Introduction
The main purpose of this article is to solve the Kato square root problem for fractional elliptic operators
L = − div α ( μ ∇ α ⋅ ) on R d (1) L=-\operatorname{div}^{\alpha}(\mu\nabla^{\alpha}\cdot)\quad\text{on }\mathbb{R}^{d} \tag*{(1)} L = − div α ( μ ∇ α ⋅ ) on R d (1)
of order 2 α 2\alpha 2 α , α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , where ∇ α \nabla^{\alpha} ∇ α and div α \operatorname{div}^{\alpha} div α are the Riesz fractional gradient and divergence (see Definition 2.1 below) and μ \mu μ are bounded, measurable and uniformly elliptic complex coefficients. We show that the domain of L 1 / 2 L^{1/2} L 1/2 coincides with the distributional fractional Sobolev space S α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^{d}) S α , 2 ( R d ) consisting of all L 2 ( R d ) L^{2}(\mathbb{R}^{d}) L 2 ( R d ) -functions whose fractional gradient belongs to L 2 ( R d ; C d ) L^{2}(\mathbb{R}^{d};\mathbb{C}^{d}) L 2 ( R d ; C d ) , and that the homogeneous estimate ∥ L 1 / 2 u ∥ L 2 ≃ ∥ ∇ α u ∥ L 2 \|L^{1/2}u\|_{L^{2}}\simeq\|\nabla^{\alpha}u\|_{L^{2}} ∥ L 1/2 u ∥ L 2 ≃ ∥ ∇ α u ∥ L 2 holds; see Theorem 1.2 below.
The Kato square root problem
The Kato square root problem. At the beginning of the 1960s, Tosio Kato asked whether, given two Hilbert spaces V ⊆ H V\subseteq H V ⊆ H with V V V dense in H H H , the domain of the square root of the maximal accretive operator L L L in H H H associated with a closed and sectorial sesquilinear form a : V × V → C a:V\times V\to\mathbb{C} a : V × V → C always coincides with the form domain, that is, whether D ( L 1 / 2 ) = V \mathcal{D}(L^{1/2})=V D ( L 1/2 ) = V . Counterexamples by Lions [49 ] and McIntosh [52 ] showed that this fails in this generality. Lions already specialized the question to elliptic operators L = − div ( μ ∇ ⋅ ) L=-\operatorname{div}(\mu\nabla\cdot) L = − div ( μ ∇ ⋅ ) with bounded measurable coefficients, and in this form it became known as the Kato square root problem. It was resolved in the affirmative on R d \mathbb{R}^{d} R d by Auscher, Hofmann, Lacey, McIntosh and Tchamitchian [10 ] in 2002.
2020 Mathematics Subject Classification. 47F10, 47A60, 35R11, 26A33.
Key words and phrases. fractional elliptic operators, Kato square root problem, Riesz fractional gradient, nonlocal operators, off-diagonal estimates.
The author was supported by Studienstiftung des deutschen Volkes .
In the last two decades the square root property has been found to be deeply connected with boundary value problems [1 , 2 , 5 , 6 , 8 , 11 , 12 , 45 ] , operator-adapted function spaces [8 , 12 , 44 ] , maximal regularity and quasilinear equations [5 , 7 , 56 ] , and many further questions [4 , 17 , 18 , 37 , 43 ] . Since the restriction from general sectorial forms to elliptic operators in divergence form is a drastic one, the question for which further classes of operators the square root property persists has been present throughout. For elliptic operators, the square root property was extended to rough boundary geometries and various boundary conditions [15 , 19 , 20 , 21 , 22 , 38 ] and to operators on submanifolds [53 ] . Using the theory of Muckenhoupt weights, it was extended to elliptic operators with degenerate coefficients [31 , 32 , 33 ] , and it was further established for Schrödinger operators [16 , 54 ] and for parabolic operators [3 , 9 , 55 ] . As a key structural feature, all of these operators are built from classical differential operators, and hence local.
Nonlocality in the Kato problem
Nonlocality has entered the Kato problem before. For example, Tolksdorf and the author [41 ] established the square root property for the generalized Stokes operator, formally given by
A u : = − div ( μ ∇ u ) + ∇ φ , div ( u ) = 0 , Au := -\operatorname{div}(\mu\nabla u) + \nabla\varphi,\qquad\operatorname{div}(u) = 0, A u := − div ( μ ∇ u ) + ∇ φ , div ( u ) = 0 ,
with bounded measurable coefficients μ \mu μ satisfying a Gårding inequality. There, the nonlocality is caused by the pressure φ \varphi φ and the main difficulty it creates is the loss of local arguments, in particular of the classical exponential off-diagonal estimates for the resolvents [63 , 64 ] . A second, and rather different, form of nonlocality appears in the resolution of the parabolic Kato problem by Auscher, Egert and Nyström [9 ] . The parabolic operator ∂ t − div x ( μ ∇ x ⋅ ) \partial_t-\operatorname{div}_x(\mu\nabla_x\cdot) ∂ t − div x ( μ ∇ x ⋅ ) is itself local. However, if the operator is reformulated via a sesquilinear form or as a first-order system, nonlocal half-order derivatives in time enter, which create a major new challenge compared with the elliptic case. Fortunately, the nonlocality is confined to the time variable. This is a key observation made in [3 ] , where a second-order approach to the Kato property for weighted parabolic operators is established.
The operator L = − div α ( μ ∇ α ⋅ ) L=-\operatorname{div}^{\alpha}(\mu\nabla^{\alpha}\cdot) L = − div α ( μ ∇ α ⋅ ) studied in this article, by contrast, is nonlocal in every direction. It is built from the Riesz fractional gradient
( ∇ α f ) ( x ) = c d , α ∫ R d ( y − x ) ( f ( y ) − f ( x ) ) ∣ x − y ∣ d + 1 + α d y ( x ∈ R d ) (2) (\nabla^{\alpha}f)(x)=c_{d,\alpha}\int_{\mathbb{R}^{d}}\frac{(y-x)(f(y)-f(x))}{|x-y|^{d+1+\alpha}}\,dy\qquad(x\in\mathbb{R}^{d}) \tag*{(2)} ( ∇ α f ) ( x ) = c d , α ∫ R d ∣ x − y ∣ d + 1 + α ( y − x ) ( f ( y ) − f ( x )) d y ( x ∈ R d ) (2)
and the fractional divergence
( div α F ) ( x ) = c d , α ∫ R d ( y − x ) ⋅ ( F ( y ) − F ( x ) ) ∣ x − y ∣ d + 1 + α d y ( x ∈ R d ) (\operatorname{div}^{\alpha}F)(x)=c_{d,\alpha}\int_{\mathbb{R}^{d}}\frac{(y-x)\cdot(F(y)-F(x))}{|x-y|^{d+1+\alpha}}\,dy\qquad(x\in\mathbb{R}^{d}) ( div α F ) ( x ) = c d , α ∫ R d ∣ x − y ∣ d + 1 + α ( y − x ) ⋅ ( F ( y ) − F ( x )) d y ( x ∈ R d )
of order α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , where the normalization constant c d , α c_{d,\alpha} c d , α is given in Definition 2.1. Šilhavý [60 ] singled out this pair as the unique fractional differential operators, up to a multiplicative constant and subject to natural continuity requirements, that are invariant under translations and rotations and homogeneous of degree α \alpha α under isotropic scaling. They are tied to the fractional Laplacian through − div α ∇ α = ( − Δ ) α -\operatorname{div}^{\alpha}\nabla^{\alpha}=(-\Delta)^{\alpha} − div α ∇ α = ( − Δ ) α , the operators ∇ α \nabla^{\alpha} ∇ α and − div α -\operatorname{div}^{\alpha} − div α are adjoint, and the classical derivatives are recovered as α → 1 \alpha\to1 α → 1 [28 , 29 , 58 ] . Moreover, they give rise to a distributional theory of fractional Sobolev spaces, which coincide with the Bessel potential spaces [26 , 29 , 30 , 61 ] , and to a theory of fractional partial differential equations [57 , 58 , 59 ] . One motivation comes from nonlocal continuum mechanics, where these operators provide models, for example, of nonlocal hyperelasticity [24 ] and, upon linearization, recover Eringen’s model of nonlocal elasticity with singular kernels [23 ] . See also [25 , 34 , 48 , 51 ] for further developments.
Main results
We impose the following assumption on the coefficients.
Assumption 1.1. The coefficients μ ∈ L ∞ ( R d ; C d × d ) \mu\in L^\infty(\mathbb{R}^d;\mathbb{C}^{d\times d}) μ ∈ L ∞ ( R d ; C d × d ) satisfy for some 0 < μ ∙ ≤ μ ∙ 0 < \mu_{\bullet} \le\mu^{\bullet} 0 < μ ∙ ≤ μ ∙ the inequalities
Re ( μ ( x ) ξ ⋅ ξ ˉ ) ≥ μ ∙ ∣ ξ ∣ 2 and ∣ μ ( x ) ∣ C d × d ≤ μ ∙ ( ξ ∈ C d , a.e. x ∈ R d ) . \operatorname{Re}(\mu(x)\xi\cdot\bar{\xi}) \ge\mu_{\bullet}|\xi|^2 \qquad\text{and}\qquad|\mu(x)|_{\mathbb{C}^{d\times d}} \le\mu^{\bullet}\qquad(\xi\in\mathbb{C}^d,\ \text{a.e. }x\in\mathbb{R}^d). Re ( μ ( x ) ξ ⋅ ξ ˉ ) ≥ μ ∙ ∣ ξ ∣ 2 and ∣ μ ( x ) ∣ C d × d ≤ μ ∙ ( ξ ∈ C d , a.e. x ∈ R d ) .
The operator L L L in (1.1) is defined as the maximal accretive operator in L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) associated with the sesquilinear form
a ( u , v ) = ∫ R d μ ∇ α u ⋅ ∇ α v ‾ d x a(u,v)=\int_{\mathbb{R}^d}\mu\nabla^\alpha u\cdot\overline{\nabla^\alpha v}\,dx a ( u , v ) = ∫ R d μ ∇ α u ⋅ ∇ α v d x
on the distributional fractional Sobolev space S α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^d) S α , 2 ( R d ) ; see § 3 . We emphasize that L L L is in general not an integro-differential operator with a kernel comparable to ∣ x − y ∣ − d − 2 α |x-y|^{-d-2\alpha} ∣ x − y ∣ − d − 2 α , even for constant coefficients; see Remark 3.2 . Our main result reads as follows.
Theorem 1.2 (The Kato square root property of L L L ). Let α 0 ∈ ( 0 , 1 ) \alpha_0\in(0,1) α 0 ∈ ( 0 , 1 ) and let μ \mu μ satisfy Assumption 1.1. Then for all α ∈ [ α 0 , 1 ) \alpha\in[\alpha_0,1) α ∈ [ α 0 , 1 ) we have D ( L 1 2 ) = S α , 2 ( R d ) D(L^{\frac12})=S^{\alpha,2}(\mathbb{R}^d) D ( L 2 1 ) = S α , 2 ( R d ) and
∥ L 1 2 u ∥ L 2 ≃ ∥ ∇ α u ∥ L 2 ( u ∈ S α , 2 ( R d ) ) , \|L^{\frac12}u\|_{L^2}\simeq\|\nabla^\alpha u\|_{L^2}\qquad(u\in S^{\alpha,2}(\mathbb{R}^d)), ∥ L 2 1 u ∥ L 2 ≃ ∥ ∇ α u ∥ L 2 ( u ∈ S α , 2 ( R d )) , where the implicit constants depend only on d d d , α 0 \alpha_0 α 0 , μ ∙ \mu_{\bullet} μ ∙ and μ ∙ \mu^{\bullet} μ ∙ .
In particular, we obtain the Kato property for all α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . Moreover, since S α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^d) S α , 2 ( R d ) coincides with the fractional Sobolev–Slobodeckij space W α , 2 ( R d ) W^{\alpha,2}(\mathbb{R}^d) W α , 2 ( R d ) (see Proposition 2.5 ), we obtain the following description.
Corollary 1.3. Under the assumptions of Theorem 1.2,
D ( L 1 2 ) = W α , 2 ( R d ) D(L^{\frac12})=W^{\alpha,2}(\mathbb{R}^d) D ( L 2 1 ) = W α , 2 ( R d ) and for all u ∈ W α , 2 ( R d ) u\in W^{\alpha,2}(\mathbb{R}^d) u ∈ W α , 2 ( R d ) we have
∥ L 1 2 u ∥ L 2 ≃ ∥ ( − Δ ) α 2 u ∥ L 2 ≃ α 1 2 ( 1 − α ) 1 2 ∥ u ∥ W ˙ α , 2 . \|L^{\frac12}u\|_{L^2}\simeq\|(-\Delta)^{\frac{\alpha}{2}}u\|_{L^2}\simeq\alpha^{\frac12}(1-\alpha)^{\frac12}\|u\|_{\dot{W}^{\alpha,2}}. ∥ L 2 1 u ∥ L 2 ≃ ∥ ( − Δ ) 2 α u ∥ L 2 ≃ α 2 1 ( 1 − α ) 2 1 ∥ u ∥ W ˙ α , 2 .
As consequences of the Kato property, we obtain the L 2 L^2 L 2 -boundedness of the Riesz transform ∇ α L − 1 2 \nabla^\alpha L^{-\frac12} ∇ α L − 2 1 , as well as holomorphic and locally Lipschitz dependence of L 1 2 L^{\frac12} L 2 1 on the coefficients μ \mu μ in the L ∞ L^\infty L ∞ -topology; see § 5 . Furthermore, a special feature of the Kato property in Theorem 1.2 is that the implicit constants are locally uniform with respect to the fractional parameter α \alpha α . In particular, our result fits into the asymptotic analysis of the fractional calculus as α → 1 \alpha\to1 α → 1 [26 , 28 ] , and we recover the classical Kato square root property for elliptic operators L 1 = − div ( μ ∇ ⋅ ) L_1=-\operatorname{div}(\mu\nabla\cdot) L 1 = − div ( μ ∇ ⋅ ) in divergence form in the limit α → 1 \alpha\to1 α → 1 .
Proposition 1.4 (The Kato property of L 1 L_1 L 1 ). Let μ \mu μ satisfy Assumption 1.1. Then we have D ( L 1 1 2 ) = W 1 , 2 ( R d ) D(L_1^{\frac12})=W^{1,2}(\mathbb{R}^d) D ( L 1 2 1 ) = W 1 , 2 ( R d ) and
∥ L 1 1 2 u ∥ L 2 ≃ ∥ ∇ u ∥ L 2 ( u ∈ W 1 , 2 ( R d ) ) , \|L_1^{\frac12}u\|_{L^2}\simeq\|\nabla u\|_{L^2}\qquad(u\in W^{1,2}(\mathbb{R}^d)), ∥ L 1 2 1 u ∥ L 2 ≃ ∥∇ u ∥ L 2 ( u ∈ W 1 , 2 ( R d )) , where the implicit constants depend only on d d d , μ ∙ \mu_{\bullet} μ ∙ and μ ∙ \mu^{\bullet} μ ∙ .
Finally, we show continuity of the square root with respect to the parameter α \alpha α for fixed coefficients μ \mu μ . To formulate such a statement, let us emphasize the fractional parameter α \alpha α in the notation of L L L by writing L α L_\alpha L α .
Theorem 1.5. Let μ \mu μ satisfy Assumption 1.1. Then for every u ∈ W 1 , 2 ( R d ) u \in W^{1,2}(\mathbb{R}^d) u ∈ W 1 , 2 ( R d ) the map
( 0 , 1 ] → L 2 ( R d ) , α ↦ L α 1 / 2 u (0,1] \to L^2(\mathbb{R}^d), \qquad\alpha\mapsto L_\alpha^{1/2}u ( 0 , 1 ] → L 2 ( R d ) , α ↦ L α 1/2 u is continuous.
Strategy and main difficulties
The main part of this paper is devoted to the Kato property in Theorem 1.2. Here, we follow the second-order approach to the Kato problem for elliptic operators [10 ] . By an abstract argument, Theorem 1.2 reduces to a square function estimate for t α L ( 1 + t 2 α L ) − 1 t^\alpha L(1+t^{2\alpha}L)^{-1} t α L ( 1 + t 2 α L ) − 1 , which is in turn reduced, by a principal part approximation, to a Carleson measure estimate, and the latter is proved by a T ( b ) T(b) T ( b ) argument. In our setting, two steps of this scheme require new ideas.
The first point concerns the off-diagonal estimates. Classical off-diagonal estimates for resolvents of elliptic operators possess an exponential decay rate and are classically derived through local arguments. To be more precise, the derivation relies on testing the resolvent equation with its solution u u u multiplied by an exponential weight, or on an iteration of local Caccioppoli inequalities. However, due to the nonlocality of ∇ α \nabla^\alpha ∇ α and the polynomial decay of its kernel, we can neither preserve exponential decay nor derive local Caccioppoli inequalities. A similar phenomenon also appeared in the treatment of the generalized Stokes operator [41 ] . Nonetheless, it was observed there that the overall strategy for proving the Kato property remains stable under weaker nonlocal off-diagonal estimates with polynomial decay, provided the order is greater than the dimension d d d .
For the Stokes operator, estimates of this type are derived via an iteration procedure involving nonlocal Caccioppoli inequalities [41 ] . In our setting, we follow this approach by first establishing a novel nonlocal Caccioppoli inequality for the resolvent of L L L in Theorem 3.6 and then deriving the nonlocal off-diagonal estimates of polynomial order ν < d + α 0 \nu< d+\alpha_0 ν < d + α 0 by an iteration process in Theorem 3.7. Here, the main difficulty lies in controlling the nonlocal bilinear remainder ∇ N L α \nabla^\alpha_{\mathrm{NL}} ∇ NL α coming from the product rule for ∇ α \nabla^\alpha ∇ α (see Lemma 2.9 and 2.10). Since α 0 > 0 \alpha_0 > 0 α 0 > 0 , the decay order of our estimates can be chosen larger than d d d , which suffices for every step of the proof of Theorem 1.2.
For comparison, in the first-order approach to the Kato property for the parabolic operator [9 ] , the authors had to deal with similar issues due to the presence of a half-order time derivative. In particular, the off-diagonal estimates in [9 ] are of restricted polynomial order and of anisotropic type due to the separation of space and time. More recently, Ataei, Egert and Nyström [3 ] succeeded in providing a second-order proof of the Kato property for weighted parabolic operators, which makes use of the anisotropy so that off-diagonal estimates are only needed for operators that involve differentiation in space and no fractional operator. In this sense, the much more involved off-diagonal decay for nonlocal derivatives could be avoided. In our setting, however, the operator is nonlocal in every direction.
The second difficulty concerns the Carleson measure argument, which relies on local T ( b ) T(b) T ( b ) -techniques. More precisely, for every cube Q Q Q , one has to construct a test function b b b in the range of ∇ α \nabla^\alpha ∇ α that points roughly in a given direction on Q Q Q and satisfies ∥ b ∥ L 2 ≲ ∣ Q ∣ 1 / 2 \lVert b\rVert_{L^2} \lesssim\lvert Q\rvert^{1/2} ∥ b ∥ L 2 ≲ ∣ Q ∣ 1/2 . In the elliptic setting, one chooses b b b approximately equal to ∇ Φ \nabla\Phi ∇Φ with Φ \Phi Φ being the product of an affine function, whose gradient points in the given direction, and a cut-off function that equals one on Q Q Q . Since ∇ \nabla ∇ is a local operator, the cut-off does not affect the values and hence the direction of ∇ Φ \nabla\Phi ∇Φ on Q Q Q . For ∇ α \nabla^\alpha ∇ α , however, the values of ∇ α Φ \nabla^\alpha\Phi ∇ α Φ on Q Q Q depend on the values of Φ \Phi Φ on all of R d \mathbb{R}^d R d , so that the local construction no longer works. Such a situation also appears in the parabolic setting [9 ] due to the presence of the half-order time derivative. Luckily there, Φ \Phi Φ has a product structure in space and time, and the time part is obtained by inverting the nonlocal half-order derivative on the Fourier side. This is possible because the time variable is one-dimensional, so that the Fourier symbol of the half-order derivative is scalar. In our setting, however, for d ≥ 2 d \ge2 d ≥ 2 the Fourier symbol m ( ξ ) = i ξ ∣ ξ ∣ α − 1 m(\xi)=i\xi|\xi|^{\alpha-1} m ( ξ ) = i ξ ∣ ξ ∣ α − 1 of ∇ α \nabla^\alpha ∇ α is vector-valued and admits no right inverse. Thus, we instead exploit the factorization ∇ α = ∇ I 1 − α \nabla^\alpha=\nabla I_{1-\alpha} ∇ α = ∇ I 1 − α through the Riesz potential, which formally gives
∇ α ( − Δ ) 1 − α 2 = ∇ . \nabla^\alpha(-\Delta)^{\frac{1-\alpha}{2}}=\nabla. ∇ α ( − Δ ) 2 1 − α = ∇.
This reduces the construction to the local gradient ∇ \nabla ∇ , for which the elliptic construction is available again. We refer to § 4.3 for the details.
Moreover, we emphasize that all steps in our procedure keep track of the dependence on the fractional parameter α \alpha α , which eventually leads to the local uniformity of the implicit constants in Theorem 1.2 . This will be the key to obtaining asymptotic results in the limit α → 1 \alpha\to1 α → 1 . Here, two effects have to be balanced: the normalization constant c d , α c_{d,\alpha} c d , α in (2) vanishes like 1 − α 1-\alpha 1 − α , while integrals of the kernel ∣ x − y ∣ − ( d + α ) |x-y|^{-(d+\alpha)} ∣ x − y ∣ − ( d + α ) , such as in (7) , grow like ( 1 − α ) − 1 (1-\alpha)^{-1} ( 1 − α ) − 1 . In our estimates, both always appear together, so that they cancel out and the constants stay uniformly bounded with respect to α \alpha α . Moreover, we show that the fractional gradient ∇ α \nabla^\alpha ∇ α converges to ∇ \nabla ∇ and that the resolvents of L = − div α ( μ ∇ α ⋅ ) L=-\operatorname{div}^{\alpha}(\mu\nabla^\alpha\cdot) L = − div α ( μ ∇ α ⋅ ) converge to the resolvents of L 1 = − div ( μ ∇ ⋅ ) L_1=-\operatorname{div}(\mu\nabla\cdot) L 1 = − div ( μ ∇ ⋅ ) . Together with the uniform square function estimate for L L L obtained in the proof of Theorem 1.2 , this yields the corresponding square function estimate for L 1 L_1 L 1 , and thus Proposition 1.4 . The same ingredients, combined with the locally uniform Kato bounds, give the continuity property in Theorem 1.5 . Finally, the holomorphic and locally Lipschitz dependence on the coefficients in Proposition 5.3 and Corollary 5.4 follows as for elliptic operators [13 ] , see also [41 ] , since the argument only uses that L L L is in divergence form and that the constants in Theorem 1.2 depend on μ \mu μ only through μ ∙ \mu_{\bullet} μ ∙ and μ ∙ \mu^{\bullet} μ ∙ .
Organization
In Section 2 we collect the required properties of the Riesz fractional calculus that are used throughout the manuscript. In Section 3 we introduce the fractional elliptic operator L L L and prove a nonlocal Caccioppoli inequality and off-diagonal estimates for its resolvent. Section 4 contains the proofs of Theorem 1.2 and Corollary 1.3 . Finally, in Section 5 we present applications of the Kato property, namely the L 2 L^2 L 2 -boundedness of the Riesz transform associated with L L L , the holomorphic and Lipschitz dependence on the coefficients, and the continuous dependence on the fractional parameter α \alpha α as well as the limit α → 1 \alpha\to1 α → 1 .
Notation
Throughout, d ≥ 1 d\ge1 d ≥ 1 denotes the dimension and N 0 = N ∪ { 0 } \mathbb{N}_0=\mathbb{N}\cup\{0\} N 0 = N ∪ { 0 } . We write ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ for the Euclidean norms on C \mathbb{C} C and C d \mathbb{C}^d C d , and a ⋅ b = ∑ j a j b j a\cdot b=\sum_j a_jb_j a ⋅ b = ∑ j a j b j for a , b ∈ C d a,b\in\mathbb{C}^d a , b ∈ C d . The open ball with center x x x and radius r r r is denoted by B = B ( x , r ) B=B(x,r) B = B ( x , r ) , and for λ > 0 \lambda>0 λ > 0 we write λ B : = B ( x , λ r ) \lambda B:=B(x,\lambda r) λ B := B ( x , λ r ) and λ Q \lambda Q λ Q for the cube with the same center as Q Q Q and λ \lambda λ times its side-length ℓ ( Q ) \ell(Q) ℓ ( Q ) . For a ball B B B and k ∈ N 0 k\in\mathbb{N}_0 k ∈ N 0 we set B k : = 2 k B B_k:=2^kB B k := 2 k B and define the dyadic annuli
C 0 ( B ) : = 2 B , C k ( B ) : = B k + 1 ∖ B k ‾ ( k ≥ 1 ) , C_0(B):=2B,\qquad C_k(B):=B_{k+1}\setminus\overline{B_k}\quad(k\ge1), C 0 ( B ) := 2 B , C k ( B ) := B k + 1 ∖ B k ( k ≥ 1 ) ,
and analogously C k ( Q ) C_k(Q) C k ( Q ) for cubes Q Q Q . For ω ∈ ( 0 , π ] \omega\in(0,\pi] ω ∈ ( 0 , π ] let
Σ ω : = { z ∈ C ∖ { 0 } : ∣ arg z ∣ < ω } \Sigma_\omega:= \{z \in\mathbb{C} \setminus\{0\} : |\arg z| < \omega\} Σ ω := { z ∈ C ∖ { 0 } : ∣ arg z ∣ < ω }
and Σ 0 : = ( 0 , ∞ ) \Sigma_0 := (0,\infty) Σ 0 := ( 0 , ∞ ) . We write a ≲ b a \lesssim b a ≲ b if a ≤ C b a \le Cb a ≤ C b for a constant C > 0 C > 0 C > 0 that depends only on parameters not at stake, and a ≃ b a \simeq b a ≃ b if a ≲ b a \lesssim b a ≲ b and b ≲ a b \lesssim a b ≲ a hold. In some situations it will be more handy to keep the notation a ≤ C b a \le Cb a ≤ C b or C a ≥ b Ca \ge b C a ≥ b . In this case C C C is generic.
For a measurable set A A A we denote by ∣ A ∣ |A| ∣ A ∣ its Lebesgue measure and by 1 A \mathbf{1}_A 1 A its indicator function. Moreover, if A A A satisfies 0 < ∣ A ∣ < ∞ 0 < |A| < \infty 0 < ∣ A ∣ < ∞ , then for f ∈ L l o c 1 ( R d ) f \in L^1_{\mathrm{loc}}(\mathbb{R}^d) f ∈ L loc 1 ( R d ) we write
(f)_A := \fint_A f\,\mathrm{d}x = \frac{1}{|A|}\int_A f\,\mathrm{d}x.
We denote by W 1 , 2 ( R d ) W^{1,2}(\mathbb{R}^d) W 1 , 2 ( R d ) the classical Sobolev space. For β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) the fractional Sobolev–Slobodeckij space with differentiability β \beta β will be denoted by W β , 2 ( R d ) W^{\beta,2}(\mathbb{R}^d) W β , 2 ( R d ) . We write its norm in integral form
∥ f ∥ W β , 2 : = ∥ f ∥ L 2 + ∥ f ∥ W ˙ β , 2 with ∥ f ∥ W ˙ β , 2 : = ( ∬ R d × R d ∣ f ( x ) − f ( y ) ∣ 2 ∣ x − y ∣ d + 2 β d x d y ) 1 2 . \|f\|_{W^{\beta,2}} := \|f\|_{L^2} + \|f\|_{\dot{W}^{\beta,2}}
\quad\text{with}\quad
\|f\|_{\dot{W}^{\beta,2}} := \left(\iint_{\mathbb{R}^d \times\mathbb{R}^d} \frac{|f(x)-f(y)|^2}{|x-y|^{d+2\beta}}\,\mathrm{d}x\,\mathrm{d}y\right)^{\frac{1}{2}}. ∥ f ∥ W β , 2 := ∥ f ∥ L 2 + ∥ f ∥ W ˙ β , 2 with ∥ f ∥ W ˙ β , 2 := ( ∬ R d × R d ∣ x − y ∣ d + 2 β ∣ f ( x ) − f ( y ) ∣ 2 d x d y ) 2 1 .
For f ∈ W β , 2 ( R d ) f \in W^{\beta,2}(\mathbb{R}^d) f ∈ W β , 2 ( R d ) we will need to calculate a counterpart of the homogeneous W ˙ β , 2 \dot{W}^{\beta,2} W ˙ β , 2 -seminorm on a measurable set A ⊆ R d A \subseteq\mathbb{R}^d A ⊆ R d which we write as
∥ f ∥ W ˙ β , 2 ( A ) : = ( ∬ A × A ∣ f ( x ) − f ( y ) ∣ 2 ∣ x − y ∣ d + 2 β d x d y ) 1 2 . \|f\|_{\dot{W}^{\beta,2}(A)} := \left(\iint_{A \times A} \frac{|f(x)-f(y)|^2}{|x-y|^{d+2\beta}}\,\mathrm{d}x\,\mathrm{d}y\right)^{\frac{1}{2}}. ∥ f ∥ W ˙ β , 2 ( A ) := ( ∬ A × A ∣ x − y ∣ d + 2 β ∣ f ( x ) − f ( y ) ∣ 2 d x d y ) 2 1 .
A glimpse into fractional calculus
In this section we introduce the fractional calculus of Šilhavý [60 ] and collect the properties of its operators that are needed later on, see also [26 , 29 , 58 ] .
We begin with the integral representations of the fractional gradient and the fractional divergence.
Definition 2.1. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) and set
c d , α : = 2 α π − d 2 Γ ( d + α + 1 2 ) Γ ( 1 − α 2 ) . c_{d,\alpha} := 2^\alpha\pi^{-\frac{d}{2}} \frac{\Gamma\left(\frac{d+\alpha+1}{2}\right)}{\Gamma\left(\frac{1-\alpha}{2}\right)}. c d , α := 2 α π − 2 d Γ ( 2 1 − α ) Γ ( 2 d + α + 1 ) . The α \alpha α -fractional gradient of f ∈ C c ∞ ( R d ) f \in C^\infty_c(\mathbb{R}^d) f ∈ C c ∞ ( R d ) and the α \alpha α -fractional divergence of F ∈ C c ∞ ( R d ; C d ) F \in C^\infty_c(\mathbb{R}^d;\mathbb{C}^d) F ∈ C c ∞ ( R d ; C d ) are
( ∇ α f ) ( x ) : = c d , α ∫ R d ( z − x ) ( f ( z ) − f ( x ) ) ∣ x − z ∣ d + α + 1 d z (\nabla^\alpha f)(x) := c_{d,\alpha} \int_{\mathbb{R}^d} \frac{(z-x)(f(z)-f(x))}{|x-z|^{d+\alpha+1}}\,\mathrm{d}z ( ∇ α f ) ( x ) := c d , α ∫ R d ∣ x − z ∣ d + α + 1 ( z − x ) ( f ( z ) − f ( x )) d z and
( div α F ) ( x ) : = c d , α ∫ R d ( z − x ) ⋅ ( F ( z ) − F ( x ) ) ∣ x − z ∣ d + α + 1 d z (\operatorname{div}^\alpha F)(x) := c_{d,\alpha} \int_{\mathbb{R}^d} \frac{(z-x)\cdot(F(z)-F(x))}{|x-z|^{d+\alpha+1}}\,\mathrm{d}z ( div α F ) ( x ) := c d , α ∫ R d ∣ x − z ∣ d + α + 1 ( z − x ) ⋅ ( F ( z ) − F ( x )) d z for x ∈ R d x \in\mathbb{R}^d x ∈ R d .
Define the Fourier transform
F ( f ) ( ξ ) : = 1 ( 2 π ) d 2 ∫ R d f ( x ) e − i x ⋅ ξ d x . \mathcal{F}(f)(\xi) := \frac{1}{(2\pi)^{\frac{d}{2}}} \int_{\mathbb{R}^d} f(x)e^{-ix\cdot\xi}\,dx. F ( f ) ( ξ ) := ( 2 π ) 2 d 1 ∫ R d f ( x ) e − i x ⋅ ξ d x .
The following spectral representation is due to Shieh and Spector [58 Thm. 1.4] .
Proposition 2.3. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . For all f ∈ C c ∞ ( R d ) f \in C^\infty_c(\mathbb{R}^d) f ∈ C c ∞ ( R d ) and all F ∈ C c ∞ ( R d ; C d ) F \in C^\infty_c(\mathbb{R}^d;\mathbb{C}^d) F ∈ C c ∞ ( R d ; C d ) we have
F ( ∇ α f ) ( ξ ) = i ξ ∣ ξ ∣ 1 − α F ( f ) ( ξ ) and F ( div α F ) ( ξ ) = i ξ ⊤ ∣ ξ ∣ 1 − α F ( F ) ( ξ ) ( ξ ∈ R d ) . \mathcal{F}(\nabla^\alpha f)(\xi)=\frac{i\xi}{|\xi|^{1-\alpha}}\mathcal{F}(f)(\xi)
\quad\text{and}\quad
\mathcal{F}(\operatorname{div}^\alpha F)(\xi)=\frac{i\xi^\top}{|\xi|^{1-\alpha}}\mathcal{F}(F)(\xi)
\quad(\xi\in\mathbb{R}^d). F ( ∇ α f ) ( ξ ) = ∣ ξ ∣ 1 − α i ξ F ( f ) ( ξ ) and F ( div α F ) ( ξ ) = ∣ ξ ∣ 1 − α i ξ ⊤ F ( F ) ( ξ ) ( ξ ∈ R d ) . In particular, we have
− div α ∇ α = ( − Δ ) α , -\operatorname{div}^\alpha\nabla^\alpha=(-\Delta)^\alpha, − div α ∇ α = ( − Δ ) α , where ( − Δ ) α = F − 1 ( ∣ ξ ∣ 2 α F ) (-\Delta)^\alpha=\mathcal{F}^{-1}(|\xi|^{2\alpha}\mathcal{F}) ( − Δ ) α = F − 1 ( ∣ ξ ∣ 2 α F ) denotes the fractional Laplacian.
Distributional fractional Sobolev spaces
As for the classical operators, ∇ α \nabla^\alpha ∇ α and − div α -\operatorname{div}^\alpha − div α are adjoint to each other, see [29 Lem. 2.5] . More precisely, for α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , f ∈ C c ∞ ( R d ) f\in C^\infty_c(\mathbb{R}^d) f ∈ C c ∞ ( R d ) and F ∈ C c ∞ ( R d ; C d ) F\in C^\infty_c(\mathbb{R}^d;\mathbb{C}^d) F ∈ C c ∞ ( R d ; C d ) we have
∫ R d ∇ α f ⋅ F d x = − ∫ R d f div α F d x . \int_{\mathbb{R}^d}\nabla^\alpha f\cdot F\,dx=-\int_{\mathbb{R}^d}f\operatorname{div}^\alpha F\,dx. ∫ R d ∇ α f ⋅ F d x = − ∫ R d f div α F d x .
This duality relation is the starting point for a distributional fractional calculus imitating the classical Sobolev spaces.
Definition 2.4. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . Given f ∈ L 2 ( R d ) f\in L^2(\mathbb{R}^d) f ∈ L 2 ( R d ) , we say that g ∈ L l o c 1 ( R d ; C d ) g\in L^1_{\mathrm{loc}}(\mathbb{R}^d;\mathbb{C}^d) g ∈ L loc 1 ( R d ; C d ) is a weak α \alpha α -gradient of f f f , written g = ∇ α f g=\nabla^\alpha f g = ∇ α f , if
∫ R d g ⋅ F d x = − ∫ R d f div α F d x ( F ∈ C c ∞ ( R d ; C d ) ) . \int_{\mathbb{R}^d}g\cdot F\,dx=-\int_{\mathbb{R}^d}f\operatorname{div}^\alpha F\,dx
\qquad(F\in C^\infty_c(\mathbb{R}^d;\mathbb{C}^d)). ∫ R d g ⋅ F d x = − ∫ R d f div α F d x ( F ∈ C c ∞ ( R d ; C d )) .
We set
S α , 2 ( R d ) : = { f ∈ L 2 ( R d ) : ∇ α f ∈ L 2 ( R d ; C d ) } with ∥ f ∥ S α , 2 : = ∥ f ∥ L 2 + ∥ ∇ α f ∥ L 2 , S^{\alpha,2}(\mathbb{R}^d):=\{f\in L^2(\mathbb{R}^d):\nabla^\alpha f\in L^2(\mathbb{R}^d;\mathbb{C}^d)\}
\quad\text{with}\quad
\|f\|_{S^{\alpha,2}}:=\|f\|_{L^2}+\|\nabla^\alpha f\|_{L^2}, S α , 2 ( R d ) := { f ∈ L 2 ( R d ) : ∇ α f ∈ L 2 ( R d ; C d )} with ∥ f ∥ S α , 2 := ∥ f ∥ L 2 + ∥ ∇ α f ∥ L 2 ,
and we write S − α , 2 ( R d ) : = ( S α , 2 ( R d ) ) ′ S^{-\alpha,2}(\mathbb{R}^d):=(S^{\alpha,2}(\mathbb{R}^d))' S − α , 2 ( R d ) := ( S α , 2 ( R d ) ) ′ for the space of bounded antilinear functionals on S α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^d) S α , 2 ( R d ) .
The space S α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^d) S α , 2 ( R d ) is the natural domain of the multipliers of Proposition 2.3 , and can be identified with the fractional Sobolev space W α , 2 ( R d ) W^{\alpha,2}(\mathbb{R}^d) W α , 2 ( R d ) .
Proposition 2.5. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . Then S α , 2 ( R d ) = W α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^d)=W^{\alpha,2}(\mathbb{R}^d) S α , 2 ( R d ) = W α , 2 ( R d ) with equivalent norms, C c ∞ ( R d ) C^\infty_c(\mathbb{R}^d) C c ∞ ( R d ) is dense in S α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^d) S α , 2 ( R d ) , and for all u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) we have
∥ ∇ α u ∥ L 2 2 = ∥ ( − Δ ) α 2 u ∥ L 2 2 ≃ α ( 1 − α ) ∥ u ∥ W ˙ α , 2 2 (4) \|\nabla^\alpha u\|_{L^2}^2=\|(-\Delta)^{\frac{\alpha}{2}}u\|_{L^2}^2\simeq\alpha(1-\alpha)\|u\|_{\dot W^{\alpha,2}}^2 \tag*{(4)} ∥ ∇ α u ∥ L 2 2 = ∥ ( − Δ ) 2 α u ∥ L 2 2 ≃ α ( 1 − α ) ∥ u ∥ W ˙ α , 2 2 (4) with implicit constants depending only on d d d .
Proof. First, density of C c ∞ ( R d ) C^\infty_c(\mathbb{R}^d) C c ∞ ( R d ) in W α , 2 ( R d ) W^{\alpha,2}(\mathbb{R}^d) W α , 2 ( R d ) and S α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^d) S α , 2 ( R d ) was proven in [35 Thm. 2.4] and [26 Thm. 25] , respectively. Now, let u ∈ C c ∞ ( R d ) u \in C^\infty_c(\mathbb{R}^d) u ∈ C c ∞ ( R d ) . Since ∣ i ξ ∣ α − 1 ∣ ξ ∣ = ∣ ξ ∣ α |i\xi|^{\alpha-1}|\xi| = |\xi|^\alpha ∣ i ξ ∣ α − 1 ∣ ξ ∣ = ∣ ξ ∣ α , Proposition 2.3 and Plancherel’s theorem give the first equality in (4) . For the second equivalence, by [35 Prop. 3.6] , we have
∥ ( − Δ ) α 2 u ∥ L 2 2 = C ( d , α ) 2 ∥ u ∥ W ˙ α , 2 2 . \lVert(-\Delta)^{\frac{\alpha}{2}}u\rVert_{L^2}^2=\frac{C(d,\alpha)}{2}\lVert u\rVert_{\dot{W}^{\alpha,2}}^2. ∥( − Δ ) 2 α u ∥ L 2 2 = 2 C ( d , α ) ∥ u ∥ W ˙ α , 2 2 .
where the constant C ( d , α ) > 0 C(d,\alpha)>0 C ( d , α ) > 0 has the closed form
C ( d , α ) = ( ∫ R d 1 − cos ( ξ 1 ) ∣ ξ ∣ d + 2 α d ξ ) − 1 = α 4 α Γ ( d 2 + α ) π d 2 Γ ( 1 − α ) , C(d,\alpha)=\left(\int_{\mathbb{R}^d}\frac{1-\cos(\xi_1)}{|\xi|^{d+2\alpha}}\,\mathrm{d}\xi\right)^{-1}=\frac{\alpha4^\alpha\Gamma\left(\frac{d}{2}+\alpha\right)}{\pi^{\frac{d}{2}}\Gamma(1-\alpha)}, C ( d , α ) = ( ∫ R d ∣ ξ ∣ d + 2 α 1 − cos ( ξ 1 ) d ξ ) − 1 = π 2 d Γ ( 1 − α ) α 4 α Γ ( 2 d + α ) ,
see also [40 Sec. 1.10] . Using the identity ( 1 − α ) Γ ( 1 − α ) = Γ ( 2 − α ) (1-\alpha)\Gamma(1-\alpha)=\Gamma(2-\alpha) ( 1 − α ) Γ ( 1 − α ) = Γ ( 2 − α ) , we get
C ( d , α ) = α ( 1 − α ) 4 α Γ ( d 2 + α ) π d 2 Γ ( 2 − α ) . C(d,\alpha)=\alpha(1-\alpha)\frac{4^\alpha\Gamma\left(\frac{d}{2}+\alpha\right)}{\pi^{\frac{d}{2}}\Gamma(2-\alpha)}. C ( d , α ) = α ( 1 − α ) π 2 d Γ ( 2 − α ) 4 α Γ ( 2 d + α ) .
Since the last factor on the right-hand side is continuous and positive in α ∈ [ 0 , 1 ] \alpha\in[0,1] α ∈ [ 0 , 1 ] , there exist implicit constants depending only on d d d such that C ( d , α ) ≃ α ( 1 − α ) C(d,\alpha)\simeq\alpha(1-\alpha) C ( d , α ) ≃ α ( 1 − α ) . Finally, due to density of C c ∞ ( R d ) C^\infty_c(\mathbb{R}^d) C c ∞ ( R d ) in both fractional spaces, it follows that S α , 2 ( R d ) = W α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^d)=W^{\alpha,2}(\mathbb{R}^d) S α , 2 ( R d ) = W α , 2 ( R d ) with equivalent norms and that (4) extends to u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) .
In particular, we can extend the spectral representation of ∇ α \nabla^\alpha ∇ α from Proposition 2.3 to S α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^d) S α , 2 ( R d ) via density. We will use this fact freely, in particular in the next statement to obtain the following embedding chain; see also [26 Prop. 33] .
Lemma 2.6. Let 0 < β < α < 1 0<\beta<\alpha<1 0 < β < α < 1 . Then we have the continuous embeddings
W 1 , 2 ( R d ) ↪ S α , 2 ( R d ) ↪ S β , 2 ( R d ) . W^{1,2}(\mathbb{R}^d)\hookrightarrow S^{\alpha,2}(\mathbb{R}^d)\hookrightarrow S^{\beta,2}(\mathbb{R}^d). W 1 , 2 ( R d ) ↪ S α , 2 ( R d ) ↪ S β , 2 ( R d ) . Proof. Since 1 + ∣ ξ ∣ β ≤ 2 + ∣ ξ ∣ α ≤ 3 + ∣ ξ ∣ 1+|\xi|^\beta\leq2+|\xi|^\alpha\leq3+|\xi| 1 + ∣ ξ ∣ β ≤ 2 + ∣ ξ ∣ α ≤ 3 + ∣ ξ ∣ , we obtain with Proposition 2.3 and Plancherel’s theorem
∥ u ∥ S β , 2 ≲ ∥ u ∥ S α , 2 ≲ ∥ u ∥ W 1 , 2 \lVert u\rVert_{S^{\beta,2}}\lesssim\lVert u\rVert_{S^{\alpha,2}}\lesssim\lVert u\rVert_{W^{1,2}} ∥ u ∥ S β , 2 ≲ ∥ u ∥ S α , 2 ≲ ∥ u ∥ W 1 , 2 for u ∈ C c ∞ ( R d ) u\in C^\infty_c(\mathbb{R}^d) u ∈ C c ∞ ( R d ) . Since C c ∞ ( R d ) C^\infty_c(\mathbb{R}^d) C c ∞ ( R d ) is a dense subspace in all of the three spaces, we conclude the proof.
An alternative representation of the fractional gradient can be given in terms of the classical gradient and the Riesz potential. More precisely, for β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) , define the Riesz potential I β I_\beta I β on C c ∞ ( R d ) C^\infty_c(\mathbb{R}^d) C c ∞ ( R d ) by
( I β f ) ( x ) : = c d , 1 − β d − β ∫ R d f ( y ) ∣ x − y ∣ d − β d y ( x ∈ R d ) . (5) (I_\beta f)(x):=\frac{c_{d,1-\beta}}{d-\beta}\int_{\mathbb{R}^d}\frac{f(y)}{|x-y|^{d-\beta}}\,\mathrm{d}y\qquad(x\in\mathbb{R}^d). \tag*{(5)} ( I β f ) ( x ) := d − β c d , 1 − β ∫ R d ∣ x − y ∣ d − β f ( y ) d y ( x ∈ R d ) . (5)
Then, we have
∇ α f = I 1 − α ∇ f = ∇ I 1 − α f , (6) \nabla^\alpha f=I_{1-\alpha}\nabla f=\nabla I_{1-\alpha}f, \tag*{(6)} ∇ α f = I 1 − α ∇ f = ∇ I 1 − α f , (6)
for f ∈ C c ∞ ( R d ) f\in C^\infty_c(\mathbb{R}^d) f ∈ C c ∞ ( R d ) , see [29 Prop. 2.2] . Since ( − Δ ) 1 − α 2 (-\Delta)^{\frac{1-\alpha}{2}} ( − Δ ) 2 1 − α is the inverse of I 1 − α I_{1-\alpha} I 1 − α , see [29 Sec. 2.3] and [62 Sec. V.1] , composing the identity above with ( − Δ ) 1 − α 2 (-\Delta)^{\frac{1-\alpha}{2}} ( − Δ ) 2 1 − α restores (formally) the local gradient
∇ α ( − Δ ) 1 − α 2 f = ∇ I 1 − α ( − Δ ) 1 − α 2 f = ∇ f . \nabla^\alpha(-\Delta)^{\frac{1-\alpha}{2}}f=\nabla I_{1-\alpha}(-\Delta)^{\frac{1-\alpha}{2}}f=\nabla f. ∇ α ( − Δ ) 2 1 − α f = ∇ I 1 − α ( − Δ ) 2 1 − α f = ∇ f .
This relation between the classical and fractional gradient in the context of Sobolev functions was shown in [48 Prop. 3.1 and 3.3] and is recorded in the following proposition.
Proposition 2.7. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . For every v ∈ W 1 , 2 ( R d ) v \in W^{1,2}(\mathbb{R}^{d}) v ∈ W 1 , 2 ( R d ) the function
u : = ( − Δ ) 1 − α 2 v u := (-\Delta)^{\frac{1-\alpha}{2}}v u := ( − Δ ) 2 1 − α v belongs to S α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^{d}) S α , 2 ( R d ) and satisfies ∇ α u = ∇ v \nabla^\alpha u = \nabla v ∇ α u = ∇ v .
We also need a fractional divergence that is defined on all of L 2 ( R d ; C d ) L^{2}(\mathbb{R}^{d};\mathbb{C}^{d}) L 2 ( R d ; C d ) .
Definition 2.8. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . The very weak fractional divergence div α : L 2 ( R d ; C d ) → S − α , 2 ( R d ) \operatorname{div}^{\alpha}: L^{2}(\mathbb{R}^{d};\mathbb{C}^{d}) \to S^{-\alpha,2}(\mathbb{R}^{d}) div α : L 2 ( R d ; C d ) → S − α , 2 ( R d ) is defined by
⟨ div α F , g ⟩ : = − ∫ R d F ⋅ ∇ α g ‾ d x ( g ∈ S α , 2 ( R d ) ) . \langle\operatorname{div}^{\alpha} F,g\rangle:= -\int_{\mathbb{R}^{d}} F \cdot\overline{\nabla^{\alpha}g}\,\mathrm{d}x \qquad(g \in S^{\alpha,2}(\mathbb{R}^{d})). ⟨ div α F , g ⟩ := − ∫ R d F ⋅ ∇ α g d x ( g ∈ S α , 2 ( R d )) .
The Leibniz rule
In contrast to the classical case, the product rule for ∇ α \nabla^\alpha ∇ α carries a genuinely nonlocal, bilinear remainder. More precisely, for f , g ∈ C c ∞ ( R d ) f,g \in C^\infty_c(\mathbb{R}^{d}) f , g ∈ C c ∞ ( R d ) we define
∇ N L α ( f , g ) ( x ) : = c d , α ∫ R d ( y − x ) ( f ( y ) − f ( x ) ) ( g ( y ) − g ( x ) ) ∣ x − y ∣ d + α + 1 d y . \nabla^\alpha_{\mathrm{NL}}(f,g)(x) := c_{d,\alpha}\int_{\mathbb{R}^{d}} \frac{(y-x)(f(y)-f(x))(g(y)-g(x))}{\lvert x-y\rvert^{d+\alpha+1}}\,\mathrm{d}y. ∇ NL α ( f , g ) ( x ) := c d , α ∫ R d ∣ x − y ∣ d + α + 1 ( y − x ) ( f ( y ) − f ( x )) ( g ( y ) − g ( x )) d y .
It is this remainder that has to be controlled in the off-diagonal estimates of § 3.1 . The following Leibniz rule for the fractional gradient can be found in [48 Lem. 2.11] ; see also [30 Thm. 4.1] .
Lemma 2.9 (Leibniz rule). Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . For φ ∈ C c ∞ ( R d ) \varphi\in C^\infty_c(\mathbb{R}^{d}) φ ∈ C c ∞ ( R d ) and u ∈ S α , 2 ( R d ) u \in S^{\alpha,2}(\mathbb{R}^{d}) u ∈ S α , 2 ( R d ) we have u φ ∈ S α , 2 ( R d ) u\varphi\in S^{\alpha,2}(\mathbb{R}^{d}) u φ ∈ S α , 2 ( R d ) and
∇ α ( u φ ) = φ ∇ α u + u ∇ α φ + ∇ N L α ( u , φ ) , \nabla^\alpha(u\varphi) = \varphi\nabla^\alpha u + u\nabla^\alpha\varphi+ \nabla^\alpha_{\mathrm{NL}}(u,\varphi), ∇ α ( u φ ) = φ ∇ α u + u ∇ α φ + ∇ NL α ( u , φ ) , where all terms on the right-hand side belong to L 2 ( R d ; C d ) L^{2}(\mathbb{R}^{d};\mathbb{C}^{d}) L 2 ( R d ; C d ) .
We end this section by providing estimates for ∇ α \nabla^\alpha ∇ α and the remainder ∇ N L α \nabla^\alpha_{\mathrm{NL}} ∇ NL α that are needed in § 3 and § 4 .
Lemma 2.10. Let α 0 ∈ ( 0 , 1 ) \alpha_{0} \in(0,1) α 0 ∈ ( 0 , 1 ) , B = B ( x 0 , r ) B = B(x_{0},r) B = B ( x 0 , r ) be a ball, ρ ∈ ( 0 , r ] \rho\in(0,r] ρ ∈ ( 0 , r ] and let φ ∈ C c ∞ ( R d ) \varphi\in C^\infty_c(\mathbb{R}^{d}) φ ∈ C c ∞ ( R d ) satisfy 0 ≤ φ ≤ 1 0 \leq\varphi\leq1 0 ≤ φ ≤ 1 , supp φ ⊂ 2 B \operatorname{supp}\varphi\subset2B supp φ ⊂ 2 B and ρ ∥ ∇ φ ∥ L ∞ ≤ M \rho\lVert\nabla\varphi\rVert_{L^\infty} \leq M ρ ∥ ∇ φ ∥ L ∞ ≤ M for some M > 0 M > 0 M > 0 . Then there exists a constant C > 0 C > 0 C > 0 depending only on d d d , M M M and α 0 \alpha_{0} α 0 such that for all α ∈ [ α 0 , 1 ) \alpha\in[\alpha_{0},1) α ∈ [ α 0 , 1 ) the following holds.
(i) For every k ∈ N 0 k \in\mathbb{N}_{0} k ∈ N 0 we have
∥ ∇ α φ ∥ L ∞ ( C k ( B ) ) ≤ C 2 − ( d + α ) k ρ α . \lVert\nabla^\alpha\varphi\rVert_{L^\infty(C_k(B))} \leq C\frac{2^{-(d+\alpha)k}}{\rho^\alpha}. ∥ ∇ α φ ∥ L ∞ ( C k ( B )) ≤ C ρ α 2 − ( d + α ) k . (ii) For u ∈ L 2 ( R d ) u \in L^{2}(\mathbb{R}^{d}) u ∈ L 2 ( R d ) we have
∥ ∇ N L α ( u , φ ) ∥ L 2 ( 2 B ) ≤ C ρ α ∑ m = 0 ∞ 2 − ( d 2 + α ) m ∥ u ∥ L 2 ( C m ( B ) ) \lVert\nabla^\alpha_{\mathrm{NL}}(u,\varphi)\rVert_{L^{2}(2B)} \leq\frac{C}{\rho^\alpha}\sum_{m=0}^{\infty}2^{-(\frac{d}{2}+\alpha)m}\lVert u\rVert_{L^{2}(C_m(B))} ∥ ∇ NL α ( u , φ ) ∥ L 2 ( 2 B ) ≤ ρ α C m = 0 ∑ ∞ 2 − ( 2 d + α ) m ∥ u ∥ L 2 ( C m ( B )) and, for every $k \geq 1,
∥ ∇ N L α ( u , φ ) ∥ L 2 ( C k ( B ) ) ≤ C ρ α ( 2 − ( d + α ) k ∥ u ∥ L 2 ( C k ( B ) ) + 2 − ( d 2 + α ) k ∥ u ∥ L 2 ( 2 B ) ) . \lVert\nabla^\alpha_{\mathrm{NL}}(u,\varphi)\rVert_{L^{2}(C_k(B))} \leq\frac{C}{\rho^\alpha}\left(2^{-(d+\alpha)k}\lVert u\rVert_{L^{2}(C_k(B))}+2^{-(\frac{d}{2}+\alpha)k}\lVert u\rVert_{L^{2}(2B)}\right). ∥ ∇ NL α ( u , φ ) ∥ L 2 ( C k ( B )) ≤ ρ α C ( 2 − ( d + α ) k ∥ u ∥ L 2 ( C k ( B )) + 2 − ( 2 d + α ) k ∥ u ∥ L 2 ( 2 B ) ) .
Proof. Since 0 ≤ φ ≤ 1 0 \le\varphi\le1 0 ≤ φ ≤ 1 and ρ ∥ ∇ φ ∥ L ∞ ≲ 1 \rho\|\nabla\varphi\|_{L^\infty} \lesssim1 ρ ∥∇ φ ∥ L ∞ ≲ 1 , we have ∣ φ ( x ) − φ ( y ) ∣ ≲ min { ∣ x − y ∣ / ρ , 1 } |\varphi(x)-\varphi(y)| \lesssim\min\{|x-y|/\rho,1\} ∣ φ ( x ) − φ ( y ) ∣ ≲ min { ∣ x − y ∣/ ρ , 1 } . Set κ ( z ) : = min { ∣ z ∣ / ρ , 1 } ∣ z ∣ − d − α \kappa(z):=\min\{|z|/\rho,1\}|z|^{-d-\alpha} κ ( z ) := min { ∣ z ∣/ ρ , 1 } ∣ z ∣ − d − α and notice that
∥ κ ∥ L 1 = 1 ρ ∫ B ( 0 , ρ ) 1 ∣ z ∣ d + α − 1 d z + ∫ R d ∖ B ( 0 , ρ ) 1 ∣ z ∣ d + α d z ≲ 1 ( 1 − α ) ρ α . (7) \|\kappa\|_{L^1}=\frac{1}{\rho}\int_{B(0,\rho)}\frac{1}{|z|^{d+\alpha-1}}\,\mathrm{d}z+\int_{\mathbb{R}^d\setminus B(0,\rho)}\frac{1}{|z|^{d+\alpha}}\,\mathrm{d}z\lesssim\frac{1}{(1-\alpha)\rho^\alpha}. \tag*{(7)} ∥ κ ∥ L 1 = ρ 1 ∫ B ( 0 , ρ ) ∣ z ∣ d + α − 1 1 d z + ∫ R d ∖ B ( 0 , ρ ) ∣ z ∣ d + α 1 d z ≲ ( 1 − α ) ρ α 1 . (7)
We divide the proof into estimating the fractional gradient and the bilinear term.
Step 1: Show (i). For every x ∈ R d x\in\mathbb{R}^d x ∈ R d we calculate
∣ ( ∇ α φ ) ( x ) ∣ ≲ c d , α ∫ R d ∣ κ ( x − y ) ∣ d y = c d , α ∥ κ ∥ L 1 ≲ 1 ρ α . |(\nabla^\alpha\varphi)(x)|\lesssim c_{d,\alpha}\int_{\mathbb{R}^d}|\kappa(x-y)|\,\mathrm{d}y=c_{d,\alpha}\|\kappa\|_{L^1}\lesssim\frac{1}{\rho^\alpha}. ∣ ( ∇ α φ ) ( x ) ∣ ≲ c d , α ∫ R d ∣ κ ( x − y ) ∣ d y = c d , α ∥ κ ∥ L 1 ≲ ρ α 1 .
where we used (2.5) and the uniform estimate (2.1). This proves (i) for k = 0 , 1 k=0,1 k = 0 , 1 . If x ∈ C k ( B ) x\in C_k(B) x ∈ C k ( B ) , for k ≥ 2 k\ge2 k ≥ 2 , we notice that φ ( x ) = 0 \varphi(x)=0 φ ( x ) = 0 . Thus,
∣ ( ∇ α φ ) ( x ) ∣ ≤ c d , α ∫ R d ∣ φ ( y ) − φ ( x ) ∣ ∣ x − y ∣ d + α d y = c d , α ∫ 2 B ∣ φ ( y ) ∣ ∣ x − y ∣ d + α d y . |(\nabla^\alpha\varphi)(x)|\le c_{d,\alpha}\int_{\mathbb{R}^d}\frac{|\varphi(y)-\varphi(x)|}{|x-y|^{d+\alpha}}\,\mathrm{d}y=c_{d,\alpha}\int_{2B}\frac{|\varphi(y)|}{|x-y|^{d+\alpha}}\,\mathrm{d}y. ∣ ( ∇ α φ ) ( x ) ∣ ≤ c d , α ∫ R d ∣ x − y ∣ d + α ∣ φ ( y ) − φ ( x ) ∣ d y = c d , α ∫ 2 B ∣ x − y ∣ d + α ∣ φ ( y ) ∣ d y .
Moreover, for x ∈ C k ( B ) x\in C_k(B) x ∈ C k ( B ) and y ∈ 2 B y\in2B y ∈ 2 B we have ∣ x − y ∣ ≳ 2 k r |x-y|\gtrsim2^k r ∣ x − y ∣ ≳ 2 k r so that with (2.1) and ρ ≤ r \rho\le r ρ ≤ r it follows
∣ ( ∇ α φ ) ( x ) ∣ ≲ c d , α ( 2 k r ) d + α ∫ 2 B ∣ φ ( y ) ∣ d y ≲ 2 − ( d + α ) k ρ α . |(\nabla^\alpha\varphi)(x)|\lesssim\frac{c_{d,\alpha}}{(2^k r)^{d+\alpha}}\int_{2B}|\varphi(y)|\,\mathrm{d}y\lesssim\frac{2^{-(d+\alpha)k}}{\rho^\alpha}. ∣ ( ∇ α φ ) ( x ) ∣ ≲ ( 2 k r ) d + α c d , α ∫ 2 B ∣ φ ( y ) ∣ d y ≲ ρ α 2 − ( d + α ) k .
Step 2: Show (ii). Notice that for all x ∈ R d x\in\mathbb{R}^d x ∈ R d we have
∣ ∇ N L α ( u , φ ) ( x ) ∣ ≤ c d , α ∫ R d ∣ u ( x ) − u ( y ) ∣ ∣ φ ( x ) − φ ( y ) ∣ ∣ x − y ∣ d + α d y ≤ c d , α ( ∣ u ( x ) ∣ ∫ R d ∣ φ ( x ) − φ ( y ) ∣ ∣ x − y ∣ d + α d y + ∫ R d ∣ u ( y ) ∣ ∣ φ ( x ) − φ ( y ) ∣ ∣ x − y ∣ d + α d y ) . \begin{aligned}
|\nabla^\alpha_{\mathrm{NL}}(u,\varphi)(x)|\le c_{d,\alpha}\int_{\mathbb{R}^d}\frac{|u(x)-u(y)||\varphi(x)-\varphi(y)|}{|x-y|^{d+\alpha}}\,\mathrm{d}y \\
&\le c_{d,\alpha}\left(|u(x)|\int_{\mathbb{R}^d}\frac{|\varphi(x)-\varphi(y)|}{|x-y|^{d+\alpha}}\,\mathrm{d}y+\int_{\mathbb{R}^d}\frac{|u(y)||\varphi(x)-\varphi(y)|}{|x-y|^{d+\alpha}}\,\mathrm{d}y\right).
\end{aligned} ∣ ∇ NL α ( u , φ ) ( x ) ∣ ≤ c d , α ∫ R d ∣ x − y ∣ d + α ∣ u ( x ) − u ( y ) ∣∣ φ ( x ) − φ ( y ) ∣ d y ≤ c d , α ( ∣ u ( x ) ∣ ∫ R d ∣ x − y ∣ d + α ∣ φ ( x ) − φ ( y ) ∣ d y + ∫ R d ∣ x − y ∣ d + α ∣ u ( y ) ∣∣ φ ( x ) − φ ( y ) ∣ d y ) .
Thus, by Step 1, for every k ∈ N 0 k\in\mathbb{N}_0 k ∈ N 0 we have the bound
∥ ∇ N L α ( u , φ ) ∥ L 2 ( C k ( B ) ) ≲ 2 − ( d + α ) k ∥ u ∥ L 2 ( C k ( B ) ) ρ α + c d , α ( ∫ C k ( B ) ( ∫ R d ∣ u ( y ) ∣ ∣ φ ( x ) − φ ( y ) ∣ ∣ x − y ∣ d + α d y ) 2 d x ) 1 / 2 \|\nabla^\alpha_{\mathrm{NL}}(u,\varphi)\|_{L^2(C_k(B))}\lesssim\frac{2^{-(d+\alpha)k}\|u\|_{L^2(C_k(B))}}{\rho^\alpha}+c_{d,\alpha}\left(\int_{C_k(B)}\left(\int_{\mathbb{R}^d}\frac{|u(y)||\varphi(x)-\varphi(y)|}{|x-y|^{d+\alpha}}\,\mathrm{d}y\right)^2\mathrm{d}x\right)^{1/2} ∥ ∇ NL α ( u , φ ) ∥ L 2 ( C k ( B )) ≲ ρ α 2 − ( d + α ) k ∥ u ∥ L 2 ( C k ( B )) + c d , α ( ∫ C k ( B ) ( ∫ R d ∣ x − y ∣ d + α ∣ u ( y ) ∣∣ φ ( x ) − φ ( y ) ∣ d y ) 2 d x ) 1/2
≲ 2 − ( d + α ) k ∥ u ∥ L 2 ( C k ( B ) ) ρ α + c d , α ∑ m = 0 ∞ I ( k , m ) , (8) \lesssim\frac{2^{-(d+\alpha)k}\|u\|_{L^2(C_k(B))}}{\rho^\alpha}+c_{d,\alpha}\sum_{m=0}^{\infty}I(k,m), \tag*{(8)} ≲ ρ α 2 − ( d + α ) k ∥ u ∥ L 2 ( C k ( B )) + c d , α m = 0 ∑ ∞ I ( k , m ) , (8)
where we set
I ( k , m ) : = ( ∫ C k ( B ) ( ∫ C m ( B ) ∣ u ( y ) ∣ ∣ φ ( x ) − φ ( y ) ∣ ∣ x − y ∣ d + α d y ) 2 d x ) 1 / 2 ( k , m ∈ N 0 ) . I(k,m):=\left(\int_{C_k(B)}\left(\int_{C_m(B)}\frac{|u(y)||\varphi(x)-\varphi(y)|}{|x-y|^{d+\alpha}}\,\mathrm{d}y\right)^2\mathrm{d}x\right)^{1/2}\qquad(k,m\in\mathbb{N}_0). I ( k , m ) := ∫ C k ( B ) ( ∫ C m ( B ) ∣ x − y ∣ d + α ∣ u ( y ) ∣∣ φ ( x ) − φ ( y ) ∣ d y ) 2 d x 1/2 ( k , m ∈ N 0 ) .
We further investigate I ( k , m ) I(k,m) I ( k , m ) by distinguishing different cases.
Case 1: k = 0 k=0 k = 0 . Assume 0 ≤ m ≤ 2 0\le m\le2 0 ≤ m ≤ 2 . We use Young’s convolution inequality and (7) to obtain
I ( 0 , m ) ≲ ∥ κ ∗ ( 1 C m ( B ) u ) ∥ L 2 ( 2 B ) ≤ ∥ κ ∥ L 1 ∥ u ∥ L 2 ( C m ( B ) ) ≲ ∥ u ∥ L 2 ( C m ( B ) ) ( 1 − α ) ρ α . I(0,m)\lesssim\|\kappa*(1_{C_m(B)}u)\|_{L^2(2B)}\le\|\kappa\|_{L^1}\|u\|_{L^2(C_m(B))}\lesssim\frac{\|u\|_{L^2(C_m(B))}}{(1-\alpha)\rho^\alpha}. I ( 0 , m ) ≲ ∥ κ ∗ ( 1 C m ( B ) u ) ∥ L 2 ( 2 B ) ≤ ∥ κ ∥ L 1 ∥ u ∥ L 2 ( C m ( B )) ≲ ( 1 − α ) ρ α ∥ u ∥ L 2 ( C m ( B )) .
In the case m > 2 m>2 m > 2 , we notice that φ ( y ) = 0 \varphi(y)=0 φ ( y ) = 0 and ∣ x − y ∣ ≳ 2 m r |x-y|\gtrsim2^m r ∣ x − y ∣ ≳ 2 m r for x ∈ 2 B x\in2B x ∈ 2 B and y ∈ C m ( B ) y\in C_m(B) y ∈ C m ( B ) . Hence, we estimate with Hölder’s inequality
I ( 0 , m ) ≲ ( ∫ 2 B ( ∫ C m ( B ) ∣ u ( y ) ∣ ∣ φ ( x ) ∣ ( 2 m r ) d + α d y ) 2 d x ) 1 2 ≤ ∣ 2 B ∣ 1 2 ( 2 m r ) d + α ∥ u ∥ L 1 ( C m ( B ) ) ≤ ∣ 2 B ∣ 1 2 ∣ C m ( B ) ∣ 1 2 ( 2 m r ) d + α ∥ u ∥ L 2 ( C m ( B ) ) \begin{aligned}
I(0,m)\lesssim\left(\int_{2B}\left(\int_{C_m(B)}\frac{|u(y)||\varphi(x)|}{(2^m r)^{d+\alpha}}\,\mathrm{d}y\right)^2\mathrm{d}x\right)^{\frac{1}{2}}\le\frac{|2B|^{\frac{1}{2}}}{(2^m r)^{d+\alpha}}\|u\|_{L^1(C_m(B))} \\
&\le\frac{|2B|^{\frac{1}{2}}|C_m(B)|^{\frac{1}{2}}}{(2^m r)^{d+\alpha}}\|u\|_{L^2(C_m(B))}
\end{aligned} I ( 0 , m ) ≲ ∫ 2 B ( ∫ C m ( B ) ( 2 m r ) d + α ∣ u ( y ) ∣∣ φ ( x ) ∣ d y ) 2 d x 2 1 ≤ ( 2 m r ) d + α ∣2 B ∣ 2 1 ∥ u ∥ L 1 ( C m ( B )) ≤ ( 2 m r ) d + α ∣2 B ∣ 2 1 ∣ C m ( B ) ∣ 2 1 ∥ u ∥ L 2 ( C m ( B ))
≲ 2 − ( d 2 + α ) m ρ α ∥ u ∥ L 2 ( C m ( B ) ) , \lesssim\frac{2^{-\left(\frac{d}{2}+\alpha\right)m}}{\rho^\alpha}\|u\|_{L^2(C_m(B))}, ≲ ρ α 2 − ( 2 d + α ) m ∥ u ∥ L 2 ( C m ( B )) ,
where we used ρ ≤ r \rho\le r ρ ≤ r in the last step. Combining both estimates in (8) , and using the uniform bound (3) , we obtain
∥ ∇ N L α ( u , φ ) ∥ L 2 ( 2 B ) ≲ ∑ m = 0 ∞ 2 − ( d 2 + α ) m ρ α ∥ u ∥ L 2 ( C m ( B ) ) . \|\nabla^\alpha_{\mathrm{NL}}(u,\varphi)\|_{L^2(2B)}\lesssim\sum_{m=0}^{\infty}\frac{2^{-\left(\frac{d}{2}+\alpha\right)m}}{\rho^\alpha}\|u\|_{L^2(C_m(B))}. ∥ ∇ NL α ( u , φ ) ∥ L 2 ( 2 B ) ≲ m = 0 ∑ ∞ ρ α 2 − ( 2 d + α ) m ∥ u ∥ L 2 ( C m ( B )) .
Case 2: k ≥ 1 k\ge1 k ≥ 1 . If m ≥ 1 m\ge1 m ≥ 1 , we get I ( k , m ) = 0 I(k,m)=0 I ( k , m ) = 0 because φ = 0 \varphi=0 φ = 0 on C m ( B ) C_m(B) C m ( B ) and C k ( B ) C_k(B) C k ( B ) . Thus, assume m = 0 m=0 m = 0 . If k = 1 k=1 k = 1 , we use Young’s convolution inequality and (7) to arrive at
I ( 1 , 0 ) ≲ ∥ κ ∗ ( 1 2 B u ) ∥ L 2 ( C 1 ( B ) ) ≲ ∥ κ ∥ L 1 ∥ u ∥ L 2 ( 2 B ) ≲ ∥ u ∥ L 2 ( 2 B ) ( 1 − α ) ρ α . I(1,0)\lesssim\|\kappa*(1_{2B}u)\|_{L^2(C_1(B))}\lesssim\|\kappa\|_{L^1}\|u\|_{L^2(2B)}\lesssim\frac{\|u\|_{L^2(2B)}}{(1-\alpha)\rho^\alpha}. I ( 1 , 0 ) ≲ ∥ κ ∗ ( 1 2 B u ) ∥ L 2 ( C 1 ( B )) ≲ ∥ κ ∥ L 1 ∥ u ∥ L 2 ( 2 B ) ≲ ( 1 − α ) ρ α ∥ u ∥ L 2 ( 2 B ) .
If k ≥ 2 k\ge2 k ≥ 2 , then φ ( x ) = 0 \varphi(x)=0 φ ( x ) = 0 and ∣ x − y ∣ ≳ 2 k r |x-y|\gtrsim2^k r ∣ x − y ∣ ≳ 2 k r for x ∈ C k ( B ) x\in C_k(B) x ∈ C k ( B ) and y ∈ 2 B y\in2B y ∈ 2 B , so that
I ( k , 0 ) ≲ ( ∫ C k ( B ) ( ∫ 2 B ∣ u ( y ) ∣ ∣ φ ( y ) ∣ ( 2 k r ) d + α d y ) 2 d x ) 1 2 ≤ ∣ C k ( B ) ∣ 1 2 ( 2 k r ) d + α ∥ u ∥ L 1 ( 2 B ) ≤ ∣ C k ( B ) ∣ 1 2 ∣ 2 B ∣ 1 2 ( 2 k r ) d + α ∥ u ∥ L 2 ( 2 B ) \begin{aligned}
I(k,0)\lesssim\left(\int_{C_k(B)}\left(\int_{2B}\frac{|u(y)||\varphi(y)|}{(2^k r)^{d+\alpha}}\,\mathrm{d}y\right)^2\mathrm{d}x\right)^{\frac{1}{2}} \\
&\le\frac{|C_k(B)|^{\frac{1}{2}}}{(2^k r)^{d+\alpha}}\|u\|_{L^1(2B)} \\
&\le\frac{|C_k(B)|^{\frac{1}{2}}|2B|^{\frac{1}{2}}}{(2^k r)^{d+\alpha}}\|u\|_{L^2(2B)}
\end{aligned} I ( k , 0 ) ≲ ( ∫ C k ( B ) ( ∫ 2 B ( 2 k r ) d + α ∣ u ( y ) ∣∣ φ ( y ) ∣ d y ) 2 d x ) 2 1 ≤ ( 2 k r ) d + α ∣ C k ( B ) ∣ 2 1 ∥ u ∥ L 1 ( 2 B ) ≤ ( 2 k r ) d + α ∣ C k ( B ) ∣ 2 1 ∣2 B ∣ 2 1 ∥ u ∥ L 2 ( 2 B )
≲ 2 − ( d 2 + α ) k ρ α ∥ u ∥ L 2 ( 2 B ) , \lesssim\frac{2^{-\left(\frac{d}{2}+\alpha\right)k}}{\rho^\alpha}\|u\|_{L^2(2B)}, ≲ ρ α 2 − ( 2 d + α ) k ∥ u ∥ L 2 ( 2 B ) ,
where we used Hölder’s inequality in the second last step and ρ ≤ r \rho\le r ρ ≤ r in the last step. Combining both cases in (8) , and using the uniform bound (3) , we obtain the estimate
∥ ∇ N L α ( u , φ ) ∥ L 2 ( C k ( B ) ) ≲ 2 − ( d + α ) k ρ α ∥ u ∥ L 2 ( C k ( B ) ) + 2 − ( d 2 + α ) k ρ α ∥ u ∥ L 2 ( 2 B ) . \|\nabla^\alpha_{\mathrm{NL}}(u,\varphi)\|_{L^2(C_k(B))}\lesssim\frac{2^{-(d+\alpha)k}}{\rho^\alpha}\|u\|_{L^2(C_k(B))}+\frac{2^{-\left(\frac{d}{2}+\alpha\right)k}}{\rho^\alpha}\|u\|_{L^2(2B)}. ∥ ∇ NL α ( u , φ ) ∥ L 2 ( C k ( B )) ≲ ρ α 2 − ( d + α ) k ∥ u ∥ L 2 ( C k ( B )) + ρ α 2 − ( 2 d + α ) k ∥ u ∥ L 2 ( 2 B ) .
This yields (ii) for every u ∈ L 2 ( R d ) u\in L^2(\mathbb{R}^d) u ∈ L 2 ( R d ) in the case k ≥ 1 k\ge1 k ≥ 1 . □
Remark 2.11. Lemma 2.10 can be stated and proved for cubes Q Q Q instead of balls B B B in a straightforward way, where r r r is replaced by ℓ ( Q ) \ell(Q) ℓ ( Q ) .
Fractional elliptic operator
In this section, we define fractional elliptic operators in divergence form and collect important properties of them. Throughout the section, let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) and μ ∈ L ∞ ( R d ; C d × d ) \mu\in L^\infty(\mathbb{R}^d;\mathbb{C}^{d\times d}) μ ∈ L ∞ ( R d ; C d × d ) satisfy Assumption 1.1. Moreover, we set
ω 0 : = arctan ( μ ∙ μ ∙ ) ∈ ( 0 , π 2 ) . (9) \omega_0 := \arctan\left(\frac{\mu^\bullet}{\mu_\bullet}\right) \in\left(0,\frac{\pi}{2}\right). \tag*{(9)} ω 0 := arctan ( μ ∙ μ ∙ ) ∈ ( 0 , 2 π ) . (9)
Definition 3.1. We define the sesquilinear form
a : S α , 2 ( R d ) × S α , 2 ( R d ) → C , a ( u , v ) : = ∫ R d μ ∇ α u ⋅ ∇ α v ‾ d x \mathbf{a}: S^{\alpha,2}(\mathbb{R}^d) \times S^{\alpha,2}(\mathbb{R}^d) \to\mathbb{C}, \qquad\mathbf{a}(u,v) := \int_{\mathbb{R}^d} \mu\nabla^\alpha u \cdot\overline{\nabla^\alpha v}\, dx a : S α , 2 ( R d ) × S α , 2 ( R d ) → C , a ( u , v ) := ∫ R d μ ∇ α u ⋅ ∇ α v d x and the associated weak fractional operator L : S α , 2 ( R d ) → S − α , 2 ( R d ) \mathcal{L}: S^{\alpha,2}(\mathbb{R}^d) \to S^{-\alpha,2}(\mathbb{R}^d) L : S α , 2 ( R d ) → S − α , 2 ( R d ) by
⟨ L u , v ⟩ : = a ( u , v ) ( u , v ∈ S α , 2 ( R d ) ) . \langle\mathcal{L}u,v\rangle:= \mathbf{a}(u,v) \qquad(u,v \in S^{\alpha,2}(\mathbb{R}^d)). ⟨ L u , v ⟩ := a ( u , v ) ( u , v ∈ S α , 2 ( R d )) . Subsequently, we define the fractional operator L L L as the maximal restriction of L \mathcal{L} L to L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) , that is
D ( L ) = { u ∈ S α , 2 ( R d ) : L u ∈ L 2 ( R d ) } with L u = L u . D(L) = \{u \in S^{\alpha,2}(\mathbb{R}^d) : \mathcal{L}u \in L^2(\mathbb{R}^d)\} \quad\text{with} \quad Lu = \mathcal{L}u. D ( L ) = { u ∈ S α , 2 ( R d ) : L u ∈ L 2 ( R d )} with Lu = L u .
For later reasoning it will be important to deduce a representation of L \mathcal{L} L as a composition of the fractional gradient ∇ α \nabla^\alpha ∇ α , the coefficients μ \mu μ and the very weak fractional divergence div α \operatorname{div}^\alpha div α from Definition 2.8.
Lemma 3.3. For all u ∈ S α , 2 ( R d ) u \in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) , we have L u = − div α ( μ ∇ α u ) Lu=-\operatorname{div}^\alpha(\mu\nabla^\alpha u) Lu = − div α ( μ ∇ α u ) .
Under Assumption 1.1, the sesquilinear form a \mathbf{a} a is bounded and satisfies
a ( u , u ) ∈ Σ ω 0 and Re a ( u , u ) ≥ μ ∙ ∥ ∇ α u ∥ L 2 2 \mathbf{a}(u,u) \in\Sigma_{\omega_0} \quad\text{and} \quad\operatorname{Re}\mathbf{a}(u,u) \ge\mu_\bullet\lVert\nabla^\alpha u\rVert_{L^2}^2 a ( u , u ) ∈ Σ ω 0 and Re a ( u , u ) ≥ μ ∙ ∥ ∇ α u ∥ L 2 2
for all u ∈ S α , 2 ( R d ) u \in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) . Hence, classical form theory yields that L L L is maximal accretive and sectorial of angle ω 0 \omega_0 ω 0 on L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) . In particular, L L L possesses a unique maximal accretive square root L 1 2 L^{\frac{1}{2}} L 2 1 , which is the operator appearing in Theorem 1.2. We refer to [39 , 47 ] for more details. Finally, we have the following uniform resolvent bounds.
Proposition 3.4. We have σ ( L ) ∪ σ ( L ‾ ) ⊂ Σ ω 0 ‾ \sigma(L) \cup\sigma(\overline{L}) \subset\overline{\Sigma_{\omega_0}} σ ( L ) ∪ σ ( L ) ⊂ Σ ω 0 and for all θ ∈ [ 0 , π − ω 0 ) \theta\in[0,\pi-\omega_0) θ ∈ [ 0 , π − ω 0 ) there exists C > 0 C>0 C > 0 such that for all λ ∈ Σ θ \lambda\in\Sigma_\theta λ ∈ Σ θ and all f ∈ L 2 ( R d ) f \in L^2(\mathbb{R}^d) f ∈ L 2 ( R d ) we have
∣ λ ∣ ∥ ( λ + L ) − 1 f ∥ L 2 + ∣ λ ∣ 1 2 ∥ ∇ α ( λ + L ) − 1 f ∥ L 2 ≤ C ∥ f ∥ L 2 . |\lambda|\lVert(\lambda+L)^{-1}f\rVert_{L^2} + |\lambda|^{\frac{1}{2}}\lVert\nabla^\alpha(\lambda+L)^{-1}f\rVert_{L^2} \le C\lVert f\rVert_{L^2}. ∣ λ ∣ ∥( λ + L ) − 1 f ∥ L 2 + ∣ λ ∣ 2 1 ∥ ∇ α ( λ + L ) − 1 f ∥ L 2 ≤ C ∥ f ∥ L 2 .
Moreover, there exists C > 0 C>0 C > 0 such that for all λ ∈ Σ θ \lambda\in\Sigma_\theta λ ∈ Σ θ and all F ∈ L 2 ( R d ; C d ) F\in L^2(\mathbb{R}^d;\mathbb{C}^d) F ∈ L 2 ( R d ; C d ) we have
∣ λ ∣ 1 2 ∥ ( λ + L ) − 1 div α F ∥ L 2 + ∥ ∇ α ( λ + L ) − 1 div α F ∥ L 2 ≤ C ∥ F ∥ L 2 . |\lambda|^{\frac{1}{2}}\|(\lambda+\mathcal{L})^{-1}\operatorname{div}^{\alpha}F\|_{L^2}+\|\nabla^{\alpha}(\lambda+\mathcal{L})^{-1}\operatorname{div}^{\alpha}F\|_{L^2}\leq C\|F\|_{L^2}. ∣ λ ∣ 2 1 ∥ ( λ + L ) − 1 div α F ∥ L 2 + ∥ ∇ α ( λ + L ) − 1 div α F ∥ L 2 ≤ C ∥ F ∥ L 2 .
The constant C C C depends only on d d d , θ \theta θ and μ ∙ , μ ∙ \mu_{\bullet},\mu^{\bullet} μ ∙ , μ ∙ .
The proof is a standard application of the Lax–Milgram lemma, and will therefore be omitted.
Finally, for x ∈ R d x\in\mathbb{R}^d x ∈ R d let μ ∗ ( x ) \mu^*(x) μ ∗ ( x ) denote the adjoint of μ ( x ) \mu(x) μ ( x ) . Then, for all u , v ∈ S α , 2 ( R d ) u,v\in S^{\alpha,2}(\mathbb{R}^d) u , v ∈ S α , 2 ( R d ) we have
∫ R d μ ∗ ∇ α u ⋅ ∇ α v ‾ d x = ∫ R d μ ∇ α v ⋅ ∇ α u ‾ d x . \int_{\mathbb{R}^d}\mu^*\nabla^\alpha u\cdot\overline{\nabla^\alpha v}\,dx=\int_{\mathbb{R}^d}\mu\nabla^\alpha v\cdot\overline{\nabla^\alpha u}\,dx. ∫ R d μ ∗ ∇ α u ⋅ ∇ α v d x = ∫ R d μ ∇ α v ⋅ ∇ α u d x .
so that μ ∗ \mu^* μ ∗ again satisfies Assumption 1.1 with the same constants μ ∙ , μ ∙ \mu_{\bullet},\mu^{\bullet} μ ∙ , μ ∙ . Thus, introducing the adjoint form
a ∗ ( u , v ) : = a ( v , u ) ‾ = ∫ R d μ ∗ ∇ α u ⋅ ∇ α v ‾ d x ( u , v ∈ S α , 2 ( R d ) ) , \mathfrak{a}^*(u,v):=\overline{\mathfrak{a}(v,u)}=\int_{\mathbb{R}^d}\mu^*\nabla^\alpha u\cdot\overline{\nabla^\alpha v}\,dx\qquad(u,v\in S^{\alpha,2}(\mathbb{R}^d)), a ∗ ( u , v ) := a ( v , u ) = ∫ R d μ ∗ ∇ α u ⋅ ∇ α v d x ( u , v ∈ S α , 2 ( R d )) ,
we denote by L ∗ \mathcal{L}^* L ∗ and L ∗ L^* L ∗ the (weak) fractional operators associated with a ∗ \mathfrak{a}^* a ∗ . These are the adjoints of L \mathcal{L} L and L L L , have the same sectoriality angle ω 0 \omega_0 ω 0 , and every result of this paper applies verbatim to them. We use this in duality arguments without further mention.
Caccioppoli inequality and off-diagonal estimates
In this subsection, we follow the approach for the generalized Stokes operator from [41 , 64 ] to first derive a nonlocal Caccioppoli inequality (Theorem 3.6) and then deduce off-diagonal estimates by an iteration scheme (Theorem 3.7). We start with the following lemma.
Lemma 3.5. Let α 0 ∈ ( 0 , 1 ) \alpha_0\in(0,1) α 0 ∈ ( 0 , 1 ) . For all θ ∈ [ 0 , π − ω 0 ) \theta\in[0,\pi-\omega_0) θ ∈ [ 0 , π − ω 0 ) and all ν ∈ ( 0 , d + α 0 ) \nu\in(0,d+\alpha_0) ν ∈ ( 0 , d + α 0 ) there exists a constant C > 0 C>0 C > 0 such that for all balls B = B ( x 0 , r ) B=B(x_0,r) B = B ( x 0 , r ) , δ ∈ ( 0 , 1 ] \delta\in(0,1] δ ∈ ( 0 , 1 ] , α ∈ [ α 0 , 1 ) \alpha\in[\alpha_0,1) α ∈ [ α 0 , 1 ) , λ ∈ Σ θ \lambda\in\Sigma_\theta λ ∈ Σ θ , f ∈ L 2 ( R d ) f\in L^2(\mathbb{R}^d) f ∈ L 2 ( R d ) and F ∈ L 2 ( R d ; C d ) F\in L^2(\mathbb{R}^d;\mathbb{C}^d) F ∈ L 2 ( R d ; C d ) the solution u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) to
λ u − div α ( μ ∇ α u ) = f + div α ( F ) in R d (10) \lambda u-\operatorname{div}^{\alpha}(\mu\nabla^\alpha u)=f+\operatorname{div}^{\alpha}(F)\quad\text{in }\mathbb{R}^d \tag*{(10)} λ u − div α ( μ ∇ α u ) = f + div α ( F ) in R d (10) satisfies
∣ λ ∣ ∥ u ∥ L 2 ( B ) 2 + ∥ ∇ α u ∥ L 2 ( B ) 2 ≤ C ∣ λ ∣ ∥ f ∥ L 2 ( 2 B ) 2 |\lambda|\|u\|_{L^2(B)}^2+\|\nabla^\alpha u\|_{L^2(B)}^2\leq\frac{C}{|\lambda|}\|f\|_{L^2(2B)}^2 ∣ λ ∣∥ u ∥ L 2 ( B ) 2 + ∥ ∇ α u ∥ L 2 ( B ) 2 ≤ ∣ λ ∣ C ∥ f ∥ L 2 ( 2 B ) 2 + C ∑ k = 0 ∞ 2 − ν k ( ∥ F ∥ L 2 ( B k + 1 ) 2 + δ ∥ ∇ α u ∥ L 2 ( B k + 1 ) 2 + ∥ u ∥ L 2 ( B k + 1 ) 2 δ r 2 α ) . {}+C\sum_{k=0}^{\infty}2^{-\nu k}\left(\|F\|_{L^2(B_{k+1})}^2+\delta\|\nabla^\alpha u\|_{L^2(B_{k+1})}^2+\frac{\|u\|_{L^2(B_{k+1})}^2}{\delta r^{2\alpha}}\right). + C k = 0 ∑ ∞ 2 − ν k ( ∥ F ∥ L 2 ( B k + 1 ) 2 + δ ∥ ∇ α u ∥ L 2 ( B k + 1 ) 2 + δ r 2 α ∥ u ∥ L 2 ( B k + 1 ) 2 ) . The constant C > 0 C>0 C > 0 depends only on d d d , θ \theta θ , α 0 \alpha_0 α 0 , ν \nu ν and μ ∙ , μ ∙ \mu_{\bullet},\mu^{\bullet} μ ∙ , μ ∙ .
Proof. Fix φ ∈ C c ∞ ( 2 B ) \varphi\in C^\infty_c(2B) φ ∈ C c ∞ ( 2 B ) satisfying 0 ≤ φ ≤ 1 0\leq\varphi\leq1 0 ≤ φ ≤ 1 , φ = 1 \varphi=1 φ = 1 on B B B and r ∥ ∇ φ ∥ L ∞ ≲ 1 r\|\nabla\varphi\|_{L^\infty}\lesssim1 r ∥∇ φ ∥ L ∞ ≲ 1 . By Lemma 2.9, we have u φ 2 ∈ S α , 2 ( R d ) u\varphi^2\in S^{\alpha,2}(\mathbb{R}^d) u φ 2 ∈ S α , 2 ( R d ) and ∇ α ( u φ 2 ) = φ 2 ∇ α u + R \nabla^\alpha(u\varphi^2)=\varphi^2\nabla^\alpha u+R ∇ α ( u φ 2 ) = φ 2 ∇ α u + R , where R = u ∇ α ( φ 2 ) + ∇ N L α ( u , φ 2 ) R=u\nabla^\alpha(\varphi^2)+\nabla^\alpha_{\mathrm{NL}}(u,\varphi^2) R = u ∇ α ( φ 2 ) + ∇ NL α ( u , φ 2 ) . Testing the resolvent equation (3.2) with u φ 2 u\varphi^2 u φ 2 leads to
λ ∫ R d ∣ u φ ∣ 2 d x + ∫ R d μ ∇ α u ⋅ ( ∇ α u ) φ 2 ‾ d x = ∫ R d f u φ 2 d x − ∫ R d F ⋅ ( ∇ α u ) φ 2 ‾ d x \lambda\int_{\mathbb{R}^d}|u\varphi|^2\,dx+\int_{\mathbb{R}^d}\mu\nabla^\alpha u\cdot\overline{(\nabla^\alpha u)\varphi^2}\,dx
=\int_{\mathbb{R}^d}fu\varphi^2\,dx-\int_{\mathbb{R}^d}F\cdot\overline{(\nabla^\alpha u)\varphi^2}\,dx λ ∫ R d ∣ u φ ∣ 2 d x + ∫ R d μ ∇ α u ⋅ ( ∇ α u ) φ 2 d x = ∫ R d f u φ 2 d x − ∫ R d F ⋅ ( ∇ α u ) φ 2 d x
− ∫ R d ( μ ∇ α u + F ) ⋅ R ‾ d x . (11) \qquad-\int_{\mathbb{R}^d}(\mu\nabla^\alpha u+F)\cdot\overline{R}\,dx. \tag*{(11)} − ∫ R d ( μ ∇ α u + F ) ⋅ R d x . (11)
Due to Assumption 1.1, we notice that
∫ R d μ ∇ α u ⋅ ( ∇ α u ) ‾ φ 2 d x ∈ Σ ω 0 . \int_{\mathbb{R}^{d}} \mu\nabla^{\alpha}u \cdot\overline{(\nabla^{\alpha}u)}\varphi^{2}\,\mathrm{d}x \in\Sigma_{\omega_{0}}. ∫ R d μ ∇ α u ⋅ ( ∇ α u ) φ 2 d x ∈ Σ ω 0 .
Since λ ∈ Σ θ \lambda\in\Sigma_{\theta} λ ∈ Σ θ and θ + ω 0 < π \theta+\omega_{0}<\pi θ + ω 0 < π , there exists C > 0 C>0 C > 0 depending only on θ \theta θ , ω 0 \omega_{0} ω 0 and μ ∙ \mu_{\bullet} μ ∙ , such that
∣ λ ∣ ∥ u φ ∥ L 2 2 + ∥ φ ∇ α u ∥ L 2 2 ≤ C ∣ λ ∫ R d ∣ u φ ∣ 2 d x + ∫ R d μ ∇ α u ⋅ ( ∇ α u ) ‾ φ 2 d x ∣ . |\lambda|\lVert u\varphi\rVert_{L^{2}}^{2}+\lVert\varphi\nabla^{\alpha}u\rVert_{L^{2}}^{2}
\leq C\left|\lambda\int_{\mathbb{R}^{d}}|u\varphi|^{2}\,\mathrm{d}x+\int_{\mathbb{R}^{d}}\mu\nabla^{\alpha}u\cdot\overline{(\nabla^{\alpha}u)}\varphi^{2}\,\mathrm{d}x\right|. ∣ λ ∣ ∥ u φ ∥ L 2 2 + ∥ φ ∇ α u ∥ L 2 2 ≤ C λ ∫ R d ∣ u φ ∣ 2 d x + ∫ R d μ ∇ α u ⋅ ( ∇ α u ) φ 2 d x .
Combining this estimate with the identity (3.3) yields the bound
∣ λ ∣ ∥ u φ ∥ L 2 2 + ∥ φ ∇ α u ∥ L 2 2 ≤ C ∫ R d ∣ f ∣ ∣ u φ 2 ∣ d x + C ∫ R d ∣ F ∣ ∣ φ 2 ∇ α u ∣ d x + C ∫ R d ( ∣ μ ∇ α u ∣ + ∣ F ∣ ) ∣ R ∣ d x . |\lambda|\lVert u\varphi\rVert_{L^{2}}^{2}+\lVert\varphi\nabla^{\alpha}u\rVert_{L^{2}}^{2}
\leq C\int_{\mathbb{R}^{d}}|f||u\varphi^{2}|\,\mathrm{d}x+C\int_{\mathbb{R}^{d}}|F||\varphi^{2}\nabla^{\alpha}u|\,\mathrm{d}x
+C\int_{\mathbb{R}^{d}}(|\mu\nabla^{\alpha}u|+|F|)|R|\,\mathrm{d}x. ∣ λ ∣ ∥ u φ ∥ L 2 2 + ∥ φ ∇ α u ∥ L 2 2 ≤ C ∫ R d ∣ f ∣∣ u φ 2 ∣ d x + C ∫ R d ∣ F ∣∣ φ 2 ∇ α u ∣ d x + C ∫ R d ( ∣ μ ∇ α u ∣ + ∣ F ∣ ) ∣ R ∣ d x .
Applying Young’s inequality to the first two terms on the right-hand side implies
C ∫ R d ∣ f ∣ ∣ u φ 2 ∣ d x + C ∫ R d ∣ F ∣ ∣ φ 2 ∇ α u ∣ d x ≤ 2 C 2 ∣ λ ∣ ∫ R d ∣ f φ ∣ 2 d x + ∣ λ ∣ 2 ∫ R d ∣ u φ ∣ 2 d x + 2 C 2 ∫ R d ∣ F φ ∣ 2 d x + 1 2 ∫ R d ∣ ( ∇ α u ) φ ∣ 2 d x . C\int_{\mathbb{R}^{d}}|f||u\varphi^{2}|\,\mathrm{d}x+C\int_{\mathbb{R}^{d}}|F||\varphi^{2}\nabla^{\alpha}u|\,\mathrm{d}x
\leq\frac{2C^{2}}{|\lambda|}\int_{\mathbb{R}^{d}}|f\varphi|^{2}\,\mathrm{d}x+\frac{|\lambda|}{2}\int_{\mathbb{R}^{d}}|u\varphi|^{2}\,\mathrm{d}x
+2C^{2}\int_{\mathbb{R}^{d}}|F\varphi|^{2}\,\mathrm{d}x+\frac{1}{2}\int_{\mathbb{R}^{d}}|(\nabla^{\alpha}u)\varphi|^{2}\,\mathrm{d}x. C ∫ R d ∣ f ∣∣ u φ 2 ∣ d x + C ∫ R d ∣ F ∣∣ φ 2 ∇ α u ∣ d x ≤ ∣ λ ∣ 2 C 2 ∫ R d ∣ f φ ∣ 2 d x + 2 ∣ λ ∣ ∫ R d ∣ u φ ∣ 2 d x + 2 C 2 ∫ R d ∣ F φ ∣ 2 d x + 2 1 ∫ R d ∣ ( ∇ α u ) φ ∣ 2 d x .
Plugging this into the inequality above and rearranging the terms yields
∣ λ ∣ ∥ u φ ∥ L 2 2 + ∥ φ ∇ α u ∥ L 2 2 ≲ 1 ∣ λ ∣ ∥ f φ ∥ L 2 2 + ∥ F φ ∥ L 2 2 + ∫ R d ( ∣ μ ∇ α u ∣ + ∣ F ∣ ) ∣ R ∣ d x ≲ 1 ∣ λ ∣ ∥ f ∥ L 2 ( 2 B ) 2 + ∥ F ∥ L 2 ( 2 B ) 2 + ∑ k = 0 ∞ a k ∥ R ∥ L 2 ( C k ( B ) ) . (12) \begin{aligned}
|\lambda|\lVert u\varphi\rVert_{L^{2}}^{2}+\lVert\varphi\nabla^{\alpha}u\rVert_{L^{2}}^{2}
&\lesssim\frac{1}{|\lambda|}\lVert f\varphi\rVert_{L^{2}}^{2}+\lVert F\varphi\rVert_{L^{2}}^{2}
+\int_{\mathbb{R}^{d}}(|\mu\nabla^{\alpha}u|+|F|)|R|\,\mathrm{d}x \\
&\lesssim\frac{1}{|\lambda|}\lVert f\rVert_{L^{2}(2B)}^{2}+\lVert F\rVert_{L^{2}(2B)}^{2}
+\sum_{k=0}^{\infty}a_{k}\lVert R\rVert_{L^{2}(C_{k}(B))}. \tag*{(12)}
\end{aligned} ∣ λ ∣ ∥ u φ ∥ L 2 2 + ∥ φ ∇ α u ∥ L 2 2 ≲ ∣ λ ∣ 1 ∥ f φ ∥ L 2 2 + ∥ F φ ∥ L 2 2 + ∫ R d ( ∣ μ ∇ α u ∣ + ∣ F ∣ ) ∣ R ∣ d x ≲ ∣ λ ∣ 1 ∥ f ∥ L 2 ( 2 B ) 2 + ∥ F ∥ L 2 ( 2 B ) 2 + k = 0 ∑ ∞ a k ∥ R ∥ L 2 ( C k ( B )) . (12)
where we set a k : = ∥ μ ∇ α u ∥ L 2 ( C k ( B ) ) + ∥ F ∥ L 2 ( C k ( B ) ) a_{k}:=\lVert\mu\nabla^{\alpha}u\rVert_{L^{2}(C_{k}(B))}+\lVert F\rVert_{L^{2}(C_{k}(B))} a k := ∥ μ ∇ α u ∥ L 2 ( C k ( B )) + ∥ F ∥ L 2 ( C k ( B )) for k ≥ 0 k\geq0 k ≥ 0 and used the properties of φ \varphi φ and Hölder’s inequality in the last step. Finally, we estimate the last term on the right-hand side. By Lemma 2.10, we have
∥ R ∥ L 2 ( 2 B ) ≲ 1 r α ∑ m = 0 ∞ 2 − ( d 2 + α ) m ∥ u ∥ L 2 ( C m ( B ) ) , \lVert R\rVert_{L^{2}(2B)}\lesssim\frac{1}{r^{\alpha}}\sum_{m=0}^{\infty}2^{-\left(\frac{d}{2}+\alpha\right)m}\lVert u\rVert_{L^{2}(C_{m}(B))}, ∥ R ∥ L 2 ( 2 B ) ≲ r α 1 m = 0 ∑ ∞ 2 − ( 2 d + α ) m ∥ u ∥ L 2 ( C m ( B )) ,
∥ R ∥ L 2 ( C k ( B ) ) ≲ 2 − ( d + α ) k r α ∥ u ∥ L 2 ( C k ( B ) ) + 2 − ( d 2 + α ) k r α ∥ u ∥ L 2 ( 2 B ) \lVert R\rVert_{L^{2}(C_{k}(B))}\lesssim\frac{2^{-\left(d+\alpha\right)k}}{r^{\alpha}}\lVert u\rVert_{L^{2}(C_{k}(B))}+\frac{2^{-\left(\frac{d}{2}+\alpha\right)k}}{r^{\alpha}}\lVert u\rVert_{L^{2}(2B)} ∥ R ∥ L 2 ( C k ( B )) ≲ r α 2 − ( d + α ) k ∥ u ∥ L 2 ( C k ( B )) + r α 2 − ( 2 d + α ) k ∥ u ∥ L 2 ( 2 B )
for k ≥ 1 k\geq1 k ≥ 1 . Thus, we have
∑ k = 0 ∞ a k ∥ R ∥ L 2 ( C k ( B ) ) ≲ a 0 r α ∑ m = 0 ∞ 2 − ( d 2 + α ) m ∥ u ∥ L 2 ( C m ( B ) ) + ∑ k = 1 ∞ ( 2 − ( d + α ) k a k r α ∥ u ∥ L 2 ( C k ( B ) ) + 2 − ( d 2 + α ) k a k r α ∥ u ∥ L 2 ( 2 B ) ) . \begin{aligned}
\sum_{k=0}^{\infty}a_{k}\lVert R\rVert_{L^{2}(C_{k}(B))}
&\lesssim\frac{a_{0}}{r^{\alpha}}\sum_{m=0}^{\infty}2^{-\left(\frac{d}{2}+\alpha\right)m}\lVert u\rVert_{L^{2}(C_{m}(B))} \\
&\quad+\sum_{k=1}^{\infty}\left(\frac{2^{-(d+\alpha)k}a_{k}}{r^{\alpha}}\lVert u\rVert_{L^{2}(C_{k}(B))}+\frac{2^{-\left(\frac{d}{2}+\alpha\right)k}a_{k}}{r^{\alpha}}\lVert u\rVert_{L^{2}(2B)}\right).
\end{aligned} k = 0 ∑ ∞ a k ∥ R ∥ L 2 ( C k ( B )) ≲ r α a 0 m = 0 ∑ ∞ 2 − ( 2 d + α ) m ∥ u ∥ L 2 ( C m ( B )) + k = 1 ∑ ∞ ( r α 2 − ( d + α ) k a k ∥ u ∥ L 2 ( C k ( B )) + r α 2 − ( 2 d + α ) k a k ∥ u ∥ L 2 ( 2 B ) ) .
For every fixed δ ∈ ( 0 , 1 ] \delta\in(0,1] δ ∈ ( 0 , 1 ] , we apply Young’s inequality on the terms on the right-hand side to obtain
∑ k = 0 ∞ a k ∥ R ∥ L 2 ( C k ( B ) ) ≲ ∑ m = 0 ∞ ( δ 2 ( ν − d − 2 α ) m a 0 2 + 2 − ν m δ r 2 α ∥ u ∥ L 2 ( C m ( B ) ) 2 ) + ∑ k = 1 ∞ ( δ 2 − ( d + α ) k a k 2 + 2 − ( d + α ) k δ r 2 α ∥ u ∥ L 2 ( C k ( B ) ) 2 + δ 2 − ν k a k 2 + 2 ( ν − d − 2 α ) k δ r 2 α ∥ u ∥ L 2 ( 2 B ) 2 ) . \sum_{k=0}^{\infty} a_k\lVert R\rVert_{L^2(C_k(B))}
\lesssim\sum_{m=0}^{\infty}\left(\delta2^{(\nu-d-2\alpha)m}a_0^2+\frac{2^{-\nu m}}{\delta r^{2\alpha}}\lVert u\rVert_{L^2(C_m(B))}^2\right)
+\sum_{k=1}^{\infty}\left(\delta2^{-(d+\alpha)k}a_k^2+\frac{2^{-(d+\alpha)k}}{\delta r^{2\alpha}}\lVert u\rVert_{L^2(C_k(B))}^2+\delta2^{-\nu k}a_k^2+\frac{2^{(\nu-d-2\alpha)k}}{\delta r^{2\alpha}}\lVert u\rVert_{L^2(2B)}^2\right). k = 0 ∑ ∞ a k ∥ R ∥ L 2 ( C k ( B )) ≲ m = 0 ∑ ∞ ( δ 2 ( ν − d − 2 α ) m a 0 2 + δ r 2 α 2 − ν m ∥ u ∥ L 2 ( C m ( B )) 2 ) + k = 1 ∑ ∞ ( δ 2 − ( d + α ) k a k 2 + δ r 2 α 2 − ( d + α ) k ∥ u ∥ L 2 ( C k ( B )) 2 + δ 2 − ν k a k 2 + δ r 2 α 2 ( ν − d − 2 α ) k ∥ u ∥ L 2 ( 2 B ) 2 ) .
Since ν < d + α 0 \nu<d+\alpha_0 ν < d + α 0 , we have ∑ k = 0 ∞ 2 ( ν − d − 2 α ) k ≤ ∑ k = 0 ∞ 2 ( ν − d − 2 α 0 ) k < ∞ \sum_{k=0}^{\infty}2^{(\nu-d-2\alpha)k}\leq\sum_{k=0}^{\infty}2^{(\nu-d-2\alpha_0)k}<\infty ∑ k = 0 ∞ 2 ( ν − d − 2 α ) k ≤ ∑ k = 0 ∞ 2 ( ν − d − 2 α 0 ) k < ∞ . Hence, we estimate
∑ k = 0 ∞ a k ∥ R ∥ L 2 ( C k ( B ) ) ≲ δ a 0 2 + ∑ m = 0 ∞ 2 − ν m δ r 2 α ∥ u ∥ L 2 ( C m ( B ) ) 2 + ∑ k = 1 ∞ ( δ 2 − ν k a k 2 + 2 − ( d + α ) k δ r 2 α ∥ u ∥ L 2 ( C k ( B ) ) 2 ) + ∥ u ∥ L 2 ( 2 B ) 2 δ r 2 α \sum_{k=0}^{\infty}a_k\lVert R\rVert_{L^2(C_k(B))}
\lesssim\delta a_0^2+\sum_{m=0}^{\infty}\frac{2^{-\nu m}}{\delta r^{2\alpha}}\lVert u\rVert_{L^2(C_m(B))}^2+\sum_{k=1}^{\infty}\left(\delta2^{-\nu k}a_k^2+\frac{2^{-(d+\alpha)k}}{\delta r^{2\alpha}}\lVert u\rVert_{L^2(C_k(B))}^2\right)+\frac{\lVert u\rVert_{L^2(2B)}^2}{\delta r^{2\alpha}} k = 0 ∑ ∞ a k ∥ R ∥ L 2 ( C k ( B )) ≲ δ a 0 2 + m = 0 ∑ ∞ δ r 2 α 2 − ν m ∥ u ∥ L 2 ( C m ( B )) 2 + k = 1 ∑ ∞ ( δ 2 − ν k a k 2 + δ r 2 α 2 − ( d + α ) k ∥ u ∥ L 2 ( C k ( B )) 2 ) + δ r 2 α ∥ u ∥ L 2 ( 2 B ) 2
≲ δ ∑ k = 0 ∞ 2 − ν k a k 2 + ∑ k = 0 ∞ 2 − ν k δ r 2 α ∥ u ∥ L 2 ( C k ( B ) ) 2 . \lesssim\delta\sum_{k=0}^{\infty}2^{-\nu k}a_k^2+\sum_{k=0}^{\infty}\frac{2^{-\nu k}}{\delta r^{2\alpha}}\lVert u\rVert_{L^2(C_k(B))}^2. ≲ δ k = 0 ∑ ∞ 2 − ν k a k 2 + k = 0 ∑ ∞ δ r 2 α 2 − ν k ∥ u ∥ L 2 ( C k ( B )) 2 .
Inserting this estimate into (12) and using φ = 1 \varphi=1 φ = 1 on B B B and C k ( B ) ⊂ B k + 1 C_k(B)\subset B_{k+1} C k ( B ) ⊂ B k + 1 proves the claim. □
Theorem 3.6 (Nonlocal Caccioppoli inequality). Let α 0 ∈ ( 0 , 1 ) \alpha_0\in(0,1) α 0 ∈ ( 0 , 1 ) . For all θ ∈ [ 0 , π − ω 0 ) \theta\in[0,\pi-\omega_0) θ ∈ [ 0 , π − ω 0 ) and all ν ∈ ( 0 , d + α 0 ) \nu\in(0,d+\alpha_0) ν ∈ ( 0 , d + α 0 ) there exists a constant C > 0 C>0 C > 0 such that for all balls B = B ( x 0 , r ) B=B(x_0,r) B = B ( x 0 , r ) , α ∈ [ α 0 , 1 ) \alpha\in[\alpha_0,1) α ∈ [ α 0 , 1 ) , λ ∈ Σ θ \lambda\in\Sigma_\theta λ ∈ Σ θ , f ∈ L 2 ( R d ) f\in L^2(\mathbb{R}^d) f ∈ L 2 ( R d ) and F ∈ L 2 ( R d ; C d ) F\in L^2(\mathbb{R}^d;\mathbb{C}^d) F ∈ L 2 ( R d ; C d ) the solution u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) to (3.2) satisfies
∣ λ ∣ ∑ ℓ = 0 ∞ 2 − ν ℓ ∥ u ∥ L 2 ( B ℓ ) 2 + ∑ ℓ = 0 ∞ 2 − ν ℓ ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 ≤ C ∑ ℓ = 0 ∞ 2 − ν ℓ ( 1 ∣ λ ∣ ∥ f ∥ L 2 ( B ℓ + 1 ) 2 + ∥ F ∥ L 2 ( B ℓ + 1 ) 2 + 1 r 2 α ∥ u ∥ L 2 ( B ℓ + 1 ) 2 ) . |\lambda|\sum_{\ell=0}^{\infty}2^{-\nu\ell}\lVert u\rVert_{L^2(B_\ell)}^2+\sum_{\ell=0}^{\infty}2^{-\nu\ell}\lVert\nabla^\alpha u\rVert_{L^2(B_\ell)}^2
\leq C\sum_{\ell=0}^{\infty}2^{-\nu\ell}\left(\frac{1}{|\lambda|}\lVert f\rVert_{L^2(B_{\ell+1})}^2+\lVert F\rVert_{L^2(B_{\ell+1})}^2+\frac{1}{r^{2\alpha}}\lVert u\rVert_{L^2(B_{\ell+1})}^2\right). ∣ λ ∣ ℓ = 0 ∑ ∞ 2 − ν ℓ ∥ u ∥ L 2 ( B ℓ ) 2 + ℓ = 0 ∑ ∞ 2 − ν ℓ ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 ≤ C ℓ = 0 ∑ ∞ 2 − ν ℓ ( ∣ λ ∣ 1 ∥ f ∥ L 2 ( B ℓ + 1 ) 2 + ∥ F ∥ L 2 ( B ℓ + 1 ) 2 + r 2 α 1 ∥ u ∥ L 2 ( B ℓ + 1 ) 2 ) . The constant C > 0 C>0 C > 0 depends only on d d d , α 0 \alpha_0 α 0 , θ \theta θ , ν \nu ν and μ ∙ , μ ∙ \mu_\bullet,\mu^\bullet μ ∙ , μ ∙ .
Proof. Let ν ′ : = d + α 0 + ν 2 ∈ ( ν , d + α 0 ) \nu' := \frac{d+\alpha_0+\nu}{2}\in(\nu,d+\alpha_0) ν ′ := 2 d + α 0 + ν ∈ ( ν , d + α 0 ) . We apply Lemma 3.5 with ν ′ \nu' ν ′ to the balls B ℓ B_\ell B ℓ , ℓ ∈ N 0 \ell\in\mathbb{N}_0 ℓ ∈ N 0 , whose radii are 2 ℓ r ≥ r 2^\ell r\geq r 2 ℓ r ≥ r and whose dilation satisfy ( B ℓ ) k + 1 = B k + ℓ + 1 (B_\ell)_{k+1}=B_{k+\ell+1} ( B ℓ ) k + 1 = B k + ℓ + 1 to obtain
∣ λ ∣ ∥ u ∥ L 2 ( B ℓ ) 2 + ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 ≤ C ∣ λ ∣ ∥ f ∥ L 2 ( B ℓ + 1 ) 2 + C ∑ k = 0 ∞ 2 − ν ′ k ( ∥ F ∥ L 2 ( B k + ℓ + 1 ) 2 + δ ∥ ∇ α u ∥ L 2 ( B k + ℓ + 1 ) 2 + ∥ u ∥ L 2 ( B k + ℓ + 1 ) 2 δ r 2 α ) . |\lambda|\lVert u\rVert_{L^2(B_\ell)}^2+\lVert\nabla^\alpha u\rVert_{L^2(B_\ell)}^2\leq\frac{C}{|\lambda|}\lVert f\rVert_{L^2(B_{\ell+1})}^2
+C\sum_{k=0}^{\infty}2^{-\nu'k}\left(\lVert F\rVert_{L^2(B_{k+\ell+1})}^2+\delta\lVert\nabla^\alpha u\rVert_{L^2(B_{k+\ell+1})}^2+\frac{\lVert u\rVert_{L^2(B_{k+\ell+1})}^2}{\delta r^{2\alpha}}\right). ∣ λ ∣ ∥ u ∥ L 2 ( B ℓ ) 2 + ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 ≤ ∣ λ ∣ C ∥ f ∥ L 2 ( B ℓ + 1 ) 2 + C k = 0 ∑ ∞ 2 − ν ′ k ( ∥ F ∥ L 2 ( B k + ℓ + 1 ) 2 + δ ∥ ∇ α u ∥ L 2 ( B k + ℓ + 1 ) 2 + δ r 2 α ∥ u ∥ L 2 ( B k + ℓ + 1 ) 2 ) . After the index shift n = k + ℓ n=k+\ell n = k + ℓ we obtain
∣ λ ∣ ∥ u ∥ L 2 ( B ℓ ) 2 + ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 ≤ C ∣ λ ∣ ∥ f ∥ L 2 ( B ℓ + 1 ) 2 |\lambda|\lVert u\rVert_{L^2(B_\ell)}^2+\lVert\nabla^\alpha u\rVert_{L^2(B_\ell)}^2\leq\frac{C}{|\lambda|}\lVert f\rVert_{L^2(B_{\ell+1})}^2 ∣ λ ∣ ∥ u ∥ L 2 ( B ℓ ) 2 + ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 ≤ ∣ λ ∣ C ∥ f ∥ L 2 ( B ℓ + 1 ) 2
+ C ∑ n = ℓ ∞ 2 − ν ′ ( n − ℓ ) ( ∥ F ∥ L 2 ( B n + 1 ) 2 + δ ∥ ∇ α u ∥ L 2 ( B n + 1 ) 2 + ∥ u ∥ L 2 ( B n + 1 ) 2 δ r 2 α ) . + C \sum_{n=\ell}^{\infty} 2^{-\nu'(n-\ell)} \left( \|F\|_{L^2(B_{n+1})}^2 + \delta\|\nabla^\alpha u\|_{L^2(B_{n+1})}^2 + \frac{\|u\|_{L^2(B_{n+1})}^2}{\delta r^{2\alpha}} \right). + C n = ℓ ∑ ∞ 2 − ν ′ ( n − ℓ ) ( ∥ F ∥ L 2 ( B n + 1 ) 2 + δ ∥ ∇ α u ∥ L 2 ( B n + 1 ) 2 + δ r 2 α ∥ u ∥ L 2 ( B n + 1 ) 2 ) .
We multiply both sides by 2 − ν ℓ 2^{-\nu\ell} 2 − ν ℓ and sum over all ℓ ∈ N 0 \ell\in\mathbb{N}_0 ℓ ∈ N 0 . Since, for non-negative ( b n ) n (b_n)_n ( b n ) n ,
∑ ℓ = 0 ∞ 2 − ν ℓ ∑ n = ℓ ∞ 2 − ν ′ ( n − ℓ ) b n = ∑ n = 0 ∞ 2 − ν ′ n b n ∑ ℓ = 0 n 2 ( ν ′ − ν ) ℓ ≤ C 1 ∑ n = 0 ∞ 2 − ν n b n \sum_{\ell=0}^{\infty} 2^{-\nu\ell} \sum_{n=\ell}^{\infty} 2^{-\nu'(n-\ell)} b_n
= \sum_{n=0}^{\infty} 2^{-\nu'n} b_n \sum_{\ell=0}^{n} 2^{(\nu'-\nu)\ell}
\le C_1 \sum_{n=0}^{\infty} 2^{-\nu n} b_n ℓ = 0 ∑ ∞ 2 − ν ℓ n = ℓ ∑ ∞ 2 − ν ′ ( n − ℓ ) b n = n = 0 ∑ ∞ 2 − ν ′ n b n ℓ = 0 ∑ n 2 ( ν ′ − ν ) ℓ ≤ C 1 n = 0 ∑ ∞ 2 − ν n b n
with C 1 = C 1 ( d , α 0 , ν ) > 0 C_1=C_1(d,\alpha_0,\nu)>0 C 1 = C 1 ( d , α 0 , ν ) > 0 , we arrive at
∣ λ ∣ ∑ ℓ = 0 ∞ 2 − ν ℓ ∥ u ∥ L 2 ( B ℓ ) 2 + ∑ ℓ = 0 ∞ 2 − ν ℓ ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 ≤ C ∣ λ ∣ ∑ ℓ = 0 ∞ 2 − ν ℓ ∥ f ∥ L 2 ( B ℓ + 1 ) 2 + C C 1 ∑ n = 0 ∞ 2 − ν n ( ∥ F ∥ L 2 ( B n + 1 ) 2 + δ ∥ ∇ α u ∥ L 2 ( B n + 1 ) 2 + ∥ u ∥ L 2 ( B n + 1 ) 2 δ r 2 α ) . \begin{aligned}
|\lambda| \sum_{\ell=0}^{\infty} 2^{-\nu\ell}\|u\|_{L^2(B_\ell)}^2
+\sum_{\ell=0}^{\infty} 2^{-\nu\ell}\|\nabla^\alpha u\|_{L^2(B_\ell)}^2
\le\frac{C}{|\lambda|}\sum_{\ell=0}^{\infty}2^{-\nu\ell}\|f\|_{L^2(B_{\ell+1})}^2 \\
&+CC_1\sum_{n=0}^{\infty}2^{-\nu n}\left(\|F\|_{L^2(B_{n+1})}^2+\delta\|\nabla^\alpha u\|_{L^2(B_{n+1})}^2+\frac{\|u\|_{L^2(B_{n+1})}^2}{\delta r^{2\alpha}}\right).
\end{aligned} ∣ λ ∣ ℓ = 0 ∑ ∞ 2 − ν ℓ ∥ u ∥ L 2 ( B ℓ ) 2 + ℓ = 0 ∑ ∞ 2 − ν ℓ ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 ≤ ∣ λ ∣ C ℓ = 0 ∑ ∞ 2 − ν ℓ ∥ f ∥ L 2 ( B ℓ + 1 ) 2 + C C 1 n = 0 ∑ ∞ 2 − ν n ( ∥ F ∥ L 2 ( B n + 1 ) 2 + δ ∥ ∇ α u ∥ L 2 ( B n + 1 ) 2 + δ r 2 α ∥ u ∥ L 2 ( B n + 1 ) 2 ) .
Using the identity
C C 1 δ ∑ n = 0 ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n + 1 ) 2 = 2 ν C C 1 δ ∑ ℓ = 1 ∞ 2 − ν ℓ ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 CC_1\delta\sum_{n=0}^{\infty}2^{-\nu n}\|\nabla^\alpha u\|_{L^2(B_{n+1})}^2
=2^\nu CC_1\delta\sum_{\ell=1}^{\infty}2^{-\nu\ell}\|\nabla^\alpha u\|_{L^2(B_\ell)}^2 C C 1 δ n = 0 ∑ ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n + 1 ) 2 = 2 ν C C 1 δ ℓ = 1 ∑ ∞ 2 − ν ℓ ∥ ∇ α u ∥ L 2 ( B ℓ ) 2
on the right-hand side, we can absorb the term with the fractional gradients into the left-hand side by choosing δ = min { ( 2 ν + 1 C C 1 ) − 1 , 2 − 1 } \delta=\min\{(2^{\nu+1}CC_1)^{-1},2^{-1}\} δ = min {( 2 ν + 1 C C 1 ) − 1 , 2 − 1 } to conclude
∣ λ ∣ ∑ ℓ = 0 ∞ 2 − ν ℓ ∥ u ∥ L 2 ( B ℓ ) 2 + ∑ ℓ = 0 ∞ 2 − ν ℓ ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 ≤ C ∑ ℓ = 0 ∞ 2 − ν ℓ ( 1 ∣ λ ∣ ∥ f ∥ L 2 ( B ℓ + 1 ) 2 + ∥ F ∥ L 2 ( B ℓ + 1 ) 2 + 1 r 2 α ∥ u ∥ L 2 ( B ℓ + 1 ) 2 ) \begin{aligned}
|\lambda|\sum_{\ell=0}^{\infty}2^{-\nu\ell}\|u\|_{L^2(B_\ell)}^2
+\sum_{\ell=0}^{\infty}2^{-\nu\ell}\|\nabla^\alpha u\|_{L^2(B_\ell)}^2 \\
&\le C\sum_{\ell=0}^{\infty}2^{-\nu\ell}\left(\frac{1}{|\lambda|}\|f\|_{L^2(B_{\ell+1})}^2+\|F\|_{L^2(B_{\ell+1})}^2+\frac{1}{r^{2\alpha}}\|u\|_{L^2(B_{\ell+1})}^2\right)
\end{aligned} ∣ λ ∣ ℓ = 0 ∑ ∞ 2 − ν ℓ ∥ u ∥ L 2 ( B ℓ ) 2 + ℓ = 0 ∑ ∞ 2 − ν ℓ ∥ ∇ α u ∥ L 2 ( B ℓ ) 2 ≤ C ℓ = 0 ∑ ∞ 2 − ν ℓ ( ∣ λ ∣ 1 ∥ f ∥ L 2 ( B ℓ + 1 ) 2 + ∥ F ∥ L 2 ( B ℓ + 1 ) 2 + r 2 α 1 ∥ u ∥ L 2 ( B ℓ + 1 ) 2 )
for some constant C > 0 C>0 C > 0 depending only on d , α 0 , θ , ν d,\alpha_0,\theta,\nu d , α 0 , θ , ν and μ ∙ , μ ∙ \mu_\bullet,\mu^\bullet μ ∙ , μ ∙ .
Theorem 3.7 (L2 off-diagonal estimates). Let α 0 ∈ ( 0 , 1 ) \alpha_0\in(0,1) α 0 ∈ ( 0 , 1 ) . For all θ ∈ [ 0 , π − ω 0 ) \theta\in[0,\pi-\omega_0) θ ∈ [ 0 , π − ω 0 ) and all ν ∈ ( 0 , d + α 0 ) \nu\in(0,d+\alpha_0) ν ∈ ( 0 , d + α 0 ) there exists a constant C > 0 C>0 C > 0 such that for all balls B = B ( x 0 , r ) B=B(x_0,r) B = B ( x 0 , r ) , α ∈ [ α 0 , 1 ) \alpha\in[\alpha_0,1) α ∈ [ α 0 , 1 ) , λ ∈ Σ θ \lambda\in\Sigma_\theta λ ∈ Σ θ , f ∈ L 2 ( R d ) f\in L^2(\mathbb{R}^d) f ∈ L 2 ( R d ) and F ∈ L 2 ( R d ; C d ) F\in L^2(\mathbb{R}^d;\mathbb{C}^d) F ∈ L 2 ( R d ; C d ) the solution u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) to (3.2) satisfies
∑ n = 0 ∞ 2 − ν n ∥ u ∥ L 2 ( B n ) 2 + 1 ∣ λ ∣ ∑ n = 0 ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n ) 2 \sum_{n=0}^{\infty}2^{-\nu n}\|u\|_{L^2(B_n)}^2
+\frac{1}{|\lambda|}\sum_{n=0}^{\infty}2^{-\nu n}\|\nabla^\alpha u\|_{L^2(B_n)}^2 n = 0 ∑ ∞ 2 − ν n ∥ u ∥ L 2 ( B n ) 2 + ∣ λ ∣ 1 n = 0 ∑ ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n ) 2 ≤ C ∣ λ ∣ ∑ n = 0 ∞ ∑ k = 0 n ( ∏ s = 0 k − 1 C ∣ λ ∣ ( 2 s r ) 2 α ) 2 − ν ( n − k ) ( 1 ∣ λ ∣ ∥ f ∥ L 2 ( B n + 1 ) 2 + ∥ F ∥ L 2 ( B n + 1 ) 2 ) . \le\frac{C}{|\lambda|}\sum_{n=0}^{\infty}\sum_{k=0}^{n}\left(\prod_{s=0}^{k-1}\frac{C}{|\lambda|(2^s r)^{2\alpha}}\right)2^{-\nu(n-k)}
\left(\frac{1}{|\lambda|}\|f\|_{L^2(B_{n+1})}^2+\|F\|_{L^2(B_{n+1})}^2\right). ≤ ∣ λ ∣ C n = 0 ∑ ∞ k = 0 ∑ n ( s = 0 ∏ k − 1 ∣ λ ∣ ( 2 s r ) 2 α C ) 2 − ν ( n − k ) ( ∣ λ ∣ 1 ∥ f ∥ L 2 ( B n + 1 ) 2 + ∥ F ∥ L 2 ( B n + 1 ) 2 ) . The constant C > 0 C>0 C > 0 depends only on d , α 0 , θ , ν d,\alpha_0,\theta,\nu d , α 0 , θ , ν and μ ∙ , μ ∙ \mu_\bullet,\mu^\bullet μ ∙ , μ ∙ .
Proof. For r > 0 r>0 r > 0 and B = B ( x 0 , r ) B=B(x_0,r) B = B ( x 0 , r ) set
S ( r ) = ∑ ℓ = 0 ∞ 2 − ν ℓ ∥ u ∥ L 2 ( B ℓ ) 2 and T ( r ) = ∑ ℓ = 0 ∞ 2 − ν ℓ ( 1 ∣ λ ∣ ∥ f ∥ L 2 ( B ℓ + 1 ) 2 + ∥ F ∥ L 2 ( B ℓ + 1 ) 2 ) . (13) S(r)=\sum_{\ell=0}^{\infty}2^{-\nu\ell}\|u\|_{L^2(B_\ell)}^2
\qquad\text{and}\qquad
T(r)=\sum_{\ell=0}^{\infty}2^{-\nu\ell}\left(\frac{1}{|\lambda|}\|f\|_{L^2(B_{\ell+1})}^2+\|F\|_{L^2(B_{\ell+1})}^2\right). \tag*{(13)} S ( r ) = ℓ = 0 ∑ ∞ 2 − ν ℓ ∥ u ∥ L 2 ( B ℓ ) 2 and T ( r ) = ℓ = 0 ∑ ∞ 2 − ν ℓ ( ∣ λ ∣ 1 ∥ f ∥ L 2 ( B ℓ + 1 ) 2 + ∥ F ∥ L 2 ( B ℓ + 1 ) 2 ) . (13)
By Theorem 3.6 , we have
S ( r ) + 1 ∣ λ ∣ ∑ n = 0 ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n ) 2 ≤ C ∣ λ ∣ T ( r ) + C ∣ λ ∣ r 2 α S ( 2 r ) . S(r)+\frac{1}{|\lambda|}\sum_{n=0}^{\infty}2^{-\nu n}\lVert\nabla^\alpha u\rVert_{L^2(B_n)}^2\leq\frac{C}{|\lambda|}T(r)+\frac{C}{|\lambda|r^{2\alpha}}S(2r). S ( r ) + ∣ λ ∣ 1 n = 0 ∑ ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n ) 2 ≤ ∣ λ ∣ C T ( r ) + ∣ λ ∣ r 2 α C S ( 2 r ) .
Applying this inequality successively with growing radii 2 r , 4 r , … , 2 m r 2r,4r,\ldots,2^m r 2 r , 4 r , … , 2 m r for m ∈ N m\in\mathbb{N} m ∈ N to the S S S -terms, we obtain
S ( r ) + 1 ∣ λ ∣ ∑ n = 0 ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n ) 2 ≤ C ∣ λ ∣ ∑ k = 0 m − 1 ( ∏ s = 0 k − 1 C ∣ λ ∣ ( 2 s r ) 2 α ) T ( 2 k r ) + ( ∏ s = 0 m − 1 C ∣ λ ∣ ( 2 s r ) 2 α ) S ( 2 m r ) . \begin{aligned}
S(r)+\frac{1}{|\lambda|}\sum_{n=0}^{\infty}2^{-\nu n}\lVert\nabla^\alpha u\rVert_{L^2(B_n)}^2
&\leq\frac{C}{|\lambda|}\sum_{k=0}^{m-1}\left(\prod_{s=0}^{k-1}\frac{C}{|\lambda|(2^s r)^{2\alpha}}\right)T(2^k r)\\
&\quad+\left(\prod_{s=0}^{m-1}\frac{C}{|\lambda|(2^s r)^{2\alpha}}\right)S(2^m r).
\end{aligned} S ( r ) + ∣ λ ∣ 1 n = 0 ∑ ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n ) 2 ≤ ∣ λ ∣ C k = 0 ∑ m − 1 ( s = 0 ∏ k − 1 ∣ λ ∣ ( 2 s r ) 2 α C ) T ( 2 k r ) + ( s = 0 ∏ m − 1 ∣ λ ∣ ( 2 s r ) 2 α C ) S ( 2 m r ) .
Since the prefactor ∏ s = 0 m − 1 C ∣ λ ∣ ( 2 s r ) 2 α → 0 \prod_{s=0}^{m-1}\frac{C}{|\lambda|(2^s r)^{2\alpha}}\to0 ∏ s = 0 m − 1 ∣ λ ∣ ( 2 s r ) 2 α C → 0 as m → ∞ m\to\infty m → ∞ and S ( 2 m r ) ≲ ∥ u ∥ L 2 2 < ∞ S(2^m r)\lesssim\lVert u\rVert_{L^2}^2<\infty S ( 2 m r ) ≲ ∥ u ∥ L 2 2 < ∞ for all m ∈ N m\in\mathbb{N} m ∈ N , we obtain in the limit
S ( r ) + 1 ∣ λ ∣ ∑ n = 0 ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n ) 2 ≤ C ∣ λ ∣ ∑ k = 0 ∞ ( ∏ s = 0 k − 1 C ∣ λ ∣ ( 2 s r ) 2 α ) T ( 2 k r ) . S(r)+\frac{1}{|\lambda|}\sum_{n=0}^{\infty}2^{-\nu n}\lVert\nabla^\alpha u\rVert_{L^2(B_n)}^2
\leq\frac{C}{|\lambda|}\sum_{k=0}^{\infty}\left(\prod_{s=0}^{k-1}\frac{C}{|\lambda|(2^s r)^{2\alpha}}\right)T(2^k r). S ( r ) + ∣ λ ∣ 1 n = 0 ∑ ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n ) 2 ≤ ∣ λ ∣ C k = 0 ∑ ∞ ( s = 0 ∏ k − 1 ∣ λ ∣ ( 2 s r ) 2 α C ) T ( 2 k r ) .
Finally, by the substitution ℓ + k = n \ell+k=n ℓ + k = n , we observe
T ( 2 k r ) = ∑ n = k ∞ 2 − ν ( n − k ) ( 1 ∣ λ ∣ ∥ f ∥ L 2 ( B n + 1 ) 2 + ∥ F ∥ L 2 ( B n + 1 ) 2 ) . T(2^k r)=\sum_{n=k}^{\infty}2^{-\nu(n-k)}\left(\frac{1}{|\lambda|}\lVert f\rVert_{L^2(B_{n+1})}^2+\lVert F\rVert_{L^2(B_{n+1})}^2\right). T ( 2 k r ) = n = k ∑ ∞ 2 − ν ( n − k ) ( ∣ λ ∣ 1 ∥ f ∥ L 2 ( B n + 1 ) 2 + ∥ F ∥ L 2 ( B n + 1 ) 2 ) .
Inserting this into the inequality above yields
S ( r ) + 1 ∣ λ ∣ ∑ n = 0 ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n ) 2 ≤ C ∣ λ ∣ ∑ k = 0 ∞ ∑ n = k ∞ ( ∏ s = 0 k − 1 C ∣ λ ∣ ( 2 s r ) 2 α ) 2 − ν ( n − k ) × ( 1 ∣ λ ∣ ∥ f ∥ L 2 ( B n + 1 ) 2 + ∥ F ∥ L 2 ( B n + 1 ) 2 ) = C ∣ λ ∣ ∑ n = 0 ∞ ∑ k = 0 n ( ∏ s = 0 k − 1 C ∣ λ ∣ ( 2 s r ) 2 α ) 2 − ν ( n − k ) × ( 1 ∣ λ ∣ ∥ f ∥ L 2 ( B n + 1 ) 2 + ∥ F ∥ L 2 ( B n + 1 ) 2 ) , \begin{aligned}
S(r)+\frac{1}{|\lambda|}\sum_{n=0}^{\infty}2^{-\nu n}\lVert\nabla^\alpha u\rVert_{L^2(B_n)}^2
&\leq\frac{C}{|\lambda|}\sum_{k=0}^{\infty}\sum_{n=k}^{\infty}\left(\prod_{s=0}^{k-1}\frac{C}{|\lambda|(2^s r)^{2\alpha}}\right)2^{-\nu(n-k)}\\
&\qquad\qquad\qquad\qquad\times\left(\frac{1}{|\lambda|}\lVert f\rVert_{L^2(B_{n+1})}^2+\lVert F\rVert_{L^2(B_{n+1})}^2\right)\\
&=\frac{C}{|\lambda|}\sum_{n=0}^{\infty}\sum_{k=0}^{n}\left(\prod_{s=0}^{k-1}\frac{C}{|\lambda|(2^s r)^{2\alpha}}\right)2^{-\nu(n-k)}\\
&\qquad\qquad\qquad\qquad\times\left(\frac{1}{|\lambda|}\lVert f\rVert_{L^2(B_{n+1})}^2+\lVert F\rVert_{L^2(B_{n+1})}^2\right),
\end{aligned} S ( r ) + ∣ λ ∣ 1 n = 0 ∑ ∞ 2 − ν n ∥ ∇ α u ∥ L 2 ( B n ) 2 ≤ ∣ λ ∣ C k = 0 ∑ ∞ n = k ∑ ∞ ( s = 0 ∏ k − 1 ∣ λ ∣ ( 2 s r ) 2 α C ) 2 − ν ( n − k ) × ( ∣ λ ∣ 1 ∥ f ∥ L 2 ( B n + 1 ) 2 + ∥ F ∥ L 2 ( B n + 1 ) 2 ) = ∣ λ ∣ C n = 0 ∑ ∞ k = 0 ∑ n ( s = 0 ∏ k − 1 ∣ λ ∣ ( 2 s r ) 2 α C ) 2 − ν ( n − k ) × ( ∣ λ ∣ 1 ∥ f ∥ L 2 ( B n + 1 ) 2 + ∥ F ∥ L 2 ( B n + 1 ) 2 ) ,
which proves the claim. □
Corollary 3.8. Let α 0 ∈ ( 0 , 1 ) \alpha_0\in(0,1) α 0 ∈ ( 0 , 1 ) . For all ν ∈ ( 0 , d + α 0 ) \nu\in(0,d+\alpha_0) ν ∈ ( 0 , d + α 0 ) there exists a constant C > 0 C>0 C > 0 such that for all t > 0 t>0 t > 0 , α ∈ [ α 0 , 1 ) \alpha\in[\alpha_0,1) α ∈ [ α 0 , 1 ) , f ∈ L 2 ( R d ) f\in L^2(\mathbb{R}^d) f ∈ L 2 ( R d ) , F ∈ L 2 ( R d ; C d ) F\in L^2(\mathbb{R}^d;\mathbb{C}^d) F ∈ L 2 ( R d ; C d ) and balls B = B ( x 0 , r ) B=B(x_0,r) B = B ( x 0 , r ) with r ≥ t r\geq t r ≥ t we have
∥ t α ∇ α ( 1 + t 2 α L ) − 1 f ∥ L 2 ( B ) 2 + ∥ t α ( 1 + t 2 α L ) − 1 div α F ∥ L 2 ( B ) 2 ≤ C ∑ n = 0 ∞ 2 − ν n ( ∥ f ∥ L 2 ( C n ( B ) ) 2 + ∥ F ∥ L 2 ( C n ( B ) ) 2 ) . \lVert t^\alpha\nabla^\alpha(1+t^{2\alpha}\mathcal{L})^{-1}f\rVert_{L^2(B)}^2+\lVert t^\alpha(1+t^{2\alpha}\mathcal{L})^{-1}\operatorname{div}^\alpha F\rVert_{L^2(B)}^2
\leq C\sum_{n=0}^{\infty}2^{-\nu n}\left(\lVert f\rVert_{L^2(C_n(B))}^2+\lVert F\rVert_{L^2(C_n(B))}^2\right). ∥ t α ∇ α ( 1 + t 2 α L ) − 1 f ∥ L 2 ( B ) 2 + ∥ t α ( 1 + t 2 α L ) − 1 div α F ∥ L 2 ( B ) 2 ≤ C n = 0 ∑ ∞ 2 − ν n ( ∥ f ∥ L 2 ( C n ( B )) 2 + ∥ F ∥ L 2 ( C n ( B )) 2 ) . The constant C > 0 C>0 C > 0 depends only on d d d , α 0 \alpha_0 α 0 , ν \nu ν and μ ∙ , μ ∙ \mu_\bullet,\mu^\bullet μ ∙ , μ ∙ .
The proof follows the same line of reasoning as in [41 Cor. 2.6] and will therefore be omitted.
Proof of the main result
This section is devoted to proving our main result, Theorem 1.2 . In particular, we develop a second-order approach, which follows closely the algorithm of [41 ] ; see also [10 ] and [39 ] . Throughout this section, we fix α 0 ∈ ( 0 , 1 ) \alpha_{0} \in(0,1) α 0 ∈ ( 0 , 1 ) , α ∈ [ α 0 , 1 ) \alpha\in[\alpha_{0},1) α ∈ [ α 0 , 1 ) and μ ∈ L ∞ ( R d ; C d × d ) \mu\in L^{\infty}(\mathbb{R}^{d};\mathbb{C}^{d \times d}) μ ∈ L ∞ ( R d ; C d × d ) satisfying Assumption 1.1. Moreover, unless stated otherwise, all implicit constants suppressed by ≲ \lesssim ≲ depend only on d d d , α 0 \alpha_{0} α 0 , μ ∙ \mu_{\bullet} μ ∙ and μ ∙ \mu^{\bullet} μ ∙ . In particular, they are independent of α ∈ [ α 0 , 1 ) \alpha\in[\alpha_{0},1) α ∈ [ α 0 , 1 ) .
Reduction to a square function estimate
It is well known that the Kato property has a deep connection to certain square function estimates. In particular, the Kato property in Theorem 1.2 with constants depending only on d d d , α 0 \alpha_{0} α 0 and μ ∙ , μ ∙ \mu_{\bullet},\mu^{\bullet} μ ∙ , μ ∙ is equivalent to the square function estimates
∫ 0 ∞ ∥ t α L ( 1 + t 2 α L ) − 1 u ∥ L 2 2 d t t ≤ C ∥ ∇ α u ∥ L 2 2 , ∫ 0 ∞ ∥ t α L ∗ ( 1 + t 2 α L ∗ ) − 1 u ∥ L 2 2 d t t ≤ C ∥ ∇ α u ∥ L 2 2 , (14) \begin{aligned}
\int_{0}^{\infty} \left\|t^{\alpha}L(1+t^{2\alpha}L)^{-1}u\right\|_{L^{2}}^{2}\,\frac{\mathrm{d}t}{t} &\leq C\left\|\nabla^{\alpha}u\right\|_{L^{2}}^{2},\\
\int_{0}^{\infty} \left\|t^{\alpha}L^{*}(1+t^{2\alpha}L^{*})^{-1}u\right\|_{L^{2}}^{2}\,\frac{\mathrm{d}t}{t} &\leq C\left\|\nabla^{\alpha}u\right\|_{L^{2}}^{2}, \tag*{(14)}
\end{aligned} ∫ 0 ∞ t α L ( 1 + t 2 α L ) − 1 u L 2 2 t d t ∫ 0 ∞ t α L ∗ ( 1 + t 2 α L ∗ ) − 1 u L 2 2 t d t ≤ C ∥ ∇ α u ∥ L 2 2 , ≤ C ∥ ∇ α u ∥ L 2 2 , (14)
for every u ∈ S α , 2 ( R d ) u \in\mathcal{S}^{\alpha,2}(\mathbb{R}^{d}) u ∈ S α , 2 ( R d ) , where C > 0 C > 0 C > 0 depends only on d d d , α 0 \alpha_{0} α 0 and μ ∙ , μ ∙ \mu_{\bullet},\mu^{\bullet} μ ∙ , μ ∙ ; see [39 Prop. 12.7] with the substitution s = t α s=t^{\alpha} s = t α . Recall that L ∗ L^{*} L ∗ is again a fractional operator associated to the coefficients μ ∗ \mu^{*} μ ∗ that satisfy Assumption 1.1 with the same ellipticity constants, see § 3 . Hence, for Theorem 1.2 , it suffices to show the first square function estimate in (4.1).
As a first step, we rewrite the square function estimate (4.1) in terms of ∇ α u \nabla^{\alpha}u ∇ α u . For this purpose, we use the identity from Lemma 3.3 to obtain
t α L ( 1 + t 2 α L ) − 1 u = t α ( 1 + t 2 α L ) − 1 L u = − t α ( 1 + t 2 α L ) − 1 div α ( μ ∇ α u ) . (15) t^{\alpha}L(1+t^{2\alpha}L)^{-1}u=t^{\alpha}(1+t^{2\alpha}L)^{-1}Lu=-t^{\alpha}(1+t^{2\alpha}L)^{-1}\operatorname{div}^{\alpha}(\mu\nabla^{\alpha}u). \tag*{(15)} t α L ( 1 + t 2 α L ) − 1 u = t α ( 1 + t 2 α L ) − 1 Lu = − t α ( 1 + t 2 α L ) − 1 div α ( μ ∇ α u ) . (15)
Employing the uniform L 2 L^{2} L 2 -bounds from Proposition 3.4 , we can define the right-hand side as a bounded operator on vector-valued functions.
Definition 4.1. For t > 0 t>0 t > 0 we define the operator
Θ t : L 2 ( R d ; C d ) → L 2 ( R d ) , Θ t F : = − t α ( 1 + t 2 α L ) − 1 div α ( μ F ) . \Theta_{t}:L^{2}(\mathbb{R}^{d};\mathbb{C}^{d})\to L^{2}(\mathbb{R}^{d}),\qquad\Theta_{t}F:=-t^{\alpha}(1+t^{2\alpha}L)^{-1}\operatorname{div}^{\alpha}(\mu F). Θ t : L 2 ( R d ; C d ) → L 2 ( R d ) , Θ t F := − t α ( 1 + t 2 α L ) − 1 div α ( μ F ) .
Hence, using (4.2), we see that the square function estimate (4.1) is equivalent to
∫ 0 ∞ ∥ Θ t ( ∇ α u ) ∥ L 2 2 d t t ≲ ∥ ∇ α u ∥ L 2 2 ( u ∈ S α , 2 ( R d ) ) . (16) \int_{0}^{\infty}\left\|\Theta_{t}(\nabla^{\alpha}u)\right\|_{L^{2}}^{2}\,\frac{\mathrm{d}t}{t}\lesssim\left\|\nabla^{\alpha}u\right\|_{L^{2}}^{2}\qquad(u\in\mathcal{S}^{\alpha,2}(\mathbb{R}^{d})). \tag*{(16)} ∫ 0 ∞ ∥ Θ t ( ∇ α u ) ∥ L 2 2 t d t ≲ ∥ ∇ α u ∥ L 2 2 ( u ∈ S α , 2 ( R d )) . (16)
Principal part approximation
The idea of the principal part approximation is to approximate the operator Θ t \Theta_{t} Θ t in an averaged sense in order to reduce the square function estimate (4.3) to a Carleson measure estimate. For this purpose, we next introduce the following dyadic structure: let □ = ⋃ j ∈ Z □ 2 j \Box=\bigcup_{j\in\mathbb{Z}}\Box_{2^{j}} □ = ⋃ j ∈ Z □ 2 j where □ 2 j = { 2 j ( k + ( 0 , 1 ] d ) : k ∈ Z d } \Box_{2^{j}}=\{2^{j}(k+(0,1]^{d}):k\in\mathbb{Z}^{d}\} □ 2 j = { 2 j ( k + ( 0 , 1 ] d ) : k ∈ Z d } . A dyadic cube Q ∈ □ 2 j Q\in\Box_{2^{j}} Q ∈ □ 2 j has side-length ℓ ( Q ) = 2 j \ell(Q)=2^{j} ℓ ( Q ) = 2 j and measure ∣ Q ∣ = 2 j d |Q|=2^{jd} ∣ Q ∣ = 2 j d . Moreover define □ t = □ 2 j \square_t=\square_{2^j} □ t = □ 2 j for 2 j − 1 < t ≤ 2 j 2^{j-1}<t\le2^j 2 j − 1 < t ≤ 2 j . Then the dyadic averaging operator A t A_t A t on L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) is defined via
A_tu(x):=\fint_Q u(y)\,dy
for every x ∈ R d x\in\mathbb{R}^d x ∈ R d and t > 0 t>0 t > 0 , where Q Q Q is the unique dyadic cube belonging to □ t \square_t □ t and containing the point x x x . Note that A t 2 = A t A_t^2=A_t A t 2 = A t . Furthermore, when a scalar operator, such as A t A_t A t , is applied to a vector-valued function, we agree from now on to apply the operator to each component separately.
As a first step of the principal part approximation, we want to extend the domain of Θ t \Theta_t Θ t from Definition 4.1 to bounded vector fields. For this purpose, recall that ( t α ( 1 + t 2 α L ) − 1 div α ) t > 0 (t^\alpha(1+t^{2\alpha}\mathcal{L})^{-1}\operatorname{div}^\alpha)_{t>0} ( t α ( 1 + t 2 α L ) − 1 div α ) t > 0 satisfies the nonlocal off-diagonal bounds from Corollary 3.8 , which also hold for balls replaced by cubes, see Remark 3.9 . Since Θ t \Theta_t Θ t arises from that operator family applied to the bounded multiplication operator μ \mu μ , these estimates are transferred to Θ t \Theta_t Θ t and will be recorded in the next lemma.
Lemma 4.2. For all ν ∈ ( 0 , d + α 0 ) \nu\in(0,d+\alpha_0) ν ∈ ( 0 , d + α 0 ) there exists a constant C > 0 C>0 C > 0 depending only on d , α 0 , ν d,\alpha_0,\nu d , α 0 , ν and μ ∗ , μ ∗ \mu_\ast,\mu^\ast μ ∗ , μ ∗ such that for all t > 0 t>0 t > 0 , F ∈ L 2 ( R d ; C d ) F\in L^2(\mathbb{R}^d;\mathbb{C}^d) F ∈ L 2 ( R d ; C d ) and cubes Q Q Q with side-length ℓ ( Q ) ≥ t \ell(Q)\ge t ℓ ( Q ) ≥ t we have
∥ Θ t F ∥ L 2 ( Q ) 2 ≤ C ∑ n = 0 ∞ 2 − ν n ∥ F ∥ L 2 ( C n ( Q ) ) 2 . \|\Theta_tF\|_{L^2(Q)}^2\le C\sum_{n=0}^{\infty}2^{-\nu n}\|F\|_{L^2(C_n(Q))}^2. ∥ Θ t F ∥ L 2 ( Q ) 2 ≤ C n = 0 ∑ ∞ 2 − ν n ∥ F ∥ L 2 ( C n ( Q )) 2 .
Lemma 4.3. Let b ∈ L ∞ ( R d ; C d ) b\in L^\infty(\mathbb{R}^d;\mathbb{C}^d) b ∈ L ∞ ( R d ; C d ) . Then for every t > 0 t>0 t > 0 and cube Q ⊂ R d Q\subset\mathbb{R}^d Q ⊂ R d with sides parallel to the coordinate axes the limit
Θ t b : = lim j → ∞ Θ t ( 1 2 j Q b ) \Theta_tb:=\lim_{j\to\infty}\Theta_t(1_{2^jQ}b) Θ t b := j → ∞ lim Θ t ( 1 2 j Q b ) exists in L l o c 2 ( R d ) L^2_{\mathrm{loc}}(\mathbb{R}^d) L loc 2 ( R d ) and is independent of the underlying cube Q Q Q .
The proof is similar to [41 Prop. 4.2] and only needs the off-diagonal estimates from Lemma 4.2 . The remaining details are omitted for brevity.
Definition 4.4. Identify the standard unit vectors e 1 , … , e d ∈ C d e_1,\ldots,e_d\in\mathbb{C}^d e 1 , … , e d ∈ C d with the respective constant functions on R d \mathbb{R}^d R d . For t > 0 t>0 t > 0 we define the principal part of Θ t \Theta_t Θ t by
[ γ t ] j : = Θ t ( e j ) ∈ L l o c 2 ( R d ) [\gamma_t]_j:=\Theta_t(e_j)\in L^2_{\mathrm{loc}}(\mathbb{R}^d) [ γ t ] j := Θ t ( e j ) ∈ L loc 2 ( R d ) for j = 1 , … , d j=1,\ldots,d j = 1 , … , d .
To state the main reduction step of this subsection, we briefly recapitulate the notion of a Carleson measure.
Definition 4.5. A Borel measure Λ \Lambda Λ on R + d + 1 \mathbb{R}^{d+1}_+ R + d + 1 is called a Carleson measure if there exists a constant C > 0 C>0 C > 0 such that
Λ ( R ( Q ) ) ≤ C ∣ Q ∣ ( Q ∈ □ ) , \Lambda(R(Q))\le C|Q|\qquad(Q\in\square), Λ ( R ( Q )) ≤ C ∣ Q ∣ ( Q ∈ □ ) , where R ( Q ) = ( 0 , ℓ ( Q ) ] × Q R(Q)=(0,\ell(Q)]\times Q R ( Q ) = ( 0 , ℓ ( Q )] × Q .
Proposition 4.6. If
∫ 0 ∞ ∥ ( Θ t − γ t ⋅ A t ) ( ∇ α u ) ∥ L 2 2 d t t ≲ ∥ ∇ α u ∥ L 2 2 (17) \int_0^\infty\|(\Theta_t-\gamma_t\cdot A_t)(\nabla^\alpha u)\|_{L^2}^2\frac{dt}{t}\lesssim\|\nabla^\alpha u\|_{L^2}^2 \tag*{(17)} ∫ 0 ∞ ∥ ( Θ t − γ t ⋅ A t ) ( ∇ α u ) ∥ L 2 2 t d t ≲ ∥ ∇ α u ∥ L 2 2 (17)
holds for every u ∈ S α , 2 ( R d ) u \in\mathrm{S}^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) and ∣ γ t ( x ) ∣ 2 d x d t t |\gamma_t(x)|^2 \frac{dx\,dt}{t} ∣ γ t ( x ) ∣ 2 t d x d t is a Carleson measure, then (4.3) is valid for all u ∈ S α , 2 ( R d ) u \in\mathrm{S}^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) .
Its proof is an application of Carleson’s lemma, following the same reasoning as [41 Prop. 4.5] . The rest of this subsection is devoted to verifying (4.4). This will be done in the same fashion as in [41 Sec. 4] . For this purpose we make use of the following identity: for a fixed cube Q ∈ D t Q \in\mathbb{D}_t Q ∈ D t and F ∈ L 2 ( R d ; C d ) F \in L^2(\mathbb{R}^d;\mathbb{C}^d) F ∈ L 2 ( R d ; C d ) we have for x ∈ Q x \in Q x ∈ Q
( γ t ⋅ A t ) F ( x ) = ∑ j = 1 d Θ t ( e j ) ( x ) ( F j ) Q = Θ t ( ∑ j = 1 d ( F j ) Q e j ) ( x ) = Θ t ( ( F ) Q ) ( x ) . (\gamma_t \cdot A_t)F(x) = \sum_{j=1}^{d} \Theta_t(e_j)(x)(F_j)_Q = \Theta_t\left(\sum_{j=1}^{d}(F_j)_Qe_j\right)(x) = \Theta_t((F)_Q)(x). ( γ t ⋅ A t ) F ( x ) = j = 1 ∑ d Θ t ( e j ) ( x ) ( F j ) Q = Θ t ( j = 1 ∑ d ( F j ) Q e j ) ( x ) = Θ t (( F ) Q ) ( x ) .
Furthermore, by Lemma 4.3, the following holds on Q Q Q
( γ t ⋅ A t ) F = lim k → ∞ Θ t ( 1 2 k Q ( F ) Q ) = lim k → ∞ ∑ n = 0 k − 1 Θ t ( 1 C n ( Q ) ( F ) Q ) = ∑ n = 0 ∞ Θ t ( 1 C n ( Q ) ( F ) Q ) . (18) \begin{aligned}
(\gamma_t \cdot A_t)F = \lim_{k\to\infty}\Theta_t(1_{2^kQ}(F)_Q) \\
&= \lim_{k\to\infty}\sum_{n=0}^{k-1}\Theta_t(1_{C_n(Q)}(F)_Q) = \sum_{n=0}^{\infty}\Theta_t(1_{C_n(Q)}(F)_Q). \tag*{(18)}
\end{aligned} ( γ t ⋅ A t ) F = k → ∞ lim Θ t ( 1 2 k Q ( F ) Q ) = k → ∞ lim n = 0 ∑ k − 1 Θ t ( 1 C n ( Q ) ( F ) Q ) = n = 0 ∑ ∞ Θ t ( 1 C n ( Q ) ( F ) Q ) . (18)
Lemma 4.7. For all t > 0 t>0 t > 0 and F ∈ L 2 ( R d ; C d ) F \in L^2(\mathbb{R}^d;\mathbb{C}^d) F ∈ L 2 ( R d ; C d ) we have
∥ ( γ t ⋅ A t ) F ∥ L 2 ≲ ∥ F ∥ L 2 . \|(\gamma_t \cdot A_t)F\|_{L^2} \lesssim\|F\|_{L^2}. ∥ ( γ t ⋅ A t ) F ∥ L 2 ≲ ∥ F ∥ L 2 .
Proof. For fixed t > 0 t>0 t > 0 it is enough to show the bound
∥ ( γ t ⋅ A t ) F ∥ L 2 ( Q ) 2 ≲ ∥ F ∥ L 2 ( Q ) 2 \|(\gamma_t \cdot A_t)F\|_{L^2(Q)}^2 \lesssim\|F\|_{L^2(Q)}^2 ∥ ( γ t ⋅ A t ) F ∥ L 2 ( Q ) 2 ≲ ∥ F ∥ L 2 ( Q ) 2
for all Q ∈ D t Q \in\mathbb{D}_t Q ∈ D t . Since ℓ ( Q ) = 2 j ≥ t \ell(Q)=2^j \ge t ℓ ( Q ) = 2 j ≥ t , using the identity (4.5) and Lemma 4.2 yields
∥ ( γ t ⋅ A t ) F ∥ L 2 ( Q ) ≲ ∑ n = 0 ∞ ∥ Θ t ( 1 C n ( Q ) ( F ) Q ) ∥ L 2 ( Q ) ≲ ∑ n = 0 ∞ 2 − ν 2 n ∥ ( F ) Q ∥ L 2 ( C n ( Q ) ) \|(\gamma_t \cdot A_t)F\|_{L^2(Q)} \lesssim\sum_{n=0}^{\infty}\|\Theta_t(1_{C_n(Q)}(F)_Q)\|_{L^2(Q)} \lesssim\sum_{n=0}^{\infty}2^{-\frac{\nu}{2}n}\|(F)_Q\|_{L^2(C_n(Q))} ∥ ( γ t ⋅ A t ) F ∥ L 2 ( Q ) ≲ n = 0 ∑ ∞ ∥ Θ t ( 1 C n ( Q ) ( F ) Q ) ∥ L 2 ( Q ) ≲ n = 0 ∑ ∞ 2 − 2 ν n ∥ ( F ) Q ∥ L 2 ( C n ( Q ))
for all 0 < ν < d + α 0 0<\nu<d+\alpha_0 0 < ν < d + α 0 . Due to Hölder’s inequality, we have
∥ ( F ) Q ∥ L 2 ( C n ( Q ) ) ≤ ∣ C n ( Q ) ∣ 1 2 ∣ Q ∣ ∫ Q ∣ F ∣ d x ≤ ∣ C n ( Q ) ∣ 1 2 ∣ Q ∣ 1 2 ∥ F ∥ L 2 ( Q ) ≃ 2 d 2 n ∥ F ∥ L 2 ( Q ) . \|(F)_Q\|_{L^2(C_n(Q))} \le\frac{|C_n(Q)|^{\frac12}}{|Q|}\int_Q |F|\,dx \le\frac{|C_n(Q)|^{\frac12}}{|Q|^{\frac12}}\|F\|_{L^2(Q)} \simeq2^{\frac{d}{2}n}\|F\|_{L^2(Q)}. ∥ ( F ) Q ∥ L 2 ( C n ( Q )) ≤ ∣ Q ∣ ∣ C n ( Q ) ∣ 2 1 ∫ Q ∣ F ∣ d x ≤ ∣ Q ∣ 2 1 ∣ C n ( Q ) ∣ 2 1 ∥ F ∥ L 2 ( Q ) ≃ 2 2 d n ∥ F ∥ L 2 ( Q ) .
Hence, we derive
∥ ( γ t ⋅ A t ) F ∥ L 2 ( Q ) ≲ ∑ n = 0 ∞ 2 ( d 2 − ν 2 ) n ∥ F ∥ L 2 ( Q ) ≃ ∥ F ∥ L 2 ( Q ) , \|(\gamma_t \cdot A_t)F\|_{L^2(Q)} \lesssim\sum_{n=0}^{\infty}2^{(\frac{d}{2}-\frac{\nu}{2})n}\|F\|_{L^2(Q)} \simeq\|F\|_{L^2(Q)}, ∥ ( γ t ⋅ A t ) F ∥ L 2 ( Q ) ≲ n = 0 ∑ ∞ 2 ( 2 d − 2 ν ) n ∥ F ∥ L 2 ( Q ) ≃ ∥ F ∥ L 2 ( Q ) ,
where we chose ν = d + α 0 2 > d \nu=d+\frac{\alpha_0}{2}>d ν = d + 2 α 0 > d in the last step. □
The following lemma presents an approximation of Θ t \Theta_t Θ t in an averaged sense in space but pointwise in time t t t . Note that such a control by a fractional Sobolev norm was also used in the resolution of the parabolic Kato problem and for the generalized Stokes operator, see [9 Sec. 7.3] and [41 Lem. 4.7] .
Lemma 4.8. For all t > 0 t>0 t > 0 we have
∥ ( Θ t − γ t ⋅ A t ) F ∥ L 2 ≲ t β ∥ F ∥ W ˙ β , 2 \|(\Theta_t-\gamma_t\cdot A_t)F\|_{L^2} \lesssim t^\beta\|F\|_{\dot W^{\beta,2}} ∥ ( Θ t − γ t ⋅ A t ) F ∥ L 2 ≲ t β ∥ F ∥ W ˙ β , 2 ( F ∈ S β , 2 ( R d ; C d ) ) (F \in\mathrm{S}^{\beta,2}(\mathbb{R}^d;\mathbb{C}^d)) ( F ∈ S β , 2 ( R d ; C d )) for every β ∈ ( 0 , α 0 2 ) \beta\in(0,\frac{\alpha_0}{2}) β ∈ ( 0 , 2 α 0 ) . The implicit constant depends only on d , α 0 , β d,\alpha_0,\beta d , α 0 , β and μ ∙ , μ ∙ \mu_\bullet,\mu^\bullet μ ∙ , μ ∙ .
Proof. Fix t > 0 t>0 t > 0 . We argue as in [41 Lem. 4.7] in two steps. First we work on a fixed dyadic cube, and then sum over a partition of R d \mathbb{R}^{d} R d .
Step 1: Estimate on a fixed cube Q ∈ □ t Q\in\square_t Q ∈ □ t . Fix a cube Q ∈ □ t Q\in\square_t Q ∈ □ t . Since Θ t \Theta_t Θ t is a bounded operator on L 2 L^2 L 2 , we have
Θ t F = ∑ ℓ = 0 ∞ Θ t ( 1 C ℓ ( Q ) F ) . \Theta_tF=\sum_{\ell=0}^{\infty}\Theta_t\left(1_{C_\ell(Q)}F\right). Θ t F = ℓ = 0 ∑ ∞ Θ t ( 1 C ℓ ( Q ) F ) .
By subtracting the identity (4.5) and taking the L 2 ( Q ) L^2(Q) L 2 ( Q ) -norm, we obtain
∥ ( Θ t − γ t ⋅ A t ) F ∥ L 2 ( Q ) ≤ ∑ ℓ = 0 ∞ ∥ Θ t ( 1 C ℓ ( Q ) ( F − ( F ) Q ) ) ∥ L 2 ( Q ) . \left\|\left(\Theta_t-\gamma_t\cdot\mathcal{A}_t\right)F\right\|_{L^2(Q)}
\leq\sum_{\ell=0}^{\infty}\left\|\Theta_t\left(1_{C_\ell(Q)}\left(F-(F)_Q\right)\right)\right\|_{L^2(Q)}. ∥ ( Θ t − γ t ⋅ A t ) F ∥ L 2 ( Q ) ≤ ℓ = 0 ∑ ∞ Θ t ( 1 C ℓ ( Q ) ( F − ( F ) Q ) ) L 2 ( Q ) .
Because ℓ ( Q ) ≥ t \ell(Q)\geq t ℓ ( Q ) ≥ t , we can apply Lemma 4.2 to derive
∥ ( Θ t − γ t ⋅ A t ) F ∥ L 2 ( Q ) ≲ ∑ ℓ = 0 ∞ ( ∑ n = 0 ∞ 2 − ν n ∥ 1 C ℓ ( Q ) ( F − ( F ) Q ) ∥ L 2 ( C n ( Q ) ) 2 ) 1 2 ≤ ∑ ℓ = 0 ∞ 2 − ν 2 ℓ ∥ F − ( F ) Q ∥ L 2 ( 2 ℓ + 1 Q ) . (19) \begin{aligned}
\left\|\left(\Theta_t-\gamma_t\cdot\mathcal{A}_t\right)F\right\|_{L^2(Q)}
&\lesssim\sum_{\ell=0}^{\infty}\left(\sum_{n=0}^{\infty}2^{-\nu n}\left\|1_{C_\ell(Q)}\left(F-(F)_Q\right)\right\|_{L^2(C_n(Q))}^{2}\right)^{\frac{1}{2}}\\
&\leq\sum_{\ell=0}^{\infty}2^{-\frac{\nu}{2}\ell}\left\|F-(F)_Q\right\|_{L^2(2^{\ell+1}Q)}. \tag*{(19)}
\end{aligned} ∥ ( Θ t − γ t ⋅ A t ) F ∥ L 2 ( Q ) ≲ ℓ = 0 ∑ ∞ ( n = 0 ∑ ∞ 2 − ν n 1 C ℓ ( Q ) ( F − ( F ) Q ) L 2 ( C n ( Q )) 2 ) 2 1 ≤ ℓ = 0 ∑ ∞ 2 − 2 ν ℓ ∥ F − ( F ) Q ∥ L 2 ( 2 ℓ + 1 Q ) . (19)
Next, we introduce a telescoping sum to estimate
∥ F − ( F ) Q ∥ L 2 ( 2 ℓ + 1 Q ) ≤ ∥ F − ( F ) 2 ℓ + 1 Q ∥ L 2 ( 2 ℓ + 1 Q ) + ∑ j = 1 ℓ + 1 ∥ ( F ) 2 j Q − ( F ) 2 j − 1 Q ∥ L 2 ( 2 ℓ + 1 Q ) = ∥ F − ( F ) 2 ℓ + 1 Q ∥ L 2 ( 2 ℓ + 1 Q ) + ∑ j = 1 ℓ + 1 ∣ 2 ℓ + 1 Q ∣ 1 2 ∣ ( F ) 2 j Q − ( F ) 2 j − 1 Q ∣ = ∥ F − ( F ) 2 ℓ + 1 Q ∥ L 2 ( 2 ℓ + 1 Q ) + ∑ j = 1 ℓ + 1 ∣ 2 ℓ + 1 Q ∣ 1 2 ∣ 2 j − 1 Q ∣ ∣ ∫ 2 j − 1 Q ( F − ( F ) 2 j Q ) d x ∣ ≲ ∑ j = 1 ℓ + 1 2 d 2 ( ℓ + 1 − j ) ∥ F − ( F ) 2 j Q ∥ L 2 ( 2 j Q ) , \begin{aligned}
\left\|F-(F)_Q\right\|_{L^2(2^{\ell+1}Q)}
&\leq\left\|F-(F)_{2^{\ell+1}Q}\right\|_{L^2(2^{\ell+1}Q)}
+\sum_{j=1}^{\ell+1}\left\|(F)_{2^jQ}-(F)_{2^{j-1}Q}\right\|_{L^2(2^{\ell+1}Q)}\\
&=\left\|F-(F)_{2^{\ell+1}Q}\right\|_{L^2(2^{\ell+1}Q)}
+\sum_{j=1}^{\ell+1}|2^{\ell+1}Q|^{\frac{1}{2}}\left|(F)_{2^jQ}-(F)_{2^{j-1}Q}\right|\\
&=\left\|F-(F)_{2^{\ell+1}Q}\right\|_{L^2(2^{\ell+1}Q)}
+\sum_{j=1}^{\ell+1}\frac{|2^{\ell+1}Q|^{\frac{1}{2}}}{|2^{j-1}Q|}\left|\int_{2^{j-1}Q}\left(F-(F)_{2^jQ}\right)\,dx\right|\\
&\lesssim\sum_{j=1}^{\ell+1}2^{\frac{d}{2}(\ell+1-j)}\left\|F-(F)_{2^jQ}\right\|_{L^2(2^jQ)},
\end{aligned} ∥ F − ( F ) Q ∥ L 2 ( 2 ℓ + 1 Q ) ≤ F − ( F ) 2 ℓ + 1 Q L 2 ( 2 ℓ + 1 Q ) + j = 1 ∑ ℓ + 1 ( F ) 2 j Q − ( F ) 2 j − 1 Q L 2 ( 2 ℓ + 1 Q ) = F − ( F ) 2 ℓ + 1 Q L 2 ( 2 ℓ + 1 Q ) + j = 1 ∑ ℓ + 1 ∣ 2 ℓ + 1 Q ∣ 2 1 ( F ) 2 j Q − ( F ) 2 j − 1 Q = F − ( F ) 2 ℓ + 1 Q L 2 ( 2 ℓ + 1 Q ) + j = 1 ∑ ℓ + 1 ∣ 2 j − 1 Q ∣ ∣ 2 ℓ + 1 Q ∣ 2 1 ∫ 2 j − 1 Q ( F − ( F ) 2 j Q ) d x ≲ j = 1 ∑ ℓ + 1 2 2 d ( ℓ + 1 − j ) F − ( F ) 2 j Q L 2 ( 2 j Q ) ,
where we used Hölder’s inequality in the last step. Now, we apply the fractional Poincaré inequality from [46 Lem. 2.2] componentwise for β ∈ ( 0 , α 0 2 ) \beta\in\left(0,\frac{\alpha_0}{2}\right) β ∈ ( 0 , 2 α 0 ) to get
∥ F − ( F ) 2 j Q ∥ L 2 ( 2 j Q ) ≲ ∣ 2 j Q ∣ β d ∥ F ∥ W ˙ β , 2 ( 2 j Q ) ≲ t β 2 β j ∥ F ∥ W ˙ β , 2 ( 2 j Q ) (20) \left\|F-(F)_{2^jQ}\right\|_{L^2(2^jQ)}
\lesssim|2^jQ|^{\frac{\beta}{d}}\|F\|_{\dot{W}^{\beta,2}(2^jQ)}
\lesssim t^\beta2^{\beta j}\|F\|_{\dot{W}^{\beta,2}(2^jQ)} \tag*{(20)} F − ( F ) 2 j Q L 2 ( 2 j Q ) ≲ ∣ 2 j Q ∣ d β ∥ F ∥ W ˙ β , 2 ( 2 j Q ) ≲ t β 2 β j ∥ F ∥ W ˙ β , 2 ( 2 j Q ) (20)
for all j ∈ N j\in\mathbb{N} j ∈ N and implicit constants depending only on d d d and β \beta β . This leads to the bound
∥ F − ( F ) Q ∥ L 2 ( 2 ℓ + 1 Q ) ≲ t β ∑ j = 1 ℓ + 1 2 d 2 ( ℓ + 1 − j ) + β j ∥ F ∥ W ˙ β , 2 ( 2 j Q ) . \left\|F-(F)_Q\right\|_{L^2(2^{\ell+1}Q)}
\lesssim t^\beta\sum_{j=1}^{\ell+1}2^{\frac{d}{2}(\ell+1-j)+\beta j}\|F\|_{\dot{W}^{\beta,2}(2^jQ)}. ∥ F − ( F ) Q ∥ L 2 ( 2 ℓ + 1 Q ) ≲ t β j = 1 ∑ ℓ + 1 2 2 d ( ℓ + 1 − j ) + β j ∥ F ∥ W ˙ β , 2 ( 2 j Q ) .
Inserting this estimate into (19) , squaring both sides and applying Hölder’s inequality, we obtain for 0 < ν ′ < ν 0<\nu'<\nu 0 < ν ′ < ν that
∥ ( Θ t − γ t ⋅ A t ) F ∥ L 2 ( Q ) 2 ≲ t 2 β ∑ ℓ = 0 ∞ 2 − ν ′ ℓ ( ∑ j = 1 ℓ + 1 2 d 2 ( ℓ + 1 − j ) + β j ∥ F ∥ W ˙ β , 2 ( 2 j Q ) ) 2 (21) \lVert(\Theta_t-\gamma_t\cdot A_t)F\rVert_{L^2(Q)}^2\lesssim t^{2\beta}\sum_{\ell=0}^{\infty}2^{-\nu'\ell}\left(\sum_{j=1}^{\ell+1}2^{\frac{d}{2}(\ell+1-j)+\beta j}\lVert F\rVert_{\dot{W}^{\beta,2}(2^jQ)}\right)^2 \tag*{(21)} ∥( Θ t − γ t ⋅ A t ) F ∥ L 2 ( Q ) 2 ≲ t 2 β ℓ = 0 ∑ ∞ 2 − ν ′ ℓ ( j = 1 ∑ ℓ + 1 2 2 d ( ℓ + 1 − j ) + β j ∥ F ∥ W ˙ β , 2 ( 2 j Q ) ) 2 (21)
≲ t 2 β ∑ ℓ = 0 ∞ ( ℓ + 1 ) 2 − ν ′ ℓ ∑ j = 1 ℓ + 1 2 d ( ℓ + 1 − j ) + 2 β j ∥ F ∥ W ˙ β , 2 ( 2 j Q ) 2 . \lesssim t^{2\beta}\sum_{\ell=0}^{\infty}(\ell+1)2^{-\nu'\ell}\sum_{j=1}^{\ell+1}2^{d(\ell+1-j)+2\beta j}\lVert F\rVert_{\dot{W}^{\beta,2}(2^jQ)}^2. ≲ t 2 β ℓ = 0 ∑ ∞ ( ℓ + 1 ) 2 − ν ′ ℓ j = 1 ∑ ℓ + 1 2 d ( ℓ + 1 − j ) + 2 β j ∥ F ∥ W ˙ β , 2 ( 2 j Q ) 2 .
Step 2: Conclusion. Summing over all Q ∈ D t Q\in\mathcal{D}_t Q ∈ D t in (21) gives us
∥ ( Θ t − γ t ⋅ A t ) F ∥ L 2 2 ≲ t 2 β ∑ ℓ = 0 ∞ ( ℓ + 1 ) 2 − ν ′ ℓ ∑ j = 1 ℓ + 1 2 d ( ℓ + 1 − j ) + 2 β j ∑ Q ∈ D t ∥ F ∥ W ˙ β , 2 ( 2 j Q ) 2 . \lVert(\Theta_t-\gamma_t\cdot A_t)F\rVert_{L^2}^2\lesssim t^{2\beta}\sum_{\ell=0}^{\infty}(\ell+1)2^{-\nu'\ell}\sum_{j=1}^{\ell+1}2^{d(\ell+1-j)+2\beta j}\sum_{Q\in\mathcal{D}_t}\lVert F\rVert_{\dot{W}^{\beta,2}(2^jQ)}^2. ∥( Θ t − γ t ⋅ A t ) F ∥ L 2 2 ≲ t 2 β ℓ = 0 ∑ ∞ ( ℓ + 1 ) 2 − ν ′ ℓ j = 1 ∑ ℓ + 1 2 d ( ℓ + 1 − j ) + 2 β j Q ∈ D t ∑ ∥ F ∥ W ˙ β , 2 ( 2 j Q ) 2 .
By a counting argument, we observe that
∑ Q ∈ D t ∥ F ∥ W ˙ β , 2 ( 2 j Q ) 2 ≤ ∬ R d × R d ( ∑ Q ∈ D t 1 2 j Q ( x ) ) ∣ F ( x ) − F ( y ) ∣ 2 ∣ x − y ∣ d + 2 β d x d y ≲ 2 d j ∥ F ∥ W ˙ β , 2 ( R d ) 2 \sum_{Q\in\mathcal{D}_t}\lVert F\rVert_{\dot{W}^{\beta,2}(2^jQ)}^2\leq\iint_{\mathbb{R}^d\times\mathbb{R}^d}\left(\sum_{Q\in\mathcal{D}_t}\mathbf{1}_{2^jQ}(x)\right)\frac{|F(x)-F(y)|^2}{|x-y|^{d+2\beta}}\,dx\,dy\lesssim2^{dj}\lVert F\rVert_{\dot{W}^{\beta,2}(\mathbb{R}^d)}^2 Q ∈ D t ∑ ∥ F ∥ W ˙ β , 2 ( 2 j Q ) 2 ≤ ∬ R d × R d Q ∈ D t ∑ 1 2 j Q ( x ) ∣ x − y ∣ d + 2 β ∣ F ( x ) − F ( y ) ∣ 2 d x d y ≲ 2 d j ∥ F ∥ W ˙ β , 2 ( R d ) 2
for all j ∈ N j\in\mathbb{N} j ∈ N . Thus, we conclude
∥ ( Θ t − γ t ⋅ A t ) F ∥ L 2 2 ≲ t 2 β ∑ ℓ = 0 ∞ ( ℓ + 1 ) 2 − ν ′ ℓ ∑ j = 1 ℓ + 1 2 d ( ℓ + 1 ) + 2 β j ∥ F ∥ W ˙ β , 2 ( R d ) 2 \lVert(\Theta_t-\gamma_t\cdot A_t)F\rVert_{L^2}^2\lesssim t^{2\beta}\sum_{\ell=0}^{\infty}(\ell+1)2^{-\nu'\ell}\sum_{j=1}^{\ell+1}2^{d(\ell+1)+2\beta j}\lVert F\rVert_{\dot{W}^{\beta,2}(\mathbb{R}^d)}^2 ∥( Θ t − γ t ⋅ A t ) F ∥ L 2 2 ≲ t 2 β ℓ = 0 ∑ ∞ ( ℓ + 1 ) 2 − ν ′ ℓ j = 1 ∑ ℓ + 1 2 d ( ℓ + 1 ) + 2 β j ∥ F ∥ W ˙ β , 2 ( R d ) 2
≲ t 2 β ∥ F ∥ W ˙ β , 2 ( R d ) 2 ∑ ℓ = 0 ∞ ( ℓ + 1 ) 2 ( d + 2 β − ν ′ ) ℓ \lesssim t^{2\beta}\lVert F\rVert_{\dot{W}^{\beta,2}(\mathbb{R}^d)}^2\sum_{\ell=0}^{\infty}(\ell+1)2^{(d+2\beta-\nu')\ell} ≲ t 2 β ∥ F ∥ W ˙ β , 2 ( R d ) 2 ℓ = 0 ∑ ∞ ( ℓ + 1 ) 2 ( d + 2 β − ν ′ ) ℓ
≲ t 2 β ∥ F ∥ W ˙ β , 2 ( R d ) 2 , \lesssim t^{2\beta}\lVert F\rVert_{\dot{W}^{\beta,2}(\mathbb{R}^d)}^2, ≲ t 2 β ∥ F ∥ W ˙ β , 2 ( R d ) 2 ,
where the series converges for every β ∈ ( 0 , α 0 2 ) \beta\in\left(0,\frac{\alpha_0}{2}\right) β ∈ ( 0 , 2 α 0 ) since we can choose d + 2 β < ν ′ < d + α 0 d+2\beta<\nu'<d+\alpha_0 d + 2 β < ν ′ < d + α 0 .
To prove (17) , we will first smooth out the “harsh” approximation to get a control via quadratic estimates of some well-behaved operator and afterwards remove the smoothing again to gain back the full estimate. For this purpose, we introduce the following smoothing operators.
Definition 4.9. For t > 0 t>0 t > 0 we define the following operators on L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) :
P t = ( 1 + t 2 α ( − Δ ) α ) − 1 and Q t = t α ∇ α ( 1 + t 2 α ( − Δ ) α ) − 1 . P_t=(1+t^{2\alpha}(-\Delta)^\alpha)^{-1}\qquad\text{and}\qquad Q_t=t^\alpha\nabla^\alpha(1+t^{2\alpha}(-\Delta)^\alpha)^{-1}. P t = ( 1 + t 2 α ( − Δ ) α ) − 1 and Q t = t α ∇ α ( 1 + t 2 α ( − Δ ) α ) − 1 .
As a reminder, the reader should keep in mind that we agreed to use the same notation for extensions of operators on vector fields by acting componentwise.
Lemma 4.10. Let ( P t ) t > 0 (P_t)_{t>0} ( P t ) t > 0 and ( Q t ) t > 0 (Q_t)_{t>0} ( Q t ) t > 0 be as in Definition 4.9. Then we have the following assertions.
(1) The operator families ( P t ) t > 0 (P_t)_{t>0} ( P t ) t > 0 and ( Q t ) t > 0 (Q_t)_{t>0} ( Q t ) t > 0 are uniformly L 2 L^2 L 2 -bounded. Moreover, for every u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) , we have t α P t ∇ α u = Q t u t^\alpha P_t\nabla^\alpha u=Q_tu t α P t ∇ α u = Q t u .
(2) We have the adjoint square function estimate
∫ 0 ∞ ∥ Q t ∗ ∇ α u ∥ L 2 2 d t t = 1 2 α ∥ ∇ α u ∥ L 2 2 ( u ∈ S α , 2 ( R d ) ) . \int_0^\infty\lVert Q_t^*\nabla^\alpha u\rVert_{L^2}^2\frac{dt}{t}=\frac{1}{2\alpha}\lVert\nabla^\alpha u\rVert_{L^2}^2\qquad(u\in S^{\alpha,2}(\mathbb{R}^d)). ∫ 0 ∞ ∥ Q t ∗ ∇ α u ∥ L 2 2 t d t = 2 α 1 ∥ ∇ α u ∥ L 2 2 ( u ∈ S α , 2 ( R d )) .
(3) The following Calderón reproducing formula holds
2 α ∫ 0 ∞ Q t Q t ∗ ∇ α u d t t = ∇ α u ( u ∈ S α , 2 ( R d ) ) . 2\alpha\int_{0}^{\infty} Q_tQ_t^{*}\nabla^{\alpha}u \,\frac{\mathrm{d}t}{t}=\nabla^{\alpha}u \qquad(u\in S^{\alpha,2}(\mathbb{R}^{d})). 2 α ∫ 0 ∞ Q t Q t ∗ ∇ α u t d t = ∇ α u ( u ∈ S α , 2 ( R d )) .
Proof. We prove the three assertions separately.
Step 1: Show (1). By Proposition 3.4 with μ = Id C d × d \mu=\operatorname{Id}_{\mathbb{C}^{d\times d}} μ = Id C d × d being the identity matrix, we observe that
∥ P t u ∥ L 2 + ∥ Q t u ∥ L 2 ≲ ∥ u ∥ L 2 \|P_tu\|_{L^2}+\|Q_tu\|_{L^2}\lesssim\|u\|_{L^2} ∥ P t u ∥ L 2 + ∥ Q t u ∥ L 2 ≲ ∥ u ∥ L 2
for some implicit constant depending only on d d d . Furthermore, for all u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^{d}) u ∈ S α , 2 ( R d ) we have the identity t α P t ∇ α u = Q t u t^{\alpha}P_t\nabla^{\alpha}u=Q_tu t α P t ∇ α u = Q t u by taking the Fourier transform, see Proposition 2.3.
Step 2: Show (2). Notice that Q t ∗ = − t α ( 1 + t 2 α ( − Δ ) α ) − 1 div α Q_t^{*}=-t^{\alpha}(1+t^{2\alpha}(-\Delta)^{\alpha})^{-1}\operatorname{div}^{\alpha} Q t ∗ = − t α ( 1 + t 2 α ( − Δ ) α ) − 1 div α , where ( − Δ ) α = − div α ∇ α (-\Delta)^{\alpha}=-\operatorname{div}^{\alpha}\nabla^{\alpha} ( − Δ ) α = − div α ∇ α has to be understood in the sense of Lemma 3.3. Hence, we compute
Q t ∗ ∇ α u = − t α ( 1 + t 2 α ( − Δ ) α ) − 1 div α ( ∇ α u ) = t α ( − Δ ) α ( 1 + t 2 α ( − Δ ) α ) − 1 u (22) Q_t^{*}\nabla^{\alpha}u=-t^{\alpha}(1+t^{2\alpha}(-\Delta)^{\alpha})^{-1}\operatorname{div}^{\alpha}(\nabla^{\alpha}u)=t^{\alpha}(-\Delta)^{\alpha}(1+t^{2\alpha}(-\Delta)^{\alpha})^{-1}u \tag*{(22)} Q t ∗ ∇ α u = − t α ( 1 + t 2 α ( − Δ ) α ) − 1 div α ( ∇ α u ) = t α ( − Δ ) α ( 1 + t 2 α ( − Δ ) α ) − 1 u (22)
for every u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^{d}) u ∈ S α , 2 ( R d ) . Define ψ α ( r ) : = r α ( 1 + r 2 α ) − 1 \psi_{\alpha}(r):=r^{\alpha}(1+r^{2\alpha})^{-1} ψ α ( r ) := r α ( 1 + r 2 α ) − 1 for r > 0 r>0 r > 0 . Then, we get with Plancherel’s and Fubini’s theorems
∫ 0 ∞ ∥ Q t ∗ ∇ α u ∥ L 2 2 d t t = ∫ 0 ∞ ∥ t α ( − Δ ) α ( 1 + t 2 α ( − Δ ) α ) − 1 u ∥ L 2 2 d t t = ∫ 0 ∞ ∫ R d ∣ ∣ ξ ∣ α ψ α ( t ∣ ξ ∣ ) F ( u ) ( ξ ) ∣ 2 d ξ d t t = ∫ R d ( ∫ 0 ∞ ∣ ψ α ( t ∣ ξ ∣ ) ∣ 2 d t t ) ∣ ∣ ξ ∣ α F ( u ) ( ξ ) ∣ 2 d ξ . \begin{aligned}
\int_{0}^{\infty}\|Q_t^{*}\nabla^{\alpha}u\|_{L^2}^{2}\frac{\mathrm{d}t}{t}
=\int_{0}^{\infty}\|t^{\alpha}(-\Delta)^{\alpha}(1+t^{2\alpha}(-\Delta)^{\alpha})^{-1}u\|_{L^2}^{2}\frac{\mathrm{d}t}{t} \\
&=\int_{0}^{\infty}\int_{\mathbb{R}^{d}}|\lvert\xi\rvert^{\alpha}\psi_{\alpha}(t\lvert\xi\rvert)\mathcal{F}(u)(\xi)|^{2}\,\frac{\mathrm{d}\xi\mathrm{d}t}{t} \\
&=\int_{\mathbb{R}^{d}}\left(\int_{0}^{\infty}|\psi_{\alpha}(t\lvert\xi\rvert)|^{2}\frac{\mathrm{d}t}{t}\right)|\lvert\xi\rvert^{\alpha}\mathcal{F}(u)(\xi)|^{2}\,\mathrm{d}\xi.
\end{aligned} ∫ 0 ∞ ∥ Q t ∗ ∇ α u ∥ L 2 2 t d t = ∫ 0 ∞ ∥ t α ( − Δ ) α ( 1 + t 2 α ( − Δ ) α ) − 1 u ∥ L 2 2 t d t = ∫ 0 ∞ ∫ R d ∣ ∣ ξ ∣ α ψ α ( t ∣ ξ ∣) F ( u ) ( ξ ) ∣ 2 t d ξ d t = ∫ R d ( ∫ 0 ∞ ∣ ψ α ( t ∣ ξ ∣) ∣ 2 t d t ) ∣ ∣ ξ ∣ α F ( u ) ( ξ ) ∣ 2 d ξ .
Using the substitution t ∣ ξ ∣ ↦ t t\lvert\xi\rvert\mapsto t t ∣ ξ ∣ ↦ t , Proposition 2.3 with ∣ ξ ∣ α = ∣ ξ ∣ ∣ ξ ∣ α − 1 \lvert\xi\rvert^{\alpha}=\lvert\xi\rvert\lvert\xi\rvert^{\alpha-1} ∣ ξ ∣ α = ∣ ξ ∣ ∣ ξ ∣ α − 1 and Plancherel’s theorem leaves us with
∫ 0 ∞ ∥ Q t ∗ ∇ α u ∥ L 2 2 d t t = ∫ R d ( ∫ 0 ∞ ∣ ψ α ( t ) ∣ 2 d t t ) ∣ ∣ ξ ∣ α − 1 F ( u ) ( ξ ) ∣ 2 d ξ = 1 2 α ∥ ∇ α u ∥ L 2 2 , \int_{0}^{\infty}\|Q_t^{*}\nabla^{\alpha}u\|_{L^2}^{2}\frac{\mathrm{d}t}{t}
=\int_{\mathbb{R}^{d}}\left(\int_{0}^{\infty}|\psi_{\alpha}(t)|^{2}\frac{\mathrm{d}t}{t}\right)|\lvert\xi\rvert^{\alpha-1}\mathcal{F}(u)(\xi)|^{2}\,\mathrm{d}\xi
=\frac{1}{2\alpha}\|\nabla^{\alpha}u\|_{L^2}^{2}, ∫ 0 ∞ ∥ Q t ∗ ∇ α u ∥ L 2 2 t d t = ∫ R d ( ∫ 0 ∞ ∣ ψ α ( t ) ∣ 2 t d t ) ∣ ∣ ξ ∣ α − 1 F ( u ) ( ξ ) ∣ 2 d ξ = 2 α 1 ∥ ∇ α u ∥ L 2 2 ,
where we used the integral identity
∫ 0 ∞ ∣ ψ α ( t ) ∣ 2 d t t = ∫ 0 ∞ t 2 α ( 1 + t 2 α ) − 2 d t t = 1 2 α ∫ 0 ∞ s ( 1 + s ) − 2 d s s = 1 2 α . \int_{0}^{\infty}|\psi_{\alpha}(t)|^{2}\frac{\mathrm{d}t}{t}
=\int_{0}^{\infty}t^{2\alpha}(1+t^{2\alpha})^{-2}\frac{\mathrm{d}t}{t}
=\frac{1}{2\alpha}\int_{0}^{\infty}s(1+s)^{-2}\frac{\mathrm{d}s}{s}
=\frac{1}{2\alpha}. ∫ 0 ∞ ∣ ψ α ( t ) ∣ 2 t d t = ∫ 0 ∞ t 2 α ( 1 + t 2 α ) − 2 t d t = 2 α 1 ∫ 0 ∞ s ( 1 + s ) − 2 s d s = 2 α 1 .
Step 3: Show (3). Applying Q t Q_t Q t to (4.9) yields
Q t Q t ∗ ∇ α u = t 2 α ( − Δ ) α ( 1 + t 2 α ( − Δ ) α ) − 2 ∇ α u , Q_tQ_t^{*}\nabla^{\alpha}u=t^{2\alpha}(-\Delta)^{\alpha}(1+t^{2\alpha}(-\Delta)^{\alpha})^{-2}\nabla^{\alpha}u, Q t Q t ∗ ∇ α u = t 2 α ( − Δ ) α ( 1 + t 2 α ( − Δ ) α ) − 2 ∇ α u ,
for u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^{d}) u ∈ S α , 2 ( R d ) by commuting the corresponding symbols in Fourier space with Proposition 2.3. Hence, (3) follows, for example, from the Calderón reproducing formula of the negative Laplacian ( − Δ ) (-\Delta) ( − Δ ) , see [39 Thm. 6.16] . □
To prove (17) , we decompose
Θ t − γ t ⋅ A t = Θ t ( 1 − P t ) + ( Θ t − γ t ⋅ A t ) P t − γ t ⋅ A t ( 1 − P t ) (23) \Theta_t-\gamma_t\cdot A_t=\Theta_t(1-P_t)+(\Theta_t-\gamma_t\cdot A_t)P_t-\gamma_t\cdot A_t(1-P_t) \tag*{(23)} Θ t − γ t ⋅ A t = Θ t ( 1 − P t ) + ( Θ t − γ t ⋅ A t ) P t − γ t ⋅ A t ( 1 − P t ) (23)
and treat the three terms separately.
Proposition 4.11. For every u ∈ S α , 2 ( R d ) u\in\mathrm{S}^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) we have
∫ 0 ∞ ∥ Θ t ( 1 − P t ) ∇ α u ∥ L 2 2 d t t ≲ ∥ ∇ α u ∥ L 2 2 . \int_0^\infty\left\|\Theta_t(1-P_t)\nabla^\alpha u\right\|_{L^2}^2\frac{\mathrm{d}t}{t}\lesssim\left\|\nabla^\alpha u\right\|_{L^2}^2. ∫ 0 ∞ ∥ Θ t ( 1 − P t ) ∇ α u ∥ L 2 2 t d t ≲ ∥ ∇ α u ∥ L 2 2 .
Proof. By Definition 4.1, (15) and (22) , we have for every u ∈ S α , 2 ( R d ) u\in\mathrm{S}^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) the identity
Θ t ( 1 − P t ) ∇ α u = Θ t ∇ α ( 1 − P t ) u = − t α ( 1 + t 2 α L ) − 1 div α ( μ ∇ α ( 1 − P t ) u ) = t α L ( 1 + t 2 α L ) − 1 t 2 α ( − Δ ) α ( 1 + t 2 α ( − Δ ) α ) − 1 u = ( 1 − ( 1 + t 2 α L ) − 1 ) Q t ∗ ∇ α u . \begin{aligned}
\Theta_t(1-P_t)\nabla^\alpha u=\Theta_t\nabla^\alpha(1-P_t)u=-t^\alpha(1+t^{2\alpha}\mathcal{L})^{-1}\operatorname{div}_\alpha\left(\mu\nabla^\alpha(1-P_t)u\right) \\
&=t^\alpha\mathcal{L}(1+t^{2\alpha}\mathcal{L})^{-1}t^{2\alpha}(-\Delta)^\alpha(1+t^{2\alpha}(-\Delta)^\alpha)^{-1}u \\
&=\left(1-(1+t^{2\alpha}\mathcal{L})^{-1}\right)Q_t^*\nabla^\alpha u.
\end{aligned} Θ t ( 1 − P t ) ∇ α u = Θ t ∇ α ( 1 − P t ) u = − t α ( 1 + t 2 α L ) − 1 div α ( μ ∇ α ( 1 − P t ) u ) = t α L ( 1 + t 2 α L ) − 1 t 2 α ( − Δ ) α ( 1 + t 2 α ( − Δ ) α ) − 1 u = ( 1 − ( 1 + t 2 α L ) − 1 ) Q t ∗ ∇ α u .
Combining this with Proposition 3.4 yields the estimate
∥ Θ t ( 1 − P t ) ∇ α u ∥ L 2 ≲ ∥ Q t ∗ ∇ α u ∥ L 2 , \left\|\Theta_t(1-P_t)\nabla^\alpha u\right\|_{L^2}\lesssim\left\|Q_t^*\nabla^\alpha u\right\|_{L^2}, ∥ Θ t ( 1 − P t ) ∇ α u ∥ L 2 ≲ ∥ Q t ∗ ∇ α u ∥ L 2 ,
which concludes the claim by Lemma 4.10. □
Proposition 4.12 (Smoothed principal part approximation). For all F ∈ L 2 ( R d ; C d ) F\in L^2(\mathbb{R}^d;\mathbb{C}^d) F ∈ L 2 ( R d ; C d ) we have
∫ 0 ∞ ∥ ( Θ t − γ t ⋅ A t ) P t F ∥ L 2 2 d t t ≲ ∥ F ∥ L 2 2 . \int_0^\infty\left\|(\Theta_t-\gamma_t\cdot A_t)P_tF\right\|_{L^2}^2\frac{\mathrm{d}t}{t}\lesssim\left\|F\right\|_{L^2}^2. ∫ 0 ∞ ∥ ( Θ t − γ t ⋅ A t ) P t F ∥ L 2 2 t d t ≲ ∥ F ∥ L 2 2 .
Proof. Set β = α 0 4 \beta=\frac{\alpha_0}{4} β = 4 α 0 . By Lemma 4.8 and Proposition 2.5, we have
∥ ( Θ t − γ t ⋅ A t ) P t F ∥ L 2 ≲ t β ∥ P t F ∥ W ˙ β , 2 ≃ ∥ t β ( − Δ ) β 2 ( 1 + t 2 α ( − Δ ) α ) − 1 F ∥ L 2 . \left\|(\Theta_t-\gamma_t\cdot A_t)P_tF\right\|_{L^2}\lesssim t^\beta\left\|P_tF\right\|_{\dot{W}^{\beta,2}}\simeq\left\|t^\beta(-\Delta)^{\frac{\beta}{2}}(1+t^{2\alpha}(-\Delta)^\alpha)^{-1}F\right\|_{L^2}. ∥ ( Θ t − γ t ⋅ A t ) P t F ∥ L 2 ≲ t β ∥ P t F ∥ W ˙ β , 2 ≃ t β ( − Δ ) 2 β ( 1 + t 2 α ( − Δ ) α ) − 1 F L 2 .
Finally, using Plancherel and Fubini’s theorems yields
∫ 0 ∞ ∥ ( Θ t − γ t ⋅ A t ) P t F ∥ L 2 2 d t t ≲ ∫ 0 ∞ ∥ t β ( − Δ ) β 2 ( 1 + t 2 α ( − Δ ) α ) − 1 F ∥ L 2 2 d t t = ∫ 0 ∞ ∫ R d ∣ t β ∣ ξ ∣ β ( 1 + t 2 α ∣ ξ ∣ 2 α ) − 1 F ( F ) ( ξ ) ∣ 2 d ξ d t t = ∫ R d ( ∫ 0 ∞ ( t ∣ ξ ∣ ) 2 β ( 1 + ( t ∣ ξ ∣ ) 2 α ) − 2 d t t ) ∣ F ( F ) ( ξ ) ∣ 2 d ξ = ( ∫ 0 ∞ t 2 β ( 1 + t 2 α ) − 2 d t t ) ∥ F ∥ L 2 2 . \begin{aligned}
\int_0^\infty\left\|(\Theta_t-\gamma_t\cdot A_t)P_tF\right\|_{L^2}^2\frac{\mathrm{d}t}{t}\lesssim\int_0^\infty\left\|t^\beta(-\Delta)^{\frac{\beta}{2}}(1+t^{2\alpha}(-\Delta)^\alpha)^{-1}F\right\|_{L^2}^2\frac{\mathrm{d}t}{t} \\
&=\int_0^\infty\int_{\mathbb{R}^d}\left|t^\beta|\xi|^\beta(1+t^{2\alpha}|\xi|^{2\alpha})^{-1}\mathcal{F}(F)(\xi)\right|^2\frac{\mathrm{d}\xi\mathrm{d}t}{t} \\
&=\int_{\mathbb{R}^d}\left(\int_0^\infty(t|\xi|)^{2\beta}(1+(t|\xi|)^{2\alpha})^{-2}\frac{\mathrm{d}t}{t}\right)|\mathcal{F}(F)(\xi)|^2\mathrm{d}\xi\\
&=\left(\int_0^\infty t^{2\beta}(1+t^{2\alpha})^{-2}\frac{\mathrm{d}t}{t}\right)\left\|F\right\|_{L^2}^2.
\end{aligned} ∫ 0 ∞ ∥ ( Θ t − γ t ⋅ A t ) P t F ∥ L 2 2 t d t ≲ ∫ 0 ∞ t β ( − Δ ) 2 β ( 1 + t 2 α ( − Δ ) α ) − 1 F L 2 2 t d t = ∫ 0 ∞ ∫ R d t β ∣ ξ ∣ β ( 1 + t 2 α ∣ ξ ∣ 2 α ) − 1 F ( F ) ( ξ ) 2 t d ξ d t = ∫ R d ( ∫ 0 ∞ ( t ∣ ξ ∣ ) 2 β ( 1 + ( t ∣ ξ ∣ ) 2 α ) − 2 t d t ) ∣ F ( F ) ( ξ ) ∣ 2 d ξ = ( ∫ 0 ∞ t 2 β ( 1 + t 2 α ) − 2 t d t ) ∥ F ∥ L 2 2 .
Using the estimate of the integral
∫ 0 ∞ t 2 β ( 1 + t 2 α ) − 2 d t t ≤ ∫ 0 1 t 2 β d t t + ∫ 1 ∞ t − 2 α d t t = 1 2 β + 1 2 α ≤ 5 2 α 0 , \int_0^\infty t^{2\beta}(1+t^{2\alpha})^{-2}\frac{\mathrm{d}t}{t}\leq\int_0^1t^{2\beta}\frac{\mathrm{d}t}{t}+\int_1^\infty t^{-2\alpha}\frac{\mathrm{d}t}{t}=\frac{1}{2\beta}+\frac{1}{2\alpha}\leq\frac{5}{2\alpha_0}, ∫ 0 ∞ t 2 β ( 1 + t 2 α ) − 2 t d t ≤ ∫ 0 1 t 2 β t d t + ∫ 1 ∞ t − 2 α t d t = 2 β 1 + 2 α 1 ≤ 2 α 0 5 ,
we obtain the claim with a constant that depends only on d d d , α 0 \alpha_0 α 0 and μ 0 \mu_0 μ 0 , μ \mu μ .
To handle the last term of the splitting (4.10), we need the following auxiliary lemma, which is also crucial for the T ( b ) T(b) T ( b ) argument. We refer to [39 Lem. 13.6] for a local version.
Lemma 4.13. Let α 0 ∈ ( 0 , 1 ) \alpha_0 \in(0,1) α 0 ∈ ( 0 , 1 ) . There exists C > 0 C>0 C > 0 such that for all 0 < γ < α 0 0<\gamma<\alpha_0 0 < γ < α 0 , Q ∈ D Q\in\mathbb{D} Q ∈ D , α ∈ [ α 0 , 1 ) \alpha\in[\alpha_0,1) α ∈ [ α 0 , 1 ) and all u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) we have
∣ ∫ Q ∇ α u d x ∣ 2 ≤ C ℓ ( Q ) 2 α 3 ( ∑ k = 0 ∞ 2 − γ k ∫ 2 k + 1 Q ∣ u ∣ 2 d x ) 1 3 ( ∫ Q ∣ ∇ α u ∣ 2 d x ) 2 3 . \left|\int_Q \nabla^\alpha u\,\mathrm{d}x\right|^2 \le\frac{C}{\ell(Q)^{\frac{2\alpha}{3}}}\left(\sum_{k=0}^{\infty}2^{-\gamma k}\int_{2^{k+1}Q}|u|^2\,\mathrm{d}x\right)^{\frac{1}{3}}\left(\int_Q|\nabla^\alpha u|^2\,\mathrm{d}x\right)^{\frac{2}{3}}. ∫ Q ∇ α u d x 2 ≤ ℓ ( Q ) 3 2 α C ( k = 0 ∑ ∞ 2 − γ k ∫ 2 k + 1 Q ∣ u ∣ 2 d x ) 3 1 ( ∫ Q ∣ ∇ α u ∣ 2 d x ) 3 2 . The constant C > 0 C>0 C > 0 depends only on d d d and α 0 \alpha_0 α 0 .
Proof. Let a ∈ ( 0 , 1 ) a\in(0,1) a ∈ ( 0 , 1 ) be chosen later. Fix u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) , Q ∈ D Q\in\mathbb{D} Q ∈ D with side-length ℓ = ℓ ( Q ) \ell=\ell(Q) ℓ = ℓ ( Q ) , and a function η ∈ C 0 ∞ ( Q ) \eta\in C^\infty_0(Q) η ∈ C 0 ∞ ( Q ) with 0 ≤ η ≤ 1 0\le\eta\le1 0 ≤ η ≤ 1 and η = 1 \eta=1 η = 1 on ( 1 − a ) Q (1-a)Q ( 1 − a ) Q that satisfies a ℓ ∥ ∇ η ∥ L ∞ ≤ C a\ell\|\nabla\eta\|_{L^\infty}\le C a ℓ ∥∇ η ∥ L ∞ ≤ C for some constant C C C that only depends on d d d . Applying Definition 2.4 to F ′ = η e j F^\prime=\eta e_j F ′ = η e j for every 1 ≤ j ≤ d 1\le j\le d 1 ≤ j ≤ d , we get ∫ η ∇ α u d x = − ∫ ( ∇ α η ) u d x \int\eta\nabla^\alpha u\,\mathrm{d}x=-\int(\nabla^\alpha\eta)u\,\mathrm{d}x ∫ η ∇ α u d x = − ∫ ( ∇ α η ) u d x . Thus, with the splitting ∇ α u = η ∇ α u + ( 1 − η ) ∇ α u \nabla^\alpha u=\eta\nabla^\alpha u+(1-\eta)\nabla^\alpha u ∇ α u = η ∇ α u + ( 1 − η ) ∇ α u we obtain
∫ Q ∇ α u d x = ∫ Q η ∇ α u d x + ∫ Q ( 1 − η ) ∇ α u d x = − 1 ∣ Q ∣ ∫ R d ( ∇ α η ) u d x + ∫ Q ( 1 − η ) ∇ α u d x . \int_Q\nabla^\alpha u\,\mathrm{d}x=\int_Q\eta\nabla^\alpha u\,\mathrm{d}x+\int_Q(1-\eta)\nabla^\alpha u\,\mathrm{d}x=-\frac{1}{|Q|}\int_{\mathbb{R}^d}(\nabla^\alpha\eta)u\,\mathrm{d}x+\int_Q(1-\eta)\nabla^\alpha u\,\mathrm{d}x. ∫ Q ∇ α u d x = ∫ Q η ∇ α u d x + ∫ Q ( 1 − η ) ∇ α u d x = − ∣ Q ∣ 1 ∫ R d ( ∇ α η ) u d x + ∫ Q ( 1 − η ) ∇ α u d x .
Set E : = Q ∖ ( 1 − a ) Q E:=Q\setminus(1-a)Q E := Q ∖ ( 1 − a ) Q . With Hölder’s inequality we derive
∣ ∫ Q ∇ α u d x ∣ ≤ 1 ∣ Q ∣ ∫ R d ∣ ∇ α η ∣ ∣ u ∣ d x + C ∫ Q 1 E ∣ ∇ α u ∣ d x ≤ 1 ∣ Q ∣ ∫ R d ∣ ∇ α η ∣ ∣ u ∣ d x + C ( ∣ E ∣ ∣ Q ∣ ) 1 2 ( ∫ Q ∣ ∇ α u ∣ 2 d x ) 1 2 . \begin{aligned}
\left|\int_Q\nabla^\alpha u\,\mathrm{d}x\right|\le\frac{1}{|Q|}\int_{\mathbb{R}^d}|\nabla^\alpha\eta||u|\,\mathrm{d}x+C\int_Q1_E|\nabla^\alpha u|\,\mathrm{d}x \\
&\le\frac{1}{|Q|}\int_{\mathbb{R}^d}|\nabla^\alpha\eta||u|\,\mathrm{d}x+C\left(\frac{|E|}{|Q|}\right)^{\frac12}\left(\int_Q|\nabla^\alpha u|^2\,\mathrm{d}x\right)^{\frac12}.
\end{aligned} ∫ Q ∇ α u d x ≤ ∣ Q ∣ 1 ∫ R d ∣ ∇ α η ∣∣ u ∣ d x + C ∫ Q 1 E ∣ ∇ α u ∣ d x ≤ ∣ Q ∣ 1 ∫ R d ∣ ∇ α η ∣∣ u ∣ d x + C ( ∣ Q ∣ ∣ E ∣ ) 2 1 ( ∫ Q ∣ ∇ α u ∣ 2 d x ) 2 1 .
For the first term on the right-hand side, we apply Lemma 2.10, see also Remark 2.11, to obtain
∫ R d ∣ ∇ α η ∣ ∣ u ∣ d x = ∑ k = 0 ∞ ∫ C k ( Q ) ∣ ∇ α η ∣ ∣ u ∣ d x ≤ C ∑ k = 0 ∞ 2 − ( d + α 0 ) k ( a ℓ ) α ∫ 2 k + 1 Q ∣ u ∣ d x , \int_{\mathbb{R}^d}|\nabla^\alpha\eta||u|\,\mathrm{d}x=\sum_{k=0}^{\infty}\int_{C_k(Q)}|\nabla^\alpha\eta||u|\,\mathrm{d}x\le C\sum_{k=0}^{\infty}\frac{2^{-(d+\alpha_0)k}}{(a\ell)^\alpha}\int_{2^{k+1}Q}|u|\,\mathrm{d}x, ∫ R d ∣ ∇ α η ∣∣ u ∣ d x = k = 0 ∑ ∞ ∫ C k ( Q ) ∣ ∇ α η ∣∣ u ∣ d x ≤ C k = 0 ∑ ∞ ( a ℓ ) α 2 − ( d + α 0 ) k ∫ 2 k + 1 Q ∣ u ∣ d x ,
which leads to the estimate
∣ ∫ Q ∇ α u d x ∣ ≤ C ( a ℓ ) α ∑ k = 0 ∞ 2 − α 0 k ∫ 2 k + 1 Q ∣ u ∣ d x + C ( ∣ E ∣ ∣ Q ∣ ) 1 2 ( ∫ Q ∣ ∇ α u ∣ 2 d x ) 1 2 . \left|\int_Q\nabla^\alpha u\,\mathrm{d}x\right|\le\frac{C}{(a\ell)^\alpha}\sum_{k=0}^{\infty}2^{-\alpha_0k}\int_{2^{k+1}Q}|u|\,\mathrm{d}x+C\left(\frac{|E|}{|Q|}\right)^{\frac12}\left(\int_Q|\nabla^\alpha u|^2\,\mathrm{d}x\right)^{\frac12}. ∫ Q ∇ α u d x ≤ ( a ℓ ) α C k = 0 ∑ ∞ 2 − α 0 k ∫ 2 k + 1 Q ∣ u ∣ d x + C ( ∣ Q ∣ ∣ E ∣ ) 2 1 ( ∫ Q ∣ ∇ α u ∣ 2 d x ) 2 1 .
Squaring both sides, estimating ∣ E ∣ ∣ Q ∣ = ( 1 − ( 1 − a ) d ) ≤ a d ≤ a α d \frac{|E|}{|Q|}=(1-(1-a)^d)\le ad\le a^\alpha d ∣ Q ∣ ∣ E ∣ = ( 1 − ( 1 − a ) d ) ≤ a d ≤ a α d by Bernoulli’s inequality, and using Hölder’s inequality for 0 < γ < α 0 0<\gamma<\alpha_0 0 < γ < α 0 , we arrive at
∣ ∫ Q ∇ α u d x ∣ 2 ≤ C ( X a 2 α ℓ 2 α + a α Y ) \left|\int_Q\nabla^\alpha u\,\mathrm{d}x\right|^2\le C\left(\frac{X}{a^{2\alpha}\ell^{2\alpha}}+a^\alpha Y\right) ∫ Q ∇ α u d x 2 ≤ C ( a 2 α ℓ 2 α X + a α Y )
for some constant C > 0 C>0 C > 0 depending on d d d and α 0 \alpha_0 α 0 , where
X : = ∑ k = 0 ∞ 2 − γ k ∫ 2 k + 1 Q ∣ u ∣ 2 d x and Y : = ∫ Q ∣ ∇ α u ∣ 2 d x . X:=\sum_{k=0}^{\infty}2^{-\gamma k}\int_{2^{k+1}Q}|u|^2\,\mathrm{d}x
\qquad\text{and}\qquad
Y:=\int_Q|\nabla^\alpha u|^2\,\mathrm{d}x. X := k = 0 ∑ ∞ 2 − γ k ∫ 2 k + 1 Q ∣ u ∣ 2 d x and Y := ∫ Q ∣ ∇ α u ∣ 2 d x .
Optimizing the inequality in the parameter a a a leads to the value
a 0 : = ( 2 X ℓ 2 α Y ) 1 3 α . a_0 := \left(\frac{2X}{\ell^{2\alpha}Y}\right)^{\frac{1}{3\alpha}}. a 0 := ( ℓ 2 α Y 2 X ) 3 α 1 .
If a 0 < 1 a_0<1 a 0 < 1 , then we can choose a = a 0 a=a_0 a = a 0 so that the inequality becomes
\left|\fint_Q \nabla^\alpha u\,dx\right|^2 \le\frac{C}{\ell^{\frac{2\alpha}{3}}}X^{\frac{1}{3}}Y^{\frac{2}{3}},
which proves the claim. In the case a 0 ≥ 1 a_0\ge1 a 0 ≥ 1 , this condition reads
\fint_Q |\nabla^\alpha u|^2\,dx \le\frac{2}{\ell^{2\alpha}}\sum_{k=0}^{\infty}2^{-\gamma k}\int_{2^{k+1}Q}|u|^2\,dx.
Hence, using Jensen’s inequality, we directly calculate
\begin{aligned}
\left|\fint_Q \nabla^\alpha u\,dx\right|^2
\le\left(\fint_Q|\nabla^\alpha u|^2\,dx\right)^{\frac{1}{3}}
\left(\fint_Q|\nabla^\alpha u|^2\,dx\right)^{\frac{2}{3}} \\
&\le\frac{2^{\frac{1}{3}}}{\ell^{\frac{2\alpha}{3}}}
\left(\sum_{k=0}^{\infty}2^{-\gamma k}\int_{2^{k+1}Q}|u|^2\,dx\right)^{\frac{1}{3}}
\left(\fint_Q|\nabla^\alpha u|^2\,dx\right)^{\frac{2}{3}},
\end{aligned}
so that the claim follows once again. □ \square □
Proposition 4.14. For every u ∈ S α , 2 ( R d ) u\in\mathrm{S}^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) we have
∫ 0 ∞ ∥ A t ( 1 − P t ) ∇ α u ∥ L 2 2 d t t ≲ ∥ ∇ α u ∥ L 2 2 . (24) \int_0^\infty\|\mathcal{A}_t(1-P_t)\nabla^\alpha u\|_{L^2}^2\,\frac{dt}{t}\lesssim\|\nabla^\alpha u\|_{L^2}^2. \tag*{(24)} ∫ 0 ∞ ∥ A t ( 1 − P t ) ∇ α u ∥ L 2 2 t d t ≲ ∥ ∇ α u ∥ L 2 2 . (24) Furthermore, we have the same bounds for γ t ⋅ A t ( 1 − P t ) \gamma_t\cdot\mathcal{A}_t(1-P_t) γ t ⋅ A t ( 1 − P t ) .
Proof. We follow the same line of reasoning as in [39 Prop. 13.13] , see also [14 Prop. 5.7] . Since A t 2 = A t \mathcal{A}_t^2=\mathcal{A}_t A t 2 = A t and γ t ⋅ A t \gamma_t\cdot\mathcal{A}_t γ t ⋅ A t is uniformly L 2 L^2 L 2 -bounded by Lemma 4.7, it suffices to prove (4.11).
Step 1: Reduction to a Schur-type bound. In this step, we show that (4.11) follows from
∥ A s ( 1 − P s ) Q t f ∥ L 2 ≲ ζ ( s t − 1 ) ∥ f ∥ L 2 ( s , t > 0 , f ∈ L 2 ( R d ) ) , (25) \|\mathcal{A}_s(1-P_s)Q_tf\|_{L^2}\lesssim\zeta(st^{-1})\|f\|_{L^2}\qquad(s,t>0,\ f\in L^2(\mathbb{R}^d)), \tag*{(25)} ∥ A s ( 1 − P s ) Q t f ∥ L 2 ≲ ζ ( s t − 1 ) ∥ f ∥ L 2 ( s , t > 0 , f ∈ L 2 ( R d )) , (25)
where ζ ( τ ) : = min ( τ α , τ − α + τ − α 3 ) ∈ L 1 ( ( 0 , ∞ ) ; d τ τ ) \zeta(\tau):=\min(\tau^\alpha,\tau^{-\alpha}+\tau^{-\frac{\alpha}{3}})\in L^1((0,\infty);\frac{d\tau}{\tau}) ζ ( τ ) := min ( τ α , τ − α + τ − 3 α ) ∈ L 1 (( 0 , ∞ ) ; τ d τ ) . In fact, by Lemma 4.10 we have
A s ( 1 − P s ) ∇ α u = 2 α ∫ 0 ∞ A s ( 1 − P s ) Q t Q t ∗ ∇ α u d t t . \mathcal{A}_s(1-P_s)\nabla^\alpha u
=2\alpha\int_0^\infty\mathcal{A}_s(1-P_s)Q_tQ_t^*\nabla^\alpha u\,\frac{dt}{t}. A s ( 1 − P s ) ∇ α u = 2 α ∫ 0 ∞ A s ( 1 − P s ) Q t Q t ∗ ∇ α u t d t .
Combining this representation with (4.12) yields the estimate
∫ 0 ∞ ∥ A s ( 1 − P s ) ∇ α u ∥ L 2 2 d s s ≲ ∫ 0 ∞ ( ∫ 0 ∞ ∥ A s ( 1 − P s ) Q t Q t ∗ ∇ α u ∥ L 2 d t t ) 2 d s s \int_0^\infty\|\mathcal{A}_s(1-P_s)\nabla^\alpha u\|_{L^2}^2\,\frac{ds}{s}
\lesssim\int_0^\infty\left(\int_0^\infty
\|\mathcal{A}_s(1-P_s)Q_tQ_t^*\nabla^\alpha u\|_{L^2}\,\frac{dt}{t}\right)^2\frac{ds}{s} ∫ 0 ∞ ∥ A s ( 1 − P s ) ∇ α u ∥ L 2 2 s d s ≲ ∫ 0 ∞ ( ∫ 0 ∞ ∥ A s ( 1 − P s ) Q t Q t ∗ ∇ α u ∥ L 2 t d t ) 2 s d s
≲ ∫ 0 ∞ ( ∫ 0 ∞ ζ ( s t − 1 ) ∥ Q t ∗ ∇ α u ∥ L 2 d t t ) 2 d s s . \lesssim\int_0^\infty\left(\int_0^\infty
\zeta(st^{-1})\|Q_t^*\nabla^\alpha u\|_{L^2}\,\frac{dt}{t}\right)^2\frac{ds}{s}. ≲ ∫ 0 ∞ ( ∫ 0 ∞ ζ ( s t − 1 ) ∥ Q t ∗ ∇ α u ∥ L 2 t d t ) 2 s d s .
Finally, using Hölder’s inequality, the substitution τ = s t − 1 \tau=st^{-1} τ = s t − 1 and Fubini’s theorem gives us
∫ 0 ∞ ∥ A s ( 1 − P s ) ∇ α u ∥ L 2 2 d s s ≲ ∫ 0 ∞ ( ∫ 0 ∞ ζ ( s t − 1 ) d t t ) ( ∫ 0 ∞ ζ ( s t − 1 ) ∥ Q t ∗ ∇ α u ∥ L 2 2 d t t ) d s s = ( ∫ 0 ∞ ζ ( τ ) d τ τ ) ( ∫ 0 ∞ ζ ( s t − 1 ) ∥ Q t ∗ ∇ α u ∥ L 2 2 d t t ) d s s = ( ∫ 0 ∞ ζ ( τ ) d τ τ ) 2 ( ∫ 0 ∞ ∥ Q t ∗ ∇ α u ∥ L 2 2 d t t ) . \begin{aligned}
\int_{0}^{\infty}\|\mathcal{A}_{s}(1-P_{s})\nabla^{\alpha}u\|_{L^{2}}^{2}\,\frac{\mathrm{d}s}{s}
&\lesssim\int_{0}^{\infty}\left(\int_{0}^{\infty}\zeta(st^{-1})\,\frac{\mathrm{d}t}{t}\right)\left(\int_{0}^{\infty}\zeta(st^{-1})\|Q_{t}^{*}\nabla^{\alpha}u\|_{L^{2}}^{2}\,\frac{\mathrm{d}t}{t}\right)\frac{\mathrm{d}s}{s}\\
&=\left(\int_{0}^{\infty}\zeta(\tau)\,\frac{\mathrm{d}\tau}{\tau}\right)\left(\int_{0}^{\infty}\zeta(st^{-1})\|Q_{t}^{*}\nabla^{\alpha}u\|_{L^{2}}^{2}\,\frac{\mathrm{d}t}{t}\right)\frac{\mathrm{d}s}{s}\\
&=\left(\int_{0}^{\infty}\zeta(\tau)\,\frac{\mathrm{d}\tau}{\tau}\right)^{2}\left(\int_{0}^{\infty}\|Q_{t}^{*}\nabla^{\alpha}u\|_{L^{2}}^{2}\,\frac{\mathrm{d}t}{t}\right).
\end{aligned} ∫ 0 ∞ ∥ A s ( 1 − P s ) ∇ α u ∥ L 2 2 s d s ≲ ∫ 0 ∞ ( ∫ 0 ∞ ζ ( s t − 1 ) t d t ) ( ∫ 0 ∞ ζ ( s t − 1 ) ∥ Q t ∗ ∇ α u ∥ L 2 2 t d t ) s d s = ( ∫ 0 ∞ ζ ( τ ) τ d τ ) ( ∫ 0 ∞ ζ ( s t − 1 ) ∥ Q t ∗ ∇ α u ∥ L 2 2 t d t ) s d s = ( ∫ 0 ∞ ζ ( τ ) τ d τ ) 2 ( ∫ 0 ∞ ∥ Q t ∗ ∇ α u ∥ L 2 2 t d t ) .
Using the quadratic estimates from Lemma 4.10 and the bound
∫ 0 ∞ ζ ( τ ) d τ τ ≤ ∫ 0 1 τ α d τ τ + ∫ 1 ∞ τ − α + τ − α 3 d τ τ ≤ 5 α ≤ 5 α 0 , \int_{0}^{\infty}\zeta(\tau)\,\frac{\mathrm{d}\tau}{\tau}\leq\int_{0}^{1}\tau^{\alpha}\,\frac{\mathrm{d}\tau}{\tau}+\int_{1}^{\infty}\tau^{-\alpha}+\tau^{-\frac{\alpha}{3}}\,\frac{\mathrm{d}\tau}{\tau}\leq\frac{5}{\alpha}\leq\frac{5}{\alpha_{0}}, ∫ 0 ∞ ζ ( τ ) τ d τ ≤ ∫ 0 1 τ α τ d τ + ∫ 1 ∞ τ − α + τ − 3 α τ d τ ≤ α 5 ≤ α 0 5 ,
we conclude (24) from the inequality above.
Step 2: Show (25) . Using the Fourier description from Proposition 2.3 , we obtain the identities
P s Q t = F − 1 ( 1 1 + s 2 α ∣ ξ ∣ 2 α ∣ ξ ∣ 1 − α ( 1 + t 2 α ∣ ξ ∣ 2 α ) i t α ξ s α F ) = t α s α P t Q s P_{s}Q_{t}=\mathcal{F}^{-1}\left(\frac{1}{1+s^{2\alpha}|\xi|^{2\alpha}|\xi|^{1-\alpha}(1+t^{2\alpha}|\xi|^{2\alpha})}\,\frac{it^{\alpha}\xi}{s^{\alpha}}\mathcal{F}\right)=\frac{t^{\alpha}}{s^{\alpha}}P_{t}Q_{s} P s Q t = F − 1 ( 1 + s 2 α ∣ ξ ∣ 2 α ∣ ξ ∣ 1 − α ( 1 + t 2 α ∣ ξ ∣ 2 α ) 1 s α i t α ξ F ) = s α t α P t Q s
and
( 1 − P s ) Q t = F − 1 ( s 2 α ∣ ξ ∣ 2 α 1 + s 2 α ∣ ξ ∣ 2 α ∣ ξ ∣ 1 − α ( 1 + t 2 α ∣ ξ ∣ 2 α ) i t α ξ s α F ) = s α t α ( 1 − P t ) Q s . (1-P_{s})Q_{t}=\mathcal{F}^{-1}\left(\frac{s^{2\alpha}|\xi|^{2\alpha}}{1+s^{2\alpha}|\xi|^{2\alpha}|\xi|^{1-\alpha}(1+t^{2\alpha}|\xi|^{2\alpha})}\,\frac{it^{\alpha}\xi}{s^{\alpha}}\mathcal{F}\right)=\frac{s^{\alpha}}{t^{\alpha}}(1-P_{t})Q_{s}. ( 1 − P s ) Q t = F − 1 ( 1 + s 2 α ∣ ξ ∣ 2 α ∣ ξ ∣ 1 − α ( 1 + t 2 α ∣ ξ ∣ 2 α ) s 2 α ∣ ξ ∣ 2 α s α i t α ξ F ) = t α s α ( 1 − P t ) Q s .
When s ≤ t s\leq t s ≤ t we obtain (25) by uniform L 2 L^{2} L 2 -bounds as follows
∥ A s ( 1 − P s ) Q t f ∥ L 2 ≲ ∥ ( 1 − P s ) Q t f ∥ L 2 = s α t α ∥ ( 1 − P t ) Q s f ∥ L 2 ≲ s α t α ∥ f ∥ L 2 . \|\mathcal{A}_{s}(1-P_{s})Q_{t}f\|_{L^{2}}\lesssim\|(1-P_{s})Q_{t}f\|_{L^{2}}=\frac{s^{\alpha}}{t^{\alpha}}\|(1-P_{t})Q_{s}f\|_{L^{2}}\lesssim\frac{s^{\alpha}}{t^{\alpha}}\|f\|_{L^{2}}. ∥ A s ( 1 − P s ) Q t f ∥ L 2 ≲ ∥ ( 1 − P s ) Q t f ∥ L 2 = t α s α ∥ ( 1 − P t ) Q s f ∥ L 2 ≲ t α s α ∥ f ∥ L 2 .
When t ≤ s t\leq s t ≤ s , we first estimate
∥ A s ( 1 − P s ) Q t f ∥ L 2 ≲ ∥ A s Q t f ∥ L 2 + ∥ P s Q t f ∥ L 2 = ∥ A s Q t f ∥ L 2 + t α s α ∥ P t Q s f ∥ L 2 (26) \begin{aligned}
\|\mathcal{A}_{s}(1-P_{s})Q_{t}f\|_{L^{2}}\lesssim\|\mathcal{A}_{s}Q_{t}f\|_{L^{2}}+\|P_{s}Q_{t}f\|_{L^{2}} \\
&=\|\mathcal{A}_{s}Q_{t}f\|_{L^{2}}+\frac{t^{\alpha}}{s^{\alpha}}\|P_{t}Q_{s}f\|_{L^{2}} \tag*{(26)}
\end{aligned} ∥ A s ( 1 − P s ) Q t f ∥ L 2 ≲ ∥ A s Q t f ∥ L 2 + ∥ P s Q t f ∥ L 2 = ∥ A s Q t f ∥ L 2 + s α t α ∥ P t Q s f ∥ L 2 (26)
≲ ∥ A s Q t f ∥ L 2 + t α s α ∥ f ∥ L 2 . \lesssim\|\mathcal{A}_{s}Q_{t}f\|_{L^{2}}+\frac{t^{\alpha}}{s^{\alpha}}\|f\|_{L^{2}}. ≲ ∥ A s Q t f ∥ L 2 + s α t α ∥ f ∥ L 2 .
To control the term ∥ A s Q t f ∥ L 2 \|\mathcal{A}_{s}Q_{t}f\|_{L^{2}} ∥ A s Q t f ∥ L 2 on the right-hand side, we first decompose the L 2 L^{2} L 2 -norm by dyadic cubes to get
\|\mathcal{A}_{s}Q_{t}f\|_{L^{2}}^{2}=\sum_{Q\in\mathcal{D}_{s}}|Q|\left|\fint_{Q}Q_{t}f\,\mathrm{d}x\right|^{2}=\sum_{Q\in\mathcal{D}_{s}}|Q|\left|\fint_{Q}t^{\alpha}\nabla^{\alpha}P_{t}f\,\mathrm{d}x\right|^{2}.
For γ = α 0 2 \gamma=\frac{\alpha_{0}}{2} γ = 2 α 0 we apply Lemma 4.13 and Hölder’s inequality to get
∥ A s Q t f ∥ L 2 2 ≲ ∑ Q ∈ D s ∣ Q ∣ s 2 α 3 ( ∑ k = 0 ∞ 2 − α 0 2 k ∫ 2 k + 1 Q ∣ t α P t f ∣ 2 d x ) 1 3 ( ∫ Q ∣ t α ∇ α P t f ∣ 2 d x ) 2 3 ≤ 1 s 2 α 3 ( ∑ Q ∈ □ s ∑ k = 0 ∞ 2 − ( d + α 0 2 ) k ∫ 2 k Q ∣ t α P t f ∣ 2 d x ) 1 3 ( ∑ Q ∈ □ s ∫ Q ∣ t α ∇ α P t f ∣ 2 d x ) 2 3 . \begin{aligned}
\|\mathcal{A}_{s}Q_{t}f\|_{L^{2}}^{2}\lesssim\sum_{Q\in\mathcal{D}_{s}}\frac{|Q|}{s^{\frac{2\alpha}{3}}}\left(\sum_{k=0}^{\infty}2^{-\frac{\alpha_{0}}{2}k}\int_{2^{k+1}Q}|t^{\alpha}P_{t}f|^{2}\,\mathrm{d}x\right)^{\frac{1}{3}}\left(\int_{Q}|t^{\alpha}\nabla^{\alpha}P_{t}f|^{2}\,\mathrm{d}x\right)^{\frac{2}{3}} \\
&\leq\frac{1}{s^{\frac{2\alpha}{3}}}\left(\sum_{Q\in\Box_s}\sum_{k=0}^{\infty}2^{-(d+\frac{\alpha_0}{2})k}\int_{2^kQ}|t^\alpha P_tf|^2\,\mathrm{d}x\right)^{\frac{1}{3}}\left(\sum_{Q\in\Box_s}\int_Q|t^\alpha\nabla^\alpha P_tf|^2\,\mathrm{d}x\right)^{\frac{2}{3}}.
\end{aligned} ∥ A s Q t f ∥ L 2 2 ≲ Q ∈ D s ∑ s 3 2 α ∣ Q ∣ ( k = 0 ∑ ∞ 2 − 2 α 0 k ∫ 2 k + 1 Q ∣ t α P t f ∣ 2 d x ) 3 1 ( ∫ Q ∣ t α ∇ α P t f ∣ 2 d x ) 3 2 ≤ s 3 2 α 1 Q ∈ □ s ∑ k = 0 ∑ ∞ 2 − ( d + 2 α 0 ) k ∫ 2 k Q ∣ t α P t f ∣ 2 d x 3 1 Q ∈ □ s ∑ ∫ Q ∣ t α ∇ α P t f ∣ 2 d x 3 2 .
With the observation ∑ Q ∈ □ s 1 2 k Q ( x ) ≲ 2 k d \sum_{Q\in\Box_s}\mathbf{1}_{2^kQ}(x)\lesssim2^{kd} ∑ Q ∈ □ s 1 2 k Q ( x ) ≲ 2 k d for all k ∈ N 0 k\in\mathbb{N}_0 k ∈ N 0 , we conclude
∥ A s Q t f ∥ L 2 2 ≲ 1 s 2 α 3 ( ∑ k = 0 ∞ 2 − α 0 2 k ∫ R d ∣ t α P t f ∣ 2 d x ) 1 3 ( ∫ R d ∣ t α ∇ α P t f ∣ 2 d x ) 2 3 \|A_sQ_tf\|_{L^2}^2\lesssim\frac{1}{s^{\frac{2\alpha}{3}}}\left(\sum_{k=0}^{\infty}2^{-\frac{\alpha_0}{2}k}\int_{\mathbb{R}^d}|t^\alpha P_tf|^2\,\mathrm{d}x\right)^{\frac{1}{3}}\left(\int_{\mathbb{R}^d}|t^\alpha\nabla^\alpha P_tf|^2\,\mathrm{d}x\right)^{\frac{2}{3}} ∥ A s Q t f ∥ L 2 2 ≲ s 3 2 α 1 ( k = 0 ∑ ∞ 2 − 2 α 0 k ∫ R d ∣ t α P t f ∣ 2 d x ) 3 1 ( ∫ R d ∣ t α ∇ α P t f ∣ 2 d x ) 3 2
≲ ( t s ) 2 α 3 ∥ P t f ∥ L 2 2 3 ∥ Q t f ∥ L 2 4 3 \lesssim\left(\frac{t}{s}\right)^{\frac{2\alpha}{3}}\|P_tf\|_{L^2}^{\frac{2}{3}}\|Q_tf\|_{L^2}^{\frac{4}{3}} ≲ ( s t ) 3 2 α ∥ P t f ∥ L 2 3 2 ∥ Q t f ∥ L 2 3 4
≲ ( t s ) 2 α 3 ∥ f ∥ L 2 2 . \lesssim\left(\frac{t}{s}\right)^{\frac{2\alpha}{3}}\|f\|_{L^2}^2. ≲ ( s t ) 3 2 α ∥ f ∥ L 2 2 .
This together with (26) completes the proof of (25) in the case t ≤ s t\leq s t ≤ s . □
Combining Propositions 4.11, 4.12 and 4.14 with the splitting (23) establishes the full principal part approximation.
Proposition 4.15 (Principal part approximation). For all u ∈ S α , 2 ( R d ) u\in\mathrm{S}^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) we have
∫ 0 ∞ ∥ ( Θ t − γ t ⋅ A t ) ∇ α u ∥ L 2 2 d t t ≲ ∥ ∇ α u ∥ L 2 2 . \int_0^\infty\|(\Theta_t-\gamma_t\cdot A_t)\nabla_\alpha u\|_{L^2}^2\frac{\mathrm{d}t}{t}\lesssim\|\nabla_\alpha u\|_{L^2}^2. ∫ 0 ∞ ∥ ( Θ t − γ t ⋅ A t ) ∇ α u ∥ L 2 2 t d t ≲ ∥ ∇ α u ∥ L 2 2 .
The Carleson measure argument
By Proposition 4.15 and Proposition 4.6, it remains to show that ∣ γ t ( x ) ∣ 2 d x d t t |\gamma_t(x)|^2\frac{\mathrm{d}x\mathrm{d}t}{t} ∣ γ t ( x ) ∣ 2 t d x d t is a Carleson measure. Following the usual procedure, we will first establish a directional Carleson measure estimate, where we restrict the values of the principal part to a cone in C d \mathbb{C}^d C d .
Definition 4.16. Let ε > 0 \varepsilon>0 ε > 0 . For ξ ∈ C d \xi\in\mathbb{C}^d ξ ∈ C d with ∣ ξ ∣ = 1 |\xi|=1 ∣ ξ ∣ = 1 , define the open cone with central axis ξ \xi ξ by
Γ ξ ε : = { w ∈ C d ∖ { 0 } : ∣ w ∣ w ∣ − ξ ∣ < ε } . \Gamma_\xi^\varepsilon:=\left\{w\in\mathbb{C}^d\setminus\{0\}:\left|\frac{w}{|w|}-\xi\right|<\varepsilon\right\}. Γ ξ ε := { w ∈ C d ∖ { 0 } : ∣ w ∣ w − ξ < ε } .
Proposition 4.17 (Directional Carleson measure estimate). There is a choice of ε > 0 \varepsilon>0 ε > 0 , depending only on d d d , α 0 \alpha_0 α 0 , μ ∙ \mu_\bullet μ ∙ and μ ∗ \mu^\ast μ ∗ , such that for every ξ ∈ C d \xi\in\mathbb{C}^d ξ ∈ C d with ∣ ξ ∣ = 1 |\xi|=1 ∣ ξ ∣ = 1 we have
∬ R ( Q ) ∣ γ t ( x ) ∣ 2 1 Γ ξ ε ( γ t ( x ) ) d x d t t ≲ ∣ Q ∣ ( Q ∈ □ ) . \iint_{R(Q)}|\gamma_t(x)|^2\mathbf{1}_{\Gamma_\xi^\varepsilon}(\gamma_t(x))\frac{\mathrm{d}x\mathrm{d}t}{t}\lesssim|Q|\qquad(Q\in\Box). ∬ R ( Q ) ∣ γ t ( x ) ∣ 2 1 Γ ξ ε ( γ t ( x )) t d x d t ≲ ∣ Q ∣ ( Q ∈ □ ) .
Taking the directional Carleson measure estimate from Proposition 4.17 for granted, we are ready to prove Theorem 1.2.
Proof of Theorem 1.2. Let ε > 0 \varepsilon>0 ε > 0 be as in Proposition 4.17. Then we cover C d ∖ { 0 } \mathbb{C}^d\setminus\{0\} C d ∖ { 0 } by finitely many cones Γ ξ 1 ε , … , Γ ξ N ε \Gamma_{\xi_1}^\varepsilon,\ldots,\Gamma_{\xi_N}^\varepsilon Γ ξ 1 ε , … , Γ ξ N ε , where N N N depends only on d d d and ε \varepsilon ε . The estimate from Proposition 4.17 then yields
∬ R ( Q ) ∣ γ t ( x ) ∣ 2 d x d t t ≤ ∬ R ( Q ) ∑ j = 1 N ∣ γ t ( x ) ∣ 2 1 Γ ξ j ε ( γ t ( x ) ) d x d t t ≲ N ∣ Q ∣ \iint_{R(Q)}|\gamma_t(x)|^2\frac{\mathrm{d}x\mathrm{d}t}{t}\leq\iint_{R(Q)}\sum_{j=1}^N|\gamma_t(x)|^2\mathbf{1}_{\Gamma_{\xi_j}^\varepsilon}(\gamma_t(x))\frac{\mathrm{d}x\mathrm{d}t}{t}\lesssim N|Q| ∬ R ( Q ) ∣ γ t ( x ) ∣ 2 t d x d t ≤ ∬ R ( Q ) j = 1 ∑ N ∣ γ t ( x ) ∣ 2 1 Γ ξ j ε ( γ t ( x )) t d x d t ≲ N ∣ Q ∣
for every cube Q ∈ □ Q\in\Box Q ∈ □ as required. Thus, ∣ γ t ( x ) ∣ 2 d x d t t |\gamma_t(x)|^2\frac{\mathrm{d}x\mathrm{d}t}{t} ∣ γ t ( x ) ∣ 2 t d x d t defines a Carleson measure by Definition 4.5. Combining this with Propositions 4.15 and 4.6, we obtain (4.3), which proves the claim; see Section 4.1. □
The remainder of this subsection is devoted to the proof of Proposition 4.17 , which is based on the construction of a T ( b ) T(b) T ( b ) -test function. More precisely, for given Q ∈ □ Q \in\square Q ∈ □ and ξ ∈ C d \xi\in\mathbb{C}^{d} ξ ∈ C d with ∣ ξ ∣ = 1 |\xi| = 1 ∣ ξ ∣ = 1 , we strive for a construction of functions b Q ε b_{Q}^{\varepsilon} b Q ε such that
∣ w ∣ ≲ ∣ w ⋅ ( A t b Q ε ) ( x ) ∣ ( w ∈ Γ ξ ε ) |w| \lesssim\left|w \cdot(\mathcal{A}_{t}b_{Q}^{\varepsilon})(x)\right| \qquad(w \in\Gamma_{\xi}^{\varepsilon}) ∣ w ∣ ≲ w ⋅ ( A t b Q ε ) ( x ) ( w ∈ Γ ξ ε )
for a hopefully large set of ( t , x ) ∈ R ( Q ) (t,x) \in R(Q) ( t , x ) ∈ R ( Q ) . The test function b Q ε b_{Q}^{\varepsilon} b Q ε should have the following properties:
It should belong to the range of ∇ α \nabla^{\alpha} ∇ α to make the principal part approximation accessible.
It should satisfy the L 2 L^{2} L 2 -bound ∥ b Q ε ∥ L 2 ≲ ∣ Q ∣ 1 2 \|b_{Q}^{\varepsilon}\|_{L^{2}} \lesssim|Q|^{\frac{1}{2}} ∥ b Q ε ∥ L 2 ≲ ∣ Q ∣ 2 1 .
It should point roughly in the direction of ξ ˉ ∈ C d \bar{\xi} \in\mathbb{C}^{d} ξ ˉ ∈ C d on Q Q Q .
This will be achieved by the ansatz
b Q ε ≃ ∇ α ( 1 + ε 2 α ℓ ( Q ) 2 α L ) − 1 Φ . b_{Q}^{\varepsilon} \simeq\nabla^{\alpha}\left(1+\varepsilon^{2\alpha}\ell(Q)^{2\alpha}L\right)^{-1}\Phi. b Q ε ≃ ∇ α ( 1 + ε 2 α ℓ ( Q ) 2 α L ) − 1 Φ.
By construction, this function satisfies the first point. Moreover, since ( 1 + ε 2 α ℓ ( Q ) 2 α L ) − 1 \left(1+\varepsilon^{2\alpha}\ell(Q)^{2\alpha}L\right)^{-1} ( 1 + ε 2 α ℓ ( Q ) 2 α L ) − 1 is an approximation of the identity for small ε \varepsilon ε , properties (2) and (3) hold roughly if Φ \Phi Φ is chosen such that ∥ ∇ α Φ ∥ L 2 ≲ ∣ Q ∣ 1 2 \|\nabla^{\alpha}\Phi\|_{L^{2}} \lesssim|Q|^{\frac{1}{2}} ∥ ∇ α Φ ∥ L 2 ≲ ∣ Q ∣ 2 1 and ∇ α Φ = ξ ˉ \nabla^{\alpha}\Phi=\bar{\xi} ∇ α Φ = ξ ˉ on Q Q Q . As explained in the introduction, the elliptic construction of Φ \Phi Φ as an affine function times a cut-off function is not admissible in our fractional situation. Thus, we instead use the factorization ∇ α = ∇ I 1 − α \nabla^{\alpha}=\nabla I_{1-\alpha} ∇ α = ∇ I 1 − α . More precisely, composing with the fractional Laplacian restores the local gradient
∇ α ( − Δ ) 1 − α 2 = ∇ I 1 − α ( − Δ ) 1 − α 2 = ∇ , \nabla^{\alpha}(-\Delta)^{\frac{1-\alpha}{2}}=\nabla I_{1-\alpha}(-\Delta)^{\frac{1-\alpha}{2}}=\nabla, ∇ α ( − Δ ) 2 1 − α = ∇ I 1 − α ( − Δ ) 2 1 − α = ∇ ,
see Proposition 2.7 . Therefore, we choose Φ = ( − Δ ) 1 − α 2 h \Phi=(-\Delta)^{\frac{1-\alpha}{2}}h Φ = ( − Δ ) 2 1 − α h , where h h h is the elliptic ansatz, so that ∇ α Φ = ∇ h \nabla^{\alpha}\Phi=\nabla h ∇ α Φ = ∇ h has the required properties. This will be made rigorous in the proof of the following proposition.
Proposition 4.18. There exists ε 0 ∈ ( 0 , 1 ] \varepsilon_{0}\in(0,1] ε 0 ∈ ( 0 , 1 ] depending only on d , α 0 , μ ∙ , μ ∙ d,\alpha_{0},\mu_{\bullet},\mu^{\bullet} d , α 0 , μ ∙ , μ ∙ such that for all ξ ∈ C d \xi\in\mathbb{C}^{d} ξ ∈ C d with ∣ ξ ∣ = 1 |\xi|=1 ∣ ξ ∣ = 1 , all ε ∈ ( 0 , ε 0 ] \varepsilon\in(0,\varepsilon_{0}] ε ∈ ( 0 , ε 0 ] and all cubes Q ∈ □ Q\in\square Q ∈ □ there exists b Q ε ∈ L 2 ( R d ; C d ) b_{Q}^{\varepsilon}\in L^{2}(\mathbb{R}^{d};\mathbb{C}^{d}) b Q ε ∈ L 2 ( R d ; C d ) with the following properties:
(B1) ∥ b Q ε ∥ L 2 ≲ ∣ Q ∣ 1 2 \|b_{Q}^{\varepsilon}\|_{L^{2}}\lesssim|Q|^{\frac{1}{2}} ∥ b Q ε ∥ L 2 ≲ ∣ Q ∣ 2 1 ,
(B2) Re ( ξ ˉ ⋅ ∫ Q b Q ε d x ) ≥ 1 \operatorname{Re}\left(\bar{\xi}\cdot\int_{Q}b_{Q}^{\varepsilon}\,dx\right)\geq1 Re ( ξ ˉ ⋅ ∫ Q b Q ε d x ) ≥ 1 ,
(B3) ∬ R ( Q ) ∣ ( γ t ( x ) ⋅ A t ) b Q ε ( x ) ∣ 2 d x d t t ≲ ∣ Q ∣ ε 2 \iint_{R(Q)}|(\gamma_{t}(x)\cdot\mathcal{A}_{t})b_{Q}^{\varepsilon}(x)|^{2}\,\frac{dx\,dt}{t}\lesssim\frac{|Q|}{\varepsilon^{2}} ∬ R ( Q ) ∣ ( γ t ( x ) ⋅ A t ) b Q ε ( x ) ∣ 2 t d x d t ≲ ε 2 ∣ Q ∣ .
Proof. We fix ξ ∈ C d \xi\in\mathbb{C}^{d} ξ ∈ C d with ∣ ξ ∣ = 1 |\xi|=1 ∣ ξ ∣ = 1 , Q ∈ □ Q\in\square Q ∈ □ and abbreviate by x Q ∈ R d x_{Q}\in\mathbb{R}^{d} x Q ∈ R d and ℓ : = ℓ ( Q ) \ell:=\ell(Q) ℓ := ℓ ( Q ) its center and side-length. Similarly, we simplify notation by omitting ε \varepsilon ε and Q Q Q when constructing the test function b = b Q ε b=b_{Q}^{\varepsilon} b = b Q ε . Fix η ∈ C c ∞ ( R d ) \eta\in C_{c}^{\infty}(\mathbb{R}^{d}) η ∈ C c ∞ ( R d ) with supp η ⊂ 2 Q \operatorname{supp}\eta\subset2Q supp η ⊂ 2 Q , η ≡ 1 \eta\equiv1 η ≡ 1 on Q Q Q and ∥ η ∥ L ∞ + ℓ ∥ ∇ η ∥ L ∞ ≲ 1 \|\eta\|_{L^{\infty}}+\ell\|\nabla\eta\|_{L^{\infty}}\lesssim1 ∥ η ∥ L ∞ + ℓ ∥∇ η ∥ L ∞ ≲ 1 . Define
h ( x ) : = ξ ˉ ⋅ ( x − x Q ) η ( x ) ∈ C c ∞ ( R d ) , h(x):=\bar{\xi}\cdot(x-x_{Q})\eta(x)\in C_{c}^{\infty}(\mathbb{R}^{d}), h ( x ) := ξ ˉ ⋅ ( x − x Q ) η ( x ) ∈ C c ∞ ( R d ) ,
and set Φ : = ( − Δ ) 1 − α 2 h \Phi:=(-\Delta)^{\frac{1-\alpha}{2}}h Φ := ( − Δ ) 2 1 − α h , which belongs to S α , 2 ( R d ) S^{\alpha,2}(\mathbb{R}^{d}) S α , 2 ( R d ) due to Proposition 2.7 . Finally, we define our test function
b : = 2 ∇ α ( 1 + ε 2 α ℓ 2 α L ) − 1 Φ . (27) b:=2\nabla^{\alpha}(1+\varepsilon^{2\alpha}\ell^{2\alpha}L)^{-1}\Phi. \tag*{(27)} b := 2 ∇ α ( 1 + ε 2 α ℓ 2 α L ) − 1 Φ. (27)
By construction, we have with (15)
1 2 b − ∇ α Φ = ∇ α ( ( 1 + ε 2 α ℓ 2 α L ) − 1 − 1 ) Φ = − ∇ α ( ε 2 α ℓ 2 α L ( 1 + ε 2 α ℓ 2 α L ) − 1 ) Φ = ε 2 α ℓ 2 α ∇ α ( 1 + ε 2 α ℓ 2 α L ) − 1 div α ( μ ∇ α Φ ) . (28) \begin{aligned}
\frac{1}{2}b-\nabla^\alpha\Phi
&=\nabla^\alpha\left((1+\varepsilon^{2\alpha}\ell^{2\alpha}L)^{-1}-1\right)\Phi\\
&=-\nabla^\alpha\left(\varepsilon^{2\alpha}\ell^{2\alpha}L(1+\varepsilon^{2\alpha}\ell^{2\alpha}L)^{-1}\right)\Phi\\
&=\varepsilon^{2\alpha}\ell^{2\alpha}\nabla^\alpha(1+\varepsilon^{2\alpha}\ell^{2\alpha}L)^{-1}\operatorname{div}^\alpha(\mu\nabla^\alpha\Phi). \tag*{(28)}
\end{aligned} 2 1 b − ∇ α Φ = ∇ α ( ( 1 + ε 2 α ℓ 2 α L ) − 1 − 1 ) Φ = − ∇ α ( ε 2 α ℓ 2 α L ( 1 + ε 2 α ℓ 2 α L ) − 1 ) Φ = ε 2 α ℓ 2 α ∇ α ( 1 + ε 2 α ℓ 2 α L ) − 1 div α ( μ ∇ α Φ ) . (28)
We are now ready to prove the properties (B1), (B2) and (B3).
Step 1: Show (B1). Using Proposition 2.7, we calculate
( ∇ α Φ ) ( x ) = ∇ α ( − Δ ) 1 − α 2 h ( x ) = ∇ h ( x ) = ξ ˉ η ( x ) + ξ ˉ ⋅ ( x − x Q ) ∇ η ( x ) . (29) (\nabla^\alpha\Phi)(x)=\nabla^\alpha(-\Delta)^{\frac{1-\alpha}{2}}h(x)=\nabla h(x)=\bar{\xi}\eta(x)+\bar{\xi}\cdot(x-x_Q)\nabla\eta(x). \tag*{(29)} ( ∇ α Φ ) ( x ) = ∇ α ( − Δ ) 2 1 − α h ( x ) = ∇ h ( x ) = ξ ˉ η ( x ) + ξ ˉ ⋅ ( x − x Q ) ∇ η ( x ) . (29)
Since ∣ x − x Q ∣ ≤ d ℓ |x-x_Q|\le\sqrt{d}\ell ∣ x − x Q ∣ ≤ d ℓ for every x ∈ 2 Q x\in2Q x ∈ 2 Q and ℓ ∥ ∇ η ∥ L ∞ ≲ 1 \ell\|\nabla\eta\|_{L^\infty}\lesssim1 ℓ ∥∇ η ∥ L ∞ ≲ 1 , we obtain the pointwise bound
∣ ( ∇ α Φ ) ( x ) ∣ ≤ ∣ ξ ˉ η ( x ) ∣ + ∣ ξ ˉ ⋅ ( x − x Q ) ∇ η ( x ) ∣ ≲ ( 1 + d ℓ ℓ ) 1 2 Q ( x ) , |(\nabla^\alpha\Phi)(x)|\le|\bar{\xi}\eta(x)|+|\bar{\xi}\cdot(x-x_Q)\nabla\eta(x)|\lesssim\left(1+\frac{\sqrt{d}\ell}{\ell}\right)1_{2Q}(x), ∣ ( ∇ α Φ ) ( x ) ∣ ≤ ∣ ξ ˉ η ( x ) ∣ + ∣ ξ ˉ ⋅ ( x − x Q ) ∇ η ( x ) ∣ ≲ ( 1 + ℓ d ℓ ) 1 2 Q ( x ) ,
and subsequently
∥ ∇ α Φ ∥ L 2 ≲ ∥ 1 2 Q ∥ L 2 ≲ ∣ Q ∣ 1 2 . (30) \|\nabla^\alpha\Phi\|_{L^2}\lesssim\|1_{2Q}\|_{L^2}\lesssim|Q|^{\frac{1}{2}}. \tag*{(30)} ∥ ∇ α Φ ∥ L 2 ≲ ∥ 1 2 Q ∥ L 2 ≲ ∣ Q ∣ 2 1 . (30)
Combining this estimate with the identity (28) and the uniform L 2 L^2 L 2 -bounds from Proposition 3.4, we get
∥ b − 2 ∇ α Φ ∥ L 2 ≲ ∥ μ ∇ α Φ ∥ L 2 ≲ ∣ Q ∣ 1 2 , \|b-2\nabla^\alpha\Phi\|_{L^2}\lesssim\|\mu\nabla^\alpha\Phi\|_{L^2}\lesssim|Q|^{\frac{1}{2}}, ∥ b − 2 ∇ α Φ ∥ L 2 ≲ ∥ μ ∇ α Φ ∥ L 2 ≲ ∣ Q ∣ 2 1 ,
which concludes (B1).
Step 2: Show (B2). We see from (29) that ∇ α Φ = ξ ˉ \nabla^\alpha\Phi=\bar{\xi} ∇ α Φ = ξ ˉ on Q Q Q . Hence, by (28) , we have
\fint_Q b-2\bar{\xi}\,dx=\fint_Q b-2\nabla^\alpha\Phi\,dx=2\varepsilon^{2\alpha}\ell^{2\alpha}\fint_Q\nabla^\alpha u\,dx, \tag*{(31)}
where we set u : = ( 1 + ε 2 α ℓ 2 α L ) − 1 div α ( μ ∇ α Φ ) u:=(1+\varepsilon^{2\alpha}\ell^{2\alpha}L)^{-1}\operatorname{div}^\alpha(\mu\nabla^\alpha\Phi) u := ( 1 + ε 2 α ℓ 2 α L ) − 1 div α ( μ ∇ α Φ ) . The uniform L 2 L^2 L 2 -bounds from Proposition 3.4 in combination with (30) give us
∥ u ∥ L 2 ≲ ∣ Q ∣ 1 2 ε α ℓ α and ∥ ∇ α u ∥ L 2 ≲ ∣ Q ∣ 1 2 ε 2 α ℓ 2 α , \|u\|_{L^2}\lesssim\frac{|Q|^{\frac{1}{2}}}{\varepsilon^\alpha\ell^\alpha}
\qquad\text{and}\qquad
\|\nabla^\alpha u\|_{L^2}\lesssim\frac{|Q|^{\frac{1}{2}}}{\varepsilon^{2\alpha}\ell^{2\alpha}}, ∥ u ∥ L 2 ≲ ε α ℓ α ∣ Q ∣ 2 1 and ∥ ∇ α u ∥ L 2 ≲ ε 2 α ℓ 2 α ∣ Q ∣ 2 1 ,
so that estimating the average on the right-hand side of (31) by means of Lemma 4.13 with γ = α 0 2 \gamma=\frac{\alpha_0}{2} γ = 2 α 0 leads us to
\begin{aligned}
\left|\fint_Q b-2\bar{\xi}\,dx\right|^2
\le\frac{4C\varepsilon^{4\alpha}\ell^{4\alpha}}{\ell^{\frac{2\alpha}{3}}}
\left(\sum_{k=0}^{\infty}2^{-\frac{\alpha_0}{2}k}\fint_{2^{k+1}Q}|u|^2\,dx\right)^{\frac{1}{3}}
\left(\fint_Q|\nabla^\alpha u|^2\,dx\right)^{\frac{2}{3}} \\
&\le\frac{4C\varepsilon^{4\alpha}\ell^{4\alpha}}{\ell^{\frac{2\alpha}{3}}|Q|}
\|u\|_{L^2}^{\frac{2}{3}}\|\nabla^\alpha u\|_{L^2}^{\frac{4}{3}} \\
&\le C\varepsilon^{\frac{2\alpha}{3}},
\end{aligned}
where C C C varies in every step. Since ∣ ξ ∣ = 1 |\xi|=1 ∣ ξ ∣ = 1 , we conclude that
\operatorname{Re}\left(\bar{\xi}\cdot\fint_Q b\,dx\right)
=2+\operatorname{Re}\left(\bar{\xi}\cdot\fint_Q b-2\bar{\xi}\,dx\right)
\ge2-\left|\fint_Q b-2\bar{\xi}\,dx\right|
≥ 2 − C ε 2 α 3 \ge2-\sqrt{C\varepsilon^{\frac{2\alpha}{3}}} ≥ 2 − C ε 3 2 α
≥ 2 − C ε 2 α 0 3 . \ge2-\sqrt{C\varepsilon^{\frac{2\alpha_0}{3}}}. ≥ 2 − C ε 3 2 α 0 .
Now, (B2) follows upon taking ε \varepsilon ε small enough.
Step 3: Show (B3). Recall from (4.14) that b ∈ R ( ∇ α ) b\in\mathcal{R}(\nabla^\alpha) b ∈ R ( ∇ α ) . Hence, we can apply the principal part approximation from Proposition 4.15 and use (B1) to bound
1 2 ∭ R ( Q ) ∣ ( γ t ( x ) ⋅ A t ) b ( x ) ∣ 2 d x d t t ≤ ∫ 0 ℓ ( ∥ ( γ t ⋅ A t − Θ t ) b ∥ L 2 2 + ∥ Θ t b ∥ L 2 2 ) d t t ≲ ∥ b ∥ L 2 2 + ∫ 0 ℓ ∥ Θ t b ∥ L 2 2 d t t ≲ ∣ Q ∣ + ∫ 0 ℓ ∥ Θ t b ∥ L 2 2 d t t . (32) \begin{aligned}
\frac{1}{2}\iiint_{R(Q)}\left|(\gamma_t(x)\cdot A_t)b(x)\right|^2\frac{\mathrm{d}x\mathrm{d}t}{t}
&\leq\int_0^\ell\left(\|(\gamma_t\cdot A_t-\Theta_t)b\|_{L^2}^2+\|\Theta_t b\|_{L^2}^2\right)\frac{\mathrm{d}t}{t}\\
&\lesssim\|b\|_{L^2}^2+\int_0^\ell\|\Theta_t b\|_{L^2}^2\frac{\mathrm{d}t}{t}\\
&\lesssim|Q|+\int_0^\ell\|\Theta_t b\|_{L^2}^2\frac{\mathrm{d}t}{t}. \tag*{(32)}
\end{aligned} 2 1 ∭ R ( Q ) ∣ ( γ t ( x ) ⋅ A t ) b ( x ) ∣ 2 t d x d t ≤ ∫ 0 ℓ ( ∥ ( γ t ⋅ A t − Θ t ) b ∥ L 2 2 + ∥ Θ t b ∥ L 2 2 ) t d t ≲ ∥ b ∥ L 2 2 + ∫ 0 ℓ ∥ Θ t b ∥ L 2 2 t d t ≲ ∣ Q ∣ + ∫ 0 ℓ ∥ Θ t b ∥ L 2 2 t d t . (32)
By definition of Θ t \Theta_t Θ t and b b b , we have
Θ t b = − 2 t α ( 1 + t 2 α L ) − 1 div α ( μ ∇ α ( 1 + ε 2 α ℓ 2 α L ) − 1 Φ ) = 2 t 2 α ( 1 + t 2 α L ) − 1 L ( 1 + ε 2 α ℓ 2 α L ) − 1 Φ = − t α ( 1 + t 2 α L ) − 1 ( 1 + ε 2 α ℓ 2 α L ) − 1 div α ( 2 μ ∇ α Φ ) . \begin{aligned}
\Theta_t b&=-2t^\alpha(1+t^{2\alpha}\mathcal{L})^{-1}\operatorname{div}^\alpha\left(\mu\nabla^\alpha(1+\varepsilon^{2\alpha}\ell^{2\alpha}\mathcal{L})^{-1}\Phi\right)\\
&=2t^{2\alpha}(1+t^{2\alpha}\mathcal{L})^{-1}\mathcal{L}(1+\varepsilon^{2\alpha}\ell^{2\alpha}\mathcal{L})^{-1}\Phi\\
&=-t^\alpha(1+t^{2\alpha}\mathcal{L})^{-1}(1+\varepsilon^{2\alpha}\ell^{2\alpha}\mathcal{L})^{-1}\operatorname{div}^\alpha(2\mu\nabla^\alpha\Phi).
\end{aligned} Θ t b = − 2 t α ( 1 + t 2 α L ) − 1 div α ( μ ∇ α ( 1 + ε 2 α ℓ 2 α L ) − 1 Φ ) = 2 t 2 α ( 1 + t 2 α L ) − 1 L ( 1 + ε 2 α ℓ 2 α L ) − 1 Φ = − t α ( 1 + t 2 α L ) − 1 ( 1 + ε 2 α ℓ 2 α L ) − 1 div α ( 2 μ ∇ α Φ ) .
Employing the uniform L 2 L^2 L 2 -bounds of Proposition 3.4 once again and (4.17), we obtain
∥ Θ t b ∥ L 2 2 ≲ t 2 α ε 2 α ℓ 2 α ∥ 2 μ ∇ α Φ ∥ L 2 2 ≲ t 2 α ε 2 α ℓ 2 α ∣ Q ∣ . \|\Theta_t b\|_{L^2}^2\lesssim\frac{t^{2\alpha}}{\varepsilon^{2\alpha}\ell^{2\alpha}}\|2\mu\nabla^\alpha\Phi\|_{L^2}^2\lesssim\frac{t^{2\alpha}}{\varepsilon^{2\alpha}\ell^{2\alpha}}|Q|. ∥ Θ t b ∥ L 2 2 ≲ ε 2 α ℓ 2 α t 2 α ∥2 μ ∇ α Φ ∥ L 2 2 ≲ ε 2 α ℓ 2 α t 2 α ∣ Q ∣.
Finally, integration in t t t leaves us with
∫ 0 ℓ ∥ Θ t b ∥ L 2 2 d t t ≲ ∫ 0 ℓ t 2 α ε 2 α ℓ 2 α ∣ Q ∣ d t t ≤ ∣ Q ∣ 2 α ε 2 α ≤ ∣ Q ∣ 2 α 0 ε 2 , \int_0^\ell\|\Theta_t b\|_{L^2}^2\frac{\mathrm{d}t}{t}\lesssim\int_0^\ell\frac{t^{2\alpha}}{\varepsilon^{2\alpha}\ell^{2\alpha}}|Q|\frac{\mathrm{d}t}{t}\leq\frac{|Q|}{2\alpha\varepsilon^{2\alpha}}\leq\frac{|Q|}{2\alpha_0\varepsilon^2}, ∫ 0 ℓ ∥ Θ t b ∥ L 2 2 t d t ≲ ∫ 0 ℓ ε 2 α ℓ 2 α t 2 α ∣ Q ∣ t d t ≤ 2 α ε 2 α ∣ Q ∣ ≤ 2 α 0 ε 2 ∣ Q ∣ ,
which we insert in (4.19) to conclude (B3) provided ε ≤ 1 \varepsilon\leq1 ε ≤ 1 . □
Proof of Proposition 4.17. Fix ε = ε 0 \varepsilon=\varepsilon_0 ε = ε 0 from Proposition 4.18. Given the test functions b Q ε b_Q^\varepsilon b Q ε , the directional Carleson measure estimate follows from a stopping time argument with a John–Nirenberg lemma for Carleson measures, as in [10 ] , in the form of [39 Lem. 14.8 and Lem. 14.10] . In particular, the argument depends only on d d d , ε 0 \varepsilon_0 ε 0 and the constants from (B1)–(B3), and therefore applies verbatim. □
Proof of Corollary 1.3. The claim follows directly from Theorem 1.2 and Proposition 2.5. □
Applications and consequences of the Kato property
In this section, we present different applications of the Kato square root property for our fractional operator. These include L 2 L^2 L 2 -boundedness of the Riesz transform associated to L L L (Section 5.1) and topological mapping properties of L 1 2 L^{\frac{1}{2}} L 2 1 with respect to the coefficients μ \mu μ (Section 5.2) and the fractional parameter α \alpha α (Section 5.3). In the latter, we will also study the behavior of the square root as α → 1 \alpha\to1 α → 1 and recover the Kato square root property for elliptic operators − div ( μ ∇ ⋅ ) -\operatorname{div}(\mu\nabla\cdot) − div ( μ ∇ ⋅ ) from [10 ] .
The Riesz transform associated to L
As an immediate consequence of Theorem 1.2, we have the following properties of the square root.
Lemma 5.1. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . The square root L 1 2 : S α , 2 ( R d ) → L 2 ( R d ) L^{\frac{1}{2}}:S^{\alpha,2}(\mathbb{R}^d)\to L^2(\mathbb{R}^d) L 2 1 : S α , 2 ( R d ) → L 2 ( R d ) is injective but not surjective.
Proof. For injectivity, we use Theorem 1.2, Proposition 2.3 and Plancherel’s theorem to observe that if L 1 2 u = 0 L^{\frac{1}{2}}u=0 L 2 1 u = 0 for some u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) , then
0 = ∥ L 1 2 u ∥ L 2 ≃ ∥ ∇ α u ∥ L 2 ≃ ∥ ∣ ξ ∣ α F ( u ) ∥ L 2 . 0=\left\|L^{\frac{1}{2}}u\right\|_{L^2}\simeq\left\|\nabla^\alpha u\right\|_{L^2}\simeq\left\||\xi|^\alpha\mathcal{F}(u)\right\|_{L^2}. 0 = L 2 1 u L 2 ≃ ∥ ∇ α u ∥ L 2 ≃ ∥ ∣ ξ ∣ α F ( u ) ∥ L 2 .
As a consequence F ( u ) = 0 \mathcal{F}(u)=0 F ( u ) = 0 , which implies u = 0 u=0 u = 0 . Next, assume L 1 2 L^{\frac{1}{2}} L 2 1 is surjective. Then, the open mapping theorem implies that L 1 2 L^{\frac{1}{2}} L 2 1 is an isomorphism and
∥ u ∥ S α , 2 ≃ ∥ L 1 2 u ∥ L 2 ≃ ∥ ∇ α u ∥ L 2 \|u\|_{S^{\alpha,2}}\simeq\left\|L^{\frac{1}{2}}u\right\|_{L^2}\simeq\left\|\nabla^\alpha u\right\|_{L^2} ∥ u ∥ S α , 2 ≃ L 2 1 u L 2 ≃ ∥ ∇ α u ∥ L 2
for all u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) . However, for λ > 0 \lambda>0 λ > 0 and u ∈ C c ∞ ( R d ) ∖ { 0 } u\in C^\infty_c(\mathbb{R}^d)\setminus\{0\} u ∈ C c ∞ ( R d ) ∖ { 0 } , we obtain
λ d 2 ∥ u ∥ L 2 = ∥ u ( ⋅ / λ ) ∥ L 2 ≲ ∥ ∇ α [ u ( ⋅ / λ ) ] ∥ L 2 = λ d 2 − α ∥ ∇ α u ∥ L 2 . \lambda^{\frac{d}{2}}\|u\|_{L^2}=\|u(\cdot/\lambda)\|_{L^2}\lesssim\|\nabla^\alpha[u(\cdot/\lambda)]\|_{L^2}=\lambda^{\frac{d}{2}-\alpha}\|\nabla^\alpha u\|_{L^2}. λ 2 d ∥ u ∥ L 2 = ∥ u ( ⋅ / λ ) ∥ L 2 ≲ ∥ ∇ α [ u ( ⋅ / λ )] ∥ L 2 = λ 2 d − α ∥ ∇ α u ∥ L 2 .
Thus, dividing by λ d 2 \lambda^{\frac{d}{2}} λ 2 d and letting λ → ∞ \lambda\to\infty λ → ∞ , we conclude ∥ u ∥ L 2 = 0 \|u\|_{L^2}=0 ∥ u ∥ L 2 = 0 , and thus, u = 0 u=0 u = 0 , which is a contradiction. □
Theorem 5.2 (Riesz transform on L 2 L^2 L 2 ). Let α 0 ∈ ( 0 , 1 ) \alpha_0\in(0,1) α 0 ∈ ( 0 , 1 ) and α ∈ [ α 0 , 1 ) \alpha\in[\alpha_0,1) α ∈ [ α 0 , 1 ) . We can extend the fractional Riesz transform
∇ α L − 1 2 : R ( L 1 2 ) → L 2 ( R d ; C d ) \nabla^\alpha L^{-\frac{1}{2}}:\mathcal{R}(L^{\frac{1}{2}})\to L^2(\mathbb{R}^d;\mathbb{C}^d) ∇ α L − 2 1 : R ( L 2 1 ) → L 2 ( R d ; C d ) to a bounded linear operator on L 2 L^2 L 2 satisfying
∥ ∇ α L − 1 2 u ∥ L 2 ≃ ∥ u ∥ L 2 ( u ∈ L 2 ( R d ) ) (33) \|\nabla^\alpha L^{-\frac{1}{2}}u\|_{L^2}\simeq\|u\|_{L^2}\qquad(u\in L^2(\mathbb{R}^d)) \tag*{(33)} ∥ ∇ α L − 2 1 u ∥ L 2 ≃ ∥ u ∥ L 2 ( u ∈ L 2 ( R d )) (33) where the implicit constants depend only on d d d , α 0 \alpha_0 α 0 and μ ∙ , μ ∙ \mu_\bullet,\mu^\bullet μ ∙ , μ ∙ .
Proof. By Lemma 5.1, we notice that ∇ α L − 1 2 : R ( L 1 2 ) → L 2 ( R d ; C d ) \nabla^\alpha L^{-\frac{1}{2}}:\mathcal{R}(L^{\frac{1}{2}})\to L^2(\mathbb{R}^d;\mathbb{C}^d) ∇ α L − 2 1 : R ( L 2 1 ) → L 2 ( R d ; C d ) is well defined. Moreover, by Theorem 1.2, we observe that for every u = L 1 2 v ∈ R ( L 1 2 ) u=L^{\frac{1}{2}}v\in\mathcal{R}(L^{\frac{1}{2}}) u = L 2 1 v ∈ R ( L 2 1 ) the two-sided estimate
∥ ∇ α L − 1 2 u ∥ L 2 = ∥ ∇ α v ∥ L 2 ≃ ∥ L 1 2 v ∥ L 2 = ∥ u ∥ L 2 \|\nabla^\alpha L^{-\frac{1}{2}}u\|_{L^2}=\|\nabla^\alpha v\|_{L^2}\simeq\|L^{\frac{1}{2}}v\|_{L^2}=\|u\|_{L^2} ∥ ∇ α L − 2 1 u ∥ L 2 = ∥ ∇ α v ∥ L 2 ≃ ∥ L 2 1 v ∥ L 2 = ∥ u ∥ L 2
holds, where the implicit constants depend only on d d d , α 0 \alpha_0 α 0 and μ ∙ , μ ∙ \mu_\bullet,\mu^\bullet μ ∙ , μ ∙ . Finally, since
L 2 ( R d ) = R ( L ) ‾ ⊂ R ( L 1 2 ) ⊂ L 2 ( R d ) , L^2(\mathbb{R}^d)=\overline{\mathcal{R}(L)}\subset\mathcal{R}(L^{\frac{1}{2}})\subset L^2(\mathbb{R}^d), L 2 ( R d ) = R ( L ) ⊂ R ( L 2 1 ) ⊂ L 2 ( R d ) ,
see [42 Prop. 2.1.1] , we can extend ∇ α L − 1 2 \nabla^\alpha L^{-\frac{1}{2}} ∇ α L − 2 1 to a bounded operator on L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) via density, which satisfies (5.1) for every u ∈ L 2 ( R d ) u\in L^2(\mathbb{R}^d) u ∈ L 2 ( R d ) . □
Holomorphic and Lipschitz dependence
A particular application that Kato had in mind was the Lipschitz dependence of the square roots L 1 2 L^{\frac{1}{2}} L 2 1 with respect to the coefficients μ \mu μ measured in the L ∞ L^\infty L ∞ -topology, which is proven via holomorphy. To formulate such properties and statements, let us emphasize the coefficients in the notation of L L L by writing L μ L_\mu L μ for the fractional operator with coefficients μ \mu μ .
Proposition 5.3. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . The set
O : = { μ : μ satisfies Assumption 1.1 for some μ ∙ , μ ∙ > 0 } \mathcal{O}:=\{\mu:\mu\text{ satisfies Assumption 1.1 for some }\mu_{\bullet},\mu^{\bullet}>0\} O := { μ : μ satisfies Assumption 1.1 for some μ ∙ , μ ∙ > 0 } is open in the L ∞ L^\infty L ∞ -topology and the map
O ⟶ L ( S α , 2 ( R d ) , L 2 ( R d ) ) , μ ↦ L μ 1 2 \mathcal{O}\longrightarrow\mathcal{L}(S^{\alpha,2}(\mathbb{R}^d),L^2(\mathbb{R}^d)),\qquad\mu\mapsto L_\mu^{\frac{1}{2}} O ⟶ L ( S α , 2 ( R d ) , L 2 ( R d )) , μ ↦ L μ 2 1 is holomorphic, that is, for every μ ∈ O \mu\in\mathcal{O} μ ∈ O and M ∈ L ∞ ( R d ; C d × d ) M\in L^\infty(\mathbb{R}^d;\mathbb{C}^{d\times d}) M ∈ L ∞ ( R d ; C d × d ) there exists r > 0 r>0 r > 0 such that the map
{ z ∈ C : ∣ z ∣ < r } ⟶ L ( S α , 2 ( R d ) , L 2 ( R d ) ) , z ↦ L μ + z M 1 2 \{z\in\mathbb{C}:|z|<r\}\longrightarrow\mathcal{L}(S^{\alpha,2}(\mathbb{R}^d),L^2(\mathbb{R}^d)),\qquad z\mapsto L_{\mu+zM}^{\frac{1}{2}} { z ∈ C : ∣ z ∣ < r } ⟶ L ( S α , 2 ( R d ) , L 2 ( R d )) , z ↦ L μ + z M 2 1 is holomorphic.
This kind of holomorphy of μ ↦ L μ 1 2 \mu\mapsto L_\mu^{\frac{1}{2}} μ ↦ L μ 2 1 implies the above-mentioned local Lipschitz property for small perturbations, which reads as follows.
Corollary 5.4. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) and μ 1 ∈ O \mu_1\in\mathcal{O} μ 1 ∈ O . Then there exist δ > 0 \delta>0 δ > 0 and a constant C > 0 C>0 C > 0 depending only on d d d , α \alpha α , ( μ 1 ) ∙ , ( μ 1 ) ∙ (\mu_1)_{\bullet},(\mu_1)^{\bullet} ( μ 1 ) ∙ , ( μ 1 ) ∙ and δ \delta δ such that for all μ 2 ∈ O \mu_2\in\mathcal{O} μ 2 ∈ O with ∥ μ 2 − μ 1 ∥ L ∞ ( R d ; C d × d ) < δ \|\mu_2-\mu_1\|_{L^\infty(\mathbb{R}^d;\mathbb{C}^{d\times d})}<\delta ∥ μ 2 − μ 1 ∥ L ∞ ( R d ; C d × d ) < δ we have
∥ L μ 2 1 2 − L μ 1 1 2 ∥ L ( S α , 2 , L 2 ) ≤ C ∥ μ 2 − μ 1 ∥ L ∞ ( R d ; C d × d ) . \|L_{\mu_2}^{\frac{1}{2}}-L_{\mu_1}^{\frac{1}{2}}\|_{\mathcal{L}(S^{\alpha,2},L^2)}\leq C\|\mu_2-\mu_1\|_{L^\infty(\mathbb{R}^d;\mathbb{C}^{d\times d})}. ∥ L μ 2 2 1 − L μ 1 2 1 ∥ L ( S α , 2 , L 2 ) ≤ C ∥ μ 2 − μ 1 ∥ L ∞ ( R d ; C d × d ) .
We emphasize that this is a nontrivial result even for the class of symmetric coefficients μ \mu μ . Indeed, already for elliptic operators in divergence form, there is no known proof that establishes the Lipschitz estimate for symmetric coefficients by remaining in the class of operators with symmetric coefficients. All existing proofs use in a crucial way the square root property for operators with nonsymmetric coefficients.
The proofs of both statements follow verbatim from the corresponding proofs for the generalized Stokes operator (see [41 Sec. 6] ), which are based only on the divergence form structure and on the fact that the constants in Theorem 1.2 depend on μ \mu μ only through μ ∙ \mu_{\bullet} μ ∙ and μ ∙ \mu^{\bullet} μ ∙ . We therefore omit them and leave the details to the interested reader.
Continuous dependence on alpha and the limit alpha to 1
In this subsection we use the local uniformity in α \alpha α in Theorem 1.2 to recover the Kato square root property for second-order elliptic operators in divergence form in the limit α → 1 \alpha\to1 α → 1 , and show continuity of the square root with respect to α \alpha α .
Throughout this subsection, μ \mu μ is fixed and satisfies Assumption 1.1. We write L α L_\alpha L α for the operator L L L from Definition 3.1 to indicate its dependence on α \alpha α .
Definition 5.5. Define the sesquilinear form
a 1 : W 1 , 2 ( R d ) × W 1 , 2 ( R d ) ⟶ C , a 1 ( u , v ) : = ∫ R d μ ∇ u ⋅ ∇ v ‾ d x . \mathfrak{a}_1:W^{1,2}(\mathbb{R}^d)\times W^{1,2}(\mathbb{R}^d)\longrightarrow\mathbb{C},\qquad\mathfrak{a}_1(u,v):=\int_{\mathbb{R}^d}\mu\nabla u\cdot\overline{\nabla v}\,dx. a 1 : W 1 , 2 ( R d ) × W 1 , 2 ( R d ) ⟶ C , a 1 ( u , v ) := ∫ R d μ ∇ u ⋅ ∇ v d x .
Additionally, we define the weak elliptic operator L 1 : W 1 , 2 ( R d ) → ( W 1 , 2 ( R d ) ) ′ L_1: W^{1,2}(\mathbb{R}^d) \to(W^{1,2}(\mathbb{R}^d))' L 1 : W 1 , 2 ( R d ) → ( W 1 , 2 ( R d ) ) ′ by
⟨ L 1 u , v ⟩ : = a 1 ( u , v ) ( u , v ∈ W 1 , 2 ( R d ) ) \langle L_1u,v\rangle:= \mathbf{a}_1(u,v) \qquad(u,v\in W^{1,2}(\mathbb{R}^d)) ⟨ L 1 u , v ⟩ := a 1 ( u , v ) ( u , v ∈ W 1 , 2 ( R d ))
and define
D ( L 1 ) : = { u ∈ W 1 , 2 ( R d ) : L 1 u ∈ L 2 ( R d ) } with L 1 u = L 1 u . D(L_1):=\{u\in W^{1,2}(\mathbb{R}^d): L_1u\in L^2(\mathbb{R}^d)\}\quad\text{with}\quad L_1u=\mathcal{L}_1u. D ( L 1 ) := { u ∈ W 1 , 2 ( R d ) : L 1 u ∈ L 2 ( R d )} with L 1 u = L 1 u .
As in the fractional case, L 1 L_1 L 1 is maximal accretive and sectorial of angle ω 0 \omega_0 ω 0 from (9) on L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) . In particular, L 1 L_1 L 1 satisfies the uniform resolvent bounds from Proposition 3.4 for α = 1 \alpha=1 α = 1 and possesses a unique maximal accretive square root L 1 1 / 2 L_1^{1/2} L 1 1/2 . We will need the following two convergence results; see also [28 Prop. 4.4 and Lem. 5.3] . We will use the conventions
∇ 1 : = ∇ , div 1 : = div , S 1 , 2 ( R d ) : = W 1 , 2 ( R d ) \nabla^1:=\nabla,\qquad\operatorname{div}^1:=\operatorname{div},\qquad S^{1,2}(\mathbb{R}^d):=W^{1,2}(\mathbb{R}^d) ∇ 1 := ∇ , div 1 := div , S 1 , 2 ( R d ) := W 1 , 2 ( R d )
to unify the notation.
Lemma 5.6. Let α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] , f ∈ W 1 , 2 ( R d ) f\in W^{1,2}(\mathbb{R}^d) f ∈ W 1 , 2 ( R d ) and F ∈ C ∞ ( R d ; C d ) F\in C^\infty(\mathbb{R}^d;\mathbb{C}^d) F ∈ C ∞ ( R d ; C d ) . Then, we have
∇ β f → ∇ α f in L 2 ( R d ; C d ) and div β F → div α F in L 2 ( R d ) \nabla^\beta f\to\nabla^\alpha f\quad\text{in }L^2(\mathbb{R}^d;\mathbb{C}^d)
\qquad\text{and}\qquad
\operatorname{div}^\beta F\to\operatorname{div}^\alpha F\quad\text{in }L^2(\mathbb{R}^d) ∇ β f → ∇ α f in L 2 ( R d ; C d ) and div β F → div α F in L 2 ( R d ) as β → α \beta\to\alpha β → α .
Proof. Since W 1 , 2 ( R d ) ⊂ S α , 2 ( R d ) W^{1,2}(\mathbb{R}^d)\subset S^{\alpha,2}(\mathbb{R}^d) W 1 , 2 ( R d ) ⊂ S α , 2 ( R d ) for all α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] by Lemma 2.6 , Proposition 2.3 together with Plancherel’s theorem yields
∥ ∇ β f − ∇ α f ∥ L 2 2 = ∫ R d ∣ ∣ ξ ∣ β − ∣ ξ ∣ α ∣ 2 ∣ F ( f ) ( ξ ) ∣ 2 d ξ . \left\|\nabla^\beta f-\nabla^\alpha f\right\|_{L^2}^2
=
\int_{\mathbb{R}^d}\left|\lvert\xi\rvert^\beta-\lvert\xi\rvert^\alpha\right|^2\left|\mathcal{F}(f)(\xi)\right|^2\,\mathrm{d}\xi. ∇ β f − ∇ α f L 2 2 = ∫ R d ∣ ξ ∣ β − ∣ ξ ∣ α 2 ∣ F ( f ) ( ξ ) ∣ 2 d ξ .
The integrand tends to zero as β → α \beta\to\alpha β → α and is bounded by 4 ( 1 + ∣ ξ ∣ 2 ) ∣ F ( f ) ( ξ ) ∣ 2 4(1+\lvert\xi\rvert^2)\lvert\mathcal{F}(f)(\xi)\rvert^2 4 ( 1 + ∣ ξ ∣ 2 ) ∣ F ( f ) ( ξ ) ∣ 2 , which is integrable, because f ∈ W 1 , 2 ( R d ) f\in W^{1,2}(\mathbb{R}^d) f ∈ W 1 , 2 ( R d ) . Hence, the claim follows from dominated convergence. The same argument applies to the fractional divergence. □
Lemma 5.7. Let α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] and ( α n ) n ⊂ ( 0 , 1 ] (\alpha_n)_n\subset(0,1] ( α n ) n ⊂ ( 0 , 1 ] with α n → α \alpha_n\to\alpha α n → α , s > 0 s>0 s > 0 and f ∈ L 2 ( R d ) f\in L^2(\mathbb{R}^d) f ∈ L 2 ( R d ) . Then
( 1 + s L α n ) − 1 f → ( 1 + s L α ) − 1 f and ∇ α n ( 1 + s L α n ) − 1 f → ∇ α ( 1 + s L α ) − 1 f (1+sL_{\alpha_n})^{-1}f\to(1+sL_\alpha)^{-1}f
\qquad\text{and}\qquad
\nabla^{\alpha_n}(1+sL_{\alpha_n})^{-1}f\to\nabla^\alpha(1+sL_\alpha)^{-1}f ( 1 + s L α n ) − 1 f → ( 1 + s L α ) − 1 f and ∇ α n ( 1 + s L α n ) − 1 f → ∇ α ( 1 + s L α ) − 1 f in L 2 L^2 L 2 as $n\to\infty.
Proof. The proof comes in two steps. The first step shows that the claimed convergence holds weakly in L 2 L^2 L 2 . Subsequently, we upgrade the weak convergence to a strong convergence in L 2 L^2 L 2 in the second step.
Step 1: Weak convergence in L 2 L^2 L 2 . Set u n : = ( 1 + s L α n ) − 1 f u_n:=(1+sL_{\alpha_n})^{-1}f u n := ( 1 + s L α n ) − 1 f . By Definition 3.1, we have u n ∈ S α n , 2 ( R d ) u_n\in S^{\alpha_n,2}(\mathbb{R}^d) u n ∈ S α n , 2 ( R d ) and
∫ R d u n v ‾ d x + s ∫ R d μ ∇ α n u n ⋅ ∇ α n v ‾ d x = ∫ R d f v ‾ d x (34) \int_{\mathbb{R}^d}u_n\overline{v}\,\mathrm{d}x+s\int_{\mathbb{R}^d}\mu\nabla^{\alpha_n}u_n\cdot\overline{\nabla^{\alpha_n}v}\,\mathrm{d}x
=
\int_{\mathbb{R}^d}f\overline{v}\,\mathrm{d}x \tag*{(34)} ∫ R d u n v d x + s ∫ R d μ ∇ α n u n ⋅ ∇ α n v d x = ∫ R d f v d x (34)
for every v ∈ S α n , 2 ( R d ) v\in S^{\alpha_n,2}(\mathbb{R}^d) v ∈ S α n , 2 ( R d ) . Testing with v = u n v=u_n v = u n and using Assumption 1.1 yields
∥ u n ∥ L 2 2 + s μ ∙ ∥ ∇ α n u n ∥ L 2 2 ≤ ∫ R d ∣ u n ∣ 2 d x + s Re ∫ R d μ ∇ α n u n ⋅ ∇ α n u n ‾ d x ≤ ∣ ∫ R d f u n ‾ d x ∣ . \|u_n\|_{L^2}^2+s\mu_\bullet\|\nabla^{\alpha_n}u_n\|_{L^2}^2
\leq
\int_{\mathbb{R}^d}|u_n|^2\,\mathrm{d}x+s\operatorname{Re}\int_{\mathbb{R}^d}\mu\nabla^{\alpha_n}u_n\cdot\overline{\nabla^{\alpha_n}u_n}\,\mathrm{d}x
\leq
\left|\int_{\mathbb{R}^d}f\overline{u_n}\,\mathrm{d}x\right|. ∥ u n ∥ L 2 2 + s μ ∙ ∥ ∇ α n u n ∥ L 2 2 ≤ ∫ R d ∣ u n ∣ 2 d x + s Re ∫ R d μ ∇ α n u n ⋅ ∇ α n u n d x ≤ ∫ R d f u n d x .
Applying Young’s inequality gets us
∥ u n ∥ L 2 2 + s μ ∙ ∥ ∇ α n u n ∥ L 2 2 ≤ ∥ f ∥ L 2 2 2 + ∥ u n ∥ L 2 2 2 \|u_n\|_{L^2}^2+s\mu_\bullet\|\nabla^{\alpha_n}u_n\|_{L^2}^2\le\frac{\|f\|_{L^2}^2}{2}+\frac{\|u_n\|_{L^2}^2}{2} ∥ u n ∥ L 2 2 + s μ ∙ ∥ ∇ α n u n ∥ L 2 2 ≤ 2 ∥ f ∥ L 2 2 + 2 ∥ u n ∥ L 2 2
and hence
∥ u n ∥ L 2 2 + 2 s μ ∙ ∥ ∇ α n u n ∥ L 2 2 ≤ ∥ f ∥ L 2 2 . (35) \|u_n\|_{L^2}^2+2s\mu_\bullet\|\nabla^{\alpha_n}u_n\|_{L^2}^2\le\|f\|_{L^2}^2. \tag*{(35)} ∥ u n ∥ L 2 2 + 2 s μ ∙ ∥ ∇ α n u n ∥ L 2 2 ≤ ∥ f ∥ L 2 2 . (35)
Thus, after passing to subsequences ( u n k ) k (u_{n_k})_k ( u n k ) k and ( ∇ α n k u n k ) k (\nabla^{\alpha_{n_k}}u_{n_k})_k ( ∇ α n k u n k ) k , there exist u ∈ L 2 ( R d ) u\in L^2(\mathbb{R}^d) u ∈ L 2 ( R d ) and G ∈ L 2 ( R d ; C d ) G\in L^2(\mathbb{R}^d;\mathbb{C}^d) G ∈ L 2 ( R d ; C d ) such that
u n k ⇀ u and ∇ α n k u n k ⇀ G weakly in L 2 . (36) u_{n_k}\rightharpoonup u\quad\text{and}\quad\nabla^{\alpha_{n_k}}u_{n_k}\rightharpoonup G\quad\text{weakly in }L^2. \tag*{(36)} u n k ⇀ u and ∇ α n k u n k ⇀ G weakly in L 2 . (36)
First, we show u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) with ∇ α u = G \nabla^\alpha u=G ∇ α u = G . For F ∈ C c ∞ ( R d ; C d ) F\in C^\infty_c(\mathbb{R}^d;\mathbb{C}^d) F ∈ C c ∞ ( R d ; C d ) , Definition 2.4 gives
∫ R d ∇ α n k u n k ⋅ F d x = − ∫ R d u n k div α n k F d x . \int_{\mathbb{R}^d}\nabla^{\alpha_{n_k}}u_{n_k}\cdot F\,\mathrm{d}x=-\int_{\mathbb{R}^d}u_{n_k}\operatorname{div}^{\alpha_{n_k}}F\,\mathrm{d}x. ∫ R d ∇ α n k u n k ⋅ F d x = − ∫ R d u n k div α n k F d x .
Observe that the left-hand side tends to ∫ R d G ⋅ F d x \int_{\mathbb{R}^d}G\cdot F\,\mathrm{d}x ∫ R d G ⋅ F d x by (5.4). Moreover, on the right-hand side, u n k ⇀ u u_{n_k}\rightharpoonup u u n k ⇀ u weakly in L 2 L^2 L 2 and div α n k F → div α F \operatorname{div}^{\alpha_{n_k}}F\to\operatorname{div}^\alpha F div α n k F → div α F strongly in L 2 L^2 L 2 by Lemma 5.6. Thus, it converges to − ∫ R d u div α F d x -\int_{\mathbb{R}^d}u\operatorname{div}^\alpha F\,\mathrm{d}x − ∫ R d u div α F d x . Hence, we obtain
∫ R d G ⋅ F d x = − ∫ R d u div α F d x \int_{\mathbb{R}^d}G\cdot F\,\mathrm{d}x=-\int_{\mathbb{R}^d}u\operatorname{div}^\alpha F\,\mathrm{d}x ∫ R d G ⋅ F d x = − ∫ R d u div α F d x
for every F ∈ C c ∞ ( R d ; C d ) F\in C^\infty_c(\mathbb{R}^d;\mathbb{C}^d) F ∈ C c ∞ ( R d ; C d ) , which shows u ∈ S α , 2 ( R d ) u\in S^{\alpha,2}(\mathbb{R}^d) u ∈ S α , 2 ( R d ) with ∇ α u = G \nabla^\alpha u=G ∇ α u = G . Next, let v ∈ C c ∞ ( R d ) v\in C^\infty_c(\mathbb{R}^d) v ∈ C c ∞ ( R d ) and split
∫ R d μ ∇ α n k u n k ⋅ ∇ α n k v ‾ d x = ∫ R d μ ∇ α n k u n k ⋅ ∇ α v ‾ d x + ∫ R d μ ∇ α n k u n k ⋅ ( ∇ α n k v ‾ − ∇ α v ‾ ) d x . \int_{\mathbb{R}^d}\mu\nabla^{\alpha_{n_k}}u_{n_k}\cdot\overline{\nabla^{\alpha_{n_k}}v}\,\mathrm{d}x
=
\int_{\mathbb{R}^d}\mu\nabla^{\alpha_{n_k}}u_{n_k}\cdot\overline{\nabla^\alpha v}\,\mathrm{d}x
+
\int_{\mathbb{R}^d}\mu\nabla^{\alpha_{n_k}}u_{n_k}\cdot\left(\overline{\nabla^{\alpha_{n_k}}v}-\overline{\nabla^\alpha v}\right)\,\mathrm{d}x. ∫ R d μ ∇ α n k u n k ⋅ ∇ α n k v d x = ∫ R d μ ∇ α n k u n k ⋅ ∇ α v d x + ∫ R d μ ∇ α n k u n k ⋅ ( ∇ α n k v − ∇ α v ) d x .
Notice that the first term on the right-hand side tends to ∫ R d μ ∇ α u ⋅ ∇ α v ‾ d x \int_{\mathbb{R}^d}\mu\nabla^\alpha u\cdot\overline{\nabla^\alpha v}\,\mathrm{d}x ∫ R d μ ∇ α u ⋅ ∇ α v d x by (5.4). Furthermore, the second term on the right-hand side is bounded by
∣ ∫ R d μ ∇ α n k u n k ⋅ ( ∇ α n k v ‾ − ∇ α v ‾ ) d x ∣ ≤ μ ( sup k ∈ N ∥ ∇ α n k u n k ∥ L 2 ) ∥ ∇ α n k v − ∇ α v ∥ L 2 , \left|\int_{\mathbb{R}^d}\mu\nabla^{\alpha_{n_k}}u_{n_k}\cdot\left(\overline{\nabla^{\alpha_{n_k}}v}-\overline{\nabla^\alpha v}\right)\,\mathrm{d}x\right|
\le
\mu\left(\sup_{k\in\mathbb{N}}\|\nabla^{\alpha_{n_k}}u_{n_k}\|_{L^2}\right)\|\nabla^{\alpha_{n_k}}v-\nabla^\alpha v\|_{L^2}, ∫ R d μ ∇ α n k u n k ⋅ ( ∇ α n k v − ∇ α v ) d x ≤ μ ( k ∈ N sup ∥ ∇ α n k u n k ∥ L 2 ) ∥ ∇ α n k v − ∇ α v ∥ L 2 ,
where the supremum is finite, because ( ∇ α n k u n k ) k (\nabla^{\alpha_{n_k}}u_{n_k})_k ( ∇ α n k u n k ) k converges weakly in L 2 L^2 L 2 . Hence, it converges to zero as k → ∞ k\to\infty k → ∞ by Lemma 5.6. As a consequence, we obtain
∫ R d μ ∇ α n k u n k ⋅ ∇ α n k v ‾ d x → ∫ R d μ ∇ α u ⋅ ∇ α v ‾ d x \int_{\mathbb{R}^d}\mu\nabla^{\alpha_{n_k}}u_{n_k}\cdot\overline{\nabla^{\alpha_{n_k}}v}\,\mathrm{d}x
\to
\int_{\mathbb{R}^d}\mu\nabla^\alpha u\cdot\overline{\nabla^\alpha v}\,\mathrm{d}x ∫ R d μ ∇ α n k u n k ⋅ ∇ α n k v d x → ∫ R d μ ∇ α u ⋅ ∇ α v d x
for k → ∞ k\to\infty k → ∞ . Finally, passing to the limit in (5.2) yields
∫ R d u v ‾ d x + s ∫ R d μ ∇ α u ⋅ ∇ α v ‾ d x = ∫ R d f v ‾ d x \int_{\mathbb{R}^d}u\overline{v}\,\mathrm{d}x+s\int_{\mathbb{R}^d}\mu\nabla^\alpha u\cdot\overline{\nabla^\alpha v}\,\mathrm{d}x
=
\int_{\mathbb{R}^d}f\overline{v}\,\mathrm{d}x ∫ R d u v d x + s ∫ R d μ ∇ α u ⋅ ∇ α v d x = ∫ R d f v d x
for every v ∈ C ∞ ( R d ) v\in C^\infty(\mathbb{R}^d) v ∈ C ∞ ( R d ) , and by density for all v ∈ S α , 2 ( R d ) v\in S^{\alpha,2}(\mathbb{R}^d) v ∈ S α , 2 ( R d ) . Hence, u = ( 1 + s L α ) − 1 f u=(1+sL_\alpha)^{-1}f u = ( 1 + s L α ) − 1 f and
( 1 + s L α n k ) − 1 f ⇀ ( 1 + s L α ) − 1 f (1+sL_{\alpha_{n_k}})^{-1}f\rightharpoonup(1+sL_\alpha)^{-1}f ( 1 + s L α n k ) − 1 f ⇀ ( 1 + s L α ) − 1 f and ∇ α n k ( 1 + s L α n k ) − 1 f ⇀ ∇ α ( 1 + s L α ) − 1 f \nabla^{\alpha_{n_k}}(1+sL_{\alpha_{n_k}})^{-1}f\rightharpoonup\nabla^\alpha(1+sL_\alpha)^{-1}f ∇ α n k ( 1 + s L α n k ) − 1 f ⇀ ∇ α ( 1 + s L α ) − 1 f
weakly in L 2 L^2 L 2 . Since every subsequence of ( u n , ∇ α n u n ) n (u_n,\nabla^{\alpha_n}u_n)_n ( u n , ∇ α n u n ) n satisfies the uniform bound (5.3), the argument above applies to every subsequence. More precisely, every subsequence of ( u n , ∇ α n u n ) n (u_n,\nabla^{\alpha_n}u_n)_n ( u n , ∇ α n u n ) n has a further subsequence which converges weakly to ( u , ∇ α u ) (u,\nabla^\alpha u) ( u , ∇ α u ) in L 2 L^2 L 2 , where u = ( 1 + s L α ) − 1 f u=(1+sL_\alpha)^{-1}f u = ( 1 + s L α ) − 1 f . As these limits are the same for all subsequences, the whole sequences ( u n ) n (u_n)_n ( u n ) n and ( ∇ α n u n ) n (\nabla^{\alpha_n}u_n)_n ( ∇ α n u n ) n has to converge weakly to u u u and ∇ α u \nabla^\alpha u ∇ α u , respectively.
Step 2: Strong convergence in L 2 L^2 L 2 . Testing the resolvent equations for u n u_n u n and u u u with u n u_n u n and u u u , respectively, and taking real parts yields
∥ u n ∥ L 2 2 + s q ( ∇ α n u n ) = Re ⟨ f , u n ⟩ and ∥ u ∥ L 2 2 + s q ( ∇ α u ) = Re ⟨ f , u ⟩ , \|u_n\|_{L^2}^2+sq(\nabla^{\alpha_n}u_n)=\operatorname{Re}\langle f,u_n\rangle\quad\text{and}\quad\|u\|_{L^2}^2+sq(\nabla^\alpha u)=\operatorname{Re}\langle f,u\rangle, ∥ u n ∥ L 2 2 + s q ( ∇ α n u n ) = Re ⟨ f , u n ⟩ and ∥ u ∥ L 2 2 + s q ( ∇ α u ) = Re ⟨ f , u ⟩ ,
where we set
q ( F ) : = Re ∫ R d μ F ⋅ F ‾ d x ( F ∈ L 2 ( R d ; C d ) ) . q(F):=\operatorname{Re}\int_{\mathbb{R}^d}\mu F\cdot\overline{F}\,\mathrm{d}x \quad(F\in L^2(\mathbb{R}^d;\mathbb{C}^d)). q ( F ) := Re ∫ R d μ F ⋅ F d x ( F ∈ L 2 ( R d ; C d )) .
Subtracting both identities and writing G n : = ∇ α n u n G_n:=\nabla^{\alpha_n}u_n G n := ∇ α n u n and G : = ∇ α u G:=\nabla^\alpha u G := ∇ α u , we obtain
∥ u n ∥ L 2 2 − ∥ u ∥ L 2 2 + s q ( G n ) − s q ( G ) = Re ⟨ f , u n − u ⟩ . \|u_n\|_{L^2}^2-\|u\|_{L^2}^2+sq(G_n)-sq(G)=\operatorname{Re}\langle f,u_n-u\rangle. ∥ u n ∥ L 2 2 − ∥ u ∥ L 2 2 + s q ( G n ) − s q ( G ) = Re ⟨ f , u n − u ⟩ .
Using the identities
∥ u n ∥ L 2 2 − ∥ u ∥ L 2 2 = ∥ u n − u ∥ L 2 2 + 2 Re ⟨ u n − u , u ⟩ \|u_n\|_{L^2}^2-\|u\|_{L^2}^2=\|u_n-u\|_{L^2}^2+2\operatorname{Re}\langle u_n-u,u\rangle ∥ u n ∥ L 2 2 − ∥ u ∥ L 2 2 = ∥ u n − u ∥ L 2 2 + 2 Re ⟨ u n − u , u ⟩
and
q ( G n ) − q ( G ) = q ( G n − G ) + Re ∫ R d μ G ⋅ G n − G ‾ d x + Re ∫ R d μ ( G n − G ) ⋅ G ‾ d x q(G_n)-q(G)=q(G_n-G)+\operatorname{Re}\int_{\mathbb{R}^d}\mu G\cdot\overline{G_n-G}\,\mathrm{d}x+\operatorname{Re}\int_{\mathbb{R}^d}\mu(G_n-G)\cdot\overline{G}\,\mathrm{d}x q ( G n ) − q ( G ) = q ( G n − G ) + Re ∫ R d μ G ⋅ G n − G d x + Re ∫ R d μ ( G n − G ) ⋅ G d x
in the equation above, we arrive after rearranging at
∥ u n − u ∥ L 2 2 + s q ( G n − G ) = Re ⟨ f , u n − u ⟩ − 2 Re ⟨ u n − u , u ⟩ − s Re ∫ R d μ G ⋅ G n − G ‾ d x − s Re ∫ R d μ ( G n − G ) ⋅ G ‾ d x . \begin{aligned}
\|u_n-u\|_{L^2}^2+sq(G_n-G)
={}&\operatorname{Re}\langle f,u_n-u\rangle-2\operatorname{Re}\langle u_n-u,u\rangle\\
&-s\operatorname{Re}\int_{\mathbb{R}^d}\mu G\cdot\overline{G_n-G}\,\mathrm{d}x
-s\operatorname{Re}\int_{\mathbb{R}^d}\mu(G_n-G)\cdot\overline{G}\,\mathrm{d}x.
\end{aligned} ∥ u n − u ∥ L 2 2 + s q ( G n − G ) = Re ⟨ f , u n − u ⟩ − 2 Re ⟨ u n − u , u ⟩ − s Re ∫ R d μ G ⋅ G n − G d x − s Re ∫ R d μ ( G n − G ) ⋅ G d x .
Because f , u ∈ L 2 ( R d ) f,u\in L^2(\mathbb{R}^d) f , u ∈ L 2 ( R d ) and G = ∇ α u ∈ L 2 ( R d ; C d ) G=\nabla^\alpha u\in L^2(\mathbb{R}^d;\mathbb{C}^d) G = ∇ α u ∈ L 2 ( R d ; C d ) , the right-hand side of the equality above vanishes for n → ∞ n\to\infty n → ∞ due to the weak convergence from Step 1. Hence, using Assumption 1.1, we get
∥ u n − u ∥ L 2 2 + s μ ∙ ∥ G n − G ∥ L 2 2 ≤ ∥ u n − u ∥ L 2 2 + s q ( G n − G ) ⟶ 0 \|u_n-u\|_{L^2}^2+s\mu_\bullet\|G_n-G\|_{L^2}^2\leq\|u_n-u\|_{L^2}^2+sq(G_n-G)\longrightarrow0 ∥ u n − u ∥ L 2 2 + s μ ∙ ∥ G n − G ∥ L 2 2 ≤ ∥ u n − u ∥ L 2 2 + s q ( G n − G ) ⟶ 0
as n → ∞ n\to\infty n → ∞ , which proves the claim. □
Combining these two convergence lemmata with the uniform Kato property from Theorem 1.2 recovers the classical Kato property for the local elliptic operator L 1 L_1 L 1 .
Proof of Proposition 1.4. As argued in Section 4.1, it is well known that the Kato property of L 1 L_1 L 1 follows from the square function estimate
∫ 0 ∞ ∥ t L 1 ( 1 + t 2 L 1 ) − 1 u ∥ L 2 2 d t t ≲ ∥ ∇ u ∥ L 2 2 , (37) \int_0^\infty\|tL_1(1+t^2L_1)^{-1}u\|_{L^2}^2\,\frac{\mathrm{d}t}{t}\lesssim\|\nabla u\|_{L^2}^2, \tag*{(37)} ∫ 0 ∞ ∥ t L 1 ( 1 + t 2 L 1 ) − 1 u ∥ L 2 2 t d t ≲ ∥∇ u ∥ L 2 2 , (37)
for all u ∈ W 1 , 2 ( R d ) u\in W^{1,2}(\mathbb{R}^d) u ∈ W 1 , 2 ( R d ) , see for instance [39 Cor. 12.9] . Recall from the proof of Theorem 1.2 that for all α 0 ∈ ( 0 , 1 ) \alpha_0\in(0,1) α 0 ∈ ( 0 , 1 ) and α ∈ [ α 0 , 1 ) \alpha\in[\alpha_0,1) α ∈ [ α 0 , 1 ) , we proved the square function estimate
∫ 0 ∞ ∥ t α L α ( 1 + t 2 α L α ) − 1 u ∥ L 2 2 d t t ≤ C ∥ ∇ α u ∥ L 2 2 ( u ∈ S α , 2 ( R d ) ) (38) \int_0^\infty\|t^\alpha L_\alpha(1+t^{2\alpha}L_\alpha)^{-1}u\|_{L^2}^2\,\frac{\mathrm{d}t}{t}\leq C\|\nabla^\alpha u\|_{L^2}^2 \quad(u\in S^{\alpha,2}(\mathbb{R}^d)) \tag*{(38)} ∫ 0 ∞ ∥ t α L α ( 1 + t 2 α L α ) − 1 u ∥ L 2 2 t d t ≤ C ∥ ∇ α u ∥ L 2 2 ( u ∈ S α , 2 ( R d )) (38)
for C > 0 C>0 C > 0 depending only on d d d , α 0 \alpha_0 α 0 and μ ∙ \mu_\bullet μ ∙ , μ ∙ \mu^\bullet μ ∙ . To this end, fix α 0 : = 1 2 \alpha_0:=\frac{1}{2} α 0 := 2 1 . Let u ∈ W 1 , 2 ( R d ) u\in W^{1,2}(\mathbb{R}^d) u ∈ W 1 , 2 ( R d ) and ( α n ) n ⊂ [ 1 2 , 1 ) (\alpha_n)_n\subset[\frac{1}{2},1) ( α n ) n ⊂ [ 2 1 , 1 ) with α n → 1 \alpha_n\to1 α n → 1 . Using t L α n ( 1 + t 2 L α n ) − 1 = t − 1 ( 1 − ( 1 + t 2 L α n ) − 1 ) tL_{\alpha_n}(1+t^2L_{\alpha_n})^{-1}=t^{-1}(1-(1+t^2L_{\alpha_n})^{-1}) t L α n ( 1 + t 2 L α n ) − 1 = t − 1 ( 1 − ( 1 + t 2 L α n ) − 1 ) and Lemma 5.7 with s = t 2 s=t^2 s = t 2 , we have
t L α n ( 1 + t 2 L α n ) − 1 u → t L 1 ( 1 + t 2 L 1 ) − 1 u in L 2 ( R d ) . tL_{\alpha_n}(1+t^2L_{\alpha_n})^{-1}u\to tL_1(1+t^2L_1)^{-1}u\quad\text{in }L^2(\mathbb{R}^d). t L α n ( 1 + t 2 L α n ) − 1 u → t L 1 ( 1 + t 2 L 1 ) − 1 u in L 2 ( R d ) .
Combining this with Fatou’s lemma, (5.6) and Lemma 5.6, we obtain
∫ 0 ∞ ∥ t L 1 ( 1 + t 2 L 1 ) − 1 u ∥ L 2 2 d t t ≤ lim inf n → ∞ ∫ 0 ∞ ∥ t L α n ( 1 + t 2 L α n ) − 1 u ∥ L 2 2 d t t = lim inf n → ∞ α n ∫ 0 ∞ ∥ t α n L α n ( 1 + t 2 α n L α n ) − 1 u ∥ L 2 2 d t t ≤ C lim inf n → ∞ ∥ ∇ α n u ∥ L 2 2 = C ∥ ∇ u ∥ L 2 2 , \begin{aligned}
\int_0^\infty\left\|tL_1(1+t^2L_1)^{-1}u\right\|_{L^2}^2\frac{dt}{t}
&\le\liminf_{n\to\infty}\int_0^\infty\left\|tL_{\alpha_n}(1+t^2L_{\alpha_n})^{-1}u\right\|_{L^2}^2\frac{dt}{t}\\
&=\liminf_{n\to\infty}\alpha_n\int_0^\infty\left\|t^{\alpha_n}L_{\alpha_n}(1+t^{2\alpha_n}L_{\alpha_n})^{-1}u\right\|_{L^2}^2\frac{dt}{t}\\
&\le C\liminf_{n\to\infty}\left\|\nabla_{\alpha_n}u\right\|_{L^2}^2\\
&=C\left\|\nabla u\right\|_{L^2}^2,
\end{aligned} ∫ 0 ∞ t L 1 ( 1 + t 2 L 1 ) − 1 u L 2 2 t d t ≤ n → ∞ lim inf ∫ 0 ∞ t L α n ( 1 + t 2 L α n ) − 1 u L 2 2 t d t = n → ∞ lim inf α n ∫ 0 ∞ t α n L α n ( 1 + t 2 α n L α n ) − 1 u L 2 2 t d t ≤ C n → ∞ lim inf ∥ ∇ α n u ∥ L 2 2 = C ∥ ∇ u ∥ L 2 2 ,
where C > 0 C>0 C > 0 is uniform with respect to the sequence ( α n ) n ∈ N (\alpha_n)_{n\in\mathbb{N}} ( α n ) n ∈ N . This finishes the proof. □
While the previous statement recovers the square root property at α = 1 \alpha=1 α = 1 via square function estimates, we can actually prove more. In fact, we show that the square root L α 1 2 u L_{\alpha}^{\frac{1}{2}}u L α 2 1 u is continuous in α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] . However, since the domain D ( L α 1 2 ) = S α , 2 ( R d ) D(L_{\alpha}^{\frac{1}{2}})=S^{\alpha,2}(\mathbb{R}^d) D ( L α 2 1 ) = S α , 2 ( R d ) depends on α \alpha α , we need one fixed common function space. Fortunately, the fractional Sobolev spaces form a continuous embedding chain due to Lemma 2.6, so that
⋂ α ∈ ( 0 , 1 ] D ( L α 1 2 ) = W 1 , 2 ( R d ) \bigcap_{\alpha\in(0,1]}D(L_{\alpha}^{\frac{1}{2}})=W^{1,2}(\mathbb{R}^d) α ∈ ( 0 , 1 ] ⋂ D ( L α 2 1 ) = W 1 , 2 ( R d )
is the natural choice.
Proof of Theorem 1.5. Fix u ∈ W 1 , 2 ( R d ) u\in W^{1,2}(\mathbb{R}^d) u ∈ W 1 , 2 ( R d ) . Let α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] and ( α n ) ⊂ ( 0 , 1 ] (\alpha_n)\subset(0,1] ( α n ) ⊂ ( 0 , 1 ] with α n → α \alpha_n\to\alpha α n → α . Without loss of generality, we assume α n ≥ α 2 \alpha_n\ge\frac{\alpha}{2} α n ≥ 2 α for every n ∈ N n\in\mathbb{N} n ∈ N . We divide the proof into two steps.
Step 1: Smoothed version. For ε > 0 \varepsilon>0 ε > 0 , we show the convergence
L α n 1 2 ( 1 + ε L α n ) − 1 u → L α 1 2 ( 1 + ε L α ) − 1 u L_{\alpha_n}^{\frac{1}{2}}(1+\varepsilon L_{\alpha_n})^{-1}u\to L_{\alpha}^{\frac{1}{2}}(1+\varepsilon L_{\alpha})^{-1}u L α n 2 1 ( 1 + ε L α n ) − 1 u → L α 2 1 ( 1 + ε L α ) − 1 u
in L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) . For this purpose, we use the Balakrishnan representation formula [42 Prop. 3.1.12]
L α n 1 2 ( 1 + ε L α n ) − 1 u = 1 π ∫ 0 ∞ t − 1 2 ( t + L α n ) − 1 L α n ( 1 + ε L α n ) − 1 u d t . (39) L_{\alpha_n}^{\frac{1}{2}}(1+\varepsilon L_{\alpha_n})^{-1}u=\frac{1}{\pi}\int_0^\infty t^{-\frac{1}{2}}(t+L_{\alpha_n})^{-1}L_{\alpha_n}(1+\varepsilon L_{\alpha_n})^{-1}u\,dt. \tag*{(39)} L α n 2 1 ( 1 + ε L α n ) − 1 u = π 1 ∫ 0 ∞ t − 2 1 ( t + L α n ) − 1 L α n ( 1 + ε L α n ) − 1 u d t . (39)
Combining the identity L α n ( s + L α n ) − 1 = 1 − s ( s + L α n ) − 1 L_{\alpha_n}(s+L_{\alpha_n})^{-1}=1-s(s+L_{\alpha_n})^{-1} L α n ( s + L α n ) − 1 = 1 − s ( s + L α n ) − 1 for every s > 0 s>0 s > 0 with the uniform L 2 L^2 L 2 -bounds of Proposition 3.4 yields
∥ ( t + L α n ) − 1 L α n ( 1 + ε L α n ) − 1 u ∥ L 2 ≲ min { 1 , 1 t ε } ∥ u ∥ L 2 . \left\|(t+L_{\alpha_n})^{-1}L_{\alpha_n}(1+\varepsilon L_{\alpha_n})^{-1}u\right\|_{L^2}\lesssim\min\left\{1,\frac{1}{t\varepsilon}\right\}\|u\|_{L^2}. ( t + L α n ) − 1 L α n ( 1 + ε L α n ) − 1 u L 2 ≲ min { 1 , tε 1 } ∥ u ∥ L 2 .
Moreover, by Lemma 5.7, the limit
( t + L α n ) − 1 L α n ( 1 + ε L α n ) − 1 u → ( t + L α ) − 1 L α ( 1 + ε L α ) − 1 u (t+L_{\alpha_n})^{-1}L_{\alpha_n}(1+\varepsilon L_{\alpha_n})^{-1}u\to(t+L_{\alpha})^{-1}L_{\alpha}(1+\varepsilon L_{\alpha})^{-1}u ( t + L α n ) − 1 L α n ( 1 + ε L α n ) − 1 u → ( t + L α ) − 1 L α ( 1 + ε L α ) − 1 u
holds in L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) as n → ∞ n \to\infty n → ∞ . Hence, we can apply dominated convergence to obtain
L α n 1 2 ( 1 + ε L α n ) − 1 u = 1 π ∫ 0 ∞ t − 1 2 ( t + L α n ) − 1 L α n ( 1 + ε L α n ) − 1 u d t → 1 π ∫ 0 ∞ t − 1 2 ( t + L α ) − 1 L α ( 1 + ε L α ) − 1 u d t = L α 1 2 ( 1 + ε L α ) − 1 u \begin{aligned}
L_{\alpha_n}^{\frac{1}{2}}(1+\varepsilon L_{\alpha_n})^{-1}u
= \frac{1}{\pi}\int_0^\infty t^{-\frac{1}{2}}(t+L_{\alpha_n})^{-1}L_{\alpha_n}(1+\varepsilon L_{\alpha_n})^{-1}u\,\mathrm{d}t \\
&\to\frac{1}{\pi}\int_0^\infty t^{-\frac{1}{2}}(t+L_\alpha)^{-1}L_\alpha(1+\varepsilon L_\alpha)^{-1}u\,\mathrm{d}t
= L_\alpha^{\frac{1}{2}}(1+\varepsilon L_\alpha)^{-1}u
\end{aligned} L α n 2 1 ( 1 + ε L α n ) − 1 u = π 1 ∫ 0 ∞ t − 2 1 ( t + L α n ) − 1 L α n ( 1 + ε L α n ) − 1 u d t → π 1 ∫ 0 ∞ t − 2 1 ( t + L α ) − 1 L α ( 1 + ε L α ) − 1 u d t = L α 2 1 ( 1 + ε L α ) − 1 u
as n → ∞ n \to\infty n → ∞ for every ε > 0 \varepsilon> 0 ε > 0 .
Step 2: Conclusion. Because L β L_\beta L β is sectorial of angle ω 0 \omega_0 ω 0 as defined in (9) for all β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] , we have the approximation property
L β 1 2 ( 1 + ε L β ) − 1 u = ( 1 + ε L β ) − 1 L β 1 2 u → L β 1 2 u (40) L_\beta^{\frac{1}{2}}(1+\varepsilon L_\beta)^{-1}u
= (1+\varepsilon L_\beta)^{-1}L_\beta^{\frac{1}{2}}u
\to L_\beta^{\frac{1}{2}}u \tag*{(40)} L β 2 1 ( 1 + ε L β ) − 1 u = ( 1 + ε L β ) − 1 L β 2 1 u → L β 2 1 u (40)
in L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) as ε → 0 \varepsilon\to0 ε → 0 , see for instance [42 Prop. 2.1.1] . Now, fix some arbitrary ε > 0 \varepsilon> 0 ε > 0 . Using the triangle inequality, we estimate
∥ L α 1 2 u − L α n 1 2 u ∥ L 2 ≤ ∥ L α 1 2 u − L α 1 2 ( 1 + ε L α ) − 1 u ∥ L 2 + ∥ L α 1 2 ( 1 + ε L α ) − 1 u − L α n 1 2 ( 1 + ε L α n ) − 1 u ∥ L 2 + ∥ L α n 1 2 u − L α n 1 2 ( 1 + ε L α n ) − 1 u ∥ L 2 = : I ( ε ) + I I ( n , ε ) + I I I ( n , ε ) . (41) \begin{aligned}
\|L_\alpha^{\frac{1}{2}}u-L_{\alpha_n}^{\frac{1}{2}}u\|_{L^2}
&\leq\|L_\alpha^{\frac{1}{2}}u-L_\alpha^{\frac{1}{2}}(1+\varepsilon L_\alpha)^{-1}u\|_{L^2}\\
&\quad+\|L_\alpha^{\frac{1}{2}}(1+\varepsilon L_\alpha)^{-1}u-L_{\alpha_n}^{\frac{1}{2}}(1+\varepsilon L_{\alpha_n})^{-1}u\|_{L^2}\\
&\quad+\|L_{\alpha_n}^{\frac{1}{2}}u-L_{\alpha_n}^{\frac{1}{2}}(1+\varepsilon L_{\alpha_n})^{-1}u\|_{L^2}\\
&=: \mathrm{I}(\varepsilon)+\mathrm{II}(n,\varepsilon)+\mathrm{III}(n,\varepsilon). \tag*{(41)}
\end{aligned} ∥ L α 2 1 u − L α n 2 1 u ∥ L 2 ≤ ∥ L α 2 1 u − L α 2 1 ( 1 + ε L α ) − 1 u ∥ L 2 + ∥ L α 2 1 ( 1 + ε L α ) − 1 u − L α n 2 1 ( 1 + ε L α n ) − 1 u ∥ L 2 + ∥ L α n 2 1 u − L α n 2 1 ( 1 + ε L α n ) − 1 u ∥ L 2 =: I ( ε ) + II ( n , ε ) + III ( n , ε ) . (41)
By Step 1, we know that I I ( n , ε ) → 0 \mathrm{II}(n,\varepsilon) \to0 II ( n , ε ) → 0 as n → ∞ n \to\infty n → ∞ . Furthermore, by Theorem 1.2 with α 0 = α 2 \alpha_0=\frac{\alpha}{2} α 0 = 2 α (or Proposition 1.4 if α n = 1 \alpha_n=1 α n = 1 or α = 1 \alpha=1 α = 1 ), we have
I I I ( n , ε ) ≲ ∥ ∇ α n ( u − ( 1 + ε L α n ) − 1 u ) ∥ L 2 , \mathrm{III}(n,\varepsilon)\lesssim\|\nabla^{\alpha_n}(u-(1+\varepsilon L_{\alpha_n})^{-1}u)\|_{L^2}, III ( n , ε ) ≲ ∥ ∇ α n ( u − ( 1 + ε L α n ) − 1 u ) ∥ L 2 ,
where the implicit constant does not depend on n n n and ε \varepsilon ε . Thus, applying Lemma 5.6 and Lemma 5.7 , we see that
lim sup n → ∞ I I I ( n , ε ) ≲ lim sup n → ∞ ∥ ∇ α n ( u − ( 1 + ε L α n ) − 1 u ) ∥ L 2 = ∥ ∇ α ( u − ( 1 + ε L α ) − 1 u ) ∥ L 2 \begin{aligned}
\limsup_{n\to\infty}\mathrm{III}(n,\varepsilon)
\lesssim\limsup_{n\to\infty}\|\nabla^{\alpha_n}(u-(1+\varepsilon L_{\alpha_n})^{-1}u)\|_{L^2} \\
&= \|\nabla^\alpha(u-(1+\varepsilon L_\alpha)^{-1}u)\|_{L^2}
\end{aligned} n → ∞ lim sup III ( n , ε ) ≲ n → ∞ lim sup ∥ ∇ α n ( u − ( 1 + ε L α n ) − 1 u ) ∥ L 2 = ∥ ∇ α ( u − ( 1 + ε L α ) − 1 u ) ∥ L 2
≲ ∥ L α 1 2 u − L α 1 2 ( 1 + ε L α ) − 1 u ∥ L 2 , \lesssim\|L_\alpha^{\frac{1}{2}}u-L_\alpha^{\frac{1}{2}}(1+\varepsilon L_\alpha)^{-1}u\|_{L^2}, ≲ ∥ L α 2 1 u − L α 2 1 ( 1 + ε L α ) − 1 u ∥ L 2 ,
where we used Theorem 1.2 (or Proposition 1.4 if α = 1 \alpha=1 α = 1 ) in the last step. Using both observations on I I \mathrm{II} II and I I I \mathrm{III} III in (5.9), we get
lim sup n → ∞ ∥ L α 1 2 u − L α n 1 2 u ∥ L 2 ≤ I ( ε ) + lim sup n → ∞ I I I ( n , ε ) ≲ I ( ε ) . \limsup_{n\to\infty}\|L_\alpha^{\frac{1}{2}}u-L_{\alpha_n}^{\frac{1}{2}}u\|_{L^2}
\leq\mathrm{I}(\varepsilon)+\limsup_{n\to\infty}\mathrm{III}(n,\varepsilon)
\lesssim\mathrm{I}(\varepsilon). n → ∞ lim sup ∥ L α 2 1 u − L α n 2 1 u ∥ L 2 ≤ I ( ε ) + n → ∞ lim sup III ( n , ε ) ≲ I ( ε ) .
Since this holds for every ε > 0 \varepsilon>0 ε > 0 , using (5.8) and passing to ε → 0 \varepsilon\to0 ε → 0 yields
lim sup n → ∞ ∥ L α 1 2 u − L α n 1 2 u ∥ L 2 = 0 , \limsup_{n\to\infty}\|L_\alpha^{\frac{1}{2}}u-L_{\alpha_n}^{\frac{1}{2}}u\|_{L^2}=0, n → ∞ lim sup ∥ L α 2 1 u − L α n 2 1 u ∥ L 2 = 0 ,
which proves the claim. □
AI disclosure statement. The author used AI tools for language editing and for checking clarity and internal consistency. The mathematical ideas, proofs and results were developed by the author, who takes full responsibility for the manuscript.
References
Karlsruhe Institute of Technology, Department of Mathematics, 76131 Karlsruhe, Germany
Email address: luca.haardt@kit.edu