Introduction
The main motivation for the present work arises in the theory of spontaneous stochasticity for turbulent flows [1 ] . Although the pre-history of this notion lies in the classic study of Richardson on 2-particle turbulent dispersion [2 ] , the concept was first clearly formulated in the work of Bernard, Gawędzki, and Kupiainen [3 ] . Their fundamental observation was that the Hölder exponent of the infinite-Reynolds velocity field must be ≤ 1 / 3 \le1/3 ≤ 1/3 , according to a theorem of Onsager [4 ] , and this regularity is too low to guarantee uniqueness of the solutions to the ODE for Lagrangian particle trajectories. The authors of [3 ] showed that this non-uniqueness manifests in remnant stochasticity of Brownian particles (e.g. dye molecules) in the limit of large Reynolds and Péclet numbers, at least for the soluble model of Kraichnan [5 ] , even though the limiting evolution equations for a fixed velocity realization are deterministic. Furthermore, they showed that such “Lagrangian spontaneous stochasticity” is necessary in general for a scalar advected by any turbulent flow, either passively or actively, to exhibit anomalous dissipation in the infinite Reynolds & Péclet limit [6 ] . More recently, it was pointed out by Mailybaev [7 , 8 ] that there is a closely allied phenomenon of “Eulerian spontaneous stochasticity” in which the entire velocity field obtained by solving the deterministic Navier-Stokes equations remains random in the infinite-Reynolds limit, even for initial data with vanishing randomness. This result was anticipated in the profound work of Lorenz [9 ] on turbulence predictability, who argued that “certain formally deterministic fluid systems which possess many scales of motion are observationally indistinguishable from indeterministic systems.” Spontaneous stochasticity provides a precise mathematical formulation of Lorenz’ ideas, in which the “formally deterministic fluid system” exhibiting random solutions is governed by the ideal Euler equations with fixed initial data, obtained in the simultaneous limit of infinite Reynolds number and vanishing noise. The mathematical fact which underlies this persistent randomness is the non-uniqueness of solutions to the Cauchy initial-value problem for the limiting Euler equations, even with additional admissibility constraints imposed such as Hölder regularity and local dissipativity [10 , 11 , 12 ] . The prototypical example of a Kelvin-Helmholtz unstable vortex sheet provides strong empirical evidence for this phenomenon [13 , 14 ] .
The vanishing noise considered in the work of Lorenz was small-scale external perturbations, such as the proverbial flap of a butterfly’s wing [15 ] . Intrinsic thermal noise has been more recently investigated, following the pioneering work of Betchov [16 , 17 ] and Ruelle [18 ] . The distinct feature of thermal noise is that it is unavoidable, since it is a necessary consequence of the local thermodynamic equilibrium required for validity of a hydrodynamic description. The simplest fluid dynamical model which incorporates thermal fluctuation effects is Landau-Lifschitz fluctuating hydrodynamics [19 ] , a stochastic modification of the Navier-Stokes equations. For incompressible fluids in the limit of low Mach numbers, this model takes the form in any space dimension d ≥ 2 d \ge2 d ≥ 2
∂ t u + π Λ ( u ⋅ ∇ ) u = − ∇ p + ν Λ Δ u + 2 ν Λ k B T ρ ∇ ⋅ η , ∇ ⋅ u = 0 , (1) \partial_t \mathbf{u} + \pi_{\Lambda}(\mathbf{u} \cdot\nabla)\mathbf{u} = -\nabla p + \nu_{\Lambda}\Delta\mathbf{u} + \sqrt{\frac{2\nu_{\Lambda}k_{\mathrm{B}}T}{\rho}}\nabla\cdot\boldsymbol{\eta}, \qquad\nabla\cdot\mathbf{u} = 0, \tag*{(1)} ∂ t u + π Λ ( u ⋅ ∇ ) u = − ∇ p + ν Λ Δ u + ρ 2 ν Λ k B T ∇ ⋅ η , ∇ ⋅ u = 0 , (1)
where π Λ \pi_{\Lambda} π Λ is the projection to Fourier modes with wavenumbers ∣ k ∣ < Λ |\mathbf{k}| < \Lambda ∣ k ∣ < Λ , ν Λ \nu_{\Lambda} ν Λ is (scale-dependent) kinematic viscosity, T T T is absolute temperature, k B k_{\mathrm{B}} k B is Boltzmann’s constant, ρ \rho ρ is mass density, and τ = ( 2 ν Λ k B T / ρ ) 1 / 2 η \boldsymbol{\tau} = (2\nu_{\Lambda}k_{\mathrm{B}}T/\rho)^{1/2}\boldsymbol{\eta} τ = ( 2 ν Λ k B T / ρ ) 1/2 η is a fluctuating stress modeled as a Gaussian random matrix field, both symmetric and traceless, with mean zero and covariance
⟨ η i j ( x , t ) η k l ( x ′ , t ′ ) ⟩ = ( δ i k δ j l + δ i l δ j k − 2 3 δ i j δ k l ) δ Λ d ( x − x ′ ) δ ( t − t ′ ) . (2) \left\langle\eta_{ij}(\mathbf{x},t)\eta_{kl}(\mathbf{x}',t')\right\rangle= \left(\delta_{ik}\delta_{jl}+\delta_{il}\delta_{jk}-\frac{2}{3}\delta_{ij}\delta_{kl}\right)\delta_{\Lambda}^{d}(\mathbf{x}-\mathbf{x}')\delta(t-t'). \tag*{(2)} ⟨ η ij ( x , t ) η k l ( x ′ , t ′ ) ⟩ = ( δ ik δ j l + δ i l δ j k − 3 2 δ ij δ k l ) δ Λ d ( x − x ′ ) δ ( t − t ′ ) . (2)
The kinematic pressure p p p is obtained from the incompressibility constraint ∇ ⋅ u = 0 \nabla\cdot\mathbf{u} = 0 ∇ ⋅ u = 0 . For example, see [20 , 21 ] . Although the model in (1),(2) appears superficially to be given by a stochastic partial differential equation, such a continuum interpretation is not the customary one in the physics literature. Instead, the model is understood to contain an explicit high-wavenumber cutoff Λ \Lambda Λ , which may be freely chosen in the range ℓ ν − 1 ≪ Λ ≪ λ m f p − 1 \ell_{\nu}^{-1} \ll\Lambda\ll\lambda_{mfp}^{-1} ℓ ν − 1 ≪ Λ ≪ λ m f p − 1 where ℓ ν \ell_{\nu} ℓ ν is the gradient length of the flow and λ m f p \lambda_{mfp} λ m f p is the molecular mean free path length of the fluid. In the language of theoretical physics, the model is an “effective field theory” which describes molecular fluids only at large length-scales > ∣ Λ ∣ − 1 > |\Lambda|^{-1} > ∣Λ ∣ − 1 [22 , 23 ] . In particular, the subscript on δ Λ d \delta_{\Lambda}^{d} δ Λ d in (2) indicates that this is not a true delta-function, but only an effective delta-function for band-limited functions restricted to wavenumbers ∣ k ∣ < Λ |\mathbf{k}| < \Lambda ∣ k ∣ < Λ . For more discussion, see Section 2.
Simulations of turbulence in the presence of thermal noise have detected sensitivity of large-scale flow features in individual flow realizations to such noise, even at high Reynolds numbers [24 ] . The prior work of one of us [25 ] studied systematically the limit of high Reynolds number R e = U L / ν Re = UL/\nu R e = U L / ν for the dynamics (1), which, after non-dimensionalization based on large-scale flow parameters, takes the form
∂ t u + π ^ Λ ( u ⋅ ∇ ) u = − ∇ p + 1 R e Δ u + 2 θ η R e 3 ( d + 1 ) / 4 ∇ ⋅ η + F f . (3) \partial_t\mathbf{u}+\hat{\pi}_{\Lambda}(\mathbf{u}\cdot\nabla)\mathbf{u}=-\nabla p+\frac{1}{Re}\Delta\mathbf{u}+\sqrt{\frac{2\theta_{\eta}}{Re^{3(d+1)/4}}}\nabla\cdot\boldsymbol{\eta}+\mathbf{F}\mathbf{f}. \tag*{(3)} ∂ t u + π ^ Λ ( u ⋅ ∇ ) u = − ∇ p + R e 1 Δ u + R e 3 ( d + 1 ) /4 2 θ η ∇ ⋅ η + Ff . (3)
See section 2 for detailed discussion. Here we note only that θ η \theta_{\eta} θ η is the ratio of thermal energy to the energy of a Kolmogorov-scale eddy, which is a generally tiny quantity of order 10 − 8 10^{-8} 1 0 − 8 or smaller and fixed as R e → ∞ Re \to\infty R e → ∞ . We have added as well a large-scale body force f \mathbf{f} f to drive turbulence, with a dimensionless amplitude F F F which is also fixed as R e → ∞ Re \to\infty R e → ∞ . It is important to note, however, that the dimensionless wavenumber cutoff Λ ^ = Λ L \hat{\Lambda}=\Lambda L Λ ^ = Λ L diverges together with R e Re R e , e.g. Λ ^ ∝ R e 3 / 4 \hat{\Lambda}\propto Re^{3/4} Λ ^ ∝ R e 3/4 if the cutoff is fixed relative to the Kolmogorov dissipation scale η \eta η or Λ ^ ∝ R e \hat{\Lambda}\propto Re Λ ^ ∝ R e if instead the cutoff is fixed relative to the molecular mean free path λ m f p \lambda_{mfp} λ m f p . Therefore, formally, the limiting dynamics as R e → ∞ Re \to\infty R e → ∞ is described by the forced continuum Euler equations
∂ t u + ∇ ⋅ ( u ⊗ u ) = − ∇ p + F f , (4) \partial_t\mathbf{u}+\nabla\cdot(\mathbf{u}\otimes\mathbf{u})=-\nabla p+\mathbf{F}\mathbf{f}, \tag*{(4)} ∂ t u + ∇ ⋅ ( u ⊗ u ) = − ∇ p + Ff , (4)
with both the viscous diffusion and the thermal fluctuations vanishing in the limit. To investigate numerically the limiting behavior for R e ≫ 1 Re \gg1 R e ≫ 1 , the authors of [25 ] studied a simplified “shell model” version of eq.(1) which allowed them to simulate a broad range of Reynolds numbers 10 6 10^{6} 1 0 6 –10 8 10^{8} 1 0 8 . For quasi-singular initial data u 0 \mathbf{u}_{0} u 0 with K41 spectra over a broad range of wave-numbers, or with even more singular initial data u 0 \mathbf{u}_{0} u 0 selected from a steady-state turbulent simulation, the transition probability P R e ( u , t ∣ u 0 , t 0 ) P_{Re}(\mathbf{u},t|\mathbf{u}_{0},t_{0}) P R e ( u , t ∣ u 0 , t 0 ) was found to become R e Re R e -independent for sufficiently large Reynolds number. Most remarkably, the limiting statistics were found to be very universal, independent of the selected sequence R e k → ∞ Re_k \to\infty R e k → ∞ , the precise details of the thermal noise model, etc. This universality is one of the hallmarks of spontaneous stochasticity [14 , 26 , 27 ] .
The empirical observation by [25 ] of non-trivial infinite-R e Re R e limits of transition probabilities for fixed initial data u 0 \mathbf{u}_0 u 0 , strongly suggests the existence of non-unique solutions of the Cauchy problem for the limiting Euler equation (4). Such a non-uniqueness is consistent with the recent results from convex integration constructions, [10 , 11 , 12 ] , that provide a mathematical underpinning for spontaneous stochasticity. However, the Euler solutions provided by convex integration do not correspond to infinite-Reynolds limits. Furthermore, no mathematical justification was offered in the numerical study of high-R e Re R e turbulence [25 ] that the limiting velocity realizations should indeed be (weak) solutions of the ideal Euler equation (4). The purpose of the present paper is to fill this gap and to prove that infinite-R e Re R e limiting probability measures exist for the Landau-Lifshitz dynamics (3) that are supported indeed on such weak solutions of the Euler equations. We shall model our proofs closely after the work of Vishik & Fursikov [28 ] , who generalized Leray’s construction of weak solutions of the incompressible Navier-Stokes equation to space-time statistical solutions of incompressible Navier-Stokes equation with random initial data. In fact, the monograph [28 ] developed a very general approach, which can be applied to a great many models described by PDE’s or stochastic PDE’s. Unlike the work on Navier-Stokes, however, we cannot derive the necessary regularity for the limit from a priori energy estimates. As is well-known, the compactness that follows from kinetic energy bounds alone does not suffice to yield standard weak Euler solutions in the infinite-R e Re R e limit, but instead only still more generalized notions of Euler solution such as the “dissipative solutions” in the sense of Lions [29 ] or measure-valued weak solutions [30 ] . While it cannot be ruled out that such generalized weak solutions are physically relevant [31 , 32 ] , it was pointed out by Chen & Glimm [33 ] that standard weak solutions can be obtained in the infinite-R e Re R e limit if additional conditions are satisfied that may be empirically verified. See also following work [34 , 35 , 36 , 37 ] . We follow a version of this strategy here, invoking a condition of spatial Besov regularity uniform in the Reynolds number that was first employed to study inviscid limits in [35 ] but which in fact goes back to a remark of Isett [37 ] . Recently, Fjordholm et al. [38 ] have proved an inviscid limit result with a similar assumption, but in the framework of statistical solutions in the Foiaş-Prodi sense, which involves single-time measures on spatial velocity fields. See Remark 1 below.
Theorem Statement. We now formulate in precise mathematical terms the main result of this work. We take as flow domain the d d d -dimensional torus T d = [ 0 , 2 π ] d \mathbb{T}^d = [0,2\pi]^d T d = [ 0 , 2 π ] d , d ≥ 2 d \ge2 d ≥ 2 , in order to avoid issues with boundaries. Indeed, the proper formulation of fluctuating hydrodynamics in the presence of solid walls is still a subject under active investigation [39 ] . We consider the incompressible fluid velocity field u ν \mathbf{u}^{\nu} u ν governed by the Landau-Lifshitz-Navier-Stokes (LLNS) equations written in the dimensionless form (3), or
d u ν = P [ − ∇ ⋅ π Λ ( u ν ⊗ u ν ) + ν Δ u ν + π Λ f ] d t + ν κ P ∇ ⋅ π Λ d ξ , θ > 0 (5) \mathrm{d}\mathbf{u}^{\nu} = \mathcal{P}\left[-\nabla\cdot\pi_{\Lambda}(\mathbf{u}^{\nu}\otimes\mathbf{u}^{\nu})+\nu\Delta\mathbf{u}^{\nu}+\pi_{\Lambda}\mathbf{f}\right]\mathrm{d}t+\nu^{\kappa}\mathcal{P}\nabla\cdot\pi_{\Lambda}\mathrm{d}\xi,\qquad\theta>0 \tag*{(5)} d u ν = P [ − ∇ ⋅ π Λ ( u ν ⊗ u ν ) + ν Δ u ν + π Λ f ] d t + ν κ P ∇ ⋅ π Λ d ξ , θ > 0 (5)
where we follow standard practice in the mathematics literature of denoting ν = 1 / R e \nu=1/Re ν = 1/ R e and considering the infinite Reynolds number limit as an “inviscid limit.” We have, for simplicity, set θ η = F = 1 \theta_\eta=F=1 θ η = F = 1 , since these finite constants are fixed in the limit ν → 0 \nu\to0 ν → 0 and play no essential role in the argument. The projection π Λ \pi_{\Lambda} π Λ is defined in terms of the Fourier transform ( F f ) ( k ) = 1 ( 2 π ) d ∫ T d e − i k ⋅ x f ( x ) d d x (\mathcal{F}f)(\mathbf{k})=\frac{1}{(2\pi)^d}\int_{\mathbb{T}^d}e^{-i\mathbf{k}\cdot\mathbf{x}}f(\mathbf{x})\,\mathrm{d}^d x ( F f ) ( k ) = ( 2 π ) d 1 ∫ T d e − i k ⋅ x f ( x ) d d x by
π Λ f ( x ) = ∑ k ∈ Z d , ∣ k ∣ ∞ < Λ e i k ⋅ x ( F f ) ( k ) , (6) \pi_{\Lambda}f(\mathbf{x})=\sum_{\mathbf{k}\in\mathbb{Z}^d,\ |\mathbf{k}|_{\infty}<\Lambda}e^{i\mathbf{k}\cdot\mathbf{x}}(\mathcal{F}f)(\mathbf{k}), \tag*{(6)} π Λ f ( x ) = k ∈ Z d , ∣ k ∣ ∞ < Λ ∑ e i k ⋅ x ( F f ) ( k ) , (6)
where for technical reasons of the proof related to Lizorkin representation of function spaces, we use the max-norm ∣ k ∣ ∞ = max i = 1 d ∣ k i ∣ |\mathbf{k}|_{\infty}=\max_{i=1}^{d}|k_i| ∣ k ∣ ∞ = max i = 1 d ∣ k i ∣ and Λ = 2 m \Lambda=2^m Λ = 2 m with m ( ν ) ∈ N m(\nu)\in\mathbb{N} m ( ν ) ∈ N such that Λ − α → 0 \Lambda^{-\alpha}\to0 Λ − α → 0 as ν → 0 \nu\to0 ν → 0 , 3 / 4 ≤ α ≤ 1 3/4\le\alpha\le1 3/4 ≤ α ≤ 1 . The proof would work for a wider range of α \alpha α but, as discussed in the Introduction Section 2, only the above range is of physical relevance. We have eliminated the pressure which enforces incompressibility in (3) by the Leray–Hodge projection P ≡ I − ∇ Δ − 1 div \mathcal{P}\equiv I-\nabla\Delta^{-1}\operatorname{div} P ≡ I − ∇ Δ − 1 div , given in Fourier representation by
( F P u ) i ( k ) = ∑ j = 1 d ( δ i j − k i k j ∣ k ∣ 2 ) ( F u ) j ( k ) , i = 1 , … , d . (\mathcal{F}\mathcal{P}\mathbf{u})_i(\mathbf{k})=\sum_{j=1}^{d}\left(\delta_{ij}-\frac{k_i k_j}{|\mathbf{k}|^2}\right)(\mathcal{F}\mathbf{u})_j(\mathbf{k}),\qquad i=1,\ldots,d. ( F P u ) i ( k ) = j = 1 ∑ d ( δ ij − ∣ k ∣ 2 k i k j ) ( F u ) j ( k ) , i = 1 , … , d .
To represent the random forcing by thermal noise, we take a probability space ( Ω , F , P ) (\Omega,\mathcal{F},\mathbb{P}) ( Ω , F , P ) , an increasing filtration ( F t : t ∈ [ 0 , T ] ) (\mathcal{F}_t:t\in[0,T]) ( F t : t ∈ [ 0 , T ]) and an F t \mathcal{F}_t F t -adapted cylindrical Brownian motion ξ \xi ξ on the Hilbert space L 2 ( T d ) L^2(\mathbb{T}^d) L 2 ( T d ) [40 ] , with ⟨ ξ ( t ) , h ⟩ \langle\xi(t),h\rangle ⟨ ξ ( t ) , h ⟩ for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , h ∈ L 2 ( T d ) h\in L^2(\mathbb{T}^d) h ∈ L 2 ( T d ) taking values in the space of symmetric, traceless d × d d\times d d × d matrices, with mean E [ ⟨ ξ ( t ) , h ⟩ ] = 0 \mathbb{E}[\langle\xi(t),h\rangle]=0 E [⟨ ξ ( t ) , h ⟩] = 0 and covariance
E [ ξ i j ( t ) ξ k l ( t ′ ) , h ⟩ ] = ( δ i k δ j l + δ i l δ j k − 2 3 δ i j δ k l ) ⟨ h , h ′ ⟩ ( t ∧ t ′ ) . \mathbb{E}[\xi_{ij}(t)\xi_{kl}(t'),h\rangle]=\left(\delta_{ik}\delta_{jl}+\delta_{il}\delta_{jk}-\frac{2}{3}\delta_{ij}\delta_{kl}\right)\langle h,h'\rangle(t\wedge t'). E [ ξ ij ( t ) ξ k l ( t ′ ) , h ⟩] = ( δ ik δ j l + δ i l δ j k − 3 2 δ ij δ k l ) ⟨ h , h ′ ⟩ ( t ∧ t ′ ) .
The cylindrical Brownian motion can be represented [40 ] as ⟨ ξ ( t ) , h ⟩ = 1 ( 2 π ) d ∫ T d ξ ( x , t ) h ( x ) d d x \langle\xi(t),h\rangle=\frac{1}{(2\pi)^d}\int_{\mathbb{T}^d}\xi(x,t)h(x)\,d^dx ⟨ ξ ( t ) , h ⟩ = ( 2 π ) d 1 ∫ T d ξ ( x , t ) h ( x ) d d x with
ξ ( x , t ) = ∑ k ∈ Z d ξ i j k ( t ) e i k ⋅ x (7) \xi(x,t)=\sum_{k\in\mathbb{Z}^d}\xi_{ij}^k(t)e^{ik\cdot x} \tag*{(7)} ξ ( x , t ) = k ∈ Z d ∑ ξ ij k ( t ) e ik ⋅ x (7)
where ξ i j k ( t ) \xi_{ij}^k(t) ξ ij k ( t ) for k ∈ Z d k\in\mathbb{Z}^d k ∈ Z d are an infinite set of complex Brownian motions on the probability space ( Ω , F , P ) (\Omega,\mathcal{F},\mathbb{P}) ( Ω , F , P ) , with covariance
E [ ξ i j k ( t ) ξ k l k ′ ( t ′ ) ] = ( δ i k δ j l + δ i l δ j k − 2 3 δ i j δ k l ) δ k + k ′ , 0 ( t ∧ t ′ ) . \mathbb{E}[\xi_{ij}^k(t)\xi_{kl}^{k'}(t')]=\left(\delta_{ik}\delta_{jl}+\delta_{il}\delta_{jk}-\frac{2}{3}\delta_{ij}\delta_{kl}\right)\delta_{k+k',0}(t\wedge t'). E [ ξ ij k ( t ) ξ k l k ′ ( t ′ )] = ( δ ik δ j l + δ i l δ j k − 3 2 δ ij δ k l ) δ k + k ′ , 0 ( t ∧ t ′ ) .
and thus independent except for the constraints ξ i j k ( t ) ‾ = ξ i j − k ( t ) \overline{\xi_{ij}^k(t)}=\xi_{ij}^{-k}(t) ξ ij k ( t ) = ξ ij − k ( t ) , ξ i j k ( t ) = ξ j i k ( t ) \xi_{ij}^k(t)=\xi_{ji}^k(t) ξ ij k ( t ) = ξ j i k ( t ) and ∑ i = 1 d ξ i i k ( t ) = 0 \sum_{i=1}^d\xi_{ii}^k(t)=0 ∑ i = 1 d ξ ii k ( t ) = 0 . The representation (7) is particularly useful in conjunction with the projection π Λ \pi_\Lambda π Λ which simply restricts the summation over k \mathbf{k} k to ∣ k ∣ ∞ ≤ Λ |\mathbf{k}|_\infty\leq\Lambda ∣ k ∣ ∞ ≤ Λ . The prefactor ν κ \nu^\kappa ν κ of the stochastic term originates from careful non-dimensionalization of (1) with Taylor’s relation (see § 2 ), but our proof would work with any factor vanishing as ν → 0 \nu\to0 ν → 0 . We include also in (5) a deterministic stirring force f \mathbf{f} f to agitate turbulence. In addition to some mild regularity which allows for existence of solutions to (5), we require also
( F f ) ( 0 ) = 1 ( 2 π ) d ∫ T d f ( x ) d d x = 0 , (\mathcal{F}\mathbf{f})(0)=\frac{1}{(2\pi)^d}\int_{\mathbb{T}^d}\mathbf{f}(\mathbf{x})\,d^d x=0, ( F f ) ( 0 ) = ( 2 π ) d 1 ∫ T d f ( x ) d d x = 0 ,
so that there is no bulk acceleration of the fluid.
The physical necessity of the high-wavenumber cutoff Λ \Lambda Λ is crucial for the solution theory of the Landau–Lifschitz equations. In fact, the equation (5) is not an SPDE but instead a finite set of Itō SDE’s with an energy-conserving quadratic nonlinearity, of a type very well-studied mathematically. For example, the lecture notes of Flandoli [41 ] , Section 3, show that strong stochastic solutions u ν \mathbf{u}^\nu u ν of (5) exist globally in time for f = 0 \mathbf{f}=0 f = 0 and are pathwise unique, at any ν > 0 \nu>0 ν > 0 . In fact, under very modest regularity assumptions, the same is true with a force f ≠ 0 \mathbf{f}\ne0 f = 0 . Thus, for fixed initial data u 0 ν = π Λ u 0 \mathbf{u}_0^\nu=\pi_\Lambda\mathbf{u}_0 u 0 ν = π Λ u 0 , there is a well-defined distribution P u 0 ν P_{\mathbf{u}_0}^\nu P u 0 ν on solutions of (5) induced by the measure P \mathbb{P} P . If the initial data satisfies the zero-momentum condition
( F u 0 ) ( 0 ) = 1 ( 2 π ) d ∫ T d u 0 ( x ) d d x = 0 , (8) (\mathcal{F}\mathbf{u}_0)(0)=\frac{1}{(2\pi)^d}\int_{\mathbb{T}^d}\mathbf{u}_0(\mathbf{x})\,d^d x=0, \tag*{(8)} ( F u 0 ) ( 0 ) = ( 2 π ) d 1 ∫ T d u 0 ( x ) d d x = 0 , (8)
then it is furthermore easy to show [41 ] that
( F u ) ( 0 , t ) = 1 ( 2 π ) d ∫ T d u ( x , t ) d d x = 0 , ∀ t ∈ [ 0 , T ] , P u 0 ν -a.e. u (9) (\mathcal{F}\mathbf{u})(0,t)=\frac{1}{(2\pi)^d}\int_{\mathbb{T}^d}\mathbf{u}(\mathbf{x},t)\,d^d x=0,\qquad\forall t\in[0,T],\qquad P_{\mathbf{u}_0}^\nu\text{-a.e. }\mathbf{u} \tag*{(9)} ( F u ) ( 0 , t ) = ( 2 π ) d 1 ∫ T d u ( x , t ) d d x = 0 , ∀ t ∈ [ 0 , T ] , P u 0 ν -a.e. u (9)
For the purposes of studying spontaneous stochasticity, we are interested in non-smooth initial velocity data u 0 ∉ C 1 , α ( T d ) \mathbf{u}_0\notin C^{1,\alpha}(\mathbb{T}^d) u 0 ∈ / C 1 , α ( T d ) , 0 < α < 1 0<\alpha<1 0 < α < 1 , for which local existence of a smooth Euler solution is not guaranteed. This restriction would not be necessary, if the smooth Euler solution should develop a singularity in finite-time for initial data u 0 ∈ C 1 , α ( T d ) \mathbf{u}_0\in C^{1,\alpha}(\mathbb{T}^d) u 0 ∈ C 1 , α ( T d ) . However, singular initial data are natural, for example, if the velocity field u 0 \mathbf{u}_0 u 0 is chosen from a steady-state turbulent ensemble at R e = ∞ Re=\infty R e = ∞ . More modest regularity such as u 0 ∈ L p ( T d ) \mathbf{u}_0\in L^p(\mathbb{T}^d) u 0 ∈ L p ( T d ) , p > 2 p>2 p > 2 is nevertheless expected for such initial data, which can in fact be shown [41 ] to be preserved by the dynamics (5). We are thus motivated more generally to consider initial data u 0 \mathbf{u}_0 u 0 distributed according to a Borel measure μ \mu μ on L p ( T d ) L^p(\mathbb{T}^d) L p ( T d ) and satisfying (8) μ \mu μ -a.s. We denote as P μ ν P^\nu_\mu P μ ν the distribution over the solutions of (5) induced by the product probability measure P × μ \mathbb{P}\times\mu P × μ and, for simplicity, we denote expectation over the latter measure by E μ \mathbb{E}_\mu E μ . The initial data of primary interest for the study of spontaneous stochasticity are Dirac measures of the form μ = δ u 0 \mu=\delta_{\mathbf{u}_0} μ = δ u 0 , but we may always restrict attention subsequently to such a special case.
A principal ingredient of our theorem is the following hypothesis on the family of distributions P μ ν P_\mu^\nu P μ ν :
sup ν > 0 ∫ ∥ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 d P μ ν ( u ) < ∞ . (10) \sup_{\nu>0}\int\lVert\mathbf{u}\rVert_{L^r([0,T],B^\sigma_{p,\infty}(\mathbb{T}^d))}^{2}\,\mathrm{d}P_\mu^\nu(\mathbf{u})<\infty. \tag*{(10)} ν > 0 sup ∫ ∥ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) 2 d P μ ν ( u ) < ∞. (10)
with 2 < r , p < ∞ 2<r,p<\infty 2 < r , p < ∞ , 0 < σ < 1 0<\sigma<1 0 < σ < 1 . This condition is a probabilistic version of the assumption of uniform boundedness of Besov norms for incompressible Navier-Stokes solutions u ν ∈ L 3 ( [ 0 , T ] , B ∞ , ∞ 3 ( T d ) ) \mathbf{u}^\nu\in L^3([0,T],B^3_{\infty,\infty}(\mathbb{T}^d)) u ν ∈ L 3 ([ 0 , T ] , B ∞ , ∞ 3 ( T d )) , invoked in several works [35 , 36 , 37 ] to derive sufficient compactness of the sequence u ν \mathbf{u}^\nu u ν to guarantee strong convergence of suitable subsequences to weak Euler solutions. We are far from deriving such regularity a priori (but see [42 ] ) and, as in those previous works, the primary justification is experimental and numerical observations on turbulent velocity fields. We shall discuss in the following section 2 the available empirical evidence that uniform Besov regularity (10) holds even in the presence of thermal noise.
With this background, we can now state our main result:
Theorem 1. Let P μ ν P_\mu^\nu P μ ν , ν > 0 \nu>0 ν > 0 be the distribution on zero-space-mean Itô strong solutions to (5) for d ≥ 2 d\geq2 d ≥ 2 , supposing that the deterministic, zero-mean forcing has regularity f ∈ C 0 ( [ 0 , T ] × T d ) \mathbf{f}\in C^0([0,T]\times\mathbb{T}^d) f ∈ C 0 ([ 0 , T ] × T d ) and that the initial distribution μ \mu μ is supported on L p ( T d ) L^p(\mathbb{T}^d) L p ( T d ) , p > 2 p>2 p > 2 . We assume that the initial data satisfy ∇ ⋅ u 0 = 0 \nabla\cdot\mathbf{u}_0=0 ∇ ⋅ u 0 = 0 in the distributional sense, for μ \mu μ –a.e. u 0 \mathbf{u}_0 u 0 . We furthermore assume the uniform regularity condition (10) for the same p p p and 2 < r < ∞ 2<r<\infty 2 < r < ∞ , 0 < σ < 1 0<\sigma<1 0 < σ < 1 , with Λ = 2 m ( ν ) ∼ ν − α \Lambda=2^{m(\nu)}\sim\nu^{-\alpha} Λ = 2 m ( ν ) ∼ ν − α for m : ( 0 , ∞ ) → N m:(0,\infty)\to\mathbb{N} m : ( 0 , ∞ ) → N and 3 / 4 ≤ α ≤ 1 3/4\leq\alpha\leq1 3/4 ≤ α ≤ 1 . Then the following statements hold:
(i) The family of measures P μ ν P_\mu^\nu P μ ν , ν > 0 \nu>0 ν > 0 is tight on the Borel σ \sigma σ -algebra B ( Z r , p σ , β ) \mathcal{B}(\mathcal{Z}_{r,p}^{\sigma,\beta}) B ( Z r , p σ , β ) of the Banach space
Z r , p σ , β : = L r ( [ 0 , T ] , L p ( T d ) ) ∩ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ ( T d ) ) , β ≥ 1 + σ + 2 d . (11) \mathcal{Z}_{r,p}^{\sigma,\beta}:=L^r([0,T],L^p(\mathbb{T}^d))\cap C^0([0,T],B^{\sigma}_{p/2,\infty}(\mathbb{T}^d)),\qquad\beta\geq1+\sigma+2d. \tag*{(11)} Z r , p σ , β := L r ([ 0 , T ] , L p ( T d )) ∩ C 0 ([ 0 , T ] , B p /2 , ∞ σ ( T d )) , β ≥ 1 + σ + 2 d . (11) Thus, there exists a subsequence ν k → 0 \nu_k\to0 ν k → 0 such that P μ ν k P_\mu^{\nu_k} P μ ν k converges weakly to some measure P μ P_\mu P μ , i.e.
∫ φ d P μ ν k → ∫ φ d P μ as ν k → 0 , ∀ φ ∈ C b 0 ( Z r , p σ , β ) , \int\varphi\,\mathrm{d}P_\mu^{\nu_k}\to\int\varphi\,\mathrm{d}P_\mu\ \text{as }\nu_k\to0,\qquad\forall\varphi\in C_b^0(\mathcal{Z}_{r,p}^{\sigma,\beta}), ∫ φ d P μ ν k → ∫ φ d P μ as ν k → 0 , ∀ φ ∈ C b 0 ( Z r , p σ , β ) , the latter space consisting of functions φ \varphi φ bounded and continuous in the strong norm topology on Z r , p σ , β \mathcal{Z}_{r,p}^{\sigma,\beta} Z r , p σ , β .
(ii) P μ P_\mu P μ is supported on
Z ~ r , p σ , β : = L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) ∩ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \widetilde{\mathcal{Z}}_{r,p}^{\sigma,\beta}:=L^r([0,T],B^\sigma_{p,\infty}(\mathbb{T}^d))\cap C^0([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)) Z r , p σ , β := L r ([ 0 , T ] , B p , ∞ σ ( T d )) ∩ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) and furthermore the zero-mean condition holds, so that
∫ T d u ( x , t ) d d x = 0 , ∀ t ∈ [ 0 , T ] P μ –a.e. u . \int_{\mathbb{T}^d}\mathbf{u}(\mathbf{x},t)\,\mathrm{d}^d x=0,\qquad\forall t\in[0,T]\quad P_\mu\text{–a.e. }\mathbf{u}. ∫ T d u ( x , t ) d d x = 0 , ∀ t ∈ [ 0 , T ] P μ –a.e. u . (iii) P μ P_\mu P μ is supported on standard weak solutions of the incompressible Euler equations, so that distributionally in space-time
∂ t u = P [ − ∇ ⋅ ( u ⊗ u ) + f ] , ∇ ⋅ u = 0 , P μ –a.e. u \partial_t\mathbf{u}=\mathbb{P}[-\nabla\cdot(\mathbf{u}\otimes\mathbf{u})+\mathbf{f}],\qquad\nabla\cdot\mathbf{u}=0,\qquad P_\mu\text{–a.e. }\mathbf{u} ∂ t u = P [ − ∇ ⋅ ( u ⊗ u ) + f ] , ∇ ⋅ u = 0 , P μ –a.e. u (iv) Because of (iii), the trace γ 0 : u ↦ u ( ⋅ , 0 ) \gamma_0:\mathbf{u}\mapsto\mathbf{u}(\cdot,0) γ 0 : u ↦ u ( ⋅ , 0 ) to initial data is well-defined as a continuous map
γ 0 : Z r , p σ , β → B p / 2 , ∞ σ − β ( T d ) \gamma_0:\mathcal{Z}_{r,p}^{\sigma,\beta}\to B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d) γ 0 : Z r , p σ , β → B p /2 , ∞ σ − β ( T d ) and the pushforward measure P μ ( 0 ) = γ 0 ∗ P μ P_\mu^{(0)}=\gamma_0*P_\mu P μ ( 0 ) = γ 0 ∗ P μ is defined on B ( B p / 2 , ∞ σ − β ( T d ) ) \mathcal{B}(B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)) B ( B p /2 , ∞ σ − β ( T d )) . In fact, P μ ( 0 ) P_\mu^{(0)} P μ ( 0 ) is supported on L p ( T d ) L^p(\mathbb{T}^d) L p ( T d ) and when restricted to B ( L p ( T d ) ) \mathcal{B}(L^p(\mathbb{T}^d)) B ( L p ( T d )) it coincides with μ \mu μ .
We make a sequence of remarks concerning this result:
Remark 1. Important physical consequences of the theorem are obtained for the special case of deterministic (Dirac delta) initial data, μ = δ u 0 \mu=\delta_{\mathbf{u}_0} μ = δ u 0 . In that case, the implication is that any measure P u 0 P_{\mathbf{u}_0} P u 0 obtained as a weak limit for ν → 0 \nu\to0 ν → 0 is supported on weak Euler solutions u \mathbf{u} u , all of which have the same initial datum u ( ⋅ , 0 ) = u 0 \mathbf{u}(\cdot,0)=\mathbf{u}_0 u ( ⋅ , 0 ) = u 0 . If the measure P u 0 P_{\mathbf{u}_0} P u 0 is non-trivial for suitable non-smooth initial data u 0 \mathbf{u}_0 u 0 , as suggested by numerical evidence [25 ] , then such non-triviality requires non-unique solutions of the corresponding Cauchy problem for Euler. This is the essence of the concept of “spontaneous stochasticity”. Our proof can be easily modified for the stochastic shell models employed in the numerical studies [25 ] and, in fact, the additional regularity assumption (10) should be unnecessary in that case [43 ] . We remark that the same implications cannot be drawn using the method of statistical solutions of Foiaş-Proḑ, as in the recent work [38 ] . The statistical Euler solutions constructed in that work are probability measures P μ ( t ) P_{\mu}^{(t)} P μ ( t ) on a function space such as L 2 ( T d ) L^{2}(\mathbb{T}^{d}) L 2 ( T d ) which are parameterized by time t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] and which satisfy, in a weak suitable sense, the forward Kolmogorov or Liouville functional equation for an ensemble of incompressible Euler equations. Such statistical solutions should be obtained from our space-time statistical solutions as pushforward measures P μ ( t ) = γ t ∗ P μ P_{\mu}^{(t)}=\gamma_{t*}P_{\mu} P μ ( t ) = γ t ∗ P μ where γ t : u ↦ u ( ⋅ , t ) \gamma_t:\mathbf{u}\mapsto\mathbf{u}(\cdot,t) γ t : u ↦ u ( ⋅ , t ) is the time-t t t trace. For rigorous results of such type, see [28 ] , Theorem XI.1.1. It is likely that the approach of [38 ] can be extended to the inviscid limit of the Landau-Lifshitz system and spontaneous stochasticity would then manifest as non-trivial measures P u 0 ( t ) P_{u_0}^{(t)} P u 0 ( t ) for t > 0 t>0 t > 0 with appropriate deterministic initial data μ = δ u 0 \mu=\delta_{u_0} μ = δ u 0 . However, this result would not immediately yield the conclusion on non-unique solutions of Euler equations without the knowledge that P μ ( t ) = γ t ∗ P μ P_{\mu}^{(t)}=\gamma_{t*}P_{\mu} P μ ( t ) = γ t ∗ P μ for a space-time statistical solution P μ P_{\mu} P μ of the type provided by our theorem.
Remark 2. An essential feature of spontaneous stochasticity is the universality of the limiting measure P u 0 P_{u_0} P u 0 which should be independent of the particular subsequence ν → 0 \nu\to0 ν → 0 and in fact independent of the particular type of vanishing noise and regularization (which need not arise from thermal fluctuations or molecular viscosity). We obtain existence of a limit measure from tightness alone, so that we can guarantee convergence only along suitable subsequences, whose limits could in principle differ between subsequences. Previous rigorous examples of universal spontaneous statistics in model problems [44 , 45 ] have relied on very strong dynamical mixing properties at all scales, which imply in particular a broad spectrum of positive Lyapunov exponents tending to infinite values as ν → 0 \nu\to0 ν → 0 . Since current mathematical methods for realistic systems like the Landau-Lifshitz equations (5) yield only a positive lower bound for the leading Lyapunov exponent [46 , 47 ] , it does not seem feasible presently to prove the uniqueness or “universality” of the limit measures whose existence we establish. See [48 ] for a numerical RG approach.
Remark 3. We conjecture that the limit probability measures of the theorem are supported not only on weak Euler solutions but in fact on solutions which are locally dissipative. If true, then this would mean that the non-uniqueness of dissipative Euler solutions proved by convex integration methods [10 , 11 , 12 ] is in fact physically realizable. Unlike for deterministic Navier-Stokes solutions [49 ] , it is not so elementary to prove for the Landau-Lifshitz equations that the weak Euler solutions obtained in the infinite-R e Re R e limit are dissipative. The essential difficulty can be understood by considering the local kinetic energy balance for the Landau-Lifshitz equations, which we state here for equation (1) written in dimensional variables:
d ( 1 2 ∣ u ∣ 2 ) + ∇ ⋅ [ ( 1 2 ∣ u ∣ 2 + p ) u − ν Λ ∇ ( 1 2 ∣ u ∣ 2 ) ] d t = − ν Λ ∣ ∇ u ∣ 2 d t + 2 ν Λ k B T ρ u ⋅ ( ∇ ⋅ d ξ ) + 2 ν Λ k B T ρ 1 L 3 ∑ ∣ k ∣ < Λ ∣ k ∣ 2 d t . (12) \mathrm{d}\left(\frac{1}{2}|\mathbf{u}|^2\right)+\nabla\cdot\left[\left(\frac{1}{2}|\mathbf{u}|^2+p\right)\mathbf{u}-\nu_{\Lambda}\nabla\left(\frac{1}{2}|\mathbf{u}|^2\right)\right]\mathrm{d}t
=
-\nu_{\Lambda}|\nabla\mathbf{u}|^2\mathrm{d}t+\sqrt{\frac{2\nu_{\Lambda}k_{\mathrm{B}}T}{\rho}}\mathbf{u}\cdot(\nabla\cdot\mathrm{d}\boldsymbol{\xi})+\frac{2\nu_{\Lambda}k_{\mathrm{B}}T}{\rho}\frac{1}{L^3}\sum_{|\mathbf{k}|<\Lambda}|\mathbf{k}|^2\mathrm{d}t. \tag*{(12)} d ( 2 1 ∣ u ∣ 2 ) + ∇ ⋅ [ ( 2 1 ∣ u ∣ 2 + p ) u − ν Λ ∇ ( 2 1 ∣ u ∣ 2 ) ] d t = − ν Λ ∣∇ u ∣ 2 d t + ρ 2 ν Λ k B T u ⋅ ( ∇ ⋅ d ξ ) + ρ 2 ν Λ k B T L 3 1 ∣ k ∣ < Λ ∑ ∣ k ∣ 2 d t . (12)
Here L L L is the side-length of the periodic flow domain. For the derivation, see e.g. [41 ] . Note that this equation reduces only approximately to the energy balance for deterministic Navier-Stokes, if one chooses the wavenumber cutoff Λ \Lambda Λ in a broad range ≳ 1 / η \gtrsim1/\eta ≳ 1/ η , with η \eta η the Kolmogorov length, and if one considers only typical value of the thermal noise term proportional to ξ \boldsymbol{\xi} ξ . See [50 ] , Chapter II.E, for more details. It seems likely to us that the limit measures of our theorem can be proved to be supported on dissipative Euler solutions by using large deviations estimates similar to those of [51 ] (although note that these authors do not consider the infinite-R e Re R e limit, but instead treat the “hydrodynamic scaling limit” [52 , 53 , 54 ] , which was shown in [55 , 56 ] to be unphysical). It is reasonable to expect, however, that if one considers not the R e → ∞ Re\to\infty R e → ∞ limit itself, but instead the large deviations from that limit, then non-dissipative Euler solutions will be relevant. In fact, the hyperbolic scaling limit of the TASEP (totally asymmetric exclusion process) that leads to the inviscid Burgers equation is quite similar to the infinite Reynolds limit, and there the non-dissipative weak solutions of inviscid Burgers equation govern the large deviations probabilities [57 , 58 , 59 ] .
Remark 4. We conjecture also that the limiting probability measures of our theorem give the path distributions of a Markov process in time. In fact, in all rigorous examples of which we are aware, the limiting spontaneously stochastic process is Markov [60 , 61 , 62 , 44 , 45 ] . Furthermore, the Landau-Lifshitz dynamics (5) for ν > 0 \nu> 0 ν > 0 is proved to be Markov by standard arguments [41 ] and this property is expected to be preserved in the limit ν → 0 \nu\to0 ν → 0 under reasonable assumptions [63 ] .
Remark 5. A comment should be made about the regularity assumptions on our initial data u 0 \mathbf{u}_0 u 0 . As discussed earlier, it is convenient to choose initial data that are non-smooth, so that spontaneous stochasticity appears immediately, rather than after some finite time sufficient, hypothetically, for the solution to blow-up. However, the regularity u 0 ∈ L p ( T d ) \mathbf{u}_0 \in L^p(\mathbb{T}^d) u 0 ∈ L p ( T d ) is much rougher than required, and it would be much more natural physically to take u 0 ∈ B p , ∞ σ ( T d ) \mathbf{u}_0 \in B^{\sigma}_{p,\infty}(\mathbb{T}^d) u 0 ∈ B p , ∞ σ ( T d ) for some p > 2 p > 2 p > 2 and 0 < σ < 1 0 < \sigma< 1 0 < σ < 1 . We did not do so, because our proof of part (iv) of the theorem requires separability of L p ( T d ) L^p(\mathbb{T}^d) L p ( T d ) , whereas B p , ∞ σ ( T d ) B^\sigma_{p,\infty}(\mathbb{T}^d) B p , ∞ σ ( T d ) is non-separable. Two comments are useful. First, there is reason to believe that dissipative weak solutions of the Euler equations will have some spatial Besov regularity immediately at any time t > 0 t > 0 t > 0 , even if the initial data at t = 0 t = 0 t = 0 is rougher [42 ] . Second, we have checked that our theorem remains true if we employ the Besov space B p , q σ ( T d ) B^\sigma_{p,q}(\mathbb{T}^d) B p , q σ ( T d ) for q < ∞ q < \infty q < ∞ rather than for q = ∞ q = \infty q = ∞ , as stated above. Since the Besov spaces are separable for p , q < ∞ p,q < \infty p , q < ∞ , we can in that case choose initial data u 0 ∈ B p , q σ ( T d ) \mathbf{u}_0 \in B^\sigma_{p,q}(\mathbb{T}^d) u 0 ∈ B p , q σ ( T d ) and our current technique of proof applies. Furthermore, part (ii) of the theorem can then be strengthened by replacing E ~ r , p σ , β \widetilde{\mathcal{E}}^{\sigma,\beta}_{r,p} E r , p σ , β instead with
E r , p , q σ , β , γ = L r ( [ 0 , T ] , B p , q σ ( T d ) ) ∩ C 0 , γ ( [ 0 , T ] , B p / 2 , q / 2 σ − β ( T d ) ) \mathcal{E}^{\sigma,\beta,\gamma}_{r,p,q}=L^r\left([0,T],B^\sigma_{p,q}(\mathbb{T}^d)\right)\cap C^{0,\gamma}\left([0,T],B^{\sigma-\beta}_{p/2,q/2}(\mathbb{T}^d)\right) E r , p , q σ , β , γ = L r ( [ 0 , T ] , B p , q σ ( T d ) ) ∩ C 0 , γ ( [ 0 , T ] , B p /2 , q /2 σ − β ( T d ) )
which supplies the realizations of the limit measure with some Hölder continuity in time.
Remark 6. A similar comment is required about the assumed regularity of the force f \mathbf{f} f in our theorem. In fact, for parts (i),(ii) the proofs go through easily assuming only f ∈ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \mathbf{f} \in C^0([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)) f ∈ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) rather than our more stringent assumption f ∈ C 0 ( [ 0 , T ] × T d ) \mathbf{f} \in C^0([0,T]\times\mathbb{T}^d) f ∈ C 0 ([ 0 , T ] × T d ) . It seems possible that, with some further work, we could mathematically strengthen our theorem to require only f ∈ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \mathbf{f} \in C^0([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)) f ∈ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) for all of its statements. However, even the condition f ∈ C 0 ( [ 0 , T ] × T d ) \mathbf{f} \in C^0([0,T]\times\mathbb{T}^d) f ∈ C 0 ([ 0 , T ] × T d ) has dubious physical relevance. To be clear, a physical forcing should represent stirring only at large scales, so that the force should be extremely smooth in space, e.g. f ∈ C 0 ( [ 0 , T ] , S ( T d ) ) \mathbf{f} \in C^0([0,T],\mathcal{S}(\mathbb{T}^d)) f ∈ C 0 ([ 0 , T ] , S ( T d )) . Besov regularity of the type (10) that we have hypothesized is very unlikely to hold when this fluid is forced with increasing strength at decreasing length scales.
Remark 7. Finally, we make a comment about the dependence of our results on the space dimensionality d d d . The non-dimensionalization of the equation (3) assumes for any space dimension d d d the relation ε ∼ U 3 / L \varepsilon\sim U^3/L ε ∼ U 3 / L of Taylor [64 ] , which is an expression of anomalous dissipation and energy cascade. See section 2. In fact, for d > 3 d > 3 d > 3 there are both theoretical arguments and numerical evidence [65 , 66 , 67 , 68 ] that a turbulent energy cascade to small scales exists, just as in d = 3 d = 3 d = 3 . If the Taylor relation is, in fact, modified by an R e Re R e -dependent factor, then the exponent κ = 3 ( d + 2 ) / 8 \kappa= 3(d+2)/8 κ = 3 ( d + 2 ) /8 of the ν \nu ν -dependent prefactor in the thermal noise will be slightly modified but remain positive. However, as is well-known, turbulence phenomenology is quite different for d = 2 d = 2 d = 2 , where there is no energy cascade to small scales but instead an enstrophy cascade [69 ] . In addition, the consequences of thermal noise are much more profound for d = 2 d = 2 d = 2 , where theory predicts that the renormalized viscosity ν Λ \nu_\Lambda ν Λ should depend logarithmically both upon the wavenumber cutoff Λ \Lambda Λ and the flow domain size L L L [20 ] . The logarithmic divergence with L L L has been observed in some lattice-gas simulations of d = 2 d = 2 d = 2 laminar shear flow [70 ] . We conjecture that this logarithmic divergence in a forced turbulent flow in d = 2 d = 2 d = 2 will depend not upon L L L but instead upon the Kraichnan dissipation scale that terminates the enstrophy cascade range. Although we have stated our mathematical theorem for d ≥ 2 d \ge2 d ≥ 2 , many of our assumptions and results must be modified to have physical relevance for d = 2 d = 2 d = 2 .
Theoretical and Empirical Background
In this section, prior to proving our main theorem, we give some necessary background on Landau-Lifschitz fluctuating hydrodynamics, its physical justification, empirical verification and rigorous status. We shall discuss also the non-dimensionalization of the Landau-Lifschitz equation with large-scale flow parameters that underlies our theorem and the empirical basis of the uniform Besov regularity which is its main hypothesis. Readers who are interested mainly in the mathematical results can skip this section on a first reading and go directly to section 3 where the proofs are presented.
Fluctuating hydrodynamics was originally proposed on a phenomenological basis by Landau & Lifschitz [19 ] , who added a fluctuating stress with statistics governed by the fluctuation-dissipation relation (2) to the fluid equations, as the simplest hydrodynamic model to reproduce the known Boltzmann-Einstein single-time statistics in local thermodynamic equilibrium. It was later pointed out in [71 ] that the original Landau-Lifschitz theory reproduced the correct local equilibrium fluctuations only if transport coefficients are state-independent. This observation triggered efforts to derive fluctuating hydrodynamics microscopically by formal methods of statistical physics [72 ] , with the conclusion that the correct noise given by the fluctuation-dissipation relation must be multiplicative [73 ] . See also [74 , 75 ] . We are aware of no mathematically rigorous derivation of the Landau & Lifschitz equations for nonlinear fluctuating hydrodynamics, as necessary to describe turbulent flow. The closest result of which we are aware is a rigorous derivation of a large deviations rate function for a microscopic lattice gas in a “hydrodynamic scaling limit” at small Knudsen and Mach numbers but for arbitrary (fixed) Reynolds number [54 ] . The large-deviations rate function obtained in this work corresponds formally to that obtained from the incompressible Landau-Lifschitz theory in a weak-noise limit [76 , 51 ] . Unfortunately, subsequent work [55 , 56 ] has shown that the assumed “hydrodynamic scaling limit” is physically unrealistic for turbulent flows and probably never applies anywhere in Nature. On the other hand, the Landau-Lifschitz theory predicts non-trivial effects of thermal noise, especially in flow states out of global thermodynamic equilibrium, which are extensively confirmed by laboratory experiments for many laminar flows [77 ] . Since the main physical assumption of the Landau-Lifschitz theory is a local equilibrium property, it is reasonable to expect that fluctuating hydrodynamics applies also to turbulent flows as long as this fundamental property continues to hold there.
A basic conclusion of physicists’ microscopic derivations [72 , 75 ] of nonlinear Landau-Lifschitz equations is that, despite superficial appearances, they are not stochastic partial differential equations. Instead, these equations contain a specific high-wavenumber cutoff Λ \Lambda Λ and are designed to describe only the length scales > Λ − 1 > \Lambda^{-1} > Λ − 1 . In the language of physics, the Landau-Lifschitz hydrodynamics equations are low-wavenumber “effective field theories” [22 , 23 ] . A standard feature of such theories is that the cutoff Λ \Lambda Λ is arbitrary to a large degree but that the parameters of the model depend upon Λ \Lambda Λ or “run” with Λ \Lambda Λ according to renormalization-group flow equations, so that the objective predictions of the model are Λ \Lambda Λ -independent. All of these features are nicely illustrated by the incompressible model (1) . Here we may note that this model follows from the general Landau-Lifschitz theory for a simple Newtonian fluid in the limit of small Mach numbers, as discussed carefully in [21 ] . A renormalization group analysis of Forster, Nelson & Stephen [20 ] studied the Λ \Lambda Λ -dependence of the viscosity ν Λ \nu_\Lambda ν Λ in the model (1) for a fluid in global thermal equilibrium (no mean motion) with the well-known energy spectrum E ( k ) ∝ k B T ρ k 2 E(k) \propto\frac{k_{\mathrm{B}}T}{\rho} k^2 E ( k ) ∝ ρ k B T k 2 of thermal velocity fluctuations at absolute temperature T T T . They found for d = 3 d = 3 d = 3 (3D) that the renormalization of the viscosity by thermal fluctuations is rather weak and of the form ν Λ = ν − O ( Λ k B T / ρ ν ) \nu_\Lambda= \nu- O(\Lambda k_{\mathrm{B}}T/\rho\nu) ν Λ = ν − O ( Λ k B T / ρ ν ) where ν \nu ν is the measured or “macroscopic” viscosity for very small Λ \Lambda Λ . These same results should hold also for 3D turbulent flows if one takes Λ > 1 / η \Lambda> 1/\eta Λ > 1/ η , where η = ν 3 / 4 / ε 1 / 4 \eta= \nu^{3/4}/\varepsilon^{1/4} η = ν 3/4 / ε 1/4 is the Kolmogorov dissipation length defined in terms of the macroscopic viscosity ν \nu ν and the mean rate ε \varepsilon ε of kinetic energy injection by the external force (per unit time and per unit mass). Indeed, numerical simulations of 3D turbulence which include thermal fluctuations show that the energy spectrum E ( k ) ∝ k B T ρ k 2 E(k) \propto\frac{k_{\mathrm{B}}T}{\rho} k^2 E ( k ) ∝ ρ k B T k 2 of these fluctuations become dominant just below the Kolmogorov length [55 , 56 , 78 ] . In that case, any choice of Λ − 1 \Lambda^{-1} Λ − 1 below the Kolmogorov length η \eta η and well above the mean-free-path length λ m f p \lambda_{\mathrm{mfp}} λ mfp of the fluid should suffice to yield (1) as a valid model of turbulent flow, with roughly the same choice of ν Λ \nu_\Lambda ν Λ as in thermal equilibrium. Note, however, that this model might break down locally in spacetime due to extreme turbulent fluctuations at sufficiently high R e Re R e [55 ] . Although the model (1) has solutions which are mathematically well-defined globally in time [41 ] , the solutions might develop extremely energetic small scales ∼ Λ − 1 \sim\Lambda^{-1} ∼ Λ − 1 which invalidate the physical derivations [72 , 75 ] .
An important application of Theorem 1 is to solutions of the Landau-Lifschitz equations (1) that appear as typical realizations selected from a turbulent flow. Therefore its statement is formulated in terms of the equations (3) non-dimensionalized in the standard manner for the inertial scales of high-R e Re R e turbulence. As usual, one introduces dimensionless variables
x ^ = x / L , t ^ = t / ( L / U ) , u ^ = u / U , p ^ = p / U 2 \hat{x} = x/L,\qquad\hat{t} = t/(L/U),\qquad\hat{u} = u/U,\qquad\hat{p} = p/U^2 x ^ = x / L , t ^ = t / ( L / U ) , u ^ = u / U , p ^ = p / U 2
where L L L is the outer or stirring length (defined by the scale of the forcing f \mathbf{f} f ) and U U U is the velocity of the energy-containing eddies (for example, the rms turbulent velocity u r m s u_{\mathrm{rms}} u rms ). After introducing these rescalings and omitting the hats, for convenience, the equation (3) is obtained. In performing the non-dimensionalization, we have furthermore assumed Taylor’s relation ε ∼ U 3 / L \varepsilon\sim U^3/L ε ∼ U 3 / L [64 ] , sometimes called the “zeroth-law of turbulence.” Although there is considerable empirical support for this relation [79 , 80 ] , it is possible that ε ^ = ε / ( U 3 / L ) \hat{\varepsilon} = \varepsilon/(U^3/L) ε ^ = ε / ( U 3 / L ) vanishes slowly with R e Re R e [35 , 81 , 82 , 83 ] and this could change the R e Re R e -dependencies slightly, but without vitiating our analysis. The dimensionless number groups that appear in (3) in addition to R e Re R e are the parameters θ η = k B T / ρ u η 2 η d \theta_\eta= k_{\mathrm{B}}T/\rho u_\eta^2\eta^d θ η = k B T / ρ u η 2 η d (where u η = ( ε ν ) 1 / 4 u_\eta= (\varepsilon\nu)^{1/4} u η = ( ε ν ) 1/4 is the Kolmogorov velocity) and F = L f r m s / U 2 F = Lf_{\mathrm{rms}}/U^2 F = L f rms / U 2 (where f r m s f_{\mathrm{rms}} f rms is the rms amplitude of the external force). The infinite-R e Re R e limit is one with U / L ≫ ν / L 2 U/L \gg\nu/L^2 U / L ≫ ν / L 2 due to large L L L and U U U 1 and fixed fluid parameters such as ρ \rho ρ , T T T , and ν \nu ν , so that θ η \theta_\eta θ η is held fixed. With Taylor’s relation F ∼ f r m s U / ε F \sim f_{\mathrm{rms}}U/\varepsilon F ∼ f rms U / ε measures the “inefficiency” of the force for energy input and will be assumed fixed as well. Finally, Λ \Lambda Λ is kept fixed, but note that Λ ^ → ∞ \hat{\Lambda} \to\infty Λ ^ → ∞ . For example, Λ ^ ∼ R e 3 / 4 \hat{\Lambda} \sim Re^{3/4} Λ ^ ∼ R e 3/4 if η Λ \eta\Lambda η Λ is held fixed, or, using the standard kinetic theory result ν ∼ c s λ m f p \nu\sim c_s\lambda_{\mathrm{mfp}} ν ∼ c s λ mfp , then instead Λ ^ ∼ R e \hat{\Lambda} \sim Re Λ ^ ∼ R e if λ m f p Λ \lambda_{\mathrm{mfp}}\Lambda λ mfp Λ is held fixed.
We consider finally the empirical basis of the a priori Besov regularity that is assumed in Theorem 1. The primary focus of interest is a measure P ∗ \mathbb{P}_* P ∗ on spacetime velocity fields, e.g. a statistically stationary turbulent solution with time-independent single-instant marginals or perhaps a decaying turbulent solution with time-dependent marginals. It was shown already some time ago by a simple Borel-Cantelli argument [84 ] that power-law scaling (or even upper bounds) of ensemble-average absolute structure functions
S p ( r ) : = E ∗ [ ∣ δ u ( r ) ∣ p ] ∼ C p U p ( r L ) ζ p (13) S_p(r) := \mathbb{E}_*\left[|\delta u(r)|^p\right] \sim C_p U^p\left(\frac{r}{L}\right)^{\zeta_p} \tag*{(13)} S p ( r ) := E ∗ [ ∣ δ u ( r ) ∣ p ] ∼ C p U p ( L r ) ζ p (13)
obtained in the limit R e → ∞ Re \to\infty R e → ∞ , together with E ∗ ∣ u ∣ p < ∞ \mathbb{E}_*|u|^p < \infty E ∗ ∣ u ∣ p < ∞ , implies that u ( ⋅ ; ω ) ∈ B p , ∞ σ ( T d ) u(\cdot;\omega) \in B_{p,\infty}^{\sigma}(\mathbb{T}^d) u ( ⋅ ; ω ) ∈ B p , ∞ σ ( T d ) for any choice σ < σ p : = ζ p / p \sigma< \sigma_p := \zeta_p/p σ < σ p := ζ p / p and for P ∗ \mathbb{P}_* P ∗ -a.e. sample point ω \omega ω . In that case, for any ϵ > 0 \epsilon> 0 ϵ > 0
∥ δ u ( r ) ∥ p p : = 1 V ∫ T d ∣ δ u ( x ; r ) ∣ p d d x ≤ C p ( ω ) U p ( r L ) ζ p − p ϵ , P ∗ -a.e. ω (14) \|\delta u(r)\|_p^p := \frac{1}{V}\int_{\mathbb{T}^d}|\delta u(x;r)|^p\,\mathrm{d}^d x \le C_p(\omega)U^p\left(\frac{r}{L}\right)^{\zeta_p-p\epsilon},\qquad\mathbb{P}_*\text{-a.e. }\omega \tag*{(14)} ∥ δ u ( r ) ∥ p p := V 1 ∫ T d ∣ δ u ( x ; r ) ∣ p d d x ≤ C p ( ω ) U p ( L r ) ζ p − p ϵ , P ∗ -a.e. ω (14)
with the optimal constant given by the Besov semi-norm
C p ( ω ) = ∥ u ( ⋅ ; ω ) ∥ B p , ∞ σ p − ϵ p : = sup r < L ∥ δ u ( r ) ∥ p p U p ( r / L ) ζ p − p ϵ , P ∗ -a.e. ω . (15) C_p(\omega) = \|\mathbf{u}(\cdot;\omega)\|_{B_{p,\infty}^{\sigma_p-\epsilon}}^p := \sup_{r<L}\frac{\|\delta u(r)\|_p^p}{U^p(r/L)^{\zeta_p-p\epsilon}},\qquad\mathbb{P}_*\text{-a.e. }\omega. \tag*{(15)} C p ( ω ) = ∥ u ( ⋅ ; ω ) ∥ B p , ∞ σ p − ϵ p := r < L sup U p ( r / L ) ζ p − p ϵ ∥ δ u ( r ) ∥ p p , P ∗ -a.e. ω . (15)
These considerations apply at each time instant, implying that (14) holds for increments δ u ( x , t ; r ) \delta u(\mathbf{x},t;r) δ u ( x , t ; r ) at each time t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] with the constant C p ( ω , t ) = ∥ u ( ⋅ , t ; ω ) ∥ B p , ∞ σ p − ϵ p C_p(\omega,t) = \|\mathbf{u}(\cdot,t;\omega)\|_{B_{p,\infty}^{\sigma_p-\epsilon}}^p C p ( ω , t ) = ∥ u ( ⋅ , t ; ω ) ∥ B p , ∞ σ p − ϵ p . In that case, the reasonable physical assumptions that
1 T ∫ 0 T E ∗ ∣ u ( ⋅ , t ) ∣ p d t < ∞ , 1 T ∫ 0 T E ∗ [ C p ( t ) ] d t < ∞ (16) \frac{1}{T}\int_0^T \mathbb{E}_*|\mathbf{u}( \cdot,t)|^p\,\mathrm{d}t < \infty,\qquad\frac{1}{T}\int_0^T \mathbb{E}_*[C_p(t)]\,\mathrm{d}t < \infty \tag*{(16)} T 1 ∫ 0 T E ∗ ∣ u ( ⋅ , t ) ∣ p d t < ∞ , T 1 ∫ 0 T E ∗ [ C p ( t )] d t < ∞ (16)
are equivalent to the mathematical statement that E ∗ [ ∥ u ∥ L p ( [ 0 , T ] , B p , ∞ σ p − ϵ ) p ] < ∞ \mathbb{E}_*\left[\lVert\mathbf{u}\rVert_{L^p([0,T],B_{p,\infty}^{\sigma_p-\epsilon})}^p\right]<\infty E ∗ [ ∥ u ∥ L p ([ 0 , T ] , B p , ∞ σ p − ϵ ) p ] < ∞ . The above considerations for the ideal limit R e = ∞ Re=\infty R e = ∞ are easily extended to the realistic case of large but finite R e Re R e . Indeed, with E ∗ \mathbb{E}_* E ∗ replaced by E ∗ ν \mathbb{E}_*^\nu E ∗ ν , the relation (13) is observed to hold both in laboratory experiments and in numerical simulations but only over an “inertial range” of length-scales η p < r < L \eta_p<r<L η p < r < L . For example, see [85 , 86 , 87 ] . From simulations, S p ν ( r ) S_p^\nu(r) S p ν ( r ) takes on much smaller values C p ′ ⟨ ∣ ∇ u ∣ p ⟩ r p C'_p\langle\lvert\nabla\mathbf{u}\rvert^p\rangle r^p C p ′ ⟨∣ ∇ u ∣ p ⟩ r p in the “dissipation range” r < η p r<\eta_p r < η p . In that case, (14),(15) still hold for each ν = 1 / R e \nu=1/Re ν = 1/ R e and P ∗ ν \mathbb{P}_*^\nu P ∗ ν -a.e. ω \omega ω , with the supremum in (15) achieved in a range η p ( ω ) < r < L \eta_p(\omega)<r<L η p ( ω ) < r < L . Thus, the conditions (16) are plausibly assumed to hold uniformly in ν = 1 / R e \nu=1/Re ν = 1/ R e so that sup ν > 0 E ∗ ν [ ∥ u ∥ L p ( [ 0 , T ] , B p , ∞ σ p − ϵ ) p ] < ∞ \sup_{\nu>0}\mathbb{E}_*^\nu\left[\lVert\mathbf{u}\rVert_{L^p([0,T],B_{p,\infty}^{\sigma_p-\epsilon})}^p\right]<\infty sup ν > 0 E ∗ ν [ ∥ u ∥ L p ([ 0 , T ] , B p , ∞ σ p − ϵ ) p ] < ∞ , which for p > 2 p>2 p > 2 is stronger than the fundamental assumption 10 of Theorem 1 .
The previous considerations do not explicitly include thermal noise. Physical experiments do include all molecular effects implicitly, of course, but there are no observations at sub-Kolmogorov scales, so that only the inertial-range scaling (13) can be confirmed in a range η p ≪ r ≪ L \eta_p\ll r\ll L η p ≪ r ≪ L [85 , 86 , 87 ] and near-dissipation range drop-off for r ∼ η p r\sim\eta_p r ∼ η p [88 , 89 ] . Since velocity fluctuations from thermal noise will increase with the wavenumber cutoff Λ \Lambda Λ and eventually will even exceed the low-Mach turbulent fluctuations, there is a serious concern that our uniform a priori bounds will be violated. The only empirical source of information about the effects of thermal noise on turbulence come from the few numerical simulations that have so far included them [56 , 78 , 90 ] . Out of this small set of studies, only [90 ] so far has studied the effects of thermal noise on turbulent structure functions, using the Direct Simulation Monte Carlo (DSMC) method for d = 3 d=3 d = 3 . Most useful for us, [90 ] found that their numerical results are well-described by a simple “superposition model”, according to which the velocity coarse-grained over length-scales < Λ − 1 <\Lambda^{-1} < Λ − 1 can be represented as
u ( x , t ) = u N S ( x , t ) + ( F k B T ρ Λ 3 ) 1 / 2 N ( x , t ) (17) \mathbf{u}(\mathbf{x},t)=\mathbf{u}^{NS}(\mathbf{x},t)+\left(\frac{Fk_B T}{\rho}\Lambda^3\right)^{1/2}\mathbf{N}(\mathbf{x},t) \tag*{(17)} u ( x , t ) = u N S ( x , t ) + ( ρ F k B T Λ 3 ) 1/2 N ( x , t ) (17)
where u N S ( x , t ) \mathbf{u}^{NS}(\mathbf{x},t) u N S ( x , t ) is the deterministic Navier-Stokes solution and N ( x , t ) \mathbf{N}(\mathbf{x},t) N ( x , t ) is a standard normal random variable. The latter is chosen independently for every discrete value of x \mathbf{x} x and t t t in the DSMC simulation and independent of the Navier-Stokes solution. In the DSMC method the quantity F F F is the “simulation ratio,” a large factor which prescribes the number of true molecules represented by each DSMC computational molecule, but in the physically realistic limit F F F is a number of order unity. This model yields the following predictions for the first few even-order velocity structure functions at separations Δ r > 1 \Delta r>1 Δ r > 1 :
S 2 ( r ) = S 2 N S ( r ) + 2 F k B T ρ Λ 3 (18) S_2(r)=S_2^{NS}(r)+2\frac{Fk_B T}{\rho}\Lambda^3 \tag*{(18)} S 2 ( r ) = S 2 N S ( r ) + 2 ρ F k B T Λ 3 (18)
S 4 ( r ) = S 4 N S ( r ) + 12 ( F k B T ρ Λ 3 ) S 2 N S ( r ) + 12 ( F k B T ρ Λ 3 ) 2 (19) S_4(r)=S_4^{NS}(r)+12\left(\frac{Fk_B T}{\rho}\Lambda^3\right)S_2^{NS}(r)+12\left(\frac{Fk_B T}{\rho}\Lambda^3\right)^2 \tag*{(19)} S 4 ( r ) = S 4 N S ( r ) + 12 ( ρ F k B T Λ 3 ) S 2 N S ( r ) + 12 ( ρ F k B T Λ 3 ) 2 (19)
S 6 ( r ) = S 6 N S ( r ) + 30 ( F k B T ρ Λ 3 ) S 4 N S ( r ) + 180 ( F k B T ρ Λ 3 ) 2 S 2 N S ( r ) + 120 ( F k B T ρ Λ 3 ) 3 (20) \begin{aligned}
S_6(r)={}&S_6^{NS}(r)+30\left(\frac{Fk_B T}{\rho}\Lambda^3\right)S_4^{NS}(r)\\
&+180\left(\frac{Fk_B T}{\rho}\Lambda^3\right)^2S_2^{NS}(r)+120\left(\frac{Fk_B T}{\rho}\Lambda^3\right)^3 \tag*{(20)}
\end{aligned} S 6 ( r ) = S 6 N S ( r ) + 30 ( ρ F k B T Λ 3 ) S 4 N S ( r ) + 180 ( ρ F k B T Λ 3 ) 2 S 2 N S ( r ) + 120 ( ρ F k B T Λ 3 ) 3 (20)
See [90 ] , Eq. (6) . Note that the superposition model becomes open to question if one considers structure functions of higher order p ≫ 1 p\gg1 p ≫ 1 , since these quantities are dominated by very rare, extreme turbulent fluctuations which penetrate to very tiny length scales even smaller than Λ − 1 \Lambda^{-1} Λ − 1 and thus disrupt the assumption of local equilibrium at scale Λ \Lambda Λ [55 ] . We employ the model here as the best available representation of the effects of thermal noise on structure functions of low orders, conveniently summarizing the numerical observations in [90 ] and also the related observations for a shell model in [55 ] . Finally, we note that for distances Δ r < 1 \Delta r<1 Δ r < 1 the structure functions decrease rapidly below the above values, with S p ( r ) = O ( ∣ Δ r ∣ p ) S_p(r)=O(\lvert\Delta r\rvert^p) S p ( r ) = O (∣ Δ r ∣ p ) , since the velocity field within the “effective field theory” is band-limited to wavenumbers < Λ <\Lambda < Λ .
To carry over our previous argument about Besov regularity of velocity realizations, we must show that the scaling relation (13) survives for all r < L r<L r < L in the presence of thermal fluctuations, at least in the modest sense of an upper bound. Thus, essentially, we must establish finiteness of the following constant:
C p : = sup r < L S p ( r ) U p ( r / L ) ζ p < ∞ . (21) C_p := \sup_{r<L} \frac{S_p(r)}{U^p(r/L)^{\zeta_p}} < \infty. \tag*{(21)} C p := r < L sup U p ( r / L ) ζ p S p ( r ) < ∞. (21)
We naturally assume finiteness of the corresponding constant for the deterministic Navier–Stokes solution: C p N S = sup r < L S p N S ( r ) U p ( r / L ) ζ p < ∞ C_p^{NS}=\sup_{r<L}\frac{S_p^{NS}(r)}{U^p(r/L)^{\zeta_p}}<\infty C p N S = sup r < L U p ( r / L ) ζ p S p N S ( r ) < ∞ . To estimate the thermal contribution we must consider the key quantity
C p t h ( R e ) : = sup Λ − 1 < r < L ( F k B T ρ Λ 3 ) p / 2 1 U p ( L r ) ζ p = ( F k B T ρ U 2 Λ 3 ) p / 2 ( Λ L ) ζ p = Θ η p / 2 ( η L ) p 3 − ζ p ( η Λ ) 3 p 2 + ζ p . (22) \begin{aligned}
C_p^{th}(Re) &:= \sup_{\Lambda^{-1}<r<L}\left(\frac{Fk_BT}{\rho}\Lambda^3\right)^{p/2}\frac{1}{U^p}\left(\frac{L}{r}\right)^{\zeta_p}\\
&= \left(\frac{Fk_BT}{\rho U^2}\Lambda^3\right)^{p/2}(\Lambda L)^{\zeta_p}
= \Theta_\eta^{p/2}\left(\frac{\eta}{L}\right)^{\frac{p}{3}-\zeta_p}(\eta\Lambda)^{\frac{3p}{2}+\zeta_p}. \tag*{(22)}
\end{aligned} C p t h ( R e ) := Λ − 1 < r < L sup ( ρ F k B T Λ 3 ) p /2 U p 1 ( r L ) ζ p = ( ρ U 2 F k B T Λ 3 ) p /2 ( Λ L ) ζ p = Θ η p /2 ( L η ) 3 p − ζ p ( η Λ ) 2 3 p + ζ p . (22)
Notice, because of the rapid decay of S p ( r ) S_p(r) S p ( r ) for Λ r < 1 \Lambda r<1 Λ r < 1 , that the supremum over all r < L r<L r < L is in fact achieved for Λ r > 1 \Lambda r>1 Λ r > 1 and this leads to the restriction of the range in (22). We have also defined Θ η : = F η \Theta_\eta:=F\eta Θ η := F η . In terms of the two constants C p N S C_p^{NS} C p N S and C p t h C_p^{th} C p t h we immediately obtain from (18)
C 2 = sup Λ − 1 < r < L S 2 ( r ) U 2 ( r / L ) ζ 2 ≤ C 2 N S + 2 C 2 t h , (23) C_2=\sup_{\Lambda^{-1}<r<L}\frac{S_2(r)}{U^2(r/L)^{\zeta_2}}\leq C_2^{NS}+2C_2^{th}, \tag*{(23)} C 2 = Λ − 1 < r < L sup U 2 ( r / L ) ζ 2 S 2 ( r ) ≤ C 2 N S + 2 C 2 t h , (23)
from (19) using the Hölder inequality S 2 N S ( r ) ≤ [ S 4 N S ( r ) ] 1 / 2 S_2^{NS}(r)\leq[S_4^{NS}(r)]^{1/2} S 2 N S ( r ) ≤ [ S 4 N S ( r ) ] 1/2
C 4 = sup Λ − 1 < r < L S 4 ( r ) U 4 ( r / L ) ζ 4 ≤ C 4 N S + 12 ( C 4 t h ) 1 / 2 ( C 4 N S ) 1 / 2 + 12 C 4 t h (24) C_4=\sup_{\Lambda^{-1}<r<L}\frac{S_4(r)}{U^4(r/L)^{\zeta_4}}\leq C_4^{NS}+12(C_4^{th})^{1/2}(C_4^{NS})^{1/2}+12C_4^{th} \tag*{(24)} C 4 = Λ − 1 < r < L sup U 4 ( r / L ) ζ 4 S 4 ( r ) ≤ C 4 N S + 12 ( C 4 t h ) 1/2 ( C 4 N S ) 1/2 + 12 C 4 t h (24)
from (20) using Hölder inequalities S 4 N S ( r ) ≤ [ S 6 N S ( r ) ] 2 / 3 S_4^{NS}(r)\leq[S_6^{NS}(r)]^{2/3} S 4 N S ( r ) ≤ [ S 6 N S ( r ) ] 2/3 , S 2 N S ( r ) ≤ [ S 6 N S ( r ) ] 1 / 3 S_2^{NS}(r)\leq[S_6^{NS}(r)]^{1/3} S 2 N S ( r ) ≤ [ S 6 N S ( r ) ] 1/3 ,
sup Λ − 1 < r < L S 6 ( r ) U 6 ( r / L ) ζ 6 ≤ C 6 N S + 30 ( C 6 t h ) 1 / 3 ( C 6 N S ) 2 / 3 + 180 ( C 6 t h ) 2 / 3 ( C 6 N S ) 1 / 3 + 120 C 6 t h , (25) \sup_{\Lambda^{-1}<r<L}\frac{S_6(r)}{U^6(r/L)^{\zeta_6}}\leq C_6^{NS}+30(C_6^{th})^{1/3}(C_6^{NS})^{2/3}+180(C_6^{th})^{2/3}(C_6^{NS})^{1/3}+120C_6^{th}, \tag*{(25)} Λ − 1 < r < L sup U 6 ( r / L ) ζ 6 S 6 ( r ) ≤ C 6 N S + 30 ( C 6 t h ) 1/3 ( C 6 N S ) 2/3 + 180 ( C 6 t h ) 2/3 ( C 6 N S ) 1/3 + 120 C 6 t h , (25)
and so forth. Clearly, the behavior of C p t h ( R e ) C_p^{th}(Re) C p t h ( R e ) for R e → ∞ Re\to\infty R e → ∞ is crucial.
To complete the argument we must appeal to some properties of the turbulent scaling exponents ζ p \zeta_p ζ p , which are known to be concave in p p p and which coincide with the K41 line p / 3 p/3 p /3 at exactly two p p p -values, p = 0 p=0 p = 0 and p = p ∗ ≐ 3 p=p_*\doteq3 p = p ∗ ≐ 3 [91 ] . Note that, because we deal here with structure functions of absolute velocity increments, the Kolmogorov 4/5th law does not apply, so that it is not true that p ∗ = 3 p_*=3 p ∗ = 3 exactly. The important consequence is that ζ p < p / 3 \zeta_p<p/3 ζ p < p /3 for p > p ∗ p>p_* p > p ∗ and we can thus infer from the bound (22) that
lim R e → ∞ C p t h ( R e ) = 0 , p > p ∗ , \lim_{Re\to\infty} C_p^{th}(Re)=0,\qquad p>p_*, R e → ∞ lim C p t h ( R e ) = 0 , p > p ∗ ,
if one assumes that η Λ \eta\Lambda η Λ is held fixed. In conjunction with the estimates (24) for C 4 C_4 C 4 , (25) for C 6 C_6 C 6 , etc. we see that all of these constants remain bounded in the presence of thermal noise and, indeed, coincide for R e ≫ 1 Re\gg1 R e ≫ 1 with the Navier–Stokes values C 4 N S , C 6 N S C_4^{NS},C_6^{NS} C 4 N S , C 6 N S , etc. The Besov regularity of the velocity realizations is thus unchanged from the deterministic case when the cutoff Λ \Lambda Λ is taken to be some fixed multiple of η − 1 \eta^{-1} η − 1 . In fact, one can also take η Λ \eta\Lambda η Λ increasing with R e Re R e at a sufficiently slow rate, since the only requirement is that C p t h ( R e ) C_p^{th}(Re) C p t h ( R e ) remain bounded as R e → ∞ Re\to\infty R e → ∞ . In that case, (22) implies that the choice
η Λ ∝ ( L η ) p 3 − ζ p 3 p 2 + ζ p ∼ R e p − 3 ζ p 6 p + 4 ζ p (26) \eta\Lambda\propto\left(\frac{L}{\eta}\right)^{\frac{\frac{p}{3}-\zeta_p}{\frac{3p}{2}+\zeta_p}}\sim Re^{\frac{p-3\zeta_p}{6p+4\zeta_p}} \tag*{(26)} η Λ ∝ ( η L ) 2 3 p + ζ p 3 p − ζ p ∼ R e 6 p + 4 ζ p p − 3 ζ p (26)
would suffice to guarantee that C p ( R e ) C_p(Re) C p ( R e ) remain bounded as R e → ∞ Re\to\infty R e → ∞ , although with a different limiting value of the constant than for the deterministic case. With either choice of Λ \Lambda Λ one can conclude that
sup ν > 0 E ∗ ν [ ∥ u ∥ L p ( [ 0 , T ] , B p , ∞ σ p − ϵ ) p ] < ∞ , p = 4 , 6 , 8 , … \sup_{\nu>0}\mathbb{E}_*^\nu\left[\|\mathbf{u}\|_{L^p([0,T],B_{p,\infty}^{\sigma_p-\epsilon})}^p\right]<\infty,\qquad p=4,6,8,\ldots ν > 0 sup E ∗ ν [ ∥ u ∥ L p ([ 0 , T ] , B p , ∞ σ p − ϵ ) p ] < ∞ , p = 4 , 6 , 8 , …
at least for even p p p that are not too large. On the other hand, for p < p ∗ p < p_* p < p ∗ instead ζ p > p / 3 \zeta_p > p/3 ζ p > p /3 , and thus lim R e → ∞ C p t h ( R e ) = ∞ \lim_{Re\to\infty} C_p^{\mathrm{th}}(Re)=\infty lim R e → ∞ C p th ( R e ) = ∞ ! In that case, one can instead replace ζ p \zeta_p ζ p with p / 3 p/3 p /3 to obtain
C ~ p t h ( R e ) : = sup Λ − 1 < r < L ( F k k B T ρ Λ 3 ) p / 2 1 U p ( L r ) p / 3 = Θ η p / 2 ( η Λ ) 11 p / 6 < ∞ , \widetilde{C}_p^{\mathrm{th}}(Re) := \sup_{\Lambda^{-1}<r<L}\left(\frac{F_k k_B T}{\rho}\Lambda^3\right)^{p/2}\frac{1}{U^p}\left(\frac{L}{r}\right)^{p/3} = \Theta_{\eta}^{p/2}(\eta\Lambda)^{11p/6}<\infty, C p th ( R e ) := Λ − 1 < r < L sup ( ρ F k k B T Λ 3 ) p /2 U p 1 ( r L ) p /3 = Θ η p /2 ( η Λ ) 11 p /6 < ∞ ,
whereas also C ~ p N S = sup r < L S p N S ( r ) U p ( r / L ) p / 3 < ∞ \widetilde{C}_p^{NS}=\sup_{r<L}\frac{S_p^{NS}(r)}{U_p(r/L)^{p/3}}<\infty C p N S = sup r < L U p ( r / L ) p /3 S p N S ( r ) < ∞ . Thus, choosing the cutoff Λ \Lambda Λ once again to be some fixed multiple of η − 1 \eta^{-1} η − 1 , one can obtain at least K41 regularity of the velocity field for p < p ∗ p<p_* p < p ∗ . In particular, using (23) for p = 2 p=2 p = 2 it is justified by the empirical observations to assume that
sup ν > 0 E μ ∗ ν [ ∥ u ∥ L 2 ( [ 0 , T ] , B 2 , ∞ 1 3 − ϵ ) 2 ] < ∞ . \sup_{\nu>0}\mathbb{E}_{\mu_*}^{\nu}\left[\lVert\mathbf{u}\rVert_{L^2([0,T],B_{2,\infty}^{\frac{1}{3}-\epsilon})}^2\right]<\infty. ν > 0 sup E μ ∗ ν [ ∥ u ∥ L 2 ([ 0 , T ] , B 2 , ∞ 3 1 − ϵ ) 2 ] < ∞.
Proof of the Main Theorem
We follow the general strategy of [28 ] , Chapter IV, who proved the existence of space-time statistical solutions of the incompressible Navier-Stokes equations, thereby obtaining a probabilistic version of Leray’s construction. We give the proofs of parts (i)-(iv) of our main theorem in successive subsections. We use frequently the convenient notation a ≲ b a\lesssim b a ≲ b with the meaning that a ≤ C b a\leq Cb a ≤ C b for some constant C C C depending possibly upon parameters that are fixed.
Proof of (i)
Proof of (i): Our proof of tightness is modeled after Theorem IV.4.3 of [28 ] and is likewise based upon their following general result:
Lemma II.3.1 of [28 ] : Let Ω 1 , Ω 2 \Omega_1,\Omega_2 Ω 1 , Ω 2 be Banach spaces, where Ω 2 \Omega_2 Ω 2 is separable and Ω 1 \Omega_1 Ω 1 is compactly embedded in Ω 2 \Omega_2 Ω 2 . The family M \mathcal{M} M of probability measures defined on B ( Ω 2 ) \mathcal{B}(\Omega_2) B ( Ω 2 ) and with support on Ω 1 \Omega_1 Ω 1 is tight on B ( Ω 2 ) \mathcal{B}(\Omega_2) B ( Ω 2 ) if for any μ ∈ M \mu\in\mathcal{M} μ ∈ M , the functional u ↦ ∥ u ∥ Ω 1 u\mapsto\lVert u\rVert_{\Omega_1} u ↦ ∥ u ∥ Ω 1 is μ \mu μ -integrable and
sup μ ∈ M ∫ ∥ u ∥ Ω 1 d μ ( u ) < ∞ (27) \sup_{\mu\in\mathcal{M}}\int\lVert u\rVert_{\Omega_1}\,d\mu(u)<\infty \tag*{(27)} μ ∈ M sup ∫ ∥ u ∥ Ω 1 d μ ( u ) < ∞ (27)
We apply this proposition with Ω 2 = Z r , p σ , β \Omega_2=\mathcal{Z}_{r,p}^{\sigma,\beta} Ω 2 = Z r , p σ , β and with Banach space Ω 1 = E r , p σ , β , γ \Omega_1=\mathcal{E}_{r,p}^{\sigma,\beta,\gamma} Ω 1 = E r , p σ , β , γ defined for 0 < γ < 1 0<\gamma<1 0 < γ < 1 by
E r , p σ , β , γ : = L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) ∩ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \mathcal{E}_{r,p}^{\sigma,\beta,\gamma}:=L^r([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^d))\cap C^{0,\gamma}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)) E r , p σ , β , γ := L r ([ 0 , T ] , B p , ∞ σ ( T d )) ∩ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d ))
where
∥ ⋅ ∥ E r , p σ , β , γ : = ∥ ⋅ ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) + ∥ ⋅ ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) . \lVert\cdot\rVert_{\mathcal{E}_{r,p}^{\sigma,\beta,\gamma}}:=\lVert\cdot\rVert_{L^r([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^d))}+\lVert\cdot\rVert_{C^{0,\gamma}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d))}. ∥ ⋅ ∥ E r , p σ , β , γ := ∥ ⋅ ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) + ∥ ⋅ ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) .
First, we note that Z r , p σ , β ⊆ L r ( [ 0 , T ] , L p ( T d ) ) \mathcal{Z}_{r,p}^{\sigma,\beta}\subseteq L^r([0,T],L^p(\mathbb{T}^d)) Z r , p σ , β ⊆ L r ([ 0 , T ] , L p ( T d )) , and the latter is a separable Banach space. By Proposition 3.25 of [92 ] , we conclude Z r , p σ , β = L r ( [ 0 , T ] , L p ( T d ) ) ∩ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \mathcal{Z}_{r,p}^{\sigma,\beta}=L^r([0,T],L^p(\mathbb{T}^d))\cap C^0([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)) Z r , p σ , β = L r ([ 0 , T ] , L p ( T d )) ∩ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) is separable. Next we argue that
E r , p σ , β , γ ↪ Z r , p σ , β \mathcal{E}_{r,p}^{\sigma,\beta,\gamma}\hookrightarrow\mathcal{Z}_{r,p}^{\sigma,\beta} E r , p σ , β , γ ↪ Z r , p σ , β
is a compact embedding for 0 < γ < min ( ( r − 2 ) / r , 1 / 2 ) 0<\gamma<\min((r-2)/r,1/2) 0 < γ < min (( r − 2 ) / r , 1/2 ) . We use the following version of the Aubin-Lions-Simons compactness result:
Theorem IV.4.1 of [28 ] : Let E 0 , E , E 1 E_0,E,E_1 E 0 , E , E 1 be Banach spaces, such that E E E is continuously embedded in E 1 E_1 E 1 and E 0 E_0 E 0 is compactly embedded in E E E . Let 1 < q < ∞ 1<q<\infty 1 < q < ∞ and M M M be a bounded set in L q ( [ 0 , T ] , E 0 ) L^q([0,T],E_0) L q ([ 0 , T ] , E 0 ) consisting of functions equicontinuous in C 0 ( [ 0 , T ] , E 1 ) C^0([0,T],E_1) C 0 ([ 0 , T ] , E 1 ) . Then M M M is relatively compact in L q ( [ 0 , T ] , E ) ∩ C 0 ( [ 0 , T ] , E 1 ) L^q([0,T],E)\cap C^0([0,T],E_1) L q ([ 0 , T ] , E ) ∩ C 0 ([ 0 , T ] , E 1 ) .
We note that the original statement of this result assumed that all of the Banach spaces E 0 E_0 E 0 , E E E , E 1 E_1 E 1 are reflexive. However, we have checked that the proof of Theorem IV.4.1 in [28 ] used reflexivity only in their Lemma IV.4.1 and this result is now well-known not to require reflexivity. For example, see Lemma II.5.15 in section 5.3 of [93 ] . We thus can apply the above proposition with q = r q=r q = r and
E 0 = B p , ∞ σ ( T d ) , E = L p ( T d ) , E 1 = B p / 2 , ∞ σ − β ( T d ) E_0 = B^\sigma_{p,\infty}(\mathbb{T}^d), \qquad E = L^p(\mathbb{T}^d), \qquad E_1 = B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d) E 0 = B p , ∞ σ ( T d ) , E = L p ( T d ) , E 1 = B p /2 , ∞ σ − β ( T d )
even though B p , ∞ σ ( T d ) B^\sigma_{p,\infty}(\mathbb{T}^d) B p , ∞ σ ( T d ) and B p / 2 , ∞ σ − β ( T d ) B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d) B p /2 , ∞ σ − β ( T d ) are not reflexive. The hypotheses of Theorem IV.4.1 in [28 ] are verified in the following lemmas.
First the compact embedding of E 0 E_0 E 0 into E E E is given by
Lemma 1. For any 1 ≤ p < ∞ 1 \le p < \infty 1 ≤ p < ∞ , 0 < σ < 1 0 < \sigma< 1 0 < σ < 1 , and d ∈ N d \in\mathbb{N} d ∈ N the embedding B p , ∞ σ ( T d ) ↪ L p ( T d ) B^\sigma_{p,\infty}(\mathbb{T}^d) \hookrightarrow L^p(\mathbb{T}^d) B p , ∞ σ ( T d ) ↪ L p ( T d ) is compact.
Proof. Following the Kolmogorov–Riesz–Fréchet theorem (Theorem 4.26 of [92 ] ) it suffices to argue that for any bounded subset Ω ⊆ B p , ∞ σ ( T d ) \Omega\subseteq B^\sigma_{p,\infty}(\mathbb{T}^d) Ω ⊆ B p , ∞ σ ( T d ) ,
lim ∣ h ∣ → 0 sup v ∈ Ω ∥ v ( ⋅ + h ) − v ( ⋅ ) ∥ L p ( T d ) = 0. \lim_{|h|\to0}\sup_{\mathbf{v}\in\Omega}\|\mathbf{v}(\cdot+h)-\mathbf{v}(\cdot)\|_{L^p(\mathbb{T}^d)}=0. ∣ h ∣ → 0 lim v ∈ Ω sup ∥ v ( ⋅ + h ) − v ( ⋅ ) ∥ L p ( T d ) = 0.
However, by definition of B p , ∞ σ ( T d ) B^\sigma_{p,\infty}(\mathbb{T}^d) B p , ∞ σ ( T d ) ,
sup v ∈ Ω ∥ v ( ⋅ + h ) − v ( ⋅ ) ∥ L p ( T d ) ≤ ∣ h ∣ σ sup v ∈ Ω ∥ v ∥ B p , ∞ σ ( T d ) → 0 as h → 0. \sup_{\mathbf{v}\in\Omega}\|\mathbf{v}(\cdot+h)-\mathbf{v}(\cdot)\|_{L^p(\mathbb{T}^d)}\le|h|^\sigma\sup_{\mathbf{v}\in\Omega}\|\mathbf{v}\|_{B^\sigma_{p,\infty}(\mathbb{T}^d)}\to0 \text{ as } h\to0. v ∈ Ω sup ∥ v ( ⋅ + h ) − v ( ⋅ ) ∥ L p ( T d ) ≤ ∣ h ∣ σ v ∈ Ω sup ∥ v ∥ B p , ∞ σ ( T d ) → 0 as h → 0.
Next, the continuous embedding of E E E into E 1 E_1 E 1 follows from
Lemma 2. For any 1 ≤ p ′ ≤ p < ∞ 1 \le p' \le p < \infty 1 ≤ p ′ ≤ p < ∞ , 0 < σ < 1 0 < \sigma< 1 0 < σ < 1 , β ≥ 2 \beta\ge2 β ≥ 2 and d ∈ N + d \in\mathbb{N}^+ d ∈ N + the embedding L p ( T d ) ↪ B p ′ , ∞ σ − β ( T d ) L^p(\mathbb{T}^d) \hookrightarrow B^{\sigma-\beta}_{p',\infty}(\mathbb{T}^d) L p ( T d ) ↪ B p ′ , ∞ σ − β ( T d ) is continuous.
Proof. From the famous Littlewood–Paley theorem (Remark 2 of 3.5.4 in [94 ] ), we have equivalence of L p ( T d ) L^p(\mathbb{T}^d) L p ( T d ) with F p , 2 0 ( T d ) F^0_{p,2}(\mathbb{T}^d) F p , 2 0 ( T d ) , the Triebel–Lizorkin space. Then we have a chain of continuous embeddings (Remark 4 of 3.5.1 in [94 ] )
F p , 2 0 ( T d ) ↪ B p ′ , 2 0 ( T d ) ↪ B p ′ , 2 σ − β ( T d ) ↪ B p ′ , ∞ σ − β ( T d ) . F^0_{p,2}(\mathbb{T}^d)\hookrightarrow B^0_{p',2}(\mathbb{T}^d)\hookrightarrow B^{\sigma-\beta}_{p',2}(\mathbb{T}^d)\hookrightarrow B^{\sigma-\beta}_{p',\infty}(\mathbb{T}^d). F p , 2 0 ( T d ) ↪ B p ′ , 2 0 ( T d ) ↪ B p ′ , 2 σ − β ( T d ) ↪ B p ′ , ∞ σ − β ( T d ) .
Finally, boundedness in C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) C^{0,\gamma}([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)) C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) immediately implies equicontinuity from [ 0 , T ] [0,T] [ 0 , T ] into B p / 2 , ∞ σ − β ( T d ) B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d) B p /2 , ∞ σ − β ( T d ) . Thus, any bounded subset of E r , p σ , β , γ \mathcal{E}^{\sigma,\beta,\gamma}_{r,p} E r , p σ , β , γ is relatively compact in Z r , p σ , β \mathcal{Z}^{\sigma,\beta}_{r,p} Z r , p σ , β by Theorem IV.4.1 of [28 ] , so that the embedding of the former into the latter is compact.
Finally, to obtain tightness from Lemma II.3.1 of [28 ] , we need the basic bound (27), or
sup ν > 0 ∫ ∥ u ∥ E r , p σ , β , γ D μ ν ( u ) < ∞ . (28) \sup_{\nu>0}\int\|\mathbf{u}\|_{\mathcal{E}^{\sigma,\beta,\gamma}_{r,p}}\,D\mu^\nu(\mathbf{u})<\infty. \tag*{(28)} ν > 0 sup ∫ ∥ u ∥ E r , p σ , β , γ D μ ν ( u ) < ∞. (28)
Such a uniform estimate with the norm ∥ ⋅ ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) \|\cdot\|_{L^r([0,T],B^\sigma_{p,\infty}(\mathbb{T}^d))} ∥ ⋅ ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) follows immediately from our main physical hypothesis (10). We spend the remainder of this subsection deriving the similar bound for the norm ∥ ⋅ ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \|\cdot\|_{C^{0,\gamma}([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d))} ∥ ⋅ ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) of the other component space in the definition of E r , p σ , β , γ \mathcal{E}^{\sigma,\beta,\gamma}_{r,p} E r , p σ , β , γ . Toward this end, we define the quantity z ν : = u ν − ν κ P ∇ ⋅ π Λ ξ \mathbf{z}^\nu:= \mathbf{u}^\nu-\nu^\kappa\mathcal{P}\nabla\cdot\pi_\Lambda\mathcal{\xi} z ν := u ν − ν κ P ∇ ⋅ π Λ ξ , in terms of which we can decompose the solutions of (5) as
u ν = z ν + ν κ P ∇ ⋅ π Λ ξ . \mathbf{u}^\nu=\mathbf{z}^\nu+\nu^\kappa\mathcal{P}\nabla\cdot\pi_\Lambda\mathcal{\xi}. u ν = z ν + ν κ P ∇ ⋅ π Λ ξ .
We obtain the required uniform estimate first for the term z ν \mathbf{z}^\nu z ν and then for the noise term ν κ P ∇ ⋅ π Λ ξ \nu^\kappa\mathcal{P}\nabla\cdot\pi_\Lambda\mathcal{\xi} ν κ P ∇ ⋅ π Λ ξ .
Estimate for z ν \mathbf{z}^\nu z ν : It follows directly from the definition of z ν \mathbf{z}^\nu z ν that it satisfies the equation
∂ t z ν = P π Λ [ − ∇ ⋅ ( u ν ⊗ u ν ) + f ] + ν Δ π Λ u ν . (29) \partial_t\mathbf{z}^\nu=\mathcal{P}_{\pi_\Lambda}\left[-\nabla\cdot(\mathbf{u}^\nu\otimes\mathbf{u}^\nu)+\mathbf{f}\right]+\nu\Delta_{\pi_\Lambda}\mathbf{u}^\nu. \tag*{(29)} ∂ t z ν = P π Λ [ − ∇ ⋅ ( u ν ⊗ u ν ) + f ] + ν Δ π Λ u ν . (29)
We shall use the integral form of this equation
z ν ( ⋅ , t ) = z ν ( ⋅ , 0 ) + ∫ 0 t ∂ t z ν ( ⋅ , s ) d s = u ν ( ⋅ , 0 ) + ∫ 0 t ∂ t z ν ( ⋅ , s ) d s . (30) z^\nu(\cdot,t)=z^\nu(\cdot,0)+\int_0^t\partial_tz^\nu(\cdot,s)\,\mathrm{d}s=u^\nu(\cdot,0)+\int_0^t\partial_tz^\nu(\cdot,s)\,\mathrm{d}s. \tag*{(30)} z ν ( ⋅ , t ) = z ν ( ⋅ , 0 ) + ∫ 0 t ∂ t z ν ( ⋅ , s ) d s = u ν ( ⋅ , 0 ) + ∫ 0 t ∂ t z ν ( ⋅ , s ) d s . (30)
to obtain our estimates. We proceed by a series of lemmas to control each term in (29).
Lemma 3. Take 2 ≤ s ≤ ∞ 2\le s\le\infty 2 ≤ s ≤ ∞ , 2 ≤ p < ∞ 2\le p<\infty 2 ≤ p < ∞ , 0 < σ < 1 0<\sigma<1 0 < σ < 1 , β ≥ 1 \beta\ge1 β ≥ 1 , and v ∈ L s ( [ 0 , T ] , B p , ∞ σ ( T d ) ) \mathbf{v}\in L^s([0,T],B^\sigma_{p,\infty}(\mathbb{T}^d)) v ∈ L s ([ 0 , T ] , B p , ∞ σ ( T d )) . Let
∀ t ∈ [ 0 , T ] : ∫ T d v ( x , t ) d x = 0. \forall t\in[0,T]:\qquad\int_{\mathbb{T}^d}\mathbf{v}(\mathbf{x},t)\,\mathrm{d}\mathbf{x}=0. ∀ t ∈ [ 0 , T ] : ∫ T d v ( x , t ) d x = 0. Then P ∇ ⋅ ( v ⊗ v ) ∈ L r ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \mathbb{P}\nabla\cdot(\mathbf{v}\otimes\mathbf{v})\in L^r([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)) P ∇ ⋅ ( v ⊗ v ) ∈ L r ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) for any 1 ≤ r ≤ s / 2 1\le r\le s/2 1 ≤ r ≤ s /2 when s < ∞ s<\infty s < ∞ , and P ∇ ⋅ ( v ⊗ v ) ∈ L r ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \mathbb{P}\nabla\cdot(\mathbf{v}\otimes\mathbf{v})\in L^r([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)) P ∇ ⋅ ( v ⊗ v ) ∈ L r ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) for any 1 ≤ r ≤ ∞ 1\le r\le\infty 1 ≤ r ≤ ∞ when s = ∞ s=\infty s = ∞ .
Proof. To control the Leray projection P \mathbb{P} P , we use the fact that it is given by a matrix-valued Fourier multiplier P i j ( k ) = δ i j − k i k j / ∣ k ∣ 2 \mathcal{P}_{ij}(\mathbf{k})=\delta_{ij}-k_i k_j/|\mathbf{k}|^2 P ij ( k ) = δ ij − k i k j /∣ k ∣ 2 which is homogeneous of degree 0. We employ the following specific estimate from Theorem 3.6.3 of [94 ] : for K > d ( 1 min ( p , 1 ) − 1 2 ) \mathcal{K}>d\left(\frac{1}{\min(p,1)}-\frac{1}{2}\right) K > d ( m i n ( p , 1 ) 1 − 2 1 )
∥ ∑ k ∈ Z d m ( k ) ( F f ) ( k ) ∥ B p , q σ ( T d ) ≤ c ∥ m ∥ K ( R d ) ∥ f ∥ B p , q σ ( T d ) . (31) \left\|\sum_{\mathbf{k}\in\mathbb{Z}^d}m(\mathbf{k)(\mathcal{F}f)(\mathbf{k})}\right\|_{B^\sigma_{p,q}(\mathbb{T}^d)}\le c\|m\|_{\mathcal{K}(\mathbb{R}^d)}\|f\|_{B^\sigma_{p,q}(\mathbb{T}^d)}. \tag*{(31)} k ∈ Z d ∑ m ( k ) ( F f ) ( k ) B p , q σ ( T d ) ≤ c ∥ m ∥ K ( R d ) ∥ f ∥ B p , q σ ( T d ) . (31)
where c c c is a constant and Fourier multipliers m m m , considered as functions on R d \mathbb{R}^d R d , are estimated by the norm
∥ m ∥ K ( R d ) : = ∥ ψ ∗ m ∥ H K ( R d ) + sup n ∈ N ∥ ψ ( ⋅ ) m ( 2 n ⋅ ) ∥ H K ( R d ) , \|m\|_{\mathcal{K}(\mathbb{R}^d)}:=\|\psi^*m\|_{H^\mathcal{K}(\mathbb{R}^d)}+\sup_{n\in\mathbb{N}}\|\psi(\cdot)m(2^n\cdot)\|_{H^\mathcal{K}(\mathbb{R}^d)}, ∥ m ∥ K ( R d ) := ∥ ψ ∗ m ∥ H K ( R d ) + n ∈ N sup ∥ ψ ( ⋅ ) m ( 2 n ⋅ ) ∥ H K ( R d ) ,
and where the smooth test functions ψ ∗ , ψ ∈ S ( R d ) \psi^*,\psi\in\mathcal{S}(\mathbb{R}^d) ψ ∗ , ψ ∈ S ( R d ) with the properties
∀ k : 0 ≤ ψ ∗ ( k ) ≤ 1 , supp ψ ∗ ⊂ { k : ∣ k ∣ ≤ 4 } , ψ ∗ ( k ) = 1 when ∣ k ∣ ≤ 2 , \forall\mathbf{k}:0\le\psi^*(\mathbf{k})\le1,\qquad\operatorname{supp}\psi^*\subset\{\mathbf{k}:|\mathbf{k}|\le4\},\qquad\psi^*(\mathbf{k})=1\text{ when }|\mathbf{k}|\le2, ∀ k : 0 ≤ ψ ∗ ( k ) ≤ 1 , supp ψ ∗ ⊂ { k : ∣ k ∣ ≤ 4 } , ψ ∗ ( k ) = 1 when ∣ k ∣ ≤ 2 ,
∀ k : 0 ≤ ψ ( k ) ≤ 1 , supp ψ ⊂ { k : 1 / 4 ≤ ∣ k ∣ ≤ 4 } , ψ ( k ) = 1 when 1 / 2 ≤ ∣ k ∣ ≤ 2. \forall\mathbf{k}:0\le\psi(\mathbf{k})\le1,\qquad\operatorname{supp}\psi\subset\{\mathbf{k}:1/4\le|\mathbf{k}|\le4\},\qquad\psi(\mathbf{k})=1\text{ when }1/2\le|\mathbf{k}|\le2. ∀ k : 0 ≤ ψ ( k ) ≤ 1 , supp ψ ⊂ { k : 1/4 ≤ ∣ k ∣ ≤ 4 } , ψ ( k ) = 1 when 1/2 ≤ ∣ k ∣ ≤ 2.
provide a spectral decomposition. In order to apply the above bound to the Leray projection, we must modify its multiplier to eliminate the singularity at ∣ k ∣ = 0 |\mathbf{k}|=0 ∣ k ∣ = 0 , by taking
( F P v ~ ) i ( k ) : = ∑ j = 1 d P ~ i j ( k ) ( F v ) j ( k ) , k ∈ Z d (32) (\mathcal{F}\widetilde{\mathbb{P}\mathbf{v}})_i(\mathbf{k}):=\sum_{j=1}^d\widetilde{\mathcal{P}}_{ij}(\mathbf{k})(\mathcal{F}\mathbf{v})_j(\mathbf{k}),\qquad\mathbf{k}\in\mathbb{Z}^d \tag*{(32)} ( F P v ) i ( k ) := j = 1 ∑ d P ij ( k ) ( F v ) j ( k ) , k ∈ Z d (32)
with modified matrix multiplier given by
P ~ i j ( k ) : = { δ i j − k i k j / ∣ k ∣ 2 , ∣ k ∣ ≥ 1 / 4 , ( δ i j − k i k j / ∣ k ∣ 2 ) [ 1 − exp ( − ∣ k ∣ 2 / ( 1 / 16 − ∣ k ∣ 2 ) ) ] , 0 < ∣ k ∣ < 1 / 4 , 0 , k = 0. \widetilde{\mathcal{P}}_{ij}(\mathbf{k}):=
\begin{cases}
\delta_{ij}-k_i k_j/|\mathbf{k}|^2, & |\mathbf{k}|\ge1/4,\\
(\delta_{ij}-k_i k_j/|\mathbf{k}|^2)[1-\exp(-|\mathbf{k}|^2/(1/16-|\mathbf{k}|^2))], & 0<|\mathbf{k}|<1/4,\\
0, & \mathbf{k}=0.
\end{cases} P ij ( k ) := ⎩ ⎨ ⎧ δ ij − k i k j /∣ k ∣ 2 , ( δ ij − k i k j /∣ k ∣ 2 ) [ 1 − exp ( − ∣ k ∣ 2 / ( 1/16 − ∣ k ∣ 2 ))] , 0 , ∣ k ∣ ≥ 1/4 , 0 < ∣ k ∣ < 1/4 , k = 0.
which is now C ∞ C^\infty C ∞ and coincident with the original multiplier P i j ( k ) \mathcal{P}_{ij}(\mathbf{k}) P ij ( k ) for ∣ k ∣ ≥ 1 / 4 |\mathbf{k}|\ge1/4 ∣ k ∣ ≥ 1/4 . The only argument k ∈ Z d \mathbf{k}\in\mathbb{Z}^d k ∈ Z d in the expression (32) affected by the modification is thus k = 0 \mathbf{k}=0 k = 0 , so that P ~ \widetilde{\mathcal{P}} P and P \mathcal{P} P are identical in their action on fields v \mathbf{v} v with zero space-mean. Hence, we can identify P ~ = P \widetilde{\mathcal{P}}=\mathcal{P} P = P . Now take K > d / 2 \mathcal{K}>d/2 K > d /2 . Using that ψ ∗ , ψ \psi^*,\psi ψ ∗ , ψ are smooth and vanish outside a compact set, and that P ~ i j ( 2 n k ) = P ~ i j ( k ) \widetilde{\mathcal{P}}_{ij}(2^n\mathbf{k})=\widetilde{\mathcal{P}}_{ij}(\mathbf{k}) P ij ( 2 n k ) = P ij ( k ) for ∣ k ∣ ≥ 1 / 4 |\mathbf{k}|\ge1/4 ∣ k ∣ ≥ 1/4 , we conclude ψ ∗ P ~ i j ψ ( ⋅ ) P ~ i j ( 2 n ⋅ ) \psi^*\widetilde{\mathcal{P}}_{ij}\psi(\cdot)\widetilde{\mathcal{P}}_{ij}(2^n\cdot) ψ ∗ P ij ψ ( ⋅ ) P ij ( 2 n ⋅ ) and their derivatives are uniformly bounded in n n n and compactly supported. Thus ∥ P ~ ∥ K ( R d ) < ∞ \|\widetilde{\mathcal{P}}\|_{\mathcal{K}(\mathbb{R}^d)}<\infty ∥ P ∥ K ( R d ) < ∞ .
By the estimate (31) we thus obtain with v \mathbf{v} v denoting v ( ⋅ , t ) \mathbf{v}(\cdot,t) v ( ⋅ , t ) for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] that
∥ P ∇ ⋅ ( v ⊗ v ) ∥ B p / 2 , ∞ σ − β ( T d ) ≲ ∥ ∇ ⋅ ( v ⊗ v ) ∥ B p / 2 , ∞ σ − β ( T d ) . \|\mathbb{P}\nabla\cdot(\mathbf{v}\otimes\mathbf{v})\|_{B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)}
\lesssim\|\nabla\cdot(\mathbf{v}\otimes\mathbf{v})\|_{B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)}. ∥ P ∇ ⋅ ( v ⊗ v ) ∥ B p /2 , ∞ σ − β ( T d ) ≲ ∥∇ ⋅ ( v ⊗ v ) ∥ B p /2 , ∞ σ − β ( T d ) .
Per Theorem 2.2 and Proposition 2.3 of [95 ] , we have the estimate for β ≥ 1 \beta\ge1 β ≥ 1
∥ ∇ ⋅ ( v ⊗ v ) ∥ B p / 2 , ∞ σ − β ( T d ) ≲ ∥ v ⊗ v ∥ B p / 2 , ∞ σ ( T d ) . \|\nabla\cdot(\mathbf{v}\otimes\mathbf{v})\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)}\lesssim\|\mathbf{v}\otimes\mathbf{v}\|_{B_{p/2,\infty}^{\sigma}(\mathbb{T}^d)}. ∥∇ ⋅ ( v ⊗ v ) ∥ B p /2 , ∞ σ − β ( T d ) ≲ ∥ v ⊗ v ∥ B p /2 , ∞ σ ( T d ) .
Since all finite-dimensional norms are equivalent, we use the natural inner product structure of tensored Hilbert space to write:
∣ v ⊗ v ∣ ≲ ∣ ⟨ v ⊗ v , v ⊗ v ⟩ R 3 ⊗ R 3 ∣ 1 / 2 = ∣ ⟨ v , v ⟩ ∣ = ∣ v ∣ 2 ⟹ ∥ v ⊗ v ∥ L p / 2 ( T d ) ≲ ∥ v ∥ L p ( T d ) 2 . |\mathbf{v}\otimes\mathbf{v}|\lesssim|\langle\mathbf{v}\otimes\mathbf{v},\mathbf{v}\otimes\mathbf{v}\rangle_{\mathbb{R}^3\otimes\mathbb{R}^3}|^{1/2}=|\langle\mathbf{v},\mathbf{v}\rangle|=|\mathbf{v}|^2\quad\Longrightarrow\quad\|\mathbf{v}\otimes\mathbf{v}\|_{L^{p/2}(\mathbb{T}^d)}\lesssim\|\mathbf{v}\|_{L^p(\mathbb{T}^d)}^2. ∣ v ⊗ v ∣ ≲ ∣ ⟨ v ⊗ v , v ⊗ v ⟩ R 3 ⊗ R 3 ∣ 1/2 = ∣ ⟨ v , v ⟩ ∣ = ∣ v ∣ 2 ⟹ ∥ v ⊗ v ∥ L p /2 ( T d ) ≲ ∥ v ∥ L p ( T d ) 2 .
For any i , j i,j i , j we have
∥ v i ( ⋅ + h ) v j ( ⋅ + h ) − v i v j ∥ L p ( T ) d ≤ ∥ v i ( ⋅ + h ) − v i ∥ L p ( T ) d ∥ v j ( ⋅ + h ) ∥ L p ( T ) d \|\mathbf{v}_i(\cdot+\mathbf{h})\mathbf{v}_j(\cdot+\mathbf{h})-\mathbf{v}_i\mathbf{v}_j\|_{L^p(\mathbb{T})^d}\le\|\mathbf{v}_i(\cdot+\mathbf{h})-\mathbf{v}_i\|_{L^p(\mathbb{T})^d}\|\mathbf{v}_j(\cdot+\mathbf{h})\|_{L^p(\mathbb{T})^d} ∥ v i ( ⋅ + h ) v j ( ⋅ + h ) − v i v j ∥ L p ( T ) d ≤ ∥ v i ( ⋅ + h ) − v i ∥ L p ( T ) d ∥ v j ( ⋅ + h ) ∥ L p ( T ) d
+ ∥ v i ∥ L p ( T ) d ∥ v j ( ⋅ + h ) − v j ∥ L p ( T ) d . \qquad+\|\mathbf{v}_i\|_{L^p(\mathbb{T})^d}\|\mathbf{v}_j(\cdot+\mathbf{h})-\mathbf{v}_j\|_{L^p(\mathbb{T})^d}. + ∥ v i ∥ L p ( T ) d ∥ v j ( ⋅ + h ) − v j ∥ L p ( T ) d .
All of the RHS terms are bounded by constants times ∣ h ∣ σ |\mathbf{h}|^\sigma ∣ h ∣ σ , so that
∥ v ⊗ v ∥ B p / 2 , ∞ σ ( T d ) ≲ ∥ v ∥ B p , ∞ σ ( T d ) 2 . \|\mathbf{v}\otimes\mathbf{v}\|_{B_{p/2,\infty}^{\sigma}(\mathbb{T}^d)}\lesssim\|\mathbf{v}\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^d)}^2. ∥ v ⊗ v ∥ B p /2 , ∞ σ ( T d ) ≲ ∥ v ∥ B p , ∞ σ ( T d ) 2 .
For s < ∞ s<\infty s < ∞ , we use linearity and monotonicity of the Lebesgue integration over time, and then apply Hölder’s inequality
∫ 0 T ∥ P ∇ ⋅ ( v ⊗ v ) ( ⋅ , t ) ∥ B p / 2 , ∞ σ − β ( T d ) r d t ≲ ∫ 0 T ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) 2 r d t ≲ ( ∫ 0 T ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) s d t ) 2 r / s , 1 ≤ r ≤ s / 2. \int_0^T\|\mathbb{P}\nabla\cdot(\mathbf{v}\otimes\mathbf{v})(\cdot,t)\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)}^r\,\mathrm{d}t\lesssim\int_0^T\|\mathbf{v}(\cdot,t)\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^d)}^{2r}\,\mathrm{d}t\lesssim\left(\int_0^T\|\mathbf{v}(\cdot,t)\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^d)}^s\,\mathrm{d}t\right)^{2r/s},\qquad1\le r\le s/2. ∫ 0 T ∥ P ∇ ⋅ ( v ⊗ v ) ( ⋅ , t ) ∥ B p /2 , ∞ σ − β ( T d ) r d t ≲ ∫ 0 T ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) 2 r d t ≲ ( ∫ 0 T ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) s d t ) 2 r / s , 1 ≤ r ≤ s /2.
Thus
∥ P ∇ ⋅ ( v ⊗ v ) ∥ L r ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ ∥ v ∥ L s ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 . \|\mathbb{P}\nabla\cdot(\mathbf{v}\otimes\mathbf{v})\|_{L^r([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d))}\lesssim\|\mathbf{v}\|_{L^s([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^d))}^2. ∥ P ∇ ⋅ ( v ⊗ v ) ∥ L r ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ ∥ v ∥ L s ([ 0 , T ] , B p , ∞ σ ( T d )) 2 .
For s = ∞ s=\infty s = ∞ we have
ess sup t ∈ [ 0 , T ] ∥ P ∇ ⋅ [ v ( ⋅ , t ) ⊗ v ( ⋅ , t ) ] ∥ B p / 2 , ∞ σ − β ( T d ) ≲ ess sup t ∈ [ 0 , T ] ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) 2 , \operatorname*{ess\,sup}_{t\in[0,T]}\|\mathbb{P}\nabla\cdot[\mathbf{v}(\cdot,t)\otimes\mathbf{v}(\cdot,t)]\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)}\lesssim\operatorname*{ess\,sup}_{t\in[0,T]}\|\mathbf{v}(\cdot,t)\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^d)}^2, t ∈ [ 0 , T ] ess sup ∥ P ∇ ⋅ [ v ( ⋅ , t ) ⊗ v ( ⋅ , t )] ∥ B p /2 , ∞ σ − β ( T d ) ≲ t ∈ [ 0 , T ] ess sup ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) 2 ,
which gives for 1 ≤ r ≤ ∞ 1\le r\le\infty 1 ≤ r ≤ ∞
∥ P ∇ ⋅ ( v ⊗ v ) ∥ L r ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ ∥ P ∇ ⋅ ( v ⊗ v ) ∥ L ∞ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ ∥ v ∥ L ∞ ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 . \|\mathbb{P}\nabla\cdot(\mathbf{v}\otimes\mathbf{v})\|_{L^r([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d))}\lesssim\|\mathbb{P}\nabla\cdot(\mathbf{v}\otimes\mathbf{v})\|_{L^\infty([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d))}\lesssim\|\mathbf{v}\|_{L^\infty([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^d))}^2. ∥ P ∇ ⋅ ( v ⊗ v ) ∥ L r ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ ∥ P ∇ ⋅ ( v ⊗ v ) ∥ L ∞ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ ∥ v ∥ L ∞ ([ 0 , T ] , B p , ∞ σ ( T d )) 2 .
Lemma 4. Let 1 ≤ r ≤ s ≤ ∞ 1\le r\le s\le\infty 1 ≤ r ≤ s ≤ ∞ , 2 ≤ p < ∞ 2\le p<\infty 2 ≤ p < ∞ , 0 < σ < 1 0<\sigma<1 0 < σ < 1 , β ≥ 2 \beta\ge2 β ≥ 2 , and v ∈ L s ( [ 0 , T ] , B p , ∞ σ ( T d ) ) \mathbf{v}\in L^s([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^d)) v ∈ L s ([ 0 , T ] , B p , ∞ σ ( T d )) . Then Δ v ∈ L r ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \Delta\mathbf{v}\in L^r([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)) Δ v ∈ L r ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) .
Proof. Again from Theorem 2.2 and Proposition 2.3 of [95 ] , we have
∥ Δ v ( ⋅ , t ) ∥ B p / 2 , ∞ σ − β ( T d ) ≲ ∥ v ( ⋅ , t ) ∥ B p / 2 , ∞ σ ( T d ) ≲ ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) , \|\Delta\mathbf{v}(\cdot,t)\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)}\lesssim\|\mathbf{v}(\cdot,t)\|_{B_{p/2,\infty}^{\sigma}(\mathbb{T}^d)}\lesssim\|\mathbf{v}(\cdot,t)\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^d)}, ∥Δ v ( ⋅ , t ) ∥ B p /2 , ∞ σ − β ( T d ) ≲ ∥ v ( ⋅ , t ) ∥ B p /2 , ∞ σ ( T d ) ≲ ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) , using for the last step the continuous embedding B p , ∞ σ ( T d ) ↪ B p / 2 , ∞ σ ( T d ) B_{p,\infty}^{\sigma}(\mathbb{T}^d)\hookrightarrow B_{p/2,\infty}^{\sigma}(\mathbb{T}^d) B p , ∞ σ ( T d ) ↪ B p /2 , ∞ σ ( T d ) . For s < ∞ s<\infty s < ∞ , we again apply the properties of the Lebesgue integral and Hölder inequality
∫ 0 T ∥ Δ v ( ⋅ , t ) ∥ B p / 2 , ∞ σ − β ( T d ) r d t ≲ ∫ 0 T ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) r d t ≲ ( ∫ 0 T ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) s d t ) r / s . \int_0^T\|\Delta\mathbf{v}(\cdot,t)\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)}^r\,\mathrm{d}t\lesssim\int_0^T\|\mathbf{v}(\cdot,t)\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^d)}^r\,\mathrm{d}t\lesssim\left(\int_0^T\|\mathbf{v}(\cdot,t)\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^d)}^s\,\mathrm{d}t\right)^{r/s}. ∫ 0 T ∥Δ v ( ⋅ , t ) ∥ B p /2 , ∞ σ − β ( T d ) r d t ≲ ∫ 0 T ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) r d t ≲ ( ∫ 0 T ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) s d t ) r / s . Thus
∥ Δ v ∥ L r ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ ∥ v ∥ L s ( [ 0 , T ] , B p , ∞ σ ( T d ) ) . \|\Delta\mathbf{v}\|_{L^r([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d))}\lesssim\|\mathbf{v}\|_{L^s([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^d))}. ∥Δ v ∥ L r ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ ∥ v ∥ L s ([ 0 , T ] , B p , ∞ σ ( T d )) . Similarly for s = ∞ s=\infty s = ∞ we have
ess sup t ∈ [ 0 , T ] ∥ Δ v ( ⋅ , t ) ∥ B p / 2 , ∞ σ − β ( T d ) ≲ ess sup t ∈ [ 0 , T ] ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) , \operatorname*{ess\,sup}_{t\in[0,T]}\|\Delta\mathbf{v}(\cdot,t)\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)}\lesssim\operatorname*{ess\,sup}_{t\in[0,T]}\|\mathbf{v}(\cdot,t)\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^d)}, t ∈ [ 0 , T ] ess sup ∥Δ v ( ⋅ , t ) ∥ B p /2 , ∞ σ − β ( T d ) ≲ t ∈ [ 0 , T ] ess sup ∥ v ( ⋅ , t ) ∥ B p , ∞ σ ( T d ) , and
∥ Δ v ∥ L r ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ ∥ Δ v ∥ L ∞ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ ∥ v ∥ L ∞ ( [ 0 , T ] , B p , ∞ σ ( T d ) ) , 1 ≤ r ≤ ∞ . \|\Delta\mathbf{v}\|_{L^r([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d))}\lesssim\|\Delta\mathbf{v}\|_{L^\infty([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d))}\lesssim\|\mathbf{v}\|_{L^\infty([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^d))},\qquad1\le r\le\infty. ∥Δ v ∥ L r ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ ∥Δ v ∥ L ∞ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ ∥ v ∥ L ∞ ([ 0 , T ] , B p , ∞ σ ( T d )) , 1 ≤ r ≤ ∞.
Finally, to control the effect of the projections π Λ \pi_{\Lambda} π Λ , we have
Lemma 5. For any σ ∈ R \sigma\in\mathbb{R} σ ∈ R , 1 ≤ p ≤ ∞ 1 \le p \le\infty 1 ≤ p ≤ ∞ , v ∈ B p , ∞ σ ( T d ) v \in B_{p,\infty}^{\sigma}(\mathbb{T}^{d}) v ∈ B p , ∞ σ ( T d ) and Λ = 2 m \Lambda= 2^{m} Λ = 2 m for some m ∈ N m \in\mathbb{N} m ∈ N , we have
∥ π Λ v ∥ B p , ∞ σ ( T d ) ↑ ∥ v ∥ B p , ∞ σ ( T d ) as m ↑ ∞ . \|\pi_{\Lambda}v\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^{d})} \uparrow\|v\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^{d})}\quad\text{as}\quad m \uparrow\infty. ∥ π Λ v ∥ B p , ∞ σ ( T d ) ↑ ∥ v ∥ B p , ∞ σ ( T d ) as m ↑ ∞.
Proof. Define corridors in Fourier space as
C 0 : = { k : ∣ k ∣ ∞ < 1 } , C j : = { k : 2 j − 1 ≤ ∣ k ∣ ∞ < 2 j } . C_{0}:=\{k:|k|_{\infty}<1\},\qquad C_{j}:=\{k:2^{j-1}\leq|k|_{\infty}<2^{j}\}. C 0 := { k : ∣ k ∣ ∞ < 1 } , C j := { k : 2 j − 1 ≤ ∣ k ∣ ∞ < 2 j } .
It then suffices to recall the Lizorkin representations (Theorem 3.5.3 of [94 ] ):
∥ v ∥ B p , ∞ σ ( T d ) = sup j ∈ N 2 σ j ∥ ∑ k ∈ C j ( F v ) ( k ) exp ( i ⟨ k , ⋅ ⟩ ) ∥ L p ( T d ) . \|v\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^{d})}=\sup_{j\in\mathbb{N}}2^{\sigma j}\left\|\sum_{k\in C_{j}}(\mathcal{F}v)(k)\exp(i\langle k,\cdot\rangle)\right\|_{L^{p}(\mathbb{T}^{d})}. ∥ v ∥ B p , ∞ σ ( T d ) = j ∈ N sup 2 σ j k ∈ C j ∑ ( F v ) ( k ) exp ( i ⟨ k , ⋅ ⟩) L p ( T d ) .
Since ( F v ) ( k ) (\mathcal{F}v)(k) ( F v ) ( k ) is just the coefficient of the k k k th eigenbasis vector, when ∣ k ∣ ∞ < 2 m |k|_{\infty}<2^{m} ∣ k ∣ ∞ < 2 m and ∣ k ∣ ∞ ≥ 2 m |k|_{\infty}\geq2^{m} ∣ k ∣ ∞ ≥ 2 m we have ( F π Λ v ) ( k ) = ( F v ) ( k ) (\mathcal{F}\pi_{\Lambda}v)(k)=(\mathcal{F}v)(k) ( F π Λ v ) ( k ) = ( F v ) ( k ) and ( F π Λ v ) ( k ) = 0 (\mathcal{F}\pi_{\Lambda}v)(k)=0 ( F π Λ v ) ( k ) = 0 respectively, and the sum in the Lizorkin representation vanishes on C j C_{j} C j when j > m j>m j > m . Thus
∥ π Λ v ∥ B p , ∞ σ ( T d ) = sup j ≤ m 2 σ j ∥ ∑ k ∈ C j ( F v ) ( k ) exp ( i ⟨ k , ⋅ ⟩ ) ∥ L p ( T d ) \|\pi_{\Lambda}v\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^{d})}=\sup_{j\leq m}2^{\sigma j}\left\|\sum_{k\in C_{j}}(\mathcal{F}v)(k)\exp(i\langle k,\cdot\rangle)\right\|_{L^{p}(\mathbb{T}^{d})} ∥ π Λ v ∥ B p , ∞ σ ( T d ) = j ≤ m sup 2 σ j k ∈ C j ∑ ( F v ) ( k ) exp ( i ⟨ k , ⋅ ⟩) L p ( T d )
and monotonic convergence of ∥ π Λ v ∥ B p , ∞ σ ( T d ) \|\pi_{\Lambda}v\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^{d})} ∥ π Λ v ∥ B p , ∞ σ ( T d ) to ∥ v ∥ B p , ∞ σ ( T d ) \|v\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^{d})} ∥ v ∥ B p , ∞ σ ( T d ) becomes obvious. □ \square □
Combining all of these results, we infer from Lemma 3 and Lemma 5 that
∥ P ∇ ⋅ π Λ ( u ′ ⊗ u ′ ) ∥ L r / 2 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ ∥ u ′ ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 . \|\mathcal{P}\nabla\cdot\pi_{\Lambda}(u^{\prime}\otimes u^{\prime})\|_{L^{r/2}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\lesssim\|u^{\prime}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}^{2}. ∥ P ∇ ⋅ π Λ ( u ′ ⊗ u ′ ) ∥ L r /2 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ ∥ u ′ ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) 2 .
and from Lemma 4 and Lemma 5 that
∥ Δ π Λ u ′ ∥ L r / 2 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ ∥ u ′ ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) . \|\Delta\pi_{\Lambda}u^{\prime}\|_{L^{r/2}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\lesssim\|u^{\prime}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}. ∥Δ π Λ u ′ ∥ L r /2 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ ∥ u ′ ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) .
Finally, by Lemma 5 we evidently have
∥ π Λ f ∥ L r / 2 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ ∥ f ∥ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) . \|\pi_{\Lambda}f\|_{L^{r/2}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\lesssim\|f\|_{C^{0}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}. ∥ π Λ f ∥ L r /2 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ ∥ f ∥ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) .
As all of our estimates above are P × μ \mathbb{P}\times\mu P × μ almost sure, we get
∥ ∂ t z ν ∥ L r / 2 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ ∥ u ν ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 + ∥ u ν ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) + ∥ f ∥ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \|\partial_{t}z^{\nu}\|_{L^{r/2}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\lesssim\|u^{\nu}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}^{2}+\|u^{\nu}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}+\|f\|_{C^{0}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))} ∥ ∂ t z ν ∥ L r /2 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ ∥ u ν ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) 2 + ∥ u ν ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) + ∥ f ∥ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d ))
almost surely.
Next we define Δ t , s z ν : = z ν ( ⋅ , t ) − z ν ( ⋅ , s ) \Delta_{t,s}z^{\nu}:=z^{\nu}(\cdot,t)-z^{\nu}(\cdot,s) Δ t , s z ν := z ν ( ⋅ , t ) − z ν ( ⋅ , s ) and note by (30) and Hölder inequality that
∥ Δ t , s z ν ∥ B p / 2 , ∞ σ − β ( T d ) ≤ ∫ s t ∥ ∂ t z ν ( ⋅ , τ ) ∥ B p / 2 , ∞ σ − β ( T d ) d τ \|\Delta_{t,s}z^{\nu}\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})}\leq\int_{s}^{t}\|\partial_{t}z^{\nu}(\cdot,\tau)\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})}\,\mathrm{d}\tau ∥ Δ t , s z ν ∥ B p /2 , ∞ σ − β ( T d ) ≤ ∫ s t ∥ ∂ t z ν ( ⋅ , τ ) ∥ B p /2 , ∞ σ − β ( T d ) d τ
≲ ∣ t − s ∣ ( r − 2 ) / r ∥ ∂ t z ν ∥ L r / 2 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) . \lesssim|t-s|^{(r-2)/r}\|\partial_{t}z^{\nu}\|_{L^{r/2}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}. ≲ ∣ t − s ∣ ( r − 2 ) / r ∥ ∂ t z ν ∥ L r /2 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) .
Recalling the definition of the semi-norm
∥ z ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) : = sup t , s ∈ [ 0 , T ] ∥ z ( ⋅ , t ) − z ( ⋅ , s ) ∥ B p / 2 , ∞ σ − β ( T d ) ∣ t − s ∣ γ \|z\|_{C^{0,\gamma}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}:=\sup_{t,s\in[0,T]}\frac{\|z(\cdot,t)-z(\cdot,s)\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})}}{|t-s|^{\gamma}} ∥ z ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) := t , s ∈ [ 0 , T ] sup ∣ t − s ∣ γ ∥ z ( ⋅ , t ) − z ( ⋅ , s ) ∥ B p /2 , ∞ σ − β ( T d )
we obtain for γ ≤ ( r − 2 ) / r \gamma\le(r-2)/r γ ≤ ( r − 2 ) / r
E μ [ ∥ z ν ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ] ≲ E μ [ ∥ ∂ t z ν ∥ L r / 2 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ] \mathbb{E}_{\mu}\left[\left\|z^{\nu}\right\|_{C^{0,\gamma}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\right]\lesssim\mathbb{E}_{\mu}\left[\left\|\partial_t z^{\nu}\right\|_{L^{r/2}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\right] E μ [ ∥ z ν ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ] ≲ E μ [ ∥ ∂ t z ν ∥ L r /2 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ]
≲ ∫ ∥ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 d P μ ν ( u ) + ( ∫ ∥ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 d P μ ν ( u ) ) 1 / 2 + ∥ f ∥ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) . \begin{aligned}
\lesssim\int\left\|u\right\|_{L^r([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}^{2}\,\mathrm{d}P^{\nu}_{\mu}(u)+\left(\int\left\|u\right\|_{L^r([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}^{2}\,\mathrm{d}P^{\nu}_{\mu}(u)\right)^{1/2} \\
&+\left\|f\right\|_{C^0([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}.
\end{aligned} ≲ ∫ ∥ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) 2 d P μ ν ( u ) + ( ∫ ∥ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) 2 d P μ ν ( u ) ) 1/2 + ∥ f ∥ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) .
We need a similar bound for the C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) C^0([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})) C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) norm. We use the expression (30) and bound separately the contributions from z ν ( ⋅ , 0 ) z^{\nu}(\cdot,0) z ν ( ⋅ , 0 ) and ∫ 0 t ∂ t z ν ( ⋅ , s ) d s \int_0^t\partial_t z^{\nu}(\cdot,s)\,\mathrm{d}s ∫ 0 t ∂ t z ν ( ⋅ , s ) d s . From Lemmas 2 and 5 and from our assumption u 0 ∈ L p ( T d ) u_0\in L^p(\mathbb{T}^{d}) u 0 ∈ L p ( T d ) , we have u ν ( ⋅ , 0 ) = π Λ u 0 u^{\nu}(\cdot,0)=\pi_{\Lambda}u_0 u ν ( ⋅ , 0 ) = π Λ u 0 bounded uniformly across ν \nu ν in B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}) B p /2 , ∞ σ − β ( T d ) . Then we have
∥ ∫ 0 t ∂ t z ν ( ⋅ , s ) d s ∥ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≤ ∫ 0 T ∥ ∂ t z ν ( ⋅ , s ) ∥ B p / 2 , ∞ σ − β ( T d ) d s ≲ ∥ ∂ t z ν ∥ L r / 2 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) . \left\|\int_0^t\partial_t z^{\nu}(\cdot,s)\,\mathrm{d}s\right\|_{C^0([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\le\int_0^T\left\|\partial_t z^{\nu}(\cdot,s)\right\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})}\,\mathrm{d}s\lesssim\left\|\partial_t z^{\nu}\right\|_{L^{r/2}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}. ∫ 0 t ∂ t z ν ( ⋅ , s ) d s C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≤ ∫ 0 T ∥ ∂ t z ν ( ⋅ , s ) ∥ B p /2 , ∞ σ − β ( T d ) d s ≲ ∥ ∂ t z ν ∥ L r /2 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) .
We have already estimated the last term, and all the estimates are almost sure in P × μ \mathbb{P}\times\mu P × μ . Thus we obtain
E μ [ ∥ z ν ∥ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ] ≲ ∫ ∥ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 d P μ ν ( u ) + ( ∫ ∥ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 d P μ ν ( u ) ) 1 / 2 + ∥ f ∥ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) + ∥ u 0 ∥ L p ( T d ) . \begin{aligned}
\mathbb{E}_{\mu}\left[\left\|z^{\nu}\right\|_{C^0([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\right]\lesssim\int\left\|u\right\|_{L^r([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}^{2}\,\mathrm{d}P^{\nu}_{\mu}(u)+\left(\int\left\|u\right\|_{L^r([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}^{2}\,\mathrm{d}P^{\nu}_{\mu}(u)\right)^{1/2} \\
&+\left\|f\right\|_{C^0([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}+\left\|u_0\right\|_{L^p(\mathbb{T}^{d})}.
\end{aligned} E μ [ ∥ z ν ∥ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ] ≲ ∫ ∥ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) 2 d P μ ν ( u ) + ( ∫ ∥ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) 2 d P μ ν ( u ) ) 1/2 + ∥ f ∥ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) + ∥ u 0 ∥ L p ( T d ) .
Finally, we conclude from our empirical hypothesis (10) that for γ ≤ ( r − 2 ) / r \gamma\le(r-2)/r γ ≤ ( r − 2 ) / r
sup ν > 0 E μ [ ∥ z ν ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ] < ∞ . \sup_{\nu>0}\mathbb{E}_{\mu}\left[\left\|z^{\nu}\right\|_{C^{0,\gamma}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\right]<\infty. ν > 0 sup E μ [ ∥ z ν ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ] < ∞.
Estimate of P ∇ ⋅ π Λ ξ \mathcal{P}\nabla\cdot\pi_{\Lambda}\xi P ∇ ⋅ π Λ ξ : Next, we argue as before from Lemma 3 that
∥ P ∇ ⋅ [ π Λ Δ t , s ξ ] ∥ B p / 2 , ∞ σ − β ( T d ) ≲ ∥ π Λ Δ t , s ξ ∥ B p / 2 , ∞ σ − β + 1 ( T d ) . \left\|\mathcal{P}\nabla\cdot\left[\pi_{\Lambda}\Delta_{t,s}\xi\right]\right\|_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})}\lesssim\left\|\pi_{\Lambda}\Delta_{t,s}\xi\right\|_{B_{p/2,\infty}^{\sigma-\beta+1}(\mathbb{T}^{d})}. ∥ P ∇ ⋅ [ π Λ Δ t , s ξ ] ∥ B p /2 , ∞ σ − β ( T d ) ≲ ∥ π Λ Δ t , s ξ ∥ B p /2 , ∞ σ − β + 1 ( T d ) .
Using the representation (7), we can write
π Λ Δ t , s ξ k = ∑ ∣ k ∣ ∞ ≤ 2 m ( ν ) Δ t , s ξ k exp ( i ⟨ k , ⋅ ⟩ ) . \pi_{\Lambda}\Delta_{t,s}\xi^{k}=\sum_{\lvert k\rvert_{\infty}\le2^{m(\nu)}}\Delta_{t,s}\xi^{k}\exp(i\langle k,\cdot\rangle). π Λ Δ t , s ξ k = ∣ k ∣ ∞ ≤ 2 m ( ν ) ∑ Δ t , s ξ k exp ( i ⟨ k , ⋅ ⟩) .
where the complex matrix-valued Brownian motions ξ k \xi^{k} ξ k are mutually independent except for the constraints ξ ‾ i j k ( t ) = ξ i j − k ( t ) \overline{\xi}_{ij}^{k}(t)=\xi_{ij}^{-k}(t) ξ ij k ( t ) = ξ ij − k ( t ) , ξ j i k ( t ) = ξ ‾ i j k ( t ) \xi_{ji}^{k}(t)=\overline{\xi}_{ij}^{k}(t) ξ j i k ( t ) = ξ ij k ( t ) and ∑ i = 1 d ξ i i k ( t ) = 0 \sum_{i=1}^{d}\xi_{ii}^{k}(t)=0 ∑ i = 1 d ξ ii k ( t ) = 0 . Using a Littlewood–Paley form of the Besov norm,
∥ π Λ Δ t , s ξ k ∥ B p / 2 , ∞ σ − β + 1 ( T d ) = sup j ∈ N 2 ( σ − β + 1 ) j ∥ ∑ ∣ k ∣ ∞ ≤ 2 m ( ν ) φ j ( k ) Δ t , s ξ k exp ( i ⟨ k , ⋅ ⟩ R d ) ∥ L p / 2 ( T d ) , \left\|\pi_{\Lambda}\Delta_{t,s}\xi^{k}\right\|_{B_{p/2,\infty}^{\sigma-\beta+1}(\mathbb{T}^{d})}=\sup_{j\in\mathbb{N}}2^{(\sigma-\beta+1)j}\left\|\sum_{\lvert k\rvert_{\infty}\le2^{m(\nu)}}\varphi_j(k)\Delta_{t,s}\xi^{k}\exp(i\langle k,\cdot\rangle_{\mathbb{R}^{d}})\right\|_{L^{p/2}(\mathbb{T}^{d})}, π Λ Δ t , s ξ k B p /2 , ∞ σ − β + 1 ( T d ) = j ∈ N sup 2 ( σ − β + 1 ) j ∣ k ∣ ∞ ≤ 2 m ( ν ) ∑ φ j ( k ) Δ t , s ξ k exp ( i ⟨ k , ⋅ ⟩ R d ) L p /2 ( T d ) ,
where ( φ j ) j ∈ N (\varphi_j)_{j\in\mathbb{N}} ( φ j ) j ∈ N is a smooth partition of unity such that the φ j ∈ S ( R d ) \varphi_j\in\mathcal{S}(\mathbb{R}^{d}) φ j ∈ S ( R d ) are uniformly bounded and supported on 2 j − 1 < ∣ k ∣ ∞ < 2 j + 1 2^{j-1}<\lvert k\rvert_{\infty}<2^{j+1} 2 j − 1 < ∣ k ∣ ∞ < 2 j + 1 . See [94 ] , section 3.5.1. From the Minkowski inequality,
∥ ∥ ∑ ∣ k ∣ ∞ ≤ 2 m ( ν ) φ j ( k ) Δ t , s ξ k exp ( i ⟨ k , ⋅ ⟩ R d ) ∥ L p / 2 ( T d ) ∥ ≤ ∑ ∣ k ∣ ∞ ≤ 2 m ( ν ) φ j ( k ) ∥ Δ t , s ξ k ∥ ∥ exp ( i ⟨ k , ⋅ ⟩ R d ) ∥ L p / 2 ( T d ) , \left\|\left\|\sum_{\lvert k\rvert_{\infty}\le2^{m(\nu)}}\varphi_j(k)\Delta_{t,s}\xi^{k}\exp(i\langle k,\cdot\rangle_{\mathbb{R}^{d}})\right\|_{L^{p/2}(\mathbb{T}^{d})}\right\|
\le\sum_{\lvert k\rvert_{\infty}\le2^{m(\nu)}}\varphi_j(k)\left\|\Delta_{t,s}\xi^{k}\right\|\left\|\exp(i\langle k,\cdot\rangle_{\mathbb{R}^{d}})\right\|_{L^{p/2}(\mathbb{T}^{d})}, ∣ k ∣ ∞ ≤ 2 m ( ν ) ∑ φ j ( k ) Δ t , s ξ k exp ( i ⟨ k , ⋅ ⟩ R d ) L p /2 ( T d ) ≤ ∣ k ∣ ∞ ≤ 2 m ( ν ) ∑ φ j ( k ) Δ t , s ξ k ∥ exp ( i ⟨ k , ⋅ ⟩ R d ) ∥ L p /2 ( T d ) ,
where ∥ exp ( i ⟨ k , ⋅ ⟩ R d ) ∥ L p / 2 ( T d ) \left\|\exp(i\langle k,\cdot\rangle_{\mathbb{R}^{d}})\right\|_{L^{p/2}(\mathbb{T}^{d})} ∥ exp ( i ⟨ k , ⋅ ⟩ R d ) ∥ L p /2 ( T d ) is constant in k k k . Thus
∥ π Λ Δ t , s ξ k ∥ B p / 2 , ∞ σ − β + 1 , ∞ ( T d ) ≲ sup j ∈ N 2 ( σ − β + 1 ) j ∑ ∣ k ∣ ∞ ≤ 2 m ( ν ) φ j ( k ) ∥ Δ t , s ξ k ∥ , \left\|\pi_{\Lambda}\Delta_{t,s}\xi^{k}\right\|_{B_{p/2,\infty}^{\sigma-\beta+1,\infty}(\mathbb{T}^{d})}\lesssim\sup_{j\in\mathbb{N}}2^{(\sigma-\beta+1)j}\sum_{\lvert k\rvert_{\infty}\le2^{m(\nu)}}\varphi_j(k)\left\|\Delta_{t,s}\xi^{k}\right\|, π Λ Δ t , s ξ k B p /2 , ∞ σ − β + 1 , ∞ ( T d ) ≲ j ∈ N sup 2 ( σ − β + 1 ) j ∣ k ∣ ∞ ≤ 2 m ( ν ) ∑ φ j ( k ) Δ t , s ξ k ,
and we get
sup j ∈ N 2 ( σ − β + 1 ) j ∑ ∣ k ∣ ∞ ≤ 2 m ( ν ) φ j ( k ) ∣ Δ t , s ξ k ∣ ≲ sup j ∈ N 2 ( σ − β + 1 ) j ∑ ∣ k ∣ ∞ ≤ 2 m ( ν ) 1 { 2 j − 1 < ∣ k ∣ ∞ < 2 j + 1 } ( k ) ∣ Δ t , s ξ k ∣ \sup_{j\in\mathbb{N}}2^{(\sigma-\beta+1)j}\sum_{\lvert k\rvert_\infty\leq2^{m(\nu)}}\varphi_j(k)\lvert\Delta_{t,s}\xi^k\rvert
\lesssim\sup_{j\in\mathbb{N}}2^{(\sigma-\beta+1)j}\sum_{\lvert k\rvert_\infty\leq2^{m(\nu)}}\mathbf{1}_{\{2^{j-1}<\lvert k\rvert_\infty<2^{j+1}\}}(k)\lvert\Delta_{t,s}\xi^k\rvert j ∈ N sup 2 ( σ − β + 1 ) j ∣ k ∣ ∞ ≤ 2 m ( ν ) ∑ φ j ( k ) ∣ Δ t , s ξ k ∣ ≲ j ∈ N sup 2 ( σ − β + 1 ) j ∣ k ∣ ∞ ≤ 2 m ( ν ) ∑ 1 { 2 j − 1 < ∣ k ∣ ∞ < 2 j + 1 } ( k ) ∣ Δ t , s ξ k ∣
≲ sup j ≤ m ( ν ) 2 ( σ − β + 1 + d ) j max 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 ∣ Δ t , s ξ k ∣ . \lesssim\sup_{j\leq m(\nu)}2^{(\sigma-\beta+1+d)j}\max_{2^{j-1}<\lvert k\rvert_\infty\leq2^{j+1}}\lvert\Delta_{t,s}\xi^k\rvert. ≲ j ≤ m ( ν ) sup 2 ( σ − β + 1 + d ) j 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 max ∣ Δ t , s ξ k ∣ .
From the definition of the Hölder semi-norm ∥ f ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) : = sup t ≠ s ∈ [ 0 , T ] ∥ Δ t , s f ∥ B p / 2 , ∞ σ − β ( T d ) ∣ t − s ∣ γ \lVert f\rVert_{C^{0,\gamma}([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d))}:=\sup_{t\ne s\in[0,T]}\frac{\lVert\Delta_{t,s}f\rVert_{B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)}}{\lvert t-s\rvert^\gamma} ∥ f ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) := sup t = s ∈ [ 0 , T ] ∣ t − s ∣ γ ∥ Δ t , s f ∥ B p /2 , ∞ σ − β ( T d ) , we thus obtain finally
∥ P ∇ ⋅ π Λ ξ ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ≲ sup j ≤ m ( ν ) 2 ( σ − β + 1 + d ) j max 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 ∥ ξ k ∥ C 0 , γ ( [ 0 , T ] , R d ) ≤ ∑ j = 0 m ( ν ) 2 ( σ − β + 1 + d ) j max 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 ∥ ξ k ∥ C 0 , γ ( [ 0 , T ] , R d ) . (33) \begin{aligned}
\lVert\mathcal{P}\nabla\cdot\pi_\Lambda\xi\rVert_{C^{0,\gamma}([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d))}
\lesssim\sup_{j\leq m(\nu)}2^{(\sigma-\beta+1+d)j}\max_{2^{j-1}<\lvert k\rvert_\infty\leq2^{j+1}}\lVert\xi^k\rVert_{C^{0,\gamma}([0,T],\mathbb{R}^d)} \\
&\leq\sum_{j=0}^{m(\nu)}2^{(\sigma-\beta+1+d)j}\max_{2^{j-1}<\lvert k\rvert_\infty\leq2^{j+1}}\lVert\xi^k\rVert_{C^{0,\gamma}([0,T],\mathbb{R}^d)}. \tag*{(33)}
\end{aligned} ∥ P ∇ ⋅ π Λ ξ ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ≲ j ≤ m ( ν ) sup 2 ( σ − β + 1 + d ) j 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 max ∥ ξ k ∥ C 0 , γ ([ 0 , T ] , R d ) ≤ j = 0 ∑ m ( ν ) 2 ( σ − β + 1 + d ) j 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 max ∥ ξ k ∥ C 0 , γ ([ 0 , T ] , R d ) . (33)
Note that all of our estimates for P ∇ ⋅ π Λ ξ \mathcal{P}\nabla\cdot\pi_\Lambda\xi P ∇ ⋅ π Λ ξ are P \mathbb{P} P -almost sure.
Taking expectations
E [ ∥ P ∇ ⋅ π Λ ξ ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ] ≤ ∑ j = 0 m ( ν ) 2 ( σ − β + 1 + d ) j E [ max 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 ∥ ξ k ∥ C 0 , γ ( [ 0 , T ] , R d ) ] . \mathbb{E}\left[\lVert\mathcal{P}\nabla\cdot\pi_\Lambda\xi\rVert_{C^{0,\gamma}([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d))}\right]
\leq\sum_{j=0}^{m(\nu)}2^{(\sigma-\beta+1+d)j}\mathbb{E}\left[\max_{2^{j-1}<\lvert k\rvert_\infty\leq2^{j+1}}\lVert\xi^k\rVert_{C^{0,\gamma}([0,T],\mathbb{R}^d)}\right]. E [ ∥ P ∇ ⋅ π Λ ξ ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ] ≤ j = 0 ∑ m ( ν ) 2 ( σ − β + 1 + d ) j E [ 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 max ∥ ξ k ∥ C 0 , γ ([ 0 , T ] , R d ) ] .
However,
E [ max 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 ∥ ξ k ∥ C 0 , γ ( [ 0 , T ] , R d ) ] ≤ ∑ 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 E [ ∥ ξ k ∥ C 0 , γ ( [ 0 , T ] , R d ) ] \mathbb{E}\left[\max_{2^{j-1}<\lvert k\rvert_\infty\leq2^{j+1}}\lVert\xi^k\rVert_{C^{0,\gamma}([0,T],\mathbb{R}^d)}\right]
\leq\sum_{2^{j-1}<\lvert k\rvert_\infty\leq2^{j+1}}\mathbb{E}\left[\lVert\xi^k\rVert_{C^{0,\gamma}([0,T],\mathbb{R}^d)}\right] E [ 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 max ∥ ξ k ∥ C 0 , γ ([ 0 , T ] , R d ) ] ≤ 2 j − 1 < ∣ k ∣ ∞ ≤ 2 j + 1 ∑ E [ ∥ ξ k ∥ C 0 , γ ([ 0 , T ] , R d ) ]
≲ 2 j d E [ ∥ ξ ~ ∥ C 0 , γ ( [ 0 , T ] , R d ) ] . \lesssim2^{jd}\mathbb{E}\left[\lVert\widetilde{\xi}\rVert_{C^{0,\gamma}([0,T],\mathbb{R}^d)}\right]. ≲ 2 j d E [ ∥ ξ ∥ C 0 , γ ([ 0 , T ] , R d ) ] .
where in the last line we used the fact that the random variables ξ k \xi^k ξ k for k ≠ 0 k\neq0 k = 0 are all identically distributed and ξ ~ \widetilde{\xi} ξ is a single representative. We now recall the classical result that E [ ∥ ξ ~ ∥ C 0 , γ ( [ 0 , T ] , R d ) n ] < ∞ \mathbb{E}\left[\lVert\widetilde{\xi}\rVert_{C^{0,\gamma}([0,T],\mathbb{R}^d)}^n\right]<\infty E [ ∥ ξ ∥ C 0 , γ ([ 0 , T ] , R d ) n ] < ∞ for any γ < 1 / 2 \gamma<1/2 γ < 1/2 and n ∈ N n\in\mathbb{N} n ∈ N and, in fact, E [ e ε ∥ ξ ~ ∥ C 0 , γ ( [ 0 , T ] , R d ) 2 ] < ∞ \mathbb{E}\left[e^{\varepsilon\lVert\widetilde{\xi}\rVert_{C^{0,\gamma}([0,T],\mathbb{R}^d)}^2}\right]<\infty E [ e ε ∥ ξ ∥ C 0 , γ ([ 0 , T ] , R d ) 2 ] < ∞ for sufficiently small ε > 0 \varepsilon>0 ε > 0 [96 ] . The latter statement is essentially a consequence of Fernique’s theorem [97 ] , modulo some technical issues due to the non-separability of the Hölder space C 0 , γ ( [ 0 , T ] , R d ) C^{0,\gamma}([0,T],\mathbb{R}^d) C 0 , γ ([ 0 , T ] , R d ) ; e.g. see [98 ] . From this fact and the condition σ − β + 1 + 2 d < 0 \sigma-\beta+1+2d<0 σ − β + 1 + 2 d < 0 , we thus deduce for any γ < 1 / 2 \gamma<1/2 γ < 1/2 that
E [ ∥ P ∇ ⋅ π Λ ξ ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ] ≲ ∑ j = 0 m ( ν ) 2 ( σ − β + 1 + 2 d ) j ∝ 1 − ν − ( σ − β + 1 + 2 d ) α \mathbb{E}\left[\lVert\mathcal{P}\nabla\cdot\pi_\Lambda\xi\rVert_{C^{0,\gamma}([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d))}\right]
\lesssim\sum_{j=0}^{m(\nu)}2^{(\sigma-\beta+1+2d)j}
\propto1-\nu^{-(\sigma-\beta+1+2d)\alpha} E [ ∥ P ∇ ⋅ π Λ ξ ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ] ≲ j = 0 ∑ m ( ν ) 2 ( σ − β + 1 + 2 d ) j ∝ 1 − ν − ( σ − β + 1 + 2 d ) α
so that the lefthand side is uniformly bounded in ν \nu ν .
To get a similar estimate for the C 0 ( [ 0 , T ] , R d ) C^0([0,T],\mathbb{R}^d) C 0 ([ 0 , T ] , R d ) norm, we use an almost identical argument, which yields
E [ ∥ P ∇ ⋅ π Λ ξ ∥ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ] ≲ ∑ j = 0 m ( ν ) 2 ( σ − β + 1 + 2 d ) j E [ ∥ ξ ~ ∥ C 0 ( [ 0 , T ] , R d ) ] . \mathbb{E}\left[\lVert\mathcal{P}\nabla\cdot\pi_\Lambda\xi\rVert_{C^0([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d))}\right]
\lesssim\sum_{j=0}^{m(\nu)}2^{(\sigma-\beta+1+2d)j}\mathbb{E}\left[\lVert\widetilde{\xi}\rVert_{C^0([0,T],\mathbb{R}^d)}\right]. E [ ∥ P ∇ ⋅ π Λ ξ ∥ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ] ≲ j = 0 ∑ m ( ν ) 2 ( σ − β + 1 + 2 d ) j E [ ∥ ξ ∥ C 0 ([ 0 , T ] , R d ) ] .
The boundedness of the latter expectation is an even more classical result in probability theory, which can be deduced from Doob’s maximal inequality for submartingales or alternatively from more elementary considerations for Brownian motion. We obtain once again a bound uniform in ν \nu ν .
Finally, we can conclude that for any γ < 1 / 2 \gamma<1/2 γ < 1/2
sup ν > 0 E [ ∥ P ∇ ⋅ π Λ ξ ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ] < ∞ . \sup_{\nu>0}\mathbb{E}\left[\lVert\mathcal{P}\nabla\cdot\pi_\Lambda\xi\rVert_{C^{0,\gamma}([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d))}\right]<\infty. ν > 0 sup E [ ∥ P ∇ ⋅ π Λ ξ ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ] < ∞.
It is noteworthy that the prefactor ν α \nu^\alpha ν α played no role in this estimate.
Estimate of u ν = z ν + ν κ Q ∇ ⋅ π Λ ξ \mathbf{u}^{\nu} = \mathbf{z}^{\nu} + \nu^{\kappa}\mathcal{Q}\nabla\cdot\pi_{\Lambda}\xi u ν = z ν + ν κ Q ∇ ⋅ π Λ ξ : Putting together the previous bounds, we obtain
sup ν > 0 ∫ ∥ u ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) d P μ ν ( u ) = sup ν > 0 E μ [ ∥ u ν ∥ C 0 , γ ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ] < ∞ , \sup_{\nu>0}\int\|\mathbf{u}\|_{C^{0,\gamma}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\,dP_{\mu}^{\nu}(\mathbf{u})=\sup_{\nu>0}\mathbb{E}_{\mu}\left[\|\mathbf{u}^{\nu}\|_{C^{0,\gamma}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))}\right]<\infty, ν > 0 sup ∫ ∥ u ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) d P μ ν ( u ) = ν > 0 sup E μ [ ∥ u ν ∥ C 0 , γ ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ] < ∞ ,
for any γ < min { ( r − 2 ) / r , 1 / 2 } \gamma<\min\{(r-2)/r,1/2\} γ < min {( r − 2 ) / r , 1/2 } . Combined with our hypothesis (10), we then get that
sup ν > 0 ∫ ∥ u ∥ E r , p σ , β d P μ ν ( u ) < ∞ . \sup_{\nu>0}\int\|\mathbf{u}\|_{\mathcal{E}_{r,p}^{\sigma,\beta}}\,dP_{\mu}^{\nu}(\mathbf{u})<\infty. ν > 0 sup ∫ ∥ u ∥ E r , p σ , β d P μ ν ( u ) < ∞.
By Lemma II.3.1 of [28 ] , we conclude that the family of measures P μ ν P_{\mu}^{\nu} P μ ν , ν > 0 \nu>0 ν > 0 are tight on Z r , p σ , β \mathcal{Z}_{r,p}^{\sigma,\beta} Z r , p σ , β and there exists a subsequence ν k → 0 \nu_k\to0 ν k → 0 so that P μ ν k → P μ P_{\mu}^{\nu_k}\to P_{\mu} P μ ν k → P μ weakly.
Proof of (ii)
We establish here support properties of the weak limit measures in (i). In particular, we will show that P μ P_{\mu} P μ is supported on a subset of Z r , p σ , β \mathcal{Z}_{r,p}^{\sigma,\beta} Z r , p σ , β , so that realizations have some spatial Besov regularity. Support on C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) C^{0}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})) C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) is immediate from (i), so that we focus on showing that P μ P_{\mu} P μ is supported on the measurable subset L r ( [ 0 , T ] , B p , ∞ σ − β ( T d ) ) L^{r}([0,T],B_{p,\infty}^{\sigma-\beta}(\mathbb{T}^{d})) L r ([ 0 , T ] , B p , ∞ σ − β ( T d )) of L r ( [ 0 , T ] , L p ( T d ) ) L^{r}([0,T],L^{p}(\mathbb{T}^{d})) L r ([ 0 , T ] , L p ( T d )) . Note that measurability follows from Lemma 1. We proceed analogously to Theorem IV.5.1 of [28 ] , making use of their simple lemma:
Lemma II.3.2 of [28 ] : Let Ω \Omega Ω be a metric space and measures μ \mu μ , μ n \mu_n μ n , n ∈ N n\in\mathbb{N} n ∈ N be defined on B ( Ω ) \mathcal{B}(\Omega) B ( Ω ) and μ n → μ \mu_n\to\mu μ n → μ weakly as n → ∞ n\to\infty n → ∞ . Let f ( u ) f(u) f ( u ) be a continuous function on Ω \Omega Ω and sup n ∫ ∣ f ( u ) ∣ d μ n ( u ) ≤ C \sup_n\int|f(u)|d\mu_n(u)\le C sup n ∫ ∣ f ( u ) ∣ d μ n ( u ) ≤ C . Then, ∫ ∣ f ( u ) ∣ d μ ( u ) ≤ C \int|f(u)|d\mu(u)\le C ∫ ∣ f ( u ) ∣ d μ ( u ) ≤ C .
We first claim that u ↦ ∥ π Λ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) \mathbf{u}\mapsto\|\pi_{\Lambda}\mathbf{u}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))} u ↦ ∥ π Λ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) for Λ = 2 m \Lambda=2^{m} Λ = 2 m and m ∈ N m\in\mathbb{N} m ∈ N is continuous on Z r , p σ , β \mathcal{Z}_{r,p}^{\sigma,\beta} Z r , p σ , β . It suffices to observe the following bound by way of the Lizorkin representation:
∥ π Λ u ∥ B p , ∞ σ ( T d ) = sup j < m ( ν ) 2 σ j ∥ ∑ k ∈ C j ( F u ) ( k ) exp ( i ⟨ k , ⋅ ⟩ ) ∥ L p ( T d ) \|\pi_{\Lambda}\mathbf{u}\|_{B_{p,\infty}^{\sigma}(\mathbb{T}^{d})}
=\sup_{j<m(\nu)}2^{\sigma j}\left\|\sum_{k\in C_j}(\mathcal{F}\mathbf{u})(k)\exp(i\langle k,\cdot\rangle)\right\|_{L^{p}(\mathbb{T}^{d})} ∥ π Λ u ∥ B p , ∞ σ ( T d ) = j < m ( ν ) sup 2 σ j k ∈ C j ∑ ( F u ) ( k ) exp ( i ⟨ k , ⋅ ⟩) L p ( T d )
≤ 2 σ m ( ν ) ∥ [ ∑ j < m ( ν ) ∣ ∑ k ∈ C j ( F u ) ( k ) exp ( i ⟨ k , ⋅ ⟩ ) ∣ 2 ] 1 / 2 ∥ L p ( T d ) \le2^{\sigma m(\nu)}
\left\|\left[\sum_{j<m(\nu)}\left|\sum_{k\in C_j}(\mathcal{F}\mathbf{u})(k)\exp(i\langle k,\cdot\rangle)\right|^{2}\right]^{1/2}\right\|_{L^{p}(\mathbb{T}^{d})} ≤ 2 σ m ( ν ) j < m ( ν ) ∑ k ∈ C j ∑ ( F u ) ( k ) exp ( i ⟨ k , ⋅ ⟩) 2 1/2 L p ( T d )
≤ 2 σ m ( ν ) ∥ u ∥ F p , 2 0 ( T d ) ≲ ∥ u ∥ L p ( T d ) . \le2^{\sigma m(\nu)}\|\mathbf{u}\|_{F_{p,2}^{0}(\mathbb{T}^{d})}\lesssim\|\mathbf{u}\|_{L^{p}(\mathbb{T}^{d})}. ≤ 2 σ m ( ν ) ∥ u ∥ F p , 2 0 ( T d ) ≲ ∥ u ∥ L p ( T d ) .
The last estimate follows by the Paley–Littlewood theorem (Remark 2 of 3.5.4 of [94 ] ).
Now by Lemma 5 the map m ↦ ∥ π 2 m u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) m\mapsto\|\pi_{2^{m}}\mathbf{u}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))} m ↦ ∥ π 2 m u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) is nondecreasing. Fixing m ∈ N + m\in\mathbb{N}^{+} m ∈ N + and setting Λ = 2 m \Lambda=2^{m} Λ = 2 m , we have:
∫ ∥ π Λ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) d P μ ν ( u ) ≤ sup ν > 0 ∫ ∥ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) d P μ ν ( u ) = : M . \int\|\pi_{\Lambda}\mathbf{u}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}\,dP_{\mu}^{\nu}(\mathbf{u})
\le\sup_{\nu>0}\int\|\mathbf{u}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}\,dP_{\mu}^{\nu}(\mathbf{u})=:M. ∫ ∥ π Λ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) d P μ ν ( u ) ≤ ν > 0 sup ∫ ∥ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) d P μ ν ( u ) =: M .
Then by Lemma II.3.2 of [28 ] we have for any weak limit measure
∫ ∥ π Λ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) d P μ ( u ) ≤ M . \int\|\pi_{\Lambda}\mathbf{u}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}\,dP_{\mu}(\mathbf{u})\le M. ∫ ∥ π Λ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) d P μ ( u ) ≤ M .
Finally, a Beppo–Levi monotone convergence argument gives us
∫ ∥ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) d P μ ( u ) = lim Λ → ∞ ∫ ∥ π Λ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) d P μ ( u ) ≤ M . \int\|\mathbf{u}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}\,dP_{\mu}(\mathbf{u})
=\lim_{\Lambda\to\infty}\int\|\pi_{\Lambda}\mathbf{u}\|_{L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))}\,dP_{\mu}(\mathbf{u})\le M. ∫ ∥ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) d P μ ( u ) = Λ → ∞ lim ∫ ∥ π Λ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) d P μ ( u ) ≤ M .
We conclude that P μ P_{\mu} P μ is supported on Z ~ r , p σ , β : = L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) ∩ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \widetilde{\mathcal{Z}}_{r,p}^{\sigma,\beta}:=L^{r}([0,T],B_{p,\infty}^{\sigma}(\mathbb{T}^{d}))\cap C^{0}([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})) Z r , p σ , β := L r ([ 0 , T ] , B p , ∞ σ ( T d )) ∩ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) .
We next show that P μ P_\mu P μ is supported on zero-mean functions, i.e.
P μ ( { u : u ∈ Z r , p σ , β and ∫ T d u ( x , t ) d d x = 0 for every t ∈ [ 0 , T ] } ) = 1. P_\mu\left(\left\{\mathbf{u}:\mathbf{u}\in Z_{r,p}^{\sigma,\beta}\ \text{and}\ \int_{\mathbb{T}^d}\mathbf{u}(\mathbf{x},t)\,\mathrm{d}^d x=0\ \text{for every }t\in[0,T]\right\}\right)=1. P μ ( { u : u ∈ Z r , p σ , β and ∫ T d u ( x , t ) d d x = 0 for every t ∈ [ 0 , T ] } ) = 1.
The map
u ↦ min ( max t ∈ [ 0 , T ] ∣ ∫ T d u ( x , t ) d d x ∣ , 1 ) \mathbf{u}\mapsto\min\left(\max_{t\in[0,T]}\left|\int_{\mathbb{T}^d}\mathbf{u}(\mathbf{x},t)\,\mathrm{d}^d x\right|,1\right) u ↦ min ( t ∈ [ 0 , T ] max ∫ T d u ( x , t ) d d x , 1 )
is continuous on Z r , p σ , β ⊂ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) Z_{r,p}^{\sigma,\beta}\subset C^0([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)) Z r , p σ , β ⊂ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) since 1 ∈ B ( p − 2 ) / p , 1 β − σ ( T d ) 1\in B_{(p-2)/p,1}^{\beta-\sigma}(\mathbb{T}^d) 1 ∈ B ( p − 2 ) / p , 1 β − σ ( T d ) , the Banach dual of B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d) B p /2 , ∞ σ − β ( T d ) [Section 3.5.6 of [94 ] ]. The map is also evidently bounded. Thus we apply part (i) and observe taking ν → 0 \nu\to0 ν → 0
0 = ∫ min ( max t ∈ [ 0 , T ] ∣ ∫ T d u ( x , t ) d d x ∣ , 1 ) d P μ ν ( u ) → ∫ min ( max t ∈ [ 0 , T ] ∣ ∫ T d u ( x , t ) d d x ∣ , 1 ) d P μ ( u ) = 0. 0=\int\min\left(\max_{t\in[0,T]}\left|\int_{\mathbb{T}^d}\mathbf{u}(\mathbf{x},t)\,\mathrm{d}^d x\right|,1\right)\,\mathrm{d}P_\mu^\nu(\mathbf{u})\to\int\min\left(\max_{t\in[0,T]}\left|\int_{\mathbb{T}^d}\mathbf{u}(\mathbf{x},t)\,\mathrm{d}^d x\right|,1\right)\,\mathrm{d}P_\mu(\mathbf{u})=0. 0 = ∫ min ( t ∈ [ 0 , T ] max ∫ T d u ( x , t ) d d x , 1 ) d P μ ν ( u ) → ∫ min ( t ∈ [ 0 , T ] max ∫ T d u ( x , t ) d d x , 1 ) d P μ ( u ) = 0.
The integrand on the RHS is nonnegative so that we conclude that for P μ P_\mu P μ -a.e. u \mathbf{u} u
min ( max t ∈ [ 0 , T ] ∣ ∫ T d u ( x , t ) d d x ∣ , 1 ) = 0 ⟹ ∀ t ∈ [ 0 , T ] : ∫ T d u ( x , t ) d d x = 0. \min\left(\max_{t\in[0,T]}\left|\int_{\mathbb{T}^d}\mathbf{u}(\mathbf{x},t)\,\mathrm{d}^d x\right|,1\right)=0
\qquad\Longrightarrow\qquad
\forall t\in[0,T]:\quad\int_{\mathbb{T}^d}\mathbf{u}(\mathbf{x},t)\,\mathrm{d}^d x=0. min ( t ∈ [ 0 , T ] max ∫ T d u ( x , t ) d d x , 1 ) = 0 ⟹ ∀ t ∈ [ 0 , T ] : ∫ T d u ( x , t ) d d x = 0.
Proof of (iii)
We prove here that realizations of the limiting measure are weak solutions of the incompressible Euler equations. We follow the basic strategy of [28 ] , who constructed space-time statistical solutions of incompressible Navier-Stokes, using in their Theorem IV.5.2 the idea of a “Leray residual”. Similarly, we define here for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , f ∈ C 0 ( [ 0 , T ] × T d ) \mathbf{f}\in C^0([0,T]\times\mathbb{T}^d) f ∈ C 0 ([ 0 , T ] × T d ) , the Euler residual E t f : Z r , p σ , β × S ( T d ) → R E_t^{\mathbf{f}}:Z_{r,p}^{\sigma,\beta}\times\mathcal{S}(\mathbb{T}^d)\to\mathbb{R} E t f : Z r , p σ , β × S ( T d ) → R
E t f ( u , φ ) : = ⟨ u ( ⋅ , t ) , φ ⟩ − ⟨ u ( ⋅ , 0 ) , φ ⟩ + ∫ 0 t ⟨ P ∇ ⋅ ( u ⊗ u ) ( ⋅ , s ) , φ ⟩ d s − ∫ 0 t ⟨ P f ( ⋅ , s ) , φ ⟩ d s . (34) E_t^{\mathbf{f}}(\mathbf{u},\varphi):=\langle\mathbf{u}(\cdot,t),\varphi\rangle-\langle\mathbf{u}(\cdot,0),\varphi\rangle+\int_0^t\langle\mathcal{P}\nabla\cdot(\mathbf{u}\otimes\mathbf{u})(\cdot,s),\varphi\rangle\,\mathrm{d}s-\int_0^t\langle\mathcal{P}\mathbf{f}(\cdot,s),\varphi\rangle\,\mathrm{d}s. \tag*{(34)} E t f ( u , φ ) := ⟨ u ( ⋅ , t ) , φ ⟩ − ⟨ u ( ⋅ , 0 ) , φ ⟩ + ∫ 0 t ⟨ P ∇ ⋅ ( u ⊗ u ) ( ⋅ , s ) , φ ⟩ d s − ∫ 0 t ⟨ P f ( ⋅ , s ) , φ ⟩ d s . (34)
The conditions
E t f ( u , φ ) = 0 , ∀ t ∈ [ 0 , T ] , ∀ φ ∈ S ( T d ) (35) E_t^{\mathbf{f}}(\mathbf{u},\varphi)=0,\qquad\forall t\in[0,T],\ \forall\varphi\in\mathcal{S}(\mathbb{T}^d) \tag*{(35)} E t f ( u , φ ) = 0 , ∀ t ∈ [ 0 , T ] , ∀ φ ∈ S ( T d ) (35)
thus state that the velocity field u \mathbf{u} u is a solution of Euler equations in a suitable weak sense, sometimes termed a “mild solution.” Note that above and hereafter we use S ( T d ) \mathcal{S}(\mathbb{T}^d) S ( T d ) as a simplified notation for S ( T d , R d ) \mathcal{S}(\mathbb{T}^d,\mathbb{R}^d) S ( T d , R d ) . The dual product in the first two terms of (34) may be taken as an element of B p / 2 , ∞ σ − β ( T d ) = ( B 2 / ( 2 − p ) , 1 β − σ ( T d ) ) ∗ B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)=(B_{2/(2-p),1}^{\beta-\sigma}(\mathbb{T}^d))^* B p /2 , ∞ σ − β ( T d ) = ( B 2/ ( 2 − p ) , 1 β − σ ( T d ) ) ∗ acting on an element of its predual, since S ( T d ) ⊆ B 2 / ( 2 − p ) , 1 β − σ ( T d ) \mathcal{S}(\mathbb{T}^d)\subseteq B_{2/(2-p),1}^{\beta-\sigma}(\mathbb{T}^d) S ( T d ) ⊆ B 2/ ( 2 − p ) , 1 β − σ ( T d ) . The same is true in the fourth term, where ⟨ f ( ⋅ , s ) , P φ ⟩ \langle\mathbf{f}(\cdot,s),\mathcal{P}\varphi\rangle ⟨ f ( ⋅ , s ) , P φ ⟩ is then continuous in s s s , so that the integral is well-defined. Finally, for the third term, we note from (32) that ( F P φ ) ( k ) (\mathcal{F}\mathcal{P}\varphi)(\mathbf{k}) ( F P φ ) ( k ) decays faster than any inverse power of ∣ k ∣ |\mathbf{k}| ∣ k ∣ if φ ∈ S ( T d ) \varphi\in\mathcal{S}(\mathbb{T}^d) φ ∈ S ( T d ) . Hence, P φ ∈ S ( T d ) \mathcal{P}\varphi\in\mathcal{S}(\mathbb{T}^d) P φ ∈ S ( T d ) and in particular ∇ P φ \nabla\mathcal{P}\varphi ∇ P φ is a bounded function on T d \mathbb{T}^d T d . Thus, this term is well-defined for any element u ∈ L r ( [ 0 , T ] , L p ( T d ) ) ⊆ Z r , p σ , β \mathbf{u}\in L^r([0,T],L^p(\mathbb{T}^d))\subseteq Z_{r,p}^{\sigma,\beta} u ∈ L r ([ 0 , T ] , L p ( T d )) ⊆ Z r , p σ , β .
We note some basic properties of the Euler residual that are needed below:
Lemma 6. The following continuity and boundedness properties hold for all f ∈ C 0 ( [ 0 , T ] × T d ) \mathbf{f}\in C^0([0,T]\times\mathbb{T}^d) f ∈ C 0 ([ 0 , T ] × T d ) :
(i) For all fixed t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , φ ∈ S ( T d ) \varphi\in\mathcal{S}(\mathbb{T}^d) φ ∈ S ( T d ) , the map u ↦ E t f ( u , φ ) \mathbf{u}\mapsto E_t^{\mathbf{f}}(\mathbf{u},\varphi) u ↦ E t f ( u , φ ) is strongly continuous on Z r , p σ , β Z_{r,p}^{\sigma,\beta} Z r , p σ , β .
(ii) For all fixed t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , u ∈ Z r , p σ , β \mathbf{u}\in Z_{r,p}^{\sigma,\beta} u ∈ Z r , p σ , β , the map φ ↦ E t f ( u , φ ) \varphi\mapsto E_t^{\mathbf{f}}(\mathbf{u},\varphi) φ ↦ E t f ( u , φ ) is linear and continuous on S ( T d ) \mathcal{S}(\mathbb{T}^d) S ( T d ) ;
(iii) For all fixed u ∈ Z r , p σ , β \mathbf{u}\in Z_{r,p}^{\sigma,\beta} u ∈ Z r , p σ , β , φ ∈ S ( T d ) \varphi\in\mathcal{S}(\mathbb{T}^d) φ ∈ S ( T d ) , the map t ↦ E t f ( u , φ ) t\mapsto E_t^{\mathbf{f}}(\mathbf{u},\varphi) t ↦ E t f ( u , φ ) is continuous on [ 0 , T ] [0,T] [ 0 , T ] .
(iv) For all fixed t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , φ ∈ S ( T d ) \varphi\in\mathcal{S}(\mathbb{T}^d) φ ∈ S ( T d ) and for some positive constants c c c , c ′ c' c ′ , c ′ ′ c'' c ′′ dependent on those choices
∣ E t f ( u , φ ) ∣ ≤ c + c ′ ∥ u ∥ Z r , p σ , β + c ′ ′ ∥ u ∥ Z r , p σ , β 2 . \left|E_t^{\mathbf{f}}(\mathbf{u},\varphi)\right|\le c+c'\|\mathbf{u}\|_{Z_{r,p}^{\sigma,\beta}}+c''\|\mathbf{u}\|_{Z_{r,p}^{\sigma,\beta}}^2. E t f ( u , φ ) ≤ c + c ′ ∥ u ∥ Z r , p σ , β + c ′′ ∥ u ∥ Z r , p σ , β 2 .
Proof. (i) Consider a sequence u n → u \mathbf{u}_n\to\mathbf{u} u n → u strong in Z r , p σ , β Z_{r,p}^{\sigma,\beta} Z r , p σ , β . Noting Z r , p σ , β ⊂ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) Z_{r,p}^{\sigma,\beta}\subset C^0([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)) Z r , p σ , β ⊂ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) , ⟨ u n ( ⋅ , s ) , φ ⟩ → ⟨ u ( ⋅ , s ) , φ ⟩ \langle\mathbf{u}_n(\cdot,s),\varphi\rangle\to\langle\mathbf{u}(\cdot,s),\varphi\rangle ⟨ u n ( ⋅ , s ) , φ ⟩ → ⟨ u ( ⋅ , s ) , φ ⟩ for any s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] and φ ∈ S ( T d ) ⊆ B 2 / ( 2 − p ) , 1 β − σ ( T d ) \varphi\in\mathcal{S}(\mathbb{T}^d)\subseteq B_{2/(2-p),1}^{\beta-\sigma}(\mathbb{T}^d) φ ∈ S ( T d ) ⊆ B 2/ ( 2 − p ) , 1 β − σ ( T d ) , since strong convergence in B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d) B p /2 , ∞ σ − β ( T d ) implies weak-* convergence and uniform convergence in [ 0 , T ] [0,T] [ 0 , T ] implies pointwise convergence. For the third term, we observe the inequalities
∣ ∫ 0 t ⟨ P ∇ ⋅ ( u n ⊗ u n − u ⊗ u ) ( ⋅ , s ) , φ ⟩ d s ∣ = ∣ ∫ 0 t ⟨ u n ⊗ u n − u ⊗ u , ∇ P φ ⟩ d s ∣ = ∣ ∫ 0 t ⟨ u n ⊗ ( u n − u ) − ( u − u n ) ⊗ u , ∇ P φ ⟩ d s ∣ \begin{aligned}
\left|\int_0^t \langle\mathscr{P}\nabla\cdot(\mathbf{u}_n\otimes\mathbf{u}_n-\mathbf{u}\otimes\mathbf{u})(\cdot,s),\varphi\rangle\,\mathrm{d}s\right|
=\left|\int_0^t\langle\mathbf{u}_n\otimes\mathbf{u}_n-\mathbf{u}\otimes\mathbf{u},\nabla\mathscr{P}\varphi\rangle\,\mathrm{d}s\right| \\
&=\left|\int_0^t\langle\mathbf{u}_n\otimes(\mathbf{u}_n-\mathbf{u})-(\mathbf{u}-\mathbf{u}_n)\otimes\mathbf{u},\nabla\mathscr{P}\varphi\rangle\,\mathrm{d}s\right|
\end{aligned} ∫ 0 t ⟨ P ∇ ⋅ ( u n ⊗ u n − u ⊗ u ) ( ⋅ , s ) , φ ⟩ d s = ∫ 0 t ⟨ u n ⊗ u n − u ⊗ u , ∇ P φ ⟩ d s = ∫ 0 t ⟨ u n ⊗ ( u n − u ) − ( u − u n ) ⊗ u , ∇ P φ ⟩ d s
≲ ∫ 0 t ∥ u n ∥ L 2 ( T d ) ∥ u n − u ∥ L 2 ( T d ) d s + ∫ 0 t ∥ u n − u ∥ L 2 ( T d ) ∥ u ∥ L 2 ( T d ) d s ≤ ∥ u n ∥ L r ( [ 0 , T ] , L p ( T d ) ) ∥ u n − u ∥ L r ( [ 0 , T ] , L p ( T d ) ) + ∥ u − u n ∥ L r ( [ 0 , T ] , L p ( T d ) ) ∥ u ∥ L r ( [ 0 , T ] , L p ( T d ) ) \begin{aligned}
\lesssim\int_0^t\|\mathbf{u}_n\|_{L^2(\mathbb{T}^d)}\|\mathbf{u}_n-\mathbf{u}\|_{L^2(\mathbb{T}^d)}\,\mathrm{d}s
+\int_0^t\|\mathbf{u}_n-\mathbf{u}\|_{L^2(\mathbb{T}^d)}\|\mathbf{u}\|_{L^2(\mathbb{T}^d)}\,\mathrm{d}s \\
&\leq\|\mathbf{u}_n\|_{L^r([0,T],L^p(\mathbb{T}^d))}\|\mathbf{u}_n-\mathbf{u}\|_{L^r([0,T],L^p(\mathbb{T}^d))}
+\|\mathbf{u}-\mathbf{u}_n\|_{L^r([0,T],L^p(\mathbb{T}^d))}\|\mathbf{u}\|_{L^r([0,T],L^p(\mathbb{T}^d))}
\end{aligned} ≲ ∫ 0 t ∥ u n ∥ L 2 ( T d ) ∥ u n − u ∥ L 2 ( T d ) d s + ∫ 0 t ∥ u n − u ∥ L 2 ( T d ) ∥ u ∥ L 2 ( T d ) d s ≤ ∥ u n ∥ L r ([ 0 , T ] , L p ( T d )) ∥ u n − u ∥ L r ([ 0 , T ] , L p ( T d )) + ∥ u − u n ∥ L r ([ 0 , T ] , L p ( T d )) ∥ u ∥ L r ([ 0 , T ] , L p ( T d ))
for r , p > 2 r,p>2 r , p > 2 . Recalling Z r , p σ , β ⊆ L r ( [ 0 , T ] , L p ( T d ) ) Z^{\sigma,\beta}_{r,p}\subseteq L^r([0,T],L^p(\mathbb{T}^d)) Z r , p σ , β ⊆ L r ([ 0 , T ] , L p ( T d )) , the expression above vanishes in the limit of n → ∞ n\to\infty n → ∞ . The last term of (34) is constant in u \mathbf{u} u and so trivially continuous.
(ii) Linearity is obvious. Recall that φ n → φ \varphi_n\to\varphi φ n → φ in S ( R d ) \mathcal{S}(\mathbb{R}^d) S ( R d ) if the sequence of functions and all of their derivatives converge uniformly on T d \mathbb{T}^d T d . Convergence of the first two terms of (34) is immediate from the fact that B p / 2 , ∞ σ − β ( T d ) ⊆ S ′ ( T d ) B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)\subseteq\mathcal{S}'(\mathbb{T}^d) B p /2 , ∞ σ − β ( T d ) ⊆ S ′ ( T d ) . For the third term, similarly as in (i), we note that
∣ ∫ 0 t ⟨ P ∇ ⋅ ( u ⊗ u ) ( ⋅ , s ) , φ − φ n ⟩ d s ∣ = ∣ ∫ 0 t ⟨ u ⊗ u , ∇ P ( φ − φ n ) ⟩ d s ∣ ≤ ∫ 0 t ∣ u ⊗ u ∣ d s ⋅ ∥ ∇ P ( φ − φ n ) ∥ ∞ ≤ ∥ u ∥ Z r , p σ , β 2 ∥ ∇ P ( φ − φ n ) ∥ ∞ \begin{aligned}
\left|\int_0^t\langle\mathscr{P}\nabla\cdot(\mathbf{u}\otimes\mathbf{u})(\cdot,s),\varphi-\varphi_n\rangle\,\mathrm{d}s\right|
=\left|\int_0^t\langle\mathbf{u}\otimes\mathbf{u},\nabla\mathscr{P}(\varphi-\varphi_n)\rangle\,\mathrm{d}s\right| \\
&\leq\int_0^t|\mathbf{u}\otimes\mathbf{u}|\,\mathrm{d}s\cdot\|\nabla\mathscr{P}(\varphi-\varphi_n)\|_\infty
\leq\|\mathbf{u}\|_{Z^{\sigma,\beta}_{r,p}}^2\|\nabla\mathscr{P}(\varphi-\varphi_n)\|_\infty
\end{aligned} ∫ 0 t ⟨ P ∇ ⋅ ( u ⊗ u ) ( ⋅ , s ) , φ − φ n ⟩ d s = ∫ 0 t ⟨ u ⊗ u , ∇ P ( φ − φ n )⟩ d s ≤ ∫ 0 t ∣ u ⊗ u ∣ d s ⋅ ∥∇ P ( φ − φ n ) ∥ ∞ ≤ ∥ u ∥ Z r , p σ , β 2 ∥∇ P ( φ − φ n ) ∥ ∞
so that this term vanishes as n → ∞ n\to\infty n → ∞ . Given the regularity f ∈ C 0 ( [ 0 , T ] × T d ) \mathbf{f}\in C^0([0,T]\times\mathbb{T}^d) f ∈ C 0 ([ 0 , T ] × T d ) , the argument for the fourth term is very similar to that above.
(iii) Because u ∈ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) \mathbf{u}\in C^0([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)) u ∈ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) and φ ∈ S ( T d ) ⊆ B σ − β 2 ( 2 − p ) , 1 ( T d ) \varphi\in\mathcal{S}(\mathbb{T}^d)\subseteq B^{\frac{\sigma-\beta}{2(2-p)},1}(\mathbb{T}^d) φ ∈ S ( T d ) ⊆ B 2 ( 2 − p ) σ − β , 1 ( T d ) , continuity of ⟨ u ( ⋅ , t ) , φ ⟩ \langle\mathbf{u}(\cdot,t),\varphi\rangle ⟨ u ( ⋅ , t ) , φ ⟩ in t t t is immediate. Regarding the time-integral terms, we note from the preceding estimates that the integrand functions ⟨ ( u ⊗ u ) ( ⋅ , s ) , ∇ P φ ⟩ \langle(\mathbf{u}\otimes\mathbf{u})(\cdot,s),\nabla\mathscr{P}\varphi\rangle ⟨( u ⊗ u ) ( ⋅ , s ) , ∇ P φ ⟩ and ⟨ f ( ⋅ , s ) , P φ ⟩ \langle\mathbf{f}(\cdot,s),\mathscr{P}\varphi\rangle ⟨ f ( ⋅ , s ) , P φ ⟩ are L 1 L^1 L 1 and the integrals are even absolutely continuous.
(iv) The following simple estimates
∣ ⟨ u ( ⋅ , s ) , φ ⟩ ∣ ≲ ∥ u ( ⋅ , s ) ∥ B p / 2 , ∞ σ − β ( T d ) ≤ ∥ u ∥ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) , s = 0 , t |\langle\mathbf{u}(\cdot,s),\varphi\rangle|\lesssim\|\mathbf{u}(\cdot,s)\|_{B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d)}\leq\|\mathbf{u}\|_{C^0([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d))},\qquad s=0,t ∣ ⟨ u ( ⋅ , s ) , φ ⟩ ∣ ≲ ∥ u ( ⋅ , s ) ∥ B p /2 , ∞ σ − β ( T d ) ≤ ∥ u ∥ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) , s = 0 , t
∣ ∫ 0 t ⟨ P ∇ ⋅ ( u ⊗ u ) , φ ⟩ d s ∣ = ∣ ∫ 0 t ⟨ u ⊗ u , ∇ P φ ⟩ d s ∣ ≲ ∫ 0 t ∥ u ∥ L 2 ( T d ) 2 d s ≲ ∥ u ∥ L r ( [ 0 , T ] , L p ( T d ) ) 2 . \left|\int_0^t\langle\mathscr{P}\nabla\cdot(\mathbf{u}\otimes\mathbf{u}),\varphi\rangle\,\mathrm{d}s\right|
=\left|\int_0^t\langle\mathbf{u}\otimes\mathbf{u},\nabla\mathscr{P}\varphi\rangle\,\mathrm{d}s\right|
\lesssim\int_0^t\|\mathbf{u}\|_{L^2(\mathbb{T}^d)}^2\,\mathrm{d}s
\lesssim\|\mathbf{u}\|_{L^r([0,T],L^p(\mathbb{T}^d))}^2. ∫ 0 t ⟨ P ∇ ⋅ ( u ⊗ u ) , φ ⟩ d s = ∫ 0 t ⟨ u ⊗ u , ∇ P φ ⟩ d s ≲ ∫ 0 t ∥ u ∥ L 2 ( T d ) 2 d s ≲ ∥ u ∥ L r ([ 0 , T ] , L p ( T d )) 2 .
explain the terms linear and quadratic in ∥ u ∥ Z r , p σ , β \|\mathbf{u}\|_{Z^{\sigma,\beta}_{r,p}} ∥ u ∥ Z r , p σ , β . The other expressions in (34) yield the constant term. □
Let us denote by C b , B 0 ( Z r , p σ , β , R ) C^0_{b,B}(Z^{\sigma,\beta}_{r,p},\mathbb{R}) C b , B 0 ( Z r , p σ , β , R ) the set of bounded, continuous functions on Z r , p σ , β Z^{\sigma,\beta}_{r,p} Z r , p σ , β whose support in addition is bounded in the Z r , p σ , β Z^{\sigma,\beta}_{r,p} Z r , p σ , β norm. For any η ∈ C b , B 0 ( Z r , p σ , β , R ) \eta\in C^0_{b,B}(Z^{\sigma,\beta}_{r,p},\mathbb{R}) η ∈ C b , B 0 ( Z r , p σ , β , R ) , the mapping u ↦ E t f ( u , φ ) η ( u ) \mathbf{u}\mapsto E_t^f(\mathbf{u},\varphi)\eta(\mathbf{u}) u ↦ E t f ( u , φ ) η ( u ) is C b , B 0 ( Z r , p σ , β , R ) C^0_{b,B}(Z^{\sigma,\beta}_{r,p},\mathbb{R}) C b , B 0 ( Z r , p σ , β , R ) because of Lemma 6 (i)&(iv). Thus from part (i) of the Main Theorem, we have
∫ E t f ( u , φ ) η ( u ) d P μ ν ( u ) → ∫ E t f ( u , φ ) η ( u ) d P μ ( u ) \int E_t^f(\mathbf{u},\varphi)\eta(\mathbf{u})\,\mathrm{d}P_\mu^\nu(\mathbf{u})
\to
\int E_t^f(\mathbf{u},\varphi)\eta(\mathbf{u})\,\mathrm{d}P_\mu(\mathbf{u}) ∫ E t f ( u , φ ) η ( u ) d P μ ν ( u ) → ∫ E t f ( u , φ ) η ( u ) d P μ ( u )
as ν → 0 \nu\to0 ν → 0 . The strategy is to show that the sequence on the left in fact converges to 0, thus yielding
∫ E t f ( u , φ ) η ( u ) d P μ ( u ) = 0. \int E_t^{\mathrm f}(\mathbf{u},\varphi)\eta(\mathbf{u})\,\mathrm{d}P_\mu(\mathbf{u})=0. ∫ E t f ( u , φ ) η ( u ) d P μ ( u ) = 0.
This latter condition for all η ∈ C b , B 0 ( Z r , p σ , β , R ) \eta\in C^0_{b,B}(\mathcal{Z}^{\sigma,\beta}_{r,p},\mathbb{R}) η ∈ C b , B 0 ( Z r , p σ , β , R ) will then be shown to imply that
E t f ( u , φ ) = 0 P μ -a.e. u E_t^{\mathrm f}(\mathbf{u},\varphi)=0 \quad P_\mu\text{-a.e. }\mathbf{u} E t f ( u , φ ) = 0 P μ -a.e. u
These two statements are proved in the next two lemmas.
Lemma 7. For any sequence ν → 0 \nu\to0 ν → 0 and for any fixed t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , φ ∈ S ( R d ) \varphi\in\mathcal{S}(\mathbb{R}^d) φ ∈ S ( R d ) , η ∈ C b , B 0 ( Z r , p σ , β , R ) \eta\in C^0_{b,B}(\mathcal{Z}^{\sigma,\beta}_{r,p},\mathbb{R}) η ∈ C b , B 0 ( Z r , p σ , β , R )
lim ν → 0 ∫ E t f ( u , φ ) η ( u ) d P μ ν ( u ) = 0. \lim_{\nu\to0}\int E_t^{\mathrm f}(\mathbf{u},\varphi)\eta(\mathbf{u})\,\mathrm{d}P_\mu^\nu(\mathbf{u})=0. ν → 0 lim ∫ E t f ( u , φ ) η ( u ) d P μ ν ( u ) = 0.
Proof. The proof is based on the definition of P μ ν P_\mu^\nu P μ ν
∫ E t f ( u , φ ) η ( u ) d P μ ν ( u ) = E μ [ E t f ( u ν , φ ) η ( u ν ) ] \int E_t^{\mathrm f}(\mathbf{u},\varphi)\eta(\mathbf{u})\,\mathrm{d}P_\mu^\nu(\mathbf{u})
=\mathbb{E}_\mu\left[E_t^{\mathrm f}(\mathbf{u}^\nu,\varphi)\eta(\mathbf{u}^\nu)\right] ∫ E t f ( u , φ ) η ( u ) d P μ ν ( u ) = E μ [ E t f ( u ν , φ ) η ( u ν ) ]
and the following expression for the Euler residual of solutions u ν \mathbf{u}^\nu u ν of (5) :
E t f ( u ν , φ ) = ∫ 0 t ⟨ P ∇ ⋅ [ ( u ν ⊗ u ν ) − π Λ ( u ν ⊗ u ν ) ] , φ ⟩ d s + ∫ 0 t ν ⟨ Δ π Λ u ν , φ ⟩ d s + ν κ ∫ 0 t ⟨ P ∇ ⋅ d π Λ ξ ( ⋅ , s ) , φ ⟩ d s − ∫ 0 t ⟨ f − π Λ f , φ ⟩ d s . (36) \begin{aligned}
E_t^{\mathrm f}(\mathbf{u}^\nu,\varphi)
={}&\int_0^t\langle\mathscr{P}\nabla\cdot[(\mathbf{u}^\nu\otimes\mathbf{u}^\nu)-\pi_\Lambda(\mathbf{u}^\nu\otimes\mathbf{u}^\nu)],\varphi\rangle\,\mathrm{d}s\\
&+\int_0^t\nu\langle\Delta\pi_\Lambda\mathbf{u}^\nu,\varphi\rangle\,\mathrm{d}s\\
&+\nu^\kappa\int_0^t\langle\mathscr{P}\nabla\cdot\mathrm{d}\pi_\Lambda\xi(\cdot,s),\varphi\rangle\,\mathrm{d}s-\int_0^t\langle\mathbf{f}-\pi_\Lambda\mathbf{f},\varphi\rangle\,\mathrm{d}s. \tag*{(36)}
\end{aligned} E t f ( u ν , φ ) = ∫ 0 t ⟨ P ∇ ⋅ [( u ν ⊗ u ν ) − π Λ ( u ν ⊗ u ν )] , φ ⟩ d s + ∫ 0 t ν ⟨ Δ π Λ u ν , φ ⟩ d s + ν κ ∫ 0 t ⟨ P ∇ ⋅ d π Λ ξ ( ⋅ , s ) , φ ⟩ d s − ∫ 0 t ⟨ f − π Λ f , φ ⟩ d s . (36)
In fact, the function η ∈ C b , B 0 ( Z r , p σ , β , R ) \eta\in C^0_{b,B}(\mathcal{Z}^{\sigma,\beta}_{r,p},\mathbb{R}) η ∈ C b , B 0 ( Z r , p σ , β , R ) plays no essential role in any estimates, aside from its boundedness. For the first term on the RHS of (36) we have
∣ ∫ 0 t ⟨ P ∇ ⋅ [ ( u ν ⊗ u ν ) − π Λ ( u ν ⊗ u ν ) ] , φ ⟩ d s ∣ = ∣ ∫ 0 t ⟨ ( u ν ⊗ u ν ) − π Λ ( u ν ⊗ u ν ) , ∇ P φ ⟩ d s ∣ ≲ ∥ ( u ν ⊗ u ν ) − π Λ ( u ν ⊗ u ν ) ∥ L r / 2 ( [ 0 , T ] , L p / 2 ( T d ) ) . \left|\int_0^t\left\langle\mathscr{P}\nabla\cdot[(\mathbf{u}^\nu\otimes\mathbf{u}^\nu)-\pi_\Lambda(\mathbf{u}^\nu\otimes\mathbf{u}^\nu)],\varphi\right\rangle\,\mathrm{d}s\right|
=\left|\int_0^t\left\langle(\mathbf{u}^\nu\otimes\mathbf{u}^\nu)-\pi_\Lambda(\mathbf{u}^\nu\otimes\mathbf{u}^\nu),\nabla\mathscr{P}\varphi\right\rangle\,\mathrm{d}s\right|
\lesssim\left\|(\mathbf{u}^\nu\otimes\mathbf{u}^\nu)-\pi_\Lambda(\mathbf{u}^\nu\otimes\mathbf{u}^\nu)\right\|_{L^{r/2}([0,T],L^{p/2}(\mathbb{T}^d))}. ∫ 0 t ⟨ P ∇ ⋅ [( u ν ⊗ u ν ) − π Λ ( u ν ⊗ u ν )] , φ ⟩ d s = ∫ 0 t ⟨ ( u ν ⊗ u ν ) − π Λ ( u ν ⊗ u ν ) , ∇ P φ ⟩ d s ≲ ∥ ( u ν ⊗ u ν ) − π Λ ( u ν ⊗ u ν ) ∥ L r /2 ([ 0 , T ] , L p /2 ( T d )) .
By the Lizorkin representation
∥ ( u ⊗ u ) − π Λ ( u ⊗ u ) ∥ L r / 2 ( [ 0 , T ] , L p / 2 ( T d ) ) ≤ 2 − σ m ( ν ) ∥ u ⊗ u ∥ L r / 2 ( [ 0 , T ] , B p / 2 , ∞ σ ( T d ) ) ≲ 2 − σ m ( ν ) ∥ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 (37) \left\|(\mathbf{u}\otimes\mathbf{u})-\pi_\Lambda(\mathbf{u}\otimes\mathbf{u})\right\|_{L^{r/2}([0,T],L^{p/2}(\mathbb{T}^d))}
\le2^{-\sigma m(\nu)}\left\|\mathbf{u}\otimes\mathbf{u}\right\|_{L^{r/2}([0,T],B^{\sigma}_{p/2,\infty}(\mathbb{T}^d))}
\lesssim2^{-\sigma m(\nu)}\left\|\mathbf{u}\right\|^2_{L^r([0,T],B^{\sigma}_{p,\infty}(\mathbb{T}^d))} \tag*{(37)} ∥ ( u ⊗ u ) − π Λ ( u ⊗ u ) ∥ L r /2 ([ 0 , T ] , L p /2 ( T d )) ≤ 2 − σ m ( ν ) ∥ u ⊗ u ∥ L r /2 ([ 0 , T ] , B p /2 , ∞ σ ( T d )) ≲ 2 − σ m ( ν ) ∥ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) 2 (37)
This is sufficient to get
∣ E μ [ η ( u ν ) [ ∫ 0 t ⟨ P ∇ ⋅ [ ( u ν ⊗ u ν ) − π Λ ( u ν ⊗ u ν ) ] , φ ⟩ d s ] ] ∣ ≲ 2 − σ m ( ν ) ∫ ∥ u ∥ L r ( [ 0 , T ] , B p , ∞ σ ( T d ) ) 2 d P μ ν ( u ) ⟶ 0. (38) \left|\mathbb{E}_\mu\left[\eta(\mathbf{u}^\nu)\left[\int_0^t\left\langle\mathscr{P}\nabla\cdot[(\mathbf{u}^\nu\otimes\mathbf{u}^\nu)-\pi_\Lambda(\mathbf{u}^\nu\otimes\mathbf{u}^\nu)],\varphi\right\rangle\,\mathrm{d}s\right]\right]\right|
\lesssim2^{-\sigma m(\nu)}\int\left\|\mathbf{u}\right\|^2_{L^r([0,T],B^\sigma_{p,\infty}(\mathbb{T}^d))}\,\mathrm{d}P_\mu^\nu(\mathbf{u})
\longrightarrow0. \tag*{(38)} E μ [ η ( u ν ) [ ∫ 0 t ⟨ P ∇ ⋅ [( u ν ⊗ u ν ) − π Λ ( u ν ⊗ u ν )] , φ ⟩ d s ] ] ≲ 2 − σ m ( ν ) ∫ ∥ u ∥ L r ([ 0 , T ] , B p , ∞ σ ( T d )) 2 d P μ ν ( u ) ⟶ 0. (38)
We can similarly write
∣ ∫ 0 t ⟨ Δ π Λ u ν , φ ⟩ d s ∣ ≲ ∥ u ν ∥ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) ∥ Δ π Λ φ ∥ B ( p − 2 ) / p , 1 β − σ ( T d ) . \left|\int_0^t\langle\Delta\pi_\Lambda\mathbf{u}^\nu,\varphi\rangle\,\mathrm{d}s\right|
\lesssim\left\|\mathbf{u}^\nu\right\|_{C^0([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d))}\left\|\Delta\pi_\Lambda\varphi\right\|_{B^{\beta-\sigma}_{(p-2)/p,1}(\mathbb{T}^d)}. ∫ 0 t ⟨ Δ π Λ u ν , φ ⟩ d s ≲ ∥ u ν ∥ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) ∥ Δ π Λ φ ∥ B ( p − 2 ) / p , 1 β − σ ( T d ) .
Thus for the second term on the RHS of (36) we have
∣ E μ [ η ( u ν ) [ ν ∫ ∫ 0 t ⟨ Δ π Λ u ν , φ ⟩ d s ] ] ∣ ≲ ν ∫ ∥ u ∥ C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) d P μ ν ( u ) ⟶ 0 , \left|\mathbb{E}_\mu\left[\eta(\mathbf{u}^\nu)\left[\nu\int\int_0^t\langle\Delta\pi_\Lambda\mathbf{u}^\nu,\varphi\rangle\,\mathrm{d}s\right]\right]\right|
\lesssim\nu\int\left\|\mathbf{u}\right\|_{C^0([0,T],B^{\sigma-\beta}_{p/2,\infty}(\mathbb{T}^d))}\,\mathrm{d}P_\mu^\nu(\mathbf{u})\longrightarrow0, E μ [ η ( u ν ) [ ν ∫ ∫ 0 t ⟨ Δ π Λ u ν , φ ⟩ d s ] ] ≲ ν ∫ ∥ u ∥ C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) d P μ ν ( u ) ⟶ 0 ,
The integral on the righthand side is finite by the tightness estimate in part (i) of the theorem and the whole term vanishes due to the coefficient ν \nu ν .
For the third term of (36) we have
∫ 0 t ⟨ P ∇ ⋅ π Λ d ξ ( ⋅ , s ) , φ ⟩ = − ∫ 0 t ⟨ π Λ d ξ ( ⋅ , s ) , ∇ P φ ⟩ \int_{0}^{t}\langle\mathcal{P}\nabla\cdot\pi_{\Lambda}\mathrm{d}\xi(\cdot,s),\varphi\rangle=-\int_{0}^{t}\langle\pi_{\Lambda}\mathrm{d}\xi(\cdot,s),\nabla\mathcal{P}\varphi\rangle ∫ 0 t ⟨ P ∇ ⋅ π Λ d ξ ( ⋅ , s ) , φ ⟩ = − ∫ 0 t ⟨ π Λ d ξ ( ⋅ , s ) , ∇ P φ ⟩
Of course, from Cauchy-Schwartz
∣ E [ ∫ 0 t ⟨ π Λ d ξ ( ⋅ , s ) , ∇ P φ ⟩ ] ∣ ≤ ( E [ ∫ 0 t ⟨ π Λ d ξ ( ⋅ , s ) , ∇ P φ ⟩ ] 2 ) 1 / 2 . \left|\mathbb{E}\left[\int_{0}^{t}\langle\pi_{\Lambda}\mathrm{d}\xi(\cdot,s),\nabla\mathcal{P}\varphi\rangle\right]\right|\leq\left(\mathbb{E}\left[\int_{0}^{t}\langle\pi_{\Lambda}\mathrm{d}\xi(\cdot,s),\nabla\mathcal{P}\varphi\rangle\right]^{2}\right)^{1/2}. E [ ∫ 0 t ⟨ π Λ d ξ ( ⋅ , s ) , ∇ P φ ⟩ ] ≤ ( E [ ∫ 0 t ⟨ π Λ d ξ ( ⋅ , s ) , ∇ P φ ⟩ ] 2 ) 1/2 .
Then using the Fourier representation of ξ \xi ξ , a simple calculation by Itô isometry yields:
E [ ∣ ∫ 0 t ⟨ π Λ d ξ ( ⋅ , s ) , ∇ P φ ⟩ ∣ 2 ] ≤ E [ ∣ ∫ 0 t ∑ ∣ k ∣ ∞ < 2 m ( ν ) ∑ i , j , ℓ = 1 d P j ℓ ( k ) k i ( F φ ) ℓ ( k ) d ξ k i j ( s ) ∣ 2 ] = ∑ ∣ k ∣ ∞ < 2 m ( ν ) ∫ 0 t ∣ ∑ i , j , ℓ = 1 d [ P j ℓ ( k ) k i + P i ℓ ( k ) k j ] ( F φ ) ℓ ( k ) ∣ 2 d s \begin{aligned}
\mathbb{E}\left[\left|\int_{0}^{t}\langle\pi_{\Lambda}\mathrm{d}\xi(\cdot,s),\nabla\mathcal{P}\varphi\rangle\right|^{2}\right]\leq\mathbb{E}\left[\left|\int_{0}^{t}\sum_{|k|_{\infty}<2^{m(\nu)}}\sum_{i,j,\ell=1}^{d}\mathcal{P}_{j\ell}(k)k_{i}(\mathcal{F}\varphi)_{\ell}(k)\mathrm{d}\xi_{k}^{ij}(s)\right|^{2}\right] \\
&=\sum_{|k|_{\infty}<2^{m(\nu)}}\int_{0}^{t}\left|\sum_{i,j,\ell=1}^{d}\left[\mathcal{P}_{j\ell}(k)k_{i}+\mathcal{P}_{i\ell}(k)k_{j}\right](\mathcal{F}\varphi)_{\ell}(k)\right|^{2}\mathrm{d}s
\end{aligned} E [ ∫ 0 t ⟨ π Λ d ξ ( ⋅ , s ) , ∇ P φ ⟩ 2 ] ≤ E ∫ 0 t ∣ k ∣ ∞ < 2 m ( ν ) ∑ i , j , ℓ = 1 ∑ d P j ℓ ( k ) k i ( F φ ) ℓ ( k ) d ξ k ij ( s ) 2 = ∣ k ∣ ∞ < 2 m ( ν ) ∑ ∫ 0 t i , j , ℓ = 1 ∑ d [ P j ℓ ( k ) k i + P i ℓ ( k ) k j ] ( F φ ) ℓ ( k ) 2 d s
≲ ∑ ∣ k ∣ ∞ < 2 m ( ν ) ∣ k ∣ 2 ∣ ( F φ ) ( k ) ∣ 2 ≲ ∥ ∇ φ ∥ L 2 ( T d ) 2 . \lesssim\sum_{|k|_{\infty}<2^{m(\nu)}}|k|^{2}|(\mathcal{F}\varphi)(k)|^{2}\lesssim\|\nabla\varphi\|_{L^{2}(\mathbb{T}^{d})}^{2}. ≲ ∣ k ∣ ∞ < 2 m ( ν ) ∑ ∣ k ∣ 2 ∣ ( F φ ) ( k ) ∣ 2 ≲ ∥∇ φ ∥ L 2 ( T d ) 2 .
Thus evidently
E μ [ η ( u ν ) ν κ [ ∫ 0 t ⟨ P ∇ ⋅ π Λ d ξ ( ⋅ , s ) , φ ⟩ ] ] → 0. \mathbb{E}_{\mu}\left[\eta(u^{\nu})\nu^{\kappa}\left[\int_{0}^{t}\langle\mathcal{P}\nabla\cdot\pi_{\Lambda}\mathrm{d}\xi(\cdot,s),\varphi\rangle\right]\right]\to0. E μ [ η ( u ν ) ν κ [ ∫ 0 t ⟨ P ∇ ⋅ π Λ d ξ ( ⋅ , s ) , φ ⟩ ] ] → 0.
We note here that the factor ν κ \nu^{\kappa} ν κ could be replaced by any expression vanishing as ν → 0 \nu\to0 ν → 0 .
For the last term of (36) write
∣ ∫ 0 t ⟨ f − π Λ f , φ ⟩ d s ∣ = ∣ ∫ 0 t ⟨ f , φ − π Λ φ ⟩ d s ∣ ≲ ∥ f ∥ C 0 ( [ 0 , T ] , B 2 / β , ∞ σ − β ( T d ) ) ∥ φ − π Λ φ ∥ B 2 / ( 2 − β ) , 1 σ − β ( T d ) . \left|\int_{0}^{t}\langle\mathbf{f}-\pi_{\Lambda}\mathbf{f},\varphi\rangle\mathrm{d}s\right|=\left|\int_{0}^{t}\langle\mathbf{f},\varphi-\pi_{\Lambda}\varphi\rangle\mathrm{d}s\right|\lesssim\|\mathbf{f}\|_{C^{0}([0,T],B^{\sigma-\beta}_{2/\beta,\infty}(\mathbb{T}^{d}))}\|\varphi-\pi_{\Lambda}\varphi\|_{B^{\sigma-\beta}_{2/(2-\beta),1}(\mathbb{T}^{d})}. ∫ 0 t ⟨ f − π Λ f , φ ⟩ d s = ∫ 0 t ⟨ f , φ − π Λ φ ⟩ d s ≲ ∥ f ∥ C 0 ([ 0 , T ] , B 2/ β , ∞ σ − β ( T d )) ∥ φ − π Λ φ ∥ B 2/ ( 2 − β ) , 1 σ − β ( T d ) .
Since φ ∈ S ( T d ) ⊆ B 2 / ( 2 − β ) , 1 σ − β ( T d ) \varphi\in\mathcal{S}(\mathbb{T}^{d})\subseteq B^{\sigma-\beta}_{2/(2-\beta),1}(\mathbb{T}^{d}) φ ∈ S ( T d ) ⊆ B 2/ ( 2 − β ) , 1 σ − β ( T d ) , the last term vanishes as Λ → ∞ \Lambda\to\infty Λ → ∞ . □
Lemma 8. Fix t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , φ ∈ S ( R d ) \varphi\in\mathcal{S}(\mathbb{R}^{d}) φ ∈ S ( R d ) . If
∫ E t f ( u , φ ) η ( u ) d P μ ( u ) = 0 , ∀ η ∈ C b , B 0 ( Z r , p σ , β , R ) , \int E_{t}^{f}(\mathbf{u},\varphi)\eta(\mathbf{u})\,\mathrm{d}P_{\mu}(\mathbf{u})=0,\quad\forall\eta\in C^{0}_{b,B}(\mathcal{Z}^{\sigma,\beta}_{r,p},\mathbb{R}), ∫ E t f ( u , φ ) η ( u ) d P μ ( u ) = 0 , ∀ η ∈ C b , B 0 ( Z r , p σ , β , R ) , then
E t f ( u , φ ) = 0 P μ -a.e. u E_{t}^{f}(\mathbf{u},\varphi)=0\quad P_{\mu}\text{-a.e. }\mathbf{u} E t f ( u , φ ) = 0 P μ -a.e. u
Proof. We define a sequence of functions χ k ∈ C b , B 0 ( Z r , p σ , β , [ 0 , 1 ] ) \chi_{k}\in C^{0}_{b,B}(\mathcal{Z}^{\sigma,\beta}_{r,p},[0,1]) χ k ∈ C b , B 0 ( Z r , p σ , β , [ 0 , 1 ]) with χ k ( u ) ↑ 1 \chi_{k}(\mathbf{u})\uparrow1 χ k ( u ) ↑ 1 as k ↑ ∞ k\uparrow\infty k ↑ ∞ pointwise in u \mathbf{u} u . First, we define the sets for k ∈ N + k\in\mathbb{N}^{+} k ∈ N +
A k = { u ∈ Z r , p σ , β : ∥ u ∥ Z r , p σ , β ≤ k } A_{k}=\{\mathbf{u}\in\mathcal{Z}^{\sigma,\beta}_{r,p}:\|\mathbf{u}\|_{\mathcal{Z}^{\sigma,\beta}_{r,p}}\leq k\} A k = { u ∈ Z r , p σ , β : ∥ u ∥ Z r , p σ , β ≤ k }
B k = { u ∈ Z r , p σ , β : ∥ u ∥ Z r , p σ , β ≥ k + 1 } B_{k}=\{\mathbf{u}\in\mathcal{Z}^{\sigma,\beta}_{r,p}:\|\mathbf{u}\|_{\mathcal{Z}^{\sigma,\beta}_{r,p}}\geq k+1\} B k = { u ∈ Z r , p σ , β : ∥ u ∥ Z r , p σ , β ≥ k + 1 }
which are clearly disjoint and closed in the strong topology of Z r , p σ , β \mathcal{Z}^{\sigma,\beta}_{r,p} Z r , p σ , β . Since this topology for Z r , p σ , β \mathcal{Z}^{\sigma,\beta}_{r,p} Z r , p σ , β is given by a metric, it is normal and the Urysohn lemma applies. Thus, for each k ∈ N + k\in\mathbb{N}^{+} k ∈ N + there exists a function χ k ∈ C 0 ( Z r , p σ , β , [ 0 , 1 ] ) \chi_{k}\in C^{0}(\mathcal{Z}^{\sigma,\beta}_{r,p},[0,1]) χ k ∈ C 0 ( Z r , p σ , β , [ 0 , 1 ]) such that χ k ∣ A k = 1 \chi_{k}|_{A_{k}}=1 χ k ∣ A k = 1 and χ k ∣ B k = 0 \chi_{k}|_{B_{k}}=0 χ k ∣ B k = 0 . Obviously, χ k \chi_{k} χ k is supported in A k + 1 A_{k+1} A k + 1 , which is norm-bounded, and lim k ↑ ∞ χ k ↑ 1 \lim_{k\uparrow\infty}\chi_{k}\uparrow1 lim k ↑ ∞ χ k ↑ 1 . Setting now η k ( u ) : = E t f ( u , φ ) χ k ( u ) \eta_{k}(\mathbf{u}):=E_{t}^{f}(\mathbf{u},\varphi)\chi_{k}(\mathbf{u}) η k ( u ) := E t f ( u , φ ) χ k ( u ) , we observe that η k ∈ C b , B 0 ( Z r , p σ , β ( T d ) ) \eta_{k}\in C^{0}_{b,B}(\mathcal{Z}^{\sigma,\beta}_{r,p}(\mathbb{T}^{d})) η k ∈ C b , B 0 ( Z r , p σ , β ( T d )) and thus we can apply Lemma 7 to infer that
∫ ∣ E t f ( u , φ ) ∣ 2 χ k ( u ) d P μ ( u ) = ∫ ∣ E t f ( u , φ ) η k ( u ) ∣ d P μ ( u ) = 0. \int\left|E_{t}^{f}(\mathbf{u},\varphi)\right|^{2}\chi_{k}(\mathbf{u})\,\mathrm{d}P_{\mu}(\mathbf{u})=\int\left|E_{t}^{f}(\mathbf{u},\varphi)\eta_{k}(\mathbf{u})\right|\,\mathrm{d}P_{\mu}(\mathbf{u})=0. ∫ E t f ( u , φ ) 2 χ k ( u ) d P μ ( u ) = ∫ E t f ( u , φ ) η k ( u ) d P μ ( u ) = 0.
Taking the limit k → ∞ k \to\infty k → ∞ , we obtain by monotone convergence that
∫ ∣ E t f ( u , φ ) ∣ 2 d P μ ( u ) = 0. \int\left|E_t^{\mathrm{f}}(\mathbf{u},\varphi)\right|^2\,\mathrm{d}P_\mu(\mathbf{u})=0. ∫ E t f ( u , φ ) 2 d P μ ( u ) = 0.
It is immediate that E t f ( u , φ ) = 0 E_t^{\mathrm{f}}(\mathbf{u},\varphi)=0 E t f ( u , φ ) = 0 for P μ P_\mu P μ –a.e. u \mathbf{u} u . \quad □ \square □
Using the previous lemmas, we now prove that realizations of P μ P_\mu P μ are weak Euler solutions. For each t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , φ ∈ S ( R d ) \varphi\in\mathcal{S}(\mathbb{R}^d) φ ∈ S ( R d ) , we define the set
W t , φ : = { u ∈ Z r , p σ , β : E t f ( u , φ ) = 0 } , W_{t,\varphi}:=\left\{\mathbf{u}\in\mathcal{Z}_{r,p}^{\sigma,\beta}:E_t^{\mathrm{f}}(\mathbf{u},\varphi)=0\right\}, W t , φ := { u ∈ Z r , p σ , β : E t f ( u , φ ) = 0 } ,
As u ↦ E t f ( u , φ ) \mathbf{u}\mapsto E_t^{\mathrm{f}}(\mathbf{u},\varphi) u ↦ E t f ( u , φ ) is a continuous mapping W t , φ W_{t,\varphi} W t , φ is a closed set and thus measurable. From our previous analysis, P μ ( W t , φ ) = 1 P_\mu(W_{t,\varphi})=1 P μ ( W t , φ ) = 1 . In part (ii) of the theorem, we proved that P μ ( E r , p ξ , β ) = 1 P_\mu(\mathcal{E}_{r,p}^{\xi,\beta})=1 P μ ( E r , p ξ , β ) = 1 , so that
P ( W t , φ ∩ E r , p ξ , β ) = 1. P(W_{t,\varphi}\cap\mathcal{E}_{r,p}^{\xi,\beta})=1. P ( W t , φ ∩ E r , p ξ , β ) = 1.
We can thus replace W t , φ W_{t,\varphi} W t , φ with the measurable set W t , φ ∩ E r , p ξ , β W_{t,\varphi}\cap\mathcal{E}_{r,p}^{\xi,\beta} W t , φ ∩ E r , p ξ , β , so that W t , φ ⊂ E r , p ξ , β W_{t,\varphi}\subset\mathcal{E}_{r,p}^{\xi,\beta} W t , φ ⊂ E r , p ξ , β .
Let T = [ 0 , T ] ∩ Q \mathcal{T}=[0,T]\cap\mathbb{Q} T = [ 0 , T ] ∩ Q be the countable dense subset of rational times and denote by φ α ( x ) = e i k ⋅ x e i \varphi_\alpha(\mathbf{x})=e^{\mathrm{i}\mathbf{k}\cdot\mathbf{x}}\mathbf{e}_i φ α ( x ) = e i k ⋅ x e i , α ∈ A = { ( k , i ) : k ∈ Z d , i ∈ { 1 , 2 , 3 } } \alpha\in\mathcal{A}=\{(\mathbf{k},i):\mathbf{k}\in\mathbb{Z}^d,\ i\in\{1,2,3\}\} α ∈ A = {( k , i ) : k ∈ Z d , i ∈ { 1 , 2 , 3 }} , which by standard Fourier theory has dense linear span in S ( T d ) \mathcal{S}(\mathbb{T}^d) S ( T d ) . Defining
W α : = ⋂ t ∈ T W t , φ α , W : = ⋂ α ∈ A W α . W_\alpha:=\bigcap_{t\in\mathcal{T}}W_{t,\varphi_\alpha},\qquad W:=\bigcap_{\alpha\in\mathcal{A}}W_\alpha. W α := t ∈ T ⋂ W t , φ α , W := α ∈ A ⋂ W α .
we have P ( W α ) = 1 P(W_\alpha)=1 P ( W α ) = 1 and P ( W ) = 1 P(W)=1 P ( W ) = 1 , and furthermore
E t f ( u , φ α ) = 0 , ∀ t ∈ T , ∀ α ∈ A , ∀ u ∈ W . E_t^{\mathrm{f}}(\mathbf{u},\varphi_\alpha)=0,\qquad\forall t\in\mathcal{T},\ \forall\alpha\in\mathcal{A},\quad\forall\mathbf{u}\in W. E t f ( u , φ α ) = 0 , ∀ t ∈ T , ∀ α ∈ A , ∀ u ∈ W .
Since E t f ( u , φ α ) E_t^{\mathrm{f}}(\mathbf{u},\varphi_\alpha) E t f ( u , φ α ) is continuous in t t t for each α ∈ A \alpha\in\mathcal{A} α ∈ A by Lemma 6 (ii), we see that in fact
E t f ( u , φ α ) = 0 , ∀ t ∈ [ 0 , T ] , ∀ α ∈ A , ∀ u ∈ W . E_t^{\mathrm{f}}(\mathbf{u},\varphi_\alpha)=0,\qquad\forall t\in[0,T],\ \forall\alpha\in\mathcal{A},\quad\forall\mathbf{u}\in W. E t f ( u , φ α ) = 0 , ∀ t ∈ [ 0 , T ] , ∀ α ∈ A , ∀ u ∈ W .
Then, since E t f ( u , φ ) E_t^{\mathrm{f}}(\mathbf{u},\varphi) E t f ( u , φ ) is linear and continuous in φ \varphi φ by Lemma 6 (iii) and since finite linear combinations of { φ α : α ∈ A } \{\varphi_\alpha:\alpha\in\mathcal{A}\} { φ α : α ∈ A } are dense in S ( T d ) \mathcal{S}(\mathbb{T}^d) S ( T d ) , then furthermore
E t f ( u , φ ) = 0 , ∀ t ∈ [ 0 , T ] , ∀ φ ∈ S ( T d ) , ∀ u ∈ W . E_t^{\mathrm{f}}(\mathbf{u},\varphi)=0,\qquad\forall t\in[0,T],\ \forall\varphi\in\mathcal{S}(\mathbb{T}^d),\quad\forall\mathbf{u}\in W. E t f ( u , φ ) = 0 , ∀ t ∈ [ 0 , T ] , ∀ φ ∈ S ( T d ) , ∀ u ∈ W .
We conclude that velocity fields P μ P_\mu P μ –a.s. are mild Euler solutions in the sense of equation (35) .
In fact, it is not hard to show that (35) implies the standard notion of weak Euler solution. We sketch here the argument. Any ψ ∈ S ( 0 , T ) \psi\in\mathcal{S}(0,T) ψ ∈ S ( 0 , T ) must be supported on some closed interval [ h 0 , T − h 0 ] [h_0,T-h_0] [ h 0 , T − h 0 ] for 0 < h 0 < T / 2 0<h_0<T/2 0 < h 0 < T /2 . Smearing the equation (35) with the derivative function ψ ′ \psi' ψ ′ then gives
∫ 0 T E t f ( u , φ ) ψ ′ ( t ) d t = 0. (39) \int_0^T E_t^{\mathrm{f}}(\mathbf{u},\varphi)\psi'(t)\,\mathrm{d}t=0. \tag*{(39)} ∫ 0 T E t f ( u , φ ) ψ ′ ( t ) d t = 0. (39)
Note that the term ⟨ u ( ⋅ , 0 ) , φ ⟩ \langle\mathbf{u}(\cdot,0),\varphi\rangle ⟨ u ( ⋅ , 0 ) , φ ⟩ in equation (34) for the Euler residual is constant in time and thus gives no contribution to the above integral. As noted in the proof of Lemma 6 (iii), the time-integral part of E t f ( u , φ ) E_t^{\mathrm{f}}(\mathbf{u},\varphi) E t f ( u , φ ) in equation (34) given by
K t f ( u , φ ) : = ∫ 0 t ⟨ P [ ∇ ⋅ ( u ⊗ u ) − f ] ( ⋅ , s ) , φ ⟩ d s K_t^{\mathrm{f}}(\mathbf{u},\varphi):=\int_0^t\left\langle\mathcal{P}[\nabla\cdot(\mathbf{u}\otimes\mathbf{u})-\mathbf{f}](\cdot,s),\varphi\right\rangle\,\mathrm{d}s K t f ( u , φ ) := ∫ 0 t ⟨ P [ ∇ ⋅ ( u ⊗ u ) − f ] ( ⋅ , s ) , φ ⟩ d s
is in fact absolutely continuous in time t t t . Thus, integration by parts is justified and one obtains
∫ 0 T K t f ( u , φ ) ψ ′ ( t ) d t = − ∫ 0 T d d t K t f ( u , φ ) ψ ( t ) d t = − ∫ 0 T ⟨ P [ ∇ ⋅ ( u ⊗ u ) − f ] ( ⋅ , t ) , φ ⟩ ψ ( t ) d t . (40) \begin{aligned}
\int_0^T K_t^{\mathrm{f}}(\mathbf{u},\varphi)\psi'(t)\,\mathrm{d}t
&=-\int_0^T\frac{\mathrm{d}}{\mathrm{d}t}K_t^{\mathrm{f}}(\mathbf{u},\varphi)\psi(t)\,\mathrm{d}t\\
&=-\int_0^T\left\langle\mathcal{P}[\nabla\cdot(\mathbf{u}\otimes\mathbf{u})-\mathbf{f}](\cdot,t),\varphi\right\rangle\psi(t)\,\mathrm{d}t. \tag*{(40)}
\end{aligned} ∫ 0 T K t f ( u , φ ) ψ ′ ( t ) d t = − ∫ 0 T d t d K t f ( u , φ ) ψ ( t ) d t = − ∫ 0 T ⟨ P [ ∇ ⋅ ( u ⊗ u ) − f ] ( ⋅ , t ) , φ ⟩ ψ ( t ) d t . (40)
We thus obtain from the equation (39) the condition
∫ 0 T ⟨ u ( ⋅ , t ) , φ ⟩ ψ ′ ( t ) d t − ∫ 0 T ⟨ P [ ∇ ⋅ ( u ⊗ u ) − f ] ( ⋅ , t ) , φ ⟩ ψ ( t ) d t = 0. (41) \int_{0}^{T}\langle\mathbf{u}(\cdot,t),\varphi\rangle\psi'(t)\,\mathrm{d}t-\int_{0}^{T}\langle\mathcal{P}[\nabla\cdot(\mathbf{u}\otimes\mathbf{u})-\mathbf{f}](\cdot,t),\varphi\rangle\psi(t)\,\mathrm{d}t=0. \tag*{(41)} ∫ 0 T ⟨ u ( ⋅ , t ) , φ ⟩ ψ ′ ( t ) d t − ∫ 0 T ⟨ P [ ∇ ⋅ ( u ⊗ u ) − f ] ( ⋅ , t ) , φ ⟩ ψ ( t ) d t = 0. (41)
which is essentially the standard weak formulation of the Euler equations. In view of part (iv) of the theorem concerning initial data, which we prove in the next subsection, it is useful to remark that this same argument may be applied even if ψ ( 0 ) ≠ 0 \psi(0)\ne0 ψ ( 0 ) = 0 . In that case, one obtains instead
∫ 0 T ⟨ u ( ⋅ , t ) , φ ⟩ ψ ′ ( t ) d t + ⟨ u ( ⋅ , 0 ) , φ ⟩ ψ ( 0 ) − ∫ 0 T ⟨ P [ ∇ ⋅ ( u ⊗ u ) − f ] ( ⋅ , t ) , φ ⟩ ψ ( t ) d t = 0. (42) \int_{0}^{T}\langle\mathbf{u}(\cdot,t),\varphi\rangle\psi'(t)\,\mathrm{d}t+\langle\mathbf{u}(\cdot,0),\varphi\rangle\psi(0)-\int_{0}^{T}\langle\mathcal{P}[\nabla\cdot(\mathbf{u}\otimes\mathbf{u})-\mathbf{f}](\cdot,t),\varphi\rangle\psi(t)\,\mathrm{d}t=0. \tag*{(42)} ∫ 0 T ⟨ u ( ⋅ , t ) , φ ⟩ ψ ′ ( t ) d t + ⟨ u ( ⋅ , 0 ) , φ ⟩ ψ ( 0 ) − ∫ 0 T ⟨ P [ ∇ ⋅ ( u ⊗ u ) − f ] ( ⋅ , t ) , φ ⟩ ψ ( t ) d t = 0. (42)
This is a standard approach to incorporate initial data in the weak formulation of a PDE.
Finally, we note that a weak formulation of the incompressibility condition follows directly from (35). If we take φ = ∇ χ \varphi=\nabla\chi φ = ∇ χ for a scalar function χ ∈ S ( T d ) \chi\in\mathcal{S}(\mathbb{T}^{d}) χ ∈ S ( T d ) , then substituting into (35) it follows that all of the time-integral terms vanish in the equation (34) defining E t f ( u , φ ) E_t^f(\mathbf{u},\varphi) E t f ( u , φ ) , because of the Leray projection P \mathcal{P} P . The resulting equation is
⟨ u ( ⋅ , t ) , ∇ χ ⟩ = ⟨ u ( ⋅ , 0 ) , ∇ χ ⟩ , ∀ t ∈ [ 0 , T ] , ∀ χ ∈ S ( T d ) . \langle\mathbf{u}(\cdot,t),\nabla\chi\rangle=\langle\mathbf{u}(\cdot,0),\nabla\chi\rangle,\quad\forall t\in[0,T],\ \forall\chi\in\mathcal{S}(\mathbb{T}^{d}). ⟨ u ( ⋅ , t ) , ∇ χ ⟩ = ⟨ u ( ⋅ , 0 ) , ∇ χ ⟩ , ∀ t ∈ [ 0 , T ] , ∀ χ ∈ S ( T d ) .
Recall that initial data for the Landau-Lifschitz equation (5) are chosen from the measure μ \mu μ such that
⟨ u 0 , ∇ χ ⟩ = 0 , ∀ χ ∈ S ( T d ) μ -a.e. u 0 \langle\mathbf{u}_0,\nabla\chi\rangle=0,\quad\forall\chi\in\mathcal{S}(\mathbb{T}^{d})\qquad\mu\text{-a.e. }\mathbf{u}_0 ⟨ u 0 , ∇ χ ⟩ = 0 , ∀ χ ∈ S ( T d ) μ -a.e. u 0
Thus, we infer the weak incompressibility condition
⟨ u ( ⋅ , t ) , ∇ χ ⟩ = 0 , ∀ t ∈ [ 0 , T ] , ∀ χ ∈ S ( T d ) P μ -a.e. u \langle\mathbf{u}(\cdot,t),\nabla\chi\rangle=0,\quad\forall t\in[0,T],\ \forall\chi\in\mathcal{S}(\mathbb{T}^{d})\qquad P_{\mu}\text{-a.e. }\mathbf{u} ⟨ u ( ⋅ , t ) , ∇ χ ⟩ = 0 , ∀ t ∈ [ 0 , T ] , ∀ χ ∈ S ( T d ) P μ -a.e. u
if it can be shown that the distribution of u ( ⋅ , 0 ) \mathbf{u}(\cdot,0) u ( ⋅ , 0 ) induced by the measure P μ P_{\mu} P μ on u \mathbf{u} u coincides with measure μ \mu μ . This is precisely the result established in the next section.
Proof of (iv)
Our proof is modeled on Theorem IV.5.3 of [28 ] . We have already observed that the trace γ 0 u = u ( ⋅ , 0 ) \gamma_0\mathbf{u}=\mathbf{u}(\cdot,0) γ 0 u = u ( ⋅ , 0 ) is well-defined as a map γ 0 : C 0 ( [ 0 , T ] , B p / 2 , ∞ σ − β ( T d ) ) → B p / 2 , ∞ σ − β ( T d ) \gamma_0:C^0([0,T],B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}))\to B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}) γ 0 : C 0 ([ 0 , T ] , B p /2 , ∞ σ − β ( T d )) → B p /2 , ∞ σ − β ( T d ) and is strongly continuous on Z r , p σ , β \mathcal{Z}_{r,p}^{\sigma,\beta} Z r , p σ , β . Thus, the pushforward measure P μ ( 0 ) : = γ 0 ∗ P μ P_{\mu}^{(0)}:=\gamma_{0*}P_{\mu} P μ ( 0 ) := γ 0 ∗ P μ is well-defined on B ( B p / 2 , ∞ σ − β ) \mathcal{B}(B_{p/2,\infty}^{\sigma-\beta}) B ( B p /2 , ∞ σ − β ) . Our strategy to show the equivalence of P μ ( 0 ) P_{\mu}^{(0)} P μ ( 0 ) and μ \mu μ is based on the study of their characteristic functions:
S μ ( 0 ) [ v ] = ∫ exp ( i ⟨ u 0 , v ⟩ ) d P μ ( 0 ) ( u 0 ) , S μ [ v ] = ∫ exp ( i ⟨ u 0 , v ⟩ ) d μ ( u 0 ) S_{\mu}^{(0)}[\mathbf{v}]=\int\exp(i\langle\mathbf{u}_0,\mathbf{v}\rangle)\,\mathrm{d}P_{\mu}^{(0)}(\mathbf{u}_0),\qquad S_{\mu}[\mathbf{v}]=\int\exp(i\langle\mathbf{u}_0,\mathbf{v}\rangle)\,\mathrm{d}\mu(\mathbf{u}_0) S μ ( 0 ) [ v ] = ∫ exp ( i ⟨ u 0 , v ⟩) d P μ ( 0 ) ( u 0 ) , S μ [ v ] = ∫ exp ( i ⟨ u 0 , v ⟩) d μ ( u 0 )
defined here for v ∈ S ( T d ) \mathbf{v}\in\mathcal{S}(\mathbb{T}^{d}) v ∈ S ( T d ) . We note first that one can define similarly P μ ν , ( 0 ) : = γ 0 ∗ P μ ν P_{\mu}^{\nu,(0)}:=\gamma_{0*}P_{\mu}^{\nu} P μ ν , ( 0 ) := γ 0 ∗ P μ ν for ν > 0 \nu>0 ν > 0 and its characteristic function satisfies
S μ ν , ( 0 ) [ v ] : = ∫ exp ( i ⟨ u 0 , v ⟩ ) d P μ ν , ( 0 ) ( u 0 ) = ∫ exp ( i ⟨ γ 0 u , v ⟩ ) d P μ ν ( u ) S_{\mu}^{\nu,(0)}[\mathbf{v}]:=\int\exp(i\langle\mathbf{u}_0,\mathbf{v}\rangle)\,\mathrm{d}P_{\mu}^{\nu,(0)}(\mathbf{u}_0)=\int\exp(i\langle\gamma_0\mathbf{u},\mathbf{v}\rangle)\,\mathrm{d}P_{\mu}^{\nu}(\mathbf{u}) S μ ν , ( 0 ) [ v ] := ∫ exp ( i ⟨ u 0 , v ⟩) d P μ ν , ( 0 ) ( u 0 ) = ∫ exp ( i ⟨ γ 0 u , v ⟩) d P μ ν ( u )
Because exp ( i ⟨ γ 0 u , v ⟩ ) \exp(i\langle\gamma_0\mathbf{u},\mathbf{v}\rangle) exp ( i ⟨ γ 0 u , v ⟩) is a bounded, continuous function, it follows from weak convergence that
lim ν → 0 S μ ν , ( 0 ) [ v ] = ∫ exp ( i ⟨ γ 0 u , v ⟩ ) d P μ ( u ) = ∫ exp ( i ⟨ u 0 , v ⟩ ) d P μ ( 0 ) ( u 0 ) = S μ ( 0 ) [ v ] . \lim_{\nu\to0}S_{\mu}^{\nu,(0)}[\mathbf{v}]=\int\exp(i\langle\gamma_0\mathbf{u},\mathbf{v}\rangle)\,\mathrm{d}P_{\mu}(\mathbf{u})=\int\exp(i\langle\mathbf{u}_0,\mathbf{v}\rangle)\,\mathrm{d}P_{\mu}^{(0)}(\mathbf{u}_0)=S_{\mu}^{(0)}[\mathbf{v}]. ν → 0 lim S μ ν , ( 0 ) [ v ] = ∫ exp ( i ⟨ γ 0 u , v ⟩) d P μ ( u ) = ∫ exp ( i ⟨ u 0 , v ⟩) d P μ ( 0 ) ( u 0 ) = S μ ( 0 ) [ v ] .
On the other hand, we have also by the definition u 0 ν : = π Λ u 0 \mathbf{u}_0^\nu:=\pi_\Lambda\mathbf{u}_0 u 0 ν := π Λ u 0 that
S μ ν , ( 0 ) [ v ] = ∫ exp ( i ⟨ π Λ u 0 , v ⟩ ) d μ ( u 0 ) . S_{\mu}^{\nu,(0)}[\mathbf{v}]=\int\exp(i\langle\pi_\Lambda\mathbf{u}_0,\mathbf{v}\rangle)\,\mathrm{d}\mu(\mathbf{u}_0). S μ ν , ( 0 ) [ v ] = ∫ exp ( i ⟨ π Λ u 0 , v ⟩) d μ ( u 0 ) .
Since π Λ u 0 → u 0 \pi_\Lambda\mathbf{u}_0\to\mathbf{u}_0 π Λ u 0 → u 0 in L p ( T d ) L^p(\mathbb{T}^{d}) L p ( T d ) and thus almost sure in μ \mu μ , dominated convergence gives also that
lim ν → 0 S μ ν , ( 0 ) [ v ] = ∫ exp ( i ⟨ u 0 , v ⟩ ) d μ ( u 0 ) = S μ [ v ] . \lim_{\nu\to0}S_{\mu}^{\nu,(0)}[\mathbf{v}]=\int\exp(i\langle\mathbf{u}_0,\mathbf{v}\rangle)\,\mathrm{d}\mu(\mathbf{u}_0)=S_{\mu}[\mathbf{v}]. ν → 0 lim S μ ν , ( 0 ) [ v ] = ∫ exp ( i ⟨ u 0 , v ⟩) d μ ( u 0 ) = S μ [ v ] .
Comparing the two limit expressions, we conclude that
S μ ( 0 ) [ v ] = S μ [ v ] . (43) S_{\mu}^{(0)}[\mathbf{v}] = S_{\mu}[\mathbf{v}]. \tag*{(43)} S μ ( 0 ) [ v ] = S μ [ v ] . (43)
We would like to infer from this equality that P μ ( 0 ) = μ P_{\mu}^{(0)}=\mu P μ ( 0 ) = μ by applying the Bochner–Minlos theorem [99 ] . However, this theorem applies to measures on B ( S ′ ( T d ) ) \mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})) B ( S ′ ( T d )) , whereas P μ ( 0 ) P_{\mu}^{(0)} P μ ( 0 ) is defined on B ( B p / 2 , ∞ σ − β ( T d ) ) \mathcal{B}(B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})) B ( B p /2 , ∞ σ − β ( T d )) and μ \mu μ on B ( L p ( T d ) ) \mathcal{B}(L^{p}(\mathbb{T}^{d})) B ( L p ( T d )) . To resolve this difficulty we must extend both of these measures to B ( S ′ ( T d ) ) \mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})) B ( S ′ ( T d )) , which is the content of the following:
Lemma 9. (i) The set B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}) B p /2 , ∞ σ − β ( T d ) belongs to B ( S ′ ( T d ) ) \mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})) B ( S ′ ( T d )) and the measure P μ ( 0 ) P_{\mu}^{(0)} P μ ( 0 ) has a unique extension P ‾ μ ( 0 ) \overline{P}_{\mu}^{(0)} P μ ( 0 ) to B ( S ′ ( T d ) ) \mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})) B ( S ′ ( T d )) that is supported on B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}) B p /2 , ∞ σ − β ( T d ) , defined by
P ‾ μ ( 0 ) ( E ) = P μ ( 0 ) ( E ∩ B p / 2 , ∞ σ − β ( T d ) ) , ∀ E ∈ B ( S ′ ( T d ) ) . (44) \overline{P}_{\mu}^{(0)}(E)=P_{\mu}^{(0)}\left(E\cap B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})\right),\qquad\forall E\in\mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})). \tag*{(44)} P μ ( 0 ) ( E ) = P μ ( 0 ) ( E ∩ B p /2 , ∞ σ − β ( T d ) ) , ∀ E ∈ B ( S ′ ( T d )) . (44) (ii) The set L p ( T d ) L^{p}(\mathbb{T}^{d}) L p ( T d ) belongs to B ( S ′ ( T d ) ) \mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})) B ( S ′ ( T d )) and the measure μ \mu μ has a unique extension μ ‾ \overline{\mu} μ to B ( S ′ ( T d ) ) \mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})) B ( S ′ ( T d )) that is supported on L p ( T d ) L^{p}(\mathbb{T}^{d}) L p ( T d ) , defined by
μ ‾ ( E ) = μ ( E ∩ L p ( T d ) ) , ∀ E ∈ B ( S ′ ( T d ) ) . (45) \overline{\mu}(E)=\mu\left(E\cap L^{p}(\mathbb{T}^{d})\right),\qquad\forall E\in\mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})). \tag*{(45)} μ ( E ) = μ ( E ∩ L p ( T d ) ) , ∀ E ∈ B ( S ′ ( T d )) . (45)
Proof. The argument is an elaboration of Remark II.2.1 in [28 ] . (i) We prove Borel measurability of B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}) B p /2 , ∞ σ − β ( T d ) in S ′ ( T d ) \mathcal{S}'(\mathbb{T}^{d}) S ′ ( T d ) . Note first that the extended real-valued function defined on S ′ ( T d ) \mathcal{S}'(\mathbb{T}^{d}) S ′ ( T d ) by
u ↦ ∥ u ∥ B p / 2 , ∞ σ − β ( T d ) \mathbf{u}\mapsto\lVert\mathbf{u}\rVert_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})} u ↦ ∥ u ∥ B p /2 , ∞ σ − β ( T d )
is lower semicontinuous in the Fréchet topology of S ′ ( T d ) \mathcal{S}'(\mathbb{T}^{d}) S ′ ( T d ) . To see this, take u n → u \mathbf{u}_{n}\to\mathbf{u} u n → u in S ′ ( T d ) \mathcal{S}'(\mathbb{T}^{d}) S ′ ( T d ) and note
∀ φ ∈ S ( T d ) : ∣ ⟨ u n , φ ⟩ ∣ ≤ ∥ φ ∥ B 2 / ( 2 − p ) , 1 β − σ ( T d ) ∥ u n ∥ B p / 2 , ∞ σ − β ( T d ) . \forall\varphi\in\mathcal{S}(\mathbb{T}^{d}):\ \left|\langle\mathbf{u}_{n},\varphi\rangle\right|\leq\lVert\varphi\rVert_{B_{2/(2-p),1}^{\beta-\sigma}(\mathbb{T}^{d})}\lVert\mathbf{u}_{n}\rVert_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})}. ∀ φ ∈ S ( T d ) : ∣ ⟨ u n , φ ⟩ ∣ ≤ ∥ φ ∥ B 2/ ( 2 − p ) , 1 β − σ ( T d ) ∥ u n ∥ B p /2 , ∞ σ − β ( T d ) .
Taking the limit infimum as n → ∞ n\to\infty n → ∞ of both sides of this inequality gives
∀ φ ∈ S ( T d ) : ∣ ⟨ u , φ ⟩ ∣ ≤ ∥ φ ∥ B 2 / ( 2 − p ) , 1 β − σ ( T d ) lim inf n → ∞ ∥ u n ∥ B p / 2 , ∞ σ − β ( T d ) . \forall\varphi\in\mathcal{S}(\mathbb{T}^{d}):\ \left|\langle\mathbf{u},\varphi\rangle\right|\leq\lVert\varphi\rVert_{B_{2/(2-p),1}^{\beta-\sigma}(\mathbb{T}^{d})}\liminf_{n\to\infty}\lVert\mathbf{u}_{n}\rVert_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})}. ∀ φ ∈ S ( T d ) : ∣ ⟨ u , φ ⟩ ∣ ≤ ∥ φ ∥ B 2/ ( 2 − p ) , 1 β − σ ( T d ) n → ∞ lim inf ∥ u n ∥ B p /2 , ∞ σ − β ( T d ) .
Then, using the dense embedding S ( T d ) ↪ B 2 / ( 2 − p ) , 1 β − σ ( T d ) \mathcal{S}(\mathbb{T}^{d})\hookrightarrow B_{2/(2-p),1}^{\beta-\sigma}(\mathbb{T}^{d}) S ( T d ) ↪ B 2/ ( 2 − p ) , 1 β − σ ( T d ) ,
∥ u ∥ B p / 2 , ∞ σ − β ( T d ) ≡ sup φ ∈ S ( T d ) , ∥ φ ∥ B 2 / ( 2 − p ) , 1 β − σ ( T d ) = 1 ∣ ⟨ u , φ ⟩ ∣ ≤ lim inf n → ∞ ∥ u n ∥ B p / 2 , ∞ σ − β ( T d ) , \lVert\mathbf{u}\rVert_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})}\equiv\sup_{\substack{\varphi\in\mathcal{S}(\mathbb{T}^{d}),\ \lVert\varphi\rVert_{B_{2/(2-p),1}^{\beta-\sigma}(\mathbb{T}^{d})}=1}}\left|\langle\mathbf{u},\varphi\rangle\right|\leq\liminf_{n\to\infty}\lVert\mathbf{u}_{n}\rVert_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})}, ∥ u ∥ B p /2 , ∞ σ − β ( T d ) ≡ φ ∈ S ( T d ) , ∥ φ ∥ B 2/ ( 2 − p ) , 1 β − σ ( T d ) = 1 sup ∣ ⟨ u , φ ⟩ ∣ ≤ n → ∞ lim inf ∥ u n ∥ B p /2 , ∞ σ − β ( T d ) ,
which proves lower semicontinuity. As a consequence, the sets
M p σ , β ( n ) : = { u ∈ S ′ ( T d ) : ∥ u ∥ B p / 2 , ∞ σ − β ( T d ) ≤ n } M_{p}^{\sigma,\beta}(n):=\left\{\mathbf{u}\in\mathcal{S}'(\mathbb{T}^{d}):\lVert\mathbf{u}\rVert_{B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})}\leq n\right\} M p σ , β ( n ) := { u ∈ S ′ ( T d ) : ∥ u ∥ B p /2 , ∞ σ − β ( T d ) ≤ n }
are closed in the Fréchet topology of S ′ ( T d ) \mathcal{S}'(\mathbb{T}^{d}) S ′ ( T d ) and then
B p / 2 , ∞ σ − β ( T d ) = ⋃ n ∈ N M p σ , β ( n ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})=\bigcup_{n\in\mathbb{N}}M_{p}^{\sigma,\beta}(n) B p /2 , ∞ σ − β ( T d ) = n ∈ N ⋃ M p σ , β ( n )
is measurable in B ( S ′ ( T d ) ) \mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})) B ( S ′ ( T d )) . In fact, we have proved that B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}) B p /2 , ∞ σ − β ( T d ) is an F σ F^{\sigma} F σ -set in S ′ ( T d ) \mathcal{S}'(\mathbb{T}^{d}) S ′ ( T d ) . A measure P ‾ μ ( 0 ) \overline{P}_{\mu}^{(0)} P μ ( 0 ) defined on B ( S ′ ( T d ) ) \mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})) B ( S ′ ( T d )) that is supported on the measurable set B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}) B p /2 , ∞ σ − β ( T d ) must satisfy
P ‾ μ ( 0 ) ( E ) = P ‾ μ ( 0 ) ( E ∩ B p / 2 , ∞ σ − β ( T d ) ) , ∀ E ∈ B ( S ′ ( T d ) ) . \overline{P}_{\mu}^{(0)}(E)=\overline{P}_{\mu}^{(0)}\left(E\cap B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})\right),\qquad\forall E\in\mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})). P μ ( 0 ) ( E ) = P μ ( 0 ) ( E ∩ B p /2 , ∞ σ − β ( T d ) ) , ∀ E ∈ B ( S ′ ( T d )) .
If furthermore this measure restricted to B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d}) B p /2 , ∞ σ − β ( T d ) coincides with P μ ( 0 ) P_{\mu}^{(0)} P μ ( 0 ) , then it must hold that
P ‾ μ ( 0 ) ( E ) = P μ ( 0 ) ( E ∩ B p / 2 , ∞ σ − β ( T d ) ) , ∀ E ∈ B ( S ′ ( T d ) ) , \overline{P}_{\mu}^{(0)}(E)=P_{\mu}^{(0)}\left(E\cap B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^{d})\right),\qquad\forall E\in\mathcal{B}(\mathcal{S}'(\mathbb{T}^{d})), P μ ( 0 ) ( E ) = P μ ( 0 ) ( E ∩ B p /2 , ∞ σ − β ( T d ) ) , ∀ E ∈ B ( S ′ ( T d )) ,
which is formula (44). This establishes uniqueness. However, to show existence, one must show that
{ E ∩ B p / 2 , ∞ σ − β ( T d ) : E ∈ B ( S ′ ( T d ) ) } ⊆ B ( B p / 2 , ∞ σ − β ( T d ) ) , (46) \left\{ E \cap B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d) : E \in\mathcal{B}(\mathcal{S}'(\mathbb{T}^d)) \right\} \subseteq\mathcal{B}\left(B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)\right), \tag*{(46)} { E ∩ B p /2 , ∞ σ − β ( T d ) : E ∈ B ( S ′ ( T d )) } ⊆ B ( B p /2 , ∞ σ − β ( T d ) ) , (46)
so that the formula (44) is meaningful. To prove this, recall that the topology on S ′ ( T d ) \mathcal{S}'(\mathbb{T}^d) S ′ ( T d ) is generated by the basis of open sets of the form
{ Λ ˉ φ − 1 ( U ) : φ ∈ S ( T d ) , open U ⊆ R } (47) \left\{ \bar{\Lambda}_{\varphi}^{-1}(U) : \varphi\in\mathcal{S}(\mathbb{T}^d),\ \text{open } U \subseteq\mathbb{R} \right\} \tag*{(47)} { Λ ˉ φ − 1 ( U ) : φ ∈ S ( T d ) , open U ⊆ R } (47)
where Λ ˉ φ : S ′ ( T d ) → R \bar{\Lambda}_{\varphi} : \mathcal{S}'(\mathbb{T}^d) \to\mathbb{R} Λ ˉ φ : S ′ ( T d ) → R for φ ∈ S ( T d ) \varphi\in\mathcal{S}(\mathbb{T}^d) φ ∈ S ( T d ) is defined by Λ ˉ φ ( v ) : = ⟨ v , φ ⟩ \bar{\Lambda}_{\varphi}(\mathbf{v}) := \langle\mathbf{v},\varphi\rangle Λ ˉ φ ( v ) := ⟨ v , φ ⟩ for v ∈ S ′ ( T d ) \mathbf{v} \in\mathcal{S}'(\mathbb{T}^d) v ∈ S ′ ( T d ) . Clearly, one can also define a corresponding functional Λ φ : B p / 2 , ∞ σ − β ( T d ) → R \Lambda_{\varphi} : B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d) \to\mathbb{R} Λ φ : B p /2 , ∞ σ − β ( T d ) → R by the same formula Λ φ ( v ) : = ⟨ v , φ ⟩ \Lambda_{\varphi}(\mathbf{v}) := \langle\mathbf{v},\varphi\rangle Λ φ ( v ) := ⟨ v , φ ⟩ for v ∈ B p / 2 , ∞ σ − β ( T d ) \mathbf{v} \in B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d) v ∈ B p /2 , ∞ σ − β ( T d ) . It is then easy to check that
Λ ˉ φ − 1 ( U ) ∩ B p / 2 , ∞ σ − β ( T d ) = Λ φ − 1 ( U ) . \bar{\Lambda}_{\varphi}^{-1}(U) \cap B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d) = \Lambda_{\varphi}^{-1}(U). Λ ˉ φ − 1 ( U ) ∩ B p /2 , ∞ σ − β ( T d ) = Λ φ − 1 ( U ) .
However, because S ( T d ) ⊆ B ( p − 2 ) / p , 1 β − σ ( T d ) \mathcal{S}(\mathbb{T}^d) \subseteq B_{(p-2)/p,1}^{\beta-\sigma}(\mathbb{T}^d) S ( T d ) ⊆ B ( p − 2 ) / p , 1 β − σ ( T d ) , the function Λ φ \Lambda_{\varphi} Λ φ is strongly continuous and thus the set Λ φ − 1 ( U ) \Lambda_{\varphi}^{-1}(U) Λ φ − 1 ( U ) is open in B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d) B p /2 , ∞ σ − β ( T d ) . Since the Borel σ \sigma σ -algebra B ( S ′ ( T d ) ) \mathcal{B}(\mathcal{S}'(\mathbb{T}^d)) B ( S ′ ( T d )) is generated by the basis (47), the requirement (46) for existence follows by standard Boolean set algebra.
(i) The proofs of the corresponding statements for L p ( R d ) L^p(\mathbb{R}^d) L p ( R d ) and the measure μ \mu μ are almost identical and need not be repeated here. □
In order to apply this lemma, note that the function u 0 ↦ ⟨ u 0 , v ⟩ \mathbf{u}_0 \mapsto\langle\mathbf{u}_0,\mathbf{v}\rangle u 0 ↦ ⟨ u 0 , v ⟩ for v ∈ S ( R d ) \mathbf{v} \in\mathcal{S}(\mathbb{R}^d) v ∈ S ( R d ) extends to a measurable function on S ′ ( T d ) \mathcal{S}'(\mathbb{T}^d) S ′ ( T d ) . Thus, we can rewrite the equality (43) instead as the equality
S ‾ μ ( 0 ) [ v ] = S ‾ μ [ v ] (48) \overline{\mathcal{S}}_{\mu}^{(0)}[\mathbf{v}] = \overline{\mathcal{S}}_{\mu}[\mathbf{v}] \tag*{(48)} S μ ( 0 ) [ v ] = S μ [ v ] (48)
for the characteristic functions of the measures P ‾ μ ( 0 ) \overline{P}_{\mu}^{(0)} P μ ( 0 ) and μ ‾ \overline{\mu} μ defined on B ( S ′ ( T d ) ) \mathcal{B}(\mathcal{S}'(\mathbb{T}^d)) B ( S ′ ( T d )) . Finally, the Bochner-Minlos theorem implies that
P ‾ μ ( 0 ) = μ ‾ . \overline{P}_{\mu}^{(0)} = \overline{\mu}. P μ ( 0 ) = μ .
It follows from this equality that P ‾ μ ( 0 ) \overline{P}_{\mu}^{(0)} P μ ( 0 ) is in fact supported on L p ( T d ) L^p(\mathbb{T}^d) L p ( T d ) . Now, we know that P ‾ μ ( 0 ) \overline{P}_{\mu}^{(0)} P μ ( 0 ) restricted to B ( B p / 2 , ∞ σ − β ( T d ) ) \mathcal{B}(B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)) B ( B p /2 , ∞ σ − β ( T d )) coincides with P μ ( 0 ) P_{\mu}^{(0)} P μ ( 0 ) . Evidently L p ( T d ) L^p(\mathbb{T}^d) L p ( T d ) is reflexive and separable, and per Lemma 2 embeds continuously into B p / 2 , ∞ σ − β ( T d ) B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d) B p /2 , ∞ σ − β ( T d ) . Thus, by Theorem II.2.1 and Lemma II.2.1 of [28 ] ,
B ( B p / 2 , ∞ σ − β ( T d ) ) ∩ L p ( T d ) = B ( L p ( T d ) ) . (49) \mathcal{B}\left(B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d)\right) \cap L^p(\mathbb{T}^d) = \mathcal{B}(L^p(\mathbb{T}^d)). \tag*{(49)} B ( B p /2 , ∞ σ − β ( T d ) ) ∩ L p ( T d ) = B ( L p ( T d )) . (49)
Note that the statement of Lemma II.2.1 in [28 ] required that both spaces Ω 1 \Omega_1 Ω 1 , Ω 2 \Omega_2 Ω 2 be separable, whereas in our case Ω 2 = B p / 2 , ∞ σ − β ( T d ) \Omega_2 = B_{p/2,\infty}^{\sigma-\beta}(\mathbb{T}^d) Ω 2 = B p /2 , ∞ σ − β ( T d ) is non-separable. Separability and reflexivity of Ω 1 \Omega_1 Ω 1 were both needed in the proof of [28 ] , Lemma II.2.1 to extract a weakly convergent subsequence, but neither property was required of Ω 2 \Omega_2 Ω 2 . For our case Ω 1 = L p ( T d ) \Omega_1 = L^p(\mathbb{T}^d) Ω 1 = L p ( T d ) is both separable and reflexive. The equality (49) thus holds and we can further restrict P μ ( 0 ) P_{\mu}^{(0)} P μ ( 0 ) to B ( L p ( T d ) ) \mathcal{B}(L^p(\mathbb{T}^d)) B ( L p ( T d )) , which must then coincide with μ ‾ \overline{\mu} μ restricted to B ( L p ( T d ) ) \mathcal{B}(L^p(\mathbb{T}^d)) B ( L p ( T d )) , which is μ \mu μ .
Acknowledgements
We are grateful to Dmytro Bandak, Theodore Drivas, Nigel Goldenfeld, and Alexei Mailybaev for discussions of this problem. This work was funded by the Simons Foundation, via Targeted Grant in MPS-663054 and the Collaboration Grant No. MPS- 1151713.
References
Department of Applied Mathematics and Statistics, Johns Hopkins University, Baltimore, MD 21218
Email address: eyink@jhu.edu, lpeng22@jhu.edu