n ) \mathcal{F}^{-1}([\lambda_\rho]^2\widehat{u})\in L^p(\mathbb{R}^n) F − 1 ([ λ ρ ] 2 u ) ∈ L p ( R n ) by definition of X ρ , p ( R n ) X^{\rho,p}(\mathbb{R}^n) X ρ , p ( R n ) , there exists, by the density of C ∞ ( R n ) C^\infty(\mathbb{R}^n) C ∞ ( R n ) in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , a sequence ( w j ) j ⊂ C ∞ ( R n ) ⊂ S ( R n ) (w_j)_j\subset C^\infty(\mathbb{R}^n)\subset\mathcal{S}(\mathbb{R}^n) ( w j ) j ⊂ C ∞ ( R n ) ⊂ S ( R n ) such that w j → F − 1 ( [ λ ρ ] 2 u ^ ) w_j\to\mathcal{F}^{-1}([\lambda_\rho]^2\widehat{u}) w j → F − 1 ([ λ ρ ] 2 u ) in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . We define u j : = F 1 / [ λ ρ ] 2 w j u_j:=F_{1/[\lambda_\rho]^2}w_j u j := F 1/ [ λ ρ ] 2 w j for j ∈ N j\in\mathbb{N} j ∈ N , note that ( F 1 / [ λ ρ ] 2 w j ) j ⊂ S ( R n ) (F_{1/[\lambda_\rho]^2}w_j)_j\subset\mathcal{S}(\mathbb{R}^n) ( F 1/ [ λ ρ ] 2 w j ) j ⊂ S ( R n ) according to Lemma 2.11 (i), and claim that this sequence satisfies the convergence properties of (4.42). Indeed, by Proposition 2.10 (iv), F 1 / [ λ ρ ] 2 F_{1/[\lambda_\rho]^2} F 1/ [ λ ρ ] 2 is an L p L^p L p -Fourier multiplier and we obtain that
u j = F 1 / [ λ ρ ] 2 w j → F 1 / [ λ ρ ] 2 F − 1 ( [ λ ρ ] 2 u ^ ) = u in L p ( R n ) , u_j=F_{1/[\lambda_\rho]^2}w_j\to F_{1/[\lambda_\rho]^2}\mathcal{F}^{-1}([\lambda_\rho]^2\widehat{u})=u\quad\text{in }L^p(\mathbb{R}^n), u j = F 1/ [ λ ρ ] 2 w j → F 1/ [ λ ρ ] 2 F − 1 ([ λ ρ ] 2 u ) = u in L p ( R n ) ,
where the final equality is due to the identity F 1 / ⟦ λ ρ ⟧ 2 ‾ v = F − 1 ( 1 / ⟦ λ ρ ⟧ 2 v ^ ) \overline{F_{1/\llbracket\lambda_\rho\rrbracket^2}}v=F^{-1}(1/\llbracket\lambda_\rho\rrbracket^2\widehat{v}) F 1/ [ [ λ ρ ] ] 2 v = F − 1 ( 1/ [ [ λ ρ ] ] 2 v ) for any v ∈ D ( F 1 / ⟦ λ ρ ⟧ 2 ‾ ) = L p ( R n ) v\in D(\overline{F_{1/\llbracket\lambda_\rho\rrbracket^2}})=L^p(\mathbb{R}^n) v ∈ D ( F 1/ [ [ λ ρ ] ] 2 ) = L p ( R n ) , resulting from Lemma 2.4 with the choice m = 1 / ⟦ λ ρ ⟧ 2 m=1/\llbracket\lambda_\rho\rrbracket^2 m = 1/ [ [ λ ρ ] ] 2 . Furthermore, it holds with Lemma 2.5 ,
F ⟦ λ ρ ⟧ 2 ‾ u j = F ⟦ λ ρ ⟧ 2 ‾ F 1 / ⟦ λ ρ ⟧ 2 ‾ w j = w j → F − 1 ( ⟦ λ ρ ⟧ 2 u ^ ) in L p ( R n ) . \overline{F_{\llbracket\lambda_\rho\rrbracket^2}}u_j=\overline{F_{\llbracket\lambda_\rho\rrbracket^2}}\overline{F_{1/\llbracket\lambda_\rho\rrbracket^2}}w_j=w_j\to F^{-1}(\llbracket\lambda_\rho\rrbracket^2\widehat{u})\ \text{in }L^p(\mathbb{R}^n). F [ [ λ ρ ] ] 2 u j = F [ [ λ ρ ] ] 2 F 1/ [ [ λ ρ ] ] 2 w j = w j → F − 1 ( [ [ λ ρ ] ] 2 u ) in L p ( R n ) .
Hence, (100) is fulfilled and (99) is proved. □
After these preparatory considerations, we will now show that F ∣ λ ρ ∣ 2 ‾ = − Δ ρ \overline{F_{|\lambda_\rho|^2}}=-\Delta^\rho F ∣ λ ρ ∣ 2 = − Δ ρ as operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , deriving the statement from the preceding lemmas.
Proposition 4.17. Assume that ρ \rho ρ has compact support and satisfies ( H 0 ) (H0) ( H 0 ) –( H 3 ) (H3) ( H 3 ) . Then
F ∣ λ ρ ∣ 2 ‾ = − Δ ρ \overline{F_{|\lambda_\rho|^2}}=-\Delta^\rho F ∣ λ ρ ∣ 2 = − Δ ρ as operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) .
Proof. Subsequently, Step 1 proves the equality D ( F ∣ λ ρ ∣ 2 ‾ ) = D ( F ⟦ λ ρ ⟧ 2 ‾ ) D(\overline{F_{|\lambda_\rho|^2}})=D(\overline{F_{\llbracket\lambda_\rho\rrbracket^2}}) D ( F ∣ λ ρ ∣ 2 ) = D ( F [ [ λ ρ ] ] 2 ) , while Step 2 establishes that D ( F ∣ λ ρ ∣ − 2 ‾ ) = D ( F ⟦ λ ρ ⟧ − 2 ‾ ) D(\overline{F_{|\lambda_\rho|^{-2}}})=D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-2}}}) D ( F ∣ λ ρ ∣ − 2 ) = D ( F [ [ λ ρ ] ] − 2 ) . Thereafter, the proof is concluded in Step 3 .
Step 1: Similarly to the Step 2 in the proof of Lemma 4.14, applying Lemma 2.6 twice, namely with m 1 = ⟦ λ ρ ⟧ 2 m_1=\llbracket\lambda_\rho\rrbracket^2 m 1 = [ [ λ ρ ] ] 2 and m 2 = ∣ λ ρ ∣ 2 / ⟦ λ ρ ⟧ 2 m_2=|\lambda_\rho|^2/\llbracket\lambda_\rho\rrbracket^2 m 2 = ∣ λ ρ ∣ 2 / [ [ λ ρ ] ] 2 , as well as m 1 = ( 1 + ∣ λ ρ ∣ ) 2 m_1=(1+|\lambda_\rho|)^2 m 1 = ( 1 + ∣ λ ρ ∣ ) 2 and m 2 = ⟦ λ ρ ⟧ 2 / ( 1 + ∣ λ ρ ∣ ) 2 m_2=\llbracket\lambda_\rho\rrbracket^2/(1+|\lambda_\rho|)^2 m 2 = [ [ λ ρ ] ] 2 / ( 1 + ∣ λ ρ ∣ ) 2 (the choices for m 2 m_2 m 2 are indeed L p L^p L p -Fourier multipliers as squares of the L p L^p L p -Fourier multipliers in Proposition 2.10 (i) and (ii), respectively; see [17 Proposition 2.5.13] ) leads to
D ( F ( 1 + ∣ λ ρ ∣ ) 2 ‾ ) ⊂ D ( F ⟦ λ ρ ⟧ 2 ‾ ) ⊂ D ( F ∣ λ ρ ∣ 2 ‾ ) . D(\overline{F_{(1+|\lambda_\rho|)^2}})\subset D(\overline{F_{\llbracket\lambda_\rho\rrbracket^2}})\subset D(\overline{F_{|\lambda_\rho|^2}}). D ( F ( 1 + ∣ λ ρ ∣ ) 2 ) ⊂ D ( F [ [ λ ρ ] ] 2 ) ⊂ D ( F ∣ λ ρ ∣ 2 ) .
Identity in this line of inclusions follows from the observation that D ( F ∣ λ ρ ∣ 2 ‾ ) ⊂ D ( F ( 1 + ∣ λ ρ ∣ ) 2 ‾ ) D(\overline{F_{|\lambda_\rho|^2}})\subset D(\overline{F_{(1+|\lambda_\rho|)^2}}) D ( F ∣ λ ρ ∣ 2 ) ⊂ D ( F ( 1 + ∣ λ ρ ∣ ) 2 ) . To see the latter, let u ∈ D ( F ∣ λ ρ ∣ 2 ‾ ) u\in D(\overline{F_{|\lambda_\rho|^2}}) u ∈ D ( F ∣ λ ρ ∣ 2 ) , meaning that there exists ( u j ) j ⊂ S ( R n ) (u_j)_j\subset\mathcal{S}(\mathbb{R}^n) ( u j ) j ⊂ S ( R n ) such that u j → u u_j\to u u j → u and F ∣ λ ρ ∣ 2 ‾ u j → F ∣ λ ρ ∣ 2 ‾ u \overline{F_{|\lambda_\rho|^2}}u_j\to\overline{F_{|\lambda_\rho|^2}}u F ∣ λ ρ ∣ 2 u j → F ∣ λ ρ ∣ 2 u in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . Appealing to (47) in Lemma 2.11 , we know that ( F ∣ λ ρ ∣ ‾ u j ) j (\overline{F_{|\lambda_\rho|}}u_j)_j ( F ∣ λ ρ ∣ u j ) j defines a Cauchy sequence in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) , and hence converges to F ∣ λ ρ ∣ ‾ u \overline{F_{|\lambda_\rho|}}u F ∣ λ ρ ∣ u by the closedness of the operator. Therefore,
F ( 1 + ∣ λ ρ ∣ ) 2 ‾ u j = u j + 2 F ∣ λ ρ ∣ ‾ u j + F ∣ λ ρ ∣ 2 ‾ u j → u + 2 F ∣ λ ρ ∣ ‾ u + F ∣ λ ρ ∣ 2 ‾ u = F ( 1 + ∣ λ ρ ∣ ) 2 ‾ u in L p ( R n ) , (101) \overline{F_{(1+|\lambda_\rho|)^2}}u_j=u_j+2\overline{F_{|\lambda_\rho|}}u_j+\overline{F_{|\lambda_\rho|^2}}u_j\to u+2\overline{F_{|\lambda_\rho|}}u+\overline{F_{|\lambda_\rho|^2}}u=\overline{F_{(1+|\lambda_\rho|)^2}}u\quad\text{in }L^p(\mathbb{R}^n), \tag*{(101)} F ( 1 + ∣ λ ρ ∣ ) 2 u j = u j + 2 F ∣ λ ρ ∣ u j + F ∣ λ ρ ∣ 2 u j → u + 2 F ∣ λ ρ ∣ u + F ∣ λ ρ ∣ 2 u = F ( 1 + ∣ λ ρ ∣ ) 2 u in L p ( R n ) , (101)
showing u ∈ D ( F 1 + ∣ λ ρ ∣ ‾ ) u\in D(\overline{F_{1+|\lambda_\rho|}}) u ∈ D ( F 1 + ∣ λ ρ ∣ ) . Overall, we have proved that D ( F ∣ λ ρ ∣ 2 ‾ ) = D ( F ⟦ λ ρ ⟧ 2 ‾ ) D(\overline{F_{|\lambda_\rho|^2}})=D(\overline{F_{\llbracket\lambda_\rho\rrbracket^2}}) D ( F ∣ λ ρ ∣ 2 ) = D ( F [ [ λ ρ ] ] 2 ) .
Step 2: To show that D ( F ∣ λ ρ ∣ − 2 ‾ ) ⊂ D ( F ⟦ λ ρ ⟧ − 2 ‾ ) D(\overline{F_{|\lambda_\rho|^{-2}}})\subset D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-2}}}) D ( F ∣ λ ρ ∣ − 2 ) ⊂ D ( F [ [ λ ρ ] ] − 2 ) , let u ∈ D ( F ∣ λ ρ ∣ − 2 ‾ ) u\in D(\overline{F_{|\lambda_\rho|^{-2}}}) u ∈ D ( F ∣ λ ρ ∣ − 2 ) . Then, by definition, u ∈ D ( F ∣ λ ρ ∣ − 1 ‾ ) u\in D(\overline{F_{|\lambda_\rho|^{-1}}}) u ∈ D ( F ∣ λ ρ ∣ − 1 ) and F ∣ λ ρ ∣ − 1 ‾ u ∈ D ( F ∣ λ ρ ∣ − 1 ‾ ) \overline{F_{|\lambda_\rho|^{-1}}}u\in D(\overline{F_{|\lambda_\rho|^{-1}}}) F ∣ λ ρ ∣ − 1 u ∈ D ( F ∣ λ ρ ∣ − 1 ) . It follows immediately from Lemma 4.14 that u ∈ D ( F ⟦ λ ρ ⟧ − 1 ‾ ) u\in D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-1}}}) u ∈ D ( F [ [ λ ρ ] ] − 1 ) and F ∣ λ ρ ∣ − 1 ‾ u ∈ D ( F ⟦ λ ρ ⟧ − 1 ‾ ) \overline{F_{|\lambda_\rho|^{-1}}}u\in D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-1}}}) F ∣ λ ρ ∣ − 1 u ∈ D ( F [ [ λ ρ ] ] − 1 ) . Furthermore, we obtain from Lemma 4.15 that F ⟦ λ ρ ⟧ − 1 ‾ u ∈ D ( F ∣ λ ρ ∣ − 1 ‾ ) \overline{F_{\llbracket\lambda_\rho\rrbracket^{-1}}}u\in D(\overline{F_{|\lambda_\rho|^{-1}}}) F [ [ λ ρ ] ] − 1 u ∈ D ( F ∣ λ ρ ∣ − 1 ) , and, again with Lemma 4.14, we conclude that F ⟦ λ ρ ⟧ − 1 ‾ u ∈ D ( F ⟦ λ ρ ⟧ − 1 ‾ ) \overline{F_{\llbracket\lambda_\rho\rrbracket^{-1}}}u\in D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-1}}}) F [ [ λ ρ ] ] − 1 u ∈ D ( F [ [ λ ρ ] ] − 1 ) . This verifies that u ∈ D ( F ⟦ λ ρ ⟧ − 2 ‾ ) u\in D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-2}}}) u ∈ D ( F [ [ λ ρ ] ] − 2 ) and proves D ( F ∣ λ ρ ∣ − 2 ‾ ) ⊂ D ( F ⟦ λ ρ ⟧ − 2 ‾ ) D(\overline{F_{|\lambda_\rho|^{-2}}})\subset D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-2}}}) D ( F ∣ λ ρ ∣ − 2 ) ⊂ D ( F [ [ λ ρ ] ] − 2 ) .
The reverse inclusion D ( F ⟦ λ ρ ⟧ − 2 ‾ ) ⊂ D ( F ∣ λ ρ ∣ − 2 ‾ ) D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-2}}})\subset D(\overline{F_{|\lambda_\rho|^{-2}}}) D ( F [ [ λ ρ ] ] − 2 ) ⊂ D ( F ∣ λ ρ ∣ − 2 ) is derived similarly, with the roles of ∣ λ ρ ∣ |\lambda_\rho| ∣ λ ρ ∣ and ⟦ λ ρ ⟧ \llbracket\lambda_\rho\rrbracket [ [ λ ρ ] ] exchanged. First, from u ∈ D ( F ⟦ λ ρ ⟧ − 2 ‾ ) u\in D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-2}}}) u ∈ D ( F [ [ λ ρ ] ] − 2 ) , one has u ∈ D ( F ⟦ λ ρ ⟧ − 1 ‾ ) u\in D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-1}}}) u ∈ D ( F [ [ λ ρ ] ] − 1 ) and F ⟦ λ ρ ⟧ − 1 ‾ u ∈ D ( F ⟦ λ ρ ⟧ − 1 ‾ ) \overline{F_{\llbracket\lambda_\rho\rrbracket^{-1}}}u\in D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-1}}}) F [ [ λ ρ ] ] − 1 u ∈ D ( F [ [ λ ρ ] ] − 1 ) . Then, Lemma 4.14 entails that u ∈ D ( F ∣ λ ρ ∣ − 1 ‾ ) u\in D(\overline{F_{|\lambda_\rho|^{-1}}}) u ∈ D ( F ∣ λ ρ ∣ − 1 ) and F ⟦ λ ρ ⟧ − 1 ‾ u ∈ D ( F ∣ λ ρ ∣ − 1 ‾ ) \overline{F_{\llbracket\lambda_\rho\rrbracket^{-1}}}u\in D(\overline{F_{|\lambda_\rho|^{-1}}}) F [ [ λ ρ ] ] − 1 u ∈ D ( F ∣ λ ρ ∣ − 1 ) , and Lemma 4.15 combined with Lemma 4.14 yield F ∣ λ ρ ∣ − 1 ‾ u ∈ D ( F ∣ λ ρ ∣ − 1 ‾ ) \overline{F_{|\lambda_\rho|^{-1}}}u\in D(\overline{F_{|\lambda_\rho|^{-1}}}) F ∣ λ ρ ∣ − 1 u ∈ D ( F ∣ λ ρ ∣ − 1 ) , i.e., u ∈ D ( F ∣ λ ρ ∣ − 2 ‾ ) u\in D(\overline{F_{|\lambda_\rho|^{-2}}}) u ∈ D ( F ∣ λ ρ ∣ − 2 ) . This concludes Step 2 , establishing that D ( F ⟦ λ ρ ⟧ − 2 ‾ ) = D ( F ∣ λ ρ ∣ − 2 ‾ ) D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-2}}})=D(\overline{F_{|\lambda_\rho|^{-2}}}) D ( F [ [ λ ρ ] ] − 2 ) = D ( F ∣ λ ρ ∣ − 2 ) .
Step 3: Using Steps 1 and 2 , we infer from Lemma 4.16 that
D ( F ∣ λ ρ ∣ − 2 ‾ ) = D ( F ⟦ λ ρ ⟧ − 2 ‾ ) . (102) D(\overline{F_{|\lambda_\rho|^{-2}}})=D(\overline{F_{\llbracket\lambda_\rho\rrbracket^{-2}}}). \tag*{(102)} D ( F ∣ λ ρ ∣ − 2 ) = D ( F [ [ λ ρ ] ] − 2 ) . (102)
Lemma 2.7 with m 1 = ∣ λ ρ ∣ = m 2 m_1=|\lambda_\rho|=m_2 m 1 = ∣ λ ρ ∣ = m 2 (where the requisite property that F ( ∣ λ ρ ∣ F − 1 φ ) , F ( ∣ λ ρ ∣ 2 F − 1 φ ) ∈ L p ′ ( R n ) \mathcal{F}(|\lambda_\rho|\mathcal{F}^{-1}\varphi),\mathcal{F}(|\lambda_\rho|^2\mathcal{F}^{-1}\varphi)\in L^{p'}(\mathbb{R}^n) F ( ∣ λ ρ ∣ F − 1 φ ) , F ( ∣ λ ρ ∣ 2 F − 1 φ ) ∈ L p ′ ( R n ) for all φ ∈ S ( R n ) \varphi\in\mathcal{S}(\mathbb{R}^n) φ ∈ S ( R n ) is a direct consequence of Lemma 2.11 ) shows that F ∣ λ ρ ∣ − 2 ‾ = F ∣ λ ρ ∣ 2 ‾ \overline{F_{|\lambda_\rho|^{-2}}}=\overline{F_{|\lambda_\rho|^2}} F ∣ λ ρ ∣ − 2 = F ∣ λ ρ ∣ 2 , and it therefore follows along with Proposition 4.13 (iii) the identity F ∣ λ ρ ∣ 2 ‾ = − Δ ρ \overline{F_{|\lambda_\rho|^2}}=-\Delta^\rho F ∣ λ ρ ∣ 2 = − Δ ρ as unbounded operators in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) . □
With the result of Proposition 4.17 in mind, we define the square root of − Δ ρ -\Delta^\rho − Δ ρ as an unbounded operator in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) as follows.
Definition 4.18 (Square root of − Δ ρ -\Delta^\rho − Δ ρ ). For ρ \rho ρ with compact support and satisfying (H0)–(H3) we define the operator
( − Δ ρ ) 1 / 2 : D ( ( − Δ ρ ) 1 / 2 ) → L p ( R n ) (-\Delta^\rho)^{1/2}:D((-\Delta^\rho)^{1/2})\to L^p(\mathbb{R}^n) ( − Δ ρ ) 1/2 : D (( − Δ ρ ) 1/2 ) → L p ( R n ) by
D ( ( − Δ ρ ) 1 / 2 ) : = D ( F ∣ λ ρ ∣ ‾ ) , ( − Δ ρ ) 1 / 2 : = F ∣ λ ρ ∣ ‾ . D((-\Delta^\rho)^{1/2}):=D(\overline{\mathcal{F}|\lambda_\rho|}),\qquad(-\Delta^\rho)^{1/2}:=\overline{\mathcal{F}|\lambda_\rho|}. D (( − Δ ρ ) 1/2 ) := D ( F ∣ λ ρ ∣ ) , ( − Δ ρ ) 1/2 := F ∣ λ ρ ∣ .
Finally, we are ready to conclude, as also described in § 1.1 , a generalization of the well-known representation of classical Bessel potential spaces through the square root of the negative fractional Laplace operator.
Corollary 4.19 (Characterization in terms of ( − Δ ρ ) 1 / 2 (-\Delta^\rho)^{1/2} ( − Δ ρ ) 1/2 ). Assume that ρ \rho ρ has compact support and satisfies (H0)–(H3). Then
H ρ , p ( R n ) = { u ∈ L p ( R n ) : F ∣ λ ρ ∣ ‾ u ∈ L p ( R n ) } = D ( ( − Δ ρ ) 1 / 2 ) . H^{\rho,p}(\mathbb{R}^n)=\{u\in L^p(\mathbb{R}^n):\overline{\mathcal{F}|\lambda_\rho|}u\in L^p(\mathbb{R}^n)\}=D((-\Delta^\rho)^{1/2}). H ρ , p ( R n ) = { u ∈ L p ( R n ) : F ∣ λ ρ ∣ u ∈ L p ( R n )} = D (( − Δ ρ ) 1/2 ) .
Proof. The statement is a direct consequence of Definition 4.18 together with Proposition 4.14. □
Characterization in terms of ( 1 − Δ ρ ) 1 / 2 (1-\Delta^\rho)^{1/2} ( 1 − Δ ρ ) 1/2
Motivated by the Fourier characterization in Theorem 3.1, we define for ρ \rho ρ with compact support and subject to hypotheses (H0)–(H2) and (H3′) the operator
( 1 − Δ ρ ) 1 / 2 : S ′ ( R n ) → S ′ ( R n ) (1-\Delta^\rho)^{1/2}:\mathcal{S}'(\mathbb{R}^n)\to\mathcal{S}'(\mathbb{R}^n) ( 1 − Δ ρ ) 1/2 : S ′ ( R n ) → S ′ ( R n )
as
( 1 − Δ ρ ) 1 / 2 u : = F − 1 ( [ λ ρ ] u ^ ) , u ∈ S ′ ( R n ) . (103) (1-\Delta^\rho)^{1/2}u:=\mathcal{F}^{-1}([\lambda_\rho]\widehat{u}),\qquad u\in\mathcal{S}'(\mathbb{R}^n). \tag*{(103)} ( 1 − Δ ρ ) 1/2 u := F − 1 ([ λ ρ ] u ) , u ∈ S ′ ( R n ) . (103)
In fact, this definition is compatible with the framework developed in the preceding subsections. By virtue of the first identity of Lemma 2.1 ( i ) (i) ( i ) , we have the equality
⟨ ( 1 − Δ ρ ) 1 / 2 u , φ ⟩ S ′ × S = ⟨ u , F ( [ λ ρ ] F − 1 φ ) ⟩ S ′ × S = ⟨ u , F [ λ ρ ] φ ⟩ S ′ × S , u ∈ S ′ ( R n ) , φ ∈ S ( R n ) . \langle(1-\Delta^\rho)^{1/2}u,\varphi\rangle_{\mathcal{S}'\times\mathcal{S}}=\langle u,\mathcal{F}([\lambda_\rho]\mathcal{F}^{-1}\varphi)\rangle_{\mathcal{S}'\times\mathcal{S}}=\langle u,\mathcal{F}[\lambda_\rho]\varphi\rangle_{\mathcal{S}'\times\mathcal{S}},\qquad u\in\mathcal{S}'(\mathbb{R}^n),\ \varphi\in\mathcal{S}(\mathbb{R}^n). ⟨( 1 − Δ ρ ) 1/2 u , φ ⟩ S ′ × S = ⟨ u , F ([ λ ρ ] F − 1 φ ) ⟩ S ′ × S = ⟨ u , F [ λ ρ ] φ ⟩ S ′ × S , u ∈ S ′ ( R n ) , φ ∈ S ( R n ) .
In addition, the operator F [ λ ρ ] \mathcal{F}[\lambda_\rho] F [ λ ρ ] studied in Lemma 2.11 ( i ) (i) ( i ) has a well defined square in S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) , which equals 1 + F ∣ λ ρ ∣ 2 = 1 − Δ ρ 1+\mathcal{F}|\lambda_\rho|^2=1-\Delta^\rho 1 + F ∣ λ ρ ∣ 2 = 1 − Δ ρ , thanks to (4.6). Hence, ( 1 − Δ ρ ) 1 / 2 (1-\Delta^\rho)^{1/2} ( 1 − Δ ρ ) 1/2 is indeed the extension by duality to S ′ ( R n ) \mathcal{S}'(\mathbb{R}^n) S ′ ( R n ) of a square root of the operator 1 − Δ ρ 1-\Delta^\rho 1 − Δ ρ in S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) .
With these considerations, we can reformulate Theorem 3.1 as follows.
Corollary 4.20 (Characterization in terms of ( 1 − Δ ρ ) 1 / 2 (1-\Delta^\rho)^{1/2} ( 1 − Δ ρ ) 1/2 ). Let p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and assume that ρ \rho ρ has compact support and satisfies (H0)–(H2) and (H3′). We have the identity
H ρ , p ( R n ) = { u ∈ S ′ ( R n ) : ( 1 − Δ ρ ) 1 / 2 u ∈ L p ( R n ) } . H^{\rho,p}(\mathbb{R}^n)=\{u\in\mathcal{S}'(\mathbb{R}^n):(1-\Delta^\rho)^{1/2}u\in L^p(\mathbb{R}^n)\}. H ρ , p ( R n ) = { u ∈ S ′ ( R n ) : ( 1 − Δ ρ ) 1/2 u ∈ L p ( R n )} .
Characterization as potential spaces
The subject of this section is the identification of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) with a suitable potential space, which is a generalization of the identity between H ρ s , p ( R n ) H^{\rho_s,p}(\mathbb{R}^n) H ρ s , p ( R n ) for the Riesz fractional gradient ρ s \rho_s ρ s and the Bessel potential spaces (see [33 Theorem 1.7] ). In particular, we prove that, under sufficient conditions on the kernel function ρ \rho ρ , H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) can likewise be regarded as the image of L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) under convolution with a suitable generalized potential kernel g ρ g_\rho g ρ , defined in Proposition 5.1 below.
As outlined in the introductory § 1.1 and in analogy to the case of the Riesz fractional gradient, the candidate for the function g ρ g_\rho g ρ is the Fourier inverse of
[ λ ρ ] − 1 = ( 1 + ∣ λ ρ ∣ 2 ) − 1 / 2 . [\lambda_\rho]^{-1}=(1+|\lambda_\rho|^2)^{-1/2}. [ λ ρ ] − 1 = ( 1 + ∣ λ ρ ∣ 2 ) − 1/2 .
Thus, the challenge consists in the following in proving that this Fourier inverse exists, and, moreover, that it lies in L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) . For reasons that will become apparent in Lemma 5.4, we show a more general result.
Proposition 5.1. Let p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) and ρ \rho ρ have compact support and satisfy assumptions (H0)–(H3). Then for each ℓ ∈ N \ell\in\mathbb{N} ℓ ∈ N there exists a function g ρ , ℓ ∈ L 1 ( R n ) g_{\rho,\ell} \in L^1(\mathbb{R}^n) g ρ , ℓ ∈ L 1 ( R n ) such that
g ρ , ℓ ^ = ⟦ λ ρ ⟧ − ℓ , \widehat{g_{\rho,\ell}} = \llbracket\lambda_\rho\rrbracket^{-\ell}, g ρ , ℓ = [ [ λ ρ ] ] − ℓ , and the family ( g ρ , ℓ ) ℓ (g_{\rho,\ell})_\ell ( g ρ , ℓ ) ℓ is uniformly bounded in L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) .
Proof. We proceed in several steps. We first formally introduce the candidates for the functions g ρ , ℓ g_{\rho,\ell} g ρ , ℓ , before proving that they actually constitute elements of L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) . For the latter purpose, we derive explicit pointwise upper bounds, guaranteeing that the L 1 L^1 L 1 -norm of g ρ , ℓ g_{\rho,\ell} g ρ , ℓ is bounded independently of ℓ ∈ N \ell\in\mathbb{N} ℓ ∈ N .
Step 1: construction of the functions g ρ , ℓ g_{\rho,\ell} g ρ , ℓ . * We recall initially that according to hypothesis (H3), we know that (see (2.19) together with the comments below (2.20))
∣ λ ρ ( ξ ) ∣ 2 ≥ C ∣ ξ ∣ 2 σ , ξ ∈ B 1 c , (104) |\lambda_\rho(\xi)|^2 \ge C|\xi|^{2\sigma}, \qquad\xi\in B_1^c, \tag*{(104)} ∣ λ ρ ( ξ ) ∣ 2 ≥ C ∣ ξ ∣ 2 σ , ξ ∈ B 1 c , (104)
for σ ∈ ( 0 , 1 ) \sigma\in(0,1) σ ∈ ( 0 , 1 ) from (H3) and a constant C > 0 C > 0 C > 0 . As a consequence of this and the polynomial growth of the derivatives of ∣ λ ρ ∣ 2 |\lambda_\rho|^2 ∣ λ ρ ∣ 2 , following Lemma 2.9, for each δ > 0 \delta> 0 δ > 0 , the function
h ρ δ : = e − ∣ λ ρ ∣ 2 δ / ( 4 π 2 ) δ n / 2 (105) h_\rho^\delta:= e^{-|\lambda_\rho|^2\delta/(4\pi^2)}\delta^{n/2} \tag*{(105)} h ρ δ := e − ∣ λ ρ ∣ 2 δ / ( 4 π 2 ) δ n /2 (105)
lies in S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) . Consequently, the Fourier inverse f ρ δ : = F − 1 h ρ δ f_\rho^\delta:= \mathcal{F}^{-1}h_\rho^\delta f ρ δ := F − 1 h ρ δ exists and is again an element of S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) . Furthermore, appealing to the identity
t − ℓ / 2 = 1 Γ ( ℓ 2 ) ∫ 0 ∞ e − t δ δ ℓ / 2 d δ δ , t > 0 , ℓ > 0 (106) t^{-\ell/2} = \frac{1}{\Gamma\left(\frac{\ell}{2}\right)}\int_0^\infty e^{-t\delta}\delta^{\ell/2}\frac{\mathrm{d}\delta}{\delta}, \qquad t > 0,\ \ell> 0 \tag*{(106)} t − ℓ /2 = Γ ( 2 ℓ ) 1 ∫ 0 ∞ e − t δ δ ℓ /2 δ d δ , t > 0 , ℓ > 0 (106)
it follows with the choice t = ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) / 4 π 2 t = (1+|\lambda_\rho(\xi)|^2)/4\pi^2 t = ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) /4 π 2 for ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n that
1 ( 4 π 2 ) ℓ / 2 Γ ( ℓ 2 ) ∫ 0 ∞ h ρ δ ( ξ ) e − δ / 4 π 2 δ ( ℓ − n ) / 2 d δ δ = 1 ( 4 π 2 ) ℓ / 2 Γ ( ℓ 2 ) ∫ 0 ∞ e − ∣ λ ρ ( ξ ) ∣ 2 δ / ( 4 π 2 ) e − δ / 4 π 2 δ ℓ / 2 d δ δ = 1 ( 4 π 2 ) ℓ / 2 Γ ( ℓ 2 ) ∫ 0 ∞ e − ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) δ / ( 4 π 2 ) δ ℓ / 2 d δ δ = ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) − ℓ / 2 , ξ ∈ R n , ℓ ∈ N . \begin{aligned}
\frac{1}{(4\pi^2)^{\ell/2}\Gamma\left(\frac{\ell}{2}\right)}
\int_0^\infty h_\rho^\delta(\xi)e^{-\delta/4\pi^2}\delta^{(\ell-n)/2}\frac{\mathrm{d}\delta}{\delta}
&= \frac{1}{(4\pi^2)^{\ell/2}\Gamma\left(\frac{\ell}{2}\right)}
\int_0^\infty e^{-|\lambda_\rho(\xi)|^2\delta/(4\pi^2)}e^{-\delta/4\pi^2}\delta^{\ell/2}\frac{\mathrm{d}\delta}{\delta} \\
&= \frac{1}{(4\pi^2)^{\ell/2}\Gamma\left(\frac{\ell}{2}\right)}
\int_0^\infty e^{-(1+|\lambda_\rho(\xi)|^2)\delta/(4\pi^2)}\delta^{\ell/2}\frac{\mathrm{d}\delta}{\delta} \\
&= (1+|\lambda_\rho(\xi)|^2)^{-\ell/2}, \qquad\xi\in\mathbb{R}^n,\ \ell\in\mathbb{N}.
\end{aligned} ( 4 π 2 ) ℓ /2 Γ ( 2 ℓ ) 1 ∫ 0 ∞ h ρ δ ( ξ ) e − δ /4 π 2 δ ( ℓ − n ) /2 δ d δ = ( 4 π 2 ) ℓ /2 Γ ( 2 ℓ ) 1 ∫ 0 ∞ e − ∣ λ ρ ( ξ ) ∣ 2 δ / ( 4 π 2 ) e − δ /4 π 2 δ ℓ /2 δ d δ = ( 4 π 2 ) ℓ /2 Γ ( 2 ℓ ) 1 ∫ 0 ∞ e − ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) δ / ( 4 π 2 ) δ ℓ /2 δ d δ = ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) − ℓ /2 , ξ ∈ R n , ℓ ∈ N .
Thus, defining
g ρ , ℓ ( x ) : = 1 ( 4 π 2 ) ℓ / 2 Γ ( ℓ 2 ) ∫ 0 ∞ f ρ δ ( x ) e − δ / 4 π 2 δ ( ℓ − n ) / 2 d δ δ , x ∈ R n , (107) g_{\rho,\ell}(x) := \frac{1}{(4\pi^2)^{\ell/2}\Gamma\left(\frac{\ell}{2}\right)}
\int_0^\infty f_\rho^\delta(x)e^{-\delta/4\pi^2}\delta^{(\ell-n)/2}\frac{\mathrm{d}\delta}{\delta}, \qquad x \in\mathbb{R}^n, \tag*{(107)} g ρ , ℓ ( x ) := ( 4 π 2 ) ℓ /2 Γ ( 2 ℓ ) 1 ∫ 0 ∞ f ρ δ ( x ) e − δ /4 π 2 δ ( ℓ − n ) /2 δ d δ , x ∈ R n , (107)
Fubini’s theorem, together with the fact that g ρ , ℓ ∈ L 1 ( R n ) g_{\rho,\ell} \in L^1(\mathbb{R}^n) g ρ , ℓ ∈ L 1 ( R n ) for all ℓ ∈ N \ell\in\mathbb{N} ℓ ∈ N , as will be established below, yields that
g ρ , ℓ ^ = ( 1 + ∣ λ ρ ∣ 2 ) − ℓ / 2 = ⟦ λ ρ ⟧ − ℓ . \widehat{g_{\rho,\ell}} = (1+|\lambda_\rho|^2)^{-\ell/2} = \llbracket\lambda_\rho\rrbracket^{-\ell}. g ρ , ℓ = ( 1 + ∣ λ ρ ∣ 2 ) − ℓ /2 = [ [ λ ρ ] ] − ℓ .
Step 2: sufficient conditions for g ρ , ℓ ∈ L 1 ( R n ) g_{\rho,\ell} \in L^1(\mathbb{R}^n) g ρ , ℓ ∈ L 1 ( R n ) . * From here on, C > 0 C > 0 C > 0 denotes a generic, possibly changing constant which may depend on ρ \rho ρ but is, in particular, independent of δ > 0 \delta> 0 δ > 0 and ℓ ∈ N \ell\in\mathbb{N} ℓ ∈ N . It remains to verify the property that ( g ρ , ℓ ) ℓ (g_{\rho,\ell})_\ell ( g ρ , ℓ ) ℓ is uniformly bounded in L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) . By definition, we have
f ρ δ ( x ) = δ n / 2 ∫ R n e − ∣ λ ρ ( ξ ) ∣ 2 δ / ( 4 π 2 ) e 2 π i x ⋅ ξ d ξ , x ∈ R n . f_\rho^\delta(x) = \delta^{n/2}\int_{\mathbb{R}^n} e^{-|\lambda_\rho(\xi)|^2\delta/(4\pi^2)}e^{2\pi i x\cdot\xi}\,\mathrm{d}\xi, \qquad x \in\mathbb{R}^n. f ρ δ ( x ) = δ n /2 ∫ R n e − ∣ λ ρ ( ξ ) ∣ 2 δ / ( 4 π 2 ) e 2 π i x ⋅ ξ d ξ , x ∈ R n .
We will show in Steps 3, 4 and 5 below that there exist η 1 , η 2 , η 3 ∈ [ 0 , 1 / 2 ) \eta_1,\eta_2,\eta_3 \in[0,1/2) η 1 , η 2 , η 3 ∈ [ 0 , 1/2 ) and b ∈ ( 0 , 1 ) b \in(0,1) b ∈ ( 0 , 1 ) such that
∣ f ρ δ ( x ) ∣ ≤ C ∑ i = 1 3 δ n / 2 − η i ∣ x ∣ n − b , x ∈ B 1 ∖ { 0 } , δ ∈ ( 0 , 1 ) , (108) |f_\rho^\delta(x)| \le C\sum_{i=1}^{3}\frac{\delta^{n/2-\eta_i}}{|x|^{n-b}}, \qquad x \in B_1 \setminus\{0\},\ \delta\in(0,1), \tag*{(108)} ∣ f ρ δ ( x ) ∣ ≤ C i = 1 ∑ 3 ∣ x ∣ n − b δ n /2 − η i , x ∈ B 1 ∖ { 0 } , δ ∈ ( 0 , 1 ) , (108)
∣ f ρ δ ( x ) ∣ ≤ C δ n / 2 1 ∣ x ∣ n + 1 , x ∈ B 1 c , δ ∈ ( 0 , 1 ) . (109) |f_\rho^\delta(x)| \le C\delta^{n/2}\frac{1}{|x|^{n+1}}, \qquad x \in B_1^c,\ \delta\in(0,1). \tag*{(109)} ∣ f ρ δ ( x ) ∣ ≤ C δ n /2 ∣ x ∣ n + 1 1 , x ∈ B 1 c , δ ∈ ( 0 , 1 ) . (109)
and
∥ f ρ δ ∥ L 1 ( R n ) ≤ C δ n / 2 , δ ∈ ( 1 , ∞ ) . (110) \lVert f_{\rho}^{\delta}\rVert_{L^{1}(\mathbb{R}^{n})}\le C\delta^{n/2},\qquad\delta\in(1,\infty). \tag*{(110)} ∥ f ρ δ ∥ L 1 ( R n ) ≤ C δ n /2 , δ ∈ ( 1 , ∞ ) . (110)
We continue with the proof assuming these estimates. Observing that
∫ R n ∣ g ρ , ℓ ( x ) ∣ d x ≤ 1 ( 4 π 2 ) ℓ / 2 Γ ( ℓ 2 ) [ ∫ B 1 ∫ 0 1 + ∫ B 1 c ∫ 0 1 + ∫ R n ∫ 1 ∞ ] ∣ f ρ δ ( x ) ∣ e − δ / 4 π 2 δ ( ℓ − n ) / 2 d δ δ d x , \int_{\mathbb{R}^{n}} |g_{\rho,\ell}(x)|\,\mathrm{d}x
\le\frac{1}{(4\pi^{2})^{\ell/2}\Gamma\left(\frac{\ell}{2}\right)}
\left[\int_{B_{1}}\int_{0}^{1}+\int_{B_{1}^{c}}\int_{0}^{1}+\int_{\mathbb{R}^{n}}\int_{1}^{\infty}\right]
|f_{\rho}^{\delta}(x)|e^{-\delta/4\pi^{2}}\delta^{(\ell-n)/2}\frac{\mathrm{d}\delta}{\delta}\,\mathrm{d}x, ∫ R n ∣ g ρ , ℓ ( x ) ∣ d x ≤ ( 4 π 2 ) ℓ /2 Γ ( 2 ℓ ) 1 [ ∫ B 1 ∫ 0 1 + ∫ B 1 c ∫ 0 1 + ∫ R n ∫ 1 ∞ ] ∣ f ρ δ ( x ) ∣ e − δ /4 π 2 δ ( ℓ − n ) /2 δ d δ d x ,
we then estimate using (5.5),
∫ B 1 ∫ 0 1 ∣ f ρ δ ( x ) ∣ e − δ / 4 π 2 δ ( ℓ − n ) / 2 d δ δ d x ≤ C ∑ i = 1 3 ∫ B 1 1 ∣ x ∣ n − b d x ∫ 0 1 e − δ / 4 π 2 δ ( ℓ − 2 ) / 2 − η i d δ ≤ C ∫ 0 1 δ − 1 / 2 − max i ∈ { 1 , 2 , 3 } η i d δ ≤ C , \begin{aligned}
\int_{B_{1}}\int_{0}^{1}|f_{\rho}^{\delta}(x)|e^{-\delta/4\pi^{2}}\delta^{(\ell-n)/2}\frac{\mathrm{d}\delta}{\delta}\,\mathrm{d}x
\le C\sum_{i=1}^{3}\int_{B_{1}}\frac{1}{|x|^{n-b}}\mathrm{d}x\int_{0}^{1}e^{-\delta/4\pi^{2}}\delta^{(\ell-2)/2-\eta_i}\,\mathrm{d}\delta\\
&\le C\int_{0}^{1}\delta^{-1/2-\max_{i\in\{1,2,3\}}\eta_i}\,\mathrm{d}\delta\le C,
\end{aligned} ∫ B 1 ∫ 0 1 ∣ f ρ δ ( x ) ∣ e − δ /4 π 2 δ ( ℓ − n ) /2 δ d δ d x ≤ C i = 1 ∑ 3 ∫ B 1 ∣ x ∣ n − b 1 d x ∫ 0 1 e − δ /4 π 2 δ ( ℓ − 2 ) /2 − η i d δ ≤ C ∫ 0 1 δ − 1/2 − m a x i ∈ { 1 , 2 , 3 } η i d δ ≤ C ,
using (5.6),
∫ B 1 c ∫ 0 1 ∣ f ρ δ ( x ) ∣ e − δ / 4 π 2 δ ( ℓ − n ) / 2 d δ δ d x ≤ C ∫ B 1 c 1 ∣ x ∣ n + 1 d x ∫ 0 1 e − δ / 4 π 2 δ ℓ / 2 d δ ≤ C , \int_{B_{1}^{c}}\int_{0}^{1}|f_{\rho}^{\delta}(x)|e^{-\delta/4\pi^{2}}\delta^{(\ell-n)/2}\frac{\mathrm{d}\delta}{\delta}\,\mathrm{d}x
\le C\int_{B_{1}^{c}}\frac{1}{|x|^{n+1}}\mathrm{d}x\int_{0}^{1}e^{-\delta/4\pi^{2}}\delta^{\ell/2}\,\mathrm{d}\delta\le C, ∫ B 1 c ∫ 0 1 ∣ f ρ δ ( x ) ∣ e − δ /4 π 2 δ ( ℓ − n ) /2 δ d δ d x ≤ C ∫ B 1 c ∣ x ∣ n + 1 1 d x ∫ 0 1 e − δ /4 π 2 δ ℓ /2 d δ ≤ C ,
and using (5.7) and Fubini’s theorem,
∫ R n ∫ 1 ∞ ∣ f ρ δ ( x ) ∣ e − δ / 4 π 2 δ ( ℓ − n ) / 2 d δ δ d x ≤ C ∫ 1 ∞ e − δ / 4 π 2 δ ℓ / 2 d δ δ ≤ C ∫ 0 ∞ e − δ / 4 π 2 δ ℓ / 2 d δ δ = C ( 4 π 2 ) ℓ / 2 Γ ( ℓ 2 ) . \begin{aligned}
\int_{\mathbb{R}^{n}}\int_{1}^{\infty}|f_{\rho}^{\delta}(x)|e^{-\delta/4\pi^{2}}\delta^{(\ell-n)/2}\frac{\mathrm{d}\delta}{\delta}\,\mathrm{d}x
\le C\int_{1}^{\infty}e^{-\delta/4\pi^{2}}\delta^{\ell/2}\frac{\mathrm{d}\delta}{\delta} \\
&\le C\int_{0}^{\infty}e^{-\delta/4\pi^{2}}\delta^{\ell/2}\frac{\mathrm{d}\delta}{\delta}
=C(4\pi^{2})^{\ell/2}\Gamma\left(\frac{\ell}{2}\right).
\end{aligned} ∫ R n ∫ 1 ∞ ∣ f ρ δ ( x ) ∣ e − δ /4 π 2 δ ( ℓ − n ) /2 δ d δ d x ≤ C ∫ 1 ∞ e − δ /4 π 2 δ ℓ /2 δ d δ ≤ C ∫ 0 ∞ e − δ /4 π 2 δ ℓ /2 δ d δ = C ( 4 π 2 ) ℓ /2 Γ ( 2 ℓ ) .
Altogether and along with the property of the Γ \Gamma Γ -function that
sup ℓ ∈ ( 0 , ∞ ) 1 ( 4 π 2 ) ℓ / 2 Γ ( ℓ 2 ) < ∞ , \sup_{\ell\in(0,\infty)}\frac{1}{(4\pi^{2})^{\ell/2}\Gamma\left(\frac{\ell}{2}\right)}<\infty, ℓ ∈ ( 0 , ∞ ) sup ( 4 π 2 ) ℓ /2 Γ ( 2 ℓ ) 1 < ∞ ,
we conclude
∫ R n ∣ g ρ , ℓ ( x ) ∣ d x ≤ C \int_{\mathbb{R}^{n}}|g_{\rho,\ell}(x)|\,\mathrm{d}x\le C ∫ R n ∣ g ρ , ℓ ( x ) ∣ d x ≤ C
with a constant independent of ℓ ∈ N \ell\in\mathbb{N} ℓ ∈ N .
We will now turn to the remaining proof of (5.5), (5.6) and (5.7). For the first two estimates, we will use a quantified form of the fact that, as an inverse Fourier transform of an exponential function, f ρ δ f_{\rho}^{\delta} f ρ δ decays faster than any polynomial. In particular, the corresponding approach refines the techniques applied in Proposition 4.4. To derive (5.7), we will apply a related but different strategy.
To unburden the notation, the following estimates will be carried out for the auxiliary function
Θ ρ δ : = e − ∣ λ ρ ∣ 2 δ / ( 4 π 2 ) = h ρ δ δ − n / 2 , (111) \Theta_{\rho}^{\delta}:=e^{-|\lambda_{\rho}|^{2}\delta/(4\pi^{2})}=h_{\rho}^{\delta}\delta^{-n/2}, \tag*{(111)} Θ ρ δ := e − ∣ λ ρ ∣ 2 δ / ( 4 π 2 ) = h ρ δ δ − n /2 , (111)
whose relation with f ρ δ f_{\rho}^{\delta} f ρ δ is given through f ρ δ = F − 1 h ρ δ = δ n / 2 F − 1 Θ ρ δ f_{\rho}^{\delta}=\mathcal{F}^{-1}h_{\rho}^{\delta}=\delta^{n/2}\mathcal{F}^{-1}\Theta_{\rho}^{\delta} f ρ δ = F − 1 h ρ δ = δ n /2 F − 1 Θ ρ δ .
Step 3: proof of (5.5). We will demonstrate the sought estimate for general δ ∈ ( 0 , ∞ ) \delta\in(0,\infty) δ ∈ ( 0 , ∞ ) . Using the radial nature of Θ ρ δ \Theta_{\rho}^{\delta} Θ ρ δ and, hence, exploiting the rotational invariance of the integrand, we write
F − 1 Θ ρ δ ( x ) = [ ∫ B 1 + ∫ B 1 c ] Θ ρ δ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ , x ∈ B 1 ∖ { 0 } . \mathcal{F}^{-1}\Theta_{\rho}^{\delta}(x)
=\left[\int_{B_{1}}+\int_{B_{1}^{c}}\right]\Theta_{\rho}^{\delta}(\xi)e^{2\pi i|x|\xi_{1}}\,\mathrm{d}\xi,\qquad x\in B_{1}\setminus\{0\}. F − 1 Θ ρ δ ( x ) = [ ∫ B 1 + ∫ B 1 c ] Θ ρ δ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ , x ∈ B 1 ∖ { 0 } .
It follows directly that
∣ ∫ B 1 Θ ρ δ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ ∣ ≤ C ≤ C 1 ∣ x ∣ n − b , x ∈ B 1 ∖ { 0 } , \left|\int_{B_{1}}\Theta_{\rho}^{\delta}(\xi)e^{2\pi i|x|\xi_{1}}\,\mathrm{d}\xi\right|
\le C\le C\frac{1}{|x|^{n-b}},\qquad x\in B_{1}\setminus\{0\}, ∫ B 1 Θ ρ δ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ ≤ C ≤ C ∣ x ∣ n − b 1 , x ∈ B 1 ∖ { 0 } ,
for any b ∈ ( 0 , 1 ) b \in(0,1) b ∈ ( 0 , 1 ) . For the integral over B 1 c B_1^c B 1 c , applying now the change of variables ξ ↦ δ − 1 / ( 2 σ ) ξ \xi\mapsto\delta^{-1/(2\sigma)}\xi ξ ↦ δ − 1/ ( 2 σ ) ξ and partially integrating n − 1 n-1 n − 1 times, we find the expression
∫ B 1 c Θ ρ δ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ = δ − n / ( 2 σ ) ∫ B δ 1 / ( 2 σ ) c Θ ρ δ ( ξ δ 1 / ( 2 σ ) ) e 2 π i ∣ x ∣ δ − 1 / ( 2 σ ) ξ 1 d ξ = δ − 1 / ( 2 σ ) ( − 1 ) n − 1 ( 2 π i ∣ x ∣ ) n − 1 ∫ B δ 1 / ( 2 σ ) c ∂ ξ 1 n − 1 Θ ρ δ ( ξ δ 1 / ( 2 σ ) ) e 2 π i ∣ x ∣ δ − 1 / ( 2 σ ) ξ 1 d ξ + R ( x ) (112) \begin{aligned}
\int_{B_1^c}\Theta_\rho^\delta(\xi)e^{2\pi i|x|\xi_1}\,d\xi
=\delta^{-n/(2\sigma)}\int_{B_{\delta^{1/(2\sigma)}}^c}\Theta_\rho^\delta\left(\frac{\xi}{\delta^{1/(2\sigma)}}\right)e^{2\pi i|x|\delta^{-1/(2\sigma)}\xi_1}\,d\xi\\
&=\delta^{-1/(2\sigma)}\frac{(-1)^{n-1}}{(2\pi i|x|)^{n-1}}\int_{B_{\delta^{1/(2\sigma)}}^c}\partial_{\xi_1}^{n-1}\Theta_\rho^\delta\left(\frac{\xi}{\delta^{1/(2\sigma)}}\right)e^{2\pi i|x|\delta^{-1/(2\sigma)}\xi_1}\,d\xi+R(x) \tag*{(112)}
\end{aligned} ∫ B 1 c Θ ρ δ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ = δ − n / ( 2 σ ) ∫ B δ 1/ ( 2 σ ) c Θ ρ δ ( δ 1/ ( 2 σ ) ξ ) e 2 π i ∣ x ∣ δ − 1/ ( 2 σ ) ξ 1 d ξ = δ − 1/ ( 2 σ ) ( 2 π i ∣ x ∣ ) n − 1 ( − 1 ) n − 1 ∫ B δ 1/ ( 2 σ ) c ∂ ξ 1 n − 1 Θ ρ δ ( δ 1/ ( 2 σ ) ξ ) e 2 π i ∣ x ∣ δ − 1/ ( 2 σ ) ξ 1 d ξ + R ( x ) (112)
for x ∈ B 1 ∖ { 0 } x\in B_1\setminus\{0\} x ∈ B 1 ∖ { 0 } , with the remainder
R ( x ) = δ − n / ( 2 σ ) ∑ k = 1 n − 1 ( − 1 ) k − 1 δ ( n − 1 ) / ( 2 σ ) ( 2 π i ∣ x ∣ ) n − 1 ∫ ∂ B δ 1 / ( 2 σ ) ∂ ξ 1 k − 1 Θ ρ δ ( ξ δ 1 / ( 2 σ ) ) ∂ ξ 1 n − 1 − k e 2 π i ∣ x ∣ δ − 1 / ( 2 σ ) ξ 1 d S ( ξ ) , R(x)=\delta^{-n/(2\sigma)}\sum_{k=1}^{n-1}\frac{(-1)^{k-1}\delta^{(n-1)/(2\sigma)}}{(2\pi i|x|)^{n-1}}\int_{\partial B_{\delta^{1/(2\sigma)}}}\partial_{\xi_1}^{k-1}\Theta_\rho^\delta\left(\frac{\xi}{\delta^{1/(2\sigma)}}\right)\partial_{\xi_1}^{n-1-k}e^{2\pi i|x|\delta^{-1/(2\sigma)}\xi_1}\,dS(\xi), R ( x ) = δ − n / ( 2 σ ) k = 1 ∑ n − 1 ( 2 π i ∣ x ∣ ) n − 1 ( − 1 ) k − 1 δ ( n − 1 ) / ( 2 σ ) ∫ ∂ B δ 1/ ( 2 σ ) ∂ ξ 1 k − 1 Θ ρ δ ( δ 1/ ( 2 σ ) ξ ) ∂ ξ 1 n − 1 − k e 2 π i ∣ x ∣ δ − 1/ ( 2 σ ) ξ 1 d S ( ξ ) ,
comprising all boundary integrals arising from the n − 1 n-1 n − 1 partial integrations.
Fix
b : = σ 4 − 3 σ , γ : = 1 4 − 3 σ , b:=\frac{\sigma}{4-3\sigma},\qquad\gamma:=\frac{1}{4-3\sigma}, b := 4 − 3 σ σ , γ := 4 − 3 σ 1 ,
and note that b ∈ ( 0 , σ ) ⊂ ( 0 , 1 ) b\in(0,\sigma)\subset(0,1) b ∈ ( 0 , σ ) ⊂ ( 0 , 1 ) and γ = b / σ \gamma=b/\sigma γ = b / σ . By virtue of estimate (A.9), the remainder can be bounded by
∣ R ( x ) ∣ ≤ C δ − n / ( 2 σ ) ∑ k = 1 n − 1 δ k / ( 2 σ ) ∣ x ∣ k ∫ ∂ B δ 1 / ( 2 σ ) ∣ ξ ∣ − ( k − 1 + γ σ ) d S ( ξ ) ≤ C δ − n / ( 2 σ ) ∑ k = 1 n − 1 δ k / ( 2 σ ) ∣ x ∣ k δ ( n − 1 ) / ( 2 σ ) ∫ ∂ B 1 δ − ( k − 1 + γ σ ) / ( 2 σ ) ∣ ξ ∣ − ( k − 1 + γ σ ) d S ( ξ ) ≤ C δ − γ / 2 ∣ x ∣ n − b , x ∈ B 1 ∖ { 0 } , (113) \begin{aligned}
|R(x)|\leq C\delta^{-n/(2\sigma)}\sum_{k=1}^{n-1}\frac{\delta^{k/(2\sigma)}}{|x|^k}\int_{\partial B_{\delta^{1/(2\sigma)}}}|\xi|^{-(k-1+\gamma\sigma)}\,dS(\xi) \\
&\leq C\delta^{-n/(2\sigma)}\sum_{k=1}^{n-1}\frac{\delta^{k/(2\sigma)}}{|x|^k}\delta^{(n-1)/(2\sigma)}\int_{\partial B_1}\delta^{-(k-1+\gamma\sigma)/(2\sigma)}|\xi|^{-(k-1+\gamma\sigma)}\,dS(\xi) \\
&\leq C\frac{\delta^{-\gamma/2}}{|x|^{n-b}},\qquad x\in B_1\setminus\{0\}, \tag*{(113)}
\end{aligned} ∣ R ( x ) ∣ ≤ C δ − n / ( 2 σ ) k = 1 ∑ n − 1 ∣ x ∣ k δ k / ( 2 σ ) ∫ ∂ B δ 1/ ( 2 σ ) ∣ ξ ∣ − ( k − 1 + γ σ ) d S ( ξ ) ≤ C δ − n / ( 2 σ ) k = 1 ∑ n − 1 ∣ x ∣ k δ k / ( 2 σ ) δ ( n − 1 ) / ( 2 σ ) ∫ ∂ B 1 δ − ( k − 1 + γ σ ) / ( 2 σ ) ∣ ξ ∣ − ( k − 1 + γ σ ) d S ( ξ ) ≤ C ∣ x ∣ n − b δ − γ /2 , x ∈ B 1 ∖ { 0 } , (113)
where the change of variables ξ ↦ δ 1 / ( 2 σ ) ξ \xi\mapsto\delta^{1/(2\sigma)}\xi ξ ↦ δ 1/ ( 2 σ ) ξ was used in the second step.
Fix
a : = 4 − 3 σ 4 σ a:=\frac{4-3\sigma}{4\sigma} a := 4 σ 4 − 3 σ
and note that a ∈ ( 1 / 4 , ∞ ) a\in(1/4,\infty) a ∈ ( 1/4 , ∞ ) and a b = 1 / 4 ab=1/4 ab = 1/4 . For simplicity, we assume in the following that δ a / ∣ x ∣ > δ 1 / ( 2 σ ) \delta^a/|x|>\delta^{1/(2\sigma)} δ a /∣ x ∣ > δ 1/ ( 2 σ ) , entailing that the integral over B δ 1 / ( 2 σ ) c B_{\delta^{1/(2\sigma)}}^c B δ 1/ ( 2 σ ) c in the right-hand side of (112) is subdivided into two domains: B δ a / ∣ x ∣ ∖ B δ 1 / ( 2 σ ) B_{\delta^a/|x|}\setminus B_{\delta^{1/(2\sigma)}} B δ a /∣ x ∣ ∖ B δ 1/ ( 2 σ ) and B δ a / ∣ x ∣ c B_{\delta^a/|x|}^c B δ a /∣ x ∣ c . If instead δ a / ∣ x ∣ ≤ δ 1 / ( 2 σ ) \delta^a/|x|\leq\delta^{1/(2\sigma)} δ a /∣ x ∣ ≤ δ 1/ ( 2 σ ) , one might just extend the integral to B δ a / ∣ x ∣ c B_{\delta^a/|x|}^c B δ a /∣ x ∣ c and proceed according to the steps below. In the first domain, we leave the expression unchanged, while in the second one, we partially integrate once more, revealing the identity
δ − 1 / ( 2 σ ) ( − 1 ) n − 1 ( 2 π i ∣ x ∣ ) n − 1 ∫ B δ a / ∣ x ∣ c ∂ ξ 1 n − 1 Θ ρ δ ( ξ δ 1 / ( 2 σ ) ) e 2 π i ∣ x ∣ δ − 1 / ( 2 σ ) ξ 1 d ξ = ( − 1 ) n ( 2 π i ∣ x ∣ ) n ∫ ∂ B δ a / ∣ x ∣ ∂ ξ 1 n − 1 Θ ρ δ ( ξ δ 1 / ( 2 σ ) ) e 2 π i ∣ x ∣ δ − 1 / ( 2 σ ) ξ 1 d S ( ξ ) \begin{aligned}
\delta^{-1/(2\sigma)}\frac{(-1)^{n-1}}{(2\pi i|x|)^{n-1}}\int_{B_{\delta^a/|x|}^c}\partial_{\xi_1}^{n-1}\Theta_\rho^\delta\left(\frac{\xi}{\delta^{1/(2\sigma)}}\right)e^{2\pi i|x|\delta^{-1/(2\sigma)}\xi_1}\,d\xi\\
&=\frac{(-1)^n}{(2\pi i|x|)^n}\int_{\partial B_{\delta^a/|x|}}\partial_{\xi_1}^{n-1}\Theta_\rho^\delta\left(\frac{\xi}{\delta^{1/(2\sigma)}}\right)e^{2\pi i|x|\delta^{-1/(2\sigma)}\xi_1}\,dS(\xi)
\end{aligned} δ − 1/ ( 2 σ ) ( 2 π i ∣ x ∣ ) n − 1 ( − 1 ) n − 1 ∫ B δ a /∣ x ∣ c ∂ ξ 1 n − 1 Θ ρ δ ( δ 1/ ( 2 σ ) ξ ) e 2 π i ∣ x ∣ δ − 1/ ( 2 σ ) ξ 1 d ξ = ( 2 π i ∣ x ∣ ) n ( − 1 ) n ∫ ∂ B δ a /∣ x ∣ ∂ ξ 1 n − 1 Θ ρ δ ( δ 1/ ( 2 σ ) ξ ) e 2 π i ∣ x ∣ δ − 1/ ( 2 σ ) ξ 1 d S ( ξ )
+ ( − 1 ) n ( 2 π i ∣ x ∣ ) n ∫ B δ a / ∣ x ∣ c ∂ ξ 1 n Θ ρ δ ( ξ δ 1 / ( 2 σ ) ) e 2 π i ∣ x ∣ δ − 1 / ( 2 σ ) ξ 1 d ξ , x ∈ B 1 ∖ { 0 } . \qquad+\frac{(-1)^n}{(2\pi i|x|)^n}\int_{B_{\delta^a/|x|}^c}\partial_{\xi_1}^{n}\Theta_\rho^\delta\left(\frac{\xi}{\delta^{1/(2\sigma)}}\right)e^{2\pi i|x|\delta^{-1/(2\sigma)}\xi_1}\,d\xi,\qquad x\in B_1\setminus\{0\}. + ( 2 π i ∣ x ∣ ) n ( − 1 ) n ∫ B δ a /∣ x ∣ c ∂ ξ 1 n Θ ρ δ ( δ 1/ ( 2 σ ) ξ ) e 2 π i ∣ x ∣ δ − 1/ ( 2 σ ) ξ 1 d ξ , x ∈ B 1 ∖ { 0 } .
Again with (A.9), and recalling that
γ = b / σ , a b = 1 / 4 , − 1 2 σ + a ( 1 − b ) = 1 − 2 σ 2 σ , \gamma=b/\sigma,\qquad ab=1/4,\qquad-\frac{1}{2\sigma}+a(1-b)=\frac{1-2\sigma}{2\sigma}, γ = b / σ , ab = 1/4 , − 2 σ 1 + a ( 1 − b ) = 2 σ 1 − 2 σ ,
we ultimately deduce for x ∈ B 1 ∖ { 0 } x \in B_1 \setminus\{0\} x ∈ B 1 ∖ { 0 } the estimates
δ − 1 / ( 2 σ ) 1 ∣ x ∣ n − 1 ∫ B δ a / ∣ x ∣ ∖ B δ 1 / ( 2 σ ) ∣ ∂ ξ 1 n − 1 Θ ρ δ ( ξ δ 1 / ( 2 σ ) ) ∣ d ξ ≤ C δ − 1 / ( 2 σ ) 1 ∣ x ∣ n − 1 ∫ B δ a / ∣ x ∣ ∖ B δ 1 / ( 2 σ ) 1 ∣ ξ ∣ n − 1 + b d ξ = C δ ( 1 − 2 σ ) / ( 2 σ ) 1 ∣ x ∣ n − b , (114) \delta^{-1/(2\sigma)}\frac{1}{|x|^{n-1}}\int_{B_{\delta^a/|x|\setminus B_{\delta^{1/(2\sigma)}}} \left|\partial_{\xi_1}^{n-1}\Theta_\rho^\delta\left(\frac{\xi}{\delta^{1/(2\sigma)}}\right)\right|\,d\xi\le C\delta^{-1/(2\sigma)}\frac{1}{|x|^{n-1}}\int_{B_{\delta^a/|x|}\setminus B_{\delta^{1/(2\sigma)}}}\frac{1}{|\xi|^{n-1+b}}\,d\xi
= C\delta^{(1-2\sigma)/(2\sigma)}\frac{1}{|x|^{n-b}},} \tag*{(114)} δ − 1/ ( 2 σ ) ∣ x ∣ n − 1 1 ∫ B δ a /∣ x ∣ ∖ B δ 1/ ( 2 σ ) ∂ ξ 1 n − 1 Θ ρ δ ( δ 1/ ( 2 σ ) ξ ) d ξ ≤ C δ − 1/ ( 2 σ ) ∣ x ∣ n − 1 1 ∫ B δ a /∣ x ∣ ∖ B δ 1/ ( 2 σ ) ∣ ξ ∣ n − 1 + b 1 d ξ = C δ ( 1 − 2 σ ) / ( 2 σ ) ∣ x ∣ n − b 1 , (114)
and
1 ∣ x ∣ n ∫ ∂ B δ a / ∣ x ∣ ∣ ∂ ξ 1 n − 1 Θ ρ δ ( ξ δ 1 / 2 ) ∣ d S ( ξ ) ≤ C 1 ∣ x ∣ n ∫ ∂ B δ a / ∣ x ∣ 1 ∣ ξ ∣ n − 1 + b d S ( ξ ) = C δ − 1 / 4 1 ∣ x ∣ n − b , (115) \frac{1}{|x|^n}\int_{\partial B_{\delta^a/|x|}}\left|\partial_{\xi_1}^{n-1}\Theta_\rho^\delta\left(\frac{\xi}{\delta^{1/2}}\right)\right|\,dS(\xi) \le C\frac{1}{|x|^n}\int_{\partial B_{\delta^a/|x|}}\frac{1}{|\xi|^{n-1+b}}\,dS(\xi)=C\delta^{-1/4}\frac{1}{|x|^{n-b}}, \tag*{(115)} ∣ x ∣ n 1 ∫ ∂ B δ a /∣ x ∣ ∂ ξ 1 n − 1 Θ ρ δ ( δ 1/2 ξ ) d S ( ξ ) ≤ C ∣ x ∣ n 1 ∫ ∂ B δ a /∣ x ∣ ∣ ξ ∣ n − 1 + b 1 d S ( ξ ) = C δ − 1/4 ∣ x ∣ n − b 1 , (115)
and finally,
1 ∣ x ∣ n ∫ B δ a / ∣ x ∣ c ∣ ∂ ξ 1 n Θ ρ δ ( ξ δ 1 / ( 2 σ ) ) ∣ d ξ ≤ C 1 ∣ x ∣ n ∫ B δ a / ∣ x ∣ c 1 ∣ ξ ∣ n + b d ξ = C δ − 1 / 4 1 ∣ x ∣ n − b . (116) \frac{1}{|x|^n}\int_{B_{\delta^a/|x|}^c}\left|\partial_{\xi_1}^n\Theta_\rho^\delta\left(\frac{\xi}{\delta^{1/(2\sigma)}}\right)\right|\,d\xi\le C\frac{1}{|x|^n}\int_{B_{\delta^a/|x|}^c}\frac{1}{|\xi|^{n+b}}\,d\xi=C\delta^{-1/4}\frac{1}{|x|^{n-b}}. \tag*{(116)} ∣ x ∣ n 1 ∫ B δ a /∣ x ∣ c ∂ ξ 1 n Θ ρ δ ( δ 1/ ( 2 σ ) ξ ) d ξ ≤ C ∣ x ∣ n 1 ∫ B δ a /∣ x ∣ c ∣ ξ ∣ n + b 1 d ξ = C δ − 1/4 ∣ x ∣ n − b 1 . (116)
Since γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) , all powers of δ \delta δ in (5.10), (5.12) and (5.13) have an exponent in the range ( − 1 / 2 , 0 ) (-1/2,0) ( − 1/2 , 0 ) , and because 1 − 2 σ > − σ 1-2\sigma>-\sigma 1 − 2 σ > − σ , the exponent of δ \delta δ in (5.11) lies in ( − 1 / 2 , ∞ ) (-1/2,\infty) ( − 1/2 , ∞ ) . Thus, keeping in mind in view of (5.11) that δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , it holds
∣ ∫ R n Θ ρ δ ( ξ ) e 2 π i x ⋅ ξ d ξ ∣ ≤ C ∑ i = 1 3 δ − η i 1 ∣ x ∣ n − b and ∣ f ρ δ ( x ) ∣ ≤ C ∑ i = 1 3 δ n / 2 − η i 1 ∣ x ∣ n − b , x ∈ B 1 ∖ { 0 } \left|\int_{\mathbb{R}^n}\Theta_\rho^\delta(\xi)e^{2\pi i x\cdot\xi}\,d\xi\right|\le C\sum_{i=1}^{3}\delta^{-\eta_i}\frac{1}{|x|^{n-b}}
\quad\text{and}\quad
|f_\rho^\delta(x)|\le C\sum_{i=1}^{3}\delta^{n/2-\eta_i}\frac{1}{|x|^{n-b}},
\qquad x\in B_1\setminus\{0\} ∫ R n Θ ρ δ ( ξ ) e 2 π i x ⋅ ξ d ξ ≤ C i = 1 ∑ 3 δ − η i ∣ x ∣ n − b 1 and ∣ f ρ δ ( x ) ∣ ≤ C i = 1 ∑ 3 δ n /2 − η i ∣ x ∣ n − b 1 , x ∈ B 1 ∖ { 0 }
for some η 1 , η 2 , η 3 ∈ [ 0 , 1 / 2 ) \eta_1,\eta_2,\eta_3\in[0,1/2) η 1 , η 2 , η 3 ∈ [ 0 , 1/2 ) .
Step 4: proof of (5.6). In order to prove (5.6), we fix δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) . Integrating by parts, we write
F − 1 Θ ρ δ ( x ) = ∫ R n Θ ρ δ ( ξ ) e 2 π i x ⋅ ξ d ξ = ∫ R n Θ ρ δ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ = 1 ( 2 π i ∣ x ∣ ) n + 1 [ ∫ B 1 + ∫ B 1 c ] ∂ ξ 1 n + 1 Θ ρ δ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ , x ∈ R n . \begin{aligned}
\mathcal{F}^{-1}\Theta_\rho^\delta(x)=\int_{\mathbb{R}^n}\Theta_\rho^\delta(\xi)e^{2\pi i x\cdot\xi}\,d\xi=\int_{\mathbb{R}^n}\Theta_\rho^\delta(\xi)e^{2\pi i|x|\xi_1}\,d\xi\\
&=\frac{1}{(2\pi i|x|)^{n+1}}\left[\int_{B_1}+\int_{B_1^c}\right]\partial_{\xi_1}^{n+1}\Theta_\rho^\delta(\xi)e^{2\pi i|x|\xi_1}\,d\xi,
\qquad x\in\mathbb{R}^n.
\end{aligned} F − 1 Θ ρ δ ( x ) = ∫ R n Θ ρ δ ( ξ ) e 2 π i x ⋅ ξ d ξ = ∫ R n Θ ρ δ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ = ( 2 π i ∣ x ∣ ) n + 1 1 [ ∫ B 1 + ∫ B 1 c ] ∂ ξ 1 n + 1 Θ ρ δ ( ξ ) e 2 π i ∣ x ∣ ξ 1 d ξ , x ∈ R n .
With the choice α = ( n + 1 , 0 , … , 0 ) ∈ N 0 n \alpha=(n+1,0,\ldots,0)\in\mathbb{N}_0^n α = ( n + 1 , 0 , … , 0 ) ∈ N 0 n in (A.10), we infer that
∣ ∂ ξ 1 n + 1 Θ ρ δ ( ξ ) ∣ ≤ C , ξ ∈ B 1 , \left|\partial_{\xi_1}^{n+1}\Theta_\rho^\delta(\xi)\right|\le C,\qquad\xi\in B_1, ∂ ξ 1 n + 1 Θ ρ δ ( ξ ) ≤ C , ξ ∈ B 1 ,
and from (A.8) we get
∣ ∂ ξ 1 n + 1 Θ ρ δ ( ξ ) ∣ ≤ C ∣ ξ ∣ − ( n + 1 ) , ξ ∈ B 1 c . \left|\partial_{\xi_1}^{n+1}\Theta_\rho^\delta(\xi)\right|\le C|\xi|^{-(n+1)},\qquad\xi\in B_1^c. ∂ ξ 1 n + 1 Θ ρ δ ( ξ ) ≤ C ∣ ξ ∣ − ( n + 1 ) , ξ ∈ B 1 c .
Hence, we can estimate
∣ F − 1 Θ ρ δ ( x ) ∣ ≤ C 1 ∣ x ∣ n + 1 [ 1 + ∫ B 1 c ∣ ξ ∣ − ( n + 1 ) d ξ ] ≤ C 1 ∣ x ∣ n + 1 , x ∈ B 1 c , |\mathcal{F}^{-1}\Theta_\rho^\delta(x)|\le C\frac{1}{|x|^{n+1}}\left[1+\int_{B_1^c}|\xi|^{-(n+1)}\,d\xi\right]\le C\frac{1}{|x|^{n+1}},\qquad x\in B_1^c, ∣ F − 1 Θ ρ δ ( x ) ∣ ≤ C ∣ x ∣ n + 1 1 [ 1 + ∫ B 1 c ∣ ξ ∣ − ( n + 1 ) d ξ ] ≤ C ∣ x ∣ n + 1 1 , x ∈ B 1 c ,
proving (5.6).
Step 5: proof of (5.7). As already pointed out, we resort for the proof of (5.7) to a different technique. We use that it follows from the relation f ρ δ = δ n / 2 F − 1 Θ ρ δ f_\rho^\delta=\delta^{n/2}\mathcal{F}^{-1}\Theta_\rho^\delta f ρ δ = δ n /2 F − 1 Θ ρ δ and the joint change of variables ξ ↦ δ − 1 / 2 ξ \xi\mapsto\delta^{-1/2}\xi ξ ↦ δ − 1/2 ξ , x ↦ δ 1 / 2 x x\mapsto\delta^{1/2}x x ↦ δ 1/2 x that
∫ R n ∣ f ρ δ ( x ) ∣ d x = δ n / 2 ∫ R n ∣ ( F − 1 Θ ρ δ ( δ − 1 / 2 ⋅ ) ) ( x ) ∣ d x . (117) \int_{\mathbb{R}^n}|f_\rho^\delta(x)|\,dx=\delta^{n/2}\int_{\mathbb{R}^n}|(\mathcal{F}^{-1}\Theta_\rho^\delta(\delta^{-1/2}\cdot))(x)|\,dx. \tag*{(117)} ∫ R n ∣ f ρ δ ( x ) ∣ d x = δ n /2 ∫ R n ∣ ( F − 1 Θ ρ δ ( δ − 1/2 ⋅ )) ( x ) ∣ d x . (117)
For simplicity, we define
Θ ~ ρ δ ( ξ ) : = Θ ρ δ ( δ − 1 / 2 ξ ) = e − ( ∣ ξ ∣ 2 Q ^ ρ ( δ − 1 / 2 ξ ) ) / ( 4 π 2 ) , ξ ∈ R n . \widetilde{\Theta}_\rho^\delta(\xi):=\Theta_\rho^\delta(\delta^{-1/2}\xi)=e^{-(|\xi|^2\widehat{Q}_\rho(\delta^{-1/2}\xi))/(4\pi^2)},\qquad\xi\in\mathbb{R}^n. Θ ρ δ ( ξ ) := Θ ρ δ ( δ − 1/2 ξ ) = e − ( ∣ ξ ∣ 2 Q ρ ( δ − 1/2 ξ )) / ( 4 π 2 ) , ξ ∈ R n .
It then follows from (5.14) and the rotational symmetry of Q ^ ρ \widehat{Q}_\rho Q ρ that
∥ f ρ δ ∥ L 1 ( R n ) = δ n / 2 ∥ F Θ ~ ρ δ ∥ L 1 ( R n ) . \|f_\rho^\delta\|_{L^1(\mathbb{R}^n)}=\delta^{n/2}\|\mathcal{F}\widetilde{\Theta}_\rho^\delta\|_{L^1(\mathbb{R}^n)}. ∥ f ρ δ ∥ L 1 ( R n ) = δ n /2 ∥ F Θ ρ δ ∥ L 1 ( R n ) .
Next, we choose some k > n / 2 k>n/2 k > n /2 and observe that it holds by Hölder’s inequality and Plancherel’s identity that
∥ F Θ ~ ρ δ ∥ L 1 ( R n ) ≤ ∥ ( 1 + ∣ ⋅ ∣ 2 ) − k / 2 ∥ L 2 ( R n ) ∥ ( 1 + ∣ ⋅ ∣ 2 ) k / 2 F Θ ~ ρ δ ∥ L 2 ( R n ) = ∥ ( 1 + ∣ ⋅ ∣ 2 ) − k / 2 ∥ L 2 ( R n ) ∥ Θ ~ ρ δ ∥ W 2 , k ( R n ) , \left\|\mathcal{F}\widetilde{\Theta}^{\delta}_{\rho}\right\|_{L^{1}(\mathbb{R}^{n})}\leq\left\|(1+|\cdot|^{2})^{-k/2}\right\|_{L^{2}(\mathbb{R}^{n})}\left\|(1+|\cdot|^{2})^{k/2}\mathcal{F}\widetilde{\Theta}^{\delta}_{\rho}\right\|_{L^{2}(\mathbb{R}^{n})}
=\left\|(1+|\cdot|^{2})^{-k/2}\right\|_{L^{2}(\mathbb{R}^{n})}\left\|\widetilde{\Theta}^{\delta}_{\rho}\right\|_{W^{2,k}(\mathbb{R}^{n})}, F Θ ρ δ L 1 ( R n ) ≤ ( 1 + ∣ ⋅ ∣ 2 ) − k /2 L 2 ( R n ) ( 1 + ∣ ⋅ ∣ 2 ) k /2 F Θ ρ δ L 2 ( R n ) = ( 1 + ∣ ⋅ ∣ 2 ) − k /2 L 2 ( R n ) Θ ρ δ W 2 , k ( R n ) ,
where ∥ ( 1 + ∣ ⋅ ∣ 2 ) − k / 2 ∥ L 2 ( R n ) \left\|(1+|\cdot|^{2})^{-k/2}\right\|_{L^{2}(\mathbb{R}^{n})} ( 1 + ∣ ⋅ ∣ 2 ) − k /2 L 2 ( R n ) is finite. In the last step, we also used the characterization of the integer order Sobolev spaces W k , 2 ( R n ) W^{k,2}(\mathbb{R}^{n}) W k , 2 ( R n ) as Bessel potential spaces (see, e.g. [18 Section 1.3.1] ). We conclude that in order to prove (110) , it suffices to show that
∥ Θ ~ ρ δ ∥ W k , 2 ( R n ) < ∞ . (118) \left\|\widetilde{\Theta}^{\delta}_{\rho}\right\|_{W^{k,2}(\mathbb{R}^{n})}<\infty. \tag*{(118)} Θ ρ δ W k , 2 ( R n ) < ∞. (118)
Splitting R n \mathbb{R}^{n} R n into B δ 1 / 2 B_{\delta^{1/2}} B δ 1/2 and B δ 1 / 2 c B_{\delta^{1/2}}^{c} B δ 1/2 c , we use (A.11) and (A.12) to find for all α ∈ N 0 \alpha\in\mathbb{N}_{0} α ∈ N 0 with ∣ α ∣ ≤ k |\alpha|\leq k ∣ α ∣ ≤ k the estimates
∫ B δ 1 / 2 ∣ ∂ α Θ ~ ρ δ ( ξ ) ∣ d ξ ≤ C k ∫ R n ( 1 + ∣ ξ ∣ ) 2 ∣ α ∣ e − c ρ ∣ ξ ∣ 2 d ξ \int_{B_{\delta^{1/2}}}\left|\partial^{\alpha}\widetilde{\Theta}^{\delta}_{\rho}(\xi)\right|\,\mathrm{d}\xi\leq C_{k}\int_{\mathbb{R}^{n}}(1+|\xi|)^{2|\alpha|}e^{-c_{\rho}|\xi|^{2}}\,\mathrm{d}\xi ∫ B δ 1/2 ∂ α Θ ρ δ ( ξ ) d ξ ≤ C k ∫ R n ( 1 + ∣ ξ ∣ ) 2∣ α ∣ e − c ρ ∣ ξ ∣ 2 d ξ
and
∫ B δ 1 / 2 c ∣ ∂ α Θ ~ ρ δ ( ξ ) ∣ d ξ ≤ C k ∫ R n e − c ρ ∣ ξ ∣ 2 σ d ξ . \int_{B_{\delta^{1/2}}^{c}}\left|\partial^{\alpha}\widetilde{\Theta}^{\delta}_{\rho}(\xi)\right|\,\mathrm{d}\xi\leq C_{k}\int_{\mathbb{R}^{n}}e^{-c_{\rho}|\xi|^{2\sigma}}\,\mathrm{d}\xi. ∫ B δ 1/2 c ∂ α Θ ρ δ ( ξ ) d ξ ≤ C k ∫ R n e − c ρ ∣ ξ ∣ 2 σ d ξ .
Since the right-hand sides are bounded independently of δ ∈ ( 1 , ∞ ) \delta\in(1,\infty) δ ∈ ( 1 , ∞ ) , we indeed verify (118) . □
Remark 5.2. We comment here on the difference in the constriction of the functions ( g ρ , ℓ ) ℓ (g_{\rho,\ell})_{\ell} ( g ρ , ℓ ) ℓ between the general nonlocal gradients case and the Riesz fractional gradient with kernel ρ s \rho_{s} ρ s (see § 1.1 ). Precisely, for the Riesz fractional gradient it can be quickly observed that g ρ s , ℓ g_{\rho_{s},\ell} g ρ s , ℓ is non-negative and that ∥ g ρ s , ℓ ∥ L 1 ( R n ) = 1 \|g_{\rho_{s},\ell}\|_{L^{1}(\mathbb{R}^{n})}=1 ∥ g ρ s , ℓ ∥ L 1 ( R n ) = 1 for any ℓ ∈ N \ell\in\mathbb{N} ℓ ∈ N (see [26 Chapter 7] ). In particular, the integrability already follows from the non-negativity, since then
∥ g ρ s , ℓ ∥ L 1 ( R n ) = g ^ ρ s , ℓ ( 0 ) = ( 1 + 4 π 2 ∣ ξ ∣ 2 ) − s / 2 ∣ ξ = 0 = 1 , ℓ ∈ N . \|g_{\rho_{s},\ell}\|_{L^{1}(\mathbb{R}^{n})}=\widehat{g}_{\rho_{s},\ell}(0)=(1+4\pi^{2}|\xi|^{2})^{-s/2}\big|_{\xi=0}=1,\qquad\ell\in\mathbb{N}. ∥ g ρ s , ℓ ∥ L 1 ( R n ) = g ρ s , ℓ ( 0 ) = ( 1 + 4 π 2 ∣ ξ ∣ 2 ) − s /2 ξ = 0 = 1 , ℓ ∈ N . Following the same idea for any ρ \rho ρ as in Proposition 5.1 , one encounters the problem that the non-negativity of the g ρ , ℓ g_{\rho,\ell} g ρ , ℓ seems difficult to verify in the general case. In fact, with (105) and (107) , one would like to show that f ρ δ = F − 1 h ρ δ f_{\rho}^{\delta}=\mathcal{F}^{-1}h_{\rho}^{\delta} f ρ δ = F − 1 h ρ δ is non-negative, where, as the function ∣ λ ρ ∣ |\lambda_{\rho}| ∣ λ ρ ∣ is even and with the help of the coarea formula, the expression for f ρ δ f_{\rho}^{\delta} f ρ δ can be written as an oscillating integral of cos-type, this is
f ρ δ ( x ) = 2 δ n / 2 ∫ S + n − 1 ∫ 0 ∞ r n − 1 e − r 2 q ( r ) 2 δ cos ( 2 π r ∣ x ∣ z ) d r d S ( z ) , x ∈ R n , f_{\rho}^{\delta}(x)=2\delta^{n/2}\int_{\mathbb{S}_{+}^{n-1}}\int_{0}^{\infty}r^{n-1}e^{-r^{2}q(r)^{2\delta}}\cos(2\pi r|x|z)\,\mathrm{d}r\,\mathrm{d}S(z),\qquad x\in\mathbb{R}^{n}, f ρ δ ( x ) = 2 δ n /2 ∫ S + n − 1 ∫ 0 ∞ r n − 1 e − r 2 q ( r ) 2 δ cos ( 2 π r ∣ x ∣ z ) d r d S ( z ) , x ∈ R n , denoting by q q q the radial representative of Q ^ ρ \widehat{Q}_{\rho} Q ρ . But this remains an open problem for the moment, and instead, we demonstrate that the functions g ρ , ℓ g_{\rho,\ell} g ρ , ℓ lie in L 1 ( R n ) L^{1}(\mathbb{R}^{n}) L 1 ( R n ) via explicit decay and norm estimates for f ρ δ f_{\rho}^{\delta} f ρ δ (cf. (108) , (109) and (110) ).
The positivity of the functions g ρ , ℓ g_{\rho,\ell} g ρ , ℓ is also interesting in a different regard: if g ρ , ℓ g_{\rho,\ell} g ρ , ℓ is positive, then it defines a probability distribution and ( 1 + ∣ λ ρ ∣ 2 ) − ℓ / 2 (1+|\lambda_{\rho}|^{2})^{-\ell/2} ( 1 + ∣ λ ρ ∣ 2 ) − ℓ /2 can be seen as its characteristic function (see again § 1.1 for the analogous statement in the case of the Bessel potential spaces). Connected to the above-mentioned difficulties, the question if this holds in the case of general nonlocal gradients also remains open.
We now turn to the main result of this section. It relies on the alternative characterization of H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^{n}) H ρ , p ( R n ) obtained in Theorem 3.1 together with the existence of g ρ , 1 g_{\rho,1} g ρ , 1 according to Proposition 5.1 .
Theorem 5.3 (Potential space characterization). Let p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and assume that ρ \rho ρ has compact support and satisfies assumptions ( H 0 ) (H0) ( H 0 ) –( H 3 ) (H3) ( H 3 ) . Denote by g ρ ∈ L 1 ( R n ) g_{\rho}\in L^{1}(\mathbb{R}^{n}) g ρ ∈ L 1 ( R n ) the function g ρ , 1 g_{\rho,1} g ρ , 1 from Proposition 5.1 such that
g ^ ρ = [ λ ρ ] − 1 . \widehat{g}_{\rho}=[\lambda_{\rho}]^{-1}. g ρ = [ λ ρ ] − 1 . Then
H ρ , p ( R n ) = { g ρ ∗ f : f ∈ L p ( R n ) } H^{\rho,p}(\mathbb{R}^{n})=\{g_{\rho}*f:f\in L^{p}(\mathbb{R}^{n})\} H ρ , p ( R n ) = { g ρ ∗ f : f ∈ L p ( R n )}
and, for all f ∈ L p ( R n ) f \in L^p(\mathbb{R}^n) f ∈ L p ( R n ) ,
∥ F − 1 ( ⟦ λ ρ ⟧ ( g ρ ∗ f ^ ) ) ∥ L p ( R n ) = ∥ f ∥ L p ( R n ) . \left\|\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket(\widehat{g_\rho* f})\right)\right\|_{L^p(\mathbb{R}^n)}=\|f\|_{L^p(\mathbb{R}^n)}. F − 1 ( [ [ λ ρ ] ] ( g ρ ∗ f ) ) L p ( R n ) = ∥ f ∥ L p ( R n ) .
Proof. We establish the statement by showing that the map
L p ( R n ) → H ρ , p ( R n ) , f ↦ g ρ ∗ f L^p(\mathbb{R}^n) \to H^{\rho,p}(\mathbb{R}^n),\qquad f\mapsto g_\rho* f L p ( R n ) → H ρ , p ( R n ) , f ↦ g ρ ∗ f
is a bounded linear isomorphism with inverse
H ρ , p ( R n ) → L p ( R n ) , u ↦ F − 1 ( ⟦ λ ρ ⟧ u ^ ) , H^{\rho,p}(\mathbb{R}^n) \to L^p(\mathbb{R}^n),\qquad u\mapsto\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{u}\right), H ρ , p ( R n ) → L p ( R n ) , u ↦ F − 1 ( [ [ λ ρ ] ] u ) ,
where the last expression is understood in the sense of S ′ ( R n ) \mathcal{S}'(\mathbb{R}^n) S ′ ( R n ) and lies in L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) thanks to Theorem 3.1. We first prove that g ρ ∗ f ∈ H ρ , p ( R n ) g_\rho* f\in H^{\rho,p}(\mathbb{R}^n) g ρ ∗ f ∈ H ρ , p ( R n ) for any f ∈ L p ( R n ) f\in L^p(\mathbb{R}^n) f ∈ L p ( R n ) . Firstly, by Young’s inequality, g ρ ∗ f ∈ L p ( R n ) g_\rho* f\in L^p(\mathbb{R}^n) g ρ ∗ f ∈ L p ( R n ) . Secondly, for any φ ∈ S ( R n ) \varphi\in\mathcal{S}(\mathbb{R}^n) φ ∈ S ( R n ) ,
⟨ F − 1 ( ⟦ λ ρ ⟧ ( g ρ ∗ f ^ ) ) , φ ⟩ S ′ × S = ⟨ g ρ ∗ f , F ( ⟦ λ ρ ⟧ F − 1 φ ) ⟩ = ⟨ g ρ ∗ f , F − 1 ( ⟦ λ ρ ⟧ φ ^ ) ⟩ = ⟨ f , g ρ ∗ F − 1 ( ⟦ λ ρ ⟧ φ ^ ) ⟩ = ⟨ f , F − 1 ( g ρ ^ ⟦ λ ρ ⟧ φ ^ ) ⟩ = ⟨ f , φ ⟩ , \begin{aligned}
\left\langle\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket(\widehat{g_\rho* f})\right),\varphi\right\rangle_{\mathcal{S}'\times\mathcal{S}}
&=\left\langle g_\rho* f,\mathcal{F}\left(\llbracket\lambda_\rho\rrbracket\mathcal{F}^{-1}\varphi\right)\right\rangle\\
&=\left\langle g_\rho* f,\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{\varphi}\right)\right\rangle\\
&=\left\langle f,g_\rho*\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{\varphi}\right)\right\rangle\\
&=\left\langle f,\mathcal{F}^{-1}\left(\widehat{g_\rho}\llbracket\lambda_\rho\rrbracket\widehat{\varphi}\right)\right\rangle
=\langle f,\varphi\rangle,
\end{aligned} ⟨ F − 1 ( [ [ λ ρ ] ] ( g ρ ∗ f ) ) , φ ⟩ S ′ × S = ⟨ g ρ ∗ f , F ( [ [ λ ρ ] ] F − 1 φ ) ⟩ = ⟨ g ρ ∗ f , F − 1 ( [ [ λ ρ ] ] φ ) ⟩ = ⟨ f , g ρ ∗ F − 1 ( [ [ λ ρ ] ] φ ) ⟩ = ⟨ f , F − 1 ( g ρ [ [ λ ρ ] ] φ ) ⟩ = ⟨ f , φ ⟩ ,
where we used additionally the first identity of Lemma 2.1 (i) in the second equality, and for the last equality that F − 1 ( ⟦ λ ρ ⟧ φ ^ ) ∈ S ( R n ) \mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{\varphi}\right)\in\mathcal{S}(\mathbb{R}^n) F − 1 ( [ [ λ ρ ] ] φ ) ∈ S ( R n ) . Thus, by Theorem 3.1, g ρ ∗ f ∈ H ρ , p ( R n ) g_\rho* f\in H^{\rho,p}(\mathbb{R}^n) g ρ ∗ f ∈ H ρ , p ( R n ) .
To see the injectivity of f ↦ g ρ ∗ f f\mapsto g_\rho* f f ↦ g ρ ∗ f , we assume that g ρ ∗ f = 0 g_\rho* f=0 g ρ ∗ f = 0 for some f ∈ L p ( R n ) f\in L^p(\mathbb{R}^n) f ∈ L p ( R n ) . It follows then directly from Theorem 3.1 that
0 = ∥ g ρ ∗ f ∥ H ρ , p ( R n ) ≥ c p ∥ F − 1 ( ⟦ λ ρ ⟧ g ρ ^ f ^ ) ∥ L p ( R n ) = c p ∥ f ∥ L p ( R n ) , 0=\|g_\rho* f\|_{H^{\rho,p}(\mathbb{R}^n)}\geq c_p\left\|\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{g_\rho}\widehat{f}\right)\right\|_{L^p(\mathbb{R}^n)}=c_p\|f\|_{L^p(\mathbb{R}^n)}, 0 = ∥ g ρ ∗ f ∥ H ρ , p ( R n ) ≥ c p F − 1 ( [ [ λ ρ ] ] g ρ f ) L p ( R n ) = c p ∥ f ∥ L p ( R n ) ,
implying that f ≡ 0 f\equiv0 f ≡ 0 . Consequently, f ↦ g ρ ∗ f f\mapsto g_\rho* f f ↦ g ρ ∗ f is injective.
Turning our attention to surjectivity, we note that by the density of C c ∞ ( R n ) C^\infty_c(\mathbb{R}^n) C c ∞ ( R n ) according to [8 Theorem 3.9] , there exists for every u ∈ H ρ , p ( R n ) u\in H^{\rho,p}(\mathbb{R}^n) u ∈ H ρ , p ( R n ) a sequence ( u j ) j ⊂ C c ∞ ( R n ) (u_j)_j\subset C^\infty_c(\mathbb{R}^n) ( u j ) j ⊂ C c ∞ ( R n ) such that u j → u u_j\to u u j → u in H ρ , p ( R n ) H^{\rho,p}(\mathbb{R}^n) H ρ , p ( R n ) and, with Theorem 3.1,
F − 1 ( ⟦ λ ρ ⟧ u j ^ ) → F − 1 ( ⟦ λ ρ ⟧ u ^ ) in L p ( R n ) . \mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{u_j}\right)\to\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{u}\right)\quad\text{in }L^p(\mathbb{R}^n). F − 1 ( [ [ λ ρ ] ] u j ) → F − 1 ( [ [ λ ρ ] ] u ) in L p ( R n ) .
Using that convolution with g ρ g_\rho g ρ is continuous by Young’s inequality, we infer that
g ρ ∗ F − 1 ( ⟦ λ ρ ⟧ u j ^ ) → g ρ ∗ F − 1 ( ⟦ λ ρ ⟧ u ^ ) in L p ( R n ) , g_\rho*\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{u_j}\right)\to g_\rho*\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{u}\right)\quad\text{in }L^p(\mathbb{R}^n), g ρ ∗ F − 1 ( [ [ λ ρ ] ] u j ) → g ρ ∗ F − 1 ( [ [ λ ρ ] ] u ) in L p ( R n ) ,
while
g ρ ∗ F − 1 ( ⟦ λ ρ ⟧ u j ^ ) = u j → u in L p ( R n ) . g_\rho*\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{u_j}\right)=u_j\to u\quad\text{in }L^p(\mathbb{R}^n). g ρ ∗ F − 1 ( [ [ λ ρ ] ] u j ) = u j → u in L p ( R n ) .
We conclude that g ρ ∗ F − 1 ( ⟦ λ ρ ⟧ u ^ ) = u g_\rho*\mathcal{F}^{-1}\left(\llbracket\lambda_\rho\rrbracket\widehat{u}\right)=u g ρ ∗ F − 1 ( [ [ λ ρ ] ] u ) = u . The proof is finished. □
Having the sequence of functions ( g ρ , ℓ ) ℓ (g_{\rho,\ell})_\ell ( g ρ , ℓ ) ℓ uniformly bounded in L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) at our disposal, an alternative proof of Theorem 5.3 can be obtained following the arguments in [37 Chapter V.3] (see also [26 Chapter 7] and [33 Theorem 1.7] ), where the coincidence of the nonlocal Sobolev spaces H s , p ( R n ) H^{s,p}(\mathbb{R}^n) H s , p ( R n ) with the Bessel potential spaces is shown. The only missing ingredient is the subsequent result regarding the fractions of the generalized Riesz and Bessel potentials. The proof is analogous to the one of Stein in [37 Chapter V.3, Lemma 2] in the case of the classical Riesz and Bessel potentials (see also [26 Section 7, Lemma 2.1 and Lemma 2.2] ).
Lemma 5.4. Let p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and assume that ρ \rho ρ has compact support and satisfy assumptions (H0)–(H3). Then the fractions
∣ λ ρ ∣ ⟦ λ ρ ⟧ and ⟦ λ ρ ⟧ 1 + ∣ λ ρ ∣ \frac{|\lambda_\rho|}{\llbracket\lambda_\rho\rrbracket}
\qquad\text{and}\qquad
\frac{\llbracket\lambda_\rho\rrbracket}{1+|\lambda_\rho|} [ [ λ ρ ] ] ∣ λ ρ ∣ and 1 + ∣ λ ρ ∣ [ [ λ ρ ] ] admit Fourier inverses in the sense of finite measures.
Proof. The starting point of the proof is the binomial series
( 1 − t ) 1 / 2 = 1 + ∑ j = 1 ∞ a j t j , ∣ t ∣ < 1 , (119) (1-t)^{1/2}=1+\sum_{j=1}^{\infty}a_jt^j,\qquad|t|<1, \tag*{(119)} ( 1 − t ) 1/2 = 1 + j = 1 ∑ ∞ a j t j , ∣ t ∣ < 1 , (119)
for suitable coefficients ( a j ) j ⊂ R (a_j)_j\subset\mathbb{R} ( a j ) j ⊂ R with ∑ j = 1 ∞ ∣ a j ∣ < ∞ \sum_{j=1}^{\infty}|a_j|<\infty ∑ j = 1 ∞ ∣ a j ∣ < ∞ . This enables us to write
∣ λ ρ ( ξ ) ∣ ⟦ λ ρ ( ξ ) ⟧ = 1 + ∑ j = 1 ∞ a j ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) − j , ξ ∈ R n ∖ { 0 } , \frac{|\lambda_\rho(\xi)|}{\llbracket\lambda_\rho(\xi)\rrbracket}=1+\sum_{j=1}^{\infty}a_j(1+|\lambda_\rho(\xi)|^2)^{-j},\qquad\xi\in\mathbb{R}^n\setminus\{0\}, [ [ λ ρ ( ξ ) ] ] ∣ λ ρ ( ξ ) ∣ = 1 + j = 1 ∑ ∞ a j ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) − j , ξ ∈ R n ∖ { 0 } ,
where we chose t = 1 / ⟦ λ ρ ( ξ ) ⟧ 2 t=1/\llbracket\lambda_\rho(\xi)\rrbracket^2 t = 1/ [ [ λ ρ ( ξ ) ] ] 2 for ξ ∈ R n ∖ { 0 } \xi\in\mathbb{R}^n\setminus\{0\} ξ ∈ R n ∖ { 0 } in (5.17). Let ( g ρ , ℓ ) ℓ (g_{\rho,\ell})_\ell ( g ρ , ℓ ) ℓ be the family of Proposition 5.1. Then ( g ρ , ℓ ) ℓ (g_{\rho,\ell})_\ell ( g ρ , ℓ ) ℓ is bounded in L 1 ( R n ) L^1(\mathbb{R}^n) L 1 ( R n ) and g ρ , 2 j ^ = ( 1 + ∣ λ ρ ∣ 2 ) − j \widehat{g_{\rho,2j}}=(1+|\lambda_\rho|^2)^{-j} g ρ , 2 j = ( 1 + ∣ λ ρ ∣ 2 ) − j for all j ∈ N j\in\mathbb{N} j ∈ N . Consequently, defining
μ 1 : = δ 0 + ∑ j = 1 ∞ a j g ρ , 2 j d L n , \mu_1:=\delta_0+\sum_{j=1}^{\infty}a_jg_{\rho,2j}\,d\mathcal{L}^n, μ 1 := δ 0 + j = 1 ∑ ∞ a j g ρ , 2 j d L n ,
where δ 0 \delta_0 δ 0 is the Dirac measure at the origin and L n \mathcal{L}^n L n denotes the Lebesgue measure on R n \mathbb{R}^n R n , it holds
μ 1 ( R n ) < ∞ and μ 1 ^ = ∣ λ ρ ∣ ⟦ λ ρ ⟧ . \mu_1(\mathbb{R}^n)<\infty\qquad\text{and}\qquad\widehat{\mu_1}=\frac{|\lambda_\rho|}{\llbracket\lambda_\rho\rrbracket}. μ 1 ( R n ) < ∞ and μ 1 = [ [ λ ρ ] ] ∣ λ ρ ∣ .
As for the second fraction in (5.16), we invoke Wiener’s theorem (see e.g., [26 Theorem 2.3] ), according to which for any Φ 1 ∈ L 1 ( R n ) \Phi_1\in L^1(\mathbb{R}^n) Φ 1 ∈ L 1 ( R n ) such that 1 + Φ 1 ^ ( ξ ) ≠ 0 1+\widehat{\Phi_1}(\xi)\ne0 1 + Φ 1 ( ξ ) = 0 for all ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n , there exists Φ 2 ∈ L 1 ( R n ) \Phi_2\in L^1(\mathbb{R}^n) Φ 2 ∈ L 1 ( R n ) such that
( 1 + Φ 1 ^ ) − 1 = 1 + Φ 2 ^ . (1+\widehat{\Phi_1})^{-1}=1+\widehat{\Phi_2}. ( 1 + Φ 1 ) − 1 = 1 + Φ 2 .
In particular, for
Φ 1 : = g ρ , 1 + ∑ j = 1 ∞ a j g ρ , 2 j , \Phi_1:=g_{\rho,1}+\sum_{j=1}^{\infty}a_jg_{\rho,2j}, Φ 1 := g ρ , 1 + j = 1 ∑ ∞ a j g ρ , 2 j ,
it holds
1 + Φ 1 ^ = 1 + ∣ λ ρ ∣ ⟦ λ ρ ⟧ > 0. 1+\widehat{\Phi_1}=\frac{1+|\lambda_\rho|}{\llbracket\lambda_\rho\rrbracket}>0. 1 + Φ 1 = [ [ λ ρ ] ] 1 + ∣ λ ρ ∣ > 0.
Inverting this expression with the help of Wiener’s theorem yields a function Φ 2 ∈ L 1 ( R n ) \Phi_2\in L^1(\mathbb{R}^n) Φ 2 ∈ L 1 ( R n ) such that, with
μ 2 : = δ 0 + Φ 2 d L n , \mu_2:=\delta_0+\Phi_2\,d\mathcal{L}^n, μ 2 := δ 0 + Φ 2 d L n ,
we deduce
μ 2 ^ = ⟦ λ ρ ⟧ 1 + ∣ λ ρ ∣ , \widehat{\mu_2}=\frac{\llbracket\lambda_\rho\rrbracket}{1+|\lambda_\rho|}, μ 2 = 1 + ∣ λ ρ ∣ [ [ λ ρ ] ] ,
which concludes the proof. □
We close with a remark regarding an alternative proof strategy (under slightly stronger assumptions) for the Fourier multiplier result Proposition 2.10.
Remark 5.5. In the case that ρ \rho ρ has compact support and fulfills (H0)–(H3), in order to prove Proposition 2.10, one can also argue more directly by means of Proposition 5.1 and Lemma 5.4. Indeed, for Proposition 2.10 (iii) one can use Proposition 5.1 to write for any u ∈ S ( R n ) u\in\mathcal{S}(\mathbb{R}^n) u ∈ S ( R n ) (with g ρ : = g ρ , 1 g_\rho:=g_{\rho,1} g ρ := g ρ , 1 )
F − 1 ( [ λ ρ ] − 1 u ^ ) = g ρ ∗ u , \mathcal{F}^{-1}([\lambda_\rho]^{-1}\widehat{u})=g_\rho*u, F − 1 ([ λ ρ ] − 1 u ) = g ρ ∗ u ,
and by Young’s convolution inequality it holds immediately that
∥ g ρ ∗ u ∥ L p ( R n ) ≤ ∥ g ρ ∥ L 1 ( R n ) ∥ u ∥ L p ( R n ) . \|g_\rho*u\|_{L^p(\mathbb{R}^n)}\le\|g_\rho\|_{L^1(\mathbb{R}^n)}\|u\|_{L^p(\mathbb{R}^n)}. ∥ g ρ ∗ u ∥ L p ( R n ) ≤ ∥ g ρ ∥ L 1 ( R n ) ∥ u ∥ L p ( R n ) .
A similar approach proves Proposition 2.10 (i) and (ii), since the respective fractions actually admit Fourier inverses in the sense of finite measures by Lemma 5.4.
Appendix A. Proofs of Lemmas 2.8 and A.1
Recall that λ ρ \lambda_{\rho} λ ρ , the Fourier symbol of the nonlocal gradient D 0 ρ D_{0}^{\rho} D 0 ρ , is given by the expression (see § 2.3 )
λ ρ ( ξ ) = 2 π i ξ Q ^ ρ ( ξ ) , ξ ∈ R n . \lambda_{\rho}(\xi)=2\pi i\xi\widehat{Q}_{\rho}(\xi),\qquad\xi\in\mathbb{R}^{n}. λ ρ ( ξ ) = 2 π i ξ Q ρ ( ξ ) , ξ ∈ R n .
In the following, we will denote for some multi-index α ∈ N 0 n \alpha\in\mathbb{N}_{0}^{n} α ∈ N 0 n by C α > 0 C_{\alpha}>0 C α > 0 a positive constant that may change from line to line and only depends on α \alpha α (and implicitly on the dimension n ∈ N n\in\mathbb{N} n ∈ N ).
Proof of Lemma 2.8. For α = 0 \alpha=0 α = 0 it is readily seen that all inequalities are fulfilled with suitable constants C 0 > 0 C_{0}>0 C 0 > 0 . We point out that regarding (40) , one needs to use that ∣ ξ ∣ Q ^ ρ ( ξ ) ≥ C |\xi|\widehat{Q}_{\rho}(\xi)\geq C ∣ ξ ∣ Q ρ ( ξ ) ≥ C for ξ ∈ B 1 c \xi\in B_{1}^{c} ξ ∈ B 1 c due to the assumptions and (35) , and thus, 1 + ∣ λ ρ ∣ 2 ≤ C 0 ∣ λ ρ ∣ 2 1+|\lambda_{\rho}|^{2}\leq C_{0}|\lambda_{\rho}|^{2} 1 + ∣ λ ρ ∣ 2 ≤ C 0 ∣ λ ρ ∣ 2 in B 1 c B_{1}^{c} B 1 c .
Pick now some α ∈ N 0 n ∖ { 0 } \alpha\in\mathbb{N}_{0}^{n}\setminus\{0\} α ∈ N 0 n ∖ { 0 } . We repeat that, since by assumption ρ \rho ρ has compact support, the same is true for Q ρ Q_{\rho} Q ρ , and Q ^ ρ \widehat{Q}_{\rho} Q ρ is smooth and, in particular, all subsequently computed derivatives will be well defined.
We start by recalling a version of Faà di Bruno’s formula for the derivatives of compositions, which we will repeatedly rely on it throughout the proof. Precisely, one has for any f ∈ C ∞ ( R n ) f\in C^{\infty}(\mathbb{R}^{n}) f ∈ C ∞ ( R n ) and g : R → R g:\mathbb{R}\to\mathbb{R} g : R → R such that g ∘ f ∈ C ∞ ( R ) g\circ f\in C^{\infty}(\mathbb{R}) g ∘ f ∈ C ∞ ( R ) , and any multi-index α ∈ N 0 n ∖ { 0 } \alpha\in\mathbb{N}_{0}^{n}\setminus\{0\} α ∈ N 0 n ∖ { 0 } the estimate
∣ ∂ α ( g ∘ f ) ( ξ ) ∣ ≤ C α ∑ k = 1 ∣ α ∣ ∣ ( ( ∂ k g ) ∘ f ) ( ξ ) ∣ ∑ { α 1 ′ , … , α k ′ } ∏ l = 1 k ∣ ∂ α l ′ f ( ξ ) ∣ , ξ ∈ R n . (120) \left|\partial^{\alpha}(g\circ f)(\xi)\right|\leq C_{\alpha}\sum_{k=1}^{|\alpha|}\left|((\partial^{k}g)\circ f)(\xi)\right|\sum_{\{\alpha_{1}^{\prime},\ldots,\alpha_{k}^{\prime}\}}\prod_{l=1}^{k}\left|\partial^{\alpha_{l}^{\prime}}f(\xi)\right|,\qquad\xi\in\mathbb{R}^{n}. \tag*{(120)} ∣ ∂ α ( g ∘ f ) ( ξ ) ∣ ≤ C α k = 1 ∑ ∣ α ∣ (( ∂ k g ) ∘ f ) ( ξ ) { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∂ α l ′ f ( ξ ) , ξ ∈ R n . (120)
where and the second sum runs over all partitions { α 1 ′ , … , α k ′ } \{\alpha_{1}^{\prime},\ldots,\alpha_{k}^{\prime}\} { α 1 ′ , … , α k ′ } of α \alpha α into k k k multi-indices. By a partition of α \alpha α , we mean a set { α 1 ′ , … , α N ′ } \{\alpha_{1}^{\prime},\ldots,\alpha_{N}^{\prime}\} { α 1 ′ , … , α N ′ } of multi-indices α i ′ ∈ N 0 n ∖ { 0 } \alpha_{i}^{\prime}\in\mathbb{N}_{0}^{n}\setminus\{0\} α i ′ ∈ N 0 n ∖ { 0 } , N ∈ N N\in\mathbb{N} N ∈ N with N ≤ ∣ α ∣ N\leq|\alpha| N ≤ ∣ α ∣ , such that α 1 ′ + … + α N ′ = α \alpha_{1}^{\prime}+\ldots+\alpha_{N}^{\prime}=\alpha α 1 ′ + … + α N ′ = α , where the summation is understood component-wise in the sense of multi-indices. In particular, the partitions { α 1 ′ , … , α k ′ } \{\alpha_{1}^{\prime},\ldots,\alpha_{k}^{\prime}\} { α 1 ′ , … , α k ′ } in the inner sum in (A.1) are subject to the constraint
∣ α 1 ′ ∣ + … + ∣ α k ′ ∣ = ∣ α ∣ (121) |\alpha_{1}^{\prime}|+\ldots+|\alpha_{k}^{\prime}|=|\alpha| \tag*{(121)} ∣ α 1 ′ ∣ + … + ∣ α k ′ ∣ = ∣ α ∣ (121)
for each k ∈ N k\in\mathbb{N} k ∈ N with k ≤ ∣ α ∣ k\leq|\alpha| k ≤ ∣ α ∣ . Note that a partition { α 1 ′ , … , α k ′ } \{\alpha_{1}^{\prime},\ldots,\alpha_{k}^{\prime}\} { α 1 ′ , … , α k ′ } might contain the same multi-index multiple times.
As a particular case of the general formula, we will use in the following that for positive f ∈ C ∞ ( R n ) f\in C^{\infty}(\mathbb{R}^{n}) f ∈ C ∞ ( R n ) and g = ( ⋅ ) γ g=(\cdot)^{\gamma} g = ( ⋅ ) γ with some γ ∈ R \gamma\in\mathbb{R} γ ∈ R it holds
∣ ∂ α f γ ( ξ ) ∣ ≤ C α ∑ k = 1 ∣ α ∣ ∣ f γ − k ( ξ ) ∣ ∑ { α 1 ′ , … , α k ′ } ∏ l = 1 k ∣ ∂ α l ′ f ( ξ ) ∣ , ξ ∈ R n . (122) \left|\partial^{\alpha}f^{\gamma}(\xi)\right|\leq C_{\alpha}\sum_{k=1}^{|\alpha|}|f^{\gamma-k}(\xi)|\sum_{\{\alpha_{1}^{\prime},\ldots,\alpha_{k}^{\prime}\}}\prod_{l=1}^{k}\left|\partial^{\alpha_{l}^{\prime}}f(\xi)\right|,\qquad\xi\in\mathbb{R}^{n}. \tag*{(122)} ∣ ∂ α f γ ( ξ ) ∣ ≤ C α k = 1 ∑ ∣ α ∣ ∣ f γ − k ( ξ ) ∣ { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∂ α l ′ f ( ξ ) , ξ ∈ R n . (122)
Note that we will additionally use these elementary estimates, valid for every β ∈ N 0 n \beta\in\mathbb{N}_{0}^{n} β ∈ N 0 n ,
∣ ∂ β ξ ∣ ≤ ∣ ξ ∣ 1 − ∣ β ∣ , ∣ ∂ β ∣ ξ ∣ 2 ∣ ≤ 2 ∣ ξ ∣ 2 − ∣ β ∣ , ξ ∈ R n ∖ { 0 } . (123) |\partial^{\beta}\xi|\leq|\xi|^{1-|\beta|},\qquad|\partial^{\beta}|\xi|^{2}|\leq2|\xi|^{2-|\beta|},\qquad\xi\in\mathbb{R}^{n}\setminus\{0\}. \tag*{(123)} ∣ ∂ β ξ ∣ ≤ ∣ ξ ∣ 1 − ∣ β ∣ , ∣ ∂ β ∣ ξ ∣ 2 ∣ ≤ 2∣ ξ ∣ 2 − ∣ β ∣ , ξ ∈ R n ∖ { 0 } . (123)
Now we are in the position to derive the four stated estimates (38) –(41) , which are proven separately in the following steps.
Step 1: estimate (2.21). With the aid of Leibniz’ formula we compute
∣ ∂ α ( ξ Q ^ ρ ( ξ ) ) ∣ ≤ C α ∑ β ≤ α ∣ ∂ β ξ ∣ ∣ ∂ α − β Q ^ ρ ( ξ ) ∣ , ξ ∈ R n . (124) \left|\partial^{\alpha}(\xi\widehat{Q}_{\rho}(\xi))\right|\leq C_{\alpha}\sum_{\beta\leq\alpha}|\partial^{\beta}\xi||\partial^{\alpha-\beta}\widehat{Q}_{\rho}(\xi)|,\qquad\xi\in\mathbb{R}^{n}. \tag*{(124)} ∂ α ( ξ Q ρ ( ξ )) ≤ C α β ≤ α ∑ ∣ ∂ β ξ ∣∣ ∂ α − β Q ρ ( ξ ) ∣ , ξ ∈ R n . (124)
In combination with (A.4) and (37) , i.e. ∣ ∂ α − β Q ^ ρ ( ξ ) ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ Q ^ ρ ( ξ ) |\partial^{\alpha-\beta}\widehat{Q}_{\rho}(\xi)|\leq C_{\alpha}|\xi|^{-|\alpha|+|\beta|}\widehat{Q}_{\rho}(\xi) ∣ ∂ α − β Q ρ ( ξ ) ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ Q ρ ( ξ ) for all ξ ∈ B 1 c \xi\in B_{1}^{c} ξ ∈ B 1 c , it follows that
∣ ∂ α ( ξ Q ^ ρ ( ξ ) ) ∣ ≤ C α ∑ β ≤ α ∣ ξ ∣ 1 − ∣ β ∣ ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ Q ^ ρ ( ξ ) ≤ C α ∣ ξ ∣ 1 − ∣ α ∣ Q ^ ρ ( ξ ) , ξ ∈ B 1 c . \left|\partial^{\alpha}(\xi\widehat{Q}_{\rho}(\xi))\right|\leq C_{\alpha}\sum_{\beta\leq\alpha}|\xi|^{1-|\beta|}|\xi|^{-|\alpha|+|\beta|}\widehat{Q}_{\rho}(\xi)\leq C_{\alpha}|\xi|^{1-|\alpha|}\widehat{Q}_{\rho}(\xi),\qquad\xi\in B_{1}^{c}. ∂ α ( ξ Q ρ ( ξ )) ≤ C α β ≤ α ∑ ∣ ξ ∣ 1 − ∣ β ∣ ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ Q ρ ( ξ ) ≤ C α ∣ ξ ∣ 1 − ∣ α ∣ Q ρ ( ξ ) , ξ ∈ B 1 c .
As a variant of (A.5), we also have
∣ ∂ α ( ∣ ξ ∣ Q ^ ρ ( ξ ) ) ∣ ≤ C α ∑ β ≤ α ∣ ∂ β ∣ ξ ∣ ∣ ∣ ∂ α − β Q ^ ρ ( ξ ) ∣ , ξ ∈ R n . \left|\partial^{\alpha}(|\xi|\widehat{Q}_{\rho}(\xi))\right|\leq C_{\alpha}\sum_{\beta\leq\alpha}|\partial^{\beta}|\xi||\left|\partial^{\alpha-\beta}\widehat{Q}_{\rho}(\xi)\right|,\qquad\xi\in\mathbb{R}^{n}. ∂ α ( ∣ ξ ∣ Q ρ ( ξ )) ≤ C α β ≤ α ∑ ∣ ∂ β ∣ ξ ∣∣ ∂ α − β Q ρ ( ξ ) , ξ ∈ R n .
With Faà di Bruno’s formula in (A.3) for γ = 1 / 2 \gamma= 1/2 γ = 1/2 and (A.4), one readily obtains that
∣ ∂ α ~ ∣ ξ ∣ ∣ = ∣ ∂ α ~ ( ∣ ξ ∣ 2 ) 1 / 2 ∣ ≤ C α ∑ k = 1 ∣ α ~ ∣ ( ∣ ξ ∣ 2 ) ( 1 / 2 − k ) ∑ { α ~ 1 ′ , … , α ~ k ′ } ∏ l = 1 k ∣ ∂ α ~ l ′ ∣ ξ ∣ 2 ∣ ≤ C α ∑ k = 1 ∣ α ~ ∣ ( ∣ ξ ∣ 2 ) ( 1 / 2 − k ) ∑ { α ~ 1 ′ , … , α ~ k ′ } ∏ l = 1 k ∣ ξ ∣ 2 − ∣ α ~ l ′ ∣ ≤ C α ∣ ξ ∣ 1 − ∣ α ~ ∣ , ξ ∈ R n \begin{aligned}
\left|\partial^{\widetilde{\alpha}}|\xi|\right|&=\left|\partial^{\widetilde{\alpha}}\left(|\xi|^2\right)^{1/2}\right|\le C_{\alpha}\sum_{k=1}^{|\widetilde{\alpha}|}\left(|\xi|^2\right)^{(1/2-k)}\sum_{\{\widetilde{\alpha}'_1,\ldots,\widetilde{\alpha}'_k\}}\prod_{l=1}^{k}\left|\partial^{\widetilde{\alpha}'_l}|\xi|^2\right|\\
&\le C_{\alpha}\sum_{k=1}^{|\widetilde{\alpha}|}\left(|\xi|^2\right)^{(1/2-k)}\sum_{\{\widetilde{\alpha}'_1,\ldots,\widetilde{\alpha}'_k\}}\prod_{l=1}^{k}|\xi|^{2-|\widetilde{\alpha}'_l|}\\
&\le C_{\alpha}|\xi|^{1-|\widetilde{\alpha}|},\qquad\xi\in\mathbb{R}^{n}
\end{aligned} ∂ α ∣ ξ ∣ = ∂ α ( ∣ ξ ∣ 2 ) 1/2 ≤ C α k = 1 ∑ ∣ α ∣ ( ∣ ξ ∣ 2 ) ( 1/2 − k ) { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∂ α l ′ ∣ ξ ∣ 2 ≤ C α k = 1 ∑ ∣ α ∣ ( ∣ ξ ∣ 2 ) ( 1/2 − k ) { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∣ ξ ∣ 2 − ∣ α l ′ ∣ ≤ C α ∣ ξ ∣ 1 − ∣ α ∣ , ξ ∈ R n
for any α ~ ∈ N 0 n \widetilde{\alpha}\in\mathbb{N}_{0}^{n} α ∈ N 0 n with partition { α ~ 1 ′ , … , α ~ k ′ } \{\widetilde{\alpha}'_1,\ldots,\widetilde{\alpha}'_k\} { α 1 ′ , … , α k ′ } and α ~ ≤ α \widetilde{\alpha}\le\alpha α ≤ α . Thus, another application of (37) leads to
∣ ∂ α ( ∣ ξ ∣ Q ^ ρ ( ξ ) ) ∣ ≤ C α ∑ β ≤ α ∣ ξ ∣ 1 − ∣ β ∣ ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ Q ^ ρ ( ξ ) ≤ C α ∣ ξ ∣ 1 − ∣ α ∣ Q ^ ρ ( ξ ) , ξ ∈ B 1 c . \left|\partial^{\alpha}\left(|\xi|\widehat{Q}_{\rho}(\xi)\right)\right|\le C_{\alpha}\sum_{\beta\le\alpha}|\xi|^{1-|\beta|}|\xi|^{-|\alpha|+|\beta|}\widehat{Q}_{\rho}(\xi)\le C_{\alpha}|\xi|^{1-|\alpha|}\widehat{Q}_{\rho}(\xi),\qquad\xi\in B_{1}^{c}. ∂ α ( ∣ ξ ∣ Q ρ ( ξ ) ) ≤ C α β ≤ α ∑ ∣ ξ ∣ 1 − ∣ β ∣ ∣ ξ ∣ − ∣ α ∣ + ∣ β ∣ Q ρ ( ξ ) ≤ C α ∣ ξ ∣ 1 − ∣ α ∣ Q ρ ( ξ ) , ξ ∈ B 1 c .
Step 2: estimate (39) . Now we invoke Faà di Bruno’s formula in (A.3) for general γ ∈ R \gamma\in\mathbb{R} γ ∈ R . Thus,
∣ ∂ α ( ∣ ξ ∣ Q ^ ρ ( ξ ) ) γ ∣ ≤ C α ∑ k = 1 ∣ α ∣ ( ∣ ξ ∣ Q ^ ρ ( ξ ) ) γ − k ∑ { α 1 ′ , … , α k ′ } ∏ l = 1 k ∣ ∂ α l ′ ( ∣ ξ ∣ Q ^ ρ ( ξ ) ) ∣ ≤ C α ∑ k = 1 ∣ α ∣ ( ∣ ξ ∣ Q ^ ρ ( ξ ) ) γ − k ∑ { α 1 ′ , … , α k ′ } ∏ l = 1 k ∣ ξ ∣ 1 − ∣ α l ′ ∣ Q ^ ρ ( ξ ) ≤ C α ∣ ξ ∣ γ − ∣ α ∣ Q ^ ρ γ ( ξ ) , ξ ∈ B 1 c , \begin{aligned}
\left|\partial^{\alpha}\left(|\xi|\widehat{Q}_{\rho}(\xi)\right)^{\gamma}\right|&\le C_{\alpha}\sum_{k=1}^{|\alpha|}\left(|\xi|\widehat{Q}_{\rho}(\xi)\right)^{\gamma-k}\sum_{\{\alpha'_1,\ldots,\alpha'_k\}}\prod_{l=1}^{k}\left|\partial^{\alpha'_l}\left(|\xi|\widehat{Q}_{\rho}(\xi)\right)\right|\\
&\le C_{\alpha}\sum_{k=1}^{|\alpha|}\left(|\xi|\widehat{Q}_{\rho}(\xi)\right)^{\gamma-k}\sum_{\{\alpha'_1,\ldots,\alpha'_k\}}\prod_{l=1}^{k}|\xi|^{1-|\alpha'_l|}\widehat{Q}_{\rho}(\xi)\\
&\le C_{\alpha}|\xi|^{\gamma-|\alpha|}\widehat{Q}_{\rho}^{\gamma}(\xi),\qquad\xi\in B_{1}^{c},
\end{aligned} ∂ α ( ∣ ξ ∣ Q ρ ( ξ ) ) γ ≤ C α k = 1 ∑ ∣ α ∣ ( ∣ ξ ∣ Q ρ ( ξ ) ) γ − k { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∂ α l ′ ( ∣ ξ ∣ Q ρ ( ξ ) ) ≤ C α k = 1 ∑ ∣ α ∣ ( ∣ ξ ∣ Q ρ ( ξ ) ) γ − k { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∣ ξ ∣ 1 − ∣ α l ′ ∣ Q ρ ( ξ ) ≤ C α ∣ ξ ∣ γ − ∣ α ∣ Q ρ γ ( ξ ) , ξ ∈ B 1 c ,
validating the first estimate of (39) . The second estimate of (39) , concerning bounds for ∂ α ( 1 + ∣ ξ ∣ Q ^ ρ ( ξ ) ) γ \partial^{\alpha}(1+|\xi|\widehat{Q}_{\rho}(\xi))^{\gamma} ∂ α ( 1 + ∣ ξ ∣ Q ρ ( ξ ) ) γ , is obtained similarly. It suffices to observe that ( 1 + ∣ ξ ∣ Q ^ ρ ( ξ ) ) γ − k ≤ C α ( ∣ ξ ∣ Q ^ ρ ( ξ ) ) γ − k (1+|\xi|\widehat{Q}_{\rho}(\xi))^{\gamma-k}\le C_{\alpha}(|\xi|\widehat{Q}_{\rho}(\xi))^{\gamma-k} ( 1 + ∣ ξ ∣ Q ρ ( ξ ) ) γ − k ≤ C α ( ∣ ξ ∣ Q ρ ( ξ ) ) γ − k for all γ ∈ R \gamma\in\mathbb{R} γ ∈ R and all k ∈ N k\in\mathbb{N} k ∈ N with k ≤ ∣ α ∣ k\le|\alpha| k ≤ ∣ α ∣ , where in the case that γ − k > 0 \gamma-k>0 γ − k > 0 we use that 1 + ∣ λ ρ ∣ 2 ≤ C α ∣ λ ρ ∣ 2 1+|\lambda_{\rho}|^2\le C_{\alpha}|\lambda_{\rho}|^2 1 + ∣ λ ρ ∣ 2 ≤ C α ∣ λ ρ ∣ 2 in B 1 c B_{1}^{c} B 1 c according (35) .
Step 3: estimate (40) . Once more appealing to Leibniz’ formula, entails
∣ ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ∣ ≤ C α ∑ β ≤ α ∣ ∂ β ξ ∣ 2 ∣ ∂ α − β Q ^ ρ 2 ( ξ ) ∣ , ξ ∈ R n . (125) \left|\partial^{\alpha}(1+|\lambda_{\rho}(\xi)|^2)\right|\le C_{\alpha}\sum_{\beta\le\alpha}|\partial^{\beta}\xi|^2|\partial^{\alpha-\beta}\widehat{Q}_{\rho}^{\,2}(\xi)|,\qquad\xi\in\mathbb{R}^{n}. \tag*{(125)} ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ≤ C α β ≤ α ∑ ∣ ∂ β ξ ∣ 2 ∣ ∂ α − β Q ρ 2 ( ξ ) ∣ , ξ ∈ R n . (125)
To further estimate the right-hand side, we apply Faà di Bruno’s formula in (A.3) to the function Q ^ ρ 2 \widehat{Q}_{\rho}^{\,2} Q ρ 2 . Together with the estimate (37) for the derivatives of Q ^ ρ \widehat{Q}_{\rho} Q ρ , we infer that
∣ ∂ α ~ Q ^ ρ 2 ( ξ ) ∣ ≤ C α ~ ∑ k = 1 ∣ α ~ ∣ Q ^ ρ 2 − k ( ξ ) ∑ { α ~ 1 ′ , … , α ~ k ′ } ∏ l = 1 k ∣ ∂ α ~ l ′ Q ^ ρ ( ξ ) ∣ ≤ C α ~ ∑ k = 1 ∣ α ~ ∣ Q ^ ρ 2 − k ( ξ ) ∑ { α ~ 1 ′ , … , α ~ k ′ } ∏ l = 1 k ∣ ξ ∣ − ∣ α ~ l ′ ∣ Q ^ ρ ( ξ ) = C α ~ Q ^ ρ 2 ( ξ ) ∣ ξ ∣ − ∣ α ~ ∣ , ξ ∈ B 1 c , \begin{aligned}
\left|\partial^{\widetilde{\alpha}}\widehat{Q}_{\rho}^{\,2}(\xi)\right|&\le C_{\widetilde{\alpha}}\sum_{k=1}^{|\widetilde{\alpha}|}\widehat{Q}_{\rho}^{\,2-k}(\xi)\sum_{\{\widetilde{\alpha}'_1,\ldots,\widetilde{\alpha}'_k\}}\prod_{l=1}^{k}\left|\partial^{\widetilde{\alpha}'_l}\widehat{Q}_{\rho}(\xi)\right|\\
&\le C_{\widetilde{\alpha}}\sum_{k=1}^{|\widetilde{\alpha}|}\widehat{Q}_{\rho}^{\,2-k}(\xi)\sum_{\{\widetilde{\alpha}'_1,\ldots,\widetilde{\alpha}'_k\}}\prod_{l=1}^{k}|\xi|^{-|\widetilde{\alpha}'_l|}\widehat{Q}_{\rho}(\xi)\\
&=C_{\widetilde{\alpha}}\widehat{Q}_{\rho}^{\,2}(\xi)|\xi|^{-|\widetilde{\alpha}|},\qquad\xi\in B_{1}^{c},
\end{aligned} ∂ α Q ρ 2 ( ξ ) ≤ C α k = 1 ∑ ∣ α ∣ Q ρ 2 − k ( ξ ) { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∂ α l ′ Q ρ ( ξ ) ≤ C α k = 1 ∑ ∣ α ∣ Q ρ 2 − k ( ξ ) { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∣ ξ ∣ − ∣ α l ′ ∣ Q ρ ( ξ ) = C α Q ρ 2 ( ξ ) ∣ ξ ∣ − ∣ α ∣ , ξ ∈ B 1 c ,
where the last equality invoked again the sum constraint from (A.2). Coming back to (A.6) and using also the second part in (A.4), we conclude with ∣ α − β ∣ = ∣ α ∣ − ∣ β ∣ |\alpha-\beta|=|\alpha|-|\beta| ∣ α − β ∣ = ∣ α ∣ − ∣ β ∣ for β ≤ α \beta\le\alpha β ≤ α that
∣ ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ∣ ≤ C α ∑ β ≤ α ∣ ξ ∣ 2 − ∣ β ∣ Q ^ ρ 2 ( ξ ) ∣ ξ ∣ − ( ∣ α ∣ − ∣ β ∣ ) = C α ∣ ξ ∣ 2 − ∣ α ∣ Q ^ ρ 2 ( ξ ) , ξ ∈ B 1 c . \left|\partial^{\alpha}(1+|\lambda_{\rho}(\xi)|^2)\right|\le C_{\alpha}\sum_{\beta\le\alpha}|\xi|^{2-|\beta|}\widehat{Q}_{\rho}^{\,2}(\xi)|\xi|^{-(|\alpha|-|\beta|)}=C_{\alpha}|\xi|^{2-|\alpha|}\widehat{Q}_{\rho}^{\,2}(\xi),\qquad\xi\in B_{1}^{c}. ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ≤ C α β ≤ α ∑ ∣ ξ ∣ 2 − ∣ β ∣ Q ρ 2 ( ξ ) ∣ ξ ∣ − ( ∣ α ∣ − ∣ β ∣ ) = C α ∣ ξ ∣ 2 − ∣ α ∣ Q ρ 2 ( ξ ) , ξ ∈ B 1 c .
This finishes the proof of (40) .
Step 4: estimate (41) . We apply once again Faà di Bruno’s formula in the version of (A.3). In that regard, we note that for γ ∈ R \gamma\in\mathbb{R} γ ∈ R and k ≥ 1 k\ge1 k ≥ 1 ,
( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) γ − k ≤ ∣ ξ ∣ 2 γ − 2 k Q ^ ρ 2 γ − 2 k ( ξ ) , ξ ∈ B 1 c . (126) (1+|\lambda_{\rho}(\xi)|^2)^{\gamma-k}\le|\xi|^{2\gamma-2k}\widehat{Q}_{\rho}^{\,2\gamma-2k}(\xi),\qquad\xi\in B_{1}^{c}. \tag*{(126)} ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) γ − k ≤ ∣ ξ ∣ 2 γ − 2 k Q ρ 2 γ − 2 k ( ξ ) , ξ ∈ B 1 c . (126)
where in the case that γ − k > 0 \gamma-k>0 γ − k > 0 , we used again that, by (35) , it holds 1 + ∣ λ ρ ( ξ ) ∣ 2 ≤ C ∣ λ ρ ( ξ ) ∣ 2 1+|\lambda_\rho(\xi)|^2\le C|\lambda_\rho(\xi)|^2 1 + ∣ λ ρ ( ξ ) ∣ 2 ≤ C ∣ λ ρ ( ξ ) ∣ 2 . Combining this with the just proved third estimate (40) , we obtain from (A.3) and (A.2) that
∣ ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) γ ∣ ≤ C α ∑ k = 1 ∣ α ∣ ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) γ − k ( ξ ) ∑ { α 1 ′ , … , α k ′ } ∏ l = 1 k ∣ ∂ α l ′ ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ∣ ≤ C α ∑ k = 1 ∣ α ∣ ∣ ξ ∣ 2 γ − 2 k Q ^ ρ 2 γ − 2 k ( ξ ) ∑ { α 1 ′ , … , α k ′ } ∏ l = 1 k ∣ ξ ∣ 2 − ∣ α l ′ ∣ Q ^ ρ 2 ( ξ ) ≤ C α ∣ ξ ∣ 2 γ − ∣ α ∣ Q ^ ρ 2 γ ( ξ ) , ξ ∈ B 1 c , \begin{aligned}
\left|\partial^\alpha\left(1+|\lambda_\rho(\xi)|^2\right)^\gamma\right|
\le C_\alpha\sum_{k=1}^{|\alpha|}\left(1+|\lambda_\rho(\xi)|^2\right)^{\gamma-k}(\xi)\sum_{\{\alpha'_1,\ldots,\alpha'_k\}}\prod_{l=1}^{k}\left|\partial^{\alpha'_l}\left(1+|\lambda_\rho(\xi)|^2\right)\right| \\
&\le C_\alpha\sum_{k=1}^{|\alpha|}|\xi|^{2\gamma-2k}\widehat{Q}_\rho^{2\gamma-2k}(\xi)\sum_{\{\alpha'_1,\ldots,\alpha'_k\}}\prod_{l=1}^{k}|\xi|^{2-|\alpha'_l|}\widehat{Q}_\rho^2(\xi) \\
&\le C_\alpha|\xi|^{2\gamma-|\alpha|}\widehat{Q}_\rho^{2\gamma}(\xi),\qquad\xi\in B_1^c,
\end{aligned} ∂ α ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) γ ≤ C α k = 1 ∑ ∣ α ∣ ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) γ − k ( ξ ) { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∂ α l ′ ( 1 + ∣ λ ρ ( ξ ) ∣ 2 ) ≤ C α k = 1 ∑ ∣ α ∣ ∣ ξ ∣ 2 γ − 2 k Q ρ 2 γ − 2 k ( ξ ) { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∣ ξ ∣ 2 − ∣ α l ′ ∣ Q ρ 2 ( ξ ) ≤ C α ∣ ξ ∣ 2 γ − ∣ α ∣ Q ρ 2 γ ( ξ ) , ξ ∈ B 1 c ,
concluding the proof. □
Next, we collect here some estimates of similar kind for the function (see (111) )
Θ ρ δ = e − ∣ λ ρ ∣ 2 δ / ( 4 π 2 ) \Theta_\rho^\delta=e^{-|\lambda_\rho|^2\delta/(4\pi^2)} Θ ρ δ = e − ∣ λ ρ ∣ 2 δ / ( 4 π 2 )
for any fixed δ > 0 \delta>0 δ > 0 , which were exploited in the proof of Proposition 5.1.
Lemma A.1. Let p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and α ∈ N 0 n \alpha\in\mathbb{N}_0^n α ∈ N 0 n . Assume that ρ \rho ρ has compact support and satisfies assumptions ( H 0 ) (H0) ( H 0 ) –( H 3 ) (H3) ( H 3 ) , and let γ = ( 4 − 3 σ ) − 1 ∈ ( 1 / 4 , 1 ) \gamma=(4-3\sigma)^{-1}\in(1/4,1) γ = ( 4 − 3 σ ) − 1 ∈ ( 1/4 , 1 ) , where σ \sigma σ is the constant from ( H 3 ) (H3) ( H 3 ) . Then,
∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ , ξ ∈ B 1 c , δ ∈ ( 0 , ∞ ) , (127) |\partial^\alpha\Theta_\rho^\delta(\xi)|\le C_\alpha|\xi|^{-|\alpha|},\qquad\xi\in B_1^c,\ \delta\in(0,\infty), \tag*{(127)} ∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ , ξ ∈ B 1 c , δ ∈ ( 0 , ∞ ) , (127) ∣ ∂ α Θ ρ δ ( δ − 1 / ( 2 σ ) ξ ) ∣ ≤ C α ∣ ξ ∣ − ( ∣ α ∣ + γ σ ) , ξ ∈ B δ 1 / ( 2 σ ) c , δ ∈ ( 0 , ∞ ) , (128) \left|\partial^\alpha\Theta_\rho^\delta\left(\delta^{-1/(2\sigma)}\xi\right)\right|\le C_\alpha|\xi|^{-(|\alpha|+\gamma\sigma)},\qquad\xi\in B_{\delta^{1/(2\sigma)}}^c,\ \delta\in(0,\infty), \tag*{(128)} ∂ α Θ ρ δ ( δ − 1/ ( 2 σ ) ξ ) ≤ C α ∣ ξ ∣ − ( ∣ α ∣ + γ σ ) , ξ ∈ B δ 1/ ( 2 σ ) c , δ ∈ ( 0 , ∞ ) , (128) ∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α , ξ ∈ B 1 , δ ∈ ( 0 , 1 ) , (129) |\partial^\alpha\Theta_\rho^\delta(\xi)|\le C_\alpha,\qquad\xi\in B_1,\ \delta\in(0,1), \tag*{(129)} ∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α , ξ ∈ B 1 , δ ∈ ( 0 , 1 ) , (129) ∣ ∂ α Θ ρ δ ( δ − 1 / 2 ξ ) ∣ ≤ C α ( 1 + ∣ ξ ∣ ) 2 ∣ α ∣ e − c ∣ ξ ∣ 2 , ξ ∈ B δ 1 / 2 , δ ∈ ( 1 , ∞ ) (130) \left|\partial^\alpha\Theta_\rho^\delta\left(\delta^{-1/2}\xi\right)\right|\le C_\alpha(1+|\xi|)^{2|\alpha|}e^{-c|\xi|^2},\qquad\xi\in B_{\delta^{1/2}},\ \delta\in(1,\infty) \tag*{(130)} ∂ α Θ ρ δ ( δ − 1/2 ξ ) ≤ C α ( 1 + ∣ ξ ∣ ) 2∣ α ∣ e − c ∣ ξ ∣ 2 , ξ ∈ B δ 1/2 , δ ∈ ( 1 , ∞ ) (130) and
∣ ∂ α Θ ρ δ ( δ − 1 / 2 ξ ) ∣ ≤ C α e − c ∣ ξ ∣ 2 σ , ξ ∈ B δ 1 / 2 c , δ ∈ ( 1 , ∞ ) (131) \left|\partial^\alpha\Theta_\rho^\delta\left(\delta^{-1/2}\xi\right)\right|\le C_\alpha e^{-c|\xi|^{2\sigma}},\qquad\xi\in B_{\delta^{1/2}}^c,\ \delta\in(1,\infty) \tag*{(131)} ∂ α Θ ρ δ ( δ − 1/2 ξ ) ≤ C α e − c ∣ ξ ∣ 2 σ , ξ ∈ B δ 1/2 c , δ ∈ ( 1 , ∞ ) (131) with constants C α , c > 0 C_\alpha,c>0 C α , c > 0 .
Proof of Lemma A.1. We start by deriving some auxiliary estimates for α ∈ N 0 n ∖ { 0 } \alpha\in\mathbb{N}_0^n\setminus\{0\} α ∈ N 0 n ∖ { 0 } . By Faà di Bruno’s formula in the version of (A.1), one has
∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α ∑ k = 1 ∣ α ∣ δ k Θ ρ δ ( ξ ) ∑ { α 1 ′ , … , α k ′ } ∏ l = 1 k ∣ δ α l ′ ( ∣ λ ρ ( ξ ) ∣ 2 ) ∣ , ξ ∈ R n , (132) |\partial^\alpha\Theta_\rho^\delta(\xi)|\le C_\alpha\sum_{k=1}^{|\alpha|}\delta^k\Theta_\rho^\delta(\xi)\sum_{\{\alpha'_1,\ldots,\alpha'_k\}}\prod_{l=1}^{k}|\delta^{\alpha'_l}(|\lambda_\rho(\xi)|^2)|,\qquad\xi\in\mathbb{R}^n, \tag*{(132)} ∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α k = 1 ∑ ∣ α ∣ δ k Θ ρ δ ( ξ ) { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∣ δ α l ′ ( ∣ λ ρ ( ξ ) ∣ 2 ) ∣ , ξ ∈ R n , (132)
with the sum taken over the different partitions { α 1 ′ , … , α k ′ } \{\alpha'_1,\ldots,\alpha'_k\} { α 1 ′ , … , α k ′ } of α \alpha α for k ∈ N k\in\mathbb{N} k ∈ N such that k ≤ ∣ α ∣ k\le|\alpha| k ≤ ∣ α ∣ . From (40) of Lemma 2.8, we infer that
∣ ∂ β ∣ λ ρ ( ξ ) ∣ 2 ∣ ≤ C β ∣ ξ ∣ 2 − ∣ β ∣ Q ^ ρ 2 ( ξ ) , ξ ∈ B 1 c |\partial^\beta|\lambda_\rho(\xi)|^2|\le C_\beta|\xi|^{2-|\beta|}\widehat{Q}_\rho^2(\xi),\qquad\xi\in B_1^c ∣ ∂ β ∣ λ ρ ( ξ ) ∣ 2 ∣ ≤ C β ∣ ξ ∣ 2 − ∣ β ∣ Q ρ 2 ( ξ ) , ξ ∈ B 1 c
for all β ∈ N 0 n \beta\in\mathbb{N}_0^n β ∈ N 0 n , and it follows with (A.13) that
∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α ∑ k = 1 ∣ α ∣ δ k Θ ρ δ ( ξ ) ∑ { α 1 ′ , … , α k ′ } ∏ l = 1 k ∣ ξ ∣ 2 − ∣ α l ′ ∣ Q ^ ρ 2 ( ξ ) ≤ C α ∣ ξ ∣ − ∣ α ∣ ∑ k = 1 ∣ α ∣ δ k Θ ρ δ ( ξ ) ∣ ξ ∣ 2 k Q ^ ρ 2 k ( ξ ) , ξ ∈ B 1 c , |\partial^\alpha\Theta_\rho^\delta(\xi)|\le C_\alpha\sum_{k=1}^{|\alpha|}\delta^k\Theta_\rho^\delta(\xi)\sum_{\{\alpha'_1,\ldots,\alpha'_k\}}\prod_{l=1}^{k}|\xi|^{2-|\alpha'_l|}\widehat{Q}_\rho^2(\xi)
\le C_\alpha|\xi|^{-|\alpha|}\sum_{k=1}^{|\alpha|}\delta^k\Theta_\rho^\delta(\xi)|\xi|^{2k}\widehat{Q}_\rho^{2k}(\xi),\qquad\xi\in B_1^c, ∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α k = 1 ∑ ∣ α ∣ δ k Θ ρ δ ( ξ ) { α 1 ′ , … , α k ′ } ∑ l = 1 ∏ k ∣ ξ ∣ 2 − ∣ α l ′ ∣ Q ρ 2 ( ξ ) ≤ C α ∣ ξ ∣ − ∣ α ∣ k = 1 ∑ ∣ α ∣ δ k Θ ρ δ ( ξ ) ∣ ξ ∣ 2 k Q ρ 2 k ( ξ ) , ξ ∈ B 1 c ,
where we used (A.3) in the last inequality. This proves
∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ ∑ k = 1 ∣ α ∣ δ k Θ ρ δ ( ξ ) ∣ λ ρ ( ξ ) ∣ 2 k , ξ ∈ B 1 c . (133) |\partial^\alpha\Theta_\rho^\delta(\xi)|\le C_\alpha|\xi|^{-|\alpha|}\sum_{k=1}^{|\alpha|}\delta^k\Theta_\rho^\delta(\xi)|\lambda_\rho(\xi)|^{2k},\qquad\xi\in B_1^c. \tag*{(133)} ∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α ∣ ξ ∣ − ∣ α ∣ k = 1 ∑ ∣ α ∣ δ k Θ ρ δ ( ξ ) ∣ λ ρ ( ξ ) ∣ 2 k , ξ ∈ B 1 c . (133)
which serves as the basis for verifying the estimates which are stated on the complement of a ball around the origin.
Furthermore, since ξ ↦ e − ξ \xi\mapsto e^{-\xi} ξ ↦ e − ξ decays faster than any polynomial we have that, for any l ∈ N 0 l \in\mathbb{N}_0 l ∈ N 0 ,
Θ ρ δ ( δ − 1 / ( 2 σ ) ξ ) ≤ C l 1 ( ∣ λ ρ ( δ − 1 / ( 2 σ ) ξ ) ∣ 2 δ ) γ / 2 + l , (134) \Theta_{\rho}^{\delta}\left(\delta^{-1/(2\sigma)}\xi\right) \le C_l \frac{1}{\left(\left|\lambda_{\rho}\left(\delta^{-1/(2\sigma)}\xi\right)\right|2\delta\right)^{\gamma/2+l}}, \tag*{(134)} Θ ρ δ ( δ − 1/ ( 2 σ ) ξ ) ≤ C l ( λ ρ ( δ − 1/ ( 2 σ ) ξ ) 2 δ ) γ /2 + l 1 , (134)
Θ ρ δ ( δ − 1 / 2 ξ ) ≤ C l e − ∣ ξ ∣ 2 Q ^ ρ 2 ( δ − 1 / 2 ξ ) / 2 ( ∣ λ ρ ( δ − 1 / 2 ξ ) ∣ 2 δ ) l , ξ ∈ R n . (135) \Theta_{\rho}^{\delta}\left(\delta^{-1/2}\xi\right) \le C_l \frac{e^{-|\xi|^2\widehat{Q}_{\rho}^{2}\left(\delta^{-1/2}\xi\right)/2}}{\left(\left|\lambda_{\rho}\left(\delta^{-1/2}\xi\right)\right|2\delta\right)^l}, \qquad\xi\in\mathbb{R}^n. \tag*{(135)} Θ ρ δ ( δ − 1/2 ξ ) ≤ C l ( λ ρ ( δ − 1/2 ξ ) 2 δ ) l e − ∣ ξ ∣ 2 Q ρ 2 ( δ − 1/2 ξ ) /2 , ξ ∈ R n . (135)
We now turn to the verification of the stated inequalities (A.8)–(A.12). Estimate (A.8) is a direct consequence of inserting the first inequality in (A.15) into (A.14).
To show (A.9) for α = 0 \alpha= 0 α = 0 it suffices to combine (A.15) for l = 0 l = 0 l = 0 with condition (H3) and the resulting inequality (2.19), which yields exactly
Θ ρ δ ( δ − 1 / ( 2 σ ) ξ ) ≤ C 0 ∣ ξ ∣ − γ σ , ξ ∈ B δ 1 / ( 2 σ ) c . \Theta_{\rho}^{\delta}\left(\delta^{-1/(2\sigma)}\xi\right) \le C_0|\xi|^{-\gamma\sigma}, \qquad\xi\in B_{\delta^{1/(2\sigma)}}^{c}. Θ ρ δ ( δ − 1/ ( 2 σ ) ξ ) ≤ C 0 ∣ ξ ∣ − γ σ , ξ ∈ B δ 1/ ( 2 σ ) c .
For α ∈ N 0 n ∖ { 0 } \alpha\in\mathbb{N}_0^n \setminus\{0\} α ∈ N 0 n ∖ { 0 } , we argue that applying the chain rule to (A.14) gives
∣ ∂ α Θ ρ δ ( δ − 1 / ( 2 σ ) ξ ) ∣ = ∣ ( ∂ α Θ ρ δ ) ( δ − 1 / ( 2 σ ) ξ ) ∣ δ − ∣ α ∣ / ( 2 σ ) ≤ C α ∣ ξ ∣ − ∣ α ∣ ∑ k = 1 ∣ α ∣ δ k Θ ρ δ ( δ − 1 / ( 2 σ ) ξ ) ∣ λ ρ ( δ − 1 / ( 2 σ ) ξ ) ∣ 2 k , ξ ∈ B δ 1 / ( 2 σ ) c . (136) \begin{aligned}
\left|\partial^{\alpha}\Theta_{\rho}^{\delta}\left(\delta^{-1/(2\sigma)}\xi\right)\right| = \left|\left(\partial^{\alpha}\Theta_{\rho}^{\delta}\right)\left(\delta^{-1/(2\sigma)}\xi\right)\right|\delta^{-|\alpha|/(2\sigma)} \\
&\le C_{\alpha}|\xi|^{-|\alpha|}\sum_{k=1}^{|\alpha|}\delta^k\Theta_{\rho}^{\delta}\left(\delta^{-1/(2\sigma)}\xi\right)\left|\lambda_{\rho}\left(\delta^{-1/(2\sigma)}\xi\right)\right|^{2k}, \qquad\xi\in B_{\delta^{1/(2\sigma)}}^{c}. \tag*{(136)}
\end{aligned} ∂ α Θ ρ δ ( δ − 1/ ( 2 σ ) ξ ) = ( ∂ α Θ ρ δ ) ( δ − 1/ ( 2 σ ) ξ ) δ − ∣ α ∣/ ( 2 σ ) ≤ C α ∣ ξ ∣ − ∣ α ∣ k = 1 ∑ ∣ α ∣ δ k Θ ρ δ ( δ − 1/ ( 2 σ ) ξ ) λ ρ ( δ − 1/ ( 2 σ ) ξ ) 2 k , ξ ∈ B δ 1/ ( 2 σ ) c . (136)
Furthermore, using (A.15) and then (2.19), there also holds, for any k ∈ N k \in\mathbb{N} k ∈ N ,
Θ ρ δ ( δ − 1 / ( 2 σ ) ξ ) ≤ C α 1 ∣ λ ρ ( δ − 1 / ( 2 σ ) ξ ) ∣ 2 k δ k ∣ ξ ∣ − γ σ , ξ ∈ B δ 1 / ( 2 σ ) c . \Theta_{\rho}^{\delta}\left(\delta^{-1/(2\sigma)}\xi\right) \le C_{\alpha}\frac{1}{\left|\lambda_{\rho}\left(\delta^{-1/(2\sigma)}\xi\right)\right|^{2k}\delta^k}|\xi|^{-\gamma\sigma}, \qquad\xi\in B_{\delta^{1/(2\sigma)}}^{c}. Θ ρ δ ( δ − 1/ ( 2 σ ) ξ ) ≤ C α λ ρ ( δ − 1/ ( 2 σ ) ξ ) 2 k δ k 1 ∣ ξ ∣ − γ σ , ξ ∈ B δ 1/ ( 2 σ ) c .
Combining this with (A.17), we get the refined estimate (A.9).
Next we prove (A.10) with δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) . We use that
∣ ∂ β ~ ∣ ξ ∣ 2 ∣ ≤ 2 ∣ ξ ∣ min { 2 − ∣ β ~ ∣ , 0 } ≤ 2 , ξ ∈ B 1 (137) \left|\partial^{\widetilde{\beta}}|\xi|^2\right| \le2|\xi|^{\min\{2-|\widetilde{\beta}|,0\}} \le2, \qquad\xi\in B_1 \tag*{(137)} ∂ β ∣ ξ ∣ 2 ≤ 2∣ ξ ∣ m i n { 2 − ∣ β ∣ , 0 } ≤ 2 , ξ ∈ B 1 (137)
for any β ~ ∈ N 0 n \widetilde{\beta} \in\mathbb{N}_0^n β ∈ N 0 n . Then, (A.10) with α = 0 \alpha= 0 α = 0 follows immediately from Θ ρ δ ( ξ ) ≤ 1 \Theta_{\rho}^{\delta}(\xi) \le1 Θ ρ δ ( ξ ) ≤ 1 for all ξ ∈ R n \xi\in\mathbb{R}^n ξ ∈ R n . For α ∈ N 0 n ∖ { 0 } \alpha\in\mathbb{N}_0^n \setminus\{0\} α ∈ N 0 n ∖ { 0 } , we first conclude by inserting (A.18) into (A.6), and with the smoothness of Q ^ ρ \widehat{Q}_{\rho} Q ρ , that
∣ ∂ β ∣ λ ρ ( ξ ) ∣ 2 ∣ ≤ C β ∑ β ~ ≤ β ∣ ∂ β ~ ∣ ξ ∣ 2 ∣ ∣ ∂ β − β ~ Q ^ ρ 2 ( ξ ) ∣ ≤ C β , ξ ∈ B 1 (138) \left|\partial^{\beta}|\lambda_{\rho}(\xi)|^2\right| \le C_{\beta}\sum_{\widetilde{\beta}\le\beta}\left|\partial^{\widetilde{\beta}}|\xi|^2\right|\left|\partial^{\beta-\widetilde{\beta}}\widehat{Q}_{\rho}^{2}(\xi)\right| \le C_{\beta}, \qquad\xi\in B_1 \tag*{(138)} ∂ β ∣ λ ρ ( ξ ) ∣ 2 ≤ C β β ≤ β ∑ ∂ β ∣ ξ ∣ 2 ∂ β − β Q ρ 2 ( ξ ) ≤ C β , ξ ∈ B 1 (138)
for any β ∈ N 0 n \beta\in\mathbb{N}_0^n β ∈ N 0 n . Together with (A.13), one has
∣ ∂ α Θ ρ δ ( ξ ) ∣ ≤ C α ∑ k = 1 ∣ α ∣ δ k ≤ C α , ξ ∈ B 1 , δ ∈ ( 0 , 1 ) , \left|\partial^{\alpha}\Theta_{\rho}^{\delta}(\xi)\right| \le C_{\alpha}\sum_{k=1}^{|\alpha|}\delta^k \le C_{\alpha}, \qquad\xi\in B_1,\ \delta\in(0,1), ∂ α Θ ρ δ ( ξ ) ≤ C α k = 1 ∑ ∣ α ∣ δ k ≤ C α , ξ ∈ B 1 , δ ∈ ( 0 , 1 ) ,
validating (A.10).
In view of (A.11), we note that it holds
∣ ∂ β ~ ∣ ξ ∣ 2 ∣ ≤ 2 ( 1 + ∣ ξ ∣ ) 2 , ξ ∈ R n \left|\partial^{\widetilde{\beta}}|\xi|^2\right| \le2(1+|\xi|)^2, \qquad\xi\in\mathbb{R}^n ∂ β ∣ ξ ∣ 2 ≤ 2 ( 1 + ∣ ξ ∣ ) 2 , ξ ∈ R n
for every β ~ ∈ N 0 n \widetilde{\beta} \in\mathbb{N}_0^n β ∈ N 0 n . Hence, one deduces with the Leibniz rule from the smoothness of Q ^ ρ \widehat{Q}_{\rho} Q ρ that
∣ ∂ β λ ρ ( δ − 1 / 2 ξ ) ∣ 2 ≤ C β δ − 1 ( 1 + ∣ ξ ∣ ) 2 , ξ ∈ B δ 1 / 2 , δ ∈ ( 1 , ∞ ) \left|\partial^{\beta}\lambda_{\rho}\left(\delta^{-1/2}\xi\right)\right|^2 \le C_{\beta}\delta^{-1}(1+|\xi|)^2, \qquad\xi\in B_{\delta^{1/2}},\ \delta\in(1,\infty) ∂ β λ ρ ( δ − 1/2 ξ ) 2 ≤ C β δ − 1 ( 1 + ∣ ξ ∣ ) 2 , ξ ∈ B δ 1/2 , δ ∈ ( 1 , ∞ )
for all β ∈ N 0 n \beta\in\mathbb{N}_0^n β ∈ N 0 n . It follows similar to (A.13) that
∣ ∂ α Θ ρ δ ( δ − 1 / 2 ξ ) ∣ ≤ C α Θ ρ δ ( δ − 1 / 2 ξ ) ( 1 + ∣ ξ ∣ ) 2 ∣ α ∣ , δ ∈ ( 1 , ∞ ) , ξ ∈ B δ 1 / 2 . \left|\partial^{\alpha}\Theta_{\rho}^{\delta}\left(\delta^{-1/2}\xi\right)\right| \le C_{\alpha}\Theta_{\rho}^{\delta}\left(\delta^{-1/2}\xi\right)(1+|\xi|)^{2|\alpha|}, \qquad\delta\in(1,\infty),\ \xi\in B_{\delta^{1/2}}. ∂ α Θ ρ δ ( δ − 1/2 ξ ) ≤ C α Θ ρ δ ( δ − 1/2 ξ ) ( 1 + ∣ ξ ∣ ) 2∣ α ∣ , δ ∈ ( 1 , ∞ ) , ξ ∈ B δ 1/2 .
Inserting the estimate Q ^ ρ ≥ c > 0 \widehat{Q}_{\rho} \ge c > 0 Q ρ ≥ c > 0 on B 1 B_1 B 1 and the resulting inequality Θ ρ δ ( δ − 1 / 2 ξ ) ≤ e − c ∣ ξ ∣ 2 \Theta_{\rho}^{\delta}\left(\delta^{-1/2}\xi\right) \le e^{-c|\xi|^2} Θ ρ δ ( δ − 1/2 ξ ) ≤ e − c ∣ ξ ∣ 2 for ξ ∈ B δ 1 / 2 \xi\in B_{\delta^{1/2}} ξ ∈ B δ 1/2 , finishes the proof of (A.11).
Finally, turning to (A.12), (A.14) implies that
∣ ∂ α Θ ρ δ ( δ − 1 / 2 ξ ) ∣ = ∣ ( ∂ α Θ ρ δ ) ( δ − 1 / 2 ξ ) ∣ δ − ∣ α ∣ / 2 ≤ C α ∣ ξ ∣ − ∣ α ∣ ∑ k = 1 ∣ α ∣ δ k Θ ρ δ ( δ − 1 / 2 ξ ) λ ρ ( δ − 1 / 2 ξ ) 2 k ≤ C α e − ∣ ξ ∣ 2 Q ^ ρ 2 ( δ − 1 / 2 ξ ) / 2 , ξ ∈ B δ 1 / 2 c , \begin{aligned}
\left|\partial^\alpha\Theta_\rho^\delta(\delta^{-1/2}\xi)\right|=\left|\left(\partial^\alpha\Theta_\rho^\delta\right)(\delta^{-1/2}\xi)\right|\delta^{-|\alpha|/2} \\
&\le C_\alpha|\xi|^{-|\alpha|}\sum_{k=1}^{|\alpha|}\delta^k\Theta_\rho^\delta(\delta^{-1/2}\xi)\lambda_\rho(\delta^{-1/2}\xi)^{2k} \\
&\le C_\alpha e^{-|\xi|^2\widehat{Q}_\rho^2(\delta^{-1/2}\xi)/2},\qquad\xi\in B_{\delta^{1/2}}^c,
\end{aligned} ∂ α Θ ρ δ ( δ − 1/2 ξ ) = ( ∂ α Θ ρ δ ) ( δ − 1/2 ξ ) δ − ∣ α ∣/2 ≤ C α ∣ ξ ∣ − ∣ α ∣ k = 1 ∑ ∣ α ∣ δ k Θ ρ δ ( δ − 1/2 ξ ) λ ρ ( δ − 1/2 ξ ) 2 k ≤ C α e − ∣ ξ ∣ 2 Q ρ 2 ( δ − 1/2 ξ ) /2 , ξ ∈ B δ 1/2 c ,
where we used for the last inequality that ∣ ξ ∣ − ∣ α ∣ ≤ 1 |\xi|^{-|\alpha|}\le1 ∣ ξ ∣ − ∣ α ∣ ≤ 1 for ξ ∈ B δ 1 / 2 c \xi\in B_{\delta^{1/2}}^c ξ ∈ B δ 1/2 c , together with (A.16). Moreover, it holds with (36) that
∣ ξ ∣ 2 Q ^ ρ 2 ( δ − 1 / 2 ξ ) ≥ δ 1 − σ ∣ ξ ∣ 2 σ ≥ c ∣ ξ ∣ 2 σ , |\xi|^2\widehat{Q}_\rho^2(\delta^{-1/2}\xi)\ge\delta^{1-\sigma}|\xi|^{2\sigma}\ge c|\xi|^{2\sigma}, ∣ ξ ∣ 2 Q ρ 2 ( δ − 1/2 ξ ) ≥ δ 1 − σ ∣ ξ ∣ 2 σ ≥ c ∣ ξ ∣ 2 σ ,
and so we obtain
∣ ∂ α Θ ρ δ ( δ − 1 / 2 ξ ) ∣ ≤ C α e − c ∣ ξ ∣ 2 σ , ξ ∈ B δ 1 / 2 c \left|\partial^\alpha\Theta_\rho^\delta(\delta^{-1/2}\xi)\right|\le C_\alpha e^{-c|\xi|^{2\sigma}},\qquad\xi\in B_{\delta^{1/2}}^c ∂ α Θ ρ δ ( δ − 1/2 ξ ) ≤ C α e − c ∣ ξ ∣ 2 σ , ξ ∈ B δ 1/2 c
as claimed in (A.12). This finishes the proof.
Appendix B. Proof of Lemma 4.3
In this section, we prove Lemma 4.3, stating a second-order version of the nonlocal Leibniz rule in [8 Lemma 3.8] .
Proof of Lemma 4.3. Throughout the proof, C C C will denote a constant possibly changing from line to line, and possibly depending on the kernel ρ \rho ρ .
We first turn to the proof of estimate (65) . It was established in [8 Lemma 3.8] that
D ρ ( f g ) = f D ρ g + K f g (139) D^\rho(fg)=fD^\rho g+K_fg \tag*{(139)} D ρ ( f g ) = f D ρ g + K f g (139)
for
K f g ( x ) : = ∫ R n f ( x ) − f ( y ) ∣ x − y ∣ x − y ∣ x − y ∣ ρ ( x − y ) g ( y ) d y , x ∈ R n . K_fg(x):=\int_{\mathbb{R}^n}\frac{f(x)-f(y)}{|x-y|}\frac{x-y}{|x-y|}\rho(x-y)g(y)\,dy,\qquad x\in\mathbb{R}^n. K f g ( x ) := ∫ R n ∣ x − y ∣ f ( x ) − f ( y ) ∣ x − y ∣ x − y ρ ( x − y ) g ( y ) d y , x ∈ R n .
With the change of variables y ↦ x + y y\mapsto x+y y ↦ x + y and the radial symmetry of ρ \rho ρ according to (H0), K f g K_fg K f g can be equivalently expressed as
K f g ( x ) = ∫ R n f ( x + y ) − f ( x ) ∣ y ∣ y ∣ y ∣ ρ ( y ) g ( x + y ) d y , x ∈ R n . (140) K_fg(x)=\int_{\mathbb{R}^n}\frac{f(x+y)-f(x)}{|y|}\frac{y}{|y|}\rho(y)g(x+y)\,dy,\qquad x\in\mathbb{R}^n. \tag*{(140)} K f g ( x ) = ∫ R n ∣ y ∣ f ( x + y ) − f ( x ) ∣ y ∣ y ρ ( y ) g ( x + y ) d y , x ∈ R n . (140)
Moreover, as shown in the proof of [8 Lemma 3.8] , for the nonlocal divergence operator, one finds similarly
div ρ ( f G ) = f div ρ G + K ~ f G , G ∈ C c ∞ ( R n ; R n ) (141) \operatorname{div}^\rho(fG)=f\operatorname{div}^\rho G+\widetilde{K}_{fG},\qquad G\in C^\infty_c(\mathbb{R}^n;\mathbb{R}^n) \tag*{(141)} div ρ ( f G ) = f div ρ G + K f G , G ∈ C c ∞ ( R n ; R n ) (141)
for
K ~ f G ( x ) : = ∫ R n f ( x ) − f ( y ) ∣ x − y ∣ x − y ∣ x − y ∣ ρ ( x − y ) ⋅ G ( y ) d y , x ∈ R n . \widetilde{K}_{fG}(x):=\int_{\mathbb{R}^n}\frac{f(x)-f(y)}{|x-y|}\frac{x-y}{|x-y|}\rho(x-y)\cdot G(y)\,dy,\qquad x\in\mathbb{R}^n. K f G ( x ) := ∫ R n ∣ x − y ∣ f ( x ) − f ( y ) ∣ x − y ∣ x − y ρ ( x − y ) ⋅ G ( y ) d y , x ∈ R n .
Actually, the compact support of G G G is not relevant for the validity of the equality, and so the result extends directly to G ∈ C ∞ ( R n ; R n ) G\in C^\infty(\mathbb{R}^n;\mathbb{R}^n) G ∈ C ∞ ( R n ; R n ) .
With (B.1) and (B.3), we get the equality
Δ 0 ρ ( f g ) − f Δ 0 ρ g = K ~ f D ρ g + div ρ K f g , (142) \Delta^\rho_0(fg)-f\Delta^\rho_0g=\widetilde{K}_fD^\rho g+\operatorname{div}^\rho K_fg, \tag*{(142)} Δ 0 ρ ( f g ) − f Δ 0 ρ g = K f D ρ g + div ρ K f g , (142)
and it remains to estimate the expressions on the right-hand side. Using that ρ ∈ L 1 ( R n ) \rho\in L^1(\mathbb{R}^n) ρ ∈ L 1 ( R n ) , the term K ~ f D ρ g \widetilde{K}_fD^\rho g K f D ρ g can be estimated mimicking the steps in the proof of [8 Eq. (3.4), Lemma 3.8] , yielding
∥ K ~ f D ρ g ∥ L p ( R n ) ≤ C ∥ ∇ f ∥ L ∞ ( R n ; R n ) ∥ D ρ g ∥ L p ( R n ) . (143) \left\|\widetilde{K}_fD^\rho g\right\|_{L^p(\mathbb{R}^n)}\le C\|\nabla f\|_{L^\infty(\mathbb{R}^n;\mathbb{R}^n)}\|D^\rho g\|_{L^p(\mathbb{R}^n)}. \tag*{(143)} K f D ρ g L p ( R n ) ≤ C ∥∇ f ∥ L ∞ ( R n ; R n ) ∥ D ρ g ∥ L p ( R n ) . (143)
Next, we use that with (B.2) we have
( div ρ K f g ) ( x ) = ∫ R n D ρ ( f ( x + y ) − f ( x ) ∣ y ∣ g ( x + y ) ) ⋅ y ∣ y ∣ ρ ( y ) d y . \left(\operatorname{div}^\rho K_fg\right)(x)=\int_{\mathbb{R}^n}D^\rho\left(\frac{f(x+y)-f(x)}{|y|}g(x+y)\right)\cdot\frac{y}{|y|}\rho(y)\,dy. ( div ρ K f g ) ( x ) = ∫ R n D ρ ( ∣ y ∣ f ( x + y ) − f ( x ) g ( x + y ) ) ⋅ ∣ y ∣ y ρ ( y ) d y .
Another application of [8 Lemma 3.8] gives for x , y ∈ R n x,y\in\mathbb{R}^{n} x , y ∈ R n ,
D ρ ( f ( x + y ) − f ( x ) ∣ y ∣ g ( x + y ) ) = f ( x + y ) − f ( x ) ∣ y ∣ D ρ g ( x + y ) + K f ( x + y ) − f ( x ) ∣ y ∣ g ( x + y ) , D^{\rho}\left(\frac{f(x+y)-f(x)}{|y|}g(x+y)\right)=\frac{f(x+y)-f(x)}{|y|}D^{\rho}g(x+y)+K_{\frac{f(x+y)-f(x)}{|y|}}g(x+y), D ρ ( ∣ y ∣ f ( x + y ) − f ( x ) g ( x + y ) ) = ∣ y ∣ f ( x + y ) − f ( x ) D ρ g ( x + y ) + K ∣ y ∣ f ( x + y ) − f ( x ) g ( x + y ) ,
which can be estimated with the help of
∣ f ( x + y ) − f ( x ) ∣ y ∣ D ρ g ( x + y ) ∣ ≤ ∥ ∇ f ∥ L ∞ ( R n ; R n ) ∣ D ρ g ( x + y ) ∣ \left|\frac{f(x+y)-f(x)}{|y|}D^{\rho}g(x+y)\right|\leq\|\nabla f\|_{L^{\infty}(\mathbb{R}^{n};\mathbb{R}^{n})}|D^{\rho}g(x+y)| ∣ y ∣ f ( x + y ) − f ( x ) D ρ g ( x + y ) ≤ ∥∇ f ∥ L ∞ ( R n ; R n ) ∣ D ρ g ( x + y ) ∣
and
∣ K f ( x + y ) − f ( x ) ∣ y ∣ g ( x + y ) ∣ ≤ C ∥ ∇ f ∥ W 1 , ∞ ( R n ; R n ) ( ∣ g ∣ ∗ ρ ) ( x + y ) ; \left|K_{\frac{f(x+y)-f(x)}{|y|}}g(x+y)\right|\leq C\|\nabla f\|_{W^{1,\infty}(\mathbb{R}^{n};\mathbb{R}^{n})}(|g|\ast\rho)(x+y); K ∣ y ∣ f ( x + y ) − f ( x ) g ( x + y ) ≤ C ∥∇ f ∥ W 1 , ∞ ( R n ; R n ) ( ∣ g ∣ ∗ ρ ) ( x + y ) ;
the latter follows as in the proof of [8 Eq. (3.4), Lemma 3.8] . Altogether, taking into account again the spherical symmetry of ρ \rho ρ , it follows
∣ div ρ K f g ∣ ≤ C ∥ ∇ f ∥ W 1 , ∞ ( R n ; R n ) ( ∣ D ρ g ∣ ∗ ρ + ( ∣ g ∣ ∗ ρ ) ∗ ρ ) . |\operatorname{div}^{\rho}K_{f}g|\leq C\|\nabla f\|_{W^{1,\infty}(\mathbb{R}^{n};\mathbb{R}^{n})}(|D^{\rho}g|\ast\rho+(|g|\ast\rho)\ast\rho). ∣ div ρ K f g ∣ ≤ C ∥∇ f ∥ W 1 , ∞ ( R n ; R n ) ( ∣ D ρ g ∣ ∗ ρ + ( ∣ g ∣ ∗ ρ ) ∗ ρ ) .
Finally, recalling that ρ ∈ L 1 ( R n ) \rho\in L^{1}(\mathbb{R}^{n}) ρ ∈ L 1 ( R n ) , a repeated application of Young’s convolution inequality leads to the estimate
∥ div ρ K f g ∥ L p ( R n ) ≤ C ∥ ∇ f ∥ W 1 , ∞ ( R n ; R n ) ∥ g ∥ H ρ , p ( R n ) . (144) \|\operatorname{div}^{\rho}K_{f}g\|_{L^{p}(\mathbb{R}^{n})}\leq C\|\nabla f\|_{W^{1,\infty}(\mathbb{R}^{n};\mathbb{R}^{n})}\|g\|_{H_{\rho,p}(\mathbb{R}^{n})}. \tag*{(144)} ∥ div ρ K f g ∥ L p ( R n ) ≤ C ∥∇ f ∥ W 1 , ∞ ( R n ; R n ) ∥ g ∥ H ρ , p ( R n ) . (144)
The sought inequality (65) follows now from (B.4) together with the estimates (B.5) and (B.6). □
Declarations
Declaration of Generative AI and AI-Assisted Technologies: During the preparation of this manuscript, the authors used ChatGPT, Claude, Mistral and Copilot to assist with literature search and spell-checking, and to improve the fluency of selected portions of the text. No generative AI tools were used in the development of the mathematical ideas, the derivation of the theoretical results, or the proofs. The authors are fully responsible for the content of the manuscript.
Acknowledgments
The authors are grateful to José Carlos Bellido and Guillermo Gacía-Sáez for helpful discussions on the topic. This research was part of the Joint Spanish-German Research Project “Nonlocal Gradients in Variational Analysis and Materials Modelling: Limits, Kernels, Boundaries" funded by the Agencia Estatal de Investigación (AEI) and the Deutsche Forschungsgemeinschaft (DFG). Specifically, L.H. is employed under Project PCI2024-155023-2 funded by MICIU/AEI/10.13039/501100011033/UE and the European Union. C.K. acknowledges financial support by DFG via Project 541520348. The work of C. M.-C. was supported through Project PID2024-158664NB-C2 funded by MICIU/AEI/10.13039/501100011033/FEDER/UE.
Leon Happ. Departamento de Matemáticas, Universidad Autónoma de Madrid, C/ Francisco Tomás y Valiente 7, 28049 Madrid, Spain Email address: leon.happ@uam.es
Carolin-Kreisbeck. Mathematische-Geographische Fakultät, Katholische Universität Eichstätt-Ingolstadt, Ostenstraße 28, 85072 Eichstätt, Germany Email address: carolin.kreisbeck@ku.de
Carlos Mora-Corral. Departamento de Matemáticas, Universidad Autónoma de Madrid, C/ Francisco Tomás y Valiente 7, 28049 Madrid, Spain and Instituto de Ciencias Matemáticas, CSIC-UAM-UC3M-UCM, 28049 Madrid, Spain Email address: carlos.mora@uam.de