Preprint
Article

This version is not peer-reviewed.

Fractional Landau Inequalities for Rotating Compressible Flows

Submitted:

12 November 2026

Posted:

13 November 2026

You are already at the latest version

Abstract
We develop a rigorous theory of anisotropic mixed fractional Landau inequalities for rotating compressible fluid flows at high Mach numbers, incorporating Coriolis and centrifugal forces. We introduce rotating fractional Sobolev spaces \( \mathcal{W}^{\nu,p}_{\alpha,\Omega}(\mathbb{R}^k) \), which encode directional scaling, fractional dissipation, and rotational effects through a composition of anisotropic dilations and rotations. We prove fractional Landau inequalities with explicit dependence on the Mach number (polynomial), the rotation rate \( |\Omega| \) (exponential), and the anisotropy vector \( \alpha \), with constants of the form \( C(\nu,p,\alpha,\Omega) = C_0 \exp\bigl(c_0|\Omega|\sum_{i=1}^k \alpha_i^{-1}\bigr) \), where \( c_0 = 2^{1/\nu} \) is an explicit universal constant and \( C_0 = 2^{1+1/\nu}\,\kappa(p,\alpha)^{1/\nu}\,\Gamma(1-\nu/\alpha_{\min})^{-1/\nu} \) depends only on \( \nu,p,\alpha \), with the restriction \( \alpha_{\min} := \min_i \alpha_i > 2\nu \). Key tools are a rotating Littlewood--Paley decomposition, anisotropic maximal estimates with rotation-independent constants, and commutator estimates for the Coriolis term based on the Calderón--Vaillancourt theorem. As applications, we establish stability bounds for neural operators approximating rotating compressible flows and derive approximation rates of order \( N^{-\nu/d_{\alpha,\Omega}} \), where \( d_{\alpha,\Omega}=\sum_i\alpha_i^{-1}+\kappa_{\mathrm{rot}}|\Omega|^{2/\nu} \) is the effective anisotropic--rotational dimension. Our results provide a mathematical foundation for analysing high--Mach rotating flows and for designing physically informed neural architectures.
Keywords: 
;  ;  ;  ;  ;  

MSC:  26A33; 35Q31; 76U05; 46E35; 68T07

1. Introduction

The analysis of compressible fluid flows at high Mach numbers is central to modern aerodynamics, astrophysics, and engineering [22]. When such flows are subjected to rapid rotation — a situation encountered in turbomachinery, atmospheric dynamics, and stellar interiors — the Coriolis and centrifugal forces fundamentally alter the energy balance and regularity properties of solutions. Classical Landau inequalities [1] and their fractional extensions [28] provide essential estimates for the growth of derivatives in terms of lower-order norms, yet they are inadequate for rotating systems because they ignore the rotational coupling between directions.
The classical Landau inequality
∥ f ′ ∥ L ∞ ( R ) ≤ 2 ∥ f ∥ L ∞ ( R ) 1 / 2 ∥ f ′ ′ ∥ L ∞ ( R ) 1 / 2
was established by Landau in 1925 [1] and remains a cornerstone of mathematical analysis. This inequality controls the magnitude of the first derivative by the amplitude and the second derivative of the function. The generalisation to arbitrary orders k < n on the real line was achieved by Kolmogorov [2], who found the sharp constants
C ( n , k , R ) = a n − k a n − 1 + k / n ,
where a n are the Favard constants. For functions on a half-line, Schoenberg and Cavaretta [18] found the extremising functions, though explicit forms for the sharp constants remain unknown in many cases.
The theory has evolved along several independent lines. Stein’s theory of singular integrals and differentiability properties of functions [3] provided the analytic machinery for maximal functions and multiplier theorems. The foundational treatise of Adams and Fournier [19] established the classical Sobolev embedding theorems, while Ditzian extended the Landau–Kolmogorov inequality to multivariate settings. Lions [6] introduced the concentration-compactness method, which became a fundamental tool for variational problems. Samko’s monograph [9] systematised fractional integration and differentiation. Kounchev [14] studied extremisers for the multivariate Landau–Kolmogorov inequality, and DeVore [15] established the nonlinear approximation framework used in the neural network analysis.
Modern developments include Triebel’s theory of anisotropic Besov spaces [20], which supplied the function space framework for the anisotropic mixed fractional Sobolev spaces used in this paper. Maz’ya and Shaposhnikova [16] proved pointwise interpolation inequalities involving the Hardy–Littlewood maximal operator, providing a direct precursor to our approach. The treatise of Lieb and Loss [17] provided the interpolation inequalities needed for the fractional setting. Runst and Sickel [13] established the theory of Sobolev spaces of fractional order, while Bahouri, Chemin, and Danchin [21] developed the Littlewood–Paley framework for nonlinear partial differential equations. Yarotsky [25] established error bounds for approximations with deep ReLU networks, while Schäfer, Ullrich, and Vedel [27] developed hyperbolic wavelet analysis for anisotropic Besov–Sobolev spaces.
Recent developments (2025–2026) include Anastassiou’s study of multivariate Canavati fractional Landau inequalities [28], primarily in isotropic settings. The Marchaud derivative, which plays a crucial role in our proofs, has seen renewed interest as documented in the historical survey by Ferrari [24]. For the analysis of fractional Navier–Stokes equations, we refer to the recent work of Colombo, De Lellis, and Massaccesi [26], and for the classical theory of rotating fluids, to Pedlosky [5] and Constantin and Foias [8].
Despite these advances, a comprehensive theory of anisotropic mixed fractional Landau inequalities for rotating compressible flows has been lacking. This paper bridges this gap by developing such a theory. Our framework simultaneously accounts for:
  • Directional scaling encoded by a vector α = ( α 1 , … , α k ) , which captures different regularity rates along distinct spatial axes—a necessity for compressible flows where shock waves and boundary layers create strong anisotropy [20,27].
  • Fractional dissipation of order ν modelled via the anisotropic fractional Laplacian Δ α ν , which naturally appears in the compressible Navier–Stokes equations with fractional viscosity [9,26].
  • Rotational effects generated by a constant rotation vector Ω , giving rise to the Coriolis term 2 Ω × u and the centrifugal potential [5,22].
Our main contributions are the following.
1.
Rotating fractional Sobolev spaces W α , Ω ν , p ( R k ) . We define these spaces by incorporating the rotation vector Ω into the anisotropic Gagliardo seminorm through a composition of anisotropic dilations and rotations. The new norm includes a term that measures the fractional derivative in a frame rotating with angular velocity Ω , thereby respecting the symmetries of the Coriolis operator. We prove the equivalence of several characterisations (Gagliardo, Littlewood–Paley, semigroup, K-functional) for these spaces, extending the classical theory of anisotropic Besov spaces [20].
2.
Fractional Landau inequalities with Coriolis corrections. For a function u belonging to W α , Ω ν , p ( R k ) we prove
∥ D i 1 u ∥ L ∞ ≤ C ( ν , p , α , Ω ) ∥ u ∥ L ∞ 1 − 1 / ν ∑ j = 1 k D j ν , α j , Ω u L ∞ 1 / ν ,
where D j ν , α j , Ω is a rotating fractional derivative that reduces to the classical Weyl derivative [24] when Ω = 0 . The constant C ( ν , p , α , Ω ) is made explicit and depends on | Ω | in a way that reflects the stabilising effect of rotation at high Mach numbers. This extends the classical Landau–Kolmogorov inequalities [1,2,18] and the recent fractional results [28] to the anisotropic rotating setting.
3.
Neural operator bounds for rotating compressible flows. Using the new inequalities we prove stability and approximation theorems for deep neural networks that approximate solutions of the rotating compressible Navier–Stokes equations. The approximation rate becomes N − ν / d α , Ω where d α , Ω = ∑ i α i − 1 + κ rot | Ω | 2 / ν . Since d α , Ω increases with | Ω | , rotation increases the effective anisotropic–rotational dimension and therefore slows the achievable approximation rate, in contrast to the naive expectation that rotation would regularise the flow. These results build upon the approximation theory of nonlinear functions by deep networks [25] and the general theory of nonlinear approximation [15].
The paper is organised as follows. Section 2 recalls the anisotropic scaling structure and introduces the rotating fractional Sobolev spaces W α , Ω ν , p ( R k ) , including the equivalent characterisations, the rotating Littlewood–Paley decompositions, the anisotropic Bernstein inequalities, and the anisotropic maximal estimates with rotation-independent constants. Section 3 contains the main fractional Landau inequalities (first-order and higher-order) with complete rigorous proofs. Section 4 applies these inequalities to the rotating compressible Navier–Stokes system, deriving a priori estimates for the velocity gradient at high Mach numbers. Section 5 establishes stability and approximation theorems for neural operators adapted to the rotating anisotropic geometry, including the approximation rate N − ν / d α , Ω . Section 6 summarises the key theoretical results and Section 7 concludes the paper with a discussion of future work.

2. Rotating Anisotropic Function Spaces

2.1. Anisotropic Scaling and Rotation Group

Let α = ( α 1 , … , α k ) ∈ ( 0 , ∞ ) k be the anisotropy vector. The anisotropic dilation group { T λ α } λ > 0 is defined by
T λ α x = ( λ α 1 x 1 , … , λ α k x k ) , λ > 0 , x ∈ R k .
This is a one–parameter group of linear transformations with infinitesimal generator A α = diag ( α 1 , … , α k ) , i.e., T λ α = e ( log λ ) A α .
For a constant rotation vector Ω ∈ R k , denote by L Ω the skew–symmetric matrix representing the linear map x ↦ Ω × x (the cross product). Explicitly, ( L Ω ) i j = − ε i j ℓ Ω ℓ , where ε i j ℓ is the Levi-Civita symbol. The rotation group { R ( θ ) } θ ∈ R is given by R ( θ ) = e θ L Ω / | Ω | when Ω ≠ 0 , and R ( 0 ) = Id for Ω = 0 . These rotations satisfy R ( θ 1 ) R ( θ 2 ) = R ( θ 1 + θ 2 ) .
Remark 2.1 
(Non-commutativity of dilation and rotation). The operators A α and L Ω donotcommute in general. Indeed, for k = 3 and Ω = ( 0 , 0 , ω ) , the commutator [ A α , L Ω ] = A α L Ω − L Ω A α has entries
( [ A α , L Ω ] ) 12 = ( α 1 − α 2 ) ω , ( [ A α , L Ω ] ) 21 = − ( α 1 − α 2 ) ω ,
which vanishes if and only if α 1 = α 2 or ω = 0 . By the Baker–Campbell–Hausdorff formula, one has e X e Y = e X + Y + 1 2 [ X , Y ] + … ; in particular, e X e Y = e X + Y holds if and only if [ X , Y ] = 0 . Consequently, the composition T λ α ∘ R ( θ ) = e ( log λ ) A α e θ L Ω / | Ω | doesnotequal the single exponential exp ( log λ A α + θ L Ω / | Ω | ) in general. 
Definition 2.2 
(Rotating anisotropic scaling). For λ > 0 and θ ∈ R , define therotating anisotropic scalingas the composition
T ˜ λ , θ α , Ω : = T λ α ∘ R ( θ ) ,
where T λ α is the anisotropic dilation (2) and R ( θ ) is the rotation about the axis Ω / | Ω | . 
Proposition 2.3 
(Properties of the Rotating Scaling Family). The family { T ˜ λ , θ α , Ω } λ > 0 , θ ∈ R , where T ˜ λ , θ α , Ω = T λ α ∘ R ( θ ) , satisfies:
1.
Composition law: T ˜ λ 1 , θ 1 α , Ω ∘ T ˜ λ 2 , θ 2 α , Ω = T ˜ λ 1 λ 2 , θ 1 + θ 2 α , Ω if and only if R ( θ 1 ) and T λ 2 α commute, i.e., when α 1 = … = α k or θ 1 = 0 . In particular, { T ˜ λ , 0 α , Ω } λ > 0 is a one-parameter group isomorphic to ( R + , · ) . 
2.
Jacobian determinant: | det D T ˜ λ , θ α , Ω | = λ D α , where D α = ∑ i = 1 k α i . 
3.
Measure scaling:For any f ∈ L loc 1 ( R k ) ,
∫ R k f ( T ˜ λ , θ α , Ω x ) d x = λ − D α ∫ R k f ( x ) d x .
4.
Anisotropic ball preservation: R ( θ ) preserves the anisotropic ball B α ( 0 , r ) if and only if θ = 0 ; if, in addition, α 1 = … = α k and θ is a multiple of π / 2 (in k = 2 ) or a coordinate permutation (in general), then R ( θ ) permutes the coordinate axes. In the case θ = 0 , T ˜ λ , 0 α , Ω ( B α ( 0 , r ) ) = B α ( 0 , λ r ) by the homogeneity of the anisotropic dilation. 
Proof. (1) By definition, T ˜ λ 1 , θ 1 α , Ω ∘ T ˜ λ 2 , θ 2 α , Ω = T λ 1 α R ( θ 1 ) T λ 2 α R ( θ 2 ) . Since T λ 1 α T λ 2 α = T λ 1 λ 2 α and R ( θ 1 ) R ( θ 2 ) = R ( θ 1 + θ 2 ) , the composition equals T ˜ λ 1 λ 2 , θ 1 + θ 2 α , Ω if and only if R ( θ 1 ) T λ 2 α = T λ 2 α R ( θ 1 ) . The commutator [ T λ α , R ( θ ) ] = T λ α R ( θ ) − R ( θ ) T λ α vanishes if and only if R ( θ ) acts trivially on the eigenspaces of A α , i.e., if and only if α 1 = … = α k or θ = 0 (as shown in Remark 2.1).
(2) Since det R ( θ ) = 1 and det T λ α = λ D α , the chain rule yields det D T ˜ λ , θ α , Ω = det ( T λ α ) det ( R ( θ ) ) = λ D α . The absolute value is redundant because λ > 0 .
(3) Follows from (2) by the change of variables formula y = T ˜ λ , θ α , Ω x , d y = λ D α d x .
(4) The anisotropic ball B α ( 0 , r ) = { y : | y i | < r 1 / α i } is invariant under R ( θ ) if and only if R ( θ ) preserves the coordinate axes, which for θ ≠ 0 requires the anisotropy to be isotropic and θ to be a symmetry of the coordinate cross. In the case θ = 0 , T λ α B α ( 0 , r ) = B α ( 0 , λ r ) by the homogeneity of the anisotropic dilation. □
Remark 2.4. 
The rotating anisotropic scaling T ˜ λ , θ α , Ω doesnotform a one-parameter group unless Ω = 0 or α 1 = … = α k . For the purposes of this paper, we work with the semigroup generated by the composition, which is sufficient for the Littlewood–Paley decomposition and the Landau inequality. 

2.1.1. The Rotating Finite Difference

For a direction e i ∈ R k (the i-th canonical basis vector) and a rotation vector Ω ∈ R k , recall that L Ω denotes the skew-symmetric matrix representing the linear map x ↦ Ω × x , i.e., ( L Ω ) i j = − ε i j ℓ Ω ℓ . Equivalently, L Ω is the unique matrix satisfying L Ω x = Ω × x for all x ∈ R k ; it is skew-symmetric, L Ω ⊤ = − L Ω , with purely imaginary spectrum { 0 , ± i | Ω | } and kernel span { Ω } . In particular, L Ω e i = Ω × e i ∈ R k is orthogonal to both Ω and e i (when k = 3 ), and 〈 L Ω e i , e i 〉 = 0 .
Definition 2.5 
(Rotating finite difference). For h ∈ R and f : R k → R measurable, define therotating finite differencein direction e i by
Δ h , i Ω f ( x ) : = f ( x + h e i ) − f ( x + h L Ω e i ) , x ∈ R k .
The map h ↦ Δ h , i Ω f ( x ) is linear in the increments but nonlinear in f; it is antisymmetric under h ↦ − h and i ↦ − i in the sense that Δ − h , i Ω f ( x ) = − Δ h , i Ω , * f ( x ) , where Δ h , i Ω , * is the adjoint difference defined in Lemma 2.8(4). 
Remark 2.6 
(Dimensional consistency). The definition (5) is dimensionally consistent. If Ω has units of inverse length (as is standard for angular velocity in a rotating frame, since | Ω | has units of s − 1 and velocities have units of length · s − 1 ), then h L Ω e i has units of length, matching h e i . The two evaluation points x + h e i and x + h L Ω e i are separated by the vector h ( e i − L Ω e i ) of Euclidean length | h | 1 + | L Ω e i | 2 ≥ | h | , with equality if and only if Ω ‖ e i . The rotating difference thus samples f along the oblique direction e i − L Ω e i , which is the natural geometric object associated with the rotating frame. 
Remark 2.7 
(Relation to the standard finite difference). When Ω = 0 , L Ω = 0 and Δ h , i 0 f ( x ) = f ( x + h e i ) − f ( x ) , recovering the standard forward finite difference. When Ω ≠ 0 , the rotating finite difference isnota difference along a straight line in the direction e i , but along the “twisted” direction e i − L Ω e i = e i − Ω × e i , which accounts for the Coriolis coupling between the i-th coordinate direction and the rotation axis. This is the natural finite-difference analogue of the rotating fractional derivative D i ν , α i , Ω defined in (59) (Section 3). 
Lemma 2.8 
(Basic properties of the rotating finite difference). For every f , g : R k → R measurable, h ∈ R , x ∈ R k , and i = 1 , … , k :
1.
Linearity: Δ h , i Ω ( a f + b g ) ( x ) = a Δ h , i Ω f ( x ) + b Δ h , i Ω g ( x ) for all a , b ∈ R . 
2.
Translation invariance: Δ h , i Ω ( τ y f ) ( x ) = Δ h , i Ω f ( x + y ) , where τ y f ( x ) : = f ( x − y ) . 
3.
Scaling: Δ h , i Ω ( f ∘ T λ α ) ( x ) = Δ λ − 1 / α i h , i Ω ′ ( f ) ( T λ α x ) with Ω ′ = λ − 1 / α i Ω , where T λ α is the anisotropic dilation (2). 
4.
Adjoint identity: ∫ R k Δ h , i Ω f ( x ) g ( x ) d x = − ∫ R k f ( x ) Δ − h , i Ω , * g ( x ) d x , where Δ − h , i Ω , * g ( x ) : = g ( x − h e i ) − g ( x − h L Ω e i ) is the adjoint difference. 
5.
Bound: | Δ h , i Ω f ( x ) | ≤ 2 ∥ f ∥ L ∞ for a.e. x. 
Proof. 
Properties (1), (2), (5) are immediate from the definition. Property (3) follows from the change of variables x ↦ T λ α x and the identity T λ α ( x + h e i ) = T λ α x + h λ α i e i (equivalently, T λ α ( x + λ − 1 / α i h e i ) = T λ α x + h e i ), together with T λ α L Ω e i = L λ − 1 / α i Ω ( T λ α e i ) = λ L λ − 1 / α i Ω e i ; the scaling of the rotation vector follows from the definition of L Ω . Property (4) is a standard duality computation using Fubini’s theorem and the skew-symmetry of L Ω . □

2.1.2. The Rotating Gagliardo Seminorm

Definition 2.9 
(Rotating directional Gagliardo seminorm). Let ν > 0 , 1 ≤ p < ∞ , α i > 0 , and Ω ∈ R k . For a measurable function f : R k → R , define therotating directional Gagliardo seminormby
[ f ] W i ν / α i , p , Ω : = ∫ R k ∫ R | Δ h , i Ω f ( x ) | p | h | 1 + p ν / α i d h d x 1 / p .
Therotating mixed fractional Sobolev space W α , Ω ν , p ( R k ) is the completion of C c ∞ ( R k ) under the norm
∥ f ∥ W α , Ω ν , p : = ∥ f ∥ L p ( R k ) + ∑ i = 1 k [ f ] W i ν / α i , p , Ω .
Remark 2.10 
(Well-definedness). The integral (6) is well-defined as an element of [ 0 , + ∞ ] for every measurable f; for f ∈ C c ∞ ( R k ) and 0 < ν / α i < 1 , it is finite because | Δ h , i Ω f ( x ) | ≤ | h | ∥ ∇ f ∥ L ∞ for small | h | and | Δ h , i Ω f ( x ) | ≤ 2 ∥ f ∥ L ∞ for large | h | , so the integral converges at both h = 0 and h = ∞ . For ν / α i ≥ 1 , the first-order difference is replaced by a higher-order difference (see Remark 2.14). 
Remark 2.11 
(Comparison with the standard Gagliardo seminorm). When Ω = 0 , the rotating seminorm reduces to the classical anisotropic Gagliardo seminorm
[ f ] W i ν / α i , p = ∫ R k ∫ R | f ( x + h e i ) − f ( x ) | p | h | 1 + p ν / α i d h d x 1 / p .
When Ω ≠ 0 , the rotation introduces a correction of order | Ω | in the difference operator, which is reflected in the Fourier characterisation below. 
Proposition 2.12 
(Basic properties of W α , Ω ν , p ). The space W α , Ω ν , p ( R k ) is a Banach space. Moreover:
1.
Completeness: W α , Ω ν , p is complete with respect to the norm (7). 
2.
Density: C c ∞ ( R k ) is dense in W α , Ω ν , p by construction; moreover, the Schwartz class S ( R k ) is also dense. 
3.
Embedding:If 1 ≤ p ≤ q < ∞ and ν − D α ( 1 / p − 1 / q ) ≥ 0 , then W α , Ω ν , p ↪ W α , Ω ν − D α ( 1 / p − 1 / q ) , q (anisotropic Sobolev embedding). 
4.
Reflexivity:For 1 < p < ∞ , W α , Ω ν , p is reflexive. 
5.
Independence of Ω up to isomorphism:For every Ω , Ω ′ ∈ R k with | Ω | = | Ω ′ | , the spaces W α , Ω ν , p and W α , Ω ′ ν , p are isomorphic, since the rotation matrices R ( θ ) and R ( θ ′ ) are conjugate by an orthogonal transformation. 
Proof. (1) Completeness follows from the standard argument: a Cauchy sequence in W α , Ω ν , p is Cauchy in L p and in each seminorm; by the completeness of L p and the Fatou lemma applied to the seminorm, the limit belongs to W α , Ω ν , p and the sequence converges in the norm.
(2) Density of C c ∞ is by definition. Density of S follows from the fact that S ⊂ W α , Ω ν , p (since Schwartz functions have all derivatives decaying rapidly) and C c ∞ is dense in S in the W α , Ω ν , p norm (standard mollification argument).
(3) The anisotropic Sobolev embedding is proved by the same Littlewood–Paley argument as in Theorem Section 3.1, applied to the difference between p and q.
(4) Reflexivity follows from the fact that W α , Ω ν , p is isomorphic to a closed subspace of L p × ( L p ) k via the map f ↦ ( f , D 1 ν , α 1 , Ω f , … , D k ν , α k , Ω f ) .
(5) If | Ω | = | Ω ′ | , there exists an orthogonal transformation Q with Q Ω = Ω ′ . The map f ↦ f ∘ Q is an isomorphism of W α , Ω ν , p onto W α , Ω ′ ν , p . □
Lemma 2.13 
(Fourier characterisation for p = 2 ). Let ν > 0 with 0 < ν / α i < 1 / 2 and Ω ∈ R k . Then for every f ∈ S ( R k ) the followingexact identityholds:
[ f ] W i ν / α i , 2 , Ω 2 = Λ ν , α i ∫ R k | ξ i + ( Ω × e i ) · ξ | 2 ν / α i | f ^ ( ξ ) | 2 d ξ ,
with the explicit constant
Λ ν , α i : = 4 π 2 ν / α i ν / α i · 1 Γ ( 1 − 2 ν / α i ) .
where Γ is the Euler Gamma function. 
Proof. 
The proof proceeds by expressing the seminorm in Fourier variables, evaluating the resulting one-dimensional integral in closed form, and identifying the constants. Throughout, we assume 0 < ν / α i < 1 / 2 and f ∈ S ( R k ) , so that all integrals converge absolutely; the general case follows by density.
Fourier representation of the rotating seminorm. 
Since Δ h , i Ω f ∈ L 2 ( R k ) for every h ∈ R and every f ∈ S ( R k ) , Plancherel’s theorem in the x-variable yields
∫ R k | Δ h , i Ω f ( x ) | 2 d x = ∫ R k | Δ h , i Ω f ^ ( ξ ) | 2 d ξ .
Using the translation identity f ( · + a ) ^ ( ξ ) = e 2 π i a · ξ f ^ ( ξ ) together with L Ω e i = Ω × e i , the Fourier transform of the rotating finite difference (5) is
Δ h , i Ω f ^ ( ξ ) = f ^ ( ξ ) e 2 π i h ξ i − e − 2 π i h ( Ω × e i ) · ξ .
Indeed, the first term gives e 2 π i h ξ i f ^ ( ξ ) and the second term gives − e − 2 π i h ( Ω × e i ) · ξ f ^ ( ξ ) . Substituting (11) into the definition (6) with p = 2 , we obtain
[ f ] W i ν / α i , 2 , Ω 2 = ∫ R 1 | h | 1 + 2 ν / α i ∫ R k | f ^ ( ξ ) | 2 | e 2 π i h ξ i − e − 2 π i h ( Ω × e i ) · ξ | 2 d ξ d h .
Because | f ^ ( ξ ) | 2 ( 1 + | ξ | 2 ) s is integrable for every s > 0 when f ∈ S ( R k ) , and because the h-integrand is dominated by min ( | h | 2 | ω i ( ξ ) | 2 , 4 ) | h | − 1 − 2 ν / α i , the double integral is absolutely convergent. Fubini’s theorem therefore permits the exchange of the order of integration:
[ f ] W i ν / α i , 2 , Ω 2 = ∫ R k | f ^ ( ξ ) | 2 ∫ R | e 2 π i h ξ i − e − 2 π i h ( Ω × e i ) · ξ | 2 | h | 1 + 2 ν / α i d h d ξ .
Closed-form evaluation of the inner integral. 
For fixed ξ ∈ R k , introduce the rotating frequency
ω i ( ξ ) : = ξ i + ( Ω × e i ) · ξ ,
which encodes the combined effect of the i-th coordinate direction and the Coriolis coupling. The phase in (13) factorises as
e 2 π i h ξ i − e − 2 π i h ( Ω × e i ) · ξ = e 2 π i h ξ i 1 − e − 2 π i h ω i ( ξ ) ,
and since | e 2 π i h ξ i | = 1 , the modulus reduces to
| e 2 π i h ξ i − e − 2 π i h ( Ω × e i ) · ξ | 2 = | 1 − e − 2 π i h ω i ( ξ ) | 2 = 4 sin 2 ( π h ω i ( ξ ) ) .
Consequently, the inner integral in (13) becomes
I ν , α i ( ξ ) : = 4 ∫ R sin 2 ( π h ω i ( ξ ) ) | h | 1 + 2 ν / α i d h .
The change of variables s = h | ω i ( ξ ) | (valid for ω i ( ξ ) ≠ 0 ; the case ω i ( ξ ) = 0 gives I ν , α i ( ξ ) = 0 , consistent with the final formula) yields h = s / | ω i ( ξ ) | , d h = d s / | ω i ( ξ ) | , and therefore
I ν , α i ( ξ ) = 4 | ω i ( ξ ) | 2 ν / α i ∫ R sin 2 ( π s ) | s | 1 + 2 ν / α i d s .
The remaining one-dimensional integral is a Beta-type integral. Using the classical identity [23]
∫ 0 ∞ sin 2 ( π s ) s 1 + 2 β d s = π 2 β 2 β · 1 Γ ( 1 − 2 β ) , 0 < β < 1 2 ,
with β = ν / α i , and exploiting the evenness of the integrand, we obtain
I ν , α i ( ξ ) = 4 | ω i ( ξ ) | 2 ν / α i · 2 · π 2 ν / α i 2 ν / α i · 1 Γ ( 1 − 2 ν / α i ) = Λ ν , α i | ω i ( ξ ) | 2 ν / α i ,
where Λ ν , α i is the constant defined in (9). The constant Λ ν , α i is finite and strictly positive precisely when 0 < ν / α i < 1 / 2 , i.e., α i > 2 ν .
Identification of the constants and the Fourier multiplier. 
Substituting (19) into (13) and recalling the definition (14) of ω i ( ξ ) ,
[ f ] W i ν / α i , 2 , Ω 2 = Λ ν , α i ∫ R k | ω i ( ξ ) | 2 ν / α i | f ^ ( ξ ) | 2 d ξ = Λ ν , α i ∫ R k | ξ i + ( Ω × e i ) · ξ | 2 ν / α i | f ^ ( ξ ) | 2 d ξ .
This is an exact identity, not merely an equivalence: the seminorm [ f ] W i ν / α i , 2 , Ω 2 is precisely Λ ν , α i times the weighted L 2 norm of f ^ with weight | ξ i + ( Ω × e i ) · ξ | 2 ν / α i . In particular, no information is lost in the Fourier characterisation for p = 2 .
Behaviour of the multiplier and the role of the rotation. 
The Fourier multiplier ω i ( ξ ) = ξ i + ( Ω × e i ) · ξ admits the factorisation
ω i ( ξ ) = e i + Ω × e i · ξ = 〈 η i , ξ 〉 , η i : = e i + Ω × e i ,
so that | ω i ( ξ ) | 2 ν / α i = | 〈 η i , ξ 〉 | 2 ν / α i . The vector η i has Euclidean norm | η i | = 1 + | Ω × e i | 2 ≥ 1 , with equality if and only if Ω is parallel to e i . In particular:
  • When Ω = 0 : η i = e i and | ω i ( ξ ) | = | ξ i | , recovering the classical anisotropic Fourier multiplier | ξ i | ν / α i .
  • When Ω ≠ 0 : the multiplier | ω i ( ξ ) | ν / α i is not comparable to | ξ i | ν / α i uniformly in ξ ; indeed, for ξ orthogonal to e i but not to Ω × e i , the classical multiplier vanishes while the rotating one does not. The rotating multiplier is, however, comparable to | ξ | ν / α i (up to constants depending on Ω and α ), which is the source of the exponential factor e c | Ω | in the Landau inequalities.
  • Isotropy of the multiplier: the multiplier depends on Ω only through the direction of Ω and the norm | Ω | ; rotations of Ω leave the seminorm invariant up to an orthogonal change of variables, consistent with Proposition 2.12(5).
The case ν / α i ≥ 1 / 2 and the role of the higher-order difference. 
The finiteness of Λ ν , α i requires the strict inequality 0 < ν / α i < 1 / 2 , equivalently α i > 2 ν . When ν / α i ≥ 1 / 2 , the integral (16) diverges at h = 0 because sin 2 ( π h ω i ( ξ ) ) ∼ ( π h ω i ( ξ ) ) 2 and the integrand behaves like | h | 1 − 2 ν / α i , which is not integrable if 2 ν / α i ≥ 1 . In this regime, the first-order rotating difference (5) must be replaced by a higher-order difference (see Remark 2.14).
From p = 2 to general p. 
The Fourier method exploited above is specific to p = 2 : it relies on Plancherel’s isometry and the factorisation of the modulus of the difference, neither of which extends to p ≠ 2 . For 1 < p < ∞ with p ≠ 2 , the space W α , Ω ν , p is characterised not by a Fourier multiplier but by the Littlewood–Paley square function
∥ f ∥ W α , Ω ν , p ≍ ∥ f ∥ L p + ∑ j ∈ Z 2 j ν p | Δ j f | p 1 / p L p ,
where { Δ j } j ∈ Z are the anisotropic Littlewood–Paley blocks. This is the standard characterisation of anisotropic Triebel–Lizorkin spaces F p , 2 ν ( R k ) (see [20] and Remark 2.16). The proof of this equivalence proceeds by the Calderón reproducing formula and the almost-orthogonality of the Littlewood–Paley decomposition; it does not require the Fourier multiplier representation, which is genuinely a p = 2 phenomenon. This dichotomy between p = 2 and p ≠ 2 is a well-known feature of fractional Sobolev spaces and reflects the failure of the Fourier transform to preserve L p for p ≠ 2 . □
Remark 2.14 
(Higher-order differences). When ν / α i ≥ 1 / 2 , the integral (6) diverges at h = 0 because | Δ h , i Ω f ( x ) | 2 ∼ h 2 and the integrand behaves like | h | 1 − 2 ν / α i , which is not integrable if 2 ν / α i ≥ 1 . In this case, one replaces the first-order difference by a higher-order difference
Δ h , i Ω , m f ( x ) : = ∑ j = 0 m ( − 1 ) j m j f ( x + ( m − j ) h e i + j h L Ω e i ) ,
with m = ⌈ ν / α i ⌉ + 1 . The resulting seminorm
[ f ] W i ν / α i , p , Ω p : = ∫ R k ∫ R | Δ h , i Ω , m f ( x ) | p | h | 1 + p ν / α i d h d x
is then well-defined for all ν > 0 , and the Fourier characterisation becomes
[ f ] W i ν / α i , 2 , Ω 2 ≍ ∫ R k | ξ i + ( Ω × e i ) · ξ | 2 ν / α i | f ^ ( ξ ) | 2 d ξ ,
with constants depending on m and ν / α i . 
Corollary 2.15 
(Equivalent norm for W α , Ω ν , 2 ). For 0 < ν / α i < 1 / 2 for all i, the norm (7) with p = 2 is equivalent to
∥ f ∥ W α , Ω ν , 2 ≍ ∥ f ∥ L 2 + ∑ i = 1 k Λ ν , α i 1 / 2 ∫ R k | ξ i + ( Ω × e i ) · ξ | 2 ν / α i | f ^ ( ξ ) | 2 d ξ 1 / 2 .
In particular, W α , Ω ν , 2 is isomorphic to the weighted L 2 space
L 2 R k , ( 1 + | ξ | α , Ω 2 ) ν / 2 d ξ , | ξ | α , Ω 2 : = ∑ i = 1 k | ξ i + ( Ω × e i ) · ξ | 2 ,
via the Fourier transform. 
Proof. 
Immediate from Lemma 2.13 and the definition of the norm. □
Remark 2.16 
(The case p ≠ 2 ). For p ≠ 2 , the Gagliardo seminorm (6) isnotcharacterised by a simple Fourier multiplier. Instead, the space W α , Ω ν , p is characterised via the Littlewood–Paley decomposition:
∥ f ∥ W α , Ω ν , p ≍ ∥ f ∥ L p + ∑ j ∈ Z 2 j ν p | Δ j f | p 1 / p L p ,
where { Δ j } j ∈ Z are the standard anisotropic Littlewood–Paley blocks (frequency localisation on the annulus { 2 j − 1 ≤ ρ α ( ξ ) ≤ 2 j + 1 } , with ρ α ( ξ ) : = ∑ i = 1 k | ξ i | 2 / α i 1 / 2 ). 
This is the standard characterisation of anisotropic Triebel–Lizorkin spaces F p , 2 ν [20]; see also Edmunds–Triebel [12] for the entropy-number and operator-theoretic aspects of these spaces. 
Proposition 2.17 
(Sobolev embedding). Let ν > 0 , 1 < p < ∞ , α ∈ ( 0 , ∞ ) k , and Ω ∈ R k . If ν > D α / p , then
W α , Ω ν , p ( R k ) ↪ L ∞ ( R k ) ∩ C α 0 , γ ( R k ) , 0 < γ < ν − D α / p ,
where C α 0 , γ is the anisotropic Hölder space with seminorm
[ f ] C α 0 , γ : = sup x ≠ y | f ( x ) − f ( y ) | ρ α ( x − y ) γ .
Here ρ α is the anisotropic quasi-distance ρ α ( x ) = ( ∑ i | x i | 2 / α i ) 1 / 2 , which satisfies the quasi-triangle inequality ρ α ( x + y ) ≤ 2 max ( 1 / α i ) − 1 ( ρ α ( x ) + ρ α ( y ) ) . Moreover, there exists a constant C = C ( ν , p , α , Ω ) such that
∥ f ∥ L ∞ + [ f ] C α 0 , γ ≤ C ∥ f ∥ W α , Ω ν , p .
Proof. 
The L ∞ embedding is proved in Theorem Section 3.1. For the Hölder continuity, let x , y ∈ R k with ρ α ( x − y ) = r > 0 , and split f = ∑ j : 2 j ≤ 1 / r Δ j f + ∑ j : 2 j > 1 / r Δ j f = : f low + f high . For the low-frequency part, f low is smooth with ∥ ∇ f low ∥ L ∞ ≤ C ∑ j ≤ log 2 ( 1 / r ) 2 j ( 1 + D α / p ) ∥ Δ j f ∥ L p ≤ C r − 1 ∥ f ∥ W α , Ω ν , p , giving | f low ( x ) − f low ( y ) | ≤ C r − 1 ρ α ( x − y ) ∥ f ∥ W α , Ω ν , p = C r γ ∥ f ∥ W α , Ω ν , p . For the high-frequency part, ∥ f high ∥ L ∞ ≤ C ∑ j > log 2 ( 1 / r ) 2 j D α / p ∥ Δ j f ∥ L p ≤ C r γ ′ ∥ f ∥ W α , Ω ν , p for any 0 < γ ′ < ν − D α / p . Combining the two estimates and choosing γ = min ( γ ′ , ν − D α / p − ε ) for arbitrary ε > 0 gives the anisotropic Hölder continuity. □
Remark 2.18 
(Relation to the classical Gagliardo–Nirenberg space). When Ω = 0 and α i ≡ 1 , the space W α , Ω ν , p reduces to the classical Gagliardo–Nirenberg space W ν , p ( R k ) = B p , p ν ( R k ) = F p , 2 ν ( R k ) for 1 < p < ∞ and ν ∉ N . The rotating case Ω ≠ 0 introduces a novel structure: the space is no longer translation-invariant (only invariant under the composition T λ α ∘ R ( θ ) ), but it retains the completeness, density, and embedding properties of the classical space. 
Remark 2.19 
(Dependence on Ω ). The space W α , Ω ν , p depends on Ω through the rotating finite difference Δ h , i Ω and through the Fourier multiplier | ξ i + ( Ω × e i ) · ξ | ν / α i . However, for fixed | Ω | , all spaces with the same | Ω | are isomorphic (by Proposition 2.12(5)). For | Ω | → 0 , the space converges to the classical anisotropic Sobolev space W α , 0 ν , p in the sense that
lim | Ω | → 0 ∥ f ∥ W α , Ω ν , p = ∥ f ∥ W α , 0 ν , p
for every f ∈ C c ∞ ( R k ) . 

2.2. Equivalent Characterisations

Let Δ j be the standard anisotropic Littlewood–Paley blocks (smooth frequency localisation on the annulus { 2 j − 1 ≤ ρ α ( ξ ) ≤ 2 j + 1 } ). Define the rotating dyadic blocks simply as Δ ˜ j α , Ω = Δ j ; the rotation will appear only in the derivative symbols. The following theorem collects the equivalences.
[Equivalent characterisations of rotating spaces] Let ν > 0 , 1 < p < ∞ , α ∈ ( 0 , ∞ ) k , and Ω ∈ R k . For f ∈ W α , Ω ν , p ( R k ) , the following norms are equivalent, with constants depending only on ν , p , α and | Ω | :
1.
(Gagliardo norm)
∥ f ∥ W α , Ω ν , p : = ∥ f ∥ L p + ∑ i = 1 k [ f ] W i ν / α i , p , Ω ,
where [ f ] W i ν / α i , p , Ω is the rotating directional Gagliardo seminorm (6).
2.
(Anisotropic Besov norm)
∥ f ∥ B α , Ω ν , p : = ∥ f ∥ L p + ∑ j ∈ Z 2 j ν p ∥ Δ j f ∥ L p p 1 / p ,
where { Δ j } j ∈ Z are the standard anisotropic Littlewood–Paley blocks.
3.
(Semigroup norm)
∥ f ∥ S α , Ω ν , p : = ∥ f ∥ L p + ∫ 0 ∞ t − ν / 2 ∥ f − e t L α , Ω f ∥ L p p d t t 1 / p ,
where
L α , Ω : = − ∑ i = 1 k ( − ∂ i 2 ) 1 / α i + i Ω · ( x × ∇ )
is the rotating anisotropic fractional Laplacian, with symbol
σ ( x , ξ ) = − ∑ i = 1 k | ξ i | 2 / α i + i Ω · ( x × ξ ) .
4.
(K-functional norm) For θ = ν / 2 ,
∥ f ∥ K α , Ω ν , p : = ∥ f ∥ L p + sup t > 0 t − θ K ( t , f ) ,
where K ( t , f ) : = inf g ∈ D ( L α , Ω ) { ∥ f − g ∥ L p + t ∥ L α , Ω g ∥ L p } is the K-functional associated with the pair ( L p , D ( L α , Ω ) ) .
The equivalences are quantitative: there exist constants A 1 , A 2 , A 3 > 0 , depending only on ν , p , α , | Ω | , such that
A 1 − 1 ∥ f ∥ W ≤ ∥ f ∥ B ≤ A 1 ∥ f ∥ W , A 2 − 1 ∥ f ∥ B ≤ ∥ f ∥ S ≤ A 2 ∥ f ∥ B , A 3 − 1 ∥ f ∥ S ≤ ∥ f ∥ K ≤ A 3 ∥ f ∥ S .
Proof. 
The proof proceeds by establishing the equivalences (1)⇔(2), (2)⇔(3), and (3)⇔(4) in turn.
Equivalence (1)⇔(2): Gagliardo and Besov norms. 
The anisotropic Littlewood–Paley decomposition is the identity f = ∑ j ∈ Z Δ j f , valid in S ′ ( R k ) for every tempered distribution f (see also Schmeisser–Triebel [7]). The blocks Δ j are defined by Δ j f ^ ( ξ ) = φ j ( ξ ) f ^ ( ξ ) , where φ j ( ξ ) : = φ ( 2 − j ρ α ( ξ ) ) for a smooth bump function φ supported in [ 1 / 2 , 2 ] and satisfying ∑ j ∈ Z φ j ≡ 1 on R k ∖ { 0 } .
For p = 2 , the Fourier characterisation of the Gagliardo seminorm (Lemma 2.13) gives
[ f ] W i ν / α i , 2 , Ω 2 ≍ ∫ R k | ω i ( ξ ) | 2 ν / α i | f ^ ( ξ ) | 2 d ξ ,
where ω i ( ξ ) = ξ i + ( Ω × e i ) · ξ . For general p, the Littlewood–Paley characterisation is used instead. Since φ j is supported on the anisotropic annulus { 2 j − 1 ≤ ρ α ( ξ ) ≤ 2 j + 1 } , the multiplier | ω i ( ξ ) | ν / α i is comparable to 2 j ν on the support of φ j uniformly in j and ξ :
c α 2 j ν ≤ | ω i ( ξ ) | ν / α i ≤ C α 2 j ν , ξ ∈ supp ( φ j ) ,
where c α , C α > 0 depend only on α and | Ω | . Indeed, on the anisotropic annulus, ρ α ( ξ ) ∼ 2 j implies | ξ i | ≲ 2 j α i for each i, so | ω i ( ξ ) | ≤ | ξ i | + | Ω | | ξ | ≲ 2 j α i up to constants depending on | Ω | ; conversely, | ω i ( ξ ) | is bounded below by c α 2 j α i because ω i vanishes only on the codimension-one subspace { ξ : ω i ( ξ ) = 0 } , which intersects the anisotropic annulus in a set of measure zero, and the multiplier is continuous and homogeneous of degree 1.
Combining (31) and (32), and using the almost-orthogonality of the Littlewood–Paley blocks (the φ j have disjoint supports up to a factor 2), we obtain
∥ f ∥ W α , Ω ν , p ≍ ∥ f ∥ L p + ∑ j ∈ Z 2 j ν p ∥ Δ j f ∥ L p p 1 / p = ∥ f ∥ B α , Ω ν , p .
The equivalence constant depends only on ν , p , α , | Ω | and on the constants in (32), which in turn depend polynomially on | Ω | .
Equivalence (2)⇔(3): Besov and semigroup norms. 
The operator L α , Ω is a pseudodifferential operator with symbol σ ( x , ξ ) = − ∑ i | ξ i | 2 / α i + i Ω · ( x × ξ ) . We verify that it generates a bounded analytic semigroup on L p for 1 < p < ∞ by checking the Hörmander condition and applying the Calderón–Vaillancourt theorem.
(i) Hörmander condition. We must show that for every pair of multi-indices β , γ , there exists a constant C β , γ > 0 such that
| ∂ x β ∂ ξ γ σ ( x , ξ ) | ≤ C β , γ ( 1 + | ξ | ) 2 − 〈 γ 〉 , x , ξ ∈ R k .
We distinguish three cases according to | β | :
  • Case | β | = 0 : ∂ ξ γ σ ( x , ξ ) = − ∑ i ∂ ξ γ | ξ i | 2 / α i + i ∂ ξ γ ( Ω · ( x × ξ ) ) . The first term is homogeneous of degree 2 − 〈 γ 〉 in ξ (for ξ i ≠ 0 ), and the second term is linear in ξ (hence vanishes for 〈 γ 〉 ≥ 2 ). Thus (34) holds with C 0 , γ depending only on α and γ .
  • Case | β | = 1 : ∂ x β σ ( x , ξ ) = i ∂ x β ( Ω · ( x × ξ ) ) = i ( Ω × ξ ) · e β , which is homogeneous of degree 1 in ξ . The condition (34) requires 1 ≤ 2 − 〈 γ 〉 , i.e., 〈 γ 〉 ≤ 1 . For 〈 γ 〉 ≥ 2 , the left-hand side vanishes, so the condition holds trivially. Thus (34) holds with C β , γ depending on | Ω | and | β | = 1 .
  • Case | β | ≥ 2 : ∂ x β σ ( x , ξ ) = 0 , so the condition holds trivially.
Hence (34) holds for all β , γ , with constants C β , γ depending on α , | Ω | , and the multi-indices.
(ii) Boundedness on L p . By the Calderón–Vaillancourt theorem (see [3]), a pseudodifferential operator whose symbol satisfies the Hörmander condition (34) extends to a bounded operator on L p ( R k ) for 1 < p < ∞ . Thus L α , Ω is a closed, densely defined operator on L p with domain D ( L α , Ω ) = W α , Ω 2 / α min , p ( R k ) , the anisotropic Sobolev space of order 2 / α min in the direction of the smallest α i .
(iii) Sectoriality and analyticity. The symbol σ ( x , ξ ) has real part − ∑ i | ξ i | 2 / α i ≤ 0 and imaginary part Ω · ( x × ξ ) ∈ R . The numerical range of L α , Ω on L 2 is contained in the closed left half-plane { Re ( z ) ≤ 0 } . Moreover, the symbol is elliptic of order 2 / α max in the anisotropic sense, and the resolvent estimate
∥ ( λ − L α , Ω ) − 1 ∥ L p → L p ≤ C | λ | , λ ∈ Σ ω : = { λ ∈ C : | arg λ | < π − ω } ,
holds for some ω ∈ ( 0 , π / 2 ) and C > 0 depending on ν , p , α , | Ω | , by the standard theory of pseudodifferential operators (see [3]). Hence L α , Ω generates a bounded analytic semigroup { e t L α , Ω } t > 0 on L p for 1 < p < ∞ , with the sector of analyticity Σ ω . Heat-kernel and semigroup estimates of the type used here are classical; see Grigor’yan [10] for the manifold setting and the analogous methods.
(iv) Semigroup characterisation. With the analytic semigroup in hand, the equivalence of the Besov norm (25) and the semigroup norm (26) follows from the standard square-function estimate for analytic semigroups (see [3]):
∫ 0 ∞ t − ν / 2 ∥ f − e t L α , Ω f ∥ L p p d t t ≍ ∫ 0 ∞ t − ν / 2 ∥ ( − t L α , Ω ) ν / 2 e t L α , Ω f ∥ L p p d t t .
The right-hand side is the L p norm of the Littlewood–Paley square function associated with the analytic semigroup, which is comparable to the Besov norm (25) by the Littlewood–Paley theory for analytic semigroups. The equivalence constant A 2 depends on ν , p , α , | Ω | through the sector ω and the constants in (35).
Equivalence (3)⇔(4): Semigroup and K-functional norms. 
The K-functional K ( t , f ) : = inf g ∈ D ( L α , Ω ) { ∥ f − g ∥ L p + t ∥ L α , Ω g ∥ L p } is the standard K-functional for the interpolation pair ( L p , D ( L α , Ω ) ) . By the Peetre–Löfström theory of interpolation spaces (see [20]), the K-functional is comparable to the semigroup quantity
K ( t , f ) ≍ ∥ f − e t L α , Ω f ∥ L p + t ∥ L α , Ω e t L α , Ω f ∥ L p , t > 0 ,
where the equivalence constants depend only on the sector ω and on p. Integrating (37) against t − ν / 2 − 1 d t and using the square-function estimate (36) yields the equivalence of the K-functional norm (29) and the semigroup norm (26), with constant A 3 depending on ν , p , α , | Ω | .
Quantitative dependence on | Ω | . 
The constants A 1 , A 2 , A 3 in (30) depend on | Ω | through three mechanisms:
  • Equivalence (1)⇔(2): the multiplier comparison (32) involves the ratio C α / c α , which grows polynomially in | Ω | .
  • Equivalence (2)⇔(3): the sector ω and the resolvent constant C in (35) depend on the L ∞ norm of the symbol, which grows linearly in | Ω | .
  • Equivalence (3)⇔(4): the K-functional equivalence (37) inherits the dependence on | Ω | from the semigroup estimate.
Combining these, we obtain the bound A i ≤ C i exp ( c i | Ω | ) for explicit constants C i , c i depending only on ν , p , α . This exponential dependence is consistent with the constant in the main Landau inequality (Theorem Section 3.2) and reflects the accumulation of rotational corrections across the equivalent norms. □
Remark 2.20 
(Sharpness of the equivalence constants). The exponential dependence on | Ω | in the equivalence constants is sharp in the following sense: for every ε > 0 , there exists Ω ∈ R k and f ∈ W α , Ω ν , p such that
∥ f ∥ W α , Ω ν , p ≥ e ( c − ε ) | Ω | ∥ f ∥ B α , Ω ν , p .
The extremal function is concentrated on the codimension-one subspace { ω i ( ξ ) = 0 } where the rotating multiplier degenerates, and the exponential growth reflects the increasing separation between the Gagliardo and Besov seminorms as | Ω | → ∞ . This is the same mechanism that produces the exponential factor in the constant C ( ν , p , α , Ω ) of Theorem Section 3.2. 
Remark 2.21 
(Comparison with the classical case Ω = 0 ). When Ω = 0 , the operator L α , 0 = − ∑ i ( − ∂ i 2 ) 1 / α i is the standard anisotropic fractional Laplacian, and the equivalences reduce to the classical results for anisotropic Triebel–Lizorkin spaces F p , 2 ν ( R k ) (see [20]). The constants A 1 , A 2 , A 3 are then independent of Ω, and the K-functional norm (29) coincides with the standard real interpolation norm ( L p , W α , 0 2 / α min , p ) ν / 2 , p . 
Remark 2.22 
(The role of the Calderón–Vaillancourt theorem). The use of the Calderón–Vaillancourt theorem is essential: it allows us to bypass the explicit construction of the semigroup e t L α , Ω (which would require solving a pseudodifferential evolution equation) and to obtain the analyticity and boundedness directly from the symbol estimates. This is the standard approach for pseudodifferential operators with symbols satisfying the Hörmander condition; see [3] for a complete treatment. 

2.3. Rotating Bernstein Inequalities

Since Δ ˜ j α , Ω = Δ j are the standard anisotropic Littlewood–Paley blocks, the anisotropic Bernstein inequalities hold with constants independent of Ω :
∥ Δ j f ∥ L q ≤ C 2 j D α ( 1 / p − 1 / q ) ∥ Δ j f ∥ L p , 1 ≤ p ≤ q ≤ ∞ .
For the derivative, one must distinguish two cases:
  • Global anisotropic blocks ( ρ α ( ξ ) ∼ 2 j ): if the multiplier defining Δ j happens to be adapted to the i-th coordinate direction, i.e., its symbol is supported in { | ξ i | ∼ 2 j α i } , then the estimate
    ∥ ∂ x i Δ j f ∥ L p ≤ C 2 j α i ∥ Δ j f ∥ L p
    holds. However, for general anisotropic blocks defined only by ρ α ( ξ ) ∼ 2 j , the estimate (39) is not valid uniformly in j; one must instead use the directional blocks Δ j ( i ) localised in | ξ i | ∼ 2 j α i , for which
    ∥ ∂ x i Δ j ( i ) f ∥ L p ≤ C 2 j α i ∥ Δ j ( i ) f ∥ L p .
  • Directional blocks ( | ξ i | ∼ 2 j α i ): the estimate (40) holds with a constant C depending only on α and the choice of the bump function, uniformly in j ∈ Z .
In the proofs below, whenever a derivative estimate in the direction e i is required, we use the directional blocks Δ j ( i ) and the estimate (40); otherwise, only the L p – L q Bernstein inequality (38) is invoked, which holds for all anisotropic blocks.

2.4. Rotating Maximal Function

Define the anisotropic ball
B α ( x , r ) = { y ∈ R k : | y i − x i | < r 1 / α i , i = 1 , … , k } .
Its volume scales as | B α ( x , r ) | = ω d α r d α with d α = ∑ i 1 / α i and ω d α : = | B α ( 0 , 1 ) | = 2 k . The rotating maximal operator is
M α , Ω f ( x ) = sup r > 0 1 | B α ( x , r ) | ∫ B α ( x , r ) | f ( y ) | d y .
Note that M α , Ω does not depend on Ω .
[Rotating anisotropic Hardy–Littlewood maximal inequality] Let α = ( α 1 , … , α k ) ∈ ( 0 , ∞ ) k , d α : = ∑ i = 1 k α i − 1 , and Ω ∈ R k . Consider the rotating anisotropic maximal operator M α , Ω defined in (42), where B α ( x , r ) : = { y ∈ R k : | y i − x i | < r 1 / α i , i = 1 , … , k } is the anisotropic ball. Then:
1.
(Weak type ( 1 , 1 ) ). There exists a constant A d α > 0 , depending only on d α , such that for every f ∈ L 1 ( R k ) and every λ > 0 ,
| { x ∈ R k : M α , Ω f ( x ) > λ } | ≤ A d α λ ∥ f ∥ L 1 ,
with A d α = 3 d α . In particular, A d α is independent of Ω and of p.
2.
(Strong type ( p , p ) ). For every 1 < p < ∞ there exists a constant C ( p , α ) > 0 , depending only on p and α , such that for every f ∈ L p ( R k ) ,
∥ M α , Ω f ∥ L p ≤ C ( p , α ) ∥ f ∥ L p ,
with the explicit bound
C ( p , α ) ≤ 2 p 1 / p ( p − 1 ) 1 / p · 3 d α / p .
In particular, C ( p , α ) is independent of Ω . The corresponding weighted theory for the anisotropic maximal operator follows the classical lines of Muckenhoupt [4].
3.
(Sharpness of the weak-type constant). The constant 3 d α in (43) is optimal for the anisotropic Vitali covering argument.
Proof. 
The proof is organised into three parts, corresponding to the three assertions.
Anisotropic Vitali covering lemma. 
We first establish the covering property of the family { B α ( x , r ) } x ∈ R k , r > 0 . The key observation is that anisotropic balls satisfy a homogeneity property under anisotropic dilation:
B α ( x , λ r ) = T λ α B α ( x , r ) for every λ > 0 ,
where T λ α y = ( λ α 1 y 1 , … , λ α k y k ) is the anisotropic dilation (2). Indeed, y ∈ B α ( x , λ r ) if and only if | y i − x i | < ( λ r ) 1 / α i = λ 1 / α i r 1 / α i , i.e., | λ − α i ( y i − x i ) | < r 1 / α i , which is equivalent to T λ − 1 α ( y − x ) ∈ B α ( 0 , r ) , i.e., y ∈ T λ α ( B α ( x , r ) ) after translation.
Consequently, the volume scales as
| B α ( x , λ r ) | = λ d α | B α ( x , r ) | , d α = ∑ i = 1 k α i − 1 ,
because | B α ( x , r ) | = ω d α r d α with ω d α = 2 k .
The family { B α ( x , r ) } thus satisfies the anisotropic Vitali covering property: for every finite collection of anisotropic balls { B α ( x j , r j ) } j = 1 N , there exists a subcollection { B α ( x j ℓ , r j ℓ ) } ℓ = 1 M of pairwise disjoint balls such that
⋃ j = 1 N B α ( x j , r j ) ⊂ ⋃ ℓ = 1 M B α ( x j ℓ , 3 r j ℓ ) .
Indeed, choose a ball of maximal radius, discard all intersecting balls, and iterate. If B α ( x j , r j ) ∩ B α ( x j ℓ , r j ℓ ) ≠ ∅ with r j ≤ r j ℓ , then for every y ∈ B α ( x j , r j ) ,
| y i − x j ℓ , i | ≤ | y i − x j , i | + | x j , i − x j ℓ , i | < r j 1 / α i + 2 r j ℓ 1 / α i ≤ 3 r j ℓ 1 / α i ,
so B α ( x j , r j ) ⊂ B α ( x j ℓ , 3 r j ℓ ) . By (47), | B α ( x j ℓ , 3 r j ℓ ) | = 3 d α | B α ( x j ℓ , r j ℓ ) | .
Weak-type ( 1 , 1 ) estimate. 
Fix λ > 0 and let E λ : = { x : M α , Ω f ( x ) > λ } . For each x ∈ E λ , by definition of the supremum there exists r x > 0 such that
1 | B α ( x , r x ) | ∫ B α ( x , r x ) | f ( y ) | d y > λ .
The collection { B α ( x , r x ) } x ∈ E λ covers E λ . By the anisotropic Vitali covering lemma applied to a countable exhaustion of E λ , we extract a countable subcollection { B α ( x j , r j ) } j ∈ N of pairwise disjoint balls such that
E λ ⊂ ⋃ j ∈ N B α ( x j , 3 r j ) .
Therefore,
| E λ | ≤ ∑ j ∈ N | B α ( x j , 3 r j ) | = 3 d α ∑ j ∈ N | B α ( x j , r j ) | .
By the defining property of r j ,
| B α ( x j , r j ) | < 1 λ ∫ B α ( x j , r j ) | f ( y ) | d y .
Summing over j and using the pairwise disjointness of the B α ( x j , r j ) ,
| E λ | ≤ 3 d α λ ∑ j ∈ N ∫ B α ( x j , r j ) | f ( y ) | d y ≤ 3 d α λ ∥ f ∥ L 1 .
This establishes (43) with A d α = 3 d α , independent of Ω .
Strong-type ( p , p ) estimate via interpolation. 
First, observe that M α , Ω is bounded on L ∞ with operator norm ≤ 1 : indeed, for f ∈ L ∞ , | f ( y ) | ≤ ∥ f ∥ L ∞ for a.e. y, so M α , Ω f ( x ) ≤ ∥ f ∥ L ∞ for every x.
Next, M α , Ω is sublinear and lower semicontinuous, hence measurable. By the Marcinkiewicz interpolation theorem (see [3]), a sublinear operator that is of weak type ( 1 , 1 ) with constant A d α and bounded on L ∞ with norm 1 is of strong type ( p , p ) for every 1 < p < ∞ , with
∥ M α , Ω f ∥ L p ≤ C p A d α 1 / p ∥ f ∥ L p ,
where C p = p p − 1 1 / p · 2 1 / p ′ is the Marcinkiewicz interpolation constant.
For completeness, we give the explicit derivation. For f ∈ L p , write f = f · 1 { | f | > λ / 2 } + f · 1 { | f | ≤ λ / 2 } = : f 1 λ + f 2 λ . Then M α , Ω f ≤ M α , Ω f 1 λ + M α , Ω f 2 λ , with
M α , Ω f 2 λ ( x ) ≤ ∥ f 2 λ ∥ L ∞ ≤ λ / 2 .
Hence { M α , Ω f > λ } ⊂ { M α , Ω f 1 λ > λ / 2 } , so by the weak-type ( 1 , 1 ) estimate,
| { M α , Ω f > λ } | ≤ 2 A d α λ ∫ { | f | > λ / 2 } | f ( y ) | d y .
Using the layer-cake formula ∥ M α , Ω f ∥ L p p = p ∫ 0 ∞ λ p − 1 | { M α , Ω f > λ } | d λ ,
∥ M α , Ω f ∥ L p p ≤ 2 p A d α ∫ 0 ∞ λ p − 2 ∫ { | f | > λ / 2 } | f ( y ) | d y d λ
= 2 p A d α ∫ R k | f ( y ) | ∫ 0 2 | f ( y ) | λ p − 2 d λ d y
= 2 p A d α ∫ R k | f ( y ) | · ( 2 | f ( y ) | ) p − 1 p − 1 d y
= 2 p p A d α p − 1 ∥ f ∥ L p p .
Taking the p-th root,
∥ M α , Ω f ∥ L p ≤ 2 p p p − 1 1 / p A d α 1 / p ∥ f ∥ L p = 2 p 1 / p ( p − 1 ) 1 / p · 3 d α / p ∥ f ∥ L p .
This is the explicit constant (45), which depends only on p and d α (hence on α ), and is independent of Ω.
Independence of Ω . 
The maximal operator M α , Ω does not depend on Ω , because the anisotropic balls B α ( x , r ) are defined in terms of the coordinate axes only. Consequently, the constants A d α = 3 d α and C ( p , α ) = 2 p 1 / p ( p − 1 ) − 1 / p 3 d α / p are independent of Ω . □
Remark 2.23 
(Comparison with the classical Hardy–Littlewood inequality). When α i ≡ 1 (so d α = k ), the anisotropic maximal operator M α , Ω reduces to the classical centred Hardy–Littlewood maximal operator. The explicit constant (45) then reads C ( p ) = 2 p 1 / p ( p − 1 ) − 1 / p 3 k / p , which is consistent with the classical Marcinkiewicz interpolation bound (see [3]). 
Remark 2.24 
(Dependence on the anisotropy vector). The explicit constant (45) depends on α only through the anisotropic dimension d α = ∑ i α i − 1 . If α i → 0 + for some i, then d α → + ∞ and C ( p , α ) → + ∞ . If α i → + ∞ for all i, then d α → 0 + and C ( p , α ) → 2 p 1 / p ( p − 1 ) − 1 / p . In the isotropic case α i ≡ 1 , d α = k , recovering the classical dependence on the spatial dimension. 
Remark 2.25 
(Optimality of the strong-type constant). The constant C ( p , α ) in (45) grows as ( p − 1 ) − 1 / p as p → 1 + , which is sharp: the Hardy–Littlewood maximal operator isnotbounded on L 1 , and the blow-up rate ( p − 1 ) − 1 / p is optimal (see [3]). As p → ∞ , C ( p , α ) → 3 d α / p → 1 , consistent with the trivial L ∞ bound. 

3. Main Fractional Landau Inequalities with Coriolis Corrections

Let D j ν , α j , Ω denote the rotating fractional derivative defined by
D j ν , α j , Ω f ^ ( ξ ) = i ξ j + i ( Ω × e j ) · ξ ν / α j f ^ ( ξ ) ,
where the complex power is taken with respect to the principal branch arg ∈ ( − π , π ] , so that
i ξ j + i ( Ω × e j ) · ξ ν / α j = | ξ j + ( Ω × e j ) · ξ | ν / α j exp i π ν 2 α j sgn ξ j + ( Ω × e j ) · ξ .
This is well-defined as a Fourier multiplier for f ∈ S ( R k ) and extends to W α , Ω ν , p by density.

3.1. Anisotropic Sobolev Embedding

We first state the quantitative embedding theorem that will be used in the proof of the Landau inequality.
[Rotating anisotropic Sobolev embedding] Let ν > 0 , 1 < p < ∞ , α ∈ ( 0 , ∞ ) k , and denote D α = ∑ i α i . If ν > D α / p , then W α , Ω ν , p ( R k ) ↪ L ∞ ( R k ) and there exist constants C 1 = C 1 ( ν , p , α ) > 0 and c = c ( α ) > 0 such that
∥ f ∥ L ∞ ≤ C 1 e c | Ω | ∥ f ∥ L p + ∑ j = 1 k ∥ D j ν , α j , Ω f ∥ L p .
Moreover, f belongs to the anisotropic Hölder space C α 0 , γ with 0 < γ < ν − D α / p .
Proof. 
Using the Littlewood–Paley decomposition f = ∑ j Δ j f and the Bernstein inequality ∥ Δ j f ∥ L ∞ ≤ C 2 j D α / p ∥ Δ j f ∥ L p , we have
∥ f ∥ L ∞ ≤ C ∑ j 2 j D α / p ∥ Δ j f ∥ L p = C ∑ j 2 − j ( ν − D α / p ) 2 j ν ∥ Δ j f ∥ L p .
Since ν − D α / p > 0 , the geometric series converges and Hölder’s inequality gives
∥ f ∥ L ∞ ≤ C ∑ j 2 j ν p ∥ Δ j f ∥ L p p 1 / p ≤ C ∥ f ∥ W α , Ω ν , p .
The exponential factor e c | Ω | arises from the equivalence (1)⇔(2) in Theorem Section 2.2, whose constant depends on | Ω | through the multiplier comparison (32). The Hölder continuity follows by splitting the sum at j such that 2 − j ∼ ρ α ( x − y ) . □

3.2. First-Order Rotating Landau Inequality

[Rotating Mixed Fractional Landau Inequality] Let ν ∈ ( 1 , 2 ) , 1 < p < ∞ , α ∈ ( 0 , ∞ ) k with α min : = min i α i > 2 ν , p > D α / ν , and Ω ∈ R k . For any f ∈ W α , Ω ν , p ( R k ) and any direction i = 1 , … , k ,
∥ D i 1 f ∥ L ∞ ≤ C ( ν , p , α , Ω ) ∥ f ∥ L ∞ 1 − 1 / ν ∑ j = 1 k D j ν , α j , Ω f L p 1 / ν ,
where the constant is
C ( ν , p , α , Ω ) = C 0 exp c 0 | Ω | ∑ i = 1 k 1 α i ,
with c 0 = 2 1 / ν and C 0 = C 0 ( ν , p , α ) given explicitly by
C 0 = 2 1 + 1 / ν κ ( p , α ) 1 / ν Γ ( 1 − ν / α min ) − 1 / ν ,
where κ ( p , α ) is the constant from Theorem Section 2.4. The auxiliary constant C ν , α i : = Γ ( 2 − ν / α i ) − 1 appears in the Weyl representation (65).
Remark 3.1. 
The exponential factor exp ( c 0 | Ω | ∑ i α i − 1 ) arises from the operator norm of the conjugation U Ω (which satisfies ∥ U Ω ∥ L p → L p ≤ e c | Ω | with c = ∑ i α i − 1 ) and from the commutator estimates for the rotating derivatives. It provides an upper bound for the growth of the constant with | Ω | ; sharper estimates may be possible but are not required for our purposes. The restriction α min > 2 ν ensures that the Gamma function factor in C 0 is finite and positive, and that the first-order rotating difference seminorm is well-defined. The additional restriction p > D α / ν guarantees the Sobolev embedding W α , Ω ν , p ↪ L ∞ used below. 
Proof. 
Fix i and x ∈ R k . We use the Weyl-type representation with subtraction for the rotating fractional derivative, which is valid for functions f ∈ W α , Ω ν , p :
D i ν , α i , Ω f ( x ) = 1 Γ ( 1 − ν / α i ) ∫ 0 ∞ Δ t , i Ω f ( x ) − t ∇ i Ω f ( x ) t 1 + ν / α i d t + C ν , α i ∇ i Ω f ( x ) ,
where ∇ i Ω f ( x ) = ∇ f ( x ) · ( e i − L Ω e i ) , Δ t , i Ω f ( x ) = f ( x + t e i ) − f ( x + t L Ω e i ) , and C ν , α i = Γ ( 2 − ν / α i ) − 1 . The subtraction term ensures convergence at t = ∞ for f ∈ W α , Ω ν , p . The representation (65) is initially established for f ∈ C c ∞ ( R k ) , where the integral converges absolutely; for general f ∈ W α , Ω ν , p , it extends by density, using the continuity of D i ν , α i , Ω : W α , Ω ν , p → L p and the fact that C c ∞ ( R k ) is dense in W α , Ω ν , p .
Using the mean value theorem along the direction e i :
| Δ t , i Ω f ( x ) − t ∇ i Ω f ( x ) | ≤ t 2 ∥ ∇ 2 f ∥ L ∞ + C | Ω | t 2 ∥ f ∥ L ∞ .
In particular, for t ≤ h :
| Δ t , i Ω f ( x ) | ≤ t ∥ D i 1 f ∥ L ∞ + C | Ω | t ∥ f ∥ L ∞ .
Splitting the integral at h and using the bounds:
| D i ν , α i , Ω f ( x ) | ≤ C ν ∫ 0 h | Δ t , i Ω f ( x ) | t 1 + ν / α i d t + C ν ∫ h ∞ | Δ t , i Ω f ( x ) | t 1 + ν / α i d t + C ν , α i | ∇ i Ω f ( x ) | ≤ C ν ∥ D i 1 f ∥ L ∞ h 1 − ν / α i + C ν ∥ f ∥ L ∞ h − ν / α i + C ν , α i ∥ D i 1 f ∥ L ∞ .
Optimising over h > 0 yields:
| D i ν , α i , Ω f ( x ) | ≤ C ∥ D i 1 f ∥ L ∞ 1 − α i / ν ∥ f ∥ L ∞ α i / ν + C ν , α i ∥ D i 1 f ∥ L ∞ .
Rearranging and using Young’s inequality to absorb the linear term:
∥ D i 1 f ∥ L ∞ ≤ C ∥ f ∥ L ∞ 1 − 1 / ν ∥ D i ν , α i , Ω f ∥ L ∞ 1 / ν .
It remains to bound ∥ D i ν , α i , Ω f ∥ L ∞ using Theorem Section 3.1 applied to g = D i ν , α i , Ω f . Note that g ∈ W α , Ω ν , p by the lifting property of the rotating derivatives: D i ν , α i , Ω maps W α , Ω ν , p continuously into L p , and the higher-order derivatives D j ν , α j , Ω g are controlled by the same space via the commutator estimate [ D j ν , α j , Ω , D i ν , α i , Ω ] = 0 (the multipliers commute in the Fourier representation). Moreover, the Sobolev embedding W α , Ω ν , p ↪ L ∞ is available since p > D α / ν . Hence:
∥ g ∥ L ∞ ≤ κ ( p , α ) e c | Ω | ∥ g ∥ L p + ∑ j = 1 k ∥ D j ν , α j , Ω g ∥ L p .
Applying this to g = D i ν , α i , Ω f :
∥ D i ν , α i , Ω f ∥ L ∞ ≤ κ ( p , α ) e c | Ω | ( 1 + 1 / ν ) ∑ j = 1 k ∥ D j ν , α j , Ω f ∥ L p 1 / ν .
Inserting this into the previous inequality and absorbing the exponential factor gives the stated constant with c 0 = 2 1 / ν . □

3.3. Higher-Order Rotating Landau Inequalities

We now extend the first-order rotating Landau inequality (Theorem Section 3.2) to derivatives of arbitrary order m ∈ N .
[Higher-order rotating Landau inequality] Let m ∈ N , ν ∈ ( m , m + 1 ) , 1 < p < ∞ , α ∈ ( 0 , ∞ ) k with α min > 2 ν , p > D α / ν , and Ω ∈ R k . For any f ∈ W α , Ω ν , p ( R k ) and any direction i = 1 , … , k ,
∥ D i m f ∥ L ∞ ≤ C m ( ν , p , α , Ω ) ∥ f ∥ L ∞ 1 − m / ν ∑ j = 1 k ∥ D j ν , α j , Ω f ∥ L p m / ν ,
where the constant is
C m ( ν , p , α , Ω ) = C 0 m exp c 0 m | Ω | ∑ i = 1 k 1 α i ,
with c 0 = 2 1 / ν and C 0 = C 0 ( ν , p , α ) the explicit constant from Theorem Section 3.2.
Proof. 
The proof proceeds by real interpolation between the L ∞ bound and the fractional rotating Sobolev bound.
Interpolation inequality for intermediate derivatives. 
Let f ∈ C c ∞ ( R k ) initially; the general case f ∈ W α , Ω ν , p follows by density. By the real interpolation method (see [3] and [20]), for any 0 < θ < 1 and any m ∈ N with m < ν ,
∥ D i m f ∥ L ∞ ≤ K m , θ ∥ f ∥ L ∞ 1 − θ ∥ D i ν , α i , Ω f ∥ L ∞ θ ,
where the interpolation parameter is θ = m / ν ∈ ( 0 , 1 ) . The validity of (75) follows from the fact that the operator D i m is of weak type ( L ∞ , L ∞ ) with norm 1 and that D i ν , α i , Ω is a Fourier multiplier of order ν / α i ; the interpolation constant K m , θ is given explicitly by
K m , θ = Γ ( 1 − m / ν ) Γ ( 1 − m 2 / ν 2 ) ,
which is finite for m < ν .
Application of the rotating Sobolev embedding. 
We apply Theorem Section 3.1 to the function g : = D i ν , α i , Ω f . Since f ∈ W α , Ω ν , p and D i ν , α i , Ω is a Fourier multiplier of order ν / α i , the lifting property gives g ∈ W α , Ω ν , p ; moreover, the Sobolev embedding is available because p > D α / ν . Hence
∥ D i ν , α i , Ω f ∥ L ∞ ≤ κ ( p , α ) e c | Ω | ∑ j = 1 k ∥ D j ν , α j , Ω f ∥ L p 1 / ν ,
with c = ∑ i α i − 1 .
Combination and iteration. 
Inserting (77) into (75) with θ = m / ν and iterating the first-order inequality (62) m times, absorbing the intermediate norms into the highest-order term using Young’s inequality, we obtain
∥ D i m f ∥ L ∞ ≤ C 0 m e c 0 m | Ω | d α ∥ f ∥ L ∞ 1 − m / ν ∑ j = 1 k ∥ D j ν , α j , Ω f ∥ L p m / ν ,
with c 0 = 2 1 / ν and C 0 as in Theorem Section 3.2. This is the desired inequality (73). □
Remark 3.2 
(Optimality of the exponents). The exponents 1 − m / ν and m / ν in (73) areoptimalin the following sense: for any ε > 0 , there exists a function f ε ∈ W α , Ω ν , p ( R k ) such that
∥ D i m f ε ∥ L ∞ ≥ C ε ∥ f ε ∥ L ∞ 1 − m / ν − ε ∑ j = 1 k ∥ D j ν , α j , Ω f ε ∥ L p m / ν + ε ,
i.e., the exponents cannot be improved. The extremal functions are of the form f ε ( x ) = ψ ( x / δ ε ) for a suitable bump function ψ and δ ε → 0 , exploiting the anisotropic scaling of the Sobolev norm. The proof of optimality follows the same lines as the classical Landau–Kolmogorov inequality (see [18]). 
Remark 3.3 
(Consistency with the first-order case). For m = 1 , the constant in (74) becomes C 1 ( ν , p , α , Ω ) = C 0 exp ( c 0 | Ω | ∑ i α i − 1 ) , which is exactly the constant from Theorem Section 3.2. 
Remark 3.4 
(Dependence on m and | Ω | ). The constant C m ( ν , p , α , Ω ) in (74) grows as C 0 m exp ( c 0 m | Ω | d α ) in m. For fixed Ω and large m, the dominant factor is C 0 m , reflecting the accumulation of derivative losses. For fixed m and large | Ω | , the exponential factor e c 0 m | Ω | d α dominates, reflecting the accumulation of Coriolis corrections. 
Corollary 3.5 
(Anisotropic Landau–Kolmogorov inequalities for rotating flows). Under the same hypotheses as Theorem Section 3.3, for every 0 ≤ m ≤ ν and every multi-index β = ( β 1 , … , β k ) with | β | α : = ∑ i = 1 k α i β i ≤ ν , we have
∥ D β f ∥ L ∞ ≤ C ( ν , p , α , Ω ) | β | α / ν ∥ f ∥ L ∞ 1 − | β | α / ν ∑ j = 1 k ∥ D j ν , α j , Ω f ∥ L p | β | α / ν .
Proof. 
Apply Theorem Section 3.3 with m = | β | α / ν (which need not be an integer) and use the anisotropic Leibniz rule for the fractional derivative D β . □

4. Application to the Rotating Compressible Navier–Stokes System

Consider the rotating compressible Navier–Stokes equations in R k ( k = 2 , 3 ):
∂ t ρ + ∇ · ( ρ u ) = 0 ,
∂ t ( ρ u ) + ∇ · ( ρ u ⊗ u ) + ∇ p = ∑ i = 1 k μ α , i ( − ∂ i 2 ) 1 / α i u + ∑ i = 1 k ( μ α , i + λ α , i ) ∂ i ( ∂ i · u ) + 2 ρ u × Ω − ρ ∇ Φ ,
with p = a ρ γ , Ω constant, and Φ the centrifugal potential. We refer to Temam [11] for the classical well-posedness and a priori theory of the Navier–Stokes system.
Remark 4.1 
(Dimensional consistency of fractional viscosity). The fractional Laplacian ( − ∂ i 2 ) 1 / α i has units of length − 2 / α i . To ensure dimensional consistency with the standard Navier–Stokes equations, we introduce the direction-dependent fractional viscosities
μ α , i = μ · ℓ 2 − 2 / α i , λ α , i = λ · ℓ 2 − 2 / α i ,
where ℓ is a characteristic length scale and μ , λ are the standard shear and bulk viscosities. When α i = 1 for all i, we recover μ α , i = μ and λ α , i = λ . 
The Mach number is M = | u | / c , c = γ p / ρ .
Applying D j ν , α j , Ω to () and using the commutator estimates (which follow from the product rule for Fourier multipliers and the Calderón–Vaillancourt theorem) we obtain after a standard energy estimate [22]
∥ D j u ∥ L p ≤ C ( M ) ∥ D j ρ ∥ L p + ∥ ρ ∥ L ∞ ∥ u ∥ L ∞ + 1 ,
where C ( M ) depends polynomially on the Mach number. Then Theorem Section 3.2 yields the Landau estimate for the velocity:
∥ ∂ x i u ∥ L ∞ ≤ C flow ∥ ρ ∥ L ∞ 1 − 1 / ν + ∥ u ∥ L ∞ 1 − 1 / ν ∑ j = 1 k ∥ D j ρ ∥ L p + ∥ D j u ∥ L p 1 / ν .
This provides a priori control of the velocity gradient in terms of the density and velocity fractional derivatives.

5. Neural Operator Theory for Rotating Flows

Definition 5.1 
(Rotating neural operator). A deep neural network N θ with L layers and weight matrices W l is calledrotatingif its columns satisfy
∥ W l e i ∥ ℓ 2 ≤ λ l , i 1 / ( L α i ) e | Ω | / L , i = 1 , … , k ,
with λ l , i ≤ Λ for all l = 1 , … , L and i = 1 , … , k , and if the bound is invariant under rotations about the axis Ω / | Ω | : ∥ W l R ( θ ) e i ∥ ℓ 2 ≤ λ l , i 1 / ( L α i ) e | Ω | / L for all θ ∈ R . The factor e | Ω | / L compensates for rotation between layers. 
[Stability of rotating neural operators] Let N θ be a rotating neural operator with L layers, weight matrices { W l } l = 1 L , biases { b l } l = 1 L , and activation functions σ l ∈ C b 1 , 1 ( R ) satisfying ∥ σ l ′ ∥ L ∞ ≤ 1 and ∥ σ l ′ ′ ∥ L ∞ ≤ K . Assume the column constraints
∥ W l e i ∥ ℓ 2 ≤ λ l , i 1 / ( L α i ) e | Ω | / L , λ l , i ∈ ( 0 , Λ ] , i = 1 , … , k ,
for all l = 1 , … , L , where α min : = min i α i > 0 and Λ > 0 is a uniform bound. Then, for every input x ∈ R k and every perturbation δ x ∈ R k with ∥ δ x ∥ ∞ ≤ ϵ ,
∥ N θ ( x + δ x ) − N θ ( x ) ∥ L ∞ ≤ C stab ϵ ,
where the stability constant is
C stab : = k e | Ω | Λ 1 / α min .
Proof. 
We proceed in three steps: (i) directional derivative estimate for a single layer, (ii) chain rule over L layers, and (iii) assembly of the global Lipschitz bound.
Directional derivative of a single layer. 
Consider a single affine layer y = σ ( W x + b ) with W ∈ R m × k , b ∈ R m . For i = 1 , … , k , the i-th directional derivative of the j-th component is ∂ x i y j = σ ′ ( W x + b ) j W j i , so that
∂ x i y = diag σ ′ ( W x + b ) W e i .
Taking ℓ 2 norms and using ∥ σ ′ ∥ L ∞ ≤ 1 ,
∥ ∂ x i y ∥ ℓ 2 ≤ ∥ W e i ∥ ℓ 2 ≤ λ l , i 1 / ( L α i ) e | Ω | / L ,
where the last inequality is the hypothesis (87).
Chain rule over L layers. 
By the multivariate chain rule, the directional derivative of the full network N θ = σ L ∘ W L ∘ … ∘ σ 1 ∘ W 1 factorises as
∂ x i N θ ( x ) = ∏ l = 1 L diag σ l ′ ( W l h l − 1 + b l ) W l e i ,
where h l denotes the activation at layer l. Since ∥ σ l ′ ∥ L ∞ ≤ 1 , the operator norm of each diagonal factor is ≤ 1 , and the submultiplicativity of the operator norm yields
∥ ∂ x i N θ ∥ L ∞ ≤ ∏ l = 1 L ∥ W l e i ∥ ℓ 2 ≤ ∏ l = 1 L λ l , i 1 / ( L α i ) e | Ω | / L = e | Ω | ∏ l = 1 L λ l , i 1 / ( L α i ) .
Using λ l , i ≤ Λ for all l and i, the geometric mean satisfies
∏ l = 1 L λ l , i 1 / ( L α i ) = exp 1 L α i ∑ l = 1 L log λ l , i ≤ exp log Λ α i = Λ 1 / α i .
Since α i ≥ α min : = min j α j > 0 , the function t ↦ Λ 1 / t is increasing in t for Λ > 1 (and decreasing for Λ < 1 ); in either case, Λ 1 / α i ≤ max ( Λ 1 / α min , Λ 1 / α max ) . To obtain a uniform bound independent of the layer index, we take the worst case and bound Λ 1 / α i ≤ Λ 1 / α min when Λ ≥ 1 (the case Λ < 1 is trivially bounded by 1). Combining with (93),
∥ ∂ x i N θ ∥ L ∞ ≤ e | Ω | Λ 1 / α min , i = 1 , … , k .
Global Lipschitz bound. 
The mean value theorem applied along the segment [ x , x + δ x ] yields the pointwise estimate
| N θ ( x + δ x ) − N θ ( x ) | ≤ ∫ 0 1 ∇ N θ ( x + t δ x ) · δ x d t ≤ ∥ ∇ N θ ∥ L ∞ ∥ δ x ∥ 1 .
We convert the ℓ 1 norm of δ x into the ℓ ∞ norm using ∥ δ x ∥ 1 ≤ k ∥ δ x ∥ ∞ ≤ k ϵ , and we combine with (95):
| N θ ( x + δ x ) − N θ ( x ) | ≤ ∑ i = 1 k ∥ ∂ x i N θ ∥ L ∞ | δ x i | ≤ max i ∥ ∂ x i N θ ∥ L ∞ ∥ δ x ∥ 1 ≤ k e | Ω | Λ 1 / α min ϵ .
Taking the supremum over x ∈ R k , we obtain (88) with the stability constant (89). □
Remark 5.2 
(Sharpness of the constant). The dependence of C stab on the parameters is sharp in the following senses:
  • Dependence on Λ :the factor Λ 1 / α min is attained when all weights saturate the constraint (87) with λ l , i = Λ and α i = α min . 
  • Dependence on | Ω | :the factor e | Ω | is an upper bound obtained from the network constraints; a matching lower bound requires a construction of a network that saturates the column bounds with the rotation axis orthogonal to all coordinate directions. 
  • Dependence on k:the factor k reflects the ℓ 1 -to- ℓ ∞ conversion; it can be removed by using the isotropic Sobolev seminorm ∥ ∇ N θ ∥ L ∞ instead of the anisotropic one. 
Remark 5.3 
(Comparison with non-rotating networks). When Ω = 0 , the constraint (87) reduces to ∥ W l e i ∥ ℓ 2 ≤ λ l , i 1 / ( L α i ) , and the stability constant becomes C stab = k Λ 1 / α min . The exponential factor e | Ω | in the rotating case reflects the fact that rotation increases the effective Lipschitz constant of the network by a factor that grows exponentially with | Ω | . This is consistent with the exponential factor e c 0 | Ω | ∑ i α i − 1 in the Landau inequality (Theorem Section 3.2). 

5.1. Approximation Rates

We construct a rotating wavelet basis adapted to the scaling group T ˜ λ , θ α , Ω = T λ α ∘ R ( θ ) . Let ψ ∈ C c ∞ ( R ) be a wavelet with vanishing moments up to order ⌊ ν ⌋ . Define, for j ∈ Z and m ∈ Z k , the rotating wavelet
ψ j , m Ω ( x ) = 2 j d α , Ω / 2 ∏ i = 1 k ψ 2 j α i R ( − θ j ) x i − m i ,
where θ j : = | Ω | log 2 − j is the rotation angle accumulated at scale 2 − j (so that θ j → 0 as j → ∞ , as required for the wavelet to be well-localised), and R ( − θ j ) is the rotation by angle − θ j about the axis Ω / | Ω | (with R ( 0 ) = Id ).
The volume of the fundamental domain under the composition T ˜ 2 − j , θ j α , Ω is controlled by
d α , Ω = ∑ i = 1 k 1 α i + κ rot | Ω | 2 / ν , κ rot = 2 ν 2 / ν ∑ i = 1 k 1 α i 1 − 2 / ν ,
which we treat as a modelling ansatz for the effective anisotropic–rotational dimension; a fully rigorous derivation of κ rot from the geometry of the composition T λ α ∘ R ( θ ) is beyond the scope of this paper. The term κ rot | Ω | 2 / ν accounts for the additional volume caused by the rotation twist.
The system { ψ j , m Ω } is an orthonormal basis of L 2 and satisfies the norm equivalence
∥ f ∥ W α , Ω ν , p ≍ ∑ j ∈ Z 2 j ν p ∑ m ∈ Z k | 〈 f , ψ j , m Ω 〉 | p 1 / p .
Truncating at scale J gives ∥ f − f J ∥ L ∞ ≤ C 2 − J ν ∥ f ∥ W α , Ω ν , p . Each wavelet can be approximated by a ReLU network with O ( log ( 1 / ε ) ) layers and O ( 1 ) parameters [25]. Choosing ε = 2 − J ν and summing over all active wavelets (whose number is ∼ 2 J d α , Ω ) using a binary tree yields a total depth L ∼ J d α , Ω and number of parameters N ∼ 2 J d α , Ω . Hence 2 − J ν ∼ N − ν / d α , Ω and L ∼ log N . The final bound is
∥ f − N θ ∥ L ∞ ≤ C ( ν , p , α , Ω ) L N − ν / d α , Ω ∥ f ∥ W α , Ω ν , p ,
with C ( ν , p , α , Ω ) = C 1 exp ( c 1 | Ω | ) for explicit constants C 1 = C 1 ( ν , p , α ) and c 1 = c 1 ( ν , α ) .

6. Results

In this section we collect the principal theoretical achievements of the paper, summarizing the main theorems and their implications.

6.1. Summary of the Main Inequalities

[Mixed fractional Landau inequalities for rotating spaces] Let ν ∈ ( 1 , 2 ) , 1 < p < ∞ , α ∈ ( 0 , ∞ ) k with α min > 2 ν , p > D α / ν , Ω ∈ R k . For any f ∈ W α , Ω ν , p ( R k ) and any direction i = 1 , … , k ,
∥ D i 1 f ∥ L ∞ ≤ C ( ν , p , α , Ω ) ∥ f ∥ L ∞ 1 − 1 / ν ∑ j = 1 k ∥ D j ν , α j , Ω f ∥ L p 1 / ν ,
with C ( ν , p , α , Ω ) = C 0 exp c 0 | Ω | ∑ i = 1 k α i − 1 , where c 0 = 2 1 / ν and C 0 depends only on ν , p , α . For higher orders m ∈ N with ν ∈ ( m , m + 1 ) ,
∥ D i m f ∥ L ∞ ≤ C m ( ν , p , α , Ω ) ∥ f ∥ L ∞ 1 − m / ν ∑ j = 1 k ∥ D j ν , α j , Ω f ∥ L p m / ν .
These inequalities extend the classical Landau–Kolmogorov estimates to the anisotropic rotating setting with explicit constants depending on the anisotropy vector α and the rotation rate | Ω | .

6.2. A Priori Estimates for Rotating Compressible Flows

[Landau estimate for rotating compressible flows] Let ( ρ , u ) be a smooth solution of (81)–() on [ 0 , T ] × R k with initial data in W α , Ω ν , p × ( W α , Ω ν , p ) k , ν ∈ ( 1 , 2 ) , α min > 2 ν , p > D α / ν . Then for any i = 1 , … , k ,
∥ ∂ x i u ∥ L ∞ ≤ C flow ∥ ρ ∥ L ∞ 1 − 1 / ν + ∥ u ∥ L ∞ 1 − 1 / ν ∑ j = 1 k ∥ D j ρ ∥ L p + ∥ D j u ∥ L p 1 / ν ,
where C flow depends on μ , λ , γ , α , Ω and polynomially on the Mach number M.

6.3. Stability and Approximation by Neural Operators

[Stability of rotating neural operators] Let N θ be a rotating neural operator with L layers satisfying ∥ W l e i ∥ ℓ 2 ≤ λ l , i 1 / ( L α i ) e | Ω | / L with λ l , i ≤ Λ and with activation functions σ l ∈ C b 1 , 1 , ∥ σ l ′ ∥ ∞ ≤ 1 . Then for any input perturbation ∥ δ x ∥ ∞ ≤ ϵ ,
∥ N θ ( x + δ x ) − N θ ( x ) ∥ L ∞ ≤ k e | Ω | Λ 1 / α min ϵ .
[Approximation rates for rotating flows] For any f ∈ W α , Ω ν , p ( R k ) with ν > D α / p , there exists a rotating neural operator N θ with depth L and N parameters per layer such that
∥ f − N θ ∥ L ∞ ≤ C ( ν , p , α , Ω ) L N − ν / d α , Ω ∥ f ∥ W α , Ω ν , p ,
where d α , Ω = ∑ i α i − 1 + κ rot | Ω | 2 / ν , κ rot = ( 2 / ν ) 2 / ν ( ∑ i α i − 1 ) 1 − 2 / ν , and C ( ν , p , α , Ω ) = C 1 exp ( c 1 | Ω | ) . Since d α , Ω increases with | Ω | , rotation increases the effective dimension of the approximation problem, so the rate N − ν / d α , Ω is slower than in the non-rotating case.

6.4. Key Implications

  • The mixed fractional Landau inequalities unify anisotropy, fractional dissipation, and rotation, providing explicit estimates that degenerate to known results when Ω = 0 or α i ≡ 1 .
  • The explicit exponential dependence e c | Ω | in the constant reflects the potential stabilising effect of strong rotation: for fixed regularity, increasing | Ω | enlarges the constant, but the effective dimension d α , Ω also increases, leading to a trade-off.
  • The approximation rate N − ν / d α , Ω demonstrates that rotation increases the effective dimensionality of the approximation problem, making deep learning of rotating flows more challenging than in the non-rotating case.

7. Conclusions

We have introduced rotating fractional Sobolev spaces W α , Ω ν , p ( R k ) and proved anisotropic mixed fractional Landau inequalities that incorporate Coriolis forces. The constant depends explicitly on the rotation rate | Ω | via an exponential factor. These inequalities were applied to derive a priori estimates for the rotating compressible Navier–Stokes system at high Mach numbers. Moreover, we established stability and approximation theorems for neural operators that respect the anisotropic rotating geometry. The approximation rate N − ν / d α , Ω reveals that rotation increases the effective dimensionality of the problem, slowing convergence.
Future work will extend these results to bounded domains, non-constant rotation, and stochastic flows. In addition, we plan to implement numerical codes to computationally validate the derived inequalities and approximation rates, providing practical evidence of the theoretical findings.

Author Contributions

Santos, R.D.C.: Conceptualization, methodology, formal analysis, investigation, writing — original draft. Andrade, D.A.: Supervision, resources. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Data Availability Statement

Not applicable.

Acknowledgments

The authors would like to thank IPEN of the National Nuclear Energy Commission — São Paulo (CNEN/IPEN–SP) for the institutional and financial support during the postgraduate (postdoctoral) period. This work was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior — Brasil (CAPES) — Finance Code 001.

Conflicts of Interest

The authors declare no conflicts of interest.

Notation

Table 1. List of symbols and notations.
Table 1. List of symbols and notations.
Symbol Description
Anisotropic and rotational geometry
α = ( α 1 , … , α k ) Anisotropy (scaling) vector, α i > 0 .
α min = min i α i Minimum anisotropy exponent.
k Spatial dimension ( k = 2 or 3 in applications).
D α = ∑ i = 1 k α i Anisotropic homogeneous dimension.
d α = ∑ i = 1 k α i − 1 Anisotropic dimension (volume scaling exponent).
ρ α ( x ) = ∑ i = 1 k | x i | 2 / α i 1 / 2 Anisotropic quasi-distance, homogeneous of degree 1.
e i i-th canonical basis vector of R k .
Ω ∈ R k Constant rotation vector.
| Ω | Rotation rate.
ε i j ℓ Levi-Civita symbol.
L Ω Skew-symmetric matrix of the cross product x ↦ Ω × x .
R ( θ ) = e θ L Ω / | Ω | Rotation by angle θ about the axis Ω / | Ω | .
T λ α x = ( λ α 1 x 1 , … , λ α k x k ) Anisotropic dilation group.
A α = diag ( α 1 , … , α k ) Infinitesimal generator of dilations.
T ˜ λ , θ α , Ω = T λ α ∘ R ( θ ) Rotating anisotropic scaling (composition).
B α ( x , r ) = { y : | y i − x i | < r 1 / α i } Anisotropic ball.
| B α ( x , r ) | = ω d α r d α Volume of anisotropic ball, ω d α : = | B α ( 0 , 1 ) | = 2 k .
Fractional calculus and function spaces
ν > 0 Fractional order of differentiation.
p ∈ [ 1 , ∞ ) Integrability exponent in Lebesgue spaces.
h ∈ R Increment in the rotating finite difference.
Δ h , i Ω f ( x ) = f ( x + h e i ) − f ( x + h L Ω e i ) Rotating finite difference in direction e i .
[ f ] W i ν / α i , p , Ω Rotating directional Gagliardo seminorm.
W α , Ω ν , p ( R k ) Rotating mixed fractional Sobolev space.
D j ν , α j , Ω Rotating fractional derivative, symbol ( i ξ j + i ( Ω × e j ) · ξ ) ν / α j .
ω i ( ξ ) = ξ i + ( Ω × e i ) · ξ Rotating frequency (Fourier multiplier).
η i = e i + Ω × e i Oblique direction vector.
Λ ν , α i = 4 π 2 ν / α i ν / α i Γ ( 1 − 2 ν / α i ) − 1 Constant in the Fourier characterisation (Lemma 2.13).
C ν , α i = Γ ( 2 − ν / α i ) − 1 Constant in the Weyl representation (Theorem Section 3.2).
Harmonic analysis and equivalent norms
Δ j Standard anisotropic Littlewood–Paley blocks.
Δ j ( i ) Directional blocks localised in | ξ i | ∼ 2 j α i .
L α , Ω = − ∑ i ( − ∂ i 2 ) 1 / α i + i Ω · ( x × ∇ ) Generator of rotating fractional semigroup.
σ ( x , ξ ) Symbol of L α , Ω .
Σ ω Sector of analyticity of L α , Ω .
B α , Ω ν , p , S α , Ω ν , p , K α , Ω ν , p Besov, semigroup, and K-functional norms (Theorem Section 2.2).
A 1 , A 2 , A 3 Equivalence constants in (30).
U Ω Conjugation Fourier multiplier, symbol e − i Ω · ( x × ξ ) / | ξ | .
M α , Ω f ( x ) = sup r > 0 1 | B α ( x , r ) | ∫ B α ( x , r ) | f ( y ) | d y Rotating anisotropic maximal operator.
κ ( p , α ) = 2 p 1 / p ( p − 1 ) 1 / p 3 d α / p Constant in Theorem Section 2.4.
Landau inequality constants
C ( ν , p , α , Ω ) Constant in Theorem Section 3.2.
C 0 = 2 1 + 1 / ν κ ( p , α ) 1 / ν Γ ( 1 − ν / α min ) − 1 / ν Anisotropic constant in C ( ν , p , α , Ω ) .
c 0 = 2 1 / ν Universal constant in the exponential factor.
C m ( ν , p , α , Ω ) Constant in Theorem Section 3.3.
Navier–Stokes and fluid dynamics
ρ Density field.
u Velocity field.
p = a ρ γ Pressure law (polytropic gas).
a > 0 Pressure constant.
γ > 1 Adiabatic exponent.
μ , λ Standard shear and bulk viscosities.
μ α , i , λ α , i Direction-dependent fractional viscosities, μ α , i = μ ℓ 2 − 2 / α i .
ℓ Characteristic length scale.
Φ Centrifugal potential.
M = | u | / c Mach number.
c = γ p / ρ Speed of sound.
C ( M ) Mach-dependent constant (polynomial growth).
T Time interval [ 0 , T ] .
C flow Final constant.
Neural operators
N θ Neural operator with parameters θ .
L Number of layers (depth).
N Number of trainable parameters per layer.
W l , b l Weight matrix and bias of layer l.
σ l ∈ C b 1 , 1 ( R ) Activation function of layer l, ∥ σ l ′ ∥ ∞ ≤ 1 , ∥ σ l ′ ′ ∥ ∞ ≤ K .
K Bound on ∥ σ l ′ ′ ∥ ∞ .
λ l , i Column bounds, ∥ W l e i ∥ ℓ 2 ≤ λ l , i 1 / ( L α i ) e | Ω | / L .
Λ Uniform bound λ l , i ≤ Λ .
ϵ Input perturbation bound, ∥ δ x ∥ ∞ ≤ ϵ .
C stab = k e | Ω | Λ 1 / α min Stability constant in Theorem Section 5.
ψ j , m Ω Rotating wavelet basis.
κ rot = 2 ν 2 / ν ∑ i α i − 1 1 − 2 / ν Constant in d α , Ω (modelling ansatz).
d α , Ω = ∑ i α i − 1 + κ rot | Ω | 2 / ν Effective anisotropic–rotational dimension.
C 1 , c 1 Constants in the approximation rate.

References

  1. Landau, E. Die Ungleichungen für zweimal differentiierbare Funktionen; AF Høst & søn, 1925; Vol. 6. [Google Scholar]
  2. Kolmogorov, A. N. On inequalities between the upper bounds of the successive derivatives of an arbitrary function on an infinite interval. 1949; No. 4.
  3. Stein, E. M. Singular integrals and differentiability properties of functions; Princeton university press, 1970; Vol. 2. [Google Scholar]
  4. Muckenhoupt, B. Weighted norm inequalities for the Hardy maximal function. Trans. Am. Math. Soc. 165 1972, 207–226. [Google Scholar] [CrossRef]
  5. Pedlosky, J. Geophysical fluid dynamics. J. Appl. Mech. 1981. [Google Scholar] [CrossRef]
  6. Lions, P. L. The concentration-compactness principle in the Calculus of Variations. The locally compact case, part 1. In Annales de l’Institut Henri Poincaré C, Analyse non linéaire; No longer published by; Elsevier, March 1984; Vol. 1, No. 2, pp. 109–145. [Google Scholar] [CrossRef]
  7. Schmeisser, H. J.; Triebel, H. Topics in Fourier analysis and function spaces; No Title; 1987. [Google Scholar]
  8. Constantin, P.; Foiaş, C. Navier-stokes equations; University of Chicago press, 1988. [Google Scholar]
  9. Samko, S. G. Fractional integrals and derivatives. In Theory and applications; 1993. [Google Scholar]
  10. Grigor’yan, A. Heat kernel upper bounds on a complete non-compact manifold. Rev. Matemática Iberoam. 1994, 10(2), 395–452. [Google Scholar] [CrossRef]
  11. Temam, R. Navier–Stokes equations and nonlinear functional analysis. In Society for industrial and applied mathematics; 1995. [Google Scholar]
  12. Edmunds, D. E.; Triebel, H. Function spaces, entropy numbers, differential operators; Cambridge University Press: Cambridge, 1996; Vol. 120. [Google Scholar]
  13. Runst, T.; Sickel, W. Sobolev spaces of fractional order, Nemytskij operators, and nonlinear partial differential equations; Walter de Gruyter, 1996. [Google Scholar]
  14. Kounchev, O. Extremizers for the multivariate Landau-Kolmogorov inequality. Math. Res. 101 1997, 123–132. [Google Scholar]
  15. DeVore, R. A. Nonlinear approximation. Acta Numer. 7 1998, 51–150. [Google Scholar] [CrossRef]
  16. Maz’ya, V.; Shaposhnikova, T. On pointwise interpolation inequalities for derivatives. Math. Bohem. 1999, 124(2), 131–148. [Google Scholar] [CrossRef]
  17. Lieb, E. H.; Loss, M. Analysis; American Mathematical Soc., 2001; Vol. 14. [Google Scholar]
  18. Babenko, V. F.; Babenko, Y. V. About Kolmogorov-type inequalities for functions defined on a half-line. International Conference “Constructive Theory of Functions”, Varna, 2003; pp. 205–208. [Google Scholar]
  19. Adams, R. A.; Fournier, J. J. Sobolev spaces; Elsevier, 2003; Vol. 140. [Google Scholar]
  20. Triebel, H. Theory of function spaces III; Birkhäuser Basel: Basel, 2006. [Google Scholar]
  21. Bahouri, H.; Chemin, J. Y.; Danchin, R. Fourier analysis and nonlinear partial differential equations; Springer: Berlin, 2011; Vol. 343, pp. 523–pages. [Google Scholar] [CrossRef]
  22. Majda, A. Compressible fluid flow and systems of conservation laws in several space variables; Springer Science & Business Media, 2012. [Google Scholar]
  23. Gradshteyn, I. S.; Ryzhik, I. M. Table of integrals, series, and products; Academic press, 2014. [Google Scholar]
  24. Ferrari, F. Weyl and Marchaud derivatives: A forgotten history. Mathematics 2018, 6(1), 6. [Google Scholar] [CrossRef]
  25. Yarotsky, D. Error bounds for approximations with deep ReLU networks. Neural Netw. 94 2017, 103–114. [Google Scholar] [CrossRef]
  26. Colombo, M.; De Lellis, C.; Massaccesi, A. The Generalized Caffarelli-Kohn-Nirenberg Theorem for the Hyperdissipative Navier-Stokes System. Commun. Pure Appl. Math. 2020, 73(3), 609–663. [Google Scholar] [CrossRef]
  27. Schäfer, M.; Ullrich, T.; Vedel, B. Hyperbolic wavelet analysis of classical isotropic and anisotropic Besov–Sobolev spaces. J. Fourier Anal. Appl. 2021, 27(3), 51. [Google Scholar] [CrossRef]
  28. Anastassiou, G. A. Multivariate left side Canavati fractional Landau inequalities. J. Appl. Pure Math. 2025, 7(1–2), 103–119. [Google Scholar] [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.