Introduction
History and motivation
The present paper establishes the Richardson dispersion law and the anomalous dissipation of a tracer for a random Hölder continuous velocity vectorfield. The construction is homogeneous in time and geometrically dense in scales. In addition, it exhibits sweeping control long conjectured to be the key to important aspects of the turbulent behavior: small eddies are materially transported by the larger ones. Its physics companion is simultaneously released as [26 ] . The Obukhov-Corrsin-Yaglom law for the same vectorfield is established in the simultaneously released [13 ] .
Richardson’s 1926 prediction for turbulent relative dispersion is one of the oldest predictions in classical turbulence [32 ] . Yet it has evaded rigorous mathematical treatment to this date. Richardson proposed that the effective diffusivity of a particle pair at separation r r r should grow like r 4 / 3 r^{4/3} r 4/3 , or, equivalently, that the square of the separation between the particles grows as t 3 t^3 t 3 with time:
E R t 2 ∼ t 3 . \mathbb{E}R_t^2 \sim t^3. E R t 2 ∼ t 3 .
This scaling is consistent with Kolmogorov’s later prediction of smoothness of a turbulent vectorfield being roughly 1 / 3 1/3 1/3 (also open to this date and actually requiring an adjustment due to intermittency effects). More generally, if velocity increments at scale r r r are of order r α r^\alpha r α , dimensional reasoning suggests the separation scale r ( t ) ∼ t 1 / ( 1 − α ) r(t) \sim t^{1/(1-\alpha)} r ( t ) ∼ t 1/ ( 1 − α ) and hence
E R t 2 ∼ t 2 / ( 1 − α ) . \mathbb{E}R_t^2 \sim t^{2/(1-\alpha)}. E R t 2 ∼ t 2/ ( 1 − α ) .
We refer to [22 , 34 , 19 ] for classical reviews.
The intuitive reason behind the scaling is as follows. If velocity increments across a distance r r r are of size r α r^\alpha r α , a pair at separation r r r drifts apart at speed r α r^\alpha r α , and the ordinary differential equation r ˙ = r α \dot{r}=r^{\alpha} r ˙ = r α gives r 1 − α ≈ t r^{1-\alpha}\approx t r 1 − α ≈ t , that is, r 2 ≈ t 2 / ( 1 − α ) r^2\approx t^{2/(1-\alpha)} r 2 ≈ t 2/ ( 1 − α ) ; for the Kolmogorov exponent α = 1 / 3 \alpha=1/3 α = 1/3 this is t 3 t^3 t 3 . The heuristic is not a proof, since the increments have no preferred sign: they average to zero, and only their fluctuations, accumulated over many independent eddies, push the particles apart.
Rigorous results targeting a quantitative Richardson law are scarce and quite far from expected physical reality. The classical example which allows for explicit computations is the Kraichnan model [29 ] . Here, however, the advecting velocity is Gaussian and white in time. This is much rougher than expected regularity in time for real turbulence, see, for instance, [19 , 30 ] . We also mention the work of Fannjiang [21 ] , which proves generalized Richardson laws for colored-in-time synthetic velocities in an inertial-scale limit. In the regime yielding a closed pair-dispersion law, however, the limiting dynamics are again of white-in-time Kraichnan type, and the finite-correlation Kolmogorov endpoint is not itself covered by the limit theorem.
In a different direction, much more recently, Hess-Childs and Rowan [25 ] constructed an explicit deterministic divergence-free velocity V ∈ L t ∞ C x α ( T 2 ) V\in L_t^\infty C_x^\alpha(\mathbb{T}^2) V ∈ L t ∞ C x α ( T 2 ) for which two independently diffusive tracers satisfy the sharp two-sided estimate
E R t 2 ∼ R 0 2 + μ t + t 2 / ( 1 − α ) , \mathbb{E}R_t^2\sim R_0^2+\mu t+t^{2/(1-\alpha)}, E R t 2 ∼ R 0 2 + μ t + t 2/ ( 1 − α ) ,
where μ → 0 \mu\to0 μ → 0 is the molecular diffusivity constant. In particular, at α = 1 / 3 \alpha=1/3 α = 1/3 they obtain Richardson’s t 3 t^3 t 3 law, uniformly as μ \mu μ tends to zero. This is the only genuinely colored-in-time construction yielding Richardson, however, as the authors themselves emphasize, their velocity is a deterministic, explicitly engineered two-dimensional flow, rather than a stationary random turbulence model with finite correlation time. In particular, is not time-homogeneous: the Richardson estimate is proved for trajectories launched from a distinguished initial time, rather than uniformly with respect to the time origin. Without going into technical details, we will just say that the complexity here is concentrated near t = 0 t=0 t = 0 and thins out afterwards, rather than being homogeneous in time as physical expectations would require.
More generally, the present paper builds on a recent line of rigorous work on anomalous dissipation of passive scalars and on the related breakdown of uniqueness for rough transport. The latter, as we mentioned above, has a rich and growing history that our work builds on. Drivas, Elgindi, Iyer and Jeong [16 ] opened the subject with the first velocity field, in L t 1 C x − L_t^1C_x^{-} L t 1 C x − , that provably dissipates a passive scalar, built from alternating shears refined in time in the spirit of Pierrehumbert; Colombo, Crippa and Sorella [11 ] reached the full supercritical Obukhov–Corrsin regime and showed that vanishing diffusivity selects no unique limit. Bruè and De Lellis [9 ] , using the mixing flows of Alberti, Crippa and Mazzucato [3 , 4 ] , obtained anomalous dissipation for the forced three-dimensional Navier–Stokes equations. Together with Colombo, Crippa and Sorella [8 ] , they reached L t 3 C x α L_t^3C_x^\alpha L t 3 C x α for every α < 1 / 3 \alpha<1/3 α < 1/3 ; both constructions use viscosity-dependent forcing. Armstrong and Vicol [5 ] introduced fractal homogenization and obtained anomalous diffusion for every H 1 H^1 H 1 datum, with no distinguished time and with the small scales affected by the large ones, and Burczak, Székelyhidi and Wu [10 ] obtained the zeroth law of scalar turbulence for a scalar carried by a C 1 / 3 − C^{1/3-} C 1/3 − weak solution of the Euler equations. Johansson and Sorella [27 ] achieved anomalous dissipation with an autonomous two-dimensional C α C^\alpha C α field, α < 1 \alpha< 1 α < 1 , and Hess-Childs and Rowan [24 ] built a universal total dissipator, a single L t ∞ C x α L^\infty_t C^\alpha_x L t ∞ C x α field under which every datum loses all of its variance. On the random side, Rowan [33 ] proved anomalous diffusion for the Kraichnan ensemble and constructed correlated-in-time variants with no anomalous dissipation at any fixed positive correlation time; Bagnara, Boutros, De Lellis and the second author [7 ] established regularity thresholds above which anomalous dissipation, Richardson diffusion and the Obukhov–Corrsin law are impossible for certain classes of autonomous random fields; and Colombo, Hess-Childs and Rowan [12 ] established uniqueness above Hölder exponent 1 / 2 1/2 1/2 , away from zeros, for autonomous random multiscale fields, and constructed nonuniqueness examples below this threshold in d ≥ 4 d \ge4 d ≥ 4 . For refreshing ensembles they obtained corresponding thresholds and quantitative separation estimates, including lower bounds for the spread of trajectories in their nonuniqueness constructions. With the exception of [25 ] discussed above, none of these yields a Richardson law.
Anomalous dissipation is not unrelated to Richardson dispersion, but it is far from implying it, implicitly or explicitly. By the Lagrangian fluctuation–dissipation relation of Drivas and Eyink [15 ] , anomalous dissipation for passive scalars is intimately tied to spontaneous stochasticity. Informally, spontaneous stochasticity asserts that initially nearby or coincident trajectories retain a nontrivial separation in the zero-noise limit, without prescribing a quantitative growth law. Richardson dispersion is substantially stronger: it specifies the exact scaling power E R t 2 ∼ t 3 ER_t^2 \sim t^3 E R t 2 ∼ t 3 over the inertial range. And indeed, the numerical evidence and scaling arguments show that in synthetic example the dissipation exponent can be different from 3 / 2 3/2 3/2 .
This is connected to an important issue emphasized in the physics literature, the role of sweeping in synthetic turbulence models. In standard kinematic simulations the Fourier modes representing small eddies evolve without being materially transported by the larger scales. Thomson and Devenish [35 ] argued that this absence of sweeping gives small-scale velocity increments along particle trajectories an incorrect decorrelation time and can change the predicted pair-dispersion exponent; see also the subsequent discussion in [31 , 14 ] and the analytical treatment of sweeping decorrelation by Eyink and Benveniste [18 ] . This issue is specific to synthetic Eulerian velocity models: in an actual flow, a small eddy and two nearby tracers are to leading order transported together by the larger-scale velocity. To the best of our knowledge, aside from the above, the only mathematical result taking sweeping into account is the aforementioned new model of Armstrong and Vicol [5 ] , cf. also [6 , 10 ] which brought powerful homogenization techniques to the table. Their velocity however is still a carefully engineered, deterministic, space-time periodic fractal field built from alternating shear flows, rather than a random stationary turbulence model, and their scale hierarchy is extremely lacunary: the active length scales are separated double-exponentially, rather than forming anything like a dense geometric inertial range. Their theorem concerns anomalous scalar diffusion only, not the Richardson pair-dispersion power law. In a different but related multiscale homogenization problem, the authors of [1 , 2 ] remove scale separation for a stationary random incompressible drift and prove one-particle superdiffusion, E ∣ X t ∣ 2 ∼ t ( log t ) 1 / 2 \mathbb{E}|X_t|^2 \sim t(\log t)^{1/2} E ∣ X t ∣ 2 ∼ t ( log t ) 1/2 in the critical case and t 2 / ( 2 − γ ) t^{2/(2-\gamma)} t 2/ ( 2 − γ ) for self-similar drifts with small Hurst exponent γ \gamma γ ; these are not Richardson pair-dispersion theorems either.
In the present paper and its companion in physics, [26 ] , we introduce a random velocity with prescribed spatial Hölder regularity and scale-dependent renewal times. The same hierarchy of scales is renewed throughout the time interval, with smaller eddies transported by larger ones. We prove Richardson dispersion and anomalous dissipation for this field. The tracer particles satisfy deterministic ordinary differential equations in a common realization of the velocity, all randomness comes from the velocity ensemble. We prove in Theorem 1.1 , uniformly in the ultraviolet cutoff and for every initial separation r 0 ≥ ℓ J r_0 \ge\ell_J r 0 ≥ ℓ J , that
c ( r 0 1 − α + κ t ) 2 1 − α ≤ E R t 2 ≤ C ( r 0 1 − α + t ) 2 1 − α . c\left(r_0^{1-\alpha}+\kappa t\right)^{\frac{2}{1-\alpha}} \le\mathbb{E}R_t^2 \le C\left(r_0^{1-\alpha}+t\right)^{\frac{2}{1-\alpha}}. c ( r 0 1 − α + κ t ) 1 − α 2 ≤ E R t 2 ≤ C ( r 0 1 − α + t ) 1 − α 2 .
Here, κ \kappa κ is the Kubo number, a dimensionless constant measuring the ratio between the correlation time of an eddy and its characteristic turnover time. In our model, an eddy of size ℓ \ell ℓ has lifetime τ ℓ ≈ κ ℓ 1 − α \tau_\ell\approx\kappa\ell^{1-\alpha} τ ℓ ≈ κ ℓ 1 − α , so κ \kappa κ controls how long structures persist relative to the natural scale-dependent advective time. We need κ \kappa κ to be small for the proof – perhaps, a limitation of the argument rather than a necessity as the numerical results in [26 ] show, and the left and right bound differ with κ \kappa κ . However, it does not have to vanish and depends on the parameters of the system only.
In particular, when α = 1 / 3 \alpha= 1/3 α = 1/3 and the Kubo number is fixed, in the admissible time range,
κ R t 2 ≤ E R t 2 ≤ C t 3 . \kappa R_t^2 \le\mathbb{E}R_t^2 \le Ct^3. κ R t 2 ≤ E R t 2 ≤ C t 3 .
Thus the model exhibits the Richardson cubic law with constants independent of the small-scale cutoff, for a random velocity with finite, scale-dependent correlation times and with the smaller eddies transported by the larger ones. Furthermore, we establish anomalous dissipation.
With this, let us pass to the model.
The model
Fix an integer d ≥ 2 d \ge2 d ≥ 2 and 0 < α < 1 0 < \alpha< 1 0 < α < 1 , and put β = 1 − α \beta= 1-\alpha β = 1 − α . The physical space is T d = R d / Z d \mathbb{T}^d = \mathbb{R}^d/\mathbb{Z}^d T d = R d / Z d , equipped with its flat distance. Let R > 0 R > 0 R > 0 and let ψ ∈ C c 4 ( R d ) \psi\in C_c^4(\mathbb{R}^d) ψ ∈ C c 4 ( R d ) be radial, not identically zero, and supported in B ( 0 , R ) B(0,R) B ( 0 , R ) . Let e 1 , … , e d e_1,\ldots,e_d e 1 , … , e d be the coordinate vectors and set
A = { ( p , q ) : 1 ≤ p < q ≤ d } , J p q = e q e p T − e p e q T , W p q = J p q ∇ ψ . \mathcal{A} = \{(p,q):1 \le p < q \le d\}, \qquad J^{pq}=e_q e_p^{T}-e_p e_q^{T}, \qquad W^{pq}=J^{pq}\nabla\psi. A = {( p , q ) : 1 ≤ p < q ≤ d } , J pq = e q e p T − e p e q T , W pq = J pq ∇ ψ .
Every W ν W^\nu W ν , ν ∈ A \nu\in\mathcal{A} ν ∈ A , belongs to C c 3 ( R d ; R d ) C_c^3(\mathbb{R}^d;\mathbb{R}^d) C c 3 ( R d ; R d ) and is divergence-free, because J ν J^\nu J ν is skew-symmetric. In dimension two there is a single mode, W 12 = ( − ∂ 2 ψ , ∂ 1 ψ ) = ∇ ⊥ ψ W^{12}=(-\partial_2\psi,\partial_1\psi)=\nabla^\perp\psi W 12 = ( − ∂ 2 ψ , ∂ 1 ψ ) = ∇ ⊥ ψ .
The averaged velocity increments of one scale of the model are governed by the structure function of this family. For z ∈ R d z \in\mathbb{R}^d z ∈ R d define
D ( z ) = ∑ ν ∈ A ∫ R d [ W ν ( y ) − W ν ( y − z ) ] ⊗ [ W ν ( y ) − W ν ( y − z ) ] d y . (1) D(z)=\sum_{\nu\in\mathcal{A}}\int_{\mathbb{R}^d}[W^\nu(y)-W^\nu(y-z)]\otimes[W^\nu(y)-W^\nu(y-z)]\,\mathrm{d}y. \tag*{(1)} D ( z ) = ν ∈ A ∑ ∫ R d [ W ν ( y ) − W ν ( y − z )] ⊗ [ W ν ( y ) − W ν ( y − z )] d y . (1)
Lemma 2.5 shows that, for z ≠ 0 z \ne0 z = 0 ,
D ( z ) = D L ( ∣ z ∣ ) z ^ ⊗ z ^ + D N ( ∣ z ∣ ) ( I − z ^ ⊗ z ^ ) , z ^ = z ∣ z ∣ , (2) D(z)=D_L(|z|)\widehat{z}\otimes\widehat{z}+D_N(|z|)(I-\widehat{z}\otimes\widehat{z}),\qquad\widehat{z}=\frac{z}{|z|}, \tag*{(2)} D ( z ) = D L ( ∣ z ∣ ) z ⊗ z + D N ( ∣ z ∣ ) ( I − z ⊗ z ) , z = ∣ z ∣ z , (2)
where D N D_N D N is the eigenvalue in each of the d − 1 d-1 d − 1 transverse directions, and that D N = D L + r D L ′ / ( d − 1 ) D_N=D_L+rD_L'/(d-1) D N = D L + r D L ′ / ( d − 1 ) , the Kármán–Howarth identity [28 ] , which is the divergence-free condition written for an isotropic tensor. We are assuming the following structural hypothesis on the wavelets
(ND) D L D_L D L is nondecreasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) ,
or, equivalently D N ≥ D L D_N\ge D_L D N ≥ D L . Proposition 2.6 verifies (ND) for ψ ( x ) = ( 1 − ∣ x ∣ 2 ) + 5 \psi(x)=(1-|x|^2)_+^5 ψ ( x ) = ( 1 − ∣ x ∣ 2 ) + 5 in every dimension d ≥ 2 d\ge2 d ≥ 2 .
We defer the detailed discussion of the condition (ND) to the next section, but here let us just mention that it ensures the correct sign of the renewal gains – the main contribution to our observable, and that it is rather generic for standard families of wavelets.
Recall that d ≥ 2 d\ge2 d ≥ 2 , 0 < α < 1 0<\alpha<1 0 < α < 1 , and β = 1 − α \beta=1-\alpha β = 1 − α . Let κ > 0 \kappa>0 κ > 0 be the Kubo number and define
ℓ j = 2 − j , A j = ℓ j α , λ j = ℓ j − β , τ j = κ 2 − ⌈ β j ⌉ . (3) \ell_j=2^{-j},\qquad A_j=\ell_j^\alpha,\qquad\lambda_j=\ell_j^{-\beta},\qquad\tau_j=\kappa2^{-\lceil\beta j\rceil}. \tag*{(3)} ℓ j = 2 − j , A j = ℓ j α , λ j = ℓ j − β , τ j = κ 2 − ⌈ β j ⌉ . (3)
Set
j 0 = min { j ≥ 0 : R ℓ j ≤ 1 / 4 } , J ≥ j 0 , I j = { 0 , … , 2 j − 1 } d . j_0=\min\{j\ge0:R\ell_j\le1/4\},\qquad J\ge j_0,\qquad I_j=\{0,\ldots,2^j-1\}^d. j 0 = min { j ≥ 0 : R ℓ j ≤ 1/4 } , J ≥ j 0 , I j = { 0 , … , 2 j − 1 } d .
Thus A j A_j A j is the amplitude of scale j j j , λ j = A j / ℓ j \lambda_j=A_j/\ell_j λ j = A j / ℓ j the size of its gradient, and τ j \tau_j τ j its lifetime. The ratio of lifetime to turnover time lies in ( κ / 2 , κ ] (\kappa/2,\kappa] ( κ /2 , κ ] ; the dyadic rounding aligns the renewal times across scales (Lemma 2.1 ). The integer J J J is finite but otherwise arbitrary; all constants below are independent of J J J . Put y i j = ℓ j i ( m o d Z d ) y_i^j=\ell_j i\pmod{\mathbb{Z}^d} y i j = ℓ j i ( mod Z d ) and Q j = [ 0 , ℓ j ) d Q_j=[0,\ell_j)^d Q j = [ 0 , ℓ j ) d . The periodized wavelets are
W j ν ( x ) = ∑ n ∈ Z d W ν ( ( x + n ) / ℓ j ) , x ∈ T d . W_j^\nu(x)=\sum_{n\in\mathbb{Z}^d}W^\nu((x+n)/\ell_j),\qquad x\in\mathbb{T}^d. W j ν ( x ) = n ∈ Z d ∑ W ν (( x + n ) / ℓ j ) , x ∈ T d .
Let K j = ⌈ τ j − 1 ⌉ K_j=\lceil\tau_j^{-1}\rceil K j = ⌈ τ j − 1 ⌉ and I j k = [ k τ j , ( k + 1 ) τ j ) ∩ [ 0 , 1 ) I_{jk}=[k\tau_j,(k+1)\tau_j)\cap[0,1) I j k = [ k τ j , ( k + 1 ) τ j ) ∩ [ 0 , 1 ) for 0 ≤ k < K j 0\le k<K_j 0 ≤ k < K j . For each such ( j , k ) (j,k) ( j , k ) , let ξ k j \xi_k^j ξ k j be a random variable uniformly distributed on Q j Q_j Q j . For i ∈ I j i\in I_j i ∈ I j and ν ∈ A \nu\in\mathcal{A} ν ∈ A , let η i , k j , ν \eta_{i,k}^{j,\nu} η i , k j , ν be real random variables satisfying
E η i , k j , ν = 0 , E ( η i , k j , ν ) 2 = σ 2 > 0 , ∣ η i , k j , ν ∣ ≤ M . (4) \mathbb{E}\eta_{i,k}^{j,\nu}=0,\qquad\mathbb{E}(\eta_{i,k}^{j,\nu})^2=\sigma^2>0,\qquad|\eta_{i,k}^{j,\nu}|\le M. \tag*{(4)} E η i , k j , ν = 0 , E ( η i , k j , ν ) 2 = σ 2 > 0 , ∣ η i , k j , ν ∣ ≤ M . (4)
The last condition means that the law of each coefficient is supported in [ − M , M ] [-M,M] [ − M , M ] ; we regard each coefficient as a coordinate of the data space with values in [ − M , M ] [-M,M] [ − M , M ] , so that the bound holds for every realization. All the variables ξ k j \xi_k^j ξ k j and η i , k j , ν \eta_{i,k}^{j,\nu} η i , k j , ν are independent. Independent signs ± 1 \pm1 ± 1 with probability 1 / 2 1/2 1/2 are admissible with M = σ = 1 M=\sigma=1 M = σ = 1 . We realize the data for all scales j ≥ j 0 j\ge j_0 j ≥ j 0 on one countable product probability space; E \mathbb{E} E and P \mathbb{P} P denote expectation and probability on this space. Clearly, each finite truncation uses only finitely many of these variables. Thus the data space is the product of the cubes Q j Q_j Q j , one for each window, and of the intervals [ − M , M ] [-M,M] [ − M , M ] , one for each coefficient.
The fields are defined inductively in j j j . Write u < j : = ∑ m = j 0 j − 1 u m u_{<j} := \sum_{m=j_0}^{j-1} u_m u < j := ∑ m = j 0 j − 1 u m , with u < j 0 = 0 u_{<j_0}=0 u < j 0 = 0 . On I j k I_{jk} I j k , the centers solve
P ˙ i , k j ( t ) = u < j ( P i , k j ( t ) , t ) , P i , k j ( k τ j ) = y i j + ξ k j . (5) \dot{P}^{j}_{i,k}(t)=u_{<j}(P^{j}_{i,k}(t),t),\qquad P^{j}_{i,k}(k\tau_j)=y^j_i+\xi^j_k. \tag*{(5)} P ˙ i , k j ( t ) = u < j ( P i , k j ( t ) , t ) , P i , k j ( k τ j ) = y i j + ξ k j . (5)
That is, every k τ j k\tau_j k τ j the centers are placed roughly on a lattice (with some amount of randomness, each uniformly in its own cube) and then they are carried forward by the flow of larger eddies until the next reset after the time τ j \tau_j τ j .
Define
u j ( x , t ) = A j ∑ i ∈ I j ∑ ν ∈ A η i , k j , ν W j ν ( x − P i , k j ( t ) ) ( t ∈ I j k ) , u = u ( J ) : = ∑ j = j 0 J u j . (6) u_j(x,t)=A_j\sum_{i\in I_j}\sum_{\nu\in\mathcal{A}}\eta^{j,\nu}_{i,k}W^\nu_j(x-P^j_{i,k}(t))\quad(t\in I_{jk}),\qquad u=u^{(J)}:=\sum_{j=j_0}^{J}u_j. \tag*{(6)} u j ( x , t ) = A j i ∈ I j ∑ ν ∈ A ∑ η i , k j , ν W j ν ( x − P i , k j ( t )) ( t ∈ I j k ) , u = u ( J ) := j = j 0 ∑ J u j . (6)
Values at t = 1 t=1 t = 1 are assigned by left limits; they do not affect the trajectories. Centers of generation j j j are reset at its renewal times k τ j k\tau_j k τ j . Every renewal of a coarser scale is also a scheduled renewal of all finer scales (Lemma 2.1 ).
Here and below N = { 0 , 1 , 2 , … } \mathbb{N}=\{0,1,2,\ldots\} N = { 0 , 1 , 2 , … } . By Lemma 2.1 , the union of all renewal times for the truncation at J J J is τ J N ∩ [ 0 , 1 ) \tau_J\mathbb{N}\cap[0,1) τ J N ∩ [ 0 , 1 ) . Each u j u_j u j depends only on the data of scales j 0 , … , j j_0,\ldots,j j 0 , … , j and does not depend on the truncation index J J J . Thus all the fields u ( J ) u^{(J)} u ( J ) above are consistently defined, using the same data at their shared scales.
Each u ( J ) u^{(J)} u ( J ) is the truncation at scale J J J of the random field
u ( ∞ ) : = ∑ j ≥ j 0 u j , (7) u^{(\infty)}:=\sum_{j\ge j_0}u_j, \tag*{(7)} u ( ∞ ) := j ≥ j 0 ∑ u j , (7)
defined on the same probability space: u j u_j u j does not depend on J J J , and ∣ u j ∣ ≤ C ∗ A j |u_j|\le C_*A_j ∣ u j ∣ ≤ C ∗ A j is summable (Lemma 2.2 ), so the series converges uniformly in ( x , t ) (x,t) ( x , t ) , and u ( ∞ ) u^{(\infty)} u ( ∞ ) is bounded, Borel and divergence-free, with ∥ u ( ∞ ) ( ⋅ , t ) ∥ C α ≤ C ∗ \|u^{(\infty)}(\cdot,t)\|_{C^\alpha}\le C_* ∥ u ( ∞ ) ( ⋅ , t ) ∥ C α ≤ C ∗ for every t t t . The field u ( ∞ ) u^{(\infty)} u ( ∞ ) is the object of Theorem 1.5 and of Corollary 10.10 in Subsection 10.5.
For deterministic initial positions x 0 , x ~ 0 ∈ T d x_0,\widetilde{x}_0\in\mathbb{T}^d x 0 , x 0 ∈ T d , let
x ˙ t = u ( x t , t ) , x ~ ˙ t = u ( x ~ t , t ) , R t = dist ( x t , x ~ t ) , r 0 = R 0 . (8) \dot{x}_t=u(x_t,t),\qquad\dot{\widetilde{x}}_t=u(\widetilde{x}_t,t),\qquad R_t=\operatorname{dist}(x_t,\widetilde{x}_t),\qquad r_0=R_0. \tag*{(8)} x ˙ t = u ( x t , t ) , x ˙ t = u ( x t , t ) , R t = dist ( x t , x t ) , r 0 = R 0 . (8)
When R t ≤ 1 / 4 R_t\le1/4 R t ≤ 1/4 we write z t z_t z t for the unique shortest representative in R d \mathbb{R}^d R d of x t − x ~ t x_t-\widetilde{x}_t x t − x t ; then ∣ z t ∣ = R t |z_t|=R_t ∣ z t ∣ = R t .
Main results
Theorem 1.1. Fix an integer d ≥ 2 d\ge2 d ≥ 2 and consider the model of § 1.2 . Assume (ND) and (1.4). There are constants κ 0 ∈ ( 0 , e − 1 ) \kappa_0\in(0,e^{-1}) κ 0 ∈ ( 0 , e − 1 ) and c , C , C H > 0 c,C,C_H>0 c , C , C H > 0 , depending only on d , α , ψ , R , M , σ d,\alpha,\psi,R,M,\sigma d , α , ψ , R , M , σ (C H C_H C H is the constant of (xref)(2.4) below), with the following property. For every 0 < κ ≤ κ 0 0<\kappa\le\kappa_0 0 < κ ≤ κ 0 , every J ≥ j 0 J\ge j_0 J ≥ j 0 , and every deterministic pair of initial positions with
ℓ J ≤ r 0 ≤ 1 / 8 , \ell_J\le r_0\le1/8, ℓ J ≤ r 0 ≤ 1/8 , one has δ κ : = C κ ( 1 + ∣ log κ ∣ ) ≤ 1 / 4 \delta_\kappa:=C\kappa(1+|\log\kappa|)\le1/4 δ κ := C κ ( 1 + ∣ log κ ∣ ) ≤ 1/4 , and for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] satisfying
r 0 β + C H t ≤ ( 1 / 4 ) β , (9) r_0^\beta+C_Ht\le(1/4)^\beta, \tag*{(9)} r 0 β + C H t ≤ ( 1/4 ) β , (9) ( 1 + δ κ ) E R t β ≥ r 0 β − δ κ ℓ J β + c κ t ≥ ( 1 − δ κ ) r 0 β + c κ t . (10) (1+\delta_\kappa)\mathbb{E}R_t^\beta\ge r_0^\beta-\delta_\kappa\ell_J^\beta+c\kappa t\ge(1-\delta_\kappa)r_0^\beta+c\kappa t. \tag*{(10)} ( 1 + δ κ ) E R t β ≥ r 0 β − δ κ ℓ J β + c κ t ≥ ( 1 − δ κ ) r 0 β + c κ t . (10)
Moreover, after changing c c c and C C C ,
c ( r 0 β + κ t ) 2 / β ≤ E R t 2 ≤ C ( r 0 β + t ) 2 / β . (11) c(r_{0}^{\beta}+\kappa t)^{2/\beta} \le ER_{t}^{2} \le C(r_{0}^{\beta}+t)^{2/\beta}. \tag*{(11)} c ( r 0 β + κ t ) 2/ β ≤ E R t 2 ≤ C ( r 0 β + t ) 2/ β . (11)
All constants are independent of J J J , r 0 r_{0} r 0 , and κ \kappa κ in the specified ranges.
Corollary 1.2. Let d ≥ 2 d \ge2 d ≥ 2 and α = 1 / 3 \alpha= 1/3 α = 1/3 . For times satisfying (9) and t ≥ r 0 2 / 3 / κ t \ge r_{0}^{2/3}/\kappa t ≥ r 0 2/3 / κ ,
κ κ 3 t 3 ≤ E R t 2 ≤ C t 3 . \kappa\kappa^{3}t^{3} \le ER_{t}^{2} \le Ct^{3}. κ κ 3 t 3 ≤ E R t 2 ≤ C t 3 .
Thus, for each fixed sufficiently small Kubo number, the mean-square separation has a cubic lower bound on this time range, uniformly in the small-scale cutoff; the time range is nonempty when r 0 r_{0} r 0 is small compared with a power of κ \kappa κ . The upper bound in (11) is a consequence of the Hölder continuity of u u u alone. The exponent 2 / β 2/\beta 2/ β is therefore sharp.
The passive scalar
Fix ε 0 ∈ ( 0 , 1 ] \varepsilon_{0} \in(0,1] ε 0 ∈ ( 0 , 1 ] and set, for J ≥ j 0 J \ge j_{0} J ≥ j 0 ,
μ J : = ε 0 ℓ J 1 + α . \mu_{J}:=\varepsilon_{0}\ell_{J}^{1+\alpha}. μ J := ε 0 ℓ J 1 + α .
Thus the Péclet number of the finest scale, A ˙ J ℓ J / μ J = 1 / ε 0 \dot{A}_{J}\ell_{J}/\mu_{J}=1/\varepsilon_{0} A ˙ J ℓ J / μ J = 1/ ε 0 , does not depend on J J J . For a real θ 0 ∈ L 2 ( T d ) \theta_{0} \in L^{2}(\mathbb{T}^{d}) θ 0 ∈ L 2 ( T d ) and a realization of the data, let θ = θ ( J ) \theta=\theta^{(J)} θ = θ ( J ) be the energy solution (Definition 8.1) on [ 0 , 1 ] [0,1] [ 0 , 1 ] of
∂ t θ + u ( J ) ⋅ ∇ θ = μ J Δ θ , θ ( 0 ) = θ 0 . (12) \partial_{t}\theta+u^{(J)}\cdot\nabla\theta=\mu_{J}\Delta\theta,\qquad\theta(0)=\theta_{0}. \tag*{(12)} ∂ t θ + u ( J ) ⋅ ∇ θ = μ J Δ θ , θ ( 0 ) = θ 0 . (12)
It exists and is unique, its restriction to [ 0 , t 0 ] [0,t_{0}] [ 0 , t 0 ] is the energy solution on [ 0 , t 0 ] [0,t_{0}] [ 0 , t 0 ] , ∫ θ ( t ) = ∫ θ 0 \int\theta(t)=\int\theta_{0} ∫ θ ( t ) = ∫ θ 0 , and for t 0 ∈ ( 0 , 1 ] t_{0}\in(0,1] t 0 ∈ ( 0 , 1 ] the energy identity holds (Lemma 8.2 ):
∥ θ ( J ) ( t 0 ) ∥ 2 + D J ( θ 0 ) = ∥ θ 0 ∥ 2 , D J ( θ 0 ) : = 2 μ J ∫ 0 t 0 ∥ ∇ θ ( J ) ( t ) ∥ 2 d t . (13) \|\theta^{(J)}(t_{0})\|^{2}+\mathbf{D}_{J}(\theta_{0})=\|\theta_{0}\|^{2},\qquad\mathbf{D}_{J}(\theta_{0}):=2\mu_{J}\int_{0}^{t_{0}}\|\nabla\theta^{(J)}(t)\|^{2}\,dt. \tag*{(13)} ∥ θ ( J ) ( t 0 ) ∥ 2 + D J ( θ 0 ) = ∥ θ 0 ∥ 2 , D J ( θ 0 ) := 2 μ J ∫ 0 t 0 ∥∇ θ ( J ) ( t ) ∥ 2 d t . (13)
Here and below ∥ ⋅ ∥ \|\cdot\| ∥ ⋅ ∥ is the norm of L 2 ( T d ) L^{2}(\mathbb{T}^{d}) L 2 ( T d ) (or of L 2 ( T 2 d ) L^{2}(\mathbb{T}^{2d}) L 2 ( T 2 d ) ). The quantity D J ( θ 0 ) ∈ [ 0 , ∥ θ 0 ∥ 2 ] \mathbf{D}_{J}(\theta_{0})\in[0,\|\theta_{0}\|^{2}] D J ( θ 0 ) ∈ [ 0 , ∥ θ 0 ∥ 2 ] is the variance dissipated by time t 0 t_{0} t 0 .
Anomalous dissipation
The mean dissipation is computed mode by mode under a symmetry assumption on the coefficient laws.
Definition 1.3 (Hypothesis (S)). For every scale j j j and window k k k , the coefficients η i , k j , ν \eta_{i,k}^{j,\nu} η i , k j , ν , i ∈ I j i\in I_{j} i ∈ I j , ν ∈ A \nu\in\mathcal{A} ν ∈ A , are identically distributed, and their common law is symmetric (η = d − η \eta\overset{d}{=}-\eta η = d − η ).
Independent signs ± 1 \pm1 ± 1 with probability 1 2 \frac{1}{2} 2 1 satisfy (4) and (S).
Retain the notation κ ∗ \kappa_{*} κ ∗ from Lemma 2.2 .
Theorem 1.4 (Anomalous dissipation). Let d ≥ 2 d \ge2 d ≥ 2 and assume (ND), (4) and (S). There are κ 1 ∈ ( 0 , κ ∗ ] \kappa_{1}\in(0,\kappa_{*}] κ 1 ∈ ( 0 , κ ∗ ] and c > 0 c>0 c > 0 , depending only on d d d , α \alpha α , ψ \psi ψ , R R R , M M M , σ \sigma σ , and for every κ ∈ ( 0 , κ 1 ] \kappa\in(0,\kappa_{1}] κ ∈ ( 0 , κ 1 ] a number ε 1 ( κ ) ∈ ( 0 , 1 ] \varepsilon_{1}(\kappa)\in(0,1] ε 1 ( κ ) ∈ ( 0 , 1 ] , with the following property. Consider the model with κ ∈ ( 0 , κ 1 ] \kappa\in(0,\kappa_{1}] κ ∈ ( 0 , κ 1 ] , let ε 0 ∈ ( 0 , ε 1 ( κ ) ] \varepsilon_{0}\in(0,\varepsilon_{1}(\kappa)] ε 0 ∈ ( 0 , ε 1 ( κ )] and t 0 ∈ ( 0 , 1 ] t_{0}\in(0,1] t 0 ∈ ( 0 , 1 ] , and put c A D : = c ( κ t 0 ) 2 / β c_{\mathrm{AD}}:=c(\kappa t_{0})^{2/\beta} c AD := c ( κ t 0 ) 2/ β . Then for every real θ 0 ∈ L 2 ( T d ) \theta_{0}\in L^{2}(\mathbb{T}^{d}) θ 0 ∈ L 2 ( T d ) with ∫ θ 0 = 0 \int\theta_{0}=0 ∫ θ 0 = 0 :
(a) lim inf J → ∞ E D J ( θ 0 ) ≥ c A D ∥ θ 0 ∥ 2 \liminf_{J\to\infty}\mathbb{E}\mathbf{D}_{J}(\theta_{0})\ge c_{\mathrm{AD}}\|\theta_{0}\|^{2} lim inf J → ∞ E D J ( θ 0 ) ≥ c AD ∥ θ 0 ∥ 2 ;
(b) if θ 0 ≠ 0 \theta_{0}\ne0 θ 0 = 0 , then P ( lim sup J → ∞ D J ( θ 0 ) > 0 ) ≥ c A D \mathbb{P}\left(\limsup_{J\to\infty}\mathbf{D}_{J}(\theta_{0})>0\right)\ge c_{\mathrm{AD}} P ( lim sup J → ∞ D J ( θ 0 ) > 0 ) ≥ c AD ;
(c) for every γ > 0 \gamma> 0 γ > 0 and A ≥ 1 A \ge1 A ≥ 1 there is J ∗ = J ∗ ( γ , A , κ , ε 0 , t 0 ) J_* = J_*(\gamma, A, \kappa, \varepsilon_0, t_0) J ∗ = J ∗ ( γ , A , κ , ε 0 , t 0 ) , depending also on the fixed model parameters, such that E D J ( θ 0 ) ≥ 1 2 c A D ∥ θ 0 ∥ 0 2 \mathbb{E}D_J(\theta_0) \ge\frac{1}{2}c_{\mathrm{AD}}\|\theta_0\|_0^2 E D J ( θ 0 ) ≥ 2 1 c AD ∥ θ 0 ∥ 0 2 for all J ≥ J ∗ J \ge J_* J ≥ J ∗ and every real mean-zero θ 0 ∈ H γ ( T d ) \theta_0 \in H^\gamma(\mathbb{T}^d) θ 0 ∈ H γ ( T d ) satisfying ∥ θ 0 ∥ H ˙ γ ≤ A ∥ θ 0 ∥ 0 \|\theta_0\|_{\dot H^\gamma} \le A\|\theta_0\|_0 ∥ θ 0 ∥ H ˙ γ ≤ A ∥ θ 0 ∥ 0 .
The distinction between (a) and (c) is natural: L 2 L^2 L 2 is the natural space for scalar variance, while a bound on the Sobolev-to-L 2 L^2 L 2 norm ratio gives uniformly small high-frequency tails relative to the initial variance and hence a threshold uniform over the corresponding initial data. Here the homogeneous Sobolev seminorm is
∥ θ 0 ∥ H ˙ γ 2 : = ∑ k ∈ Z d ( 2 π ∣ k ∣ ) 2 γ ∣ θ 0 ^ ( k ) ∣ 2 . \|\theta_0\|_{\dot H^\gamma}^2 := \sum_{k\in\mathbb{Z}^d}(2\pi|k|)^{2\gamma}|\widehat{\theta_0}(k)|^2. ∥ θ 0 ∥ H ˙ γ 2 := k ∈ Z d ∑ ( 2 π ∣ k ∣ ) 2 γ ∣ θ 0 ( k ) ∣ 2 .
Recall from § 1.2 that u ( J ) u^{(J)} u ( J ) is the truncation at scale J J J of the random field u ( ∞ ) u^{(\infty)} u ( ∞ ) of (7) , which is bounded, Borel, divergence-free and C α C^\alpha C α in x x x , uniformly in t t t . For μ > 0 \mu> 0 μ > 0 and a real θ 0 ∈ L 2 ( T d ) \theta_0 \in L^2(\mathbb{T}^d) θ 0 ∈ L 2 ( T d ) let θ μ \theta^\mu θ μ be the energy solution on [ 0 , 1 ] [0,1] [ 0 , 1 ] of
∂ t θ μ + u ( ∞ ) ⋅ ∇ θ μ = μ Δ θ μ , θ μ ( 0 ) = θ 0 , (14) \partial_t\theta^\mu+ u^{(\infty)}\cdot\nabla\theta^\mu= \mu\Delta\theta^\mu,\qquad\theta^\mu(0)=\theta_0, \tag*{(14)} ∂ t θ μ + u ( ∞ ) ⋅ ∇ θ μ = μ Δ θ μ , θ μ ( 0 ) = θ 0 , (14)
and for t 0 ∈ ( 0 , 1 ] t_0 \in(0,1] t 0 ∈ ( 0 , 1 ] put
D ∞ μ ( θ 0 ) : = 2 μ ∫ 0 t 0 ∥ ∇ θ μ ( t ) ∥ 2 d t . D^\mu_\infty(\theta_0):=2\mu\int_0^{t_0}\|\nabla\theta^\mu(t)\|^2\,dt. D ∞ μ ( θ 0 ) := 2 μ ∫ 0 t 0 ∥∇ θ μ ( t ) ∥ 2 d t .
The following theorem concerns this single field, which depends neither on μ \mu μ nor on ε 0 \varepsilon_0 ε 0 .
Theorem 1.5 (Anomalous dissipation for the fixed field). Assume the hypotheses of Theorem 1.4 , fix κ ∈ ( 0 , κ 1 ] \kappa\in(0,\kappa_1] κ ∈ ( 0 , κ 1 ] and ε 0 ∈ ( 0 , ε 1 ( κ ) ] \varepsilon_0 \in(0,\varepsilon_1(\kappa)] ε 0 ∈ ( 0 , ε 1 ( κ )] , let t 0 ∈ ( 0 , 1 ] t_0 \in(0,1] t 0 ∈ ( 0 , 1 ] , and let c A D c_{\mathrm{AD}} c AD be as in Theorem 1.4 . There is a ∗ ∈ ( 0 , 1 ) a_* \in(0,1) a ∗ ∈ ( 0 , 1 ) , depending only on d , α , ψ , R , M d,\alpha,\psi,R,M d , α , ψ , R , M and ε 0 \varepsilon_0 ε 0 (explicitly, (101) ), such that for every real θ 0 ∈ L 2 ( T d ) \theta_0 \in L^2(\mathbb{T}^d) θ 0 ∈ L 2 ( T d ) with ∫ θ 0 = 0 \int\theta_0=0 ∫ θ 0 = 0 :
(a) lim inf μ ↓ 0 E D ∞ μ ( θ 0 ) ≥ a ∗ c A D ∥ θ 0 ∥ 2 \liminf_{\mu\downarrow0}\mathbb{E}D^\mu_\infty(\theta_0) \ge a_*c_{\mathrm{AD}}\|\theta_0\|^2 lim inf μ ↓ 0 E D ∞ μ ( θ 0 ) ≥ a ∗ c AD ∥ θ 0 ∥ 2 ;
(b) if θ 0 ≠ 0 \theta_0\ne0 θ 0 = 0 , then P ( lim sup μ ↓ 0 D ∞ μ ( θ 0 ) > 0 ) ≥ c A D \mathbb{P}(\limsup_{\mu\downarrow0}D^\mu_\infty(\theta_0)>0)\ge c_{\mathrm{AD}} P ( lim sup μ ↓ 0 D ∞ μ ( θ 0 ) > 0 ) ≥ c AD ;
(c) for every γ > 0 \gamma> 0 γ > 0 and A ≥ 1 A \ge1 A ≥ 1 there is μ ∗ = μ ∗ ( γ , A , κ , ε 0 , t 0 ) > 0 \mu_*=\mu_*(\gamma,A,\kappa,\varepsilon_0,t_0)>0 μ ∗ = μ ∗ ( γ , A , κ , ε 0 , t 0 ) > 0 , depending also on the fixed model parameters, such that E D ∞ μ ( θ 0 ) ≥ 1 2 a ∗ c A D ∥ θ 0 ∥ 2 \mathbb{E}D^\mu_\infty(\theta_0)\ge\frac{1}{2}a_*c_{\mathrm{AD}}\|\theta_0\|^2 E D ∞ μ ( θ 0 ) ≥ 2 1 a ∗ c AD ∥ θ 0 ∥ 2 for all μ ∈ ( 0 , μ ∗ ] \mu\in(0,\mu_*] μ ∈ ( 0 , μ ∗ ] and every real mean-zero θ 0 ∈ H γ ( T d ) \theta_0\in H^\gamma(\mathbb{T}^d) θ 0 ∈ H γ ( T d ) satisfying ∥ θ 0 ∥ H ˙ γ ≤ A ∥ θ 0 ∥ \|\theta_0\|_{\dot H^\gamma}\le A\|\theta_0\| ∥ θ 0 ∥ H ˙ γ ≤ A ∥ θ 0 ∥ .
As in Theorem 1.4 , part (c) gives a threshold μ ∗ \mu_* μ ∗ uniform over all nonzero real mean-zero Sobolev data with ∥ θ 0 ∥ H ˙ γ / ∥ θ 0 ∥ ≤ A \|\theta_0\|_{\dot H^\gamma}/\|\theta_0\|\le A ∥ θ 0 ∥ H ˙ γ /∥ θ 0 ∥ ≤ A , whereas (a) also covers data belonging to no H γ H^\gamma H γ with γ > 0 \gamma>0 γ > 0 . Again the factor 1 2 \frac{1}{2} 2 1 in (c) may be replaced by any 1 − ϵ 1-\epsilon 1 − ϵ , 0 < ϵ < 1 0<\epsilon<1 0 < ϵ < 1 , with a threshold depending also on ϵ \epsilon ϵ .
The theorem is proved in Subsection 10.5 by comparing the dissipation for u ( ∞ ) u^{(\infty)} u ( ∞ ) with the dissipation for the truncation u ( J ) u^{(J)} u ( J ) at the scale J J J for which μ J ≤ μ < 2 1 + α μ J \mu_J\le\mu<2^{1+\alpha}\mu_J μ J ≤ μ < 2 1 + α μ J ; the scales finer than J J J change the dissipation by at most a constant factor (Lemma 10.8 ). Part (b) says that, on an event of probability at least c A D c_{\mathrm{AD}} c AD , a single realization of u ( ∞ ) u^{(\infty)} u ( ∞ ) dissipates the variance of θ 0 \theta_0 θ 0 anomalously, in the sense of [16 ] . For the explicit model of Proposition 2.6 this gives a deterministic statement, Corollary 10.10 in Subsection 10.5: for every d ≥ 2 d\ge2 d ≥ 2 , 0 < α < 1 0<\alpha<1 0 < α < 1 and t 0 ∈ ( 0 , 1 ] t_0\in(0,1] t 0 ∈ ( 0 , 1 ] there is a divergence-free u ∈ L ∞ ( [ 0 , 1 ] ; C α ( T d ) ) u\in L^\infty([0,1];C^\alpha(\mathbb{T}^d)) u ∈ L ∞ ([ 0 , 1 ] ; C α ( T d )) for which the energy solutions of ∂ t θ μ + u ⋅ ∇ θ μ = μ Δ θ μ \partial_t\theta^\mu+u\cdot\nabla\theta^\mu=\mu\Delta\theta^\mu ∂ t θ μ + u ⋅ ∇ θ μ = μ Δ θ μ with θ μ ( 0 ) = cos ( 2 π x 1 ) \theta^\mu(0)=\cos(2\pi x_1) θ μ ( 0 ) = cos ( 2 π x 1 ) satisfy lim sup μ ↓ 0 μ ∫ 0 t 0 ∥ ∇ θ μ ∥ 2 d t > 0 \limsup_{\mu\downarrow0}\mu\int_0^{t_0}\|\nabla\theta^\mu\|^2\,dt>0 lim sup μ ↓ 0 μ ∫ 0 t 0 ∥∇ θ μ ∥ 2 d t > 0 .
Outline of the proof
The field u u u is reset randomly at certain times depending on the scale but otherwise every next generation is deterministically carried by the flow of previous generations. Inside a window the flow is a deterministic ordinary differential equation with a C 3 C^3 C 3 right-hand side, and the mean-zero property of the coefficients can be exploited only at the instants at which a window is refreshed. Probability enters the proof through finite-dimensional averages: at each refresh time, and there, we integrate over the fresh coefficients and lattice shift of the refreshed scale with everything else held fixed. No martingales are used; where one would condition on the past, we integrate over the variables of a single window and apply Fubini’s theorem (§ 6 ).
Richardson dispersion
We start with
The symbol of the model (§ 2 ). This is the main contribution to the observable. Averaging the increment of a freshly drawn scale-j j j field between two points at distance r r r produces σ 2 A j 2 D ( z / ℓ j ) \sigma^2 A_j^2 D(z/\ell_j) σ 2 A j 2 D ( z / ℓ j ) , where D D D is the structure function ((1) ). Accordingly, for a radial test function Φ \Phi Φ of the separation we set
Γ j ( r ) : = σ 2 A j 2 2 D ( z / ℓ j ) : D 2 Φ ( z ) , ∣ z ∣ = r , \Gamma_j(r) := \frac{\sigma^2 A_j^2}{2}D(z/\ell_j) : D^2\Phi(z), \qquad|z|=r, Γ j ( r ) := 2 σ 2 A j 2 D ( z / ℓ j ) : D 2 Φ ( z ) , ∣ z ∣ = r ,
which depends on z z z only through r = ∣ z ∣ r=|z| r = ∣ z ∣ by the isotropy ((2) ); τ j 2 Γ j ( r ) \tau_j^2\Gamma_j(r) τ j 2 Γ j ( r ) is, to leading order, the mean change of Φ ( z ) \Phi(z) Φ ( z ) produced by one window of scale j j j while the particles are at distance r r r , and exactly the averaged jump, at a refresh of scale j j j , of the corrected observable introduced below (Proposition 6.6). In the body of the paper Φ \Phi Φ is replaced by its scale-adapted version Φ j \Phi_j Φ j of § 3 , which gives the precise definition ((33) ). One computes D : D 2 Φ D:D^2\Phi D : D 2 Φ explicitly; for Φ = r β \Phi=r^\beta Φ = r β it equals β r β − 2 [ ( d − 1 ) D N − α D L ] \beta r^{\beta-2}[(d-1)D_N-\alpha D_L] β r β − 2 [( d − 1 ) D N − α D L ] . The concavity of r β r^\beta r β makes the longitudinal term negative, the d − 1 d-1 d − 1 transverse directions make the other term positive, and (ND), in the form D N ≥ D L D_N \ge D_L D N ≥ D L , makes the sum positive. This is the place where (ND) is heavily used. Counting powers at r ≈ ℓ j r\approx\ell_j r ≈ ℓ j , one window of scale j j j changes the average of R β R^\beta R β by τ j 2 Γ j ( r ) ≈ τ j 2 A j 2 r β − 2 ≈ κ 2 ℓ j β ≈ κ τ j \tau_j^2\Gamma_j(r)\approx\tau_j^2 A_j^2 r^{\beta-2}\approx\kappa^2\ell_j^\beta\approx\kappa\tau_j τ j 2 Γ j ( r ) ≈ τ j 2 A j 2 r β − 2 ≈ κ 2 ℓ j β ≈ κ τ j , that is, at rate κ \kappa κ per unit time independently of the scale; β = 1 − α \beta=1-\alpha β = 1 − α is the exponent for which the gain per window is scale-free.
Test functions (§ 3 ). We would like to estimate Φ = r β \Phi=r^\beta Φ = r β , but factually, it is modified twice, due to the low frequency and high frequency cutoffs. Below the finest scale it is continued as a logarithm: there the field is Lipschitz, log r \log r log r is the observable whose derivative along a Lipschitz flow is bounded, and the continuation keeps the gain at scale J J J bounded below uniformly in r r r . This step is the source of the uniformity in J J J . Then, for each scale b b b , a copy Φ b \Phi_b Φ b is formed in which the scales much coarser than the current separation are switched off, at a ratio L L L ; this keeps every sum over scales of length O ( log L ) O(\log L) O ( log L ) at the price of an error O ( L − β ) O(L^{-\beta}) O ( L − β ) . The cutoff is nondecreasing in r r r , so that the sign computation survives it.
Commutator estimates (§ 4 ). The state of the system, a configuration, consists of the two particles and all wavelet centers; between refreshes it solves an autonomous ODE whose generator B = ∑ a B a \mathcal{B}=\sum_a\mathcal{B}_a B = ∑ a B a splits by scale. Here B a \mathcal{B}_a B a is first-order differential operator
B a F = u a ( x ) ⋅ ∇ x F + u a ( x ~ ) ⋅ ∇ x ~ + ∑ m > a ∑ i ∈ I m u a ( P i m ) ⋅ ∇ P i m F \mathcal{B}_aF=u_a(x)\cdot\nabla_xF+u_a(\tilde{x})\cdot\nabla_{\tilde{x}}+\sum_{m>a}\sum_{i\in I_m}u_a(P_i^m)\cdot\nabla_{P_i^m}F B a F = u a ( x ) ⋅ ∇ x F + u a ( x ~ ) ⋅ ∇ x ~ + m > a ∑ i ∈ I m ∑ u a ( P i m ) ⋅ ∇ P i m F
acting on functions F F F of the particles x , x ~ x,\tilde{x} x , x ~ and of the centers P i m P_i^m P i m , with the coefficients and shifts of the current windows frozen: B a \mathcal{B}_a B a transports the two particles and the centers of all finer generations m > a m>a m > a with the generation-a a a field u a u_a u a , and moves nothing else (Definition 4.2, where a query point is added so that B a \mathcal{B}_a B a also acts on the frozen fields u b u_b u b themselves). The key analytic fact, Lemma 4.3, is that differentiating the scale-b b b field along the scale-a a a flow costs one full gradient of the coarser of the two scales and never a gradient of the finer one. When a < b a<b a < b the coarse field transports the fine wavelet and the observation point together, and B a u b \mathcal{B}_a u_b B a u b involves only the increment of u a u_a u a across the support of the wavelet, a gain of a factor ℓ b / ℓ a \ell_b/\ell_a ℓ b / ℓ a over the convective estimate. This is the content of “controlled sweeping”, and it is what makes the sums over scales converge.
A corrected observable (Section 5). Once the initial computation are carried out, it becomes clear that that it is more transparent not to study Φ ( z t ) \Phi(z_t) Φ ( z t ) directly but rather the corrected observable
Q = Φ + ∑ b h b B b Φ b + ∑ a , b q a b B a B b Φ b , Q=\Phi+\sum_b h_b\mathcal{B}_b\Phi_b+\sum_{a,b}q_{ab}\mathcal{B}_a\mathcal{B}_b\Phi_b, Q = Φ + b ∑ h b B b Φ b + a , b ∑ q ab B a B b Φ b ,
where h b h_b h b is the time remaining in the current scale-b b b window and q a b q_{ab} q ab an explicit quadratic function of these phases. The weights are carefully chosen so that between refreshes d d t Q \frac{d}{dt}Q d t d Q cancels exactly to second order, mixed scales included, leaving a third-order remainder of size O ( κ 2 log 2 L ) O(\kappa^2\log^2 L) O ( κ 2 log 2 L ) and the cutoff error. The estimates are Littlewood–Paley sums: scales coarser than the separation r r r contribute r ℓ j − β r\ell_j^{-\beta} r ℓ j − β , finer ones contribute ℓ j α \ell_j^\alpha ℓ j α , and both series are geometric away from r r r .
The gains (Sections 6 and 7). This brings us back to the symbol of the model. At a refresh of scale j j j the observable jumps. Averaging over the fresh data annihilates the first-order term and every mixed term, which are linear, respectively bilinear, in independent centered families, and leaves the diagonal term τ j 2 Γ j \tau_j^2\Gamma_j τ j 2 Γ j , which is the sign computation above. At every time the scale matching the current separation contributes at rate κ \kappa κ , all other scales contribute nonnegatively, and integration in time gives κ t \kappa t κ t . Choosing L L L as a power of 1 / κ 1/\kappa 1/ κ makes the errors a fraction of the gain, and Jensen’s inequality converts the bound on E R t β \mathbb{E}R_t^\beta E R t β into one on E R t 2 \mathbb{E}R_t^2 E R t 2 .
Anomalous dissipation. The dissipation as a two-point quantity.
Let u ← \overleftarrow{u} u be the time reversal of the field on [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] (Subsection 6.2), and let ϱ s \varrho^s ϱ s solve the advection–diffusion equation on T d × T d \mathbb{T}^d\times\mathbb{T}^d T d × T d in which u ← \overleftarrow{u} u acts on each of the two variables, with diffusivity μ J \mu_J μ J , starting from the normalized indicator of the ζ \zeta ζ -neighbourhood of the diagonal. For bounded θ 0 \theta_0 θ 0 , duality and an approximate identity give (Lemma 10.1)
D J ( θ 0 ) = lim s ↓ 0 1 2 ∬ ( θ 0 ( y ) − θ 0 ( y ′ ) ) 2 ϱ s ( t 0 , y , y ′ ) d y d y ′ . \mathbf{D}_J(\theta_0)=\lim_{s\downarrow0}\frac{1}{2}\iint(\theta_0(y)-\theta_0(y'))^2\varrho^s(t_0,y,y')\,dy\,dy'. D J ( θ 0 ) = s ↓ 0 lim 2 1 ∬ ( θ 0 ( y ) − θ 0 ( y ′ ) ) 2 ϱ s ( t 0 , y , y ′ ) d y d y ′ .
This is an analytic counterpart of the Lagrangian fluctuation–dissipation relation in [15 ] . Indeed, for a fixed velocity realization, let X s x , X ~ s x X_s^x,\widetilde{X}_s^x X s x , X s x solve the backward tracer equation with independent Brownian motions:
d X s x = − u ( J ) ( X s x , t 0 − s ) d s + 2 μ J d W s , X 0 x = x , 0 ≤ s ≤ t 0 . dX_s^x=-u^{(J)}(X_s^x,t_0-s)\,ds+\sqrt{2\mu_J}\,dW_s,\qquad X_0^x=x,\qquad0\le s\le t_0. d X s x = − u ( J ) ( X s x , t 0 − s ) d s + 2 μ J d W s , X 0 x = x , 0 ≤ s ≤ t 0 .
Writing E W \mathbb{E}_W E W for expectation over these Brownian motions only, the scalar representation and dissipation identity are
θ ( J ) ( t 0 , x ) = E W θ 0 ( X t 0 x ) , D J ( θ 0 ) = 1 2 ∫ T d E W [ ( θ 0 ( X t 0 x ) − θ 0 ( X ~ t 0 x ) ) 2 ] d x . \theta^{(J)}(t_0,x)=\mathbb{E}_W\theta_0(X_{t_0}^x),\qquad D_J(\theta_0)=\frac{1}{2}\int_{\mathbb{T}^d}\mathbb{E}_W\left[(\theta_0(X_{t_0}^x)-\theta_0(\widetilde{X}_{t_0}^x))^2\right]\,dx. θ ( J ) ( t 0 , x ) = E W θ 0 ( X t 0 x ) , D J ( θ 0 ) = 2 1 ∫ T d E W [ ( θ 0 ( X t 0 x ) − θ 0 ( X t 0 x ) ) 2 ] d x .
where the pair density ϱ t 0 ς \varrho_{t_0}^{\varsigma} ϱ t 0 ς in the analytic identity above approximates, as ς ↓ 0 \varsigma\downarrow0 ς ↓ 0 , the joint law of their origins, averaged over the common starting point x x x . Under (S), the mean dissipation of a single Fourier mode e 2 π i k ⋅ x e^{2\pi i k\cdot x} e 2 π ik ⋅ x is, up to a factor, the mean of ∫ sin 2 ( π k ⋅ ( y − y ′ ) ) ϱ t 0 ς \int\sin^2(\pi k\cdot(y-y'))\varrho_{t_0}^{\varsigma} ∫ sin 2 ( π k ⋅ ( y − y ′ )) ϱ t 0 ς (Proposition 10.3 ). Theorem 1.4 is thereby reduced to showing that ϱ ς \varrho^{\varsigma} ϱ ς moves a definite amount of mass away from the diagonal by time t 0 t_0 t 0 , uniformly in J J J and ς \varsigma ς , mode by mode.
Richardson growth. Theorem 9.2 is the analogue of Theorem 1.1 for such solutions: if R = dist ( y , y ′ ) R=\operatorname{dist}(y,y') R = dist ( y , y ′ ) , then E ∫ R β ϱ ( t ) ≥ c κ t − o ( 1 ) \mathbb{E}\int R^\beta\varrho(t)\ge c\kappa t-o(1) E ∫ R β ϱ ( t ) ≥ c κ t − o ( 1 ) as J → ∞ J\to\infty J → ∞ . Its proof uses the same corrector and jump identities as the trajectory proof: the equation is tested against the corrected observable of § 5 , so that the evolution identity for weak solutions takes the place of the chain rule along trajectories, and the gains are produced at the renewal times by averaging over the fresh data. For the diffusing pair, three further ingredients are needed.
(i) Time reversal. The construction of the vectorfield u ⃗ \vec{u} u is such that it not time-reversible per se, but its backwards version carries essentially the same properties. Writing ζ ~ \widetilde{\zeta} ζ for a renewal time in the time variable of the chosen evolution, the averaging there requires that ϱ ( ζ ~ ) \varrho(\widetilde{\zeta}) ϱ ( ζ ) be independent of the fresh data. For the reversed field this holds when the lifetimes are nested (Lemma 6.4 (c)).
(ii) The diagonal. On the diagonal the two velocities coincide, so the drift cannot separate the pair. We have to show that mass leaves the diagonal by diffusion. Accordingly, Φ \Phi Φ is replaced by a function Φ ♯ \Phi^\sharp Φ ♯ with a quadratic core whose radius r ♯ r_\sharp r ♯ is a fixed multiple of ℓ J \ell_J ℓ J (Subsection 3.2), chosen so that 2 μ Δ Φ ♯ = d κ / 2 2\mu\Delta\Phi^\sharp=d\kappa/2 2 μ Δ Φ ♯ = d κ /2 on the core: there the diffusion produces a gain of the same size κ \kappa κ as the drift produces at every larger scale. Away from the core the main diffusion term 2 μ Δ Φ ♯ 2\mu\Delta\Phi^\sharp 2 μ Δ Φ ♯ is still nonnegative (Lemma 3.3 (e)) and is discarded in the lower bound. The diffusion acting on the correctors, and the terms created by the spatial cutoff, are controlled as errors, because the Péclet number of the finest scale is fixed.
(iii) Localization in scale. Just as in the argument for the Richardson dispersion, the gain of scale j j j is collected at its renewal times. It is always nonnegative, but it is bounded below by a multiple of κ \kappa κ only when the separation at that time lies in a range comparable to ℓ j \ell_j ℓ j . This is analogous to the aforementioned estimates for Γ j \Gamma_j Γ j . One therefore needs to know, in addition, that the separation at a renewal time of scale j j j is comparable to the separation during the following window. For trajectories in the Richardson dispersion argument, this follows from the speed bound (18) . Here diffusion precludes a deterministic speed bound; instead, the Hölder control of the relative drift allows us to construct exponential supersolutions and apply the parabolic comparison principle (Lemma 9.5 ). Consequently, over times shorter than a fixed multiple of a β a^{\beta} a β , only an exponentially small fraction of mass can move from separation at most a a a to separation at least 2 a 2a 2 a , or conversely.
Organization and conventions
Section 2 collects the regularity of the velocity, its matrix potentials, and the structure function. Section 3 constructs both the logarithmic and quadratic-core observables and proves their gain estimates. The scale operators and corrector estimates are developed in Sections 4 and 5, including the effect of diffusion on the correctors. Section 6 groups the fresh-data averages, time reversal, and jump identities. Section 7 completes the proof of Richardson dispersion for trajectories. We then turn to anomalous dissipation: Section 8 supplies the energy and duality estimates (their standard proofs are collected in Appendix A), and Section 9 proves separation growth for diffusing pairs. Section 10 deduces anomalous dissipation for the truncations and, using the comparison of dissipations in Subsection 10.5, for the single limiting field.
Constants denoted by C C C and c c c may change from line to line. They may depend on the dimension, on the model parameters α \alpha α , ψ \psi ψ , R R R , M M M , σ \sigma σ , and on the auxiliary cutoff functions and numbers fixed once and for all in the course of the proofs: ϑ \vartheta ϑ , χ \chi χ and ρ \rho ρ in Section 3, Υ \Upsilon Υ in Subsection 3.2, and χ ∘ \chi_{\circ} χ ∘ , λ ↑ \lambda_{\uparrow} λ ↑ , λ ↓ \lambda_{\downarrow} λ ↓ , θ \theta θ in Section 9. These auxiliary objects are themselves chosen as functions of d d d , α \alpha α , ψ \psi ψ , R R R only (the number ρ \rho ρ through the threshold v ∗ v_{*} v ∗ of Lemma 2.5 , the cutoff profiles through α \alpha α , and the remaining functions independently of all parameters), so that a constant depending on them depends only on d d d , α \alpha α , ψ \psi ψ , R R R , M M M , σ \sigma σ ; this is the dependence stated in the theorems. The constants do not depend on J J J , on κ \kappa κ , on the cutoff parameter L L L of Section 3, on the configuration, or on the realization of the random data. Tensor norms are Euclidean operator norms.
Acknowledgements
The model has been conceived by Raul Jimenez at the workshop in Castro Urdiales mathematics institute in discussions with C. De Lellis, D. N. Spergel, G. David, and M. Engelstein, and the second author of this paper. The authors are immensely thankful to aforementioned collaborators for all the discussions that led to the model and some understanding of its dynamics. The companion physics paper [26 ] is an outgrowth of this effort and the Hölder continuity proof for the small Kubo number in Lemma 2.2 is due to Camillo De Lellis. The authors used AI for some aspects of writing and to pick a particular wavelet with desired properties; we checked by hand that indeed that wavelet behaves as asserted. Both authors have been supported by the Simons Foundation, through the Geometry of Flows Initiative.
Velocity estimates and the structure function
This section collects the deterministic properties of the model and the properties of the structure function of the wavelet family. Lemma 2.1 records the comparability and nesting of the lifetimes. Lemma 2.2 establishes Hölder regularity of u u u , uniform in J J J . Lemma 2.5 establishes the isotropy of D D D , the Kármán–Howarth identity and the small-argument expansion of the structure function for every admissible wavelet; Proposition 2.6 exhibits an explicit wavelet satisfying (ND) in every dimension. The matrix-potential estimate of Lemma 2.4 controls the velocity scales omitted by a finite truncation. We begin with the elementary properties of the lifetimes τ j = 2 − ⌈ β j ⌉ κ \tau_j = 2^{-\lceil\beta j\rceil}\kappa τ j = 2 − ⌈ β j ⌉ κ introduced in Section 1.2.
Lemma 2.1 (Comparability and nesting). For every j ≥ j 0 j \ge j_0 j ≥ j 0 ,
1 2 κ ℓ j β < τ j ≤ κ ℓ j β = κ / λ j ; (15) \frac{1}{2}\kappa\ell_j^\beta< \tau_j \le\kappa\ell_j^\beta= \kappa/\lambda_j; \tag*{(15)} 2 1 κ ℓ j β < τ j ≤ κ ℓ j β = κ / λ j ; (15) and τ j / τ j + 1 ∈ { 1 , 2 } \tau_j/\tau_{j+1} \in\{1,2\} τ j / τ j + 1 ∈ { 1 , 2 } . Consequently, for a ≤ m a \le m a ≤ m , τ a / τ m \tau_a/\tau_m τ a / τ m is a nonnegative integer power of 2 2 2 , and every renewal time k τ a k\tau_a k τ a of scale a a a is a renewal time of every finer scale m ≥ a m \ge a m ≥ a .
Proof. Since ⌈ β j ⌉ ∈ [ β j , β j + 1 ) \lceil\beta j\rceil\in[\beta j,\beta j+1) ⌈ β j ⌉ ∈ [ β j , β j + 1 ) , we have 2 − ⌈ β j ⌉ ∈ ( 2 − β j / 2 , 2 − β j ] 2^{-\lceil\beta j\rceil} \in(2^{-\beta j}/2,2^{-\beta j}] 2 − ⌈ β j ⌉ ∈ ( 2 − β j /2 , 2 − β j ] , and 2 − β j = ℓ j β 2^{-\beta j}=\ell_j^\beta 2 − β j = ℓ j β . Since 0 < β < 1 0<\beta<1 0 < β < 1 , for every real x x x one has ⌈ x ⌉ ≤ ⌈ x + β ⌉ ≤ ⌈ x ⌉ + 1 \lceil x\rceil\le\lceil x+\beta\rceil\le\lceil x\rceil+1 ⌈ x ⌉ ≤ ⌈ x + β ⌉ ≤ ⌈ x ⌉ + 1 , so ⌈ β ( j + 1 ) ⌉ − ⌈ β j ⌉ ∈ { 0 , 1 } \lceil\beta(j+1)\rceil-\lceil\beta j\rceil\in\{0,1\} ⌈ β ( j + 1 )⌉ − ⌈ β j ⌉ ∈ { 0 , 1 } . Iterating gives τ a / τ m = 2 ⌈ β m ⌉ − ⌈ β a ⌉ \tau_a/\tau_m=2^{\lceil\beta m\rceil-\lceil\beta a\rceil} τ a / τ m = 2 ⌈ β m ⌉ − ⌈ β a ⌉ , and then k τ a = ( k τ a / τ m ) τ m k\tau_a=(k\tau_a/\tau_m)\tau_m k τ a = ( k τ a / τ m ) τ m with k τ a / τ m ∈ N k\tau_a/\tau_m\in\mathbb{N} k τ a / τ m ∈ N . □ \square □
Lemma 2.2 (Packing of wavelets and Hölder regularity of u u u ). There are κ ∗ > 0 \kappa_* > 0 κ ∗ > 0 and C ∗ > 0 C_* > 0 C ∗ > 0 , depending only on d , α , ψ , R , M d,\alpha,\psi,R,M d , α , ψ , R , M , such that, for κ ≤ κ ∗ \kappa\le\kappa_* κ ≤ κ ∗ , the centers of each active generation are ℓ j / 2 \ell_j/2 ℓ j /2 -separated for all t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . The number of its centers within distance R ℓ j R\ell_j R ℓ j of any point is bounded by a constant depending only on d , R d,R d , R . For 0 ≤ q ≤ 3 0\le q\le3 0 ≤ q ≤ 3 ,
A j ∑ i ∑ ν ∈ A ∣ η i , k j , ν D q W j ν ( x − P i , k j ( t ) ) ∣ ≤ C ∗ A j ℓ j − q . (16) A_j\sum_i\sum_{\nu\in\mathcal{A}} \left|\eta_{i,k}^{j,\nu}D^qW_j^\nu(x-P_{i,k}^j(t))\right|\le C_*A_j\ell_j^{-q}. \tag*{(16)} A j i ∑ ν ∈ A ∑ η i , k j , ν D q W j ν ( x − P i , k j ( t )) ≤ C ∗ A j ℓ j − q . (16) As a result,
∣ u ( x , t ) − u ( y , t ) ∣ ≤ C ∗ dist ( x , y ) α , Lip ( u ( ⋅ , t ) ) ≤ C ∗ λ J . (17) |u(x,t)-u(y,t)|\le C_*\operatorname{dist}(x,y)^\alpha,\qquad\operatorname{Lip}(u(\cdot,t))\le C_*\lambda_J. \tag*{(17)} ∣ u ( x , t ) − u ( y , t ) ∣ ≤ C ∗ dist ( x , y ) α , Lip ( u ( ⋅ , t )) ≤ C ∗ λ J . (17) For every finite J J J , all center and particle equations have unique Carathéodory solutions, measurable in the random data. Distinct particles remain distinct. As long as R t ≤ 1 / 4 R_t\le1/4 R t ≤ 1/4 ,
∣ d d t R t β ∣ ≤ C H for almost every t . (18) \left|\frac{\mathrm{d}}{\mathrm{d}t}R_t^\beta\right|\le C_H\qquad\text{for almost every }t. \tag*{(18)} d t d R t β ≤ C H for almost every t . (18) Condition (1.9) ensures R s ≤ 1 / 4 R_s\le1/4 R s ≤ 1/4 for all s ≤ t s\le t s ≤ t .
Proof. A set with mutual torus distance at least ℓ j / 2 \ell_j/2 ℓ j /2 has at most N d , R = ⌈ ( 8 R + 1 ) d ⌉ N_{d,R}=\lceil(8R+1)^d\rceil N d , R = ⌈( 8 R + 1 ) d ⌉ points in a ball of radius R ℓ j R\ell_j R ℓ j : disjoint balls of radius ℓ j / 8 \ell_j/8 ℓ j /8 around these points lie in a ball of radius ( R + 1 / 8 ) ℓ j ≤ 3 / 8 (R+1/8)\ell_j\le3/8 ( R + 1/8 ) ℓ j ≤ 3/8 ; all these balls have radius less than 1 / 2 1/2 1/2 , so their volumes are those of the corresponding Euclidean balls, and comparing volumes gives the bound. Set
K : = N d , R # A M max 0 ≤ q ≤ 3 max ν ∈ A ∥ D q W ν ∥ ∞ . K:=N_{d,R}\#\mathcal{A}M\max_{0\le q\le3}\max_{\nu\in\mathcal{A}}\|D^qW^\nu\|_\infty. K := N d , R # A M 0 ≤ q ≤ 3 max ν ∈ A max ∥ D q W ν ∥ ∞ .
Since R ℓ j ≤ 1 / 4 R\ell_j\le1/4 R ℓ j ≤ 1/4 , for each x x x and i i i at most one term of the periodization W j ν ( x − P i , k j ( t ) ) W_j^\nu(x-P_{i,k}^j(t)) W j ν ( x − P i , k j ( t )) is nonzero, and D q W j ν = ℓ j − q ( D q W ν ) ( ⋅ / ℓ j ) D^qW_j^\nu=\ell_j^{-q}(D^qW^\nu)(\cdot/\ell_j) D q W j ν = ℓ j − q ( D q W ν ) ( ⋅ / ℓ j ) term by term. Hence, at every time t t t at which the centers of generation j j j are ℓ j / 2 \ell_j/2 ℓ j /2 -separated, the packing bound gives (16) for generation j j j at time t t t with C ∗ C_* C ∗ replaced by $K. We prove by induction on j ≥ j 0 j \ge j_0 j ≥ j 0 that, for κ ≤ κ ∗ \kappa\le\kappa_* κ ≤ κ ∗ , the centers of generation j j j are ℓ j / 2 \ell_j/2 ℓ j /2 -separated at every t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] , so that (16) holds for generation j j j with constant K K K at every time. Generation j 0 j_0 j 0 is stationary between its renewals, since u < j 0 = 0 u_{<j_0}=0 u < j 0 = 0 , and a shifted grid is ℓ j 0 \ell_{j_0} ℓ j 0 -separated. Suppose the claim holds for all generations m < j m<j m < j . Then, for every t t t ,
Lip ( u < j ( ⋅ , t ) ) ≤ K ∑ m < j λ m ≤ K c β λ j , c β = ∑ n ≥ 1 2 − n β < ∞ . \operatorname{Lip}(u_{<j}(\cdot,t)) \le K\sum_{m<j}\lambda_m \le Kc_\beta\lambda_j,\qquad c_\beta=\sum_{n\ge1}2^{-n\beta}<\infty. Lip ( u < j ( ⋅ , t )) ≤ K m < j ∑ λ m ≤ K c β λ j , c β = n ≥ 1 ∑ 2 − n β < ∞.
Over an interval of length τ j ≤ κ / λ j \tau_j\le\kappa/\lambda_j τ j ≤ κ / λ j , the flow of u < j u_{<j} u < j and its inverse have Lipschitz constants at most exp ( K c β κ ) \exp(Kc_\beta\kappa) exp ( K c β κ ) . Choose κ ∗ \kappa_* κ ∗ so that this number is at most 2. A newly shifted grid is ℓ j \ell_j ℓ j -separated, hence remains ℓ j / 2 \ell_j/2 ℓ j /2 -separated throughout its window, which closes the induction. This proves (16) with C ∗ = K C_*=K C ∗ = K ; the constant is enlarged below as needed.
Summing the gradient bounds gives the Lipschitz estimate in (17) . Summing
∣ u j ( x ) − u j ( y ) ∣ ≤ C min ( ℓ j α , dist ( x , y ) ℓ j α − β ) |u_j(x)-u_j(y)|\le C\min(\ell_j^\alpha,\operatorname{dist}(x,y)\ell_j^{\alpha-\beta}) ∣ u j ( x ) − u j ( y ) ∣ ≤ C min ( ℓ j α , dist ( x , y ) ℓ j α − β )
over scales above and below dist ( x , y ) \operatorname{dist}(x,y) dist ( x , y ) gives its Hölder estimate. There are finitely many windows and variables for fixed J J J . Induction and the usual Picard construction on the intervals between renewal times give the asserted existence, uniqueness, and measurability. The flow and its inverse are Lipschitz with constants at most e C λ J t e^{C\lambda_Jt} e C λ J t , so R t ≥ r 0 e − C λ J t > 0 R_t\ge r_0e^{-C\lambda_Jt}>0 R t ≥ r 0 e − C λ J t > 0 for a distinct pair. The chain rule and the Hölder estimate give (18) . Integration up to a first exit from the local region proves the final assertion. □
Lemma 2.3 (The field between renewal times). Let κ ≤ κ ∗ \kappa\le\kappa_* κ ≤ κ ∗ , fix J J J and a realization of the data, and let ( a , b ) (a,b) ( a , b ) be an interval between consecutive points of ( τ J N ∩ [ 0 , 1 ) ) ∪ { 1 } (\tau_J\mathbb{N}\cap[0,1))\cup\{1\} ( τ J N ∩ [ 0 , 1 )) ∪ { 1 } . On ( a , b ) (a,b) ( a , b ) all shifts and coefficients are constant, all centers are C 1 C^1 C 1 in time, and u ( J ) u^{(J)} u ( J ) is C 3 C^3 C 3 in x x x and C 1 C^1 C 1 in t t t ; these derivatives have one-sided limits at a a a and b b b . Moreover ∇ ⋅ u ( J ) = 0 \nabla\cdot u^{(J)}=0 ∇ ⋅ u ( J ) = 0 and ∣ u ( J ) ∣ ≤ C ∗ ∑ j A j ≤ C ∗ ( 1 − 2 − α ) − 1 |u^{(J)}|\le C_*\sum_j A_j\le C_*(1-2^{-\alpha})^{-1} ∣ u ( J ) ∣ ≤ C ∗ ∑ j A j ≤ C ∗ ( 1 − 2 − α ) − 1 .
Proof. By Lemma 2.1 no window of any scale starts in ( a , b ) (a,b) ( a , b ) . By induction on scales, the centers of scale j j j solve P ˙ = u < j ( P , t ) \dot P=u_{<j}(P,t) P ˙ = u < j ( P , t ) , whose right-hand side is C 3 C^3 C 3 in P P P and continuous in t t t on [ a , b ] [a,b] [ a , b ] (with one-sided limits); hence the centers are C 1 C^1 C 1 on [ a , b ] [a,b] [ a , b ] , and u j = A j ∑ i , ν η ν i W ν i ( x − P i ( t ) ) u_j=A_j\sum_{i,\nu}\eta_\nu^iW_\nu^i(x-P_i(t)) u j = A j ∑ i , ν η ν i W ν i ( x − P i ( t )) is C 3 C^3 C 3 in x x x (as W ν i ∈ C 3 W_\nu^i\in C^3 W ν i ∈ C 3 ) and C 1 C^1 C 1 in t t t . Each W ν i W_\nu^i W ν i is divergence-free, and the bound follows from (16) with q = 0 q=0 q = 0 . □
Matrix potentials of the velocity tail
For a matrix field B B B , write ( div B ) i = ∑ m ∂ m B i m (\operatorname{div}B)_i=\sum_m\partial_mB_{im} ( div B ) i = ∑ m ∂ m B im and use the operator norm pointwise. The wavelet representation gives a potential for the tail with one additional power of length.
Lemma 2.4 (Matrix potential of the velocity tail). For the model with κ ≤ κ ∗ \kappa\le\kappa_* κ ≤ κ ∗ and every J ≥ j 0 J\ge j_0 J ≥ j 0 , there is a skew-symmetric matrix field B > J B_{>J} B > J on [ 0 , 1 ] × T d [0,1]\times\mathbb{T}^d [ 0 , 1 ] × T d , Borel in ( x , t ) (x,t) ( x , t ) and C 1 C^1 C 1 in x x x , such that
u ( ∞ ) − u ( J ) = div B > J , (19) u^{(\infty)}-u^{(J)}=\operatorname{div}B_{>J}, \tag*{(19)} u ( ∞ ) − u ( J ) = div B > J , (19) ∥ B > J ∥ L t , x ∞ ≤ C p o t ℓ J 1 + α , ∥ ∇ x B > J ∥ L t , x ∞ ≤ C p o t ℓ J α . \|B_{>J}\|_{L^\infty_{t,x}}\le C_{\mathrm{pot}}\ell_J^{1+\alpha},\qquad\|\nabla_xB_{>J}\|_{L^\infty_{t,x}}\le C_{\mathrm{pot}}\ell_J^\alpha. ∥ B > J ∥ L t , x ∞ ≤ C pot ℓ J 1 + α , ∥ ∇ x B > J ∥ L t , x ∞ ≤ C pot ℓ J α . :::`
The constant C p o t C_{\mathrm{pot}} C pot depends only on d d d , α \alpha α , ψ \psi ψ , R R R , M M M , and is independent of J J J , the realization, and κ ≤ κ ∗ \kappa\le\kappa_{*} κ ≤ κ ∗ .
Proof. The wavelet representation provides a matrix potential at each scale. We sum these potentials after using the packing estimate to bound them and their gradients. Put
ψ j ( x ) : = ∑ n ∈ Z d ψ ( ( x + n ) / ℓ j ) , B j ( x , t ) : = A j ℓ j ∑ i , ν η i , k j , ν J ν ∇ ψ j ( x − P i , k j ( t ) ) ( t ∈ I j , k ) . \psi_j(x) := \sum_{n\in\mathbb{Z}^d}\psi((x+n)/\ell_j), \qquad B_j(x,t) := A_j\ell_j\sum_{i,\nu}\eta_{i,k}^{j,\nu}J^\nu\nabla\psi_j(x-P_{i,k}^j(t)) \quad(t\in I_{j,k}). ψ j ( x ) := n ∈ Z d ∑ ψ (( x + n ) / ℓ j ) , B j ( x , t ) := A j ℓ j i , ν ∑ η i , k j , ν J ν ∇ ψ j ( x − P i , k j ( t )) ( t ∈ I j , k ) .
Since J ν J^\nu J ν is skew-symmetric, so is B j B_j B j ; and since ℓ j ∇ ψ j ( x ) = ∑ n ∇ ψ ( ( x + n ) / ℓ j ) \ell_j\nabla\psi_j(x)=\sum_n\nabla\psi((x+n)/\ell_j) ℓ j ∇ ψ j ( x ) = ∑ n ∇ ψ (( x + n ) / ℓ j ) and W ν = J ν ∇ ψ W^\nu=J^\nu\nabla\psi W ν = J ν ∇ ψ ,
div B j = A j ℓ j ∑ i , ν η i , k j , ν J ν ∇ ψ j ( x − P i , k j ( t ) ) = A j ∑ i , ν η i , k j , ν W ν ( x − P i , k j ( t ) ) = u j . \operatorname{div} B_j=A_j\ell_j\sum_{i,\nu}\eta_{i,k}^{j,\nu}J^\nu\nabla\psi_j(x-P_{i,k}^j(t))=A_j\sum_{i,\nu}\eta_{i,k}^{j,\nu}W^\nu(x-P_{i,k}^j(t))=u_j. div B j = A j ℓ j i , ν ∑ η i , k j , ν J ν ∇ ψ j ( x − P i , k j ( t )) = A j i , ν ∑ η i , k j , ν W ν ( x − P i , k j ( t )) = u j .
By Lemma 2.2 , at most N d , R N_{d,R} N d , R centers of generation j j j lie within distance R ℓ j R\ell_j R ℓ j of any point, and R ℓ j ≤ 1 4 R\ell_j\le\frac{1}{4} R ℓ j ≤ 4 1 ; hence
∥ B j ∥ ∞ ≤ N d , R # A M ∥ ψ ∥ ∞ A j ℓ j = C ℓ j 1 + α , \|B_j\|_\infty\le N_{d,R}\#\mathcal{A}M\|\psi\|_\infty A_j\ell_j=C\ell_j^{1+\alpha}, ∥ B j ∥ ∞ ≤ N d , R # A M ∥ ψ ∥ ∞ A j ℓ j = C ℓ j 1 + α ,
∥ ∇ x B j ∥ ∞ ≤ N d , R # A M ∥ ∇ ψ ∥ ∞ A j = C ℓ j α . \|\nabla_xB_j\|_\infty\le N_{d,R}\#\mathcal{A}M\|\nabla\psi\|_\infty A_j=C\ell_j^\alpha. ∥ ∇ x B j ∥ ∞ ≤ N d , R # A M ∥∇ ψ ∥ ∞ A j = C ℓ j α .
Both series ∑ j > J ℓ j 1 + α ≤ ℓ J 1 + α \sum_{j>J}\ell_j^{1+\alpha}\le\ell_J^{1+\alpha} ∑ j > J ℓ j 1 + α ≤ ℓ J 1 + α and ∑ j > J ℓ j α ≤ C ℓ J α \sum_{j>J}\ell_j^\alpha\le C\ell_J^\alpha ∑ j > J ℓ j α ≤ C ℓ J α converge, so B > J : = ∑ j > J B j B_{>J}:=\sum_{j>J}B_j B > J := ∑ j > J B j converges uniformly together with its x x x -gradient, is skew-symmetric, and has divergence ∑ j > J u j = u ( ∞ ) − u ( J ) \sum_{j>J}u_j=u^{(\infty)}-u^{(J)} ∑ j > J u j = u ( ∞ ) − u ( J ) .
The structure function and an explicit wavelet
Next, let us discuss the shape of the wavelets and the condition (ND). Fix unit vectors e , n ∈ R d e,n\in\mathbb{R}^d e , n ∈ R d with n ⊥ e n\perp e n ⊥ e . By isotropy, the quantities
D L ( r ) = e ⊤ D ( r e ) e , D N ( r ) = n ⊤ D ( r e ) n D_L(r)=e^\top D(re)e,\qquad D_N(r)=n^\top D(re)n D L ( r ) = e ⊤ D ( r e ) e , D N ( r ) = n ⊤ D ( r e ) n
are independent of the choice of these directions. Explicitly,
D L ( r ) = ∑ ν ∫ R d ∣ [ W ν ( y ) − W ν ( y − r e ) ] ⋅ e ∣ 2 d y , D_L(r)=\sum_\nu\int_{\mathbb{R}^d}|[W^\nu(y)-W^\nu(y-re)]\cdot e|^2\,dy, D L ( r ) = ν ∑ ∫ R d ∣ [ W ν ( y ) − W ν ( y − r e )] ⋅ e ∣ 2 d y ,
and
D N ( r ) = ∑ ν ∫ R d ∣ [ W ν ( y ) − W ν ( y − r e ) ] ⋅ n ∣ 2 d y . D_N(r)=\sum_\nu\int_{\mathbb{R}^d}|[W^\nu(y)-W^\nu(y-re)]\cdot n|^2\,dy. D N ( r ) = ν ∑ ∫ R d ∣ [ W ν ( y ) − W ν ( y − r e )] ⋅ n ∣ 2 d y .
Thus D L ( r ) D_L(r) D L ( r ) measures the integrated squared velocity increment parallel to the separation vector r e re r e , whereas D N ( r ) D_N(r) D N ( r ) measures the corresponding increment in one perpendicular direction. There are d − 1 d-1 d − 1 transverse directions.
For a fresh scale-j j j random field, multiplication by σ 2 A j 2 \sigma^2A_j^2 σ 2 A j 2 and evaluation at r / ℓ j r/\ell_j r / ℓ j give the corresponding velocity-increment variances at separation r r r .
The condition D L ′ ( r ) ≥ 0 D_L'(r)\ge0 D L ′ ( r ) ≥ 0 , r > 0 r>0 r > 0 , says that the longitudinal increment variance does not decrease as separation increases. Incompressibility and isotropy imply
D N ( r ) = D L ( r ) + r d − 1 D L ′ ( r ) . D_N(r)=D_L(r)+\frac{r}{d-1}D_L'(r). D N ( r ) = D L ( r ) + d − 1 r D L ′ ( r ) .
Consequently, (ND) is equivalent to
D N ( r ) ≥ D L ( r ) , r > 0. D_N(r)\ge D_L(r),\qquad r>0. D N ( r ) ≥ D L ( r ) , r > 0.
The proof uses (ND) to make the averaged second-order renewal gain nonnegative. Let F F F be a radial test function, writing F ( z ) = F ( r ) F(z)=F(r) F ( z ) = F ( r ) with r = ∣ z ∣ > 0 r=|z|>0 r = ∣ z ∣ > 0 . Then
D ( z ) : D 2 F ( z ) = D L ( r ) F ′ ′ ( r ) + ( d − 1 ) D N ( r ) F ′ ( r ) r , D(z):D^2F(z)=D_L(r)F^{\prime\prime}(r)+(d-1)D_N(r)\frac{F^\prime(r)}{r}, D ( z ) : D 2 F ( z ) = D L ( r ) F ′′ ( r ) + ( d − 1 ) D N ( r ) r F ′ ( r ) ,
with the colon denoting matrix inner product. Taking F ( r ) = r β F(r)=r^\beta F ( r ) = r β , where β = 1 − α ∈ ( 0 , 1 ) \beta=1-\alpha\in(0,1) β = 1 − α ∈ ( 0 , 1 ) , gives
D ( z ) : D 2 ∣ z ∣ β = β r β − 2 [ ( d − 1 ) D N ( r ) − α D L ( r ) ] . D(z):D^2|z|^\beta=\beta r^{\beta-2}\left[(d-1)D_N(r)-\alpha D_L(r)\right]. D ( z ) : D 2 ∣ z ∣ β = β r β − 2 [ ( d − 1 ) D N ( r ) − α D L ( r ) ] .
The longitudinal contribution is negative because r β r^\beta r β is concave, while the transverse contributions are positive. Under (ND), and using D L ( r ) > 0 D_L(r)>0 D L ( r ) > 0 ,
( d − 1 ) D N ( r ) − α D L ( r ) ≥ ( d − 1 − α ) D L ( r ) > 0. (d-1)D_N(r)-\alpha D_L(r)\ge(d-1-\alpha)D_L(r)>0. ( d − 1 ) D N ( r ) − α D L ( r ) ≥ ( d − 1 − α ) D L ( r ) > 0.
This is the key sign estimate: the transverse contributions more than compensate for the longitudinal concavity.
In addition, condition (ND) is used for the logarithmically continued and cutoff test functions. Writing
H F ( r ) = r F ′ ( r ) , H_F(r)=rF^\prime(r), H F ( r ) = r F ′ ( r ) ,
one obtains
D ( z ) : D 2 F ( z ) = D L ( r ) H F ′ ( r ) r + [ ( d − 1 ) D N ( r ) − D L ( r ) ] H F ( r ) r 2 . D(z):D^2F(z)=D_L(r)\frac{H_F^\prime(r)}{r}+\left[(d-1)D_N(r)-D_L(r)\right]\frac{H_F(r)}{r^2}. D ( z ) : D 2 F ( z ) = D L ( r ) r H F ′ ( r ) + [ ( d − 1 ) D N ( r ) − D L ( r ) ] r 2 H F ( r ) .
The construction ensures H F ( r ) ≥ 0 H_F(r)\ge0 H F ( r ) ≥ 0 and H F ′ ( r ) ≥ 0 H_F^\prime(r)\ge0 H F ′ ( r ) ≥ 0 . Moreover, for d ≥ 2 d\ge2 d ≥ 2 ,
( d − 1 ) D N ( r ) − D L ( r ) = ( d − 2 ) D L ( r ) + ( d − 1 ) [ D N ( r ) − D L ( r ) ] ≥ 0. (d-1)D_N(r)-D_L(r)=(d-2)D_L(r)+(d-1)\left[D_N(r)-D_L(r)\right]\ge0. ( d − 1 ) D N ( r ) − D L ( r ) = ( d − 2 ) D L ( r ) + ( d − 1 ) [ D N ( r ) − D L ( r ) ] ≥ 0.
Thus both terms in the contraction are nonnegative. Quantitative nondegeneracy at a comparable scale supplies the strictly positive gain.
All in all, (ND) ensures the correct sign of the renewal gains.
Lemma 2.5. (Structure-function properties. ) The tensor (1.1) is isotropic, and for r > 0 r>0 r > 0 it satisfies the identity
D N ( r ) = D L ( r ) + r d − 1 D L ′ ( r ) , D L ( r ) > 0. (20) D_N(r)=D_L(r)+\frac{r}{d-1}D_L^\prime(r),\qquad D_L(r)>0. \tag*{(20)} D N ( r ) = D L ( r ) + d − 1 r D L ′ ( r ) , D L ( r ) > 0. (20) In particular, (ND) is equivalent to D N ≥ D L D_N\ge D_L D N ≥ D L .
There is a constant d 0 > 0 d_0>0 d 0 > 0 such that
D L ( r ) = d 0 r 2 + o ( r 2 ) , D N ( r ) = d + 1 d − 1 d 0 r 2 + o ( r 2 ) as r ↓ 0. (21) D_L(r)=d_0r^2+o(r^2),\qquad D_N(r)=\frac{d+1}{d-1}d_0r^2+o(r^2)\quad\text{as }r\downarrow0. \tag*{(21)} D L ( r ) = d 0 r 2 + o ( r 2 ) , D N ( r ) = d − 1 d + 1 d 0 r 2 + o ( r 2 ) as r ↓ 0. (21) Consequently there are v ∗ > 0 v_\ast>0 v ∗ > 0 and c ∗ > 0 c_\ast>0 c ∗ > 0 such that
D N ( r ) − D L ( r ) ≥ c ∗ r 2 , 0 < r ≤ v ∗ . (22) D_N(r)-D_L(r)\ge c_\ast r^2,\qquad0<r\le v_\ast. \tag*{(22)} D N ( r ) − D L ( r ) ≥ c ∗ r 2 , 0 < r ≤ v ∗ . (22)
Proof. The lemma is standard isotropic incompressible covariance theory (see e.g. [19 ] which resembles some parts of its argument), but since our model is not close to what has appeared in the literature, we provide a self-contained proof. Throughout the proof, derivatives of C C C , D D D , and H H H are all with respect to the separation variable z z z .
First, by definition,
∑ p < q ( J p q ) a k ( J p q ) b ℓ = 1 2 ∑ p , q ( δ a q δ k p − δ a p δ k q ) ( δ b q δ ℓ p − δ b p δ ℓ q ) = 1 2 ( δ a b δ k ℓ − δ a ℓ δ b k − δ a ℓ δ b k + δ a b δ k ℓ ) = δ a b δ k ℓ − δ a ℓ δ b k . \begin{aligned}
\sum_{p<q}(J^{pq})_{ak}(J^{pq})_{b\ell}=\frac{1}{2}\sum_{p,q}(\delta_{aq}\delta_{kp}-\delta_{ap}\delta_{kq})(\delta_{bq}\delta_{\ell p}-\delta_{bp}\delta_{\ell q}) \\
&=\frac{1}{2}(\delta_{ab}\delta_{k\ell}-\delta_{a\ell}\delta_{bk}-\delta_{a\ell}\delta_{bk}+\delta_{ab}\delta_{k\ell}) \\
&=\delta_{ab}\delta_{k\ell}-\delta_{a\ell}\delta_{bk}.
\end{aligned} p < q ∑ ( J pq ) ak ( J pq ) b ℓ = 2 1 p , q ∑ ( δ a q δ k p − δ a p δ k q ) ( δ b q δ ℓ p − δ b p δ ℓ q ) = 2 1 ( δ ab δ k ℓ − δ a ℓ δ bk − δ a ℓ δ bk + δ ab δ k ℓ ) = δ ab δ k ℓ − δ a ℓ δ bk .
Multiplying by x k y ℓ x_k y_\ell x k y ℓ and summing over k , ℓ k,\ell k , ℓ gives, for x , y ∈ R d x,y\in\mathbb{R}^d x , y ∈ R d ,
∑ ν ∈ A ( J ν x ) a ( J ν y ) b = δ a b ( x ⋅ y ) − x b y a . \sum_{\nu\in\mathcal{A}}(J^\nu x)_a(J^\nu y)_b=\delta_{ab}(x\cdot y)-x_b y_a. ν ∈ A ∑ ( J ν x ) a ( J ν y ) b = δ ab ( x ⋅ y ) − x b y a .
In particular,
∑ ν W a ν ( y ) W b ν ( y − z ) = δ a b ∑ k ∂ k ψ ( y ) ∂ k ψ ( y − z ) − ∂ b ψ ( y ) ∂ a ψ ( y − z ) . \sum_\nu W^\nu_a(y)W^\nu_b(y-z)=\delta_{ab}\sum_k\partial_k\psi(y)\partial_k\psi(y-z)-\partial_b\psi(y)\partial_a\psi(y-z). ν ∑ W a ν ( y ) W b ν ( y − z ) = δ ab k ∑ ∂ k ψ ( y ) ∂ k ψ ( y − z ) − ∂ b ψ ( y ) ∂ a ψ ( y − z ) .
Define
H ( z ) = ∫ R d ψ ( y ) ψ ( y − z ) d y , C a b ( z ) = ∑ ν ∈ A ∫ R d W a ν ( y ) W b ν ( y − z ) d y . H(z)=\int_{\mathbb{R}^d}\psi(y)\psi(y-z)\,dy,\qquad C_{ab}(z)=\sum_{\nu\in\mathcal{A}}\int_{\mathbb{R}^d}W^\nu_a(y)W^\nu_b(y-z)\,dy. H ( z ) = ∫ R d ψ ( y ) ψ ( y − z ) d y , C ab ( z ) = ν ∈ A ∑ ∫ R d W a ν ( y ) W b ν ( y − z ) d y .
Since ψ \psi ψ is even, ψ ( y − z ) = ψ ( z − y ) \psi(y-z)=\psi(z-y) ψ ( y − z ) = ψ ( z − y ) , so H = ψ ∗ ψ H=\psi\ast\psi H = ψ ∗ ψ . Differentiating twice in z z z and integrating by parts then gives
∂ a ∂ b H ( z ) = − ∫ ∂ b ψ ( y ) ∂ a ψ ( y − z ) d y . \partial_a\partial_bH(z)=-\int\partial_b\psi(y)\partial_a\psi(y-z)\,dy. ∂ a ∂ b H ( z ) = − ∫ ∂ b ψ ( y ) ∂ a ψ ( y − z ) d y .
yielding the first identity in
C a b ( z ) = − δ a b Δ H ( z ) + ∂ a ∂ b H ( z ) , D ( z ) = 2 ( C ( 0 ) − C ( z ) ) . (23) C_{ab}(z)=-\delta_{ab}\Delta H(z)+\partial_a\partial_bH(z),\qquad D(z)=2(C(0)-C(z)). \tag*{(23)} C ab ( z ) = − δ ab Δ H ( z ) + ∂ a ∂ b H ( z ) , D ( z ) = 2 ( C ( 0 ) − C ( z )) . (23)
In particular, C a b = C b a C_{ab}=C_{ba} C ab = C ba . Also H H H is even, so its second derivatives, and hence C C C , are even. To verify the second identity, expand the increments:
D a b ( z ) = ∑ ν ∫ W a ν ( y ) W b ν ( y ) d y + ∑ ν ∫ W a ν ( y − z ) W b ν ( y − z ) d y − ∑ ν ∫ W a ν ( y ) W b ν ( y − z ) d y − ∑ ν ∫ W a ν ( y − z ) W b ν ( y ) d y = 2 ( C a b ( 0 ) − C a b ( z ) ) . \begin{aligned}
D_{ab}(z)={}&\sum_\nu\int W^\nu_a(y)W^\nu_b(y)\,dy+\sum_\nu\int W^\nu_a(y-z)W^\nu_b(y-z)\,dy\\
&-\sum_\nu\int W^\nu_a(y)W^\nu_b(y-z)\,dy-\sum_\nu\int W^\nu_a(y-z)W^\nu_b(y)\,dy\\
={}&2(C_{ab}(0)-C_{ab}(z)).
\end{aligned} D ab ( z ) = = ν ∑ ∫ W a ν ( y ) W b ν ( y ) d y + ν ∑ ∫ W a ν ( y − z ) W b ν ( y − z ) d y − ν ∑ ∫ W a ν ( y ) W b ν ( y − z ) d y − ν ∑ ∫ W a ν ( y − z ) W b ν ( y ) d y 2 ( C ab ( 0 ) − C ab ( z )) .
For any orthogonal matrix U U U , the change of variables y = U x y=Ux y = U x and radiality of ψ \psi ψ give
H ( U z ) = ∫ ψ ( U x ) ψ ( U ( x − z ) ) d x = ∫ ψ ( x ) ψ ( x − z ) d x = H ( z ) . H(Uz)=\int\psi(Ux)\psi(U(x-z))\,dx=\int\psi(x)\psi(x-z)\,dx=H(z). H ( U z ) = ∫ ψ ( U x ) ψ ( U ( x − z )) d x = ∫ ψ ( x ) ψ ( x − z ) d x = H ( z ) .
Thus H H H is radial: write H ( z ) = h ( r ) H(z)=h(r) H ( z ) = h ( r ) , where r = ∣ z ∣ r=|z| r = ∣ z ∣ . For z ≠ 0 z\ne0 z = 0 , set z ^ = z / r \widehat{z}=z/r z = z / r . Direct differentiation gives
∂ a r = z ^ a , ∂ a z ^ b = δ a b − z ^ a z ^ b r . \partial_a r=\widehat{z}_a,\qquad\partial_a\widehat{z}_b=\frac{\delta_{ab}-\widehat{z}_a\widehat{z}_b}{r}. ∂ a r = z a , ∂ a z b = r δ ab − z a z b .
Consequently,
∂ a ∂ b H = h ′ ′ ( r ) z ^ a z ^ b + h ′ ( r ) r ( δ a b − z ^ a z ^ b ) , Δ H = h ′ ′ ( r ) + ( d − 1 ) h ′ ( r ) r . \partial_a\partial_b H = h^{\prime\prime}(r)\widehat{z}_a\widehat{z}_b + \frac{h^\prime(r)}{r}(\delta_{ab}-\widehat{z}_a\widehat{z}_b), \qquad\Delta H = h^{\prime\prime}(r)+(d-1)\frac{h^\prime(r)}{r}. ∂ a ∂ b H = h ′′ ( r ) z a z b + r h ′ ( r ) ( δ ab − z a z b ) , Δ H = h ′′ ( r ) + ( d − 1 ) r h ′ ( r ) .
The covariance formula therefore reads
C ( z ) = C L ( r ) z ^ ⊗ z ^ + C N ( r ) ( I − z ^ ⊗ z ^ ) , C(z)=C_L(r)\widehat{z}\otimes\widehat{z}+C_N(r)(I-\widehat{z}\otimes\widehat{z}), C ( z ) = C L ( r ) z ⊗ z + C N ( r ) ( I − z ⊗ z ) ,
where
C L ( r ) = − ( d − 1 ) h ′ ( r ) r , C N ( r ) = − h ′ ′ ( r ) − ( d − 2 ) h ′ ( r ) r . C_L(r)=-(d-1)\frac{h^\prime(r)}{r}, \qquad C_N(r)=-h^{\prime\prime}(r)-(d-2)\frac{h^\prime(r)}{r}. C L ( r ) = − ( d − 1 ) r h ′ ( r ) , C N ( r ) = − h ′′ ( r ) − ( d − 2 ) r h ′ ( r ) .
At the origin, rotational invariance gives C ( 0 ) = c I C(0)=cI C ( 0 ) = c I for a scalar c c c : indeed the Hessian of H H H at zero is invariant under conjugation by every orthogonal matrix, so it, and hence C ( 0 ) C(0) C ( 0 ) , is a scalar matrix. It follows from (2.9) that
D ( z ) = D L ( r ) z ^ ⊗ z ^ + D N ( r ) ( I − z ^ ⊗ z ^ ) , D(z)=D_L(r)\widehat{z}\otimes\widehat{z}+D_N(r)(I-\widehat{z}\otimes\widehat{z}), D ( z ) = D L ( r ) z ⊗ z + D N ( r ) ( I − z ⊗ z ) ,
D L ( r ) = 2 ( c − C L ( r ) ) , D N ( r ) = 2 ( c − C N ( r ) ) . D_L(r)=2(c-C_L(r)), \qquad D_N(r)=2(c-C_N(r)). D L ( r ) = 2 ( c − C L ( r )) , D N ( r ) = 2 ( c − C N ( r )) .
This proves (1.2), with D N D_N D N the eigenvalue in each of the d − 1 d-1 d − 1 transverse directions.
Now, each wavelet is divergence-free, since the matrix J ν J^\nu J ν is skew-symmetric and the Hessian of ψ \psi ψ is symmetric. Therefore,
∑ a ∂ z a C a b ( z ) = − ∑ ν , a ∫ R d W a ν ( y ) ∂ y a W b ν ( y − z ) d y = 0. \sum_a\partial_{z_a}C_{ab}(z)=-\sum_{\nu,a}\int_{\mathbb{R}^d}W_a^\nu(y)\partial_{y_a}W_b^\nu(y-z)\,dy=0. a ∑ ∂ z a C ab ( z ) = − ν , a ∑ ∫ R d W a ν ( y ) ∂ y a W b ν ( y − z ) d y = 0.
There is no boundary term because the wavelets are compactly supported. Hence ∑ a ∂ a D a b = 0 \sum_a\partial_aD_{ab}=0 ∑ a ∂ a D ab = 0 .
Write A ( r ) = D L ( r ) − D N ( r ) A(r)=D_L(r)-D_N(r) A ( r ) = D L ( r ) − D N ( r ) , so that D a b = D N ( r ) δ a b + A ( r ) z ^ a z ^ b D_{ab}=D_N(r)\delta_{ab}+A(r)\widehat{z}_a\widehat{z}_b D ab = D N ( r ) δ ab + A ( r ) z a z b . The first term has divergence D N ′ ( r ) z ^ b D_N^\prime(r)\widehat{z}_b D N ′ ( r ) z b . For the second, the product rule gives
∑ a ∂ a ( A z ^ a z ^ b ) = A ′ ∑ a ( z ^ a ) 2 z ^ b + A ( ∑ a ∂ a z ^ a ) z ^ b + A ∑ a z ^ a ∂ a z ^ b = A ′ z ^ b + A ( d − 1 r ) z ^ b + 0. \begin{aligned}
\sum_a\partial_a(A\widehat{z}_a\widehat{z}_b)=A^\prime\sum_a(\widehat{z}_a)^2\widehat{z}_b+A\left(\sum_a\partial_a\widehat{z}_a\right)\widehat{z}_b+A\sum_a\widehat{z}_a\partial_a\widehat{z}_b \\
&=A^\prime\widehat{z}_b+A\left(\frac{d-1}{r}\right)\widehat{z}_b+0.
\end{aligned} a ∑ ∂ a ( A z a z b ) = A ′ a ∑ ( z a ) 2 z b + A ( a ∑ ∂ a z a ) z b + A a ∑ z a ∂ a z b = A ′ z b + A ( r d − 1 ) z b + 0.
Hence,
0 = ∑ a ∂ a D a b = [ D L ′ ( r ) + d − 1 r ( D L ( r ) − D N ( r ) ) ] z ^ b . 0=\sum_a\partial_aD_{ab}=\left[D_L^\prime(r)+\frac{d-1}{r}(D_L(r)-D_N(r))\right]\widehat{z}_b. 0 = a ∑ ∂ a D ab = [ D L ′ ( r ) + r d − 1 ( D L ( r ) − D N ( r )) ] z b .
Since z ^ \widehat{z} z is a unit vector, the scalar coefficient is zero. This proves the identity in (2.6). Moreover, r / ( d − 1 ) > 0 r/(d-1)>0 r / ( d − 1 ) > 0 , so D N ≥ D L D_N\ge D_L D N ≥ D L is equivalent to D L ′ ≥ 0 D_L^\prime\ge0 D L ′ ≥ 0 . Since D L D_L D L is continuously differentiable on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , the latter condition is equivalent to (ND).
To prove strict positivity, use isotropy to write
D L ( r ) = ∑ ν ∫ R d [ W 1 ν ( y ) − W 1 ν ( y − r e 1 ) ] 2 d y . D_L(r)=\sum_\nu\int_{\mathbb{R}^d}[W_1^\nu(y)-W_1^\nu(y-re_1)]^2\,dy. D L ( r ) = ν ∑ ∫ R d [ W 1 ν ( y ) − W 1 ν ( y − r e 1 ) ] 2 d y .
If this vanished for some r > 0 r>0 r > 0 , nonnegativity and continuity would give W 1 ν ( y ) = W 1 ν ( y − r e 1 ) W_1^\nu(y)=W_1^\nu(y-re_1) W 1 ν ( y ) = W 1 ν ( y − r e 1 ) for every y , ν y,\nu y , ν . Iteration gives W 1 ν ( y ) = W 1 ν ( y − n r e 1 ) W_1^\nu(y)=W_1^\nu(y-nre_1) W 1 ν ( y ) = W 1 ν ( y − n r e 1 ) for all positive integers n n n , and compact support then forces W 1 ν ≡ 0 W_1^\nu\equiv0 W 1 ν ≡ 0 . In particular, W 1 12 = − ∂ 2 ψ = W_1^{12}=-\partial_2\psi= W 1 12 = − ∂ 2 ψ =
Thus ψ \psi ψ is constant along lines parallel to e 2 e_2 e 2 , and compact support forces ψ ≡ 0 \psi\equiv0 ψ ≡ 0 , a contradiction. Therefore D L ( r ) > 0 D_L(r) > 0 D L ( r ) > 0 for all r > 0 r > 0 r > 0 .
Finally, put f ν = W 1 ν f_\nu= W_1^\nu f ν = W 1 ν , g ν = ∂ 1 f ν g_\nu= \partial_1 f_\nu g ν = ∂ 1 f ν , and Δ r f ν = f ν − f ν ( ⋅ − r e 1 ) \Delta_r f_\nu= f_\nu- f_\nu(\cdot- r e_1) Δ r f ν = f ν − f ν ( ⋅ − r e 1 ) . The fundamental theorem of calculus and translation invariance of the L 2 L^2 L 2 norm imply
Δ r f ν = ∫ 0 r g ν ( ⋅ − ϱ e 1 ) d ϱ , ∥ g ν ( ⋅ − ϱ e 1 ) − g ν ∥ 2 ≤ ϱ ∥ ∂ 1 g ν ∥ 2 . \Delta_r f_\nu= \int_0^r g_\nu(\cdot- \varrho e_1)\,d\varrho,\qquad\|g_\nu(\cdot-\varrho e_1)-g_\nu\|_2 \leq\varrho\|\partial_1 g_\nu\|_2. Δ r f ν = ∫ 0 r g ν ( ⋅ − ϱ e 1 ) d ϱ , ∥ g ν ( ⋅ − ϱ e 1 ) − g ν ∥ 2 ≤ ϱ ∥ ∂ 1 g ν ∥ 2 .
Consequently,
Δ r f ν = r g ν + E ν , r , ∥ E ν , r ∥ 2 ≤ 1 2 r 2 ∥ ∂ 1 g ν ∥ 2 . \Delta_r f_\nu= r g_\nu+ E_{\nu,r},\qquad\|E_{\nu,r}\|_2 \leq\frac{1}{2}r^2\|\partial_1g_\nu\|_2. Δ r f ν = r g ν + E ν , r , ∥ E ν , r ∥ 2 ≤ 2 1 r 2 ∥ ∂ 1 g ν ∥ 2 .
Expanding the squared norm and using Cauchy–Schwarz gives
D L ( r ) = ∑ ν ∥ Δ r f ν ∥ 2 2 = d 0 r 2 + O ( r 3 ) , d 0 = ∑ ν ∥ g ν ∥ 2 2 = ∑ ν ∫ R d ∣ ∂ 1 W 1 ν ( y ) ∣ 2 d y . D_L(r)=\sum_\nu\|\Delta_r f_\nu\|_2^2=d_0r^2+O(r^3),\qquad d_0=\sum_\nu\|g_\nu\|_2^2=\sum_\nu\int_{\mathbb{R}^d}|\partial_1W_1^\nu(y)|^2\,dy. D L ( r ) = ν ∑ ∥ Δ r f ν ∥ 2 2 = d 0 r 2 + O ( r 3 ) , d 0 = ν ∑ ∥ g ν ∥ 2 2 = ν ∑ ∫ R d ∣ ∂ 1 W 1 ν ( y ) ∣ 2 d y .
For the derivative, differentiate the integral itself:
D L ′ ( r ) = 2 ∑ ν ⟨ Δ r f ν , g ν ( ⋅ − r e 1 ) ⟩ = 2 ∑ ν ⟨ r g ν + O L 2 ( r 2 ) , g ν + O L 2 ( r ) ⟩ = 2 d 0 r + O ( r 2 ) . \begin{aligned}
D_L'(r)=2\sum_\nu\langle\Delta_r f_\nu,g_\nu(\cdot-re_1)\rangle\\
&=2\sum_\nu\langle rg_\nu+O_{L^2}(r^2),g_\nu+O_{L^2}(r)\rangle=2d_0r+O(r^2).
\end{aligned} D L ′ ( r ) = 2 ν ∑ ⟨ Δ r f ν , g ν ( ⋅ − r e 1 )⟩ = 2 ν ∑ ⟨ r g ν + O L 2 ( r 2 ) , g ν + O L 2 ( r )⟩ = 2 d 0 r + O ( r 2 ) .
If d 0 = 0 d_0=0 d 0 = 0 , continuity gives ∂ 1 W 1 ν ≡ 0 \partial_1W_1^\nu\equiv0 ∂ 1 W 1 ν ≡ 0 for every ν \nu ν . Each W 1 ν W_1^\nu W 1 ν is then constant along lines parallel to e 1 e_1 e 1 and hence zero by compact support. As above, W 1 12 = 0 W_1^{12}=0 W 1 12 = 0 forces ψ ≡ 0 \psi\equiv0 ψ ≡ 0 , a contradiction. Thus d 0 > 0 d_0>0 d 0 > 0 .
Substitution in (20) now gives
D N ( r ) = d 0 r 2 + 2 d 0 d − 1 r 2 + O ( r 3 ) = d + 1 d − 1 d 0 r 2 + o ( r 2 ) . D_N(r)=d_0r^2+\frac{2d_0}{d-1}r^2+O(r^3)=\frac{d+1}{d-1}d_0r^2+o(r^2). D N ( r ) = d 0 r 2 + d − 1 2 d 0 r 2 + O ( r 3 ) = d − 1 d + 1 d 0 r 2 + o ( r 2 ) .
which proves (21) . In particular,
D N ( r ) − D L ( r ) = 2 d 0 d − 1 r 2 + o ( r 2 ) . D_N(r)-D_L(r)=\frac{2d_0}{d-1}r^2+o(r^2). D N ( r ) − D L ( r ) = d − 1 2 d 0 r 2 + o ( r 2 ) .
Taking c ∗ = d 0 / ( d − 1 ) c_* = d_0/(d-1) c ∗ = d 0 / ( d − 1 ) and choosing v ∗ > 0 v_* > 0 v ∗ > 0 so that the remainder has absolute value at most c ∗ r 2 c_*r^2 c ∗ r 2 for 0 < r ≤ v ∗ 0<r\leq v_* 0 < r ≤ v ∗ proves (22) . □
The next proposition shows that (ND) is satisfied by an explicit wavelet in every dimension. The proof expresses ψ ∗ ψ \psi* \psi ψ ∗ ψ through the convolution of two nonnegative radial functions in d + 4 d+4 d + 4 variables; integrating out the four auxiliary variables produces the two radial derivatives that enter D L ′ D_L' D L ′ .
Proposition 2.6 (Explicit admissible wavelets in every dimension). For each integer d ≥ 2 d\geq2 d ≥ 2 , set
ψ ( x ) = ( 1 − ∣ x ∣ 2 ) + 5 , x ∈ R d , W ν = J ν ∇ ψ , ν ∈ A , \psi(x)=(1-|x|^2)_+^5,\qquad x\in\mathbb{R}^d,\qquad W^\nu=J^\nu\nabla\psi,\quad\nu\in\mathcal{A}, ψ ( x ) = ( 1 − ∣ x ∣ 2 ) + 5 , x ∈ R d , W ν = J ν ∇ ψ , ν ∈ A , where a + = max ( a , 0 ) a_+=\max(a,0) a + = max ( a , 0 ) . This family has the required regularity, support, and incompressibility (ψ ∈ C c 4 \psi\in C_c^4 ψ ∈ C c 4 , W ν ∈ C c 3 W^\nu\in C_c^3 W ν ∈ C c 3 , R = 1 R=1 R = 1 , hence j 0 = 2 j_0=2 j 0 = 2 ), and its combined structure function satisfies
D L ′ ( r ) > 0 ( 0 < r < 2 ) , D L ′ ( r ) = 0 ( r ≥ 2 ) . D_L'(r)>0\quad(0<r<2),\qquad D_L'(r)=0\quad(r\geq2). D L ′ ( r ) > 0 ( 0 < r < 2 ) , D L ′ ( r ) = 0 ( r ≥ 2 ) .
More precisely, let
G ( X ) = 20 π 2 ( 1 − ∣ X ∣ 2 ) + 3 , X ∈ R d + 4 , K = G ∗ R d + 4 G , G(X)=\frac{20}{\pi^2}(1-|X|^2)_+^3,\qquad X\in\mathbb{R}^{d+4},\qquad K=G*_{\mathbb{R}^{d+4}}G, G ( X ) = π 2 20 ( 1 − ∣ X ∣ 2 ) + 3 , X ∈ R d + 4 , K = G ∗ R d + 4 G ,
and write K ( X ) = k ( ∣ X ∣ ) K(X)=k(|X|) K ( X ) = k ( ∣ X ∣ ) . Then
D L ′ ( r ) = 8 ( d − 1 ) π 2 r k ( r ) ( r > 0 ) . (24) D'_L(r)=8(d-1)\pi^2 r k(r)\qquad(r>0). \tag*{(24)} D L ′ ( r ) = 8 ( d − 1 ) π 2 r k ( r ) ( r > 0 ) . (24)
The dimension d + 4 d+4 d + 4 is auxiliary and is not the dimension of the flow.
Proof. The function q ↦ q + 5 q\mapsto q_+^5 q ↦ q + 5 is C 4 C^4 C 4 . Thus ψ ∈ C c 4 ( R d ) \psi\in C_c^4(\mathbb{R}^d) ψ ∈ C c 4 ( R d ) , and
W p q ( x ) = − 10 ( 1 − ∣ x ∣ 2 ) + 4 ( x p e q − x q e p ) W^{pq}(x)=-10(1-|x|^2)_+^4(x_pe_q-x_qe_p) W pq ( x ) = − 10 ( 1 − ∣ x ∣ 2 ) + 4 ( x p e q − x q e p )
is a nonzero C c 3 C_c^3 C c 3 field supported in the closed unit ball. Skew-symmetry of J p q J^{pq} J pq gives zero divergence.
For an integrable function F F F on R d + 4 = R d × R 4 \mathbb{R}^{d+4}=\mathbb{R}^d\times\mathbb{R}^4 R d + 4 = R d × R 4 , let ( P F ) ( x ) = ∫ R 4 F ( x , y ) d y (\mathcal{P}F)(x)=\int_{\mathbb{R}^4}F(x,y)\,dy ( P F ) ( x ) = ∫ R 4 F ( x , y ) d y . Since the area of the unit sphere in R 4 \mathbb{R}^4 R 4 is 2 π 2 2\pi^2 2 π 2 , polar integration gives, for a = 1 − ∣ x ∣ 2 > 0 a=1-|x|^2>0 a = 1 − ∣ x ∣ 2 > 0 ,
( P G ) ( x ) = 20 π 2 ∫ R 4 ( a − ∣ y ∣ 2 ) + 3 d y = 20 π 2 ∣ S 3 ∣ ∫ 0 a ( a − r 2 ) 3 r 3 d r = 40 ∫ 0 a r 3 ( a − r 2 ) 3 d r = 20 ∫ 0 a q ( a − q ) 3 d q = a 5 . \begin{aligned}
(\mathcal{P}G)(x)=\frac{20}{\pi^2}\int_{\mathbb{R}^4}(a-|y|^2)_+^3\,dy
=\frac{20}{\pi^2}|\mathbb{S}^3|\int_0^{\sqrt a}(a-r^2)^3r^3\,dr \\
&=40\int_0^{\sqrt a}r^3(a-r^2)^3\,dr
=20\int_0^a q(a-q)^3\,dq=a^5.
\end{aligned} ( P G ) ( x ) = π 2 20 ∫ R 4 ( a − ∣ y ∣ 2 ) + 3 d y = π 2 20 ∣ S 3 ∣ ∫ 0 a ( a − r 2 ) 3 r 3 d r = 40 ∫ 0 a r 3 ( a − r 2 ) 3 d r = 20 ∫ 0 a q ( a − q ) 3 d q = a 5 .
Both sides vanish if a ≤ 0 a\leq0 a ≤ 0 (∣ x ∣ ≥ 1 |x|\geq1 ∣ x ∣ ≥ 1 ). Hence P G = ψ \mathcal{P}G=\psi P G = ψ and Fubini’s theorem implies
P ( G ∗ R d + 4 G ) = ( P G ) ∗ R d ( P G ) = ψ ∗ R d ψ . \mathcal{P}(G*_{\mathbb{R}^{d+4}}G)=(\mathcal{P}G)*_{\mathbb{R}^d}(\mathcal{P}G)=\psi*_{\mathbb{R}^d}\psi. P ( G ∗ R d + 4 G ) = ( P G ) ∗ R d ( P G ) = ψ ∗ R d ψ .
Write ( ψ ∗ ψ ) ( z ) = h ~ ( ∣ z ∣ 2 ) (\psi*\psi)(z)=\widetilde h(|z|^2) ( ψ ∗ ψ ) ( z ) = h ( ∣ z ∣ 2 ) ; thus H ( z ) = h ~ ( ∣ z ∣ 2 ) H(z)=\widetilde h(|z|^2) H ( z ) = h ( ∣ z ∣ 2 ) , where H = ψ ∗ ψ H=\psi*\psi H = ψ ∗ ψ is the function introduced before (23) (and h ~ ( r 2 ) = h ( r ) \widetilde h(r^2)=h(r) h ( r 2 ) = h ( r ) in the notation of the proof of Lemma 2.5). The function K K K is continuous, radial, nonnegative, and supported in the closed ball of radius two. For q = ∣ z ∣ 2 q=|z|^2 q = ∣ z ∣ 2 , the identity P K = ψ ∗ ψ \mathcal{P}K=\psi*\psi P K = ψ ∗ ψ and radiality of K K K give
h ~ ( q ) = ∫ R 4 K ( z , y ) d y = ∫ R 4 k ( q + ∣ y ∣ 2 ) d y = 2 π 2 ∫ 0 ∞ r 3 k ( q + r 2 ) d r . \widetilde h(q)=\int_{\mathbb{R}^4}K(z,y)\,dy
=\int_{\mathbb{R}^4}k\left(\sqrt{q+|y|^2}\right)\,dy
=2\pi^2\int_0^\infty r^3k\left(\sqrt{q+r^2}\right)\,dr. h ( q ) = ∫ R 4 K ( z , y ) d y = ∫ R 4 k ( q + ∣ y ∣ 2 ) d y = 2 π 2 ∫ 0 ∞ r 3 k ( q + r 2 ) d r .
Setting b = q + r 2 b=q+r^2 b = q + r 2 , so that r 3 d r = 1 2 ( b − q ) d b r^3\,dr=\frac{1}{2}(b-q)\,db r 3 d r = 2 1 ( b − q ) d b ,
h ~ ( q ) = π 2 ∫ q ∞ ( b − q ) k ( b ) d b , q ≥ 0. (25) \widetilde h(q)=\pi^2\int_q^\infty(b-q)k(\sqrt b)\,db,\qquad q\geq0. \tag*{(25)} h ( q ) = π 2 ∫ q ∞ ( b − q ) k ( b ) d b , q ≥ 0. (25)
The fundamental theorem of calculus then assures
h ~ ′ ( q ) = − π 2 ∫ q ∞ k ( b ) d b , h ~ ′ ′ ( q ) = π 2 k ( q ) ≥ 0. (26) \widetilde h'(q)=-\pi^2\int_q^\infty k(\sqrt b)\,db,\qquad
\widetilde h''(q)=\pi^2k(\sqrt q)\geq0. \tag*{(26)} h ′ ( q ) = − π 2 ∫ q ∞ k ( b ) d b , h ′′ ( q ) = π 2 k ( q ) ≥ 0. (26)
Since H ( z ) = h ~ ( ∣ z ∣ 2 ) H(z)=\widetilde h(|z|^2) H ( z ) = h ( ∣ z ∣ 2 ) , the chain rule gives
∂ i 2 H ( z ) = 2 h ~ ′ ( ∣ z ∣ 2 ) + 4 z i 2 h ~ ′ ′ ( ∣ z ∣ 2 ) . \partial_i^2H(z)=2\widetilde h'(|z|^2)+4z_i^2\widetilde h''(|z|^2). ∂ i 2 H ( z ) = 2 h ′ ( ∣ z ∣ 2 ) + 4 z i 2 h ′′ ( ∣ z ∣ 2 ) .
On the other hand, at z = r e 1 z=re_1 z = r e 1 , the covariance identity (23) yields
C 11 ( r e 1 ) = ( − Δ H + ∂ 1 2 H ) ( r e 1 ) = − ∑ i = 2 d ∂ i 2 H ( r e 1 ) = − 2 ( d − 1 ) h ~ ′ ( r 2 ) . C_{11}(re_1)=(-\Delta H+\partial_1^2H)(re_1)
=-\sum_{i=2}^d\partial_i^2H(re_1)
=-2(d-1)\widetilde h'(r^2). C 11 ( r e 1 ) = ( − Δ H + ∂ 1 2 H ) ( r e 1 ) = − i = 2 ∑ d ∂ i 2 H ( r e 1 ) = − 2 ( d − 1 ) h ′ ( r 2 ) .
Since D L ( r ) = D 11 ( r e 1 ) D_L(r)=D_{11}(re_1) D L ( r ) = D 11 ( r e 1 ) and D ( z ) = 2 ( C ( 0 ) − C ( z ) ) D(z)=2(C(0)-C(z)) D ( z ) = 2 ( C ( 0 ) − C ( z )) , it follows that
D L ( r ) = 2 ( C 11 ( 0 ) − C 11 ( r e 1 ) ) = 2 C 11 ( 0 ) + 4 ( d − 1 ) h ~ ′ ( r 2 ) . D_L(r)=2(C_{11}(0)-C_{11}(re_1))=2C_{11}(0)+4(d-1)\tilde{h}'(r^2). D L ( r ) = 2 ( C 11 ( 0 ) − C 11 ( r e 1 )) = 2 C 11 ( 0 ) + 4 ( d − 1 ) h ~ ′ ( r 2 ) .
Differentiating and using (26) proves (24) . For 0 ≤ r < 2 0\le r<2 0 ≤ r < 2 , the open unit balls centered at 0 0 0 and r e 1 re_1 r e 1 in R d + 4 \mathbb{R}^{d+4} R d + 4 have a nonempty open intersection. Both factors of the convolution are strictly positive on that set, so k ( r ) > 0 k(r)>0 k ( r ) > 0 . For r ≥ 2 r\ge2 r ≥ 2 their interiors are disjoint and k ( r ) = 0 k(r)=0 k ( r ) = 0 .
Auxiliary observables and their gains
Logarithmic continuation and scale cutoffs
The observable r β r^\beta r β , r = ∣ z ∣ r=|z| r = ∣ z ∣ , is modified in two ways before use. Below the finest scale ℓ J \ell_J ℓ J it is continued as a logarithm. There the field is Lipschitz with constant λ J \lambda_J λ J , and the continuation keeps the quantity r Φ ′ ( r ) r\Phi'(r) r Φ ′ ( r ) bounded below by a constant, which is what the gain at scale J J J requires (Lemma 3.1 , (34) ); this is the source of the uniformity in J J J , at the price of the term ℓ J β \ell_J^\beta ℓ J β in (10) . Roughly speaking, we want Φ \Phi Φ to behave like
{ r β , above a small threshold , constant + β ε β log r , below that threshold . \begin{cases}
r^\beta, & \text{above a small threshold},\\
\text{constant}+\beta\varepsilon^\beta\log r, & \text{below that threshold}.
\end{cases} { r β , constant + β ε β log r , above a small threshold , below that threshold .
The transition must preserve the sign of the averaged gain (cf. Lemma 3.1 ) and the derivative bounds used later.
Second, for each scale b b b a function Φ b \Phi_b Φ b is formed in which the scales somewhat coarser than r r r , starting roughly r L rL r L , are switched off; without this, the corrected observable of § 5 would contain one term of size κ r β \kappa r^\beta κ r β for every coarse scale and would not be uniformly bounded. The cutoff is nondecreasing in r r r , as the sign computation of Lemma 3.1 requires.
It is easier to construct H = r Φ ′ H=r\Phi' H = r Φ ′ first. To this end, retain constant v ∗ v_* v ∗ from Lemma 2.5 . Fix ρ > 0 \rho>0 ρ > 0 so small that
2 ρ < min ( v ∗ , 1 / 4 ) , ε = ρ ℓ J . (27) 2\rho<\min(v_*,1/4),\qquad\varepsilon=\rho\ell_J. \tag*{(27)} 2 ρ < min ( v ∗ , 1/4 ) , ε = ρ ℓ J . (27)
Let ϑ \vartheta ϑ be a nonnegative even smooth function with integral one, supported in the interval ( − log 2 , log 2 ) (-\log2,\log2) ( − log 2 , log 2 ) , and set g ( v ) = ∫ ϑ ( a ) max ( v − a , 0 ) d a g(v)=\int\vartheta(a)\max(v-a,0)\,\mathrm{d}a g ( v ) = ∫ ϑ ( a ) max ( v − a , 0 ) d a . Then g ( v ) ≥ max ( v , 0 ) g(v)\ge\max(v,0) g ( v ) ≥ max ( v , 0 ) , g ′ ≥ 0 g'\ge0 g ′ ≥ 0 , g ( v ) = 0 g(v)=0 g ( v ) = 0 for v ≤ − log 2 v\le-\log2 v ≤ − log 2 , and g ( v ) = v g(v)=v g ( v ) = v for v ≥ log 2 v\ge\log2 v ≥ log 2 . Define
H ( r ) = β ε β exp ( β g ( log ( r / ε ) ) ) , Φ ( r ) = ( 2 ε ) β + ∫ 2 ε r H ( ϱ ) ϱ d ϱ , r > 0. (28) H(r)=\beta\varepsilon^\beta\exp\bigl(\beta g(\log(r/\varepsilon))\bigr),\qquad\Phi(r)=(2\varepsilon)^\beta+\int_{2\varepsilon}^{r}\frac{H(\varrho)}{\varrho}\,\mathrm{d}\varrho,\quad r>0. \tag*{(28)} H ( r ) = β ε β exp ( β g ( log ( r / ε )) ) , Φ ( r ) = ( 2 ε ) β + ∫ 2 ε r ϱ H ( ϱ ) d ϱ , r > 0. (28)
We also write Φ ( z ) = Φ ( ∣ z ∣ ) \Phi(z)=\Phi(|z|) Φ ( z ) = Φ ( ∣ z ∣ ) on the punctured local separation region. Then H ′ ≥ 0 H'\ge0 H ′ ≥ 0 ,
Φ ( r ) = r β ( r ≥ 2 ε ) , Φ ( r ) ≤ r β ( r > 0 ) , r Φ ′ ( r ) = H ( r ) > 0. (29) \Phi(r)=r^\beta\ (r\ge2\varepsilon),\qquad\Phi(r)\le r^\beta\ (r>0),\qquad r\Phi'(r)=H(r)>0. \tag*{(29)} Φ ( r ) = r β ( r ≥ 2 ε ) , Φ ( r ) ≤ r β ( r > 0 ) , r Φ ′ ( r ) = H ( r ) > 0. (29)
Indeed H ( r ) ≥ β r β H(r)\ge\beta r^\beta H ( r ) ≥ β r β , and integrating backwards from 2 ε 2\varepsilon 2 ε proves the inequality. All in all,
{ Φ ( r ) = r β , r ≥ 2 ε , Φ ( r ) = C ε + β ε β log r , 0 < r ≤ ε / 2 , \begin{cases}
\Phi(r)=r^\beta, & r\ge2\varepsilon,\\
\Phi(r)=C_\varepsilon+\beta\varepsilon^\beta\log r, & 0<r\le\varepsilon/2,
\end{cases} { Φ ( r ) = r β , Φ ( r ) = C ε + β ε β log r , r ≥ 2 ε , 0 < r ≤ ε /2 ,
as desired. It has no finite extension at r = 0 r=0 r = 0 , since Φ ( r ) → − ∞ \Phi(r)\to-\infty Φ ( r ) → − ∞ as r ↓ 0 r\downarrow0 r ↓ 0 .
Let us also set the notation
S ( r ) = max ( r , ℓ J ) , S ~ ( r ) = r 2 + ℓ J 2 , S ≤ S ~ ≤ 2 S . S(r)=\max(r,\ell_J),\qquad\widetilde{S}(r)=\sqrt{r^2+\ell_J^2},\qquad S\leq\widetilde{S}\leq\sqrt{2}S. S ( r ) = max ( r , ℓ J ) , S ( r ) = r 2 + ℓ J 2 , S ≤ S ≤ 2 S .
to be used below. The explicit construction gives, for k = 1 , 2 , 3 k=1,2,3 k = 1 , 2 , 3 ,
∣ ∇ k Φ ( z ) ∣ ≤ C S ( r ) β r − k . (30) \lvert\nabla^k\Phi(z)\rvert\leq CS(r)^\beta r^{-k}. \tag*{(30)} ∣ ∇ k Φ ( z )∣ ≤ C S ( r ) β r − k . (30)
This follows by differentiating H ( r ) / r H(r)/r H ( r ) / r : all derivatives of g g g are bounded, since ϑ ∈ C c ∞ \vartheta\in C^\infty_c ϑ ∈ C c ∞ , H ( r ) ≤ C max ( r , ε ) β H(r)\leq C\max(r,\varepsilon)^\beta H ( r ) ≤ C max ( r , ε ) β , and each logarithmic differentiation contributes a factor r − 1 r^{-1} r − 1 .
Choose a smooth nonincreasing χ : [ 0 , ∞ ) → [ 0 , 1 ] \chi:[0,\infty)\to[0,1] χ : [ 0 , ∞ ) → [ 0 , 1 ] equal to one on [ 0 , 1 ] [0,1] [ 0 , 1 ] and zero on [ 2 , ∞ ) [2,\infty) [ 2 , ∞ ) . For L ≥ 2 L\geq2 L ≥ 2 , let
w j ( r ) = χ ( ℓ j L S ~ ( r ) ) , Φ j ( r ) = − ∫ r 1 w j ( ϱ ) Φ ′ ( ϱ ) d ϱ . (31) w_j(r)=\chi\left(\frac{\ell_j}{L\widetilde{S}(r)}\right),\qquad\Phi_j(r)=-\int_r^1 w_j(\varrho)\Phi'(\varrho)\,\mathrm{d}\varrho. \tag*{(31)} w j ( r ) = χ ( L S ( r ) ℓ j ) , Φ j ( r ) = − ∫ r 1 w j ( ϱ ) Φ ′ ( ϱ ) d ϱ . (31)
The functions w j w_j w j are nondecreasing in r r r ; they depend on J J J (through S ~ \widetilde{S} S ) and on L L L , which the notation suppresses. Their first two derivatives obey ∣ w j ( k ) ( r ) ∣ ≤ C r − k \lvert w_j^{(k)}(r)\rvert\leq Cr^{-k} ∣ w j ( k ) ( r )∣ ≤ C r − k for k = 1 , 2 k=1,2 k = 1 , 2 , with constants independent of j , J , L j,J,L j , J , L . Hence
∣ ∇ k Φ j ( z ) ∣ ≤ C S β r − k 1 { ℓ j ≤ 3 L S } , k = 1 , 2 , 3. (32) \lvert\nabla^k\Phi_j(z)\rvert\leq CS^\beta r^{-k}\mathbf{1}_{\{\ell_j\leq3LS\}},\qquad k=1,2,3. \tag*{(32)} ∣ ∇ k Φ j ( z )∣ ≤ C S β r − k 1 { ℓ j ≤ 3 L S } , k = 1 , 2 , 3. (32)
The factor 3 accommodates 2 S ~ ≤ 2 2 S 2\widetilde{S}\leq2\sqrt{2}S 2 S ≤ 2 2 S . Where the indicator vanishes, w j w_j w j and all its derivatives vanish.
The quantity Γ j ( r ) \Gamma_j(r) Γ j ( r ) introduced in the following lemma is the averaged jump of the corrected observable when scale j j j is refreshed while the particles are at separation r r r (Proposition 6.6 ). It is one of the core ingredients in the argument.
Lemma 3.1 (Positivity and a quantitative one-scale gain). Define, for 0 < r ≤ 1 / 4 0<r\leq1/4 0 < r ≤ 1/4 ,
Γ j ( r ) = σ 2 A j 2 2 D ( z / ℓ j ) : D 2 Φ j ( z ) , ∣ z ∣ = r . (33) \Gamma_j(r)=\frac{\sigma^2 A_j^2}{2}D(z/\ell_j):D^2\Phi_j(z),\qquad\lvert z\rvert=r. \tag*{(33)} Γ j ( r ) = 2 σ 2 A j 2 D ( z / ℓ j ) : D 2 Φ j ( z ) , ∣ z ∣ = r . (33) The functions Γ j = Γ j ( r ) \Gamma_j=\Gamma_j(r) Γ j = Γ j ( r ) are well-defined independently of the direction of z z z and Γ j ≥ 0 \Gamma_j\geq0 Γ j ≥ 0 . Moreover, the following lower bounds hold.
There is c 1 > 0 c_1>0 c 1 > 0 , independent of J , L , κ J,L,\kappa J , L , κ , such that
τ J Γ J ( r ) ≥ c 1 κ if 0 < r ≤ min ( 2 ℓ J , 1 / 4 ) . (34) \tau_J\Gamma_J(r)\geq c_1\kappa\qquad\text{if}\qquad0<r\leq\min(2\ell_J,1/4). \tag*{(34)} τ J Γ J ( r ) ≥ c 1 κ if 0 < r ≤ min ( 2 ℓ J , 1/4 ) . (34) For each fixed 0 < a < b < ∞ 0<a<b<\infty 0 < a < b < ∞ there is c a , b > 0 c_{a,b}>0 c a , b > 0 such that
τ j Γ j ( r ) ≥ c a , b κ (35) \tau_j\Gamma_j(r)\geq c_{a,b}\kappa \tag*{(35)} τ j Γ j ( r ) ≥ c a , b κ (35) whenever r ≥ 2 ε r\geq2\varepsilon r ≥ 2 ε , r / ℓ j ∈ [ a , b ] r/\ell_j\in[a,b] r / ℓ j ∈ [ a , b ] , r ≤ 1 / 4 r\leq1/4 r ≤ 1/4 , and w j ( r ) = 1 w_j(r)=1 w j ( r ) = 1 .
The hypothesis w j ( r ) = 1 w_j(r)=1 w j ( r ) = 1 in (35) holds whenever ℓ j ≤ L r \ell_j\leq Lr ℓ j ≤ L r , since S ~ ( r ) ≥ r \widetilde{S}(r)\geq r S ( r ) ≥ r and χ = 1 \chi=1 χ = 1 on [ 0 , 1 ] [0,1] [ 0 , 1 ] . In the application in Lemma 7.1 , we will use it for the selected scale which satisfies ℓ j ≤ 2 r ≤ L r \ell_j\leq2r\leq Lr ℓ j ≤ 2 r ≤ L r , so this hypothesis is fulfilled.
Proof. For a radial function F ( z ) = F ( r ) F(z)=F(r) F ( z ) = F ( r ) , r = ∣ z ∣ > 0 r=\lvert z\rvert>0 r = ∣ z ∣ > 0 , the Hessian acts by F ′ ′ ( r ) F^{\prime\prime}(r) F ′′ ( r ) on the radial direction z ^ \hat{z} z ^ and by F ′ ( r ) / r F^{\prime}(r)/r F ′ ( r ) / r on its ( d − 1 ) (d-1) ( d − 1 ) -dimensional orthogonal complement, as follows by differentiating ∇ F ( z ) = F ′ ( r ) z ^ \nabla F(z)=F^{\prime}(r)\hat{z} ∇ F ( z ) = F ′ ( r ) z ^ . Since z / ℓ z/\ell z / ℓ has the same direction as z z z for ℓ > 0 \ell>0 ℓ > 0 , (2) (proved with Lemma 2.5 ) gives
D ( z / ℓ ) : D 2 F ( z ) = D L ( r / ℓ ) F ′ ′ ( r ) + ( d − 1 ) D N ( r / ℓ ) F ′ ( r ) r . D(z/\ell):D^2F(z)=D_L(r/\ell)F^{\prime\prime}(r)+(d-1)D_N(r/\ell)\frac{F^{\prime}(r)}{r}. D ( z / ℓ ) : D 2 F ( z ) = D L ( r / ℓ ) F ′′ ( r ) + ( d − 1 ) D N ( r / ℓ ) r F ′ ( r ) .
The factor d − 1 d-1 d − 1 counts the transverse directions. Setting H F ( r ) = r F ′ ( r ) H_F(r)=rF^{\prime}(r) H F ( r ) = r F ′ ( r ) , we have
F ′ ( r ) r = H F ( r ) r 2 , F ′ ′ ( r ) = H F ′ ( r ) r − H F ( r ) r 2 . \frac{F^{\prime}(r)}{r}=\frac{H_F(r)}{r^2},\qquad F^{\prime\prime}(r)=\frac{H_F^{\prime}(r)}{r}-\frac{H_F(r)}{r^2}. r F ′ ( r ) = r 2 H F ( r ) , F ′′ ( r ) = r H F ′ ( r ) − r 2 H F ( r ) .
Then
D ( z / ℓ ) : D 2 F ( z ) = D L ( r / ℓ ) H F ′ ( r ) r + ( ( d − 1 ) D N ( r / ℓ ) − D L ( r / ℓ ) ) H F ( r ) r 2 . (36) D(z/\ell):D^2F(z)=D_L(r/\ell)\frac{H_F^{\prime}(r)}{r}+\left((d-1)D_N(r/\ell)-D_L(r/\ell)\right)\frac{H_F(r)}{r^2}. \tag*{(36)} D ( z / ℓ ) : D 2 F ( z ) = D L ( r / ℓ ) r H F ′ ( r ) + ( ( d − 1 ) D N ( r / ℓ ) − D L ( r / ℓ ) ) r 2 H F ( r ) . (36)
For F = Φ j F=\Phi_j F = Φ j , we have H F = r Φ j ′ = w j H ≥ 0 H_F=r\Phi_j^{\prime}=w_jH\geq0 H F = r Φ j ′ = w j H ≥ 0 and H F ′ = w j ′ H + w j H ′ ≥ 0 H_F^{\prime}=w_j^{\prime}H+w_jH^{\prime}\geq0 H F ′ = w j ′ H + w j H ′ ≥ 0 , since w j , H w_j,H w j , H are nonnegative and nondecreasing by construction. Moreover,
( d − 1 ) D N − D L = ( d − 2 ) D L + ( d − 1 ) ( D N − D L ) ≥ 0 , (d-1)D_N-D_L=(d-2)D_L+(d-1)(D_N-D_L)\geq0, ( d − 1 ) D N − D L = ( d − 2 ) D L + ( d − 1 ) ( D N − D L ) ≥ 0 ,
so (ND) and d ≥ 2 d\geq2 d ≥ 2 prove Γ j ≥ 0 \Gamma_j\geq0 Γ j ≥ 0 .
We next establish the quantitative bounds. Whenever r ≥ 2 ε r\geq2\varepsilon r ≥ 2 ε and w j ( r ) = 1 w_j(r)=1 w j ( r ) = 1 , we have w j ′ ( r ) = 0 w_j^{\prime}(r)=0 w j ′ ( r ) = 0 , since w j w_j w j is differentiable and bounded above by one. Since Φ \Phi Φ agrees with r β r^\beta r β on [ 2 ε , ∞ ) [2\varepsilon,\infty) [ 2 ε , ∞ ) and is smooth on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , it follows that
Φ j ′ ( r ) = w j ( r ) Φ ′ ( r ) = β r β − 1 , Φ j ′ ′ ( r ) = w j ′ ( r ) Φ ′ ( r ) + w j ( r ) Φ ′ ′ ( r ) = − α β r β − 2 . \Phi_j^{\prime}(r)=w_j(r)\Phi^{\prime}(r)=\beta r^{\beta-1},\qquad\Phi_j^{\prime\prime}(r)=w_j^{\prime}(r)\Phi^{\prime}(r)+w_j(r)\Phi^{\prime\prime}(r)=-\alpha\beta r^{\beta-2}. Φ j ′ ( r ) = w j ( r ) Φ ′ ( r ) = β r β − 1 , Φ j ′′ ( r ) = w j ′ ( r ) Φ ′ ( r ) + w j ( r ) Φ ′′ ( r ) = − α β r β − 2 .
Substituting into the radial contraction formula gives
τ j Γ j ( r ) = σ 2 τ j A j 2 2 [ D L ( r / ℓ j ) Φ j ′ ′ ( r ) + ( d − 1 ) D N ( r / ℓ j ) Φ j ′ ( r ) r ] = σ 2 β τ j ℓ j − β 2 ( r / ℓ j ) β − 2 [ ( d − 1 ) D N ( r / ℓ j ) − α D L ( r / ℓ j ) ] . (37) \begin{aligned}
\tau_j\Gamma_j(r)=\frac{\sigma^2\tau_j A_j^2}{2}\left[D_L(r/\ell_j)\Phi_j^{\prime\prime}(r)+(d-1)D_N(r/\ell_j)\frac{\Phi_j^{\prime}(r)}{r}\right] \\
&=\frac{\sigma^2\beta\tau_j\ell_j^{-\beta}}{2}(r/\ell_j)^{\beta-2}\left[(d-1)D_N(r/\ell_j)-\alpha D_L(r/\ell_j)\right]. \tag*{(37)}
\end{aligned} τ j Γ j ( r ) = 2 σ 2 τ j A j 2 [ D L ( r / ℓ j ) Φ j ′′ ( r ) + ( d − 1 ) D N ( r / ℓ j ) r Φ j ′ ( r ) ] = 2 σ 2 β τ j ℓ j − β ( r / ℓ j ) β − 2 [ ( d − 1 ) D N ( r / ℓ j ) − α D L ( r / ℓ j ) ] . (37)
Define
K ( q ) = σ 2 β 2 q β − 2 [ ( d − 1 ) D N ( q ) − α D L ( q ) ] , q > 0. K(q)=\frac{\sigma^2\beta}{2}q^{\beta-2}\left[(d-1)D_N(q)-\alpha D_L(q)\right],\qquad q>0. K ( q ) = 2 σ 2 β q β − 2 [ ( d − 1 ) D N ( q ) − α D L ( q ) ] , q > 0.
By Lemma 2.5 and (ND),
( d − 1 ) D N ( q ) − α D L ( q ) ≥ ( d − 1 − α ) D L ( q ) > 0. (d-1)D_N(q)-\alpha D_L(q)\geq(d-1-\alpha)D_L(q)>0. ( d − 1 ) D N ( q ) − α D L ( q ) ≥ ( d − 1 − α ) D L ( q ) > 0.
Thus K K K is continuous and strictly positive on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . For any fixed 0 < a < b < ∞ 0<a<b<\infty 0 < a < b < ∞ , set
c a , b = 1 2 min q ∈ [ a , b ] K ( q ) > 0. c_{a,b}=\frac{1}{2}\min_{q\in[a,b]}K(q)>0. c a , b = 2 1 q ∈ [ a , b ] min K ( q ) > 0.
Every radius satisfying the hypotheses of (35) has r ≥ 2 ε r\geq2\varepsilon r ≥ 2 ε , w j ( r ) = 1 w_j(r)=1 w j ( r ) = 1 , and r / ℓ j ∈ [ a , b ] r/\ell_j\in[a,b] r / ℓ j ∈ [ a , b ] . Hence (37) gives τ j Γ j ( r ) = ( τ j ℓ j − β ) K ( r / ℓ j ) ≥ 1 2 κ K ( r / ℓ j ) ≥ c a , b κ \tau_j\Gamma_j(r)=(\tau_j\ell_j^{-\beta})K(r/\ell_j)\geq\frac{1}{2}\kappa K(r/\ell_j)\geq c_{a,b}\kappa τ j Γ j ( r ) = ( τ j ℓ j − β ) K ( r / ℓ j ) ≥ 2 1 κ K ( r / ℓ j ) ≥ c a , b κ , proving (35) .
For (3.8), note that w J ≡ 1 w_J \equiv1 w J ≡ 1 , since ℓ J / ( L S ~ ( r ) ) ≤ 1 / L ≤ 1 / 2 \ell_J/(L\widetilde{S}(r)) \le1/L \le1/2 ℓ J / ( L S ( r )) ≤ 1/ L ≤ 1/2 . For 2 ε ≤ r ≤ min ( 2 ℓ J , 1 / 4 ) 2\varepsilon\le r \le\min(2\ell_J,1/4) 2 ε ≤ r ≤ min ( 2 ℓ J , 1/4 ) , (37) gives
τ J Γ J ( r ) ≥ κ 2 min q ∈ [ 2 ρ , 2 ] K ( q ) > 0. \tau_J\Gamma_J(r) \ge\frac{\kappa}{2}\min_{q\in[2\rho,2]}K(q)>0. τ J Γ J ( r ) ≥ 2 κ q ∈ [ 2 ρ , 2 ] min K ( q ) > 0.
For 0 < r ≤ 2 ε 0<r\le2\varepsilon 0 < r ≤ 2 ε , we have r / ℓ J ≤ 2 ρ < v ∗ r/\ell_J\le2\rho<v_* r / ℓ J ≤ 2 ρ < v ∗ . Using (15) , (36) , H ′ ≥ 0 H'\ge0 H ′ ≥ 0 , ( d − 1 ) D N − D L ≥ ( d − 1 ) ( D N − D L ) (d-1)D_N-D_L\ge(d-1)(D_N-D_L) ( d − 1 ) D N − D L ≥ ( d − 1 ) ( D N − D L ) , and (22) , we obtain
τ J Γ J ( r ) ≥ c κ ℓ J − β H ( r ) ≥ c κ ρ β , \tau_J\Gamma_J(r)\ge c\kappa\ell_J^{-\beta}H(r)\ge c\kappa\rho^\beta, τ J Γ J ( r ) ≥ c κ ℓ J − β H ( r ) ≥ c κ ρ β ,
where the last inequality uses H ( r ) ≥ β ε β H(r)\ge\beta\varepsilon^\beta H ( r ) ≥ β ε β and ε = ρ ℓ J \varepsilon=\rho\ell_J ε = ρ ℓ J . Taking c 1 : = min ( 1 2 min q ∈ [ 2 ρ , 2 ] K ( q ) , c ρ β ) c_1:=\min\left(\frac{1}{2}\min_{q\in[2\rho,2]}K(q),c\rho^\beta\right) c 1 := min ( 2 1 min q ∈ [ 2 ρ , 2 ] K ( q ) , c ρ β ) , which is independent of J , L , κ J,L,\kappa J , L , κ because ρ \rho ρ is fixed, the two bounds prove (34) . □
Quadratic continuation at the diagonal
For the diffusing pair we modify Φ \Phi Φ near the diagonal. Its logarithmic singularity is useful for the trajectory argument, but must be removed before we apply the advection–diffusion equation. We replace it by a quadratic function on a ball of radius r ♯ r_\sharp r ♯ , comparable to ℓ J \ell_J ℓ J .
The choice of this radius serves two purposes. Inside the quadratic region, diffusion gives a positive contribution of order κ \kappa κ ; outside it, the drift gives the gain already established in the preceding gain estimate. The modification preserves the derivative bounds needed for the correctors. We record these facts in Lemma 3.3 and then transfer the estimates for the corrected observable in Corollary 5.3.
Let ε 0 ∈ ( 0 , 1 ] \varepsilon_0\in(0,1] ε 0 ∈ ( 0 , 1 ] and J ≥ j 0 J\ge j_0 J ≥ j 0 , and set
μ : = μ J = ε 0 ℓ J 1 + α , r ♯ : = ( 4 β μ ε β κ ) 1 / 2 = ( 4 β ρ β ε 0 κ ) 1 / 2 ℓ J , (38) \mu:=\mu_J=\varepsilon_0\ell_J^{1+\alpha},\qquad
r_\sharp:=\left(\frac{4\beta\mu\varepsilon^\beta}{\kappa}\right)^{1/2}
=\left(\frac{4\beta\rho^\beta\varepsilon_0}{\kappa}\right)^{1/2}\ell_J, \tag*{(38)} μ := μ J = ε 0 ℓ J 1 + α , r ♯ := ( κ 4 β μ ε β ) 1/2 = ( κ 4 β ρ β ε 0 ) 1/2 ℓ J , (38)
with ε = ρ ℓ J \varepsilon=\rho\ell_J ε = ρ ℓ J as in (27) ; the second expression uses 1 + α + β = 2 1+\alpha+\beta=2 1 + α + β = 2 . We assume throughout
ε 0 ≤ κ ρ 1 + α 16 β , equivalently r ♯ ≤ ε / 2. (39) \varepsilon_0\le\frac{\kappa\rho^{1+\alpha}}{16\beta},
\qquad\text{equivalently}\qquad r_\sharp\le\varepsilon/2. \tag*{(39)} ε 0 ≤ 16 β κ ρ 1 + α , equivalently r ♯ ≤ ε /2. (39)
In particular r ♯ ≤ ℓ J / 16 r_\sharp\le\ell_J/16 r ♯ ≤ ℓ J /16 , and the ratios r ♯ / ℓ J r_\sharp/\ell_J r ♯ / ℓ J and ε / r ♯ \varepsilon/r_\sharp ε / r ♯ do not depend on J J J .
To interpolate between the quadratic core and the original function, fix a non-decreasing Υ ∈ C ∞ ( [ 0 , ∞ ) ) \Upsilon\in C^\infty([0,\infty)) Υ ∈ C ∞ ([ 0 , ∞ )) such that Υ ( s ) = s 2 \Upsilon(s)=s^2 Υ ( s ) = s 2 on [ 0 , 1 2 ] [0,\frac12] [ 0 , 2 1 ] and Υ = 1 \Upsilon=1 Υ = 1 on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) . In particular, 0 ≤ Υ ≤ 1 0\le\Upsilon\le1 0 ≤ Υ ≤ 1 and Υ ( 1 4 ) = 1 16 \Upsilon(\frac14)=\frac1{16} Υ ( 4 1 ) = 16 1 .
Definition 3.2. For r > 0 r>0 r > 0 let H ♯ ( r ) : = Υ ( r / r ♯ ) H ( r ) H^\sharp(r):=\Upsilon(r/r_\sharp)H(r) H ♯ ( r ) := Υ ( r / r ♯ ) H ( r ) , with H H H from (28) , and H ♯ ( 0 ) : = 0 H^\sharp(0):=0 H ♯ ( 0 ) := 0 ; for 0 ≤ r ≤ 1 0\le r\le1 0 ≤ r ≤ 1 let
Φ ♯ ( r ) : = ( 2 ε ) β − ∫ r 2 ε H ♯ ( ϱ ) ϱ d ϱ , \Phi^\sharp(r):=(2\varepsilon)^\beta-\int_r^{2\varepsilon}\frac{H^\sharp(\varrho)}{\varrho}\,d\varrho, Φ ♯ ( r ) := ( 2 ε ) β − ∫ r 2 ε ϱ H ♯ ( ϱ ) d ϱ , Φ J ♯ ( r ) : = − ∫ r 1 w J ( ϱ ) H ♯ ( ϱ ) ϱ d ϱ , \Phi_J^\sharp(r):=-\int_r^1w_J(\varrho)\frac{H^\sharp(\varrho)}{\varrho}\,d\varrho, Φ J ♯ ( r ) := − ∫ r 1 w J ( ϱ ) ϱ H ♯ ( ϱ ) d ϱ , with w J w_J w J from (31) (for r > 2 ε r>2\varepsilon r > 2 ε , ∫ r 2 ε \int_r^{2\varepsilon} ∫ r 2 ε means − ∫ 2 ε r -\int_{2\varepsilon}^r − ∫ 2 ε r ). Write Φ ♯ ( z ) = Φ ♯ ( ∣ z ∣ ) \Phi^\sharp(z)=\Phi^\sharp(|z|) Φ ♯ ( z ) = Φ ♯ ( ∣ z ∣ ) ,
Φ j ♯ ( z ) = Φ j ♯ ( ∣ z ∣ ) \Phi_j^\sharp(z)=\Phi_j^\sharp(|z|) Φ j ♯ ( z ) = Φ j ♯ ( ∣ z ∣ ) , and define the corresponding gain by
Γ j ♯ ( r ) : = σ 2 A j 2 2 D ( z / ℓ j ) : D 2 Φ j ♯ ( z ) , ∣ z ∣ = r . \Gamma_j^\sharp(r):=\frac{\sigma^2 A_j^2}{2}D(z/\ell_j):D^2\Phi_j^\sharp(z),\qquad|z|=r. Γ j ♯ ( r ) := 2 σ 2 A j 2 D ( z / ℓ j ) : D 2 Φ j ♯ ( z ) , ∣ z ∣ = r .
For r = ∣ z ∣ r=|z| r = ∣ z ∣ put S = max ( r , ℓ J ) S=\max(r,\ell_J) S = max ( r , ℓ J ) as before and
m ♯ = m ♯ ( r ) : = max ( r , r ♯ ) . m_\sharp=m_\sharp(r):=\max(r,r_\sharp). m ♯ = m ♯ ( r ) := max ( r , r ♯ ) .
The next lemma separates the properties of the modification from their use in the evolution argument. Parts (a)–(d) concern the test functions themselves; parts (e) and (f) give the positive contributions of diffusion and of the drift, respectively.
Lemma 3.3 (Properties of the cored test functions). Assume (39) . There are constants C , c 2 > 0 C,c_2>0 C , c 2 > 0 depending only on d , α , ψ , R , M , σ , ρ , ϑ , χ , Υ d,\alpha,\psi,R,M,\sigma,\rho,\vartheta,\chi,\Upsilon d , α , ψ , R , M , σ , ρ , ϑ , χ , Υ (not on J , L , κ , ε 0 J,L,\kappa,\varepsilon_0 J , L , κ , ε 0 ) such that the following hold.
(a) H ♯ ≥ 0 H^\sharp\geq0 H ♯ ≥ 0 and ( H ♯ ) ′ ≥ 0 (H^\sharp)'\geq0 ( H ♯ ) ′ ≥ 0 on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) ; Φ ♯ = Φ \Phi^\sharp=\Phi Φ ♯ = Φ and Φ j ♯ = Φ j \Phi_j^\sharp=\Phi_j Φ j ♯ = Φ j on [ r ♯ , 1 ] [r_\sharp,1] [ r ♯ , 1 ] ; Φ ♯ \Phi^\sharp Φ ♯ is nondecreasing and Φ ♯ ( r ) ≤ max ( r , r ♯ ) β ≤ r β + ℓ J β \Phi^\sharp(r)\leq\max(r,r_\sharp)^\beta\leq r^\beta+\ell_J^\beta Φ ♯ ( r ) ≤ max ( r , r ♯ ) β ≤ r β + ℓ J β .
(b) For 0 ≤ r ≤ r ♯ / 2 0\leq r\leq r_\sharp/2 0 ≤ r ≤ r ♯ /2 , H ♯ ( r ) = β ε β r 2 / r ♯ 2 H^\sharp(r)=\beta\varepsilon^\beta r^2/r_\sharp^2 H ♯ ( r ) = β ε β r 2 / r ♯ 2 ,
Φ ♯ ( z ) = Φ ♯ ( 0 ) + β ε β 2 r ♯ 2 ∣ z ∣ 2 , \Phi^\sharp(z)=\Phi^\sharp(0)+\frac{\beta\varepsilon^\beta}{2r_\sharp^2}|z|^2, Φ ♯ ( z ) = Φ ♯ ( 0 ) + 2 r ♯ 2 β ε β ∣ z ∣ 2 , Φ j ♯ ( z ) = Φ j ♯ ( 0 ) + β ε β 2 r ♯ 2 ∫ 0 ∣ z ∣ 2 ω ‾ j ( q ) d q , \Phi_j^\sharp(z)=\Phi_j^\sharp(0)+\frac{\beta\varepsilon^\beta}{2r_\sharp^2}\int_0^{|z|^2}\overline{\omega}_j(q)\,dq, Φ j ♯ ( z ) = Φ j ♯ ( 0 ) + 2 r ♯ 2 β ε β ∫ 0 ∣ z ∣ 2 ω j ( q ) d q , where ω ‾ j ( q ) : = χ ( ℓ j / ( L ( q + ℓ J 2 ) 1 / 2 ) ) \overline{\omega}_j(q):=\chi(\ell_j/(L(q+\ell_J^2)^{1/2})) ω j ( q ) := χ ( ℓ j / ( L ( q + ℓ J 2 ) 1/2 )) . Consequently Φ ♯ , Φ j ♯ ∈ C ∞ ( { ∣ z ∣ < 1 } ) \Phi^\sharp,\Phi_j^\sharp\in C^\infty(\{|z|<1\}) Φ ♯ , Φ j ♯ ∈ C ∞ ({ ∣ z ∣ < 1 }) ; in particular, they are smooth on the local separation region { ∣ z ∣ ≤ 1 / 4 } \{|z|\leq1/4\} { ∣ z ∣ ≤ 1/4 } , including z = 0 z=0 z = 0 , where they are used.
(c) Φ ♯ ( 0 ) ≥ − β ε β ( 1 + 4 β log ( 2 ε / r ♯ ) ) \Phi^\sharp(0)\geq-\beta\varepsilon^\beta(1+4^\beta\log(2\varepsilon/r_\sharp)) Φ ♯ ( 0 ) ≥ − β ε β ( 1 + 4 β log ( 2 ε / r ♯ )) .
(d) For 0 ≤ ∣ z ∣ ≤ 1 / 4 0\leq|z|\leq1/4 0 ≤ ∣ z ∣ ≤ 1/4 , k = 1 , 2 , 3 , 4 k=1,2,3,4 k = 1 , 2 , 3 , 4 and all j j j ,
∣ ∇ k Φ j ♯ ( z ) ∣ ≤ C S β m ♯ − k 1 { ℓ j ≤ 3 L S } , |\nabla^k\Phi_j^\sharp(z)|\leq CS^\beta m_\sharp^{-k}\mathbf{1}_{\{\ell_j\leq3LS\}}, ∣ ∇ k Φ j ♯ ( z ) ∣ ≤ C S β m ♯ − k 1 { ℓ j ≤ 3 L S } , ∣ ∇ Φ j ♯ ( z ) ∣ ≤ C S β r m ♯ − 2 1 { ℓ j ≤ 3 L S } , |\nabla\Phi_j^\sharp(z)|\leq CS^\beta r m_\sharp^{-2}\mathbf{1}_{\{\ell_j\leq3LS\}}, ∣∇ Φ j ♯ ( z ) ∣ ≤ C S β r m ♯ − 2 1 { ℓ j ≤ 3 L S } , and the same bounds without the indicator hold for Φ ♯ \Phi^\sharp Φ ♯ .
(e) Δ Φ ♯ ≥ 0 \Delta\Phi^\sharp\geq0 Δ Φ ♯ ≥ 0 on { ∣ z ∣ ≤ 1 / 4 } \{|z|\leq1/4\} { ∣ z ∣ ≤ 1/4 } , and 2 μ Δ Φ ♯ ( z ) = d κ / 2 2\mu\Delta\Phi^\sharp(z)=d\kappa/2 2 μ Δ Φ ♯ ( z ) = d κ /2 for ∣ z ∣ ≤ r ♯ / 2 |z|\leq r_\sharp/2 ∣ z ∣ ≤ r ♯ /2 .
(f) For every j j j and 0 ≤ r ≤ 1 / 4 0\leq r\leq1/4 0 ≤ r ≤ 1/4 , Γ j ♯ ( r ) ≥ 0 \Gamma_j^\sharp(r)\geq0 Γ j ♯ ( r ) ≥ 0 , and Γ j ♯ = Γ j \Gamma_j^\sharp=\Gamma_j Γ j ♯ = Γ j on ( r ♯ , 1 / 4 ] (r_\sharp,1/4] ( r ♯ , 1/4 ] . Moreover,
τ j Γ j ♯ ( r ) ≥ c 2 κ for r ∈ [ r ♯ / 4 , 3 ℓ J / 2 ] ∩ ( 0 , 1 / 4 ] , \tau_j\Gamma_j^\sharp(r)\geq c_2\kappa\quad\text{for }r\in[r_\sharp/4,3\ell_J/2]\cap(0,1/4], τ j Γ j ♯ ( r ) ≥ c 2 κ for r ∈ [ r ♯ /4 , 3 ℓ J /2 ] ∩ ( 0 , 1/4 ] , and (35) holds for Γ j ♯ \Gamma_j^\sharp Γ j ♯ with the same constant c a , b c_{a,b} c a , b .
Proof. We begin with the elementary properties in (a)–(c), then prove the derivative estimates and the lower bounds.
For (a), Υ \Upsilon Υ and H H H are nonnegative and nondecreasing (H ′ ≥ 0 H'\geq0 H ′ ≥ 0 by (29) ), hence so is their product. Since Υ = 1 \Upsilon=1 Υ = 1 on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) , H ♯ = H H^\sharp=H H ♯ = H on [ r ♯ , ∞ ) [r_\sharp,\infty) [ r ♯ , ∞ ) , which gives the two identities by comparison with (28) and (31) . Monotonicity follows from ( Φ ♯ ) ′ = (\Phi^\sharp)'= ( Φ ♯ ) ′ = H ♯ / r ≥ 0 \mathcal{H}^{\sharp}/r \ge0 H ♯ / r ≥ 0 . For r ≥ r ♯ r \ge r_{\sharp} r ≥ r ♯ , Φ ♯ ( r ) = Φ ( r ) ≤ r β \Phi^{\sharp}(r)=\Phi(r)\le r^{\beta} Φ ♯ ( r ) = Φ ( r ) ≤ r β by (29) ; for r < r ♯ r<r_{\sharp} r < r ♯ , Φ ♯ ( r ) ≤ Φ ♯ ( r ♯ ) ≤ r ♯ β \Phi^{\sharp}(r)\le\Phi^{\sharp}(r_{\sharp})\le r_{\sharp}^{\beta} Φ ♯ ( r ) ≤ Φ ♯ ( r ♯ ) ≤ r ♯ β ; finally r ♯ ≤ ℓ J r_{\sharp}\le\ell_J r ♯ ≤ ℓ J .
To prove (b), suppose that r ≤ r ♯ / 2 ≤ ε / 4 r\le r_{\sharp}/2\le\varepsilon/4 r ≤ r ♯ /2 ≤ ε /4 . We have log ( r / ε ) ≤ − log 2 \log(r/\varepsilon)\le-\log2 log ( r / ε ) ≤ − log 2 , so g = 0 g=0 g = 0 and H = β ε β H=\beta\varepsilon^{\beta} H = β ε β ; and Υ ( r / r ♯ ) = ( r / r ♯ ) 2 \Upsilon(r/r_{\sharp})=(r/r_{\sharp})^2 Υ ( r / r ♯ ) = ( r / r ♯ ) 2 . The formula for Φ ♯ \Phi^{\sharp} Φ ♯ follows by integration. For Φ j ♯ \Phi_j^{\sharp} Φ j ♯ , w j ( ϱ ) = ϖ ‾ j ( ϱ 2 ) w_j(\varrho)=\overline{\varpi}_j(\varrho^2) w j ( ϱ ) = ϖ j ( ϱ 2 ) by (31) , and ∫ 0 r 2 w j ( ϱ ) ϱ d ϱ = 1 2 ∫ 0 r 2 ϖ ‾ j \int_0^{r^2}w_j(\varrho)\varrho\,d\varrho=\frac{1}{2}\int_0^{r^2}\overline{\varpi}_j ∫ 0 r 2 w j ( ϱ ) ϱ d ϱ = 2 1 ∫ 0 r 2 ϖ j . Since χ \chi χ is smooth and q ↦ ( q + ℓ j 2 ) − 1 / 2 q\mapsto(q+\ell_j^2)^{-1/2} q ↦ ( q + ℓ j 2 ) − 1/2 is smooth on ( − ℓ j 2 , ∞ ) (-\ell_j^2,\infty) ( − ℓ j 2 , ∞ ) , ϖ ‾ j \overline{\varpi}_j ϖ j is smooth there, so both functions are smooth functions of ∣ z ∣ 2 |z|^2 ∣ z ∣ 2 near 0. Away from 0 they are radial with smooth profiles (g , ϑ , χ , Υ g,\vartheta,\chi,\Upsilon g , ϑ , χ , Υ are smooth).
For the lower bound in (c), split the defining integral at r ♯ r_{\sharp} r ♯ . On ( 0 , r ♯ ] (0,r_{\sharp}] ( 0 , r ♯ ] , H = β ε β H=\beta\varepsilon^\beta H = β ε β (as r ♯ ≤ ε / 2 r_{\sharp}\le\varepsilon/2 r ♯ ≤ ε /2 ), so
∫ 0 r ♯ H ♯ ϱ d ϱ = β ε β ∫ 0 1 Υ ( s ) s − 1 d s ≤ β ε β ( 1 8 + log 2 ) ≤ β ε β . \int_0^{r_{\sharp}}\frac{H^\sharp}{\varrho}\,d\varrho=\beta\varepsilon^\beta\int_0^1\Upsilon(s)s^{-1}\,ds
\le\beta\varepsilon^\beta\left(\frac{1}{8}+\log2\right)\le\beta\varepsilon^\beta. ∫ 0 r ♯ ϱ H ♯ d ϱ = β ε β ∫ 0 1 Υ ( s ) s − 1 d s ≤ β ε β ( 8 1 + log 2 ) ≤ β ε β .
Since g ( v ) ≤ max ( v , 0 ) + log 2 g(v)\le\max(v,0)+\log2 g ( v ) ≤ max ( v , 0 ) + log 2 , we have H ( ϱ ) ≤ β 2 β max ( ϱ , ε ) β ≤ β 4 β ε β H(\varrho)\le\beta2^\beta\max(\varrho,\varepsilon)^\beta\le\beta4^\beta\varepsilon^\beta H ( ϱ ) ≤ β 2 β max ( ϱ , ε ) β ≤ β 4 β ε β for ϱ ≤ 2 ε \varrho\le2\varepsilon ϱ ≤ 2 ε , hence
∫ r ♯ 2 ε H ϱ d ϱ ≤ β 4 β ε β log ( 2 ε / r ♯ ) . \int_{r_{\sharp}}^{2\varepsilon}\frac{H}{\varrho}\,d\varrho\le\beta4^\beta\varepsilon^\beta\log(2\varepsilon/r_{\sharp}). ∫ r ♯ 2 ε ϱ H d ϱ ≤ β 4 β ε β log ( 2 ε / r ♯ ) .
Combining these estimates and discarding the nonnegative term ( 2 ε ) β (2\varepsilon)^\beta ( 2 ε ) β proves (c).
We turn to the derivative estimates in (d). Write F = Φ j ♯ F=\Phi_j^\sharp F = Φ j ♯ , F ( z ) = f ( ∣ z ∣ ) F(z)=f(|z|) F ( z ) = f ( ∣ z ∣ ) , and H : = w j H ♯ = r f ′ \mathcal{H}:=w_jH^\sharp=rf' H := w j H ♯ = r f ′ . The proof for Φ ♯ \Phi^\sharp Φ ♯ is identical, with w j ≡ 1 w_j\equiv1 w j ≡ 1 . Away from the origin we use radial differentiation; inside the quadratic core we instead differentiate the smooth function of ∣ z ∣ 2 |z|^2 ∣ z ∣ 2 supplied by (b).
Case r ≥ r ♯ / 2 r\ge r_{\sharp}/2 r ≥ r ♯ /2 . Here m ♯ ≤ 2 r m_{\sharp}\le2r m ♯ ≤ 2 r . Radial differentiation gives
∇ k F ( z ) = ∑ m = 1 k f ( m ) ( r ) T k , m ( z ) , ∣ ∇ k F ( z ) ∣ ≤ C ∑ m ≤ k ∣ f ( m ) ( r ) ∣ r m − k . \nabla^kF(z)=\sum_{m=1}^{k}f^{(m)}(r)T_{k,m}(z),\qquad|\nabla^kF(z)|\le C\sum_{m\le k}|f^{(m)}(r)|r^{m-k}. ∇ k F ( z ) = m = 1 ∑ k f ( m ) ( r ) T k , m ( z ) , ∣ ∇ k F ( z ) ∣ ≤ C m ≤ k ∑ ∣ f ( m ) ( r ) ∣ r m − k .
where T k , m T_{k,m} T k , m is smooth away from the origin and homogeneous of degree m − k m-k m − k . These identities follow by induction, using ∇ f ( m ) ( r ) = f ( m + 1 ) ( r ) z ⃗ \nabla f^{(m)}(r)=f^{(m+1)}(r)\vec z ∇ f ( m ) ( r ) = f ( m + 1 ) ( r ) z . Since f ( m ) = ( H / r ) ( m − 1 ) f^{(m)}=(\mathcal{H}/r)^{(m-1)} f ( m ) = ( H / r ) ( m − 1 ) , it is enough to show that
∣ r i H ( i ) ( r ) ∣ ≤ C S β , 0 ≤ i ≤ 3. \left|r^i\mathcal{H}^{(i)}(r)\right|\le CS^\beta,\qquad0\le i\le3. r i H ( i ) ( r ) ≤ C S β , 0 ≤ i ≤ 3.
We verify this bound for the three factors of H = w j Υ H \mathcal{H}=w_j\Upsilon H H = w j Υ H . For H H H , the definition gives
( r d d r ) i H = H ⋅ P i ( g ′ , g ′ ′ , … ) , H ≤ β 2 β max ( r , ε ) β ≤ C S β , \left(r\frac{d}{dr}\right)^iH=H\cdot P_i(g',g'',\ldots),\qquad H\le\beta2^\beta\max(r,\varepsilon)^\beta\le CS^\beta, ( r d r d ) i H = H ⋅ P i ( g ′ , g ′′ , … ) , H ≤ β 2 β max ( r , ε ) β ≤ C S β ,
where P i P_i P i is a polynomial and the derivatives of g g g are bounded. For the core cutoff, write s = r / r ♯ s=r/r_{\sharp} s = r / r ♯ . Then r i d i d r i Υ ( r / r ♯ ) = s i Υ ( i ) ( s ) r^i\frac{d^i}{dr^i}\Upsilon(r/r_{\sharp})=s^i\Upsilon^{(i)}(s) r i d r i d i Υ ( r / r ♯ ) = s i Υ ( i ) ( s ) is bounded: for i ≥ 1 i\ge1 i ≥ 1 , the derivatives of Υ \Upsilon Υ are supported in [ 0 , 1 ] [0,1] [ 0 , 1 ] . Finally, for w j w_j w j , put ϕ = ℓ j / ( L S ~ ) \phi=\ell_j/(L\widetilde S) ϕ = ℓ j / ( L S ) . The identities
r ϕ ′ = − ϕ r 2 / S ~ 2 , r d d r ( r 2 / S ~ 2 ) = 2 r 2 ℓ j 2 / S ~ 4 r\phi'=-\phi r^2/\widetilde S^2,\qquad r\frac{d}{dr}(r^2/\widetilde S^2)=2r^2\ell_j^2/\widetilde S^4 r ϕ ′ = − ϕ r 2 / S 2 , r d r d ( r 2 / S 2 ) = 2 r 2 ℓ j 2 / S 4
show that ( r d d r ) i w j (r\frac{d}{dr})^i w_j ( r d r d ) i w j is a sum of terms χ ( p ) ( ϕ ) ϕ p \chi^{(p)}(\phi)\phi^p χ ( p ) ( ϕ ) ϕ p multiplied by bounded factors, with p ≥ 1 p \ge1 p ≥ 1 . Since ϕ ∈ [ 1 , 2 ] \phi\in[1,2] ϕ ∈ [ 1 , 2 ] on supp χ ( p ) \operatorname{supp}\chi^{(p)} supp χ ( p ) , these terms are bounded as well. The Leibniz rule now proves the required estimate for H \mathcal{H} H . Consequently,
∣ ∇ k F ∣ ≤ C S β r − k ≤ C 2 k S β m ♯ − k , |\nabla^k F| \le CS^\beta r^{-k} \le C2^kS^\beta m_\sharp^{-k}, ∣ ∇ k F ∣ ≤ C S β r − k ≤ C 2 k S β m ♯ − k ,
∣ ∇ F ∣ = H / r ≤ C S β r − 1 ≤ 4 C S β r m ♯ − 2 . |\nabla F| = \mathcal{H}/r \le CS^\beta r^{-1} \le4CS^\beta r m_\sharp^{-2}. ∣∇ F ∣ = H / r ≤ C S β r − 1 ≤ 4 C S β r m ♯ − 2 .
Case r < r ♯ / 2 r < r_\sharp/2 r < r ♯ /2 . Here m ♯ = r ♯ m_\sharp= r_\sharp m ♯ = r ♯ , S = ℓ J S = \ell_J S = ℓ J , and by (b) F ( z ) = G ( ∣ z ∣ 2 ) F(z) = G(|z|^2) F ( z ) = G ( ∣ z ∣ 2 ) with G ′ = β ε β 2 r ♯ 2 ϖ j G' = \frac{\beta\varepsilon^\beta}{2r_\sharp^2}\varpi_j G ′ = 2 r ♯ 2 β ε β ϖ j .
With ϕ ˉ ( q ) = ℓ j / ( L ( q + ℓ j 2 ) 1 / 2 ) \bar{\phi}(q) = \ell_j/(L(q+\ell_j^2)^{1/2}) ϕ ˉ ( q ) = ℓ j / ( L ( q + ℓ j 2 ) 1/2 ) one has ∣ ϕ ˉ ( n ) ( q ) ∣ ≤ C n ϕ ˉ ℓ j − 2 n |\bar{\phi}^{(n)}(q)| \le C_n\bar{\phi}\ell_j^{-2n} ∣ ϕ ˉ ( n ) ( q ) ∣ ≤ C n ϕ ˉ ℓ j − 2 n for q ≥ 0 q \ge0 q ≥ 0 , and ϕ ˉ ∈ [ 1 , 2 ] \bar{\phi} \in[1,2] ϕ ˉ ∈ [ 1 , 2 ] on supp χ ( p ) \operatorname{supp}\chi^{(p)} supp χ ( p ) , so ∣ ϖ j ∣ ≤ 1 |\varpi_j| \le1 ∣ ϖ j ∣ ≤ 1 and ∣ ϖ j ( n ) ∣ ≤ C ℓ j − 2 n |\varpi_j^{(n)}| \le C\ell_j^{-2n} ∣ ϖ j ( n ) ∣ ≤ C ℓ j − 2 n for n ≥ 1 n \ge1 n ≥ 1 . Moreover ∣ ∇ k [ G ( ∣ z ∣ 2 ) ] ∣ ≤ C ∑ ⌈ k / 2 ⌉ ≤ p ≤ k ∣ G ( p ) ( ∣ z ∣ 2 ) ∣ r 2 p − k |\nabla^k[G(|z|^2)]| \le C\sum_{\lceil k/2\rceil\le p \le k}|G^{(p)}(|z|^2)|r^{2p-k} ∣ ∇ k [ G ( ∣ z ∣ 2 )] ∣ ≤ C ∑ ⌈ k /2 ⌉ ≤ p ≤ k ∣ G ( p ) ( ∣ z ∣ 2 ) ∣ r 2 p − k . Using r ≤ r ♯ ≤ ℓ J r \le r_\sharp\le\ell_J r ≤ r ♯ ≤ ℓ J and ε ≤ ℓ J = S \varepsilon\le\ell_J = S ε ≤ ℓ J = S , we obtain the following estimates for the four derivative orders:
∣ ∇ F ∣ ≤ C ε β r / r ♯ 2 ≤ C S β r m ♯ − 2 , |\nabla F| \le C\varepsilon^\beta r/r_\sharp^2 \le CS^\beta r m_\sharp^{-2}, ∣∇ F ∣ ≤ C ε β r / r ♯ 2 ≤ C S β r m ♯ − 2 ,
∣ ∇ 2 F ∣ ≤ C ε β r ♯ − 2 ( 1 + r 2 ℓ J − 2 ) ≤ C S β m ♯ − 2 , |\nabla^2F| \le C\varepsilon^\beta r_\sharp^{-2}(1+r^2\ell_J^{-2}) \le CS^\beta m_\sharp^{-2}, ∣ ∇ 2 F ∣ ≤ C ε β r ♯ − 2 ( 1 + r 2 ℓ J − 2 ) ≤ C S β m ♯ − 2 ,
∣ ∇ 3 F ∣ ≤ C ε β r ♯ − 2 ℓ J − 2 r ( 1 + r 2 ℓ J − 2 ) ≤ C ε β r ♯ − 3 , |\nabla^3F| \le C\varepsilon^\beta r_\sharp^{-2}\ell_J^{-2}r(1+r^2\ell_J^{-2}) \le C\varepsilon^\beta r_\sharp^{-3}, ∣ ∇ 3 F ∣ ≤ C ε β r ♯ − 2 ℓ J − 2 r ( 1 + r 2 ℓ J − 2 ) ≤ C ε β r ♯ − 3 ,
∣ ∇ 4 F ∣ ≤ C ε β r ♯ − 2 ℓ J − 2 ( 1 + r 2 ℓ J − 2 + r 4 ℓ J − 4 ) ≤ C ε β r ♯ − 4 . |\nabla^4F| \le C\varepsilon^\beta r_\sharp^{-2}\ell_J^{-2}(1+r^2\ell_J^{-2}+r^4\ell_J^{-4}) \le C\varepsilon^\beta r_\sharp^{-4}. ∣ ∇ 4 F ∣ ≤ C ε β r ♯ − 2 ℓ J − 2 ( 1 + r 2 ℓ J − 2 + r 4 ℓ J − 4 ) ≤ C ε β r ♯ − 4 .
It remains to check the support restriction in (3.14). If ℓ j > 3 L S \ell_j > 3LS ℓ j > 3 L S , then ℓ j / ( L S r ) > 2 \ell_j/(LSr) > 2 ℓ j / ( L S r ) > 2 on a neighbourhood of r r r , so w j w_j w j and all derivatives of Φ j ♯ \Phi_j^\sharp Φ j ♯ vanish there, as in (32) .
For (e), the radial Laplacian gives, when r > 0 r > 0 r > 0 , Δ Φ ♯ = ( H ♯ ) ′ / r + ( d − 2 ) H ♯ / r 2 ≥ 0 \Delta\Phi^\sharp= (\mathcal{H}^\sharp)'/r+(d-2)\mathcal{H}^\sharp/r^2 \ge0 Δ Φ ♯ = ( H ♯ ) ′ / r + ( d − 2 ) H ♯ / r 2 ≥ 0 by (a) and d ≥ 2 d \ge2 d ≥ 2 . For ∣ z ∣ ≤ r ♯ / 2 |z| \le r_\sharp/2 ∣ z ∣ ≤ r ♯ /2 , (b) gives Δ Φ ♯ = d β ε β / r ♯ 2 = d κ / ( 4 μ ) \Delta\Phi^\sharp= d\beta\varepsilon^\beta/r_\sharp^2 = d\kappa/(4\mu) Δ Φ ♯ = d β ε β / r ♯ 2 = d κ / ( 4 μ ) by (38) ; by continuity Δ Φ ♯ ≥ 0 \Delta\Phi^\sharp\ge0 Δ Φ ♯ ≥ 0 at 0.
Finally, we prove (f). Nonnegativity follows from the same argument as in Lemma 3.1 : that proof uses only w j H ≥ 0 w_j\mathcal{H} \ge0 w j H ≥ 0 and ( w j H ) ′ ≥ 0 (w_j\mathcal{H})' \ge0 ( w j H ) ′ ≥ 0 , which hold for w j H ♯ w_j\mathcal{H}^\sharp w j H ♯ by (a); at r = 0 r=0 r = 0 , D ( 0 ) = 0 D(0)=0 D ( 0 ) = 0 . On ( r ♯ , 1 / 4 ] (r_\sharp,1/4] ( r ♯ , 1/4 ] , Φ j ♯ = Φ j \Phi_j^\sharp=\Phi_j Φ j ♯ = Φ j on an open interval around r r r , so the Hessians agree. For r ∈ [ r ♯ / 4 , 2 ε ) r\in[r_\sharp/4,2\varepsilon) r ∈ [ r ♯ /4 , 2 ε ) : w J ≡ 1 w_J\equiv1 w J ≡ 1 (proof of Lemma 3.1 ), r / ℓ J < 2 ρ < v ∗ r/\ell_J<2\rho<v_* r / ℓ J < 2 ρ < v ∗ , and by (36) , ( H ♯ ) ′ ≥ 0 (\mathcal{H}^\sharp)' \ge0 ( H ♯ ) ′ ≥ 0 , ( d − 1 ) D N − D L ≥ ( d − 1 ) ( D N − D L ) (d-1)D_N-D_L\ge(d-1)(D_N-D_L) ( d − 1 ) D N − D L ≥ ( d − 1 ) ( D N − D L ) and (22) ,
τ J Γ J ♯ ( r ) ≥ σ 2 τ J A J 2 2 ( d − 1 ) c ∗ ℓ J − 2 H ♯ ( r ) ≥ σ 2 ( d − 1 ) c ∗ 4 κ ℓ J − β β ε β 16 = σ 2 ( d − 1 ) c ∗ β ρ β 64 κ , \tau_J\Gamma_J^\sharp(r)\ge\frac{\sigma^2\tau_JA_J^2}{2}(d-1)c_*\ell_J^{-2}\mathcal{H}^\sharp(r)\ge\frac{\sigma^2(d-1)c_*}{4}\kappa\ell_J^{-\beta}\frac{\beta\varepsilon^\beta}{16}=\frac{\sigma^2(d-1)c_*\beta\rho^\beta}{64}\kappa, τ J Γ J ♯ ( r ) ≥ 2 σ 2 τ J A J 2 ( d − 1 ) c ∗ ℓ J − 2 H ♯ ( r ) ≥ 4 σ 2 ( d − 1 ) c ∗ κ ℓ J − β 16 β ε β = 64 σ 2 ( d − 1 ) c ∗ β ρ β κ ,
using (15) , A J 2 ℓ J − 2 = ℓ J − 2 β A_J^2\ell_J^{-2}=\ell_J^{-2\beta} A J 2 ℓ J − 2 = ℓ J − 2 β and H ♯ ( r ) ≥ Υ ( r 4 ) H ( r ) ≥ β ε β / 16 \mathcal{H}^\sharp(r)\ge\Upsilon(\frac{r}{4})\mathcal{H}(r)\ge\beta\varepsilon^\beta/16 H ♯ ( r ) ≥ Υ ( 4 r ) H ( r ) ≥ β ε β /16 . For r ∈ [ 2 ε , 3 ℓ J / 2 ] r\in[2\varepsilon,3\ell_J/2] r ∈ [ 2 ε , 3 ℓ J /2 ] , Γ J ♯ = Γ J \Gamma_J^\sharp=\Gamma_J Γ J ♯ = Γ J and (37) gives τ J Γ J ( r ) = ( τ J ℓ J − β ) K ( r / ℓ J ) ≥ κ 2 min [ 2 ρ , 3 / 2 ] K \tau_J\Gamma_J(r)=(\tau_J\ell_J^{-\beta})K(r/\ell_J)\ge\frac{\kappa}{2}\min_{[2\rho,3/2]}K τ J Γ J ( r ) = ( τ J ℓ J − β ) K ( r / ℓ J ) ≥ 2 κ min [ 2 ρ , 3/2 ] K . Take c 2 c_2 c 2 the smaller of the two constants. The last claim follows from Γ j ♯ = Γ j \Gamma_j^\sharp=\Gamma_j Γ j ♯ = Γ j for r ≥ 2 ε > r ♯ r\ge2\varepsilon>r_\sharp r ≥ 2 ε > r ♯ and Lemma 3.1 .
Commutator estimates for the scale operators
The estimates of this section are deterministic: they concern one fixed configuration of particles, centers, and coefficients, and involve no averaging and no time evolution. They bound the derivative of the scale-b b b field along the flow of scale a a a , as seen by two particles at distance r r r . Lemma 4.3 shows that each such derivative costs one full gradient of the coarser of the two scales and never a gradient of the finer one; for a < b a<b a < b this is a commutator estimate, since the coarse flow moves the fine wavelet and the observation point together and only the increment of u a u_a u a across the support of the wavelet, of size λ a ℓ b \lambda_a\ell_b λ a ℓ b , survives. We first isolate the class of configurations to which the estimates apply; Lemma 2.2 shows that the process never leaves this class (Remark 4.5 ).
Definition 4.1 (Admissible configurations, frozen fields). Fix d ≥ 2 d\ge2 d ≥ 2 , α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , ψ \psi ψ , R R R , M M M , and j 0 ≤ J j_0\le J j 0 ≤ J as in § 1.2 . A configuration is a triple S = ( x , x ~ , Θ ) S=(x,\widetilde{x},\Theta) S = ( x , x , Θ ) , where x , x ~ ∈ T d x,\widetilde{x}\in\mathbb{T}^d x , x ∈ T d and
Θ = ( ( P i j ) i ∈ I j , ( η i j , ν ) i ∈ I j , ν ∈ A ) j 0 ≤ j ≤ J , P i j ∈ T d , η i j , ν ∈ R . \Theta=\left((P_i^j)_{i\in I_j},(\eta_i^{j,\nu})_{i\in I_j,\ \nu\in\mathcal{A}}\right)_{j_0\le j\le J},\qquad P_i^j\in\mathbb{T}^d,\quad\eta_i^{j,\nu}\in\mathbb{R}. Θ = ( ( P i j ) i ∈ I j , ( η i j , ν ) i ∈ I j , ν ∈ A ) j 0 ≤ j ≤ J , P i j ∈ T d , η i j , ν ∈ R . The configuration is admissible if
(C1) ∣ η i j , ν ∣ ≤ M |\eta_i^{j,\nu}|\le M ∣ η i j , ν ∣ ≤ M for all j 0 ≤ j ≤ J j_0\le j\le J j 0 ≤ j ≤ J , i ∈ I j i\in I_j i ∈ I j , ν ∈ A \nu\in\mathcal{A} ν ∈ A ;
(C2) dist ( P i j , P i ′ j ) ≥ ℓ j / 2 \operatorname{dist}(P_i^j,P_{i'}^j)\ge\ell_j/2 dist ( P i j , P i ′ j ) ≥ ℓ j /2 for all j 0 ≤ j ≤ J j_0\le j\le J j 0 ≤ j ≤ J and all i ≠ i ′ i\ne i' i = i ′ in I j I_j I j ;
(C3) 0 < r : = dist ( x , x ~ ) ≤ 1 / 4 0<r:=\operatorname{dist}(x,\widetilde{x})\le1/4 0 < r := dist ( x , x ) ≤ 1/4 .
For a configuration S S S and j 0 ≤ j ≤ J j_0\le j\le J j 0 ≤ j ≤ J , the frozen field of generation j j j is
u j ( y ) = u j ( y ; Θ ) = A j ∑ i ∈ I j ∑ ν ∈ A η i j , ν W j ν ( y − P i j ) , y ∈ T d . (40) u_j(y)=u_j(y;\Theta)=A_j\sum_{i\in I_j}\sum_{\nu\in\mathcal{A}}\eta_i^{j,\nu}W_j^\nu(y-P_i^j),\qquad y\in\mathbb{T}^d. \tag*{(40)} u j ( y ) = u j ( y ; Θ ) = A j i ∈ I j ∑ ν ∈ A ∑ η i j , ν W j ν ( y − P i j ) , y ∈ T d . (40)
Since W ν ∈ C c 3 ( R d ; R d ) W^\nu\in C_c^3(\mathbb{R}^d;\mathbb{R}^d) W ν ∈ C c 3 ( R d ; R d ) , each u j u_j u j is a C 3 C^3 C 3 vector field on T d \mathbb{T}^d T d , and it depends on Θ \Theta Θ only through the generation-j j j data ( P i j ) i (P_i^j)_i ( P i j ) i , ( η i j , ν ) i , ν (\eta_i^{j,\nu})_{i,\nu} ( η i j , ν ) i , ν , on which it depends in a C 3 C^3 C 3 manner as well. By (C2), the packing argument in the proof of Lemma 2.2 applies to the centers ( P i j ) i (P_i^j)_i ( P i j ) i of every admissible configuration: at most N d , R N_{d,R} N d , R of them lie within distance R ℓ j R\ell_j R ℓ j of any given point. Together with (C1) and ∥ D q R ℓ j W ν ∥ ∞ < ∞ \|D^qR_\ell^jW^\nu\|_\infty<\infty ∥ D q R ℓ j W ν ∥ ∞ < ∞ for q ≤ 3 q\le3 q ≤ 3 , this gives the bound (16) with P i , k j ( t ) P_{i,k}^j(t) P i , k j ( t ) replaced by P i j P_i^j P i j , and in particular
∥ D q u j ∥ L ∞ ( T d ) ≤ C ∗ A j ℓ j − q , 0 ≤ q ≤ 3 , (41) \|D^q u_j\|_{L^\infty(\mathbb{T}^d)}\le C_*A_j\ell_j^{-q},\qquad0\le q\le3, \tag*{(41)} ∥ D q u j ∥ L ∞ ( T d ) ≤ C ∗ A j ℓ j − q , 0 ≤ q ≤ 3 , (41)
with C ∗ C_* C ∗ depending only on d , α , ψ , R , M d,\alpha,\psi,R,M d , α , ψ , R , M .
Definition 4.2 (Scale operators and two-point differences). Let E = T d × T d × T d × ∏ j = j 0 J ( T d ) I j \mathcal{E}=\mathbb{T}^d\times\mathbb{T}^d\times\mathbb{T}^d\times\prod_{j=j_0}^J(\mathbb{T}^d)^{I_j} E = T d × T d × T d × ∏ j = j 0 J ( T d ) I j be the space of tuples ( y ; x , x ~ ; P ) (y;x,\widetilde{x};P) ( y ; x , x ; P ) , P = ( P i j ) j , i P=(P_i^j)_{j,i} P = ( P i j ) j , i , consisting of a query point y y y together with the positional data of a configuration; the coefficients ( η i j , ν ) (\eta_i^{j,\nu}) ( η i j , ν ) are held fixed and regarded as parameters. For j 0 ≤ a ≤ J j_0\le a\le J j 0 ≤ a ≤ J define the C 3 C^3 C 3 vector field V a V_a V a on E \mathcal{E} E by
V a ( y ; x , x ~ ; P ) = ( u a ( y ) ; u a ( x ) , u a ( x ~ ) ; ( 1 { m > a } u a ( P i m ) ) m , i ) , V_a(y;x,\widetilde{x};P)=\left(u_a(y);u_a(x),u_a(\widetilde{x});\left(\mathbf{1}_{\{m>a\}}u_a(P_i^m)\right)_{m,i}\right), V a ( y ; x , x ; P ) = ( u a ( y ) ; u a ( x ) , u a ( x ) ; ( 1 { m > a } u a ( P i m ) ) m , i ) , and the first-order differential operator B a : = V a ⋅ ∇ B_a:=V_a\cdot\nabla B a := V a ⋅ ∇ acting on C 1 C^1 C 1 functions F F F on E \mathcal{E} E (componentwise on vector- or tensor-valued F F F ):
B a F = u a ( y ) ⋅ ∇ y F + u a ( x ) ⋅ ∇ x F + u a ( x ~ ) ⋅ ∇ x ~ F + ∑ m > a ∑ i ∈ I m u a ( P i m ) ⋅ ∇ P i m F , B : = ∑ a = j 0 J B a . (42) B_aF=u_a(y)\cdot\nabla_yF+u_a(x)\cdot\nabla_xF+u_a(\widetilde{x})\cdot\nabla_{\widetilde{x}}F
+\sum_{m>a}\sum_{i\in I_m}u_a(P_i^m)\cdot\nabla_{P_i^m}F,\qquad B:=\sum_{a=j_0}^J B_a. \tag*{(42)} B a F = u a ( y ) ⋅ ∇ y F + u a ( x ) ⋅ ∇ x F + u a ( x ) ⋅ ∇ x F + m > a ∑ i ∈ I m ∑ u a ( P i m ) ⋅ ∇ P i m F , B := a = j 0 ∑ J B a . (42)
Thus B a \mathcal{B}_a B a transports the query point and the two particles with the generation-a a a field and transports the centers of every finer generation m > a m>a m > a with the same field; it does not move the centers of generation a a a or of any coarser generation. On functions that do not depend on the query point, B \mathcal{B} B is the infinitesimal generator of the deterministic dynamics (1.5)–(1.8) between two consecutive renewal times, with the active coefficients and shifts frozen.
Each frozen field u b u_b u b can be regarded as the function ( y ; x , x ~ ; P ) ↦ u b ( y ; Θ ) (y; x,\tilde{x}; P) \mapsto u_b(y; \Theta) ( y ; x , x ~ ; P ) ↦ u b ( y ; Θ ) on E \mathcal{E} E , which is C 3 C^3 C 3 . Consequently B a u b ∈ C 2 ( E ; R d ) \mathcal{B}_a u_b \in C^2(\mathcal{E};\mathbb{R}^d) B a u b ∈ C 2 ( E ; R d ) and B c B a u b ∈ C 1 ( E ; R d ) \mathcal{B}_c\mathcal{B}_a u_b \in C^1(\mathcal{E};\mathbb{R}^d) B c B a u b ∈ C 1 ( E ; R d ) are well defined for all a , b , c ∈ { j 0 , … , J } a,b,c \in\{j_0,\ldots,J\} a , b , c ∈ { j 0 , … , J } , and B a \mathcal{B}_a B a obeys the product rule B a ( F G ) = ( B a F ) G + F B a G \mathcal{B}_a(FG)=(\mathcal{B}_aF)G+F\mathcal{B}_aG B a ( F G ) = ( B a F ) G + F B a G .
For a continuous function G G G on E \mathcal{E} E and a configuration S = ( x , x ~ , Θ ) S=(x,\tilde{x},\Theta) S = ( x , x ~ , Θ ) , the two-point difference of G G G at S S S is
δ G : = G ( x ; x , x ~ ; P ) − G ( x ~ ; x , x ~ ; P ) , (43) \delta G := G(x;x,\tilde{x};P)-G(\tilde{x};x,\tilde{x};P), \tag*{(43)} δ G := G ( x ; x , x ~ ; P ) − G ( x ~ ; x , x ~ ; P ) , (43)
that is, the query point is evaluated at the two particles while all other coordinates are held fixed. In particular δ u b = u b ( x ) − u b ( x ~ ) \delta u_b=u_b(x)-u_b(\tilde{x}) δ u b = u b ( x ) − u b ( x ~ ) , and δ ( B a u b ) \delta(\mathcal{B}_a u_b) δ ( B a u b ) , δ ( B c B a u b ) \delta(\mathcal{B}_c\mathcal{B}_a u_b) δ ( B c B a u b ) are the two-point differences of the functions B a u b \mathcal{B}_a u_b B a u b and B c B a u b \mathcal{B}_c\mathcal{B}_a u_b B c B a u b on E \mathcal{E} E .
For an admissible configuration with separation r = dist ( x , x ~ ) r=\operatorname{dist}(x,\tilde{x}) r = dist ( x , x ~ ) set
γ j = A j r min ( 1 , r / ℓ j ) = min ( λ j , A j / r ) , K a b = λ min ( a , b ) γ max ( a , b ) , (44) \gamma_j=\frac{A_j}{r}\min(1,r/\ell_j)=\min(\lambda_j,A_j/r),\qquad K_{ab}=\lambda_{\min(a,b)}\gamma_{\max(a,b)}, \tag*{(44)} γ j = r A j min ( 1 , r / ℓ j ) = min ( λ j , A j / r ) , K ab = λ m i n ( a , b ) γ m a x ( a , b ) , (44)
where j 0 ≤ j , a , b ≤ J j_0\leq j,a,b\leq J j 0 ≤ j , a , b ≤ J . Evidently, K a b K_{ab} K ab is one full gradient of the coarser of the two scales times the increment factor γ \gamma γ of the finer one. Writing the full expressions, we have
γ j = { λ j = 2 β j = 2 ( 1 − α ) j , r ≤ 2 − j , A j r = 2 − α j r = 2 ( β − 1 ) j r , r > 2 − j , \gamma_j=
\begin{cases}
\lambda_j=2^{\beta j}=2^{(1-\alpha)j}, & r\leq2^{-j},\\
\dfrac{A_j}{r}=\dfrac{2^{-\alpha j}}{r}=\dfrac{2^{(\beta-1)j}}{r}, & r>2^{-j},
\end{cases} γ j = ⎩ ⎨ ⎧ λ j = 2 β j = 2 ( 1 − α ) j , r A j = r 2 − α j = r 2 ( β − 1 ) j , r ≤ 2 − j , r > 2 − j ,
and
K a b = { λ a λ b = 2 β ( a + b ) = 2 ( 1 − α ) ( a + b ) , a < b , r ≤ 2 − b , λ a A b r = 2 β a − α b r = 2 ( 1 − α ) a − α b r , a < b , r > 2 − b , λ b λ a = 2 β ( a + b ) = 2 ( 1 − α ) ( a + b ) , a ≥ b , r ≤ 2 − a , λ b A a r = 2 β b − α a r = 2 ( 1 − α ) b − α a r , a ≥ b , r > 2 − a . K_{ab}=
\begin{cases}
\lambda_a\lambda_b=2^{\beta(a+b)}=2^{(1-\alpha)(a+b)}, & a<b,\quad r\leq2^{-b},\\
\dfrac{\lambda_a A_b}{r}=\dfrac{2^{\beta a-\alpha b}}{r}=\dfrac{2^{(1-\alpha)a-\alpha b}}{r}, & a<b,\quad r>2^{-b},\\
\lambda_b\lambda_a=2^{\beta(a+b)}=2^{(1-\alpha)(a+b)}, & a\geq b,\quad r\leq2^{-a},\\
\dfrac{\lambda_b A_a}{r}=\dfrac{2^{\beta b-\alpha a}}{r}=\dfrac{2^{(1-\alpha)b-\alpha a}}{r}, & a\geq b,\quad r>2^{-a}.
\end{cases} K ab = ⎩ ⎨ ⎧ λ a λ b = 2 β ( a + b ) = 2 ( 1 − α ) ( a + b ) , r λ a A b = r 2 β a − α b = r 2 ( 1 − α ) a − α b , λ b λ a = 2 β ( a + b ) = 2 ( 1 − α ) ( a + b ) , r λ b A a = r 2 β b − α a = r 2 ( 1 − α ) b − α a , a < b , r ≤ 2 − b , a < b , r > 2 − b , a ≥ b , r ≤ 2 − a , a ≥ b , r > 2 − a .
Lemma 4.3 (Two material derivatives of a wavelet field). Let S = ( x , x ~ , Θ ) S=(x,\tilde{x},\Theta) S = ( x , x ~ , Θ ) be an admissible configuration in the sense of Definition 4.1, with r = dist ( x , x ~ ) > 0 r=\operatorname{dist}(x,\tilde{x})>0 r = dist ( x , x ~ ) > 0 . Let u j u_j u j be the frozen fields (4.1), let B a \mathcal{B}_a B a and δ \delta δ be as in Definition 4.2, and define γ j , K a b \gamma_j,K_{ab} γ j , K ab by (4.5) with this r r r . There is a constant C > 0 C>0 C > 0 , depending only on d , α , ψ , R , M d,\alpha,\psi,R,M d , α , ψ , R , M , such that, for all a , b , c ∈ { j 0 , … , J } a,b,c\in\{j_0,\ldots,J\} a , b , c ∈ { j 0 , … , J } ,
∣ δ u b ∣ ≤ C r γ b , (45) |\delta u_b|\leq C r\gamma_b, \tag*{(45)} ∣ δ u b ∣ ≤ C r γ b , (45) ∣ δ ( B a u b ) ∣ ≤ C r K a b , (46) |\delta(\mathcal{B}_a u_b)|\leq C rK_{ab}, \tag*{(46)} ∣ δ ( B a u b ) ∣ ≤ C r K ab , (46) ∣ δ ( B c B a u b ) ∣ ≤ C r λ min ( a , b ) K c , max ( a , b ) . (47) |\delta(\mathcal{B}_c\mathcal{B}_a u_b)|\leq C r\lambda_{\min(a,b)}K_{c,\max(a,b)}. \tag*{(47)} ∣ δ ( B c B a u b ) ∣ ≤ C r λ m i n ( a , b ) K c , m a x ( a , b ) . (47)
The operators act on the enlarged configuration space before the query point is evaluated at x x x and x ~ \widetilde{x} x to take δ \delta δ . The constant is independent of J J J , κ \kappa κ , L L L , the configuration, and the indices a , b , c a,b,c a , b , c .
Remark 4.4 (Meaning of the iterated derivatives). All coefficients are held fixed. In (46) , we regard u b u_b u b as the function
( y ; x , x ~ ; P ) ⟼ u b ( y ; P b ) = A b ∑ i , ν η i b , ν W ν b ( y − P i b ) , P b = ( P i b ) i . (y;x,\widetilde{x};P)\longmapsto u_b(y;P^b)=A_b\sum_{i,\nu}\eta_i^{b,\nu}W_\nu^b(y-P_i^b),\qquad P^b=(P_i^b)_i. ( y ; x , x ; P ) ⟼ u b ( y ; P b ) = A b i , ν ∑ η i b , ν W ν b ( y − P i b ) , P b = ( P i b ) i .
It depends on y y y and P b P^b P b , but not on the separate particle coordinates x , x ~ x,\widetilde{x} x , x . The function F a b : = B a u b F_{ab}:=\mathcal{B}_a u_b F ab := B a u b is then defined on the same space. As shown in (4.11),
F a b ( y ; P ) = { A b ∑ i , ν η i b , ν [ u a ( y ; P a ) − u a ( P i b ; P a ) ] ⋅ ∇ W ν b ( y − P i b ) , a < b , ( u a ( y ; P a ) ⋅ ∇ y ) u b ( y ; P b ) , a ≥ b . (48) F_{ab}(y;P)=
\begin{cases}
A_b\sum_{i,\nu}\eta_i^{b,\nu}\bigl[u_a(y;P^a)-u_a(P_i^b;P^a)\bigr]\cdot\nabla W_\nu^b(y-P_i^b), & a<b,\\
\bigl(u_a(y;P^a)\cdot\nabla_y\bigr)u_b(y;P^b), & a\ge b.
\end{cases} \tag*{(48)} F ab ( y ; P ) = { A b ∑ i , ν η i b , ν [ u a ( y ; P a ) − u a ( P i b ; P a ) ] ⋅ ∇ W ν b ( y − P i b ) , ( u a ( y ; P a ) ⋅ ∇ y ) u b ( y ; P b ) , a < b , a ≥ b . (48)
Thus F a b F_{ab} F ab depends only on y y y and the centers of generations a , b a,b a , b , and remains independent of x , x ~ x,\widetilde{x} x , x . For a ≠ b a\ne b a = b ,
B c F a b = ( u c ( y ) ⋅ ∇ y ) F a b + 1 { c < a } ∑ k ( u c ( P k a ) ⋅ ∇ P k a ) F a b + 1 { c < b } ∑ i ( u c ( P i b ) ⋅ ∇ P i b ) F a b . \mathcal{B}_cF_{ab}=(u_c(y)\cdot\nabla_y)F_{ab}
+\mathbf{1}_{\{c<a\}}\sum_k\bigl(u_c(P_k^a)\cdot\nabla_{P_k^a}\bigr)F_{ab}
+\mathbf{1}_{\{c<b\}}\sum_i\bigl(u_c(P_i^b)\cdot\nabla_{P_i^b}\bigr)F_{ab}. B c F ab = ( u c ( y ) ⋅ ∇ y ) F ab + 1 { c < a } k ∑ ( u c ( P k a ) ⋅ ∇ P k a ) F ab + 1 { c < b } i ∑ ( u c ( P i b ) ⋅ ∇ P i b ) F ab .
If a = b a=b a = b , there is only one family of centers, and its contribution is included once. Each derivative acts on every occurrence of its coordinate: for example, when a < b a<b a < b , the P i b P_i^b P i b derivative acts both on u a ( P i b ; P a ) u_a(P_i^b;P^a) u a ( P i b ; P a ) and on the translated wavelet. Only after forming B c F a b \mathcal{B}_cF_{ab} B c F ab do we take its two-point difference, by evaluating y y y at x x x and x ~ \widetilde{x} x with all other coordinates fixed.
The two cases in (4.11) also explain the scale factor K a b K_{ab} K ab . When a < b a<b a < b , the coarse field moves the query point and the fine center together, leaving an increment across a distance of order ℓ b \ell_b ℓ b . When a ≥ b a\ge b a ≥ b , the spatial derivative falls directly on the coarser field u b u_b u b . The proof below makes these two mechanisms explicit at the next derivative as well.
Proof of Lemma 4.3. Throughout the proof the admissible configuration S S S is fixed. For a function G G G on E E E , ∥ G ∥ ∞ \|G\|_\infty ∥ G ∥ ∞ denotes the supremum over the query point y ∈ T d y\in\mathbb{T}^d y ∈ T d with the positional data of S S S fixed, D q G D^qG D q G denotes derivatives in y y y , and the Lipschitz constant of G G G refers to the variable y y y . Since dist ( x , x ~ ) = r ≤ 1 / 4 \operatorname{dist}(x,\widetilde{x})=r\le1/4 dist ( x , x ) = r ≤ 1/4 , the two points are joined by a minimizing geodesic of length r r r , and therefore every such G G G satisfies
∣ δ G ∣ ≤ min ( 2 ∥ G ∥ ∞ , r ∥ ∇ y G ∥ ∞ ) . (49) |\delta G|\le\min\left(2\|G\|_\infty,r\|\nabla_yG\|_\infty\right). \tag*{(49)} ∣ δ G ∣ ≤ min ( 2∥ G ∥ ∞ , r ∥ ∇ y G ∥ ∞ ) . (49)
Applied to D q u j D^q u_j D q u j and combined with (41) for q q q and q + 1 q+1 q + 1 , this gives, for 0 ≤ q ≤ 2 0\le q\le2 0 ≤ q ≤ 2 ,
∣ δ D q u j ∣ ≤ C A j ℓ j − q min ( 1 , r / ℓ j ) . (50) |\delta D^q u_j|\le CA_j\ell_j^{-q}\min(1,r/\ell_j). \tag*{(50)} ∣ δ D q u j ∣ ≤ C A j ℓ j − q min ( 1 , r / ℓ j ) . (50)
Equation (45) is then contained in (50) .
Products are estimated using δ ( F G ) = ( δ F ) G ( x ) + F ( x ~ ) δ G \delta(FG)=(\delta F)G(x)+F(\widetilde{x})\delta G δ ( F G ) = ( δ F ) G ( x ) + F ( x ) δ G .
Write F a b ( y ) = B a u b ( y ) F_{ab}(y)=\mathcal{B}_a u_b(y) F ab ( y ) = B a u b ( y ) . As a function on the enlarged configuration space, u b ( y ) u_b(y) u b ( y ) depends only on the query point y y y and the scale-b b b centers P i b P_i^b P i b ; it does not depend on the particle coordinates x , x ~ x,\widetilde{x} x , x or the other centers. Moreover, the field u b u_b u b depends on the positional data only through the centers P i b P_i^b P i b , and by (42) the operator B a \mathcal{B}_a B a contains the derivatives ∇ P i b \nabla_{P_i^b} ∇ P i b if and only if a < b a < b a < b . Since ∇ P i b [ W b ν ( y − P i b ) ] = − ∇ W b ν ( y − P i b ) \nabla_{P_i^b}[W_b^\nu(y-P_i^b)]=-\nabla W_b^\nu(y-P_i^b) ∇ P i b [ W b ν ( y − P i b )] = − ∇ W b ν ( y − P i b ) , (42) gives
F a b ( y ) = { A b ∑ i , ν η i b , ν [ u a ( y ) − u a ( P i b ) ] ⋅ ∇ W b ν ( y − P i b ) , a < b , ( u a ( y ) ⋅ ∇ ) u b ( y ) , a ≥ b . (51) F_{ab}(y)=
\begin{cases}
A_b\sum_{i,\nu}\eta_i^{b,\nu}[u_a(y)-u_a(P_i^b)]\cdot\nabla W_b^\nu(y-P_i^b), & a<b,\\
(u_a(y)\cdot\nabla)u_b(y), & a\ge b.
\end{cases} \tag*{(51)} F ab ( y ) = { A b ∑ i , ν η i b , ν [ u a ( y ) − u a ( P i b )] ⋅ ∇ W b ν ( y − P i b ) , ( u a ( y ) ⋅ ∇ ) u b ( y ) , a < b , a ≥ b . (51)
Start with a < b a < b a < b . Write v i ( y ) = u a ( y ) − u a ( P i b ) v_i(y)=u_a(y)-u_a(P_i^b) v i ( y ) = u a ( y ) − u a ( P i b ) . On the support of each wavelet,
∣ v i ∣ ≤ C λ a ℓ b , ∣ D v i ∣ ≤ C λ a , ∣ D 2 v i ∣ ≤ C λ a / ℓ a ≤ C λ a / ℓ b , |v_i|\le C\lambda_a\ell_b,\qquad|Dv_i|\le C\lambda_a,\qquad|D^2v_i|\le C\lambda_a/\ell_a\le C\lambda_a/\ell_b, ∣ v i ∣ ≤ C λ a ℓ b , ∣ D v i ∣ ≤ C λ a , ∣ D 2 v i ∣ ≤ C λ a / ℓ a ≤ C λ a / ℓ b ,
where the last inequality uses ℓ a ≥ ℓ b \ell_a\ge\ell_b ℓ a ≥ ℓ b .
By the product rule, a term in D q ( v i ⋅ ∇ W b ν ) D^q(v_i\cdot\nabla W_b^\nu) D q ( v i ⋅ ∇ W b ν ) with m m m derivatives on v i v_i v i is bounded by
C λ a ℓ b 1 − m ℓ b − ( q − m + 1 ) = C λ a ℓ b − q , 0 ≤ m ≤ q ≤ 2. C\lambda_a\ell_b^{1-m}\ell_b^{-(q-m+1)}=C\lambda_a\ell_b^{-q},\qquad0\le m\le q\le2. C λ a ℓ b 1 − m ℓ b − ( q − m + 1 ) = C λ a ℓ b − q , 0 ≤ m ≤ q ≤ 2.
Multiplying by A b A_b A b , using the coefficient bound, and summing over the uniformly bounded number of wavelets meeting each point proves
∥ D q F a b ∥ ∞ ≤ C A b λ a ℓ b − q , q = 0 , 1 , 2 , a < b . (52) \|D^qF_{ab}\|_\infty\le CA_b\lambda_a\ell_b^{-q},\qquad q=0,1,2,\qquad a<b. \tag*{(52)} ∥ D q F ab ∥ ∞ ≤ C A b λ a ℓ b − q , q = 0 , 1 , 2 , a < b . (52)
The cases q = 0 q=0 q = 0 and q = 1 q=1 q = 1 of (52) , together with (49) , give
∣ δ F a b ∣ ≤ min ( 2 ∥ F a b ∥ ∞ , r ∥ ∇ F a b ∥ ∞ ) ≤ C A b λ a min ( 1 , r / ℓ b ) = C r λ a γ b = C r K a b , a < b . \begin{aligned}
|\delta F_{ab}|\le\min(2\|F_{ab}\|_\infty,r\|\nabla F_{ab}\|_\infty) \\
&\le CA_b\lambda_a\min(1,r/\ell_b) \\
&=Cr\lambda_a\gamma_b=CrK_{ab},\qquad a<b.
\end{aligned} ∣ δ F ab ∣ ≤ min ( 2∥ F ab ∥ ∞ , r ∥∇ F ab ∥ ∞ ) ≤ C A b λ a min ( 1 , r / ℓ b ) = C r λ a γ b = C r K ab , a < b .
The last equality uses K a b = λ a γ b K_{ab}=\lambda_a\gamma_b K ab = λ a γ b when a < b a<b a < b . This proves (46) in this case.
For a ≥ b a\ge b a ≥ b , the second line of (4.11) gives
F a b = ( u a ⋅ ∇ ) u b . F_{ab}=(u_a\cdot\nabla)u_b. F ab = ( u a ⋅ ∇ ) u b .
The product rule for differences and (50) yield
∣ δ F a b ∣ ≤ ∣ δ u a ∣ ∥ ∇ u b ∥ ∞ + ∥ u a ∥ ∞ ∣ δ ∇ u b ∣ ≤ C A a λ b [ min ( 1 , r / ℓ a ) + min ( 1 , r / ℓ b ) ] ≤ C A a λ b min ( 1 , r / ℓ a ) = C r λ b γ a = C r K a b , a ≥ b , \begin{aligned}
|\delta F_{ab}|\le|\delta u_a|\|\nabla u_b\|_\infty+\|u_a\|_\infty|\delta\nabla u_b| \\
&\le CA_a\lambda_b[\min(1,r/\ell_a)+\min(1,r/\ell_b)] \\
&\le CA_a\lambda_b\min(1,r/\ell_a)=Cr\lambda_b\gamma_a=CrK_{ab},\qquad a\ge b,
\end{aligned} ∣ δ F ab ∣ ≤ ∣ δ u a ∣∥∇ u b ∥ ∞ + ∥ u a ∥ ∞ ∣ δ ∇ u b ∣ ≤ C A a λ b [ min ( 1 , r / ℓ a ) + min ( 1 , r / ℓ b )] ≤ C A a λ b min ( 1 , r / ℓ a ) = C r λ b γ a = C r K ab , a ≥ b ,
and hence (46) for a ≥ b a\ge b a ≥ b . We moreover record that
∥ F a b ∥ ∞ ≤ ∥ u a ∥ ∞ ∥ ∇ u b ∥ ∞ ≤ C A a λ b , a ≥ b . \|F_{ab}\|_\infty\le\|u_a\|_\infty\|\nabla u_b\|_\infty\le CA_a\lambda_b,\qquad a\ge b. ∥ F ab ∥ ∞ ≤ ∥ u a ∥ ∞ ∥∇ u b ∥ ∞ ≤ C A a λ b , a ≥ b .
Hence,
∥ F a b ∥ ∞ ≤ C λ min ( a , b ) A max ( a , b ) . (53) \|F_{ab}\|_\infty\le C\lambda_{\min(a,b)}A_{\max(a,b)}. \tag*{(53)} ∥ F ab ∥ ∞ ≤ C λ m i n ( a , b ) A m a x ( a , b ) . (53)
We now prove (47) . For this part of the proof, write
m j = min ( 1 , r / ℓ j ) , A j m j = r γ j . m_j=\min(1,r/\ell_j),\qquad A_jm_j=r\gamma_j. m j = min ( 1 , r / ℓ j ) , A j m j = r γ j .
If j ≥ k j\ge k j ≥ k , then ℓ j ≤ ℓ k \ell_j\le\ell_k ℓ j ≤ ℓ k and m k ≤ m j m_k\le m_j m k ≤ m j . We repeatedly use (49) to pass from a supremum and a spatial gradient bound to a two-point estimate. All gradients below are in the query variable, with the centers fixed.
Case 1: a < b a<b a < b and c < b c<b c < b . Here the desired bound is C r λ a λ c γ b Cr\lambda_a\lambda_c\gamma_b C r λ a λ c γ b . By definition, B c [ u a ( y ) ] = F c a ( y ) \mathcal{B}_c[u_a(y)]=F_{ca}(y) B c [ u a ( y )] = F c a ( y ) . To differentiate u a ( P i b ) u_a(P_i^b) u a ( P i b ) , however, we must account for both the evaluation point and the centers defining u a u_a u a . This expression is independent of the query coordinate, so only center derivatives contribute. Since c < b c<b c < b , the center P i b P_i^b P i b moves with velocity u c ( P i b ) u_c(P_i^b) u c ( P i b ) ; and since a < b a<b a < b , it is distinct from the scale-a a a coordinates. The chain rule therefore gives
B c [ u a ( P i b ) ] = ( u c ( P i b ) ⋅ ∇ y ) u a ( y ) ∣ y = P i b + 1 { c < a } ∑ k ( u c ( P k a ) ⋅ ∇ P k a ) u a ( y ) ∣ y = P i b = F c a ( P i b ) . \mathcal{B}_c[u_a(P_i^b)]=(u_c(P_i^b)\cdot\nabla_y)u_a(y)\big|_{y=P_i^b}
+\mathbf{1}_{\{c<a\}}\sum_k (u_c(P_k^a)\cdot\nabla_{P_k^a})u_a(y)\big|_{y=P_i^b}
=F_{ca}(P_i^b). B c [ u a ( P i b )] = ( u c ( P i b ) ⋅ ∇ y ) u a ( y ) y = P i b + 1 { c < a } k ∑ ( u c ( P k a ) ⋅ ∇ P k a ) u a ( y ) y = P i b = F c a ( P i b ) .
In the second line, the center derivatives are taken before substituting y = P i b y=P_i^b y = P i b . Thus the spatial derivative in F c a ( P i b ) F_{ca}(P_i^b) F c a ( P i b ) is supplied by the motion of the evaluation point. Similarly, since c < b c<b c < b , B c [ ∇ W b ν ( y − P i b ) ] \mathcal{B}_c[\nabla W_b^\nu(y-P_i^b)] B c [ ∇ W b ν ( y − P i b )] carries contributions from both differentiating in y y y and differentiating in centers, and
B c [ ∇ W b ν ( y − P i b ) ] = D 2 W b ν ( y − P i b ) [ u c ( y ) − u c ( P i b ) ] , \mathcal{B}_c[\nabla W_b^\nu(y-P_i^b)]
=D^2W_b^\nu(y-P_i^b)[u_c(y)-u_c(P_i^b)], B c [ ∇ W b ν ( y − P i b )] = D 2 W b ν ( y − P i b ) [ u c ( y ) − u c ( P i b )] ,
where the Hessian is taken componentwise.
Applying the product rule in (4.11), with the coefficients held fixed, yields
B c F a b ( y ) = A b ∑ i , ν η i b , ν { [ F c a ( y ) − F c a ( P i b ) ] ⋅ ∇ W b ν ( y − P i b ) + [ u a ( y ) − u a ( P i b ) ] ⊗ [ u c ( y ) − u c ( P i b ) ] : D 2 W b ν ( y − P i b ) } . (54) \mathcal{B}_cF_{ab}(y)=A_b\sum_{i,\nu}\eta_i^{b,\nu}\left\{[F_{ca}(y)-F_{ca}(P_i^b)]\cdot\nabla W_b^\nu(y-P_i^b)
+[u_a(y)-u_a(P_i^b)]\otimes[u_c(y)-u_c(P_i^b)]:D^2W_b^\nu(y-P_i^b)\right\}. \tag*{(54)} B c F ab ( y ) = A b i , ν ∑ η i b , ν { [ F c a ( y ) − F c a ( P i b )] ⋅ ∇ W b ν ( y − P i b ) + [ u a ( y ) − u a ( P i b )] ⊗ [ u c ( y ) − u c ( P i b )] : D 2 W b ν ( y − P i b ) } . (54)
Here, for vectors v , w ∈ R d v,w\in\mathbb{R}^d v , w ∈ R d , [ v ⊗ w ] : D 2 W b ν [v\otimes w]:D^2W_b^\nu [ v ⊗ w ] : D 2 W b ν denotes the vector with components ∑ p , q v p w q ∂ p ∂ q ( W b ν ) m \sum_{p,q}v_pw_q\partial_p\partial_q(W_b^\nu)_m ∑ p , q v p w q ∂ p ∂ q ( W b ν ) m , 1 ≤ m ≤ d 1\leq m\leq d 1 ≤ m ≤ d , that is, the Hessian is contracted componentwise in the vector index of W b ν W_b^\nu W b ν . Indeed, since a < b a<b a < b and c < b c<b c < b , Remark 4.4 gives
B c F a b = ( u c ( y ) ⋅ ∇ y ) F a b + 1 { c < a } ∑ k ( u c ( P k a ) ⋅ ∇ P k a ) F a b + ∑ k ( u c ( P k b ) ⋅ ∇ P k b ) F a b . \mathcal{B}_cF_{ab}=(u_c(y)\cdot\nabla_y)F_{ab}
+\mathbf{1}_{\{c<a\}}\sum_k(u_c(P_k^a)\cdot\nabla_{P_k^a})F_{ab}
+\sum_k(u_c(P_k^b)\cdot\nabla_{P_k^b})F_{ab}. B c F ab = ( u c ( y ) ⋅ ∇ y ) F ab + 1 { c < a } k ∑ ( u c ( P k a ) ⋅ ∇ P k a ) F ab + k ∑ ( u c ( P k b ) ⋅ ∇ P k b ) F ab .
Consider one summand of F a b F_{ab} F ab , suppressing the fixed coefficient:
T i ν ( y ) = [ u a ( y ) − u a ( P i b ) ] ⋅ ∇ W b ν ( y − P i b ) . T_i^\nu(y)=[u_a(y)-u_a(P_i^b)]\cdot\nabla W_b^\nu(y-P_i^b). T i ν ( y ) = [ u a ( y ) − u a ( P i b )] ⋅ ∇ W b ν ( y − P i b ) .
Write v i = u a ( y ) − u a ( P i b ) v_i=u_a(y)-u_a(P_i^b) v i = u a ( y ) − u a ( P i b ) and s i = y − P i b s_i=y-P_i^b s i = y − P i b . The query derivative contributes
( u c ( y ) ⋅ ∇ y ) T i ν = [ ( u c ( y ) ⋅ ∇ ) u a ( y ) ] ⋅ ∇ W b ν ( s i ) + [ v i ⊗ u c ( y ) ] : D 2 W b ν ( s i ) . (u_c(y)\cdot\nabla_y)T_i^\nu=[(u_c(y)\cdot\nabla)u_a(y)]\cdot\nabla W_b^\nu(s_i)+[v_i\otimes u_c(y)]:D^2W_b^\nu(s_i). ( u c ( y ) ⋅ ∇ y ) T i ν = [( u c ( y ) ⋅ ∇ ) u a ( y )] ⋅ ∇ W b ν ( s i ) + [ v i ⊗ u c ( y )] : D 2 W b ν ( s i ) .
It does not differentiate u a ( P i b ) u_a(P_i^b) u a ( P i b ) , which is independent of the query coordinate. The scale-a a a center derivatives, when c < a c<a c < a , contribute
1 { c < a } ∑ k [ ( u c ( P k a ) ⋅ ∇ P k a ) u a ( y ) − ( u c ( P k a ) ⋅ ∇ P k a ) u a ( z ) ∣ z = P i b ] ⋅ ∇ W b ν ( s i ) . \mathbf{1}_{\{c<a\}}\sum_k\left[(u_c(P_k^a)\cdot\nabla_{P_k^a})u_a(y)-(u_c(P_k^a)\cdot\nabla_{P_k^a})u_a(z)\big|_{z=P_i^b}\right]\cdot\nabla W_b^\nu(s_i). 1 { c < a } k ∑ [ ( u c ( P k a ) ⋅ ∇ P k a ) u a ( y ) − ( u c ( P k a ) ⋅ ∇ P k a ) u a ( z ) z = P i b ] ⋅ ∇ W b ν ( s i ) .
They act on the centers defining u a u_a u a , but not on the wavelet, whose center belongs to generation b b b . In the second term, the center derivative is taken with z z z fixed, before evaluation.
The scale-b b b center derivatives contribute only through P i b P_i^{b} P i b :
( u c ( P i b ) ⋅ ∇ P i b ) T i V = − [ ( u c ( P i b ) ⋅ ∇ z ) u a ( z ) ∣ z = P i b ⋅ ∇ W b ν ( s i ) (u_c(P_i^b)\cdot\nabla_{P_i^b})T_i^V=-[(u_c(P_i^b)\cdot\nabla_z)u_a(z)|_{z=P_i^b}\cdot\nabla W_b^\nu(s_i) ( u c ( P i b ) ⋅ ∇ P i b ) T i V = − [( u c ( P i b ) ⋅ ∇ z ) u a ( z ) ∣ z = P i b ⋅ ∇ W b ν ( s i )
− [ v i ⊗ u c ( P i b ) ] : D 2 W b ν ( s i ) . -\left[v_i\otimes u_c(P_i^b)\right]:D^2W_b^\nu(s_i). − [ v i ⊗ u c ( P i b ) ] : D 2 W b ν ( s i ) .
The first term differentiates the evaluation point in u a ( P i b ) u_a(P_i^b) u a ( P i b ) ; the second differentiates the translated wavelet.
Adding these three contributions, all terms multiplying ∇ W b ν ( s i ) \nabla W_b^\nu(s_i) ∇ W b ν ( s i ) combine into
[ ( u c ( y ) ⋅ ∇ ) u a ( y ) + 1 { c < a } ∑ k ( u c ( P k a ) ⋅ ∇ P k a ) u a ( y ) − ( u c ( P i b ) ⋅ ∇ z ) u a ( z ) ∣ z = P i b − 1 { c < a } ∑ k ( u c ( P k a ) ⋅ ∇ P k a ) u a ( z ) ∣ z = P i b ] ⋅ ∇ W b ν ( s i ) = [ F c a ( y ) − F c a ( P i b ) ] ⋅ ∇ W b ν ( s i ) , \begin{aligned}
&\Bigl[(u_c(y)\cdot\nabla)u_a(y)+\mathbf{1}_{\{c<a\}}\sum_k\left(u_c(P_k^a)\cdot\nabla_{P_k^a}\right)u_a(y)\\
&\quad-\left.\left(u_c(P_i^b)\cdot\nabla_z\right)u_a(z)\right|_{z=P_i^b}
-\mathbf{1}_{\{c<a\}}\sum_k\left.\left(u_c(P_k^a)\cdot\nabla_{P_k^a}\right)u_a(z)\right|_{z=P_i^b}\Bigr]\cdot\nabla W_b^\nu(s_i) \\
&= [F_{ca}(y)-F_{ca}(P_i^b)]\cdot\nabla W_b^\nu(s_i),
\end{aligned} [ ( u c ( y ) ⋅ ∇ ) u a ( y ) + 1 { c < a } k ∑ ( u c ( P k a ) ⋅ ∇ P k a ) u a ( y ) − ( u c ( P i b ) ⋅ ∇ z ) u a ( z ) z = P i b − 1 { c < a } k ∑ ( u c ( P k a ) ⋅ ∇ P k a ) u a ( z ) z = P i b ] ⋅ ∇ W b ν ( s i ) = [ F c a ( y ) − F c a ( P i b )] ⋅ ∇ W b ν ( s i ) ,
by the definition of F c a F_{ca} F c a . The two Hessian terms combine into
[ v i ⊗ ( u c ( y ) − u c ( P i b ) ) ] : D 2 W b ν ( s i ) . [v_i\otimes(u_c(y)-u_c(P_i^b))]:D^2W_b^\nu(s_i). [ v i ⊗ ( u c ( y ) − u c ( P i b ))] : D 2 W b ν ( s i ) .
Multiplying by A b η i b , ν A_b\eta_i^{b,\nu} A b η i b , ν and summing gives (54) . Thus the three coordinate-derivative contributions in the remark are regrouped into the two product-rule terms in (54) .
We next estimate the two terms in braces in (54) and their first spatial derivatives. First,
∥ ∇ F c a ∥ ∞ ≤ C λ c λ a . \|\nabla F_{ca}\|_\infty\le C\lambda_c\lambda_a. ∥∇ F c a ∥ ∞ ≤ C λ c λ a .
For c < a c<a c < a , this is (52) with q = 1 q=1 q = 1 and indices ( c , a ) (c,a) ( c , a ) . For c ≥ a c\ge a c ≥ a , differentiate F c a = ( u c ⋅ ∇ ) u a F_{ca}=(u_c\cdot\nabla)u_a F c a = ( u c ⋅ ∇ ) u a to obtain
∥ ∇ F c a ∥ ∞ ≤ C ( λ c λ a + A c λ a / ℓ a ) ≤ C λ c λ a , \|\nabla F_{ca}\|_\infty\le C(\lambda_c\lambda_a+A_c\lambda_a/\ell_a)\le C\lambda_c\lambda_a, ∥∇ F c a ∥ ∞ ≤ C ( λ c λ a + A c λ a / ℓ a ) ≤ C λ c λ a ,
since A c = ℓ c λ c A_c=\ell_c\lambda_c A c = ℓ c λ c and ℓ c ≤ ℓ a \ell_c\le\ell_a ℓ c ≤ ℓ a . On the support of a scale-b b b wavelet, put
v a , i = u a ( y ) − u a ( P i b ) , v c , i = u c ( y ) − u c ( P i b ) , E i = F c a ( y ) − F c a ( P i b ) . v_{a,i}=u_a(y)-u_a(P_i^b),\qquad v_{c,i}=u_c(y)-u_c(P_i^b),\qquad E_i=F_{ca}(y)-F_{ca}(P_i^b). v a , i = u a ( y ) − u a ( P i b ) , v c , i = u c ( y ) − u c ( P i b ) , E i = F c a ( y ) − F c a ( P i b ) .
The support condition and the preceding gradient bound imply
∣ E i ∣ ≤ C ℓ b λ a λ c , ∣ ∇ E i ∣ ≤ C λ a λ c , |E_i|\le C\ell_b\lambda_a\lambda_c,\qquad|\nabla E_i|\le C\lambda_a\lambda_c, ∣ E i ∣ ≤ C ℓ b λ a λ c , ∣∇ E i ∣ ≤ C λ a λ c ,
∣ v a , i ∣ ≤ C ℓ b λ a , ∣ ∇ v a , i ∣ ≤ C λ a , |v_{a,i}|\le C\ell_b\lambda_a,\qquad|\nabla v_{a,i}|\le C\lambda_a, ∣ v a , i ∣ ≤ C ℓ b λ a , ∣∇ v a , i ∣ ≤ C λ a ,
∣ v c , i ∣ ≤ C ℓ b λ c , ∣ ∇ v c , i ∣ ≤ C λ c . |v_{c,i}|\le C\ell_b\lambda_c,\qquad|\nabla v_{c,i}|\le C\lambda_c. ∣ v c , i ∣ ≤ C ℓ b λ c , ∣∇ v c , i ∣ ≤ C λ c .
Since ∣ ∇ W b ν ∣ ≤ C ℓ b − 1 |\nabla W_b^\nu|\le C\ell_b^{-1} ∣∇ W b ν ∣ ≤ C ℓ b − 1 and ∣ D 2 W b ν ∣ ≤ C ℓ b − 2 |D^2W_b^\nu|\le C\ell_b^{-2} ∣ D 2 W b ν ∣ ≤ C ℓ b − 2 , each term in braces in (54) is bounded by C λ a λ c C\lambda_a\lambda_c C λ a λ c . One spatial derivative costs at most one additional factor ℓ b − 1 \ell_b^{-1} ℓ b − 1 : it either differentiates E i E_i E i or one of the v v v ’s, or falls on a wavelet derivative, using ∣ ∇ D 2 W b ν ∣ ≤ C ℓ b − 3 |\nabla D^2W_b^\nu|\le C\ell_b^{-3} ∣∇ D 2 W b ν ∣ ≤ C ℓ b − 3 . Bounded overlap and the coefficient bound therefore give
∥ B c F a b ∥ ∞ ≤ C A b λ a λ c , ∥ ∇ ( B c F a b ) ∥ ∞ ≤ C A b λ a λ c / ℓ b . \|\mathcal{B}_cF_{ab}\|_\infty\le CA_b\lambda_a\lambda_c,\qquad\|\nabla(\mathcal{B}_cF_{ab})\|_\infty\le CA_b\lambda_a\lambda_c/\ell_b. ∥ B c F ab ∥ ∞ ≤ C A b λ a λ c , ∥∇ ( B c F ab ) ∥ ∞ ≤ C A b λ a λ c / ℓ b .
By (49) ,
∣ δ ( B c F a b ) ∣ ≤ C A b λ a λ c m b = C r λ a λ c γ b = C r λ a K c b . |\delta(\mathcal{B}_cF_{ab})|\le CA_b\lambda_a\lambda_cm_b=Cr\lambda_a\lambda_c\gamma_b=Cr\lambda_aK_{cb}. ∣ δ ( B c F ab ) ∣ ≤ C A b λ a λ c m b = C r λ a λ c γ b = C r λ a K c b .
Case 2: a < b a < b a < b and c ≥ b c \ge b c ≥ b . The function F a b F_{ab} F ab depends only on y y y , P a P^a P a , P b P^b P b , and B c \mathcal{B}_c B c moves neither family of centers in this case. Thus B c F a b = ( u c ⋅ ∇ ) F a b \mathcal{B}_cF_{ab}=(u_c\cdot\nabla)F_{ab} B c F ab = ( u c ⋅ ∇ ) F ab . The q = 1 , 2 q=1,2 q = 1 , 2 bounds in (52) imply
∥ ∇ F a b ∥ ∞ ≤ C λ a λ b , ∣ δ ∇ F a b ∣ ≤ C λ a λ b m b . \lVert\nabla F_{ab}\rVert_\infty\le C\lambda_a\lambda_b,\qquad|\delta\nabla F_{ab}|\le C\lambda_a\lambda_bm_b. ∥ ∇ F ab ∥ ∞ ≤ C λ a λ b , ∣ δ ∇ F ab ∣ ≤ C λ a λ b m b .
The product rule for differences and m b ≤ m c m_b\le m_c m b ≤ m c now give
∣ δ ( B c F a b ) ∣ ≤ ∣ δ u c ∣ ∥ ∇ F a b ∥ ∞ + ∥ u c ∥ ∞ ∣ δ ∇ F a b ∣ ≤ C A c λ a λ b ( m c + m b ) ≤ C r λ a λ b γ c = C r λ a K c b . \begin{aligned}
|\delta(\mathcal{B}_cF_{ab})|\le|\delta u_c|\lVert\nabla F_{ab}\rVert_\infty+\lVert u_c\rVert_\infty|\delta\nabla F_{ab}| \\
&\le CA_c\lambda_a\lambda_b(m_c+m_b)\le Cr\lambda_a\lambda_b\gamma_c=Cr\lambda_aK_{cb}.
\end{aligned} ∣ δ ( B c F ab ) ∣ ≤ ∣ δ u c ∣ ∥ ∇ F ab ∥ ∞ + ∥ u c ∥ ∞ ∣ δ ∇ F ab ∣ ≤ C A c λ a λ b ( m c + m b ) ≤ C r λ a λ b γ c = C r λ a K c b .
Cases 3 and 4: a ≥ b a\ge b a ≥ b . We must now prove the bound C r λ b K c a Cr\lambda_bK_{ca} C r λ b K c a . The product rule applied to F a b = ( u a ⋅ ∇ ) u b F_{ab}=(u_a\cdot\nabla)u_b F ab = ( u a ⋅ ∇ ) u b gives
B c F a b = F c a ⋅ ∇ u b + u a ⋅ ( B c ∇ u b ) . (55) \mathcal{B}_cF_{ab}=F_{ca}\cdot\nabla u_b+u_a\cdot(\mathcal{B}_c\nabla u_b). \tag*{(55)} B c F ab = F c a ⋅ ∇ u b + u a ⋅ ( B c ∇ u b ) . (55)
Here B c ∇ u b \mathcal{B}_c\nabla u_b B c ∇ u b means that B c \mathcal{B}_c B c acts directly on the configuration-dependent spatial derivative ∇ y u b \nabla_yu_b ∇ y u b ; we do not interchange B c \mathcal{B}_c B c and ∇ y \nabla_y ∇ y . Using (46) , (53) , and (50) , we obtain
∣ δ ( F c a ⋅ ∇ u b ) ∣ ≤ ∣ δ F c a ∣ ∥ ∇ u b ∥ ∞ + ∥ F c a ∥ ∞ ∣ δ ∇ u b ∣ ≤ C r K c a λ b + C λ min ( c , a ) A max ( c , a ) λ b m b ≤ C r λ b K c a . \begin{aligned}
|\delta(F_{ca}\cdot\nabla u_b)|\le|\delta F_{ca}|\lVert\nabla u_b\rVert_\infty+\lVert F_{ca}\rVert_\infty|\delta\nabla u_b| \\
&\le CrK_{ca}\lambda_b+C\lambda_{\min(c,a)}A_{\max(c,a)}\lambda_bm_b\le Cr\lambda_bK_{ca}.
\end{aligned} ∣ δ ( F c a ⋅ ∇ u b ) ∣ ≤ ∣ δ F c a ∣ ∥ ∇ u b ∥ ∞ + ∥ F c a ∥ ∞ ∣ δ ∇ u b ∣ ≤ C r K c a λ b + C λ m i n ( c , a ) A m a x ( c , a ) λ b m b ≤ C r λ b K c a .
For the last inequality, max ( c , a ) ≥ a ≥ b \max(c,a)\ge a\ge b max ( c , a ) ≥ a ≥ b implies m b ≤ m max ( c , a ) m_b\le m_{\max(c,a)} m b ≤ m m a x ( c , a ) , and
λ min ( c , a ) A max ( c , a ) m max ( c , a ) = r K c a . \lambda_{\min(c,a)}A_{\max(c,a)}m_{\max(c,a)}=rK_{ca}. λ m i n ( c , a ) A m a x ( c , a ) m m a x ( c , a ) = r K c a .
It remains to estimate the second term in (55) .
Case 3: a ≥ b a\ge b a ≥ b and c < b c<b c < b . Since c < b c<b c < b , the operator B c \mathcal{B}_c B c differentiates both the query point and the scale-b b b centers. As ∇ u b \nabla u_b ∇ u b depends only on these coordinates, with the coefficients held fixed,
B c ∇ u b = ( u c ( y ) ⋅ ∇ y ) ∇ u b + ∑ i ( u c ( P i b ) ⋅ ∇ P i b ) ∇ u b . \mathcal{B}_c\nabla u_b=(u_c(y)\cdot\nabla_y)\nabla u_b+\sum_i(u_c(P_i^b)\cdot\nabla_{P_i^b})\nabla u_b. B c ∇ u b = ( u c ( y ) ⋅ ∇ y ) ∇ u b + i ∑ ( u c ( P i b ) ⋅ ∇ P i b ) ∇ u b .
Using ∇ u b ( y ) = A b ∑ i , ν η i b , ν ∇ W b ν ( y − P i b ) \nabla u_b(y)=A_b\sum_{i,\nu}\eta_i^{b,\nu}\nabla W_b^\nu(y-P_i^b) ∇ u b ( y ) = A b ∑ i , ν η i b , ν ∇ W b ν ( y − P i b ) and ∇ P i b = − ∇ y \nabla_{P_i^b}=-\nabla_y ∇ P i b = − ∇ y on each translated wavelet, we obtain
G ( y ) : = B c ∇ u b ( y ) = A b ∑ i , ν η i b , ν [ u c ( y ) − u c ( P i b ) ] ⋅ D 2 W b ν ( y − P i b ) . G(y):=\mathcal{B}_c\nabla u_b(y)=A_b\sum_{i,\nu}\eta_i^{b,\nu}[u_c(y)-u_c(P_i^b)]\cdot D^2W_b^\nu(y-P_i^b). G ( y ) := B c ∇ u b ( y ) = A b i , ν ∑ η i b , ν [ u c ( y ) − u c ( P i b )] ⋅ D 2 W b ν ( y − P i b ) .
On each support (given by the wavelet), the velocity difference is bounded by C ℓ b λ c C\ell_b\lambda_c C ℓ b λ c and its spatial gradient by C λ c C\lambda_c C λ c . Differentiating the displayed formula once and using bounded overlap therefore gives
∥ G ∥ ∞ ≤ C λ b λ c , ∥ ∇ G ∥ ∞ ≤ C λ b λ c / ℓ b , ∣ δ G ∣ ≤ C λ b λ c m b . \lVert G\rVert_\infty\le C\lambda_b\lambda_c,\qquad\lVert\nabla G\rVert_\infty\le C\lambda_b\lambda_c/\ell_b,\qquad|\delta G|\le C\lambda_b\lambda_cm_b. ∥ G ∥ ∞ ≤ C λ b λ c , ∥ ∇ G ∥ ∞ ≤ C λ b λ c / ℓ b , ∣ δ G ∣ ≤ C λ b λ c m b .
Since m b ≤ m a m_b\le m_a m b ≤ m a , the product difference satisfies
∣ δ ( u a ⋅ G ) ∣ ≤ ∣ δ u a ∣ ∥ G ∥ ∞ + ∥ u a ∥ ∞ ∣ δ G ∣ ≤ C A a λ b λ c ( m a + m b ) ≤ C r λ b λ c γ a = C r λ b K c a . \begin{aligned}
|\delta(u_a\cdot G)|\le|\delta u_a|\lVert G\rVert_\infty+\lVert u_a\rVert_\infty|\delta G| \\
&\le CA_a\lambda_b\lambda_c(m_a+m_b)\le Cr\lambda_b\lambda_c\gamma_a=Cr\lambda_bK_{ca}.
\end{aligned} ∣ δ ( u a ⋅ G ) ∣ ≤ ∣ δ u a ∣ ∥ G ∥ ∞ + ∥ u a ∥ ∞ ∣ δ G ∣ ≤ C A a λ b λ c ( m a + m b ) ≤ C r λ b λ c γ a = C r λ b K c a .
Case 4: a ≥ b a \ge b a ≥ b and c ≥ b c \ge b c ≥ b . Now B c \mathcal{B}_c B c does not move the scale-b b b centers, so B c ∇ u b = ( u c ⋅ ∇ ) ∇ u b \mathcal{B}_c\nabla u_b=(u_c\cdot\nabla)\nabla u_b B c ∇ u b = ( u c ⋅ ∇ ) ∇ u b . The second term in (55) is consequently ( u a ⊗ u c ) : D 2 u b (u_a\otimes u_c):D^2u_b ( u a ⊗ u c ) : D 2 u b . Taking a difference in each of its three factors and using (50) gives
∣ δ ( ( u a ⊗ u c ) : D 2 u b ) ∣ ≤ C λ b ℓ b A a A c ( m a + m c + m b ) ≤ C λ b ℓ b A min ( c , a ) A max ( c , a ) m max ( c , a ) = C r λ b A min ( c , a ) ℓ b γ max ( c , a ) ≤ C r λ b λ min ( c , a ) γ max ( c , a ) = C r λ b K c a . \begin{aligned}
\left|\delta((u_a\otimes u_c):D^2u_b)\right|\le C\frac{\lambda_b}{\ell_b}A_aA_c(m_a+m_c+m_b) \\
&\le C\frac{\lambda_b}{\ell_b}A_{\min(c,a)}A_{\max(c,a)}m_{\max(c,a)}
=Cr\lambda_b\frac{A_{\min(c,a)}}{\ell_b}\gamma_{\max(c,a)} \\
&\le Cr\lambda_b\lambda_{\min(c,a)}\gamma_{\max(c,a)}
=Cr\lambda_bK_{ca}.
\end{aligned} δ (( u a ⊗ u c ) : D 2 u b ) ≤ C ℓ b λ b A a A c ( m a + m c + m b ) ≤ C ℓ b λ b A m i n ( c , a ) A m a x ( c , a ) m m a x ( c , a ) = C r λ b ℓ b A m i n ( c , a ) γ m a x ( c , a ) ≤ C r λ b λ m i n ( c , a ) γ m a x ( c , a ) = C r λ b K c a .
Indeed, all three increment factors are at most m max ( c , a ) m_{\max(c,a)} m m a x ( c , a ) , and
A min ( c , a ) ℓ b = λ min ( c , a ) ℓ min ( c , a ) ℓ b ≤ λ min ( c , a ) . \frac{A_{\min(c,a)}}{\ell_b}=\lambda_{\min(c,a)}\frac{\ell_{\min(c,a)}}{\ell_b}\le\lambda_{\min(c,a)}. ℓ b A m i n ( c , a ) = λ m i n ( c , a ) ℓ b ℓ m i n ( c , a ) ≤ λ m i n ( c , a ) .
Together with the estimate for the first term of (55) , this completes both cases with a ≥ b a\ge b a ≥ b , and hence proves (47) for every ordering of the indices. □ \square □
We now apply the operators of Definition 4.2 4.2 4.2 to the test functions of Section 3 3 3 . Recall that on the set of configurations with 0 < dist ( x , x ~ ) ≤ 1 / 4 0<\operatorname{dist}(x,\widetilde{x})\le1/4 0 < dist ( x , x ) ≤ 1/4 the shortest representative z = z ( x , x ~ ) ∈ R d z=z(x,\widetilde{x})\in\mathbb{R}^d z = z ( x , x ) ∈ R d of x − x ~ x-\widetilde{x} x − x is unique, satisfies ∣ z ∣ = dist ( x , x ~ ) |z|=\operatorname{dist}(x,\widetilde{x}) ∣ z ∣ = dist ( x , x ) , and depends smoothly on ( x , x ~ ) (x,\widetilde{x}) ( x , x ) , with ∇ x z = I \nabla_xz=I ∇ x z = I and ∇ x ~ z = − I \nabla_{\widetilde{x}}z=-I ∇ x z = − I . For J ≥ j 0 J\ge j_0 J ≥ j 0 and L ≥ 2 L\ge2 L ≥ 2 , the functions Φ \Phi Φ of (28) and Φ j \Phi_j Φ j of (31) , built with ε = ρ ℓ J \varepsilon=\rho\ell_J ε = ρ ℓ J , are therefore smooth functions
( x , x ~ , P ) ⟼ Φ ( z ( x , x ~ ) ) , ( x , x ~ , P ) ⟼ Φ j ( z ( x , x ~ ) ) (x,\widetilde{x},P)\longmapsto\Phi(z(x,\widetilde{x})),\qquad(x,\widetilde{x},P)\longmapsto\Phi_j(z(x,\widetilde{x})) ( x , x , P ) ⟼ Φ ( z ( x , x )) , ( x , x , P ) ⟼ Φ j ( z ( x , x ))
on that set, which depend neither on the query point nor on the centers or coefficients. (Recall that ρ > 0 \rho>0 ρ > 0 is fixed independently of J , κ , L J,\kappa,L J , κ , L so that 2 ρ < min ( v ∗ , 1 / 4 ) 2\rho<\min(v_*,1/4) 2 ρ < min ( v ∗ , 1/4 ) , where v ∗ v_* v ∗ is the threshold in (22) , and that ε = ρ ℓ J \varepsilon=\rho\ell_J ε = ρ ℓ J .) Consequently, for all a , b , c ∈ { j 0 , … , J } a,b,c\in\{j_0,\ldots,J\} a , b , c ∈ { j 0 , … , J } , the iterated derivatives B b Φ b \mathcal{B}_b\Phi_b B b Φ b , B a B b Φ b \mathcal{B}_a\mathcal{B}_b\Phi_b B a B b Φ b , and B c B a B b Φ b \mathcal{B}_c\mathcal{B}_a\mathcal{B}_b\Phi_b B c B a B b Φ b are well-defined continuous functions of the configuration, and the chain rule together with (42) and (43) gives the exact identities
B b Φ b = ∇ Φ b ( z ) ⋅ δ u b , B a B b Φ b = D 2 Φ b ( z ) [ δ u a , δ u b ] + ∇ Φ b ( z ) ⋅ δ ( B a u b ) . (56) \mathcal{B}_b\Phi_b=\nabla\Phi_b(z)\cdot\delta u_b,\qquad
\mathcal{B}_a\mathcal{B}_b\Phi_b=D^2\Phi_b(z)[\delta u_a,\delta u_b]+\nabla\Phi_b(z)\cdot\delta(\mathcal{B}_a u_b). \tag*{(56)} B b Φ b = ∇ Φ b ( z ) ⋅ δ u b , B a B b Φ b = D 2 Φ b ( z ) [ δ u a , δ u b ] + ∇ Φ b ( z ) ⋅ δ ( B a u b ) . (56)
Indeed, since z = z ( x , x ~ ) z=z(x,\widetilde{x}) z = z ( x , x ) is independent of the query point and the centers, with ∇ x z = I \nabla_xz=I ∇ x z = I and ∇ x ~ z = − I \nabla_{\widetilde{x}}z=-I ∇ x z = − I , the definition of B a \mathcal{B}_a B a gives B a z = δ u a \mathcal{B}_az=\delta u_a B a z = δ u a . Hence
B b Φ b = ∇ Φ b ( z ) ⋅ B b z = ∇ Φ b ( z ) ⋅ δ u b . \mathcal{B}_b\Phi_b=\nabla\Phi_b(z)\cdot\mathcal{B}_bz=\nabla\Phi_b(z)\cdot\delta u_b. B b Φ b = ∇ Φ b ( z ) ⋅ B b z = ∇ Φ b ( z ) ⋅ δ u b .
Next, apply B a \mathcal{B}_a B a to this product:
B a B b Φ b = ( B a [ ∇ Φ b ( z ) ] ) ⋅ δ u b + ∇ Φ b ( z ) ⋅ B a [ δ u b ] . \mathcal{B}_a\mathcal{B}_b\Phi_b=(\mathcal{B}_a[\nabla\Phi_b(z)])\cdot\delta u_b+\nabla\Phi_b(z)\cdot\mathcal{B}_a[\delta u_b]. B a B b Φ b = ( B a [ ∇ Φ b ( z )]) ⋅ δ u b + ∇ Φ b ( z ) ⋅ B a [ δ u b ] .
The chain rule handles the first term:
B a [ ∇ Φ b ( z ) ] = D 2 Φ b ( z ) [ δ u a ] , \mathcal{B}_a[\nabla\Phi_b(z)]=D^2\Phi_b(z)[\delta u_a], B a [ ∇ Φ b ( z )] = D 2 Φ b ( z ) [ δ u a ] ,
so its contribution is D 2 Φ b ( z ) [ δ u a , δ u b ] D^2\Phi_b(z)[\delta u_a,\delta u_b] D 2 Φ b ( z ) [ δ u a , δ u b ] .
For the second term, we must justify
B a [ δ u b ] = δ ( B a u b ) . \mathcal{B}_a[\delta u_b]=\delta(\mathcal{B}_au_b). B a [ δ u b ] = δ ( B a u b ) .
Recall that δ u b = u b ( x ; P b ) − u b ( x ~ ; P b ) \delta u_b=u_b(x;P^b)-u_b(\widetilde{x};P^b) δ u b = u b ( x ; P b ) − u b ( x ; P b ) . Differentiating the first evaluation gives
B a [ u b ( x ; P b ) ] = ( u a ( x ) ⋅ ∇ y ) u b ( y ; P b ) ∣ y = x + 1 { a < b } ∑ i ( u a ( P i b ) ⋅ ∇ P i b ) u b ( y ; P b ) ∣ y = x . \mathcal{B}_a[u_b(x;P^b)]
=\left.(u_a(x)\cdot\nabla_y)u_b(y;P^b)\right|_{y=x}
+\mathbf{1}_{\{a<b\}}\sum_i\left.(u_a(P_i^b)\cdot\nabla_{P_i^b})u_b(y;P^b)\right|_{y=x}. B a [ u b ( x ; P b )] = ( u a ( x ) ⋅ ∇ y ) u b ( y ; P b ) y = x + 1 { a < b } i ∑ ( u a ( P i b ) ⋅ ∇ P i b ) u b ( y ; P b ) y = x .
The first term accounts for the motion of x x x ; the second accounts for the motion of the centers defining u b u_b u b . Together they are exactly
( B a u b ) ( x ; x , x ~ ; P ) . (\mathcal{B}_au_b)(x;x,\widetilde{x};P). ( B a u b ) ( x ; x , x ; P ) .
The same calculation at x ~ \widetilde{x} x , followed by subtraction, gives the claimed identity by the definition (43) of δ \delta δ .
Similarly, for the third order, the five-term expansion
B c B a B b Φ b = D 3 Φ b ( z ) [ δ u c , δ u a , δ u b ] + D 2 Φ b ( z ) [ δ ( B c u a ) , δ u b ] + D 2 Φ b ( z ) [ δ u a , δ ( B c u b ) ] + D 2 Φ b ( z ) [ δ u c , δ ( B a u b ) ] + ∇ Φ b ( z ) ⋅ δ ( B c B a u b ) , (57) \begin{aligned}
\mathcal{B}_c\mathcal{B}_a\mathcal{B}_b\Phi_b
={}&D^3\Phi_b(z)[\delta u_c,\delta u_a,\delta u_b]
+D^2\Phi_b(z)[\delta(\mathcal{B}_cu_a),\delta u_b]\\
&+D^2\Phi_b(z)[\delta u_a,\delta(\mathcal{B}_cu_b)]
+D^2\Phi_b(z)[\delta u_c,\delta(\mathcal{B}_au_b)]\\
&+\nabla\Phi_b(z)\cdot\delta(\mathcal{B}_c\mathcal{B}_au_b), \tag*{(57)}
\end{aligned} B c B a B b Φ b = D 3 Φ b ( z ) [ δ u c , δ u a , δ u b ] + D 2 Φ b ( z ) [ δ ( B c u a ) , δ u b ] + D 2 Φ b ( z ) [ δ u a , δ ( B c u b )] + D 2 Φ b ( z ) [ δ u c , δ ( B a u b )] + ∇ Φ b ( z ) ⋅ δ ( B c B a u b ) , (57)
where every δ ( ⋅ ) \delta(\cdot) δ ( ⋅ ) is the two-point difference (43) of the corresponding function on E \mathcal{E} E , evaluated at the given configuration. To see this, apply B c \mathcal{B}_c B c to the second identity (56) to get the third-order expansion. We use the following facts. First, locally, z = x − x ~ + n z=x-\widetilde{x}+n z = x − x + n , where the integer vector n n n is constant. Thus
( B c z ) i = ∑ j u c , j ( x ) ∂ x j z i + ∑ j u c , j ( x ~ ) ∂ x ~ j z i = u c , i ( x ) − u c , i ( x ~ ) , (\mathcal{B}_cz)_i=\sum_j u_{c,j}(x)\partial_{x_j}z_i+\sum_j u_{c,j}(\widetilde{x})\partial_{\widetilde{x}_j}z_i=u_{c,i}(x)-u_{c,i}(\widetilde{x}), ( B c z ) i = j ∑ u c , j ( x ) ∂ x j z i + j ∑ u c , j ( x ) ∂ x j z i = u c , i ( x ) − u c , i ( x ) ,
so that
B c z = u c ( x ) − u c ( x ~ ) = δ u c . \mathcal{B}_cz=u_c(x)-u_c(\widetilde{x})=\delta u_c. B c z = u c ( x ) − u c ( x ) = δ u c .
It is the change in separation generated by moving both particles with the scale-c c c velocity field.
Secondly,
B c [ δ G ] = δ ( B c G ) , G = u a , u b , B a u b . \mathcal{B}_c[\delta G]=\delta(\mathcal{B}_cG),\qquad G=u_a,\ u_b,\ \mathcal{B}_au_b. B c [ δ G ] = δ ( B c G ) , G = u a , u b , B a u b .
The latter identity follows from the same evaluation chain rule used above: B c \mathcal{B}_c B c transports each particle with the field assigned to the query point, and the center derivatives act on both sides in the same way.
The product and chain rules give three terms when applied to the Hessian term:
B c ( D 2 Φ b ( z ) [ δ u a , δ u b ] ) = D 3 Φ b ( z ) [ δ u c , δ u a , δ u b ] + D 2 Φ b ( z ) [ δ ( B c u a ) , δ u b ] + D 2 Φ b ( z ) [ δ u a , δ ( B c u b ) ] . \mathcal{B}_c(D^2\Phi_b(z)[\delta u_a,\delta u_b])=D^3\Phi_b(z)[\delta u_c,\delta u_a,\delta u_b]+D^2\Phi_b(z)[\delta(\mathcal{B}_c u_a),\delta u_b]+D^2\Phi_b(z)[\delta u_a,\delta(\mathcal{B}_c u_b)]. B c ( D 2 Φ b ( z ) [ δ u a , δ u b ]) = D 3 Φ b ( z ) [ δ u c , δ u a , δ u b ] + D 2 Φ b ( z ) [ δ ( B c u a ) , δ u b ] + D 2 Φ b ( z ) [ δ u a , δ ( B c u b )] .
These arise by differentiating the Hessian and its two vector arguments, respectively. The gradient term contributes two more:
B c ( ∇ Φ b ( z ) ⋅ δ ( B a u b ) ) = D 2 Φ b ( z ) [ δ u c , δ ( B a u b ) ] + ∇ Φ b ( z ) ⋅ δ ( B c B a u b ) . \mathcal{B}_c(\nabla\Phi_b(z)\cdot\delta(\mathcal{B}_a u_b))=D^2\Phi_b(z)[\delta u_c,\delta(\mathcal{B}_a u_b)]+\nabla\Phi_b(z)\cdot\delta(\mathcal{B}_c\mathcal{B}_a u_b). B c ( ∇ Φ b ( z ) ⋅ δ ( B a u b )) = D 2 Φ b ( z ) [ δ u c , δ ( B a u b )] + ∇ Φ b ( z ) ⋅ δ ( B c B a u b ) .
Adding yields (57) . No commutation of the scale operators is used; their order is preserved throughout.
Lemma 4.6 (Operator bounds for the test functions). There is a constant C > 0 C>0 C > 0 , depending only on d d d , α \alpha α , ψ \psi ψ , R R R , M M M and on the fixed auxiliary functions ϑ \vartheta ϑ , χ \chi χ of § 3 , with the following property. Let J ≥ j 0 J\ge j_0 J ≥ j 0 and L ≥ 2 L\ge2 L ≥ 2 , and let Φ \Phi Φ and Φ j \Phi_j Φ j (j 0 ≤ j ≤ J j_0\le j\le J j 0 ≤ j ≤ J ) be the test functions defined in (28) and (31) , respectively, with ε = ρ ℓ J \varepsilon=\rho\ell_J ε = ρ ℓ J . Let S = ( x , x ~ , Θ ) S=(x,\widetilde{x},\Theta) S = ( x , x , Θ ) be any admissible configuration in the sense of Definition 4.1 , let r = dist ( x , x ~ ) r=\operatorname{dist}(x,\widetilde{x}) r = dist ( x , x ) , S = S ( r ) = max ( r , ℓ J ) S=S(r)=\max(r,\ell_J) S = S ( r ) = max ( r , ℓ J ) , and let γ j \gamma_j γ j , K a b K_{ab} K ab be given by (44) with this r r r . Then, with all left-hand sides evaluated at S S S , for all a , b , c ∈ { j 0 , … , J } a,b,c\in\{j_0,\ldots,J\} a , b , c ∈ { j 0 , … , J } ,
∣ B b Φ b ∣ ≤ C S β γ b , (58) |\mathcal{B}_b\Phi_b|\le CS^\beta\gamma_b, \tag*{(58)} ∣ B b Φ b ∣ ≤ C S β γ b , (58) ∣ B a B b Φ b ∣ ≤ C S β K a b , (59) |\mathcal{B}_a\mathcal{B}_b\Phi_b|\le CS^\beta K_{ab}, \tag*{(59)} ∣ B a B b Φ b ∣ ≤ C S β K ab , (59) ∣ B c B a B b Φ b ∣ ≤ C S β ( γ c K a b + γ a K c b + γ b K c a + λ min ( a , b ) K c , max ( a , b ) ) . (60) |\mathcal{B}_c\mathcal{B}_a\mathcal{B}_b\Phi_b|\le CS^\beta\left(\gamma_cK_{ab}+\gamma_aK_{cb}+\gamma_bK_{ca}+\lambda_{\min(a,b)}K_{c,\max(a,b)}\right). \tag*{(60)} ∣ B c B a B b Φ b ∣ ≤ C S β ( γ c K ab + γ a K c b + γ b K c a + λ m i n ( a , b ) K c , m a x ( a , b ) ) . (60) Moreover:
(i) If ℓ b > 3 L S ( r ) \ell_b>3LS(r) ℓ b > 3 L S ( r ) , then B b Φ b \mathcal{B}_b\Phi_b B b Φ b , B a B b Φ b \mathcal{B}_a\mathcal{B}_b\Phi_b B a B b Φ b , and B c B a B b Φ b \mathcal{B}_c\mathcal{B}_a\mathcal{B}_b\Phi_b B c B a B b Φ b all vanish at S S S , for every a , c ∈ { j 0 , … , J } a,c\in\{j_0,\ldots,J\} a , c ∈ { j 0 , … , J } .
(ii) For every b ∈ { j 0 , … , J } b\in\{j_0,\ldots,J\} b ∈ { j 0 , … , J } , ∣ B b Φ ∣ ≤ C S β γ b |\mathcal{B}_b\Phi|\le CS^\beta\gamma_b ∣ B b Φ∣ ≤ C S β γ b at S S S .
The constant C C C does not depend on J J J , κ \kappa κ , L L L , the configuration S S S , or the indices a , b , c a,b,c a , b , c .
Proof. Insert the bounds of Lemma 4.3 and the derivative bounds (32) into the identities (56) –(57) . The first identity in (56) gives ∣ B b Φ b ∣ ≤ C S β r − 1 ⋅ C r γ b |\mathcal{B}_b\Phi_b|\le CS^\beta r^{-1}\cdot Cr\gamma_b ∣ B b Φ b ∣ ≤ C S β r − 1 ⋅ C r γ b , which is (58) ; the second gives ∣ B a B b Φ b ∣ ≤ C S β ( γ a γ b + K a b ) ≤ C S β K a b |\mathcal{B}_a\mathcal{B}_b\Phi_b|\le CS^\beta(\gamma_a\gamma_b+K_{ab})\le CS^\beta K_{ab} ∣ B a B b Φ b ∣ ≤ C S β ( γ a γ b + K ab ) ≤ C S β K ab , which is (59) , because γ min ( a , b ) ≤ λ min ( a , b ) \gamma_{\min(a,b)}\le\lambda_{\min(a,b)} γ m i n ( a , b ) ≤ λ m i n ( a , b ) implies γ a γ b ≤ K a b \gamma_a\gamma_b\le K_{ab} γ a γ b ≤ K ab . In (57) the five terms are bounded, in the displayed order, by C S β γ c γ a γ b CS^\beta\gamma_c\gamma_a\gamma_b C S β γ c γ a γ b , K c a γ b K_{ca}\gamma_b K c a γ b , γ a K c b \gamma_aK_{cb} γ a K c b , γ c K a b \gamma_cK_{ab} γ c K ab , and λ min ( a , b ) K c , max ( a , b ) \lambda_{\min(a,b)}K_{c,\max(a,b)} λ m i n ( a , b ) K c , m a x ( a , b ) ; absorbing the first into the fourth by γ a γ b ≤ K a b \gamma_a\gamma_b\le K_{ab} γ a γ b ≤ K ab proves (60) . All constants come from Lemma 4.3 and (32) and therefore depend only on the quantities listed in the statement.
For (i), if ℓ b > 3 L S ( r ) \ell_b>3LS(r) ℓ b > 3 L S ( r ) , then ∇ k Φ b ( z ) = 0 \nabla^k\Phi_b(z)=0 ∇ k Φ b ( z ) = 0 for k = 1 , 2 , 3 k=1,2,3 k = 1 , 2 , 3 by (32) , so every term in (56) –(57) vanishes at S S S , whatever a a a and c c c are. For (ii), the first identity in (56) holds verbatim with Φ \Phi Φ in place of Φ b \Phi_b Φ b , and (30) with k = 1 k=1 k = 1 together with (45) gives ∣ B b Φ ∣ ≤ C S β γ b |\mathcal{B}_b\Phi|\le CS^\beta\gamma_b ∣ B b Φ∣ ≤ C S β γ b . □
Summation over scales and the corrected observable
This section performs the summation over scales and introduces the corrected observable. With γ j = min ( λ j , A j / r ) \gamma_j=\min(\lambda_j,A_j/r) γ j = min ( λ j , A j / r ) , the scales coarser than r r r contribute r λ j = r ℓ j − β r\lambda_j=r\ell_j^{-\beta} r λ j = r ℓ j − β and the finer ones contribute A j = ℓ j α A_j=\ell_j^\alpha A j = ℓ j α ; the two geometric series meet at ℓ j ≈ r \ell_j\approx r ℓ j ≈ r and sum to a multiple of r α r^\alpha r α , as in the Littlewood–Paley characterization of C α C^\alpha C α . The corrected observable is obtained by adding to Φ \Phi Φ lower-order terms chosen so that its time derivative between refreshes cancels exactly to second order.
The objects of this section are functions of an admissible configuration S \mathcal{S} S (Definition 4.1 ) and a vector of phases
h = ( h j ) j 0 ≤ j ≤ J ∈ ∏ j = j 0 J [ 0 , τ j ] . h=(h_j)_{j_0\leq j\leq J}\in\prod_{j=j_0}^{J}[0,\tau_j]. h = ( h j ) j 0 ≤ j ≤ J ∈ j = j 0 ∏ J [ 0 , τ j ] .
Along the process,
h j ( t ) = ( k + 1 ) τ j − t , k τ j ≤ t < ( k + 1 ) τ j , h_j(t)=(k+1)\tau_j-t,\qquad k\tau_j\leq t<(k+1)\tau_j, h j ( t ) = ( k + 1 ) τ j − t , k τ j ≤ t < ( k + 1 ) τ j ,
is the time remaining until the nominal end of the active window, even if that endpoint exceeds 1 1 1 . At t = 1 t=1 t = 1 , use the left-limit convention h j ( 1 ) = h j ( 1 − ) = K j τ j − 1 h_j(1)=h_j(1-)=K_j\tau_j-1 h j ( 1 ) = h j ( 1 − ) = K j τ j − 1 . Thus h ˙ j = − 1 \dot{h}_j=-1 h ˙ j = − 1 between renewals, and h j h_j h j jumps from 0 0 0 to τ j \tau_j τ j at each renewal of scale j j j .
The correction ∑ b h b B b Φ b \sum_b h_b\mathcal{B}_b\Phi_b ∑ b h b B b Φ b cancels the first-order evolution up to the cutoff error. Between renewals, the configuration evolves with generator B = ∑ a B a \mathcal{B}=\sum_a\mathcal{B}_a B = ∑ a B a , while h ˙ b = − 1 \dot{h}_b=-1 h ˙ b = − 1 . Hence
d d t Φ = ∑ b B b Φ , d d t ( h b B b Φ b ) = − B b Φ b + h b ∑ a B a B b Φ b . \frac{d}{dt}\Phi=\sum_b\mathcal{B}_b\Phi,\qquad\frac{d}{dt}(h_b\mathcal{B}_b\Phi_b)=-\mathcal{B}_b\Phi_b+h_b\sum_a\mathcal{B}_a\mathcal{B}_b\Phi_b. d t d Φ = b ∑ B b Φ , d t d ( h b B b Φ b ) = − B b Φ b + h b a ∑ B a B b Φ b .
Summing gives
d d t ( Φ + ∑ b h b B b Φ b ) = ∑ b B b ( Φ − Φ b ) + ∑ a , b h b B a B b Φ b . \frac{d}{dt}\left(\Phi+\sum_b h_b\mathcal{B}_b\Phi_b\right)=\sum_b\mathcal{B}_b(\Phi-\Phi_b)+\sum_{a,b}h_b\mathcal{B}_a\mathcal{B}_b\Phi_b. d t d ( Φ + b ∑ h b B b Φ b ) = b ∑ B b ( Φ − Φ b ) + a , b ∑ h b B a B b Φ b .
We will discuss this more in Proposition 5.2. The first sum is the cutoff error; the second consists of the second-order terms to be canceled by the weights q a b q_{ab} q ab . The choice of remaining time also makes the old scale-b b b first-order correction vanish just before that scale is renewed.
To cancel the remaining second-order terms, we introduce weights q a b q_{ab} q ab whose derivatives along the phase flow equal − h b -h_b − h b . We also require q a b q_{ab} q ab to vanish whenever either of the two windows ends, so that the corresponding old correction disappears at a renewal. Both requirements are met by integrating the remaining lifetime of window b b b up to the first of the two window endpoints:
q a b = ∫ 0 min ( h a , h b ) ( h b − s ) d s = h b min ( h a , h b ) − 1 2 min ( h a , h b ) 2 . (61) q_{ab}=\int_0^{\min(h_a,h_b)}(h_b-s)\,\mathrm{d}s=h_b\min(h_a,h_b)-\frac{1}{2}\min(h_a,h_b)^2. \tag*{(61)} q ab = ∫ 0 m i n ( h a , h b ) ( h b − s ) d s = h b min ( h a , h b ) − 2 1 min ( h a , h b ) 2 . (61)
Indeed,
q a b = { h a h b − 1 2 h a 2 , h a ≤ h b , 1 2 h b 2 , h a ≥ h b . q_{ab}=
\begin{cases}
h_ah_b-\frac{1}{2}h_a^2, & h_a\leq h_b,\\
\frac{1}{2}h_b^2, & h_a\geq h_b.
\end{cases} q ab = { h a h b − 2 1 h a 2 , 2 1 h b 2 , h a ≤ h b , h a ≥ h b .
The values and first derivatives of these two expressions agree on h a = h b h_a = h_b h a = h b , so q a b q_{ab} q ab is continuously differentiable on the nonnegative quadrant.
Write
D h : = − ∑ j = j 0 J ∂ ∂ h j (62) \mathcal{D}_h := -\sum_{j=j_0}^{J} \frac{\partial}{\partial h_j} \tag*{(62)} D h := − j = j 0 ∑ J ∂ h j ∂ (62)
for differentiation along the phase flow h ˙ j = − 1 \dot{h}_j=-1 h ˙ j = − 1 , with the configuration held fixed. Thus D h \mathcal{D}_h D h acts only on the phase variables, whereas B \mathcal{B} B acts on the configuration variables with the phases held fixed. Along the process, between renewals, the full time derivative is represented by D h + B \mathcal{D}_h+\mathcal{B} D h + B . The preceding formulas give
D h q a b = − h b , q b b = h b 2 2 , q a b = 0 if h a h b = 0 , 0 ≤ q a b ≤ κ 2 λ b max ( λ a , λ b ) . (63) \mathcal{D}_h q_{ab}=-h_b,\qquad q_{bb}=\frac{h_b^2}{2},\qquad q_{ab}=0\ \text{if }h_a h_b=0,\qquad0\leq q_{ab}\leq\frac{\kappa^2}{\lambda_b\max(\lambda_a,\lambda_b)}. \tag*{(63)} D h q ab = − h b , q bb = 2 h b 2 , q ab = 0 if h a h b = 0 , 0 ≤ q ab ≤ λ b max ( λ a , λ b ) κ 2 . (63)
For a ≠ b a\ne b a = b , the derivative identity follows by applying − ∂ h a − ∂ h b -\partial_{h_a}-\partial_{h_b} − ∂ h a − ∂ h b to either branch; for a = b a=b a = b , it follows directly from q b b = h b 2 / 2 q_{bb}=h_b^2/2 q bb = h b 2 /2 . Thus, along the process and away from renewals, q ˙ a b = − h b \dot{q}_{ab}=-h_b q ˙ ab = − h b . The upper bound follows from
q a b ≤ h b min ( h a , h b ) ≤ τ b min ( τ a , τ b ) ≤ κ 2 λ b max ( λ a , λ b ) . q_{ab}\leq h_b\min(h_a,h_b)\leq\tau_b\min(\tau_a,\tau_b)\leq\frac{\kappa^2}{\lambda_b\max(\lambda_a,\lambda_b)}. q ab ≤ h b min ( h a , h b ) ≤ τ b min ( τ a , τ b ) ≤ λ b max ( λ a , λ b ) κ 2 .
Consequently, differentiating the correction ∑ a , b q a b B a B b Φ b \sum_{a,b}q_{ab}\mathcal{B}_a\mathcal{B}_b\Phi_b ∑ a , b q ab B a B b Φ b between renewals produces − ∑ a , b h b B a B b Φ b -\sum_{a,b}h_b\mathcal{B}_a\mathcal{B}_b\Phi_b − ∑ a , b h b B a B b Φ b , which cancels the second-order terms above, together with the third-order remainder ∑ a , b , c q a b B c B a B b Φ b \sum_{a,b,c}q_{ab}\mathcal{B}_c\mathcal{B}_a\mathcal{B}_b\Phi_b ∑ a , b , c q ab B c B a B b Φ b .
Lemma 5.1 (Uniform summation). Let N L = 1 + log L N_L=1+\log L N L = 1 + log L . There is a constant C C C , independent of J J J , κ \kappa κ , and L ≥ 2 L\geq2 L ≥ 2 , such that for every admissible configuration S \mathcal{S} S with separation r r r and S = S ( r ) S=S(r) S = S ( r ) , and every phase vector h ∈ ∏ j [ 0 , τ j ] h\in\prod_j[0,\tau_j] h ∈ ∏ j [ 0 , τ j ] , with all sums over j 0 , … , J j_0,\ldots,J j 0 , … , J and all operator expressions evaluated at S \mathcal{S} S as in Lemma 4.6,
∣ ∑ b h b B b Φ b ∣ ≤ C κ N L S β , (64) \left|\sum_b h_b\mathcal{B}_b\Phi_b\right|\leq C\kappa N_L S^\beta, \tag*{(64)} b ∑ h b B b Φ b ≤ C κ N L S β , (64) ∣ ∑ a , b q a b B a B b Φ b ∣ ≤ C κ 2 N L 2 S β , (65) \left|\sum_{a,b}q_{ab}\mathcal{B}_a\mathcal{B}_b\Phi_b\right|\leq C\kappa^2N_L^2S^\beta, \tag*{(65)} a , b ∑ q ab B a B b Φ b ≤ C κ 2 N L 2 S β , (65) ∣ ∑ a , b , c q a b B c B a B b Φ b ∣ ≤ C κ 2 N L 2 . (66) \left|\sum_{a,b,c}q_{ab}\mathcal{B}_c\mathcal{B}_a\mathcal{B}_b\Phi_b\right|\leq C\kappa^2N_L^2. \tag*{(66)} a , b , c ∑ q ab B c B a B b Φ b ≤ C κ 2 N L 2 . (66)
Informally, the cutoff restricts the outer sum to scales satisfying ℓ b ≤ 3 L S \ell_b\leq3LS ℓ b ≤ 3 L S . Of these, only O ( N L ) O(N_L) O ( N L ) satisfy ℓ b ≥ S \ell_b\geq S ℓ b ≥ S . There may still be arbitrarily many finer scales as J J J increases, but their contributions decay geometrically and have a sum bounded independently of J J J . The factors N L N_L N L and N L 2 N_L^2 N L 2 below arise from summing over the surviving coarse scales, rather than from bounding the total number of scales.
Proof. Set
p b = γ b λ b = min ( 1 , ℓ b / S ) , γ ~ b = ∑ a ≥ b γ a , p ~ b = ∑ a ≥ b p a , d ~ b = ∑ a ≥ b γ ~ a . p_b=\frac{\gamma_b}{\lambda_b}=\min(1,\ell_b/S),\qquad\widetilde{\gamma}_b=\sum_{a\geq b}\gamma_a,\qquad\widetilde{p}_b=\sum_{a\geq b}p_a,\qquad\widetilde{d}_b=\sum_{a\geq b}\widetilde{\gamma}_a. p b = λ b γ b = min ( 1 , ℓ b / S ) , γ b = a ≥ b ∑ γ a , p b = a ≥ b ∑ p a , d b = a ≥ b ∑ γ a .
All sums here and below are restricted to the available generations. Since every available scale satisfies ℓ a ≥ ℓ J \ell_a \ge\ell_J ℓ a ≥ ℓ J , the definition S = max ( r , ℓ J ) S = \max(r,\ell_J) S = max ( r , ℓ J ) gives
γ a = { ℓ a − β , ℓ a ≥ S , ℓ a α / S , ℓ a < S . \gamma_a =
\begin{cases}
\ell_a^{-\beta}, & \ell_a \ge S, \\
\ell_a^\alpha/S, & \ell_a < S.
\end{cases} γ a = { ℓ a − β , ℓ a α / S , ℓ a ≥ S , ℓ a < S .
Indeed, when r < ℓ J r < \ell_J r < ℓ J , only the first case occurs. Since consecutive scales differ by a factor of two, summing from the scale nearest S S S towards the coarser or finer scales, respectively, gives
∑ ℓ a ≥ S γ a ≤ S − β ∑ n ≥ 0 2 − n β = S − β 1 − 2 − β , ∑ ℓ a < S γ a ≤ S − β ∑ n ≥ 0 2 − n α = S − β 1 − 2 − α . \sum_{\ell_a \ge S}\gamma_a \le S^{-\beta}\sum_{n\ge0}2^{-n\beta} = \frac{S^{-\beta}}{1-2^{-\beta}}, \qquad\sum_{\ell_a<S}\gamma_a \le S^{-\beta}\sum_{n\ge0}2^{-n\alpha} = \frac{S^{-\beta}}{1-2^{-\alpha}}. ℓ a ≥ S ∑ γ a ≤ S − β n ≥ 0 ∑ 2 − n β = 1 − 2 − β S − β , ℓ a < S ∑ γ a ≤ S − β n ≥ 0 ∑ 2 − n α = 1 − 2 − α S − β .
Here and below, sums over scales are restricted to the available generations; extending a geometric sum only increases its value. Also, γ a ≤ λ a \gamma_a \le\lambda_a γ a ≤ λ a and
∑ a < b λ a = λ b ∑ n = 1 b − j 0 2 − n β ≤ λ b 2 β − 1 . \sum_{a<b}\lambda_a = \lambda_b\sum_{n=1}^{b-j_0}2^{-n\beta} \le\frac{\lambda_b}{2^\beta-1}. a < b ∑ λ a = λ b n = 1 ∑ b − j 0 2 − n β ≤ 2 β − 1 λ b .
Combining this bound with the bound for the full sum of the γ a \gamma_a γ a yields
∑ a γ a ≤ C S − β , ∑ a < b γ a ≤ C min ( λ b , S − β ) , ∑ a < b λ a ≤ C λ b . (67) \sum_a\gamma_a \le CS^{-\beta}, \qquad\sum_{a<b}\gamma_a \le C\min(\lambda_b,S^{-\beta}), \qquad\sum_{a<b}\lambda_a \le C\lambda_b. \tag*{(67)} a ∑ γ a ≤ C S − β , a < b ∑ γ a ≤ C min ( λ b , S − β ) , a < b ∑ λ a ≤ C λ b . (67)
For the tail γ ~ b \widetilde{\gamma}_b γ b , if ℓ b ≥ S \ell_b \ge S ℓ b ≥ S , the full-sum estimate gives γ ~ b ≤ C S − β \widetilde{\gamma}_b \le CS^{-\beta} γ b ≤ C S − β . If ℓ b < S \ell_b<S ℓ b < S , every a ≥ b a\ge b a ≥ b is also below S S S , and hence
γ ~ b = ℓ b α S ∑ n = 0 J − b 2 − n α ≤ S − β 1 − 2 − α ( ℓ b S ) α , \widetilde{\gamma}_b = \frac{\ell_b^\alpha}{S}\sum_{n=0}^{J-b}2^{-n\alpha} \le\frac{S^{-\beta}}{1-2^{-\alpha}}\left(\frac{\ell_b}{S}\right)^\alpha, γ b = S ℓ b α n = 0 ∑ J − b 2 − n α ≤ 1 − 2 − α S − β ( S ℓ b ) α ,
where α + β = 1 \alpha+\beta=1 α + β = 1 was used in the last expression. Finally, splitting the sum defining K c b K_{cb} K c b at c = b c=b c = b gives
∑ c K c b = γ b ∑ c < b λ c + λ b ∑ c ≥ b γ c ≤ C λ b γ b + λ b γ ~ b ≤ C λ b γ ~ b , \sum_c K_{cb} = \gamma_b\sum_{c<b}\lambda_c+\lambda_b\sum_{c\ge b}\gamma_c \le C\lambda_b\gamma_b+\lambda_b\widetilde{\gamma}_b \le C\lambda_b\widetilde{\gamma}_b, c ∑ K c b = γ b c < b ∑ λ c + λ b c ≥ b ∑ γ c ≤ C λ b γ b + λ b γ b ≤ C λ b γ b ,
since γ b ≤ γ ~ b \gamma_b\le\widetilde{\gamma}_b γ b ≤ γ b . Thus
γ ~ b ≤ C S − β min ( 1 , ( ℓ b S ) α ) , ∑ c K c b ≤ C λ b γ ~ b . (68) \widetilde{\gamma}_b \le CS^{-\beta}\min\left(1,\left(\frac{\ell_b}{S}\right)^\alpha\right), \qquad\sum_c K_{cb}\le C\lambda_b\widetilde{\gamma}_b. \tag*{(68)} γ b ≤ C S − β min ( 1 , ( S ℓ b ) α ) , c ∑ K c b ≤ C λ b γ b . (68)
We next estimate the tails p ~ b \widetilde{p}_b p b and d ~ b \widetilde{d}_b d b . Suppose first that ℓ b ≥ S \ell_b\ge S ℓ b ≥ S . The number of indices a ≥ b a\ge b a ≥ b with ℓ a ≥ S \ell_a\ge S ℓ a ≥ S is at most
1 + ⌊ log 2 ( ℓ b / S ) ⌋ . 1+\left\lfloor\log_2(\ell_b/S)\right\rfloor. 1 + ⌊ log 2 ( ℓ b / S ) ⌋ .
For these indices, p a = 1 p_a=1 p a = 1 and γ ~ a ≤ C S − β \widetilde{\gamma}_a\le CS^{-\beta} γ a ≤ C S − β by (68) . For the remaining indices, ℓ a < S \ell_a<S ℓ a < S , we have
p a = ℓ a S , γ ~ a ≤ C S − β ( ℓ a S ) α . p_a=\frac{\ell_a}{S}, \qquad\widetilde{\gamma}_a\le CS^{-\beta}\left(\frac{\ell_a}{S}\right)^\alpha. p a = S ℓ a , γ a ≤ C S − β ( S ℓ a ) α .
Starting at the first such index, the ratios ℓ a / S \ell_a/S ℓ a / S are bounded by 1 , 2 − 1 , 2 − 2 , … 1,2^{-1},2^{-2},\ldots 1 , 2 − 1 , 2 − 2 , … . Consequently,
∑ a ≥ b ℓ a < S p a ≤ ∑ n ≥ 0 2 − n = 2 , ∑ a ≥ b ℓ a < S γ ~ a ≤ C S − β ∑ n ≥ 0 2 − n α ≤ C S − β . \sum_{\substack{a\ge b\\ \ell_a<S}}p_a\le\sum_{n\ge0}2^{-n}=2,\qquad\sum_{\substack{a\ge b\\ \ell_a<S}}\widetilde{\gamma}_a\le CS^{-\beta}\sum_{n\ge0}2^{-n\alpha}\le CS^{-\beta}. a ≥ b ℓ a < S ∑ p a ≤ n ≥ 0 ∑ 2 − n = 2 , a ≥ b ℓ a < S ∑ γ a ≤ C S − β n ≥ 0 ∑ 2 − n α ≤ C S − β .
Adding the contributions from scales above and below S S S gives
p ~ b ≤ C ( 1 + log ( ℓ b / S ) ) , d ~ b ≤ C S − β ( 1 + log ( ℓ b / S ) ) , ℓ b ≥ S . \widetilde{p}_b\le C(1+\log(\ell_b/S)),\qquad\widetilde{d}_b\le CS^{-\beta}(1+\log(\ell_b/S)),\qquad\ell_b\ge S. p b ≤ C ( 1 + log ( ℓ b / S )) , d b ≤ C S − β ( 1 + log ( ℓ b / S )) , ℓ b ≥ S .
If instead ℓ b < S \ell_b<S ℓ b < S , every index a ≥ b a\ge b a ≥ b satisfies ℓ a < S \ell_a<S ℓ a < S . Using ℓ b + n = 2 − n ℓ b \ell_{b+n}=2^{-n}\ell_b ℓ b + n = 2 − n ℓ b , we obtain
p ~ b = ℓ b S ∑ n = 0 J − b 2 − n ≤ 2 ℓ b S . \widetilde{p}_b=\frac{\ell_b}{S}\sum_{n=0}^{J-b}2^{-n}\le2\frac{\ell_b}{S}. p b = S ℓ b n = 0 ∑ J − b 2 − n ≤ 2 S ℓ b .
and, by (68) ,
d ~ b = ∑ a ≥ b γ ~ a ≤ C S − β ( ℓ b S ) α ∑ n = 0 J − b 2 − n α ≤ C S − β ( ℓ b S ) α . \widetilde{d}_b=\sum_{a\ge b}\widetilde{\gamma}_a\le CS^{-\beta}\left(\frac{\ell_b}{S}\right)^\alpha\sum_{n=0}^{J-b}2^{-n\alpha}\le CS^{-\beta}\left(\frac{\ell_b}{S}\right)^\alpha. d b = a ≥ b ∑ γ a ≤ C S − β ( S ℓ b ) α n = 0 ∑ J − b 2 − n α ≤ C S − β ( S ℓ b ) α .
Thus
p ~ b ≤ C { 1 + log ( ℓ b / S ) , ℓ b ≥ S , ℓ b / S , ℓ b < S , d ~ b ≤ C S − β { 1 + log ( ℓ b / S ) , ℓ b ≥ S , ( ℓ b / S ) α , ℓ b < S . (69) \widetilde{p}_b\le C
\begin{cases}
1+\log(\ell_b/S),&\ell_b\ge S,\\
\ell_b/S,&\ell_b<S,
\end{cases}
\qquad
\widetilde{d}_b\le CS^{-\beta}
\begin{cases}
1+\log(\ell_b/S),&\ell_b\ge S,\\
(\ell_b/S)^\alpha,&\ell_b<S.
\end{cases} \tag*{(69)} p b ≤ C { 1 + log ( ℓ b / S ) , ℓ b / S , ℓ b ≥ S , ℓ b < S , d b ≤ C S − β { 1 + log ( ℓ b / S ) , ( ℓ b / S ) α , ℓ b ≥ S , ℓ b < S . (69)
If r < ℓ J r<\ell_J r < ℓ J , then S = ℓ J S=\ell_J S = ℓ J and every available scale satisfies ℓ a ≥ S \ell_a\ge S ℓ a ≥ S ; the sums over ℓ a < S \ell_a<S ℓ a < S are empty, and the same bounds apply.
For I = { b : ℓ b ≤ 3 L S } \mathcal{I}=\{b:\ell_b\le3LS\} I = { b : ℓ b ≤ 3 L S } , split the sums into S ≤ ℓ b ≤ 3 L S S\le\ell_b\le3LS S ≤ ℓ b ≤ 3 L S and ℓ b < S \ell_b<S ℓ b < S . There are at most 1 + ⌊ log 2 ( 3 L ) ⌋ ≤ C N L 1+\lfloor\log_2(3L)\rfloor\le CN_L 1 + ⌊ log 2 ( 3 L )⌋ ≤ C N L indices in the first range. For each such index,
p b = 1 , p ~ b ≤ C N L , γ ~ b ≤ C S − β , d ~ b ≤ C S − β N L , p_b=1,\qquad\widetilde{p}_b\le CN_L,\qquad\widetilde{\gamma}_b\le CS^{-\beta},\qquad\widetilde{d}_b\le CS^{-\beta}N_L, p b = 1 , p b ≤ C N L , γ b ≤ C S − β , d b ≤ C S − β N L ,
by (68) and (69) . In the second range, the same estimates give
p b + p ~ b ≤ C ℓ b S , γ ~ b + d ~ b ≤ C S − β ( ℓ b S ) α . p_b+\widetilde{p}_b\le C\frac{\ell_b}{S},\qquad\widetilde{\gamma}_b+\widetilde{d}_b\le CS^{-\beta}\left(\frac{\ell_b}{S}\right)^\alpha. p b + p b ≤ C S ℓ b , γ b + d b ≤ C S − β ( S ℓ b ) α .
Summing these geometric tails contributes at most C C C and C S − β CS^{-\beta} C S − β , respectively. Combining the two ranges, and using N L ≥ 1 N_L\ge1 N L ≥ 1 , yields
∑ b ∈ I p b ≤ C N L , ∑ b ∈ I p ~ b ≤ C N L 2 , ∑ b ∈ I γ ~ b ≤ C S − β N L , ∑ b ∈ I d ~ b ≤ C S − β N L 2 . (70) \sum_{b\in\mathcal{I}}p_b\le CN_L,\qquad\sum_{b\in\mathcal{I}}\widetilde{p}_b\le CN_L^2,\qquad\sum_{b\in\mathcal{I}}\widetilde{\gamma}_b\le CS^{-\beta}N_L,\qquad\sum_{b\in\mathcal{I}}\widetilde{d}_b\le CS^{-\beta}N_L^2. \tag*{(70)} b ∈ I ∑ p b ≤ C N L , b ∈ I ∑ p b ≤ C N L 2 , b ∈ I ∑ γ b ≤ C S − β N L , b ∈ I ∑ d b ≤ C S − β N L 2 . (70)
By Lemma 4.6(i), every summand in (64) –(66) with b ∉ I b\notin\mathcal{I} b ∈ / I vanishes, so all sums over b b b may be restricted to I \mathcal{I} I . Using h b ≤ κ / λ b h_b\le\kappa/\lambda_b h b ≤ κ / λ b , (58) , and (70) , we obtain
∣ ∑ b h b B b Φ b ∣ ≤ C S β ∑ b ∈ I h b γ b ≤ C κ S β ∑ b ∈ I γ b λ b = C κ S β ∑ b ∈ I p b ≤ C κ N L S β . \left|\sum_b h_b\mathcal{B}_b\Phi_b\right|\le CS^\beta\sum_{b\in\mathcal{I}}h_b\gamma_b\le C\kappa S^\beta\sum_{b\in\mathcal{I}}\frac{\gamma_b}{\lambda_b}=C\kappa S^\beta\sum_{b\in\mathcal{I}}p_b\le C\kappa N_LS^\beta. b ∑ h b B b Φ b ≤ C S β b ∈ I ∑ h b γ b ≤ C κ S β b ∈ I ∑ λ b γ b = C κ S β b ∈ I ∑ p b ≤ C κ N L S β .
This proves (64) . Splitting into a < b a<b a < b and a ≥ b a\ge b a ≥ b in (63) gives
∑ a q a b K a b ≤ C κ 2 ( p b + p ~ b ) ≤ C κ 2 p ~ b . (71) \sum_a q_{ab}K_{ab}\le C\kappa^2(p_b+\widetilde{p}_b)\le C\kappa^2\widetilde{p}_b. \tag*{(71)} a ∑ q ab K ab ≤ C κ 2 ( p b + p b ) ≤ C κ 2 p b . (71)
Indeed the first part is bounded by κ 2 λ b 2 γ b ∑ a < b λ a \kappa^2\lambda_b^2\gamma_b\sum_{a<b}\lambda_a κ 2 λ b 2 γ b ∑ a < b λ a , and the second is bounded by κ 2 ∑ a ≥ b γ a / λ a \kappa^2\sum_{a\ge b}\gamma_a/\lambda_a κ 2 ∑ a ≥ b γ a / λ a . Together with (59) and the bound on ∑ b ∈ T p ~ b \sum_{b\in T}\tilde p_b ∑ b ∈ T p ~ b in (70) , this proves (65) .
To verify the cubic bound, fix b b b and sum the four terms of (60) separately. We use the consequence of (63)
q a b ≤ { κ 2 / λ b 2 , a < b , κ 2 / ( λ b λ a ) , a ≥ b . q_{ab}\le\begin{cases}
\kappa^2/\lambda_b^2, & a<b,\\
\kappa^2/(\lambda_b\lambda_a), & a\ge b.
\end{cases} q ab ≤ { κ 2 / λ b 2 , κ 2 / ( λ b λ a ) , a < b , a ≥ b .
For the first term, (67) and (71) give directly
∑ a , c q a b γ c K a b = ( ∑ c γ c ) ( ∑ a q a b K a b ) ≤ C κ 2 S − β p ~ b . \sum_{a,c}q_{ab}\gamma_cK_{ab}
=\left(\sum_c\gamma_c\right)\left(\sum_aq_{ab}K_{ab}\right)
\le C\kappa^2S^{-\beta}\tilde p_b. a , c ∑ q ab γ c K ab = ( c ∑ γ c ) ( a ∑ q ab K ab ) ≤ C κ 2 S − β p ~ b .
For the second term, first sum in c c c using (68) , and then split the a a a -sum:
∑ a , c q a b γ b K c b ≤ C λ b γ ~ b ∑ a q a b γ a ≤ C κ 2 γ ~ b ( 1 λ b ∑ a < b γ a + ∑ a ≥ b γ a λ a ) ≤ C κ 2 γ ~ b ( 1 + p ~ b ) . \begin{aligned}
\sum_{a,c}q_{ab}\gamma_bK_{cb}
\le C\lambda_b\tilde\gamma_b\sum_aq_{ab}\gamma_a \\
&\le C\kappa^2\tilde\gamma_b\left(\frac{1}{\lambda_b}\sum_{a<b}\gamma_a+\sum_{a\ge b}\frac{\gamma_a}{\lambda_a}\right)
\le C\kappa^2\tilde\gamma_b(1+\tilde p_b).
\end{aligned} a , c ∑ q ab γ b K c b ≤ C λ b γ ~ b a ∑ q ab γ a ≤ C κ 2 γ ~ b ( λ b 1 a < b ∑ γ a + a ≥ b ∑ λ a γ a ) ≤ C κ 2 γ ~ b ( 1 + p ~ b ) .
Here ∑ a < b γ a ≤ C λ b \sum_{a<b}\gamma_a\le C\lambda_b ∑ a < b γ a ≤ C λ b by (67) , and the second sum is p ~ b \tilde p_b p ~ b .
For the third term, (68) gives ∑ c K c a ≤ C λ a γ ~ a \sum_cK_{ca}\le C\lambda_a\tilde\gamma_a ∑ c K c a ≤ C λ a γ ~ a . Thus
∑ a , c q a b γ b K c a ≤ C γ b ∑ a q a b λ a γ ~ a ≤ C κ 2 γ b ( 1 λ b 2 ∑ a < b λ a γ ~ a + 1 λ b ∑ a ≥ b γ ~ a ) ≤ C κ 2 ( S − β p b + p b d ˙ b ) . \begin{aligned}
\sum_{a,c}q_{ab}\gamma_bK_{ca}
\le C\gamma_b\sum_aq_{ab}\lambda_a\tilde\gamma_a \\
&\le C\kappa^2\gamma_b\left(\frac{1}{\lambda_b^2}\sum_{a<b}\lambda_a\tilde\gamma_a+\frac{1}{\lambda_b}\sum_{a\ge b}\tilde\gamma_a\right) \\
&\le C\kappa^2\left(S^{-\beta}p_b+p_b\dot d_b\right).
\end{aligned} a , c ∑ q ab γ b K c a ≤ C γ b a ∑ q ab λ a γ ~ a ≤ C κ 2 γ b ( λ b 2 1 a < b ∑ λ a γ ~ a + λ b 1 a ≥ b ∑ γ ~ a ) ≤ C κ 2 ( S − β p b + p b d ˙ b ) .
In the last step we used γ ~ a ≤ C S − β \tilde\gamma_a\le CS^{-\beta} γ ~ a ≤ C S − β , ∑ a < b λ a ≤ C λ b \sum_{a<b}\lambda_a\le C\lambda_b ∑ a < b λ a ≤ C λ b , γ b / λ b = p b \gamma_b/\lambda_b=p_b γ b / λ b = p b , and ∑ a ≥ b γ ~ a = d ˙ b \sum_{a\ge b}\tilde\gamma_a=\dot d_b ∑ a ≥ b γ ~ a = d ˙ b .
Finally, splitting the fourth term at a = b a=b a = b ,
∑ a , c q a b λ min ( a , b ) κ max ( a , b ) = ∑ a < b q a b λ a ∑ c K c b + ∑ a ≥ b q a b λ b ∑ c K c a ≤ C ∑ a < b κ 2 λ b 2 λ a λ b γ ~ b + C ∑ a ≥ b κ 2 λ b λ a λ b λ a γ ~ a = C κ 2 ( γ ~ b λ b ∑ a < b λ a + ∑ a ≥ b γ ~ a ) ≤ C κ 2 ( γ ~ b + d ˙ b ) . \begin{aligned}
\sum_{a,c}q_{ab}\lambda_{\min(a,b)}\kappa_{\max(a,b)} \\
&=\sum_{a<b}q_{ab}\lambda_a\sum_cK_{cb}+\sum_{a\ge b}q_{ab}\lambda_b\sum_cK_{ca} \\
&\le C\sum_{a<b}\frac{\kappa^2}{\lambda_b^2}\lambda_a\lambda_b\tilde\gamma_b
+C\sum_{a\ge b}\frac{\kappa^2}{\lambda_b\lambda_a}\lambda_b\lambda_a\tilde\gamma_a \\
&=C\kappa^2\left(\frac{\tilde\gamma_b}{\lambda_b}\sum_{a<b}\lambda_a+\sum_{a\ge b}\tilde\gamma_a\right) \\
&\le C\kappa^2(\tilde\gamma_b+\dot d_b).
\end{aligned} a , c ∑ q ab λ m i n ( a , b ) κ m a x ( a , b ) = a < b ∑ q ab λ a c ∑ K c b + a ≥ b ∑ q ab λ b c ∑ K c a ≤ C a < b ∑ λ b 2 κ 2 λ a λ b γ ~ b + C a ≥ b ∑ λ b λ a κ 2 λ b λ a γ ~ a = C κ 2 ( λ b γ ~ b a < b ∑ λ a + a ≥ b ∑ γ ~ a ) ≤ C κ 2 ( γ ~ b + d ˙ b ) .
The first inequality uses q a b ≤ κ 2 / λ b 2 q_{ab}\le\kappa^2/\lambda_b^2 q ab ≤ κ 2 / λ b 2 when a < b a<b a < b and q a b ≤ κ 2 / ( λ b λ a ) q_{ab}\le\kappa^2/(\lambda_b\lambda_a) q ab ≤ κ 2 / ( λ b λ a ) when a ≥ b a\ge b a ≥ b , together with ∑ c K c b ≤ C λ b γ ~ b \sum_cK_{cb}\le C\lambda_b\tilde\gamma_b ∑ c K c b ≤ C λ b γ ~ b and ∑ c K c a ≤ C λ a γ ~ a \sum_cK_{ca}\le C\lambda_a\tilde\gamma_a ∑ c K c a ≤ C λ a γ ~ a . The last inequality follows from ∑ a < b λ a ≤ C λ b \sum_{a<b}\lambda_a\le C\lambda_b ∑ a < b λ a ≤ C λ b and the definition d ˙ b = ∑ a ≥ b γ ~ a \dot d_b=\sum_{a\ge b}\tilde\gamma_a d ˙ b = ∑ a ≥ b γ ~ a .
Sum in b ∈ I b \in\mathcal{I} b ∈ I , use γ ˙ b ≤ C S − β \dot{\gamma}_b \leq C S^{-\beta} γ ˙ b ≤ C S − β , p b ≤ 1 p_b \leq1 p b ≤ 1 , and (70) . The result is C κ 2 S − β N L 2 C\kappa^2 S^{-\beta}N_L^2 C κ 2 S − β N L 2 . Multiplication by S β S^\beta S β in (60) cancels S − β S^{-\beta} S − β , proving (66) .
The weights h b h_b h b and q a b q_{ab} q ab satisfy D h h b = − 1 \mathcal{D}_h h_b=-1 D h h b = − 1 and D h q a b = − h b \mathcal{D}_h q_{ab}=-h_b D h q ab = − h b , which produces the cancellation of the first- and second-order terms in Proposition 5.2, and q a b q_{ab} q ab vanishes whenever one of the two windows has just ended, which gives the jumps at a refresh the form of Proposition 6.6. Define the corrected observable, a function of an admissible configuration S \mathcal{S} S and a phase vector h h h ,
Q ( S , h ) = Φ + ∑ b h b B b Φ b + ∑ a , b q a b B a B b Φ b . (72) Q(\mathcal{S},h)=\Phi+\sum_b h_b\mathcal{B}_b\Phi_b+\sum_{a,b}q_{ab}\mathcal{B}_a\mathcal{B}_b\Phi_b. \tag*{(72)} Q ( S , h ) = Φ + b ∑ h b B b Φ b + a , b ∑ q ab B a B b Φ b . (72)
where Φ = Φ ( z ( x , x ~ ) ) \Phi=\Phi(z(x,\widetilde{x})) Φ = Φ ( z ( x , x )) , the operator expressions are evaluated at S \mathcal{S} S as in Lemma 4.6, and q a b = q a b ( h ) q_{ab}=q_{ab}(h) q ab = q ab ( h ) . Along the process we write Q t = Q ( S t , h ( t ) ) Q_t=Q(\mathcal{S}_t,h(t)) Q t = Q ( S t , h ( t )) , with S t \mathcal{S}_t S t the configuration of Remark 4.5 and h ( t ) h(t) h ( t ) the current phases; Q t Q_t Q t is defined at every t t t at which R t ≤ 1 / 4 R_t\leq1/4 R t ≤ 1/4 , and Q t − Q_{t-} Q t − denotes the same expression with the left limits of the window data and phases. At t = 1 t=1 t = 1 , set Q 1 = Q 1 − Q_1=Q_{1-} Q 1 = Q 1 − ; no renewal is performed at 1.
Proposition 5.2 (Evolution between renewals). As functions of an admissible configuration S \mathcal{S} S and a phase vector h h h , with D h \mathcal{D}_h D h the phase-flow derivative defined in (62) and B \mathcal{B} B the operator of Definition 4.2,
( D h + B ) Q = E 0 + E 3 , E 0 = ∑ j B j ( Φ − Φ j ) , E 3 = ∑ a , b , c q a b B c B a B b Φ b . (73) (\mathcal{D}_h+\mathcal{B})Q=E_0+E_3,\qquad E_0=\sum_j\mathcal{B}_j(\Phi-\Phi_j),\qquad E_3=\sum_{a,b,c}q_{ab}\mathcal{B}_c\mathcal{B}_a\mathcal{B}_b\Phi_b. \tag*{(73)} ( D h + B ) Q = E 0 + E 3 , E 0 = j ∑ B j ( Φ − Φ j ) , E 3 = a , b , c ∑ q ab B c B a B b Φ b . (73) Moreover, writing S = ( x , x ~ , Θ ) \mathcal{S}=(x,\widetilde{x},\Theta) S = ( x , x , Θ ) and r = dist ( x , x ~ ) r=\operatorname{dist}(x,\widetilde{x}) r = dist ( x , x ) ,
∣ E 0 ∣ ≤ C L − β , ∣ E 3 ∣ ≤ C κ 2 N L 2 , (74) |E_0|\leq C L^{-\beta},\qquad|E_3|\leq C\kappa^2N_L^2, \tag*{(74)} ∣ E 0 ∣ ≤ C L − β , ∣ E 3 ∣ ≤ C κ 2 N L 2 , (74) ∣ Q − Φ ∣ ≤ C ( κ N L + κ 2 N L 2 ) ( r β + ℓ j β ) . (75) |Q-\Phi|\leq C(\kappa N_L+\kappa^2N_L^2)(r^\beta+\ell_j^\beta). \tag*{(75)} ∣ Q − Φ∣ ≤ C ( κ N L + κ 2 N L 2 ) ( r β + ℓ j β ) . (75) Consequently, on every open interval ( s 1 , s 2 ) (s_1,s_2) ( s 1 , s 2 ) containing no renewal time of any scale and on which R t ≤ 1 / 4 R_t\leq1/4 R t ≤ 1/4 ,
d d t Q t = ( E 0 + E 3 ) ( S t , h ( t ) ) , s 1 < t < s 2 . (76) \frac{\mathrm{d}}{\mathrm{d}t}Q_t=(E_0+E_3)(\mathcal{S}_t,h(t)),\qquad s_1<t<s_2. \tag*{(76)} d t d Q t = ( E 0 + E 3 ) ( S t , h ( t )) , s 1 < t < s 2 . (76)
Proof. The identity (73) is algebraic. The operator B \mathcal{B} B acts on the configuration variables only and D h \mathcal{D}_h D h on the phases only; Φ \Phi Φ and the operator expressions in (72) do not depend on h h h , and h b h_b h b , q a b q_{ab} q ab do not depend on S \mathcal{S} S . Applying D h \mathcal{D}_h D h to (72) with D h h b = − 1 \mathcal{D}_h h_b=-1 D h h b = − 1 and D h q a b = − h b \mathcal{D}_h q_{ab}=-h_b D h q ab = − h b from (63) , and applying B = ∑ a B a \mathcal{B}=\sum_a\mathcal{B}_a B = ∑ a B a to (72) , gives
− ∑ b B b Φ b − ∑ a , b h b B a B b Φ b + B Φ + ∑ a , b h b B a B b Φ b + E 3 . -\sum_b\mathcal{B}_b\Phi_b-\sum_{a,b}h_b\mathcal{B}_a\mathcal{B}_b\Phi_b+\mathcal{B}\Phi+\sum_{a,b}h_b\mathcal{B}_a\mathcal{B}_b\Phi_b+E_3. − b ∑ B b Φ b − a , b ∑ h b B a B b Φ b + B Φ + a , b ∑ h b B a B b Φ b + E 3 .
which is (73) . In particular all mixed second-order terms cancel exactly. Since Φ j ′ = w j Φ ′ \Phi'_j=w_j\Phi' Φ j ′ = w j Φ ′ , the first identity in (56) (applied to Φ j \Phi_j Φ j and to Φ \Phi Φ ) gives B j Φ j = w j ( r ) B j Φ \mathcal{B}_j\Phi_j=w_j(r)\mathcal{B}_j\Phi B j Φ j = w j ( r ) B j Φ , so the first-order error is E 0 = ∑ j ( 1 − w j ) B j Φ E_0=\sum_j(1-w_j)\mathcal{B}_j\Phi E 0 = ∑ j ( 1 − w j ) B j Φ . Its nonzero summands have ℓ j > L S ~ ≥ L S ≥ r \ell_j>L\widetilde{S}\geq LS\geq r ℓ j > L S ≥ L S ≥ r , so γ j = λ j \gamma_j=\lambda_j γ j = λ j , and Lemma 4.6(ii) gives
∣ E 0 ∣ ≤ C S β ∑ ℓ j > L S ℓ j − β ≤ C L − β . |E_0|\leq C S^\beta\sum_{\ell_j>LS}\ell_j^{-\beta}\leq C L^{-\beta}. ∣ E 0 ∣ ≤ C S β ℓ j > L S ∑ ℓ j − β ≤ C L − β .
The bound on E 3 E_3 E 3 is (66) , and (75) follows from (64) –(65) and S β ≤ r β + ℓ j β S^\beta\leq r^\beta+\ell_j^\beta S β ≤ r β + ℓ j β .
For (76) , fix such an interval. On this interval, the active coefficients are constant, and S t S_t S t is admissible by Remark 4.5 . Writing P i m ( t ) P_i^m(t) P i m ( t ) for the active center of generation m m m , the particle and center equations read
x ˙ t = ∑ a u a ( x t ; Θ t ) , x ~ ˙ t = ∑ a u a ( x ~ t ; Θ t ) , P ˙ i m ( t ) = ∑ a < m u a ( P i m ( t ) ; Θ t ) . \dot{x}_t=\sum_a u_a(x_t;\Theta_t),\qquad\dot{\widetilde{x}}_t=\sum_a u_a(\widetilde{x}_t;\Theta_t),\qquad\dot{P}_i^m(t)=\sum_{a<m}u_a(P_i^m(t);\Theta_t). x ˙ t = a ∑ u a ( x t ; Θ t ) , x ˙ t = a ∑ u a ( x t ; Θ t ) , P ˙ i m ( t ) = a < m ∑ u a ( P i m ( t ) ; Θ t ) .
The fields depend on time through the moving centers. Between renewals, however, the current particle positions, center positions, and fixed active coefficients determine all velocities: there is no additional dependence on the time variable. In other words, the configuration equations have the form S ˙ t = V ( S t ) \dot{S}_t=V(S_t) S ˙ t = V ( S t ) for a fixed vector field V V V on each interval between renewals.
For any C 1 C^1 C 1 function F F F of the configuration, the chain rule therefore gives
d d t F ( S t ) = ∑ a u a ( x t ; Θ t ) ⋅ ∇ x F ( S t ) + ∑ a u a ( x ~ t ; Θ t ) ⋅ ∇ x ~ F ( S t ) + ∑ m ∑ i ∈ I m ∑ a < m u a ( P i m ( t ) ; Θ t ) ⋅ ∇ P i m F ( S t ) = ∑ a [ u a ( x t ; Θ t ) ⋅ ∇ x F ( S t ) + u a ( x ~ t ; Θ t ) ⋅ ∇ x ~ F ( S t ) + ∑ m > a ∑ i ∈ I m u a ( P i m ( t ) ; Θ t ) ⋅ ∇ P i m F ( S t ) ] = ∑ a B a F ( S t ) = B F ( S t ) . \begin{aligned}
\frac{d}{dt}F(S_t)
={}&\sum_a u_a(x_t;\Theta_t)\cdot\nabla_xF(S_t)+\sum_a u_a(\widetilde{x}_t;\Theta_t)\cdot\nabla_{\widetilde{x}}F(S_t)\\
&+\sum_m\sum_{i\in I_m}\sum_{a<m}u_a(P_i^m(t);\Theta_t)\cdot\nabla_{P_i^m}F(S_t)\\
={}&\sum_a\left[u_a(x_t;\Theta_t)\cdot\nabla_xF(S_t)+u_a(\widetilde{x}_t;\Theta_t)\cdot\nabla_{\widetilde{x}}F(S_t)\right.\\
&\left.\qquad+\sum_{m>a}\sum_{i\in I_m}u_a(P_i^m(t);\Theta_t)\cdot\nabla_{P_i^m}F(S_t)\right]\\
={}&\sum_a B_aF(S_t)=BF(S_t).
\end{aligned} d t d F ( S t ) = = = a ∑ u a ( x t ; Θ t ) ⋅ ∇ x F ( S t ) + a ∑ u a ( x t ; Θ t ) ⋅ ∇ x F ( S t ) + m ∑ i ∈ I m ∑ a < m ∑ u a ( P i m ( t ) ; Θ t ) ⋅ ∇ P i m F ( S t ) a ∑ [ u a ( x t ; Θ t ) ⋅ ∇ x F ( S t ) + u a ( x t ; Θ t ) ⋅ ∇ x F ( S t ) + m > a ∑ i ∈ I m ∑ u a ( P i m ( t ) ; Θ t ) ⋅ ∇ P i m F ( S t ) ] a ∑ B a F ( S t ) = B F ( S t ) .
The penultimate equality is precisely Definition 4.2 , applied to a function independent of the query point. There are no contributions from derivatives with respect to the coefficients, since those are constant on this interval. The phases satisfy h ˙ j = − 1 \dot{h}_j=-1 h ˙ j = − 1 .
Since Q Q Q is C 1 C^1 C 1 in the configuration (the operator expressions in (72) are at least C 2 C^2 C 2 , because B a u b ∈ C 2 B_au_b\in C^2 B a u b ∈ C 2 by Definition 4.2 ) and C 1 C^1 C 1 in h h h , the chain rule gives
d d t Q ( S t , h ( t ) ) = ( ( D h + B ) Q ) ( S t , h ( t ) ) , \frac{d}{dt}Q(S_t,h(t))=((D_h+B)Q)(S_t,h(t)), d t d Q ( S t , h ( t )) = (( D h + B ) Q ) ( S t , h ( t )) ,
and (73) completes the proof. □
Correctors for the quadratic-core observable
Let Q ♯ Q^\sharp Q ♯ be the observable (72) with ( Φ ♯ , Φ j ♯ ) (\Phi^\sharp,\Phi_j^\sharp) ( Φ ♯ , Φ j ♯ ) in place of ( Φ , Φ j ) (\Phi,\Phi_j) ( Φ , Φ j ) . The common corrector estimates apply to these functions as well.
Corollary 5.3. Lemma 4.6, Lemma 5.1 and Proposition 5.2 hold with ( Φ ♯ , Φ j ♯ , Q ♯ ) (\Phi^\sharp,\Phi_j^\sharp,Q^\sharp) ( Φ ♯ , Φ j ♯ , Q ♯ ) in place of ( Φ , Φ j , Q ) (\Phi,\Phi_j,Q) ( Φ , Φ j , Q ) , with constants depending in addition on Υ \Upsilon Υ , at every admissible configuration with 0 < r ≤ 1 / 4 0<r\leq1/4 0 < r ≤ 1/4 , for the forward and reversed evolutions. All correction terms of Q ♯ Q^\sharp Q ♯ , as well as E 0 E_0 E 0 and E 3 E_3 E 3 , are continuous up to x = x ~ x=\widetilde{x} x = x and vanish there; moreover, Q ♯ − Φ ♯ Q^\sharp-\Phi^\sharp Q ♯ − Φ ♯ is C 2 C^2 C 2 in ( x , x ~ ) (x,\widetilde{x}) ( x , x ) on { dist ( x , x ~ ) < 1 / 4 } \{\operatorname{dist}(x,\widetilde{x})<1/4\} { dist ( x , x ) < 1/4 } , the diagonal included.
Proof. The proofs of these results use the test functions only through (30) with k = 1 k=1 k = 1 , (32) with k ≤ 3 k\leq3 k ≤ 3 , the identity Φ j ′ = w j Φ ′ \Phi_j^\prime=w_j\Phi^\prime Φ j ′ = w j Φ ′ and the support property (i) of
Lemma 4.6. Since m ♯ ≥ r m_{\sharp} \ge r m ♯ ≥ r , (3.14) implies these bounds, and ( Φ b ♯ ) ′ = w j ( Φ b ♯ ) ′ (\Phi^{\sharp}_{b})' = w_j(\Phi^{\sharp}_{b})' ( Φ b ♯ ) ′ = w j ( Φ b ♯ ) ′ by definition. The phases satisfy 0 ≤ h b ≤ τ b ≤ κ / λ b 0 \le h_b \le\tau_b \le\kappa/\lambda_b 0 ≤ h b ≤ τ b ≤ κ / λ b in both evolutions. Every term in (56) , (57) and in E 0 = ∑ j ( 1 − w j ) ∇ Φ ♯ ( z ) ⋅ δ u j E_0 = \sum_j(1-w_j)\nabla\Phi^{\sharp}(z)\cdot\delta u_j E 0 = ∑ j ( 1 − w j ) ∇ Φ ♯ ( z ) ⋅ δ u j is a product of derivatives of Φ b ♯ \Phi_b^{\sharp} Φ b ♯ or Φ ♯ \Phi^{\sharp} Φ ♯ at z z z , which are continuous by Lemma 3.3(b), and of at least one two-point difference, which is continuous and vanishes when x = x ~ x=\widetilde{x} x = x . Finally, each correction term is a product of derivatives of Φ b ♯ \Phi_b^{\sharp} Φ b ♯ at z z z , which are smooth on { ∣ z ∣ < 1 / 4 } \{|z|<1/4\} { ∣ z ∣ < 1/4 } by Lemma 3.3(b), with values at x x x or x ~ \widetilde{x} x of the fields u b ∈ C 3 u_b\in C^3 u b ∈ C 3 and B a u b ∈ C 2 \mathcal{B}_a u_b\in C^2 B a u b ∈ C 2 (Definition 4.2); hence Q ♯ − Φ ♯ Q^{\sharp}-\Phi^{\sharp} Q ♯ − Φ ♯ is C 2 C^2 C 2 in ( x , x ~ ) (x,\widetilde{x}) ( x , x ) up to the diagonal.
Diffusion acting on the correctors
Write Y = ( x , x ~ ) Y=(x,\widetilde{x}) Y = ( x , x ) and Δ Y = Δ x + Δ x ~ \Delta_Y=\Delta_x+\Delta_{\widetilde{x}} Δ Y = Δ x + Δ x . The principal diffusive term, 2 μ Δ Φ ♯ 2\mu\Delta\Phi^{\sharp} 2 μ Δ Φ ♯ , is nonnegative. We must show that diffusion applied to the remaining terms in Q ♯ Q^{\sharp} Q ♯ is smaller than the gain of order κ \kappa κ . The following estimate is uniform in the finest scale, including at the diagonal.
Lemma 5.4. Assume (39) , ε 0 ≤ κ \varepsilon_0\le\kappa ε 0 ≤ κ and κ N L ≤ 1 \kappa N_L\le1 κ N L ≤ 1 . There is C 9 C_9 C 9 , independent of J , L , κ , ε 0 J,L,\kappa,\varepsilon_0 J , L , κ , ε 0 , such that at every configuration satisfying (C1), (C2) and 0 ≤ r ≤ 1 / 4 0\le r\le1/4 0 ≤ r ≤ 1/4 (the pair Y = ( y , y ′ ) Y=(y,y') Y = ( y , y ′ ) playing the role of the particles ( x , x ~ ) (x,\widetilde{x}) ( x , x ) ),
μ ∣ Δ Y ( Q ♯ − Φ ♯ ) ∣ ≤ C 9 κ 2 N L . \mu\lvert\Delta_Y(Q^{\sharp}-\Phi^{\sharp})\rvert\le C_9\kappa^2N_L. μ ∣ Δ Y ( Q ♯ − Φ ♯ )∣ ≤ C 9 κ 2 N L .
Proof. We estimate the first and second correctors separately, sum over the scales, and use the definition of r ♯ r_{\sharp} r ♯ to control the result near the diagonal.
Step 1: product formulas for the two-point Laplacian. For a function G G G of the query point and a function f f f of z z z ,
Δ Y δ G = δ ( Δ G ) , Δ Y f ( z ) = 2 Δ f ( z ) , ∇ x ~ f ( z ) = − ∇ x f ( z ) , \Delta_Y\delta G=\delta(\Delta G),\qquad\Delta_Y f(z)=2\Delta f(z),\qquad\nabla_{\widetilde{x}}f(z)=-\nabla_x f(z), Δ Y δ G = δ ( Δ G ) , Δ Y f ( z ) = 2Δ f ( z ) , ∇ x f ( z ) = − ∇ x f ( z ) ,
∇ x δ G = ∇ G ( x ) , ∇ x ~ δ G = − ∇ G ( x ~ ) . \nabla_x\delta G=\nabla G(x),\qquad\nabla_{\widetilde{x}}\delta G=-\nabla G(\widetilde{x}). ∇ x δ G = ∇ G ( x ) , ∇ x δ G = − ∇ G ( x ) .
The product rule gives, for a vector field f ( z ) f(z) f ( z ) and a bilinear form Λ ( z ) \Lambda(z) Λ ( z ) ,
Δ Y [ f ⋅ δ G ] = 2 Δ f ⋅ δ G + f ⋅ δ ( Δ G ) + 2 ∑ k ∂ k f ⋅ ( ∂ k G ( x ) + ∂ k G ( x ~ ) ) , \Delta_Y[f\cdot\delta G]=2\Delta f\cdot\delta G+f\cdot\delta(\Delta G)+2\sum_k\partial_k f\cdot\bigl(\partial_kG(x)+\partial_kG(\widetilde{x})\bigr), Δ Y [ f ⋅ δ G ] = 2Δ f ⋅ δ G + f ⋅ δ ( Δ G ) + 2 k ∑ ∂ k f ⋅ ( ∂ k G ( x ) + ∂ k G ( x ) ) ,
Δ Y ( Λ [ δ G 1 , δ G 2 ] ) = 2 Δ Λ [ δ G 1 , δ G 2 ] + Λ [ δ Δ G 1 , δ G 2 ] + Λ [ δ G 1 , δ Δ G 2 ] + 2 ∑ k ∂ k Λ [ ∂ k G 1 ( x ) + ∂ k G 1 ( x ~ ) , δ G 2 ] + 2 ∑ k ∂ k Λ [ δ G 1 , ∂ k G 2 ( x ) + ∂ k G 2 ( x ~ ) ] + 2 ∑ k ( Λ ( z ) [ ∂ k G 1 ( x ) , ∂ k G 2 ( x ) ] + Λ ( z ) [ ∂ k G 1 ( x ~ ) , ∂ k G 2 ( x ~ ) ] ) . \begin{aligned}
\Delta_Y\bigl(\Lambda[\delta G_1,\delta G_2]\bigr)
={}&2\Delta\Lambda[\delta G_1,\delta G_2]+\Lambda[\delta\Delta G_1,\delta G_2]+\Lambda[\delta G_1,\delta\Delta G_2]\\
&+2\sum_k\partial_k\Lambda[\partial_kG_1(x)+\partial_kG_1(\widetilde{x}),\delta G_2]\\
&+2\sum_k\partial_k\Lambda[\delta G_1,\partial_kG_2(x)+\partial_kG_2(\widetilde{x})]\\
&+2\sum_k\bigl(\Lambda(z)[\partial_kG_1(x),\partial_kG_2(x)]+\Lambda(z)[\partial_kG_1(\widetilde{x}),\partial_kG_2(\widetilde{x})]\bigr).
\end{aligned} Δ Y ( Λ [ δ G 1 , δ G 2 ] ) = 2ΔΛ [ δ G 1 , δ G 2 ] + Λ [ δ Δ G 1 , δ G 2 ] + Λ [ δ G 1 , δ Δ G 2 ] + 2 k ∑ ∂ k Λ [ ∂ k G 1 ( x ) + ∂ k G 1 ( x ) , δ G 2 ] + 2 k ∑ ∂ k Λ [ δ G 1 , ∂ k G 2 ( x ) + ∂ k G 2 ( x )] + 2 k ∑ ( Λ ( z ) [ ∂ k G 1 ( x ) , ∂ k G 2 ( x )] + Λ ( z ) [ ∂ k G 1 ( x ) , ∂ k G 2 ( x )] ) .
Step 2: derivative bounds for the velocity terms. Write m j = min ( 1 , r / ℓ j ) m_j=\min(1,r/\ell_j) m j = min ( 1 , r / ℓ j ) , as in the proof of Lemma 4.3, and let ℓ ∨ = min ( ℓ a , ℓ b ) \ell_{\vee}=\min(\ell_a,\ell_b) ℓ ∨ = min ( ℓ a , ℓ b ) . From (41) , (50) and (46) , we have
∥ ∇ u a ∥ ≤ C λ a , ∣ δ u a ∣ ≤ C r γ a ≤ C r λ a , \|\nabla u_a\|\le C\lambda_a,\qquad|\delta u_a|\le Cr^{\gamma_a}\le Cr\lambda_a, ∥∇ u a ∥ ≤ C λ a , ∣ δ u a ∣ ≤ C r γ a ≤ C r λ a ,
∣ δ Δ u a ∣ ≤ C λ a ℓ a − 1 m a , ∣ δ F a b ∣ ≤ C r K a b ≤ C r λ a λ b . |\delta\Delta u_a|\le C\lambda_a\ell_a^{-1}m_a,\qquad|\delta F_{ab}|\le CrK_{ab}\le Cr\lambda_a\lambda_b. ∣ δ Δ u a ∣ ≤ C λ a ℓ a − 1 m a , ∣ δ F ab ∣ ≤ C r K ab ≤ C r λ a λ b .
We also need the first two spatial derivatives of F a b F_{ab} F ab . If a < b a<b a < b , (52) gives
∥ ∇ F a b ∥ ≤ C λ a λ b , ∥ D 2 F a b ∥ ≤ C λ a λ b ℓ b − 1 . \|\nabla F_{ab}\|\le C\lambda_a\lambda_b,\qquad\|D^2F_{ab}\|\le C\lambda_a\lambda_b\ell_b^{-1}. ∥∇ F ab ∥ ≤ C λ a λ b , ∥ D 2 F ab ∥ ≤ C λ a λ b ℓ b − 1 .
If a ≥ b a\ge b a ≥ b , differentiate F a b = ( u a ⋅ ∇ ) u b F_{ab}=(u_a\cdot\nabla)u_b F ab = ( u a ⋅ ∇ ) u b and use (41) with q ≤ 3 q\le3 q ≤ 3 . Since ℓ a ≤ ℓ b \ell_a\le\ell_b ℓ a ≤ ℓ b ,
∥ ∇ F a b ∥ ≤ C ( λ a λ b + A a λ b ℓ b − 1 ) ≤ C λ a λ b , \|\nabla F_{ab}\|\le C(\lambda_a\lambda_b+A_a\lambda_b\ell_b^{-1})\le C\lambda_a\lambda_b, ∥∇ F ab ∥ ≤ C ( λ a λ b + A a λ b ℓ b − 1 ) ≤ C λ a λ b ,
∥ D 2 F a b ∥ ≤ C ( λ a ℓ a − 1 λ b + λ a λ b ℓ b − 1 + A a λ b ℓ b − 2 ) ≤ C λ a λ b ℓ a − 1 . \|D^2F_{ab}\|\le C(\lambda_a\ell_a^{-1}\lambda_b+\lambda_a\lambda_b\ell_b^{-1}+A_a\lambda_b\ell_b^{-2})\le C\lambda_a\lambda_b\ell_a^{-1}. ∥ D 2 F ab ∥ ≤ C ( λ a ℓ a − 1 λ b + λ a λ b ℓ b − 1 + A a λ b ℓ b − 2 ) ≤ C λ a λ b ℓ a − 1 .
Thus both cases give
∥ D 2 F a b ∥ ≤ C λ a λ b ℓ b − 1 , ∣ δ Δ F a b ∣ ≤ 2 ∥ Δ F a b ∥ ≤ C λ a λ b ℓ b − 1 . \lVert D^2F_{ab}\rVert\le C\lambda_a\lambda_b\ell_b^{-1},\qquad|\delta\Delta F_{ab}| \le2\lVert\Delta F_{ab}\rVert\le C\lambda_a\lambda_b\ell_b^{-1}. ∥ D 2 F ab ∥ ≤ C λ a λ b ℓ b − 1 , ∣ δ Δ F ab ∣ ≤ 2 ∥ Δ F ab ∥ ≤ C λ a λ b ℓ b − 1 .
The test functions satisfy (3.14). We write I = { b : ℓ b ≤ 3 L S } \mathcal{I}=\{b:\ell_b\le3LS\} I = { b : ℓ b ≤ 3 L S } ; terms with b ∉ I b\notin\mathcal{I} b ∈ / I vanish identically near the configuration.
Step 3: the first corrector. Step 1 with f = ∇ Φ b ♯ f=\nabla\Phi_b^\sharp f = ∇ Φ b ♯ and G = u b G=u_b G = u b gives
∣ Δ Y ( h b B b Φ b ♯ ) ∣ ≤ C h b ( S β m ♯ − 3 r λ b + S β r m ♯ − 2 λ b ℓ b − 1 m b + S β m ♯ − 2 λ b ) ≤ C κ S β m ♯ − 2 ( 1 + r ℓ b − 1 m b ) , \begin{aligned}
|\Delta_Y(h_b\mathcal{B}_b\Phi_b^\sharp)|\le Ch_b(S^\beta m_\sharp^{-3}r\lambda_b+S^\beta r m_\sharp^{-2}\lambda_b\ell_b^{-1}m_b+S^\beta m_\sharp^{-2}\lambda_b) \\
&\le C\kappa S^\beta m_\sharp^{-2}(1+r\ell_b^{-1}m_b),
\end{aligned} ∣ Δ Y ( h b B b Φ b ♯ ) ∣ ≤ C h b ( S β m ♯ − 3 r λ b + S β r m ♯ − 2 λ b ℓ b − 1 m b + S β m ♯ − 2 λ b ) ≤ C κ S β m ♯ − 2 ( 1 + r ℓ b − 1 m b ) ,
using r ≤ m ♯ r\le m_\sharp r ≤ m ♯ and h b λ b ≤ κ h_b\lambda_b\le\kappa h b λ b ≤ κ .
Step 4: the second corrector. Recall from (56) that
B a B b Φ b ♯ = D 2 Φ b ♯ [ δ u a , δ u b ] + ∇ Φ b ♯ ⋅ δ F a b . \mathcal{B}_a\mathcal{B}_b\Phi_b^\sharp=D^2\Phi_b^\sharp[\delta u_a,\delta u_b]+\nabla\Phi_b^\sharp\cdot\delta F_{ab}. B a B b Φ b ♯ = D 2 Φ b ♯ [ δ u a , δ u b ] + ∇ Φ b ♯ ⋅ δ F ab .
For the first term, apply the second product formula in Step 1 with Λ = D 2 Φ b ♯ \Lambda=D^2\Phi_b^\sharp Λ = D 2 Φ b ♯ . After factoring out C S β m ♯ − 2 CS^\beta m_\sharp^{-2} C S β m ♯ − 2 , its six contributions are bounded by
m ♯ − 2 r 2 λ a λ b , λ a ℓ a − 1 m a r λ b , r λ a λ b ℓ b − 1 m b , m_\sharp^{-2}r^2\lambda_a\lambda_b,\qquad\lambda_a\ell_a^{-1}m_a r\lambda_b,\qquad r\lambda_a\lambda_b\ell_b^{-1}m_b, m ♯ − 2 r 2 λ a λ b , λ a ℓ a − 1 m a r λ b , r λ a λ b ℓ b − 1 m b ,
m ♯ − 1 λ a r λ b , m ♯ − 1 r λ a λ b , λ a λ b . m_\sharp^{-1}\lambda_a r\lambda_b,\qquad m_\sharp^{-1}r\lambda_a\lambda_b,\qquad\lambda_a\lambda_b. m ♯ − 1 λ a r λ b , m ♯ − 1 r λ a λ b , λ a λ b .
For the second term, apply the first product formula with f = ∇ Φ b ♯ f=\nabla\Phi_b^\sharp f = ∇ Φ b ♯ and G = F a b G=F_{ab} G = F ab . With the same factor removed, the three contributions are bounded by
m ♯ − 1 r λ a λ b , r λ a λ b ℓ ν − 1 , λ a λ b . m_\sharp^{-1}r\lambda_a\lambda_b,\qquad r\lambda_a\lambda_b\ell_\nu^{-1},\qquad\lambda_a\lambda_b. m ♯ − 1 r λ a λ b , r λ a λ b ℓ ν − 1 , λ a λ b .
Since r ≤ m ♯ r\le m_\sharp r ≤ m ♯ , ℓ ν ≤ ℓ a , ℓ b \ell_\nu\le\ell_a,\ell_b ℓ ν ≤ ℓ a , ℓ b , and m a , m b ≤ 1 m_a,m_b\le1 m a , m b ≤ 1 , all nine terms are bounded by
C S β m ♯ − 2 λ a λ b ( 1 + r / ℓ ν ) . CS^\beta m_\sharp^{-2}\lambda_a\lambda_b(1+r/\ell_\nu). C S β m ♯ − 2 λ a λ b ( 1 + r / ℓ ν ) .
The coefficient is estimated by (63) :
q a b λ a λ b ≤ κ 2 λ a / max ( λ a , λ b ) . q_{ab}\lambda_a\lambda_b\le\kappa^2\lambda_a/\max(\lambda_a,\lambda_b). q ab λ a λ b ≤ κ 2 λ a / max ( λ a , λ b ) .
Step 5: summation over the scales. Put N : = N L + log ( S / ℓ J ) \mathcal{N}:=N_L+\log(S/\ell_J) N := N L + log ( S / ℓ J ) . Since ℓ J ≤ ℓ b ≤ 3 L S \ell_J\le\ell_b\le3LS ℓ J ≤ ℓ b ≤ 3 L S for b ∈ I b\in\mathcal{I} b ∈ I , # I ≤ C N \#\mathcal{I}\le C\mathcal{N} # I ≤ C N and # { a : a ≥ b } ≤ C N \#\{a:a\ge b\}\le C\mathcal{N} # { a : a ≥ b } ≤ C N . Also
∑ b r ℓ b − 1 m b ≤ ∑ ℓ b ≥ r r 2 ℓ b − 2 + ∑ ℓ J ≤ ℓ b < r r / ℓ b ≤ C ( 1 + r / ℓ J ) . \sum_b r\ell_b^{-1}m_b\le\sum_{\ell_b\ge r}r^2\ell_b^{-2}+\sum_{\ell_J\le\ell_b<r}r/\ell_b\le C(1+r/\ell_J). b ∑ r ℓ b − 1 m b ≤ ℓ b ≥ r ∑ r 2 ℓ b − 2 + ℓ J ≤ ℓ b < r ∑ r / ℓ b ≤ C ( 1 + r / ℓ J ) .
For b ∈ I b\in\mathcal{I} b ∈ I , by (67) ,
∑ a λ a max ( λ a , λ b ) ( 1 + r ℓ ν ) ≤ ∑ a < b λ a λ b ( 1 + r ℓ b ) + ∑ a ≥ b ( 1 + r ℓ a ) ≤ C ( N + r / ℓ J ) . \sum_a\frac{\lambda_a}{\max(\lambda_a,\lambda_b)}\left(1+\frac{r}{\ell_\nu}\right)\le\sum_{a<b}\frac{\lambda_a}{\lambda_b}\left(1+\frac{r}{\ell_b}\right)+\sum_{a\ge b}\left(1+\frac{r}{\ell_a}\right)\le C(\mathcal{N}+r/\ell_J). a ∑ max ( λ a , λ b ) λ a ( 1 + ℓ ν r ) ≤ a < b ∑ λ b λ a ( 1 + ℓ b r ) + a ≥ b ∑ ( 1 + ℓ a r ) ≤ C ( N + r / ℓ J ) .
Hence
μ ∣ Δ Y ( Q ♯ − Φ ♯ ) ∣ ≤ C μ S β m ♯ − 2 [ κ ( N + r / ℓ J ) + κ 2 N ( N + r / ℓ J ) ] . \mu|\Delta_Y(Q^\sharp-\Phi^\sharp)|\le C\mu S^\beta m_\sharp^{-2}\left[\kappa(\mathcal{N}+r/\ell_J)+\kappa^2\mathcal{N}(\mathcal{N}+r/\ell_J)\right]. μ ∣ Δ Y ( Q ♯ − Φ ♯ ) ∣ ≤ C μ S β m ♯ − 2 [ κ ( N + r / ℓ J ) + κ 2 N ( N + r / ℓ J ) ] .
Step 6: comparison with the diffusivity. We distinguish the regions above and below the finest scale. If r ≥ ℓ J r\ge\ell_J r ≥ ℓ J , then S = m ♯ = r S=m_\sharp=r S = m ♯ = r , N = N L + log ( r / ℓ J ) \mathcal{N}=N_L+\log(r/\ell_J) N = N L + log ( r / ℓ J ) with r / ℓ J ≥ 1 r/\ell_J\ge1 r / ℓ J ≥ 1 , and μ S β m ♯ − 2 = ε 0 ( r / ℓ J ) − ( 1 + α ) \mu S^\beta m_\sharp^{-2}=\varepsilon_0(r/\ell_J)^{-(1+\alpha)} μ S β m ♯ − 2 = ε 0 ( r / ℓ J ) − ( 1 + α ) ; since α > 0 \alpha>0 α > 0 , ( r / ℓ J ) − ( 1 + α ) ( 1 + log ( r / ℓ J ) ) 2 ( 1 + r / ℓ J ) ≤ C (r/\ell_J)^{-(1+\alpha)}(1+\log(r/\ell_J))^2(1+r/\ell_J)\le C ( r / ℓ J ) − ( 1 + α ) ( 1 + log ( r / ℓ J ) ) 2 ( 1 + r / ℓ J ) ≤ C for r / ℓ J ≥ 1 r/\ell_J\ge1 r / ℓ J ≥ 1 , so the bracket contributes at most C ε 0 ( κ N L + κ 2 N L 2 ) C\varepsilon_0(\kappa N_L+\kappa^2N_L^2) C ε 0 ( κ N L + κ 2 N L 2 ) . If r < ℓ j r < \ell_j r < ℓ j , then S = ℓ j S = \ell_j S = ℓ j , N = N L N = N_L N = N L , m ♯ ≥ r ♯ m_{\sharp} \ge r_{\sharp} m ♯ ≥ r ♯ , and μ ℓ j β r ♯ − 2 = κ / ( 4 β ρ β ) \mu_{\ell_j}^{\beta} r_{\sharp}^{-2} = \kappa/(4\beta\rho^{\beta}) μ ℓ j β r ♯ − 2 = κ / ( 4 β ρ β ) by (38) , so the contribution is at most C ( κ 2 N L + κ 3 N L 2 ) C(\kappa^2N_L+\kappa^3N_L^2) C ( κ 2 N L + κ 3 N L 2 ) . With ε 0 ≤ κ \varepsilon_0 \le\kappa ε 0 ≤ κ and κ N L ≤ 1 \kappa N_L \le1 κ N L ≤ 1 both are at most C κ 2 N L C\kappa^2N_L C κ 2 N L . Steps 2–6 use Lemma 4.3 , which assumes r > 0 r>0 r > 0 ; the case r = 0 r=0 r = 0 follows by continuity, since Q ♯ − Φ ♯ Q^{\sharp}-\Phi^{\sharp} Q ♯ − Φ ♯ is C 1 C^1 C 1 in Y Y Y up to the diagonal (Corollary 5.3). □
Renewal averages and time reversal
Averages over one renewal window
6.1. Averages over one renewal window. The following identities are averages over the random data of one window, that is, integrals over the shift ξ ∈ Q j \xi\in Q_j ξ ∈ Q j and the finitely many coefficients η i ν \eta_i^\nu η i ν . The role of the random lattice shift is visible in the first display of the proof: averaging over ξ \xi ξ turns a lattice sum into an integral over T d \mathbb{T}^d T d , so that one scale of the model has, on average, the translation-invariant structure function of a single wavelet.
Lemma 6.1 (Fresh-window identities). For a freshly drawn scale-j j j field
U ( x ) = A j ∑ i ∈ I ∑ ν η i ν W j ν ( x − y i j − ξ ) , U(x)=A_j\sum_{i\in\mathcal{I}}\sum_{\nu}\eta_i^\nu W_j^\nu(x-y_i^j-\xi), U ( x ) = A j i ∈ I ∑ ν ∑ η i ν W j ν ( x − y i j − ξ ) , and fixed x , y ∈ T d x,y\in\mathbb{T}^d x , y ∈ T d with shortest difference z = x − y z=x-y z = x − y satisfying ∣ z ∣ ≤ 1 / 4 |z|\le1/4 ∣ z ∣ ≤ 1/4 ,
E U ( x ) = 0 , E [ ( U ⋅ ∇ ) U ] ( x ) = 0 , (77) \mathbb{E}U(x)=0,\qquad\mathbb{E}[(U\cdot\nabla)U](x)=0, \tag*{(77)} E U ( x ) = 0 , E [( U ⋅ ∇ ) U ] ( x ) = 0 , (77) E [ ( U ( x ) − U ( y ) ) ⊗ ( U ( x ) − U ( y ) ) ] = σ 2 A j 2 D ( z / ℓ j ) . (78) \mathbb{E}[(U(x)-U(y))\otimes(U(x)-U(y))]=\sigma^2 A_j^2D(z/\ell_j). \tag*{(78)} E [( U ( x ) − U ( y )) ⊗ ( U ( x ) − U ( y ))] = σ 2 A j 2 D ( z / ℓ j ) . (78) Here E \mathbb{E} E is the average over ( ξ , η ) (\xi,\eta) ( ξ , η ) only. The identities therefore remain true when x x x and y y y are any points that do not depend on ( ξ , η ) (\xi,\eta) ( ξ , η ) , for instance the particle positions at the start of the window, which are functions of the data of earlier windows only.
Proof. Centering proves the first identity. In the quadratic expressions, independence removes terms with different spatial or mode indices. For any periodic integrable function F F F ,
E ξ ∑ i ∈ I j F ( y i j + ξ ) = ℓ j − d ∫ T d F ( y ~ ) d y ~ . \mathbb{E}_{\xi}\sum_{i\in I_j}F(y_i^j+\xi)=\ell_j^{-d}\int_{\mathbb{T}^d}F(\widetilde y)\,\mathrm{d}\widetilde y. E ξ i ∈ I j ∑ F ( y i j + ξ ) = ℓ j − d ∫ T d F ( y ) d y .
The one-point drift is therefore σ 2 A j 2 ℓ j − d ∑ ν ∫ T d ( W ν ⋅ ∇ ) W ν = 0 \sigma^2A_j^2\ell_j^{-d}\sum_{\nu}\int_{\mathbb{T}^d}(W^\nu\cdot\nabla)W^\nu=0 σ 2 A j 2 ℓ j − d ∑ ν ∫ T d ( W ν ⋅ ∇ ) W ν = 0 , because each integrand is ∇ ⋅ ( W ν ⊗ W ν ) \nabla\cdot(W^\nu\otimes W^\nu) ∇ ⋅ ( W ν ⊗ W ν ) and has integral zero.
The same unfolding gives (78) . Averaging first over the centered independent coefficients and then over the lattice shift gives
E [ ( U ( x ) − U ( y ) ) ⊗ 2 ] = σ 2 A j 2 ℓ j − d ∑ ν ∫ T d [ W j ν ( x ~ ) − W j ν ( x ~ − z ) ] ⊗ 2 d x ~ . \mathbb{E}[(U(x)-U(y))^{\otimes2}]=\sigma^2A_j^2\ell_j^{-d}\sum_{\nu}\int_{\mathbb{T}^d}[W_j^\nu(\widetilde x)-W_j^\nu(\widetilde x-z)]^{\otimes2}\,\mathrm{d}\widetilde x. E [( U ( x ) − U ( y ) ) ⊗ 2 ] = σ 2 A j 2 ℓ j − d ν ∑ ∫ T d [ W j ν ( x ) − W j ν ( x − z ) ] ⊗ 2 d x .
Set Ψ ν ( x ~ ) = W ν ( x ~ / ℓ j ) − W ν ( ( x ~ − z ) / ℓ j ) \Psi^\nu(\widetilde x)=W^\nu(\widetilde x/\ell_j)-W^\nu((\widetilde x-z)/\ell_j) Ψ ν ( x ) = W ν ( x / ℓ j ) − W ν (( x − z ) / ℓ j ) . Its support lies in two balls whose union has diameter at most 2 R ℓ j + ∣ z ∣ ≤ 3 / 4 < 1 2R\ell_j+|z|\le3/4<1 2 R ℓ j + ∣ z ∣ ≤ 3/4 < 1 , so distinct integer translates of its support are disjoint. Thus the torus integral unfolds to ∫ R d ( Ψ ν ) ⊗ 2 \int_{\mathbb{R}^d}(\Psi^\nu)^{\otimes2} ∫ R d ( Ψ ν ) ⊗ 2 . Changing variables x ~ = ℓ j y ~ \widetilde x=\ell_j\widetilde y x = ℓ j y cancels ℓ j − d \ell_j^{-d} ℓ j − d , and summing over ν \nu ν gives σ 2 A j 2 D ( z / ℓ j ) \sigma^2A_j^2D(z/\ell_j) σ 2 A j 2 D ( z / ℓ j ) , as claimed. □
Lemma 6.2 (Fresh-window identities after a measure-preserving deformation). *Let Ψ : T d → T d \Psi:\mathbb{T}^d\to\mathbb{T}^d Ψ : T d → T d be a Lebesgue-measure-preserving bijection, and U ( x ) = U(x)= U ( x ) =
A j ∑ i , ν η i ν W j ν ( x − Ψ ( y i j + ξ ) ) A_j \sum_{i,\nu} \eta_i^\nu W_j^\nu(x-\Psi(y_i^j+\xi)) A j ∑ i , ν η i ν W j ν ( x − Ψ ( y i j + ξ )) with ξ \xi ξ uniform on Q j Q_j Q j and ( η i ν ) (\eta_i^\nu) ( η i ν ) as in (1.4), all independent. Then (77) and (78) hold for U U U .
Proof. The identity E U ( x ) = 0 \mathbb{E}U(x)=0 E U ( x ) = 0 follows directly from the centering of the coefficients. For the quadratic identities, we first average over the coefficients with the shift ξ \xi ξ fixed. Independence and centering eliminate all mixed terms, while each diagonal term has coefficient σ 2 \sigma^2 σ 2 . Thus, writing
F ν ( p ) : = [ W j ν ( x − p ) − W j ν ( y − p ) ] ⊗ 2 , F_\nu(p):=[W_j^\nu(x-p)-W_j^\nu(y-p)]^{\otimes2}, F ν ( p ) := [ W j ν ( x − p ) − W j ν ( y − p ) ] ⊗ 2 ,
we obtain
E η [ ( U ( x ) − U ( y ) ) ⊗ 2 ] = σ 2 A j 2 ∑ i , ν F ν ( Ψ ( y i j + ξ ) ) . \mathbb{E}_{\eta}[(U(x)-U(y))^{\otimes2}]=\sigma^2 A_j^2\sum_{i,\nu}F_\nu(\Psi(y_i^j+\xi)). E η [( U ( x ) − U ( y ) ) ⊗ 2 ] = σ 2 A j 2 i , ν ∑ F ν ( Ψ ( y i j + ξ )) .
The deformation Ψ \Psi Ψ is still present after this first average.
We now average over the shift. For every periodic integrable function F F F , the uniform distribution of ξ \xi ξ on Q j Q_j Q j gives
E ξ ∑ i F ( Ψ ( y i j + ξ ) ) = ℓ j − d ∑ i ∫ Q j F ( Ψ ( y i j + ξ ) ) d ξ = ℓ j − d ∫ T d F ( Ψ ( q ) ) d q = ℓ j − d ∫ T d F ( p ) d p . \begin{aligned}
\mathbb{E}_{\xi}\sum_i F(\Psi(y_i^j+\xi))=\ell_j^{-d}\sum_i\int_{Q_j}F(\Psi(y_i^j+\xi))\,\mathrm{d}\xi\\
&=\ell_j^{-d}\int_{\mathbb{T}^d}F(\Psi(q))\,\mathrm{d}q=\ell_j^{-d}\int_{\mathbb{T}^d}F(p)\,\mathrm{d}p.
\end{aligned} E ξ i ∑ F ( Ψ ( y i j + ξ )) = ℓ j − d i ∑ ∫ Q j F ( Ψ ( y i j + ξ )) d ξ = ℓ j − d ∫ T d F ( Ψ ( q )) d q = ℓ j − d ∫ T d F ( p ) d p .
Here the second equality uses the tiling of T d \mathbb{T}^d T d by the cells y i j + Q j y_i^j+Q_j y i j + Q j , and the last equality uses that Ψ \Psi Ψ preserves Lebesgue measure. Applying this identity componentwise to F ν F_\nu F ν removes the deformation and yields
E [ ( U ( x ) − U ( y ) ) ⊗ 2 ] = σ 2 A j 2 ℓ j − d ∑ ν ∫ T d [ W j ν ( x − p ) − W j ν ( y − p ) ] ⊗ 2 d p . \mathbb{E}[(U(x)-U(y))^{\otimes2}]=\sigma^2 A_j^2\ell_j^{-d}\sum_\nu\int_{\mathbb{T}^d}[W_j^\nu(x-p)-W_j^\nu(y-p)]^{\otimes2}\,\mathrm{d}p. E [( U ( x ) − U ( y ) ) ⊗ 2 ] = σ 2 A j 2 ℓ j − d ν ∑ ∫ T d [ W j ν ( x − p ) − W j ν ( y − p ) ] ⊗ 2 d p .
This is precisely the covariance integral in the proof of Lemma 6.1; the same unfolding and rescaling give σ 2 A j 2 D ( z / ℓ j ) \sigma^2 A_j^2D(z/\ell_j) σ 2 A j 2 D ( z / ℓ j ) , where z = x − y z=x-y z = x − y is the shortest displacement. This proves (78) .
For the quadratic drift, the coefficient average similarly gives
E η [ ( U ⋅ ∇ ) U ] ( x ) = σ 2 A j 2 ∑ i , ν [ ( W j ν ⋅ ∇ ) W j ν ] ( x − Ψ ( y i j + ξ ) ) . \mathbb{E}_{\eta}[(U\cdot\nabla)U](x)=\sigma^2 A_j^2\sum_{i,\nu}[(W_j^\nu\cdot\nabla)W_j^\nu](x-\Psi(y_i^j+\xi)). E η [( U ⋅ ∇ ) U ] ( x ) = σ 2 A j 2 i , ν ∑ [( W j ν ⋅ ∇ ) W j ν ] ( x − Ψ ( y i j + ξ )) .
The same shift average therefore implies
E [ ( U ⋅ ∇ ) U ] ( x ) = σ 2 A j 2 ℓ j − d ∑ ν ∫ T d ( W j ν ⋅ ∇ ) W j ν ( q ) d q = 0. \mathbb{E}[(U\cdot\nabla)U](x)=\sigma^2 A_j^2\ell_j^{-d}\sum_\nu\int_{\mathbb{T}^d}(W_j^\nu\cdot\nabla)W_j^\nu(q)\,\mathrm{d}q=0. E [( U ⋅ ∇ ) U ] ( x ) = σ 2 A j 2 ℓ j − d ν ∑ ∫ T d ( W j ν ⋅ ∇ ) W j ν ( q ) d q = 0.
Indeed, each wavelet is divergence-free, so the integrand equals ∇ ⋅ ( W j ν ⊗ W j ν ) \nabla\cdot(W_j^\nu\otimes W_j^\nu) ∇ ⋅ ( W j ν ⊗ W j ν ) , whose integral vanishes by periodicity. This completes the proof of (77) . □
The reversed field and its renewal data
The dissipation of the scalar up to time t 0 t_0 t 0 is expressed through an equation driven by the time-reversed field (Lemma 10.1). This section describes the reversed field and verifies the two properties needed for the same corrector and jump identities: the structure between renewals, and the independence of the fresh data at a renewal time from the past of the reversed field.
Fix J ≥ j 0 J \ge j_0 J ≥ j 0 and t 0 ∈ ( 0 , 1 ] t_0 \in(0,1] t 0 ∈ ( 0 , 1 ] . By Lemma 2.3 , for each realization, t ↦ u j ( x , t ) t \mapsto u_j(x,t) t ↦ u j ( x , t ) has left limits u j ( x , t − ) u_j(x,t-) u j ( x , t − ) for t ∈ ( 0 , 1 ] t \in(0,1] t ∈ ( 0 , 1 ] .
Forward renewal times are denoted by ζ \zeta ζ , and their reversed counterparts by ζ ← = t 0 − ζ \overleftarrow{\zeta}=t_0-\zeta ζ = t 0 − ζ . In statements applying to either evolution, ζ ~ \widetilde{\zeta} ζ denotes a renewal time in the time variable of the chosen evolution; left limits and jumps are always taken in that time variable.
Definition 6.3 (Reversed field). For s ∈ [ 0 , t 0 ) s \in[0,t_0) s ∈ [ 0 , t 0 ) and j 0 ≤ j ≤ J j_0 \le j \le J j 0 ≤ j ≤ J let k k k be such that t 0 − s ∈ ( k τ j , ( k + 1 ) τ j ] t_0-s \in(k\tau_j,(k+1)\tau_j] t 0 − s ∈ ( k τ j , ( k + 1 ) τ j ] , and set
u ← j ( x , s ) : = − u j ( x , ( t 0 − s ) − ) , u ← : = ∑ j = j 0 J u ← j , P ← i j ( s ) : = P i , k j ( ( t 0 − s ) − ) . \overleftarrow{u}_j(x,s):=-u_j(x,(t_0-s)-),\qquad
\overleftarrow{u}:=\sum_{j=j_0}^{J}\overleftarrow{u}_j,\qquad
\overleftarrow{P}^{\,j}_i(s):=P^j_{i,k}((t_0-s)-). u j ( x , s ) := − u j ( x , ( t 0 − s ) − ) , u := j = j 0 ∑ J u j , P i j ( s ) := P i , k j (( t 0 − s ) − ) . where P i , k j ( ( k + 1 ) τ j − ) P^j_{i,k}((k+1)\tau_j-) P i , k j (( k + 1 ) τ j − ) is the left limit within window k k k ; at s = t 0 s=t_0 s = t 0 all three are defined by their left limits as s ↑ t 0 s \uparrow t_0 s ↑ t 0 . The reversed renewal times of scale j j j are the points ζ ← = t 0 − ζ \overleftarrow{\zeta}=t_0-\zeta ζ = t 0 − ζ with ζ ∈ τ j N ∩ ( 0 , t 0 ) \zeta\in\tau_j\mathbb{N}\cap(0,t_0) ζ ∈ τ j N ∩ ( 0 , t 0 ) ; the reversed windows of scale j j j are the intervals between consecutive reversed renewal times, together with the initial window starting at 0 0 0 . The reversed phase is h ← j ( s ) : = ( t 0 − s ) − k τ j \overleftarrow{h}_j(s):=(t_0-s)-k\tau_j h j ( s ) := ( t 0 − s ) − k τ j for s < t 0 s<t_0 s < t 0 , and h ← j ( t 0 ) : = h ← j ( t 0 − ) \overleftarrow{h}_j(t_0):=\overleftarrow{h}_j(t_0-) h j ( t 0 ) := h j ( t 0 − ) .
Changing the field on a finite set of times changes neither the solutions of the center equations nor the energy solutions of Section 8, so the left-limit conventions are immaterial; in particular u ← ( x , s ) = − u ( x , t 0 − s ) \overleftarrow{u}(x,s)=-u(x,t_0-s) u ( x , s ) = − u ( x , t 0 − s ) for all but finitely many s s s .
By construction 0 < h ← j ( s ) ≤ τ j 0<\overleftarrow{h}_j(s)\le\tau_j 0 < h j ( s ) ≤ τ j for s < t 0 s<t_0 s < t 0 and h ← j ( t 0 ) = 0 \overleftarrow{h}_j(t_0)=0 h j ( t 0 ) = 0 ; h ← j \overleftarrow{h}_j h j decreases with slope − 1 -1 − 1 between reversed renewals; and at a reversed renewal time ζ ← \overleftarrow{\zeta} ζ of scale j j j , h ← j ( ζ ← − ) = 0 \overleftarrow{h}_j(\overleftarrow{\zeta}-)=0 h j ( ζ − ) = 0 and h ← j ( ζ ← ) = τ j \overleftarrow{h}_j(\overleftarrow{\zeta})=\tau_j h j ( ζ ) = τ j . Every reversed window starting at a time ζ ← > 0 \overleftarrow{\zeta}>0 ζ > 0 has length exactly τ j \tau_j τ j (it corresponds to the forward window [ t 0 − ζ ← − τ j , t 0 − ζ ← ) [t_0-\overleftarrow{\zeta}-\tau_j,t_0-\overleftarrow{\zeta}) [ t 0 − ζ − τ j , t 0 − ζ ) ), and the initial reversed window has length at most τ j \tau_j τ j . Thus the reversed phases play exactly the role of the phases of Section 5.
Lemma 6.4 (Structure of the reversed field). (a) On each reversed window of scale j j j , u ← j ( x , s ) = A j ∑ i ∑ ν ( − η i , k j , ν ) W ν j ( x − P ← i j ( s ) ) \overleftarrow{u}_j(x,s)=A_j\sum_i\sum_\nu(-\eta^{j,\nu}_{i,k})W^j_\nu(x-\overleftarrow{P}^{\,j}_i(s)) u j ( x , s ) = A j ∑ i ∑ ν ( − η i , k j , ν ) W ν j ( x − P i j ( s )) with constant coefficients, and d d s P ← i j = u ← < j ( P ← i j , s ) \dfrac{d}{ds}\overleftarrow{P}^{\,j}_i=\overleftarrow{u}_{<j}(\overleftarrow{P}^{\,j}_i,s) d s d P i j = u < j ( P i j , s ) for a.e. s s s .
(b) For κ ≤ κ ∗ \kappa\le\kappa_* κ ≤ κ ∗ , the conclusions of Lemma 2.2 (separation, (16) , (17) ) hold for u ← \overleftarrow{u} u with the same constants, and the configurations ( x , x ~ , Θ ← s ) (x,\widetilde{x},\overleftarrow{\Theta}_s) ( x , x , Θ s ) with 0 < dist ( x , x ~ ) ≤ 1 / 4 0<\operatorname{dist}(x,\widetilde{x})\le1/4 0 < dist ( x , x ) ≤ 1/4 are admissible (Definition 4.1). Between consecutive reversed renewal times of all scales, the reversed configuration evolves with the generator B ← = ∑ a B ← a \overleftarrow{B}=\sum_a\overleftarrow{B}_a B = ∑ a B a of Definition 4.2, built from the frozen fields with centers P ← \overleftarrow{P} P and coefficients − η -\eta − η : explicitly, the reversed frozen field of generation j j j at time s s s is the field (4.1) formed with the centers P ← i j ( s ) \overleftarrow{P}^{\,j}_i(s) P i j ( s ) and the coefficients − η i , k j , ν -\eta^{j,\nu}_{i,k} − η i , k j , ν of the current reversed window, and it equals u ← j ( ⋅ , s ) \overleftarrow{u}_j(\cdot,s) u j ( ⋅ , s ) . Since ∣ − η ∣ = ∣ η ∣ ≤ M |-\eta|=|\eta|\le M ∣ − η ∣ = ∣ η ∣ ≤ M , condition ( C 1 ) (C1) ( C 1 ) holds for these coefficients.
(c) (Freshness. ) Let ζ ← ∈ ( 0 , t 0 ) \overleftarrow{\zeta}\in(0,t_0) ζ ∈ ( 0 , t 0 ) be a reversed renewal time of at least one scale, put ζ : = t 0 − ζ ← \zeta:=t_0-\overleftarrow{\zeta} ζ := t 0 − ζ , and let j 1 j_1 j 1 be the smallest index with ζ ∈ τ j 1 N \zeta\in\tau_{j_1}\mathbb{N} ζ ∈ τ j 1 N . The set of scales renewed at ζ ← \overleftarrow{\zeta} ζ is J ( ζ ← ) = { j 1 , … , J } \mathcal{J}(\overleftarrow{\zeta})=\{j_1,\ldots,J\} J ( ζ ) = { j 1 , … , J } . Let ω ζ ← \omega_{\overleftarrow{\zeta}} ω ζ be the data ( ξ m , η m ) (\xi^m,\eta^m) ( ξ m , η m ) of the forward windows [ ζ − τ m , ζ ) [\zeta-\tau_m,\zeta) [ ζ − τ m , ζ ) , m ∈ J ( ζ ← ) m\in\mathcal{J}(\overleftarrow{\zeta}) m ∈ J ( ζ ) , and ω ζ ← c \omega_{\overleftarrow{\zeta}}^c ω ζ c all other data. Then:
(c1) u ← \overleftarrow{u} u restricted to [ 0 , ζ ← ) [0,\overleftarrow{\zeta}) [ 0 , ζ ) is a function of ω ζ ← c \omega_{\overleftarrow{\zeta}}^c ω ζ c ;
(c2) for m ∈ J ( ζ ← ) m\in\mathcal{J}(\overleftarrow{\zeta}) m ∈ J ( ζ ) , P ← i m ( ζ ← ) = Ψ m ( y i m + ξ m ) \overleftarrow{P}^{m}_i(\overleftarrow{\zeta})=\Psi_m(y_i^m+\xi^m) P i m ( ζ ) = Ψ m ( y i m + ξ m ) , where Ψ m \Psi_m Ψ m is the forward flow map of u < m u_{<m} u < m from time ζ − τ m \zeta-\tau_m ζ − τ m to ζ \zeta ζ ; Ψ m \Psi_m Ψ m is a Lebesgue-measure-preserving C 1 C^1 C 1 diffeomorphism of T d \mathbb{T}^d T d and a function of ω ζ ← c \omega_{\overleftarrow{\zeta}}^c ω ζ c and of ( ξ a , η a ) a ∈ J ( ζ ← ) , a < m (\xi^a,\eta^a)_{a\in\mathcal{J}(\overleftarrow{\zeta}),\,a<m} ( ξ a , η a ) a ∈ J ( ζ ) , a < m .
Proof. (a) On a reversed window the forward time t 0 − s t_0-s t 0 − s runs through a single forward window ( k τ j , ( k + 1 ) τ j ] (k\tau_j,(k+1)\tau_j] ( k τ j , ( k + 1 ) τ j ] , where the coefficients are η i n , j \eta^{n,j}_i η i n , j ; for s s s off the finite set of renewal times, d d s P ( t 0 − s ) = − u < j ( P ( t 0 − s ) , t 0 − s ) \frac{d}{ds}P(t_0-s)=-u_{<j}(P(t_0-s),t_0-s) d s d P ( t 0 − s ) = − u < j ( P ( t 0 − s ) , t 0 − s ) .
(b) The fields u ← j ( ⋅ , s ) \overleftarrow{u}_j(\cdot,s) u j ( ⋅ , s ) are, up to sign, the forward fields at time ( t 0 − s ) − (t_0-s)- ( t 0 − s ) − , and the reversed centers are forward center positions or their left limits. The separation of the centers, (16) and (17) are statements about the field and the centers at one time, and they pass to left limits (left limits of ℓ j / 2 \ell_j/2 ℓ j /2 -separated configurations are ℓ j / 2 \ell_j/2 ℓ j /2 -separated). The statement about the generator is (a) together with the computation in the proof of Proposition 5.2.
(c) By Lemma 2.1, ζ ∈ τ m N \zeta\in\tau_m\mathbb{N} ζ ∈ τ m N exactly for m ≥ j 1 m\ge j_1 m ≥ j 1 . We claim that for f ∈ [ ζ , t 0 ) f\in[\zeta,t_0) f ∈ [ ζ , t 0 ) , u a ( ⋅ , f ) u_a(\cdot,f) u a ( ⋅ , f ) is a function of the data
D : = { data of all windows of scales a ′ < j 1 } ∪ { data of the windows of scales m ≥ j 1 starting at times ≥ ζ } . D:=\{\text{data of all windows of scales }a'<j_1\}
\cup\{\text{data of the windows of scales }m\ge j_1\text{ starting at times }\ge\zeta\}. D := { data of all windows of scales a ′ < j 1 } ∪ { data of the windows of scales m ≥ j 1 starting at times ≥ ζ } .
We argue by induction on a a a . For a < j 1 a<j_1 a < j 1 , u a u_a u a depends only on data of scales ≤ a < j 1 \le a<j_1 ≤ a < j 1 . For a ≥ j 1 a\ge j_1 a ≥ j 1 , the scale-a a a window containing f f f starts at some ζ ′ ∈ [ ζ , f ] \zeta'\in[\zeta,f] ζ ′ ∈ [ ζ , f ] (as ζ ∈ τ a N \zeta\in\tau_a\mathbb{N} ζ ∈ τ a N ), and u a ( ⋅ , f ) u_a(\cdot,f) u a ( ⋅ , f ) is determined by the data of that window, which lie in D D D , and by its centers, which solve P ˙ = u < a ( P , ⋅ ) \dot{P}=u_{<a}(P,\cdot) P ˙ = u < a ( P , ⋅ ) on [ ζ ′ , f ] ⊂ [ ζ , t 0 ) [\zeta',f]\subset[\zeta,t_0) [ ζ ′ , f ] ⊂ [ ζ , t 0 ) from y i + ξ y_i+\xi y i + ξ ; by the induction hypothesis u < a u_{<a} u < a on [ ζ , t 0 ) [\zeta,t_0) [ ζ , t 0 ) is a function of D D D . Left limits at f ∈ ( ζ , t 0 ] f\in(\zeta,t_0] f ∈ ( ζ , t 0 ] are limits of values on [ ζ , f ) [\zeta,f) [ ζ , f ) . Hence u ← \overleftarrow{u} u on [ 0 , ζ ← ) [0,\overleftarrow{\zeta}) [ 0 , ζ ) is a function of D D D , and D ∩ ω ζ ← c = ∅ D\cap\omega_{\overleftarrow{\zeta}}^c=\varnothing D ∩ ω ζ c = ∅ because the windows in ω ζ ← \omega_{\overleftarrow{\zeta}} ω ζ start at ζ − τ m < ζ \zeta-\tau_m<\zeta ζ − τ m < ζ . This is (c1).
For (c2): P ← i m ( ζ ← ) = P i m ( ζ − ) \overleftarrow{P}^{m}_i(\overleftarrow{\zeta})=P_i^m(\zeta-) P i m ( ζ ) = P i m ( ζ − ) with k = ζ / τ m − 1 k=\zeta/\tau_m-1 k = ζ / τ m − 1 , and P i m ( ζ − ) = Ψ m ( y i m + ξ m ) P_i^m(\zeta-)=\Psi_m(y_i^m+\xi^m) P i m ( ζ − ) = Ψ m ( y i m + ξ m ) by (5) . We claim that, for j 1 ≤ a ≤ m j_1\le a\le m j 1 ≤ a ≤ m , the restriction of u a u_a u a to [ ζ − τ a , ζ ) [\zeta-\tau_a,\zeta) [ ζ − τ a , ζ ) is a function of ω ζ ← c \omega_{\overleftarrow{\zeta}}^c ω ζ c and of ( ξ a ′ , η a ′ ) j 1 ≤ a ′ ≤ a (\xi^{a'},\eta^{a'})_{j_1\le a'\le a} ( ξ a ′ , η a ′ ) j 1 ≤ a ′ ≤ a ; since [ ζ − τ m , ζ ) ⊂ [ ζ − τ a , ζ ) [\zeta-\tau_m,\zeta)\subset[\zeta-\tau_a,\zeta) [ ζ − τ m , ζ ) ⊂ [ ζ − τ a , ζ ) for a ≤ m a\le m a ≤ m , it follows that u < m u_{<m} u < m on [ ζ − τ m , ζ ) [\zeta-\tau_m,\zeta) [ ζ − τ m , ζ ) is a function of ω ζ ← c \omega_{\overleftarrow{\zeta}}^c ω ζ c and of the fresh data of the scales in [ j 1 , m ) [j_1,m) [ j 1 , m ) .
We argue by induction on a a a . A scale a ′ < j 1 a'<j_1 a ′ < j 1 is determined, at all times, by the data of scales < j 1 <j_1 < j 1 , which belong to ω ζ ← c \omega_{\overleftarrow{\zeta}}^c ω ζ c . A scale j 1 ≤ a ≤ m j_1\le a\le m j 1 ≤ a ≤ m is, on [ ζ − τ a , ζ ) [\zeta-\tau_a,\zeta) [ ζ − τ a , ζ ) , in its window [ ζ − τ a , ζ ) [\zeta-\tau_a,\zeta) [ ζ − τ a , ζ ) , whose data ( ξ a , η a ) (\xi^a,\eta^a) ( ξ a , η a ) belong to ω ζ ← c \omega_{\overleftarrow{\zeta}}^c ω ζ c , and its centers are transported on this interval by u < a u_{<a} u < a , which by the induction hypothesis is there a function of ω ζ ← c \omega_{\overleftarrow{\zeta}}^c ω ζ c and of ( ξ a ′ , η a ′ ) j 1 ≤ a ′ < a (\xi^{a'},\eta^{a'})_{j_1\le a'<a} ( ξ a ′ , η a ′ ) j 1 ≤ a ′ < a (the restriction of u a ′ u_{a'} u a ′ to [ ζ − τ a , ζ ) ⊂ [ ζ − τ a ′ , ζ ′ ) [\zeta-\tau_a,\zeta)\subset[\zeta-\tau_{a'},\zeta') [ ζ − τ a , ζ ) ⊂ [ ζ − τ a ′ , ζ ′ ) being a function of the restriction to the larger interval). This proves the claim. On each subinterval between points of τ N \tau_{\mathbb{N}} τ N this field is C 3 C^3 C 3 in x x x , continuous in time and divergence-free (Lemma 2.3 ), so its flow is a C 1 C^1 C 1 diffeomorphism with Jacobian determinant exp ( ∫ ∇ ⋅ u < m ) = 1 \exp\left(\int\nabla\cdot u_{<m}\right)=1 exp ( ∫ ∇ ⋅ u < m ) = 1 ; composing finitely many such maps gives (c2). □ \square □
Jump identity in both directions of time
The same jump calculation applies to both families of observables. Let ( Ξ , Ξ j ) (\Xi,\Xi_j) ( Ξ , Ξ j ) be ( Φ , Φ j ) (\Phi,\Phi_j) ( Φ , Φ j ) for a distinct pair, or ( Φ ♯ , Φ j ♯ ) (\Phi^\sharp,\Phi^\sharp_j) ( Φ ♯ , Φ j ♯ ) for a pair that may lie on the diagonal. Define Q Q Q by (5.12) with ( Ξ , Ξ j ) (\Xi,\Xi_j) ( Ξ , Ξ j ) in place of ( Φ , Φ j ) (\Phi,\Phi_j) ( Φ , Φ j ) , and put
Γ j Ξ ( r ) : = σ 2 A j 2 2 D ( z / ℓ j ) : D 2 Ξ j ( z ) , ∣ z ∣ = r ; \Gamma^{\Xi}_j(r):=\frac{\sigma^2 A_j^2}{2}D(z/\ell_j):D^2\Xi_j(z),\qquad|z|=r; Γ j Ξ ( r ) := 2 σ 2 A j 2 D ( z / ℓ j ) : D 2 Ξ j ( z ) , ∣ z ∣ = r ;
For the quadratic-core functions, Γ j Ξ ( 0 ) = 0 \Gamma^{\Xi}_j(0)=0 Γ j Ξ ( 0 ) = 0 since D ( 0 ) = 0 D(0)=0 D ( 0 ) = 0 .
Proposition 6.5 (Averaged jump in both directions of time). Let ζ ~ ∈ ( 0 , t 0 ) \widetilde{\zeta}\in(0,t_0) ζ ∈ ( 0 , t 0 ) be a renewal time of the forward or reversed field, J ( ζ ~ ) \mathcal{J}(\widetilde{\zeta}) J ( ζ ) the set of renewed scales, and ω ζ ~ \omega_{\widetilde{\zeta}} ω ζ the fresh data (the data of the windows starting at ζ ~ \widetilde{\zeta} ζ in the forward case; as in Lemma 6.4 (c) in the reversed case). Let x , x ~ ∈ T d x,\widetilde{x}\in\mathbb{T}^d x , x ∈ T d be fixed points with dist ( x , x ~ ) ≤ 1 / 4 \operatorname{dist}(x,\widetilde{x})\leq1/4 dist ( x , x ) ≤ 1/4 (strictly positive for the logarithmic functions), and let S ζ ~ − , S ζ ~ S_{\widetilde{\zeta}-},S_{\widetilde{\zeta}} S ζ − , S ζ be the configurations with particles x , x ~ x,\widetilde{x} x , x and the window data before and after ζ ~ \widetilde{\zeta} ζ . Then
E ζ ~ [ Q ( S ζ ~ , h ( ζ ~ ) ) − Q ( S ζ ~ − , h ( ζ ~ − ) ) ] = ∑ j ∈ J ( ζ ~ ) τ j 2 Γ j Ξ ( dist ( x , x ~ ) ) , \mathbb{E}_{\widetilde{\zeta}}\left[Q(S_{\widetilde{\zeta}},h(\widetilde{\zeta}))-Q(S_{\widetilde{\zeta}-},h(\widetilde{\zeta}-))\right]
=\sum_{j\in\mathcal{J}(\widetilde{\zeta})}\tau_j^2\Gamma^{\Xi}_j(\operatorname{dist}(x,\widetilde{x})), E ζ [ Q ( S ζ , h ( ζ )) − Q ( S ζ − , h ( ζ − )) ] = j ∈ J ( ζ ) ∑ τ j 2 Γ j Ξ ( dist ( x , x )) , where E ζ ~ \mathbb{E}_{\widetilde{\zeta}} E ζ integrates over ω ζ ~ \omega_{\widetilde{\zeta}} ω ζ with all other data fixed.
Proof. Write J = J ( ζ ~ ) = { j 1 , … , J } \mathcal{J}=\mathcal{J}(\widetilde{\zeta})=\{j_1,\ldots,J\} J = J ( ζ ) = { j 1 , … , J } and drop the arrows. By (56) and (4.11), each first-order term uses only the data of its scale and each second-order term only the data of its two scales. Terms with all indices outside J \mathcal{J} J are therefore unchanged. Old terms with an index in J \mathcal{J} J vanish because h j ( ζ ~ − ) = 0 h_j(\widetilde{\zeta}-)=0 h j ( ζ − ) = 0 and q a b = 0 q_{ab}=0 q ab = 0 when h a h b = 0 h_a h_b=0 h a h b = 0 . In the reversed case the new centers of scale m ∈ J m\in\mathcal{J} m ∈ J are Ψ m ( y m + ξ m ) \Psi_m(y^m+\xi^m) Ψ m ( y m + ξ m ) with Ψ m \Psi_m Ψ m as in Lemma 6.4 (c2); in the forward case Ψ m = id \Psi_m=\operatorname{id} Ψ m = id . In both cases Ψ m \Psi_m Ψ m , the particles, and all data of scales outside J \mathcal{J} J are independent of ( ξ m , η m ) (\xi^m,\eta^m) ( ξ m , η m ) and of the fresh data of the scales finer than m m m . We treat the new terms.
First order, b ∈ J b\in\mathcal{J} b ∈ J : τ b ∇ Ξ b ( z ) ⋅ δ u b \tau_b\nabla\Xi_b(z)\cdot\delta u_b τ b ∇ Ξ b ( z ) ⋅ δ u b is linear in η b \eta^b η b , and every other quantity in it is independent of η b \eta^b η b . Integrating first over η b \eta^b η b gives zero.
Mixed, a ≠ b a\neq b a = b with { a , b } ∩ J ≠ ∅ \{a,b\}\cap\mathcal{J}\neq\varnothing { a , b } ∩ J = ∅ : then m : = max ( a , b ) ∈ J m:=\max(a,b)\in\mathcal{J} m := max ( a , b ) ∈ J , since J \mathcal{J} J is an upper set. By (56) and (4.11), B a B b Ξ b \mathcal{B}_a\mathcal{B}_b\Xi_b B a B b Ξ b is linear in η m \eta^m η m when all centers and the other coefficients are fixed, and the centers of the scales a , b a,b a , b do not depend on η m \eta^m η m (those of scale m m m depend on ξ m \xi^m ξ m and on Ψ m \Psi_m Ψ m ; those of the coarser scale do not involve scale m m m ). Integrating first over η m \eta^m η m gives zero.
Diagonal, a = b ∈ J a=b\in\mathcal{J} a = b ∈ J : the new term is τ b 2 2 ( D 2 Ξ b ( z ) [ δ u b , δ u b ] + ∇ Ξ b ( z ) ⋅ δ ( u b ⋅ ∇ ) u b ) \frac{\tau_b^2}{2}\left(D^2\Xi_b(z)[\delta u_b,\delta u_b]+\nabla\Xi_b(z)\cdot\delta(u_b\cdot\nabla)u_b\right) 2 τ b 2 ( D 2 Ξ b ( z ) [ δ u b , δ u b ] + ∇ Ξ b ( z ) ⋅ δ ( u b ⋅ ∇ ) u b ) . Integrating over ( ξ b , η b ) (\xi^b,\eta^b) ( ξ b , η b ) with Ψ b \Psi_b Ψ b and the particles fixed, Lemma 6.2 (the sign of the coefficients is irrelevant) gives τ b 2 2 σ 2 A b 2 D ( z / ℓ b ) : D 2 Ξ b ( z ) = τ b 2 Γ b Ξ \frac{\tau_b^2}{2}\sigma^2A_b^2D(z/\ell_b):D^2\Xi_b(z)=\tau_b^2\Gamma_b^\Xi 2 τ b 2 σ 2 A b 2 D ( z / ℓ b ) : D 2 Ξ b ( z ) = τ b 2 Γ b Ξ .
Each term is bounded, so Fubini’s theorem justifies integrating in the stated order; the remaining fresh variables are then integrated trivially. □ \square □
Jumps along the forward trajectories
Along the forward trajectories, we integrate the random data one window at a time. For a renewal time s s s , write the random data as ( ω < s , ω s , ω > s ) (\omega_{<s},\omega_s,\omega_{>s}) ( ω < s , ω s , ω > s ) , the data of the windows starting before, at, and after s s s . Every quantity attached to the process on [ 0 , s ) [0,s) [ 0 , s ) or at s − s- s − , in particular Q s − Q_{s-} Q s − , R s R_s R s , and the configuration S s − S_{s-} S s − , is a function of ω < s \omega_{<s} ω < s alone, because trajectories up to time s s s are determined by the data of the windows active before s s s ; Q s Q_s Q s depends in addition on ω s \omega_s ω s . For every bounded measurable G G G define the partial average
E s [ G ] : = ∫ G ( ω < s , ω s , ω > s ) d P s ( ω s ) , (79) \mathbb{E}_s[G] := \int G(\omega_{<s},\omega_s,\omega_{>s})\,\mathrm{d}P_s(\omega_s), \tag*{(79)} E s [ G ] := ∫ G ( ω < s , ω s , ω > s ) d P s ( ω s ) , (79)
the integral over ω s \omega_s ω s against its law P s P_s P s , a finite product of uniform measures on cubes and of coefficient laws, with all other variables held fixed. By independence and Fubini’s theorem,
E [ G ] = ∬ ( E s G ) ( ω < s , ω > s ) d P < s ( ω < s ) d P > s ( ω > s ) = E [ E s G ] . \mathbb{E}[G] = \iint(\mathbb{E}_sG)(\omega_{<s},\omega_{>s})\,\mathrm{d}P_{<s}(\omega_{<s})\,\mathrm{d}P_{>s}(\omega_{>s}) = \mathbb{E}[\mathbb{E}_sG]. E [ G ] = ∬ ( E s G ) ( ω < s , ω > s ) d P < s ( ω < s ) d P > s ( ω > s ) = E [ E s G ] .
In the last expression, E \mathbb{E} E still denotes expectation over all random data; integration over ω s \omega_s ω s contributes a factor one because E s G \mathbb{E}_sG E s G does not depend on that coordinate. Retain the definition of Γ j \Gamma_j Γ j from (33) .
Proposition 6.6 (Averaged jump at a refresh). At a renewal time s < 1 s<1 s < 1 with 0 < R s ≤ 1 / 4 0<R_s\le1/4 0 < R s ≤ 1/4 , let J ( s ) \mathcal{J}(s) J ( s ) be the set of renewed generations. Then
E s [ Q s − Q s − ] = ∑ j ∈ J ( s ) τ j 2 Γ j ( R s ) , (80) \mathbb{E}_s[Q_s-Q_{s-}] = \sum_{j\in\mathcal{J}(s)}\tau_j^2\Gamma_j(R_s), \tag*{(80)} E s [ Q s − Q s − ] = j ∈ J ( s ) ∑ τ j 2 Γ j ( R s ) , (80) an identity between functions of ω < s \omega_{<s} ω < s . At s = 0 s=0 s = 0 use Q 0 − = Φ ( r 0 ) Q_{0-}=\Phi(r_0) Q 0 − = Φ ( r 0 ) and regard all initial windows as renewals; then both sides are deterministic.
Proof. For s > 0 s>0 s > 0 , the particle positions are functions of ω < s \omega_{<s} ω < s and are independent of the fresh data. Proposition 6.5 in the forward direction, with t 0 = 1 t_0=1 t 0 = 1 and ( Ξ , Ξ j ) = ( Φ , Φ j ) (\Xi,\Xi_j)=(\Phi,\Phi_j) ( Ξ , Ξ j ) = ( Φ , Φ j ) , therefore gives (80) after fixing ω < s \omega_{<s} ω < s . At s = 0 s=0 s = 0 all scales are fresh and Q 0 − = Φ ( r 0 ) Q_{0-}=\Phi(r_0) Q 0 − = Φ ( r 0 ) . The same calculation gives zero means for the first-order and mixed terms, while the scale-j j j diagonal term contributes τ j 2 Γ j ( r 0 ) \tau_j^2\Gamma_j(r_0) τ j 2 Γ j ( r 0 ) .
Richardson dispersion of trajectories
We now count the gains. The lemma below follows from the fact that at every time some scale is comparable to the current separation, and the window of that scale contributes τ j Γ j ≥ c κ \tau_j\Gamma_j\ge c\kappa τ j Γ j ≥ c κ , while all other scales contribute nonnegatively by Lemma 3.1 .
Lemma 7.1 (Accumulation of trajectory gains). For sufficiently small κ \kappa κ , independently of J J J and L ≥ 2 L\ge2 L ≥ 2 , there is c 0 > 0 c_0>0 c 0 > 0 such that, for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and on every realization with 0 < R s ≤ 1 / 4 0<R_s\le1/4 0 < R s ≤ 1/4 for all s ∈ [ 0 , t ] s\in[0,t] s ∈ [ 0 , t ] ,
∑ j = j 0 J ∑ 0 ≤ k < K j k τ j ≤ t τ j 2 Γ j ( R k τ j ) ≥ c 0 κ t . (81) \sum_{j=j_0}^{J}\ \sum_{\substack{0\le k<K_j\\ k\tau_j\le t}}\tau_j^2\Gamma_j(R_{k\tau_j})\ge c_0\kappa t. \tag*{(81)} j = j 0 ∑ J 0 ≤ k < K j k τ j ≤ t ∑ τ j 2 Γ j ( R k τ j ) ≥ c 0 κ t . (81)
At t = 1 t = 1 t = 1 the sum includes only window starts strictly below 1 1 1 , as already enforced by k < K j k < K_j k < K j .
Proof. For s ∈ [ 0 , t ) s \in[0,t) s ∈ [ 0 , t ) let s j ( s ) = τ j ⌊ s / τ j ⌋ s_j(s) = \tau_j\lfloor s/\tau_j\rfloor s j ( s ) = τ j ⌊ s / τ j ⌋ , the start of the active scale-j j j window. Suppose first that R s ≥ ℓ J R_s \ge\ell_J R s ≥ ℓ J . Choose j j j with ℓ j ≤ R s < 2 ℓ j \ell_j \le R_s < 2\ell_j ℓ j ≤ R s < 2 ℓ j if R s ≤ ℓ j 0 R_s \le\ell_{j_0} R s ≤ ℓ j 0 (then j 0 ≤ j ≤ J j_0 \le j \le J j 0 ≤ j ≤ J , because ℓ J ≤ R s ≤ ℓ j 0 \ell_J \le R_s \le\ell_{j_0} ℓ J ≤ R s ≤ ℓ j 0 ); if R s > ℓ j 0 R_s > \ell_{j_0} R s > ℓ j 0 choose j = j 0 j = j_0 j = j 0 . In either case
1 ≤ R s / ℓ j ≤ C 1 , C 1 = max ( 2 , ( 4 ℓ j 0 ) − 1 ) . 1 \le R_s/\ell_j \le C_1, \qquad C_1 = \max(2,(4\ell_{j_0})^{-1}). 1 ≤ R s / ℓ j ≤ C 1 , C 1 = max ( 2 , ( 4 ℓ j 0 ) − 1 ) .
By (18) ,
∣ R s j ( s ) β − R s β ∣ ≤ C H τ j ≤ C H κ R s β . \left|R_{s_j(s)}^\beta- R_s^\beta\right| \le C_H\tau_j \le C_H\kappa R_s^\beta. R s j ( s ) β − R s β ≤ C H τ j ≤ C H κ R s β .
Choose κ \kappa κ so small that this implies R s j ( s ) / R s ∈ [ 1 / 2 , 2 ] R_{s_j(s)}/R_s \in[1/2,2] R s j ( s ) / R s ∈ [ 1/2 , 2 ] . It follows that R s j ( s ) / ℓ j ∈ [ 1 / 2 , 2 C 1 ] R_{s_j(s)}/\ell_j \in[1/2,2C_1] R s j ( s ) / ℓ j ∈ [ 1/2 , 2 C 1 ] and R s j ( s ) ≥ ℓ j / 2 ≥ ℓ J / 2 > 2 ε R_{s_j(s)} \ge\ell_j/2 \ge\ell_J/2 > 2\varepsilon R s j ( s ) ≥ ℓ j /2 ≥ ℓ J /2 > 2 ε . Also ℓ j ≤ 2 R s j ( s ) ≤ L S ‾ ( R s j ( s ) ) \ell_j \le2R_{s_j(s)} \le L\overline{S}(R_{s_j(s)}) ℓ j ≤ 2 R s j ( s ) ≤ L S ( R s j ( s ) ) . Thus, with r = R s j ( s ) r = R_{s_j(s)} r = R s j ( s ) , we have
2 ε < r ≤ 1 / 4 , r / ℓ j ∈ [ 1 / 2 , 2 C 1 ] , w j ( r ) = 1. 2\varepsilon< r \le1/4, \qquad r/\ell_j \in[1/2,2C_1], \qquad w_j(r) = 1. 2 ε < r ≤ 1/4 , r / ℓ j ∈ [ 1/2 , 2 C 1 ] , w j ( r ) = 1.
Applying (35) with a = 1 / 2 a = 1/2 a = 1/2 and b = 2 C 1 b = 2C_1 b = 2 C 1 therefore gives
τ j Γ j ( R s j ( s ) ) ≥ c 1 / 2 , 2 C 1 κ . \tau_j\Gamma_j(R_{s_j(s)}) \ge c_{1/2,2C_1}\kappa. τ j Γ j ( R s j ( s ) ) ≥ c 1/2 , 2 C 1 κ .
Since C 1 C_1 C 1 depends only on the fixed model parameters, this constant is independent of J , L , κ J,L,\kappa J , L , κ , and s s s ; we denote it by c c c . If R s < ℓ J R_s < \ell_J R s < ℓ J , choose j = J j = J j = J . The Lipschitz bound in (17) , applied forward and backward on this window, gives
e − C κ ≤ R s J ( s ) / R s ≤ e C κ . e^{-C\kappa} \le R_{s_J(s)}/R_s \le e^{C\kappa}. e − C κ ≤ R s J ( s ) / R s ≤ e C κ .
For small enough κ \kappa κ , R s J ( s ) ≤ 2 ℓ J R_{s_J(s)} \le2\ell_J R s J ( s ) ≤ 2 ℓ J , and (34) gives the same positive bound. Every radius used is at most 1 / 4 1/4 1/4 by the local condition. Since the remaining Γ j \Gamma_j Γ j are nonnegative, in both cases
∑ j τ j Γ j ( R s j ( s ) ) ≥ c 0 κ . \sum_j \tau_j\Gamma_j(R_{s_j(s)}) \ge c_0\kappa. j ∑ τ j Γ j ( R s j ( s ) ) ≥ c 0 κ .
Integrating the preceding inequality gives
c 0 κ t ≤ ∫ 0 t ∑ j τ j Γ j ( R s j ( s ) ) d s . c_0\kappa t \le\int_0^t \sum_j \tau_j\Gamma_j(R_{s_j(s)})\,ds. c 0 κ t ≤ ∫ 0 t j ∑ τ j Γ j ( R s j ( s ) ) d s .
For each fixed j j j , the integrand τ j Γ j ( R s j ( s ) ) \tau_j\Gamma_j(R_{s_j(s)}) τ j Γ j ( R s j ( s ) ) is constant on I j k I_{jk} I j k , with value τ j Γ j ( R k τ j ) \tau_j\Gamma_j(R_{k\tau_j}) τ j Γ j ( R k τ j ) . Consequently,
∫ 0 t ∑ j τ j Γ j ( R s j ( s ) ) d s = ∑ j ∑ 0 ≤ k < K j k τ j ≤ t ∣ I j k ∩ [ 0 , t ] ∣ τ j Γ j ( R k τ j ) ≤ ∑ j ∑ 0 ≤ k < K j k τ j ≤ t τ j 2 Γ j ( R k τ j ) , \begin{aligned}
\int_0^t \sum_j \tau_j\Gamma_j(R_{s_j(s)})\,ds
= \sum_j \sum_{\substack{0\le k<K_j\\ k\tau_j\le t}}
|I_{jk}\cap[0,t]|\,\tau_j\Gamma_j(R_{k\tau_j}) \\
&\le\sum_j \sum_{\substack{0\le k<K_j\\ k\tau_j\le t}}
\tau_j^2\Gamma_j(R_{k\tau_j}),
\end{aligned} ∫ 0 t j ∑ τ j Γ j ( R s j ( s ) ) d s = j ∑ 0 ≤ k < K j k τ j ≤ t ∑ ∣ I j k ∩ [ 0 , t ] ∣ τ j Γ j ( R k τ j ) ≤ j ∑ 0 ≤ k < K j k τ j ≤ t ∑ τ j 2 Γ j ( R k τ j ) ,
where ∣ I j k ∩ [ 0 , t ] ∣ |I_{jk}\cap[0,t]| ∣ I j k ∩ [ 0 , t ] ∣ denotes the length of the intersection. The inequality uses ∣ I j k ∩ [ 0 , t ] ∣ ≤ τ j |I_{jk}\cap[0,t]| \le\tau_j ∣ I j k ∩ [ 0 , t ] ∣ ≤ τ j and Γ j ≥ 0 \Gamma_j \ge0 Γ j ≥ 0 . This proves (81) , including windows that have started but have not yet ended at time t t t . □
Proof of Theorem 1.1
Proof of Theorem 1.1. First restrict κ \kappa κ by all the fixed smallness conditions in Lemmas 2.2 and 7.1 . Fix a time satisfying (9) . Then R s ≤ 1 / 4 R_s \le1/4 R s ≤ 1/4 for all s ∈ [ 0 , t ] s \in[0,t] s ∈ [ 0 , t ] by Lemma 2.2 , so by Remark 4.5 the configuration S s S_s S s is admissible, and Q s Q_s Q s is defined, for every s ∈ [ 0 , t ] s \in[0,t] s ∈ [ 0 , t ] and every realization.
For fixed J J J , the bound R s ≥ r 0 e − C λ J s R_s \ge r_0 e^{-C\lambda_Js} R s ≥ r 0 e − C λ J s established in the proof of Lemma 2.2 gives log R s ≥ log r 0 − C λ J t \log R_s \ge\log r_0-C\lambda_Jt log R s ≥ log r 0 − C λ J t for every s ∈ [ 0 , t ] s \in[0,t] s ∈ [ 0 , t ] . Thus Φ ( R s ) \Phi(R_s) Φ ( R s ) is bounded on [ 0 , t ] [0,t] [ 0 , t ] , uniformly in the realization. The corrector Q − Φ Q-\Phi Q − Φ is bounded by (75) , and the set of renewal times { k τ j } \{k\tau_j\} { k τ j } in [ 0 , t ] [0,t] [ 0 , t ] is finite and deterministic. On each closed interval between consecutive renewal times, s ↦ Q s s \mapsto Q_s s ↦ Q s is continuous, with the window data and phases of that interval (one-sided limits at the endpoints): the particles and centers are continuous in time by Lemma 2.3 , the phases are affine, and Q Q Q is a continuous function of the configuration and the phases. Therefore, we can integrate the deterministic identity (76) on each interval between consecutive renewal times and add the jumps at the renewal times, obtaining, on every realization,
Q t = Φ ( r 0 ) + ∑ ζ ≤ t ( Q ζ − Q ζ − ) + ∫ 0 t ( E 0 ( s ) + E 3 ( s ) ) d s Q_t=\Phi(r_0)+\sum_{\zeta\le t}(Q_\zeta-Q_{\zeta-})+\int_0^t(E_0(s)+E_3(s))\,\mathrm{d}s Q t = Φ ( r 0 ) + ζ ≤ t ∑ ( Q ζ − Q ζ − ) + ∫ 0 t ( E 0 ( s ) + E 3 ( s )) d s
with the sum over the renewal times in [ 0 , t ] [0,t] [ 0 , t ] . All terms are bounded uniformly in the realization, so we may integrate over the random data as well. For each renewal time ζ \zeta ζ , Fubini’s theorem gives E [ Q ζ − Q ζ − ] = E [ E ζ [ Q ζ − Q ζ − ] ] \mathbb{E}[Q_\zeta-Q_{\zeta-}]=\mathbb{E}[\mathbb{E}_\zeta[Q_\zeta-Q_{\zeta-}]] E [ Q ζ − Q ζ − ] = E [ E ζ [ Q ζ − Q ζ − ]] , and Proposition 6.6 evaluates the inner average. At t = 1 t=1 t = 1 use the left limits Q 1 − Q_{1-} Q 1 − if necessary. Then
E Q t = Φ ( r 0 ) + E ∑ j , k : k τ j ≤ t τ j 2 Γ j ( R k τ j ) + E ∫ 0 t ( E 0 ( s ) + E 3 ( s ) ) d s (82) \mathbb{E}Q_t=\Phi(r_0)+\mathbb{E}\sum_{\substack{j,k:\ k\tau_j\le t}}\tau_j^2\Gamma_j(R_{k\tau_j})+\mathbb{E}\int_0^t(E_0(s)+E_3(s))\,\mathrm{d}s \tag*{(82)} E Q t = Φ ( r 0 ) + E j , k : k τ j ≤ t ∑ τ j 2 Γ j ( R k τ j ) + E ∫ 0 t ( E 0 ( s ) + E 3 ( s )) d s (82)
where the window-index restrictions are as in Lemma 7.1 . The time-zero renewals are included in the sum. Since r 0 ≥ ℓ J > 2 ε r_0\ge\ell_J>2\varepsilon r 0 ≥ ℓ J > 2 ε , Φ ( r 0 ) = r 0 β \Phi(r_0)=r_0^\beta Φ ( r 0 ) = r 0 β . It follows from (74) and (81) that
E Q t ≥ r 0 β + [ c 0 κ − C L − β − C κ 2 ( 1 + log L ) 2 ] t . (83) \mathbb{E}Q_t\ge r_0^\beta+[c_0\kappa-CL^{-\beta}-C\kappa^2(1+\log L)^2]t. \tag*{(83)} E Q t ≥ r 0 β + [ c 0 κ − C L − β − C κ 2 ( 1 + log L ) 2 ] t . (83)
Choose a fixed K ≥ 1 K\ge1 K ≥ 1 sufficiently large that C / K ≤ c 0 / 4 C/K\le c_0/4 C / K ≤ c 0 /4 , and set
L = ( K / κ ) 1 / β . (84) L=(K/\kappa)^{1/\beta}. \tag*{(84)} L = ( K / κ ) 1/ β . (84)
Decrease the permissible κ 0 \kappa_0 κ 0 so that L ≥ 2 L\ge2 L ≥ 2 and C κ ( 1 + log L ) 2 ≤ c 0 / 4 C\kappa(1+\log L)^2\le c_0/4 C κ ( 1 + log L ) 2 ≤ c 0 /4 whenever κ ≤ κ 0 \kappa\le\kappa_0 κ ≤ κ 0 . This is possible because κ ( 1 + ∣ log κ ∣ ) 2 → 0 \kappa(1+|\log\kappa|)^2\to0 κ ( 1 + ∣ log κ ∣ ) 2 → 0 . Equation (83) then becomes
E Q t ≥ r 0 β + ( c 0 / 2 ) κ t . (85) \mathbb{E}Q_t\ge r_0^\beta+(c_0/2)\kappa t. \tag*{(85)} E Q t ≥ r 0 β + ( c 0 /2 ) κ t . (85)
For this choice of L L L , (75) and (29) imply
Q t ≤ ( 1 + δ κ ) R t β + δ κ ℓ J β , δ κ = C κ ( 1 + ∣ log κ ∣ ) , Q_t\le(1+\delta_\kappa)R_t^\beta+\delta_\kappa\ell_J^\beta,\qquad\delta_\kappa=C\kappa(1+|\log\kappa|), Q t ≤ ( 1 + δ κ ) R t β + δ κ ℓ J β , δ κ = C κ ( 1 + ∣ log κ ∣ ) ,
after increasing the fixed constant C C C and decreasing κ 0 \kappa_0 κ 0 if necessary. Combining this with (85) proves (10) . Make a final decrease of κ 0 \kappa_0 κ 0 to ensure δ κ ≤ 1 / 4 \delta_\kappa\le1/4 δ κ ≤ 1/4 . Since ℓ J ≤ r 0 \ell_J\le r_0 ℓ J ≤ r 0 , it follows that E R t β ≥ c ( r 0 β + κ t ) \mathbb{E}R_t^\beta\ge c(r_0^\beta+\kappa t) E R t β ≥ c ( r 0 β + κ t ) . Convexity of q ↦ q 2 / β q\mapsto q^{2/\beta} q ↦ q 2/ β now gives
E R t 2 = E [ ( R t β ) 2 / β ] ≥ ( E R t β ) 2 / β ≥ c ( r 0 β + κ t ) 2 / β . \mathbb{E}R_t^2=\mathbb{E}[(R_t^\beta)^{2/\beta}]\ge(\mathbb{E}R_t^\beta)^{2/\beta}\ge c(r_0^\beta+\kappa t)^{2/\beta}. E R t 2 = E [( R t β ) 2/ β ] ≥ ( E R t β ) 2/ β ≥ c ( r 0 β + κ t ) 2/ β .
Finally, integrating (18) gives R t β ≤ r 0 β + C H t R_t^\beta\le r_0^\beta+ C_Ht R t β ≤ r 0 β + C H t on every realization, which implies the upper bound in (11) . This completes the proof. □
Proof of Corollary 1.2. When α = 1 / 3 \alpha= 1/3 α = 1/3 , β = 2 / 3 \beta= 2/3 β = 2/3 and 2 / β = 3 2/\beta= 3 2/ β = 3 . If t ≥ r 0 β / κ t \ge r_0^\beta/\kappa t ≥ r 0 β / κ , then r 0 β ≤ κ t ≤ t r_0^\beta\le\kappa t \le t r 0 β ≤ κ t ≤ t , since κ 0 < 1 \kappa_0 < 1 κ 0 < 1 . Apply (11) . □
Advection–diffusion equations with bounded incompressible drift
We work on a torus T n \mathbb{T}^n T n (n = d n = d n = d or n = 2 d n = 2d n = 2 d ), where integrations by parts produce no boundary terms. We recall two standard facts and prove the consequences used later. Only the boundedness and the incompressibility of the drift enter, and nothing in this section is quantitative.
Notation. Functions are real-valued; ( ⋅ , ⋅ ) (\cdot,\cdot) ( ⋅ , ⋅ ) and ∥ ⋅ ∥ \| \cdot\| ∥ ⋅ ∥ are the inner product and norm of L 2 ( T n ) L^2(\mathbb{T}^n) L 2 ( T n ) , H 1 = H 1 ( T n ) H^1 = H^1(\mathbb{T}^n) H 1 = H 1 ( T n ) , H − 1 H^{-1} H − 1 is its dual, and ⟨ ⋅ , ⋅ ⟩ \langle\cdot,\cdot\rangle ⟨ ⋅ , ⋅ ⟩ the duality pairing. Let I = [ T 0 , T 1 ] I = [T_0,T_1] I = [ T 0 , T 1 ] be a compact interval and μ > 0 \mu> 0 μ > 0 . A drift is a Borel map b : I × T n → R n b : I \times\mathbb{T}^n \to\mathbb{R}^n b : I × T n → R n such that
(D) ∣ b ∣ ≤ B 0 |b| \le B_0 ∣ b ∣ ≤ B 0 ,
∫ T n b ( s , Y ) ⋅ ∇ φ ( Y ) d Y = 0 for all φ ∈ C ∞ ( T n ) and a.e. s ∈ I . \int_{\mathbb{T}^n} b(s,Y) \cdot\nabla\varphi(Y)\,dY = 0 \qquad\text{for all } \varphi\in C^\infty(\mathbb{T}^n) \text{ and a.e. }s \in I. ∫ T n b ( s , Y ) ⋅ ∇ φ ( Y ) d Y = 0 for all φ ∈ C ∞ ( T n ) and a.e. s ∈ I .
Since C ∞ ( T n ) C^\infty(\mathbb{T}^n) C ∞ ( T n ) contains a countable set that is dense in C 1 C^1 C 1 , there is one null set N ⊂ I N \subset I N ⊂ I such that the identity in (D) holds for all φ ∈ C ∞ \varphi\in C^\infty φ ∈ C ∞ and all s ∉ N s \notin N s ∈ / N ; for such s s s it holds for all φ ∈ W 1 , 1 ( T n ) \varphi\in W^{1,1}(\mathbb{T}^n) φ ∈ W 1 , 1 ( T n ) , since C ∞ C^\infty C ∞ is dense in W 1 , 1 W^{1,1} W 1 , 1 and b ( s ) ∈ L ∞ b(s) \in L^\infty b ( s ) ∈ L ∞ . If v , φ ∈ H 1 v,\varphi\in H^1 v , φ ∈ H 1 , then v φ ∈ W 1 , 1 v\varphi\in W^{1,1} v φ ∈ W 1 , 1 with ∇ ( v φ ) = φ ∇ v + v ∇ φ \nabla(v\varphi) = \varphi\nabla v + v\nabla\varphi ∇ ( v φ ) = φ ∇ v + v ∇ φ , and therefore, for s ∉ N s \notin N s ∈ / N ,
( b ( s ) ⋅ ∇ v , φ ) = − ( b ( s ) ⋅ ∇ φ , v ) , ( b ( s ) ⋅ ∇ v , v ) = 0. (86) \begin{aligned}
(b(s)\cdot\nabla v,\varphi) &= -(b(s)\cdot\nabla\varphi,v), \\
(b(s)\cdot\nabla v,v) &= 0. \tag*{(86)}
\end{aligned} ( b ( s ) ⋅ ∇ v , φ ) ( b ( s ) ⋅ ∇ v , v ) = − ( b ( s ) ⋅ ∇ φ , v ) , = 0. (86)
We consider the equation
∂ s w + b ⋅ ∇ w = μ Δ w on I × T n . (87) \partial_s w + b\cdot\nabla w = \mu\Delta w \qquad\text{on } I\times\mathbb{T}^n. \tag*{(87)} ∂ s w + b ⋅ ∇ w = μ Δ w on I × T n . (87)
Definition 8.1 (Energy solutions). An energy solution of (87) on I I I is a function w ∈ L 2 ( I ; H 1 ) w \in L^2(I;H^1) w ∈ L 2 ( I ; H 1 ) whose distributional derivative satisfies ∂ s w ∈ L 2 ( I ; H − 1 ) \partial_s w \in L^2(I;H^{-1}) ∂ s w ∈ L 2 ( I ; H − 1 ) and such that, for a.e. s ∈ I s \in I s ∈ I ,
⟨ ∂ s w ( s ) , φ ⟩ + ( b ( s ) ⋅ ∇ w ( s ) , φ ) + μ ( ∇ w ( s ) , ∇ φ ) = 0 for all φ ∈ H 1 . (88) \langle\partial_s w(s),\varphi\rangle+ (b(s)\cdot\nabla w(s),\varphi) + \mu(\nabla w(s),\nabla\varphi) = 0 \qquad\text{for all }\varphi\in H^1. \tag*{(88)} ⟨ ∂ s w ( s ) , φ ⟩ + ( b ( s ) ⋅ ∇ w ( s ) , φ ) + μ ( ∇ w ( s ) , ∇ φ ) = 0 for all φ ∈ H 1 . (88)
We use two standard facts.
(E1) The Lions–Magenes lemma. If w , w ~ ∈ L 2 ( I ; H 1 ) w,\widetilde w \in L^2(I;H^1) w , w ∈ L 2 ( I ; H 1 ) and ∂ s w , ∂ s w ~ ∈ L 2 ( I ; H − 1 ) \partial_s w,\partial_s\widetilde w \in L^2(I;H^{-1}) ∂ s w , ∂ s w ∈ L 2 ( I ; H − 1 ) , then, after modification on null sets, w , w ~ ∈ C ( I ; L 2 ) w,\widetilde w \in C(I;L^2) w , w ∈ C ( I ; L 2 ) ; there are w k ∈ C 1 ( I ; H 1 ) w_k \in C^1(I;H^1) w k ∈ C 1 ( I ; H 1 ) with w k → w w_k \to w w k → w in L 2 ( I ; H 1 ) L^2(I;H^1) L 2 ( I ; H 1 ) and in C ( I ; L 2 ) C(I;L^2) C ( I ; L 2 ) , and ∂ s w k → ∂ s w \partial_s w_k \to\partial_s w ∂ s w k → ∂ s w in L 2 ( I ; H − 1 ) L^2(I;H^{-1}) L 2 ( I ; H − 1 ) ; and s ↦ ( w ( s ) , w ~ ( s ) ) s \mapsto(w(s),\widetilde w(s)) s ↦ ( w ( s ) , w ( s )) is absolutely continuous with derivative ⟨ ∂ s w , w ~ ⟩ + ⟨ ∂ s w ~ , w ⟩ \langle\partial_s w,\widetilde w\rangle+ \langle\partial_s\widetilde w,w\rangle ⟨ ∂ s w , w ⟩ + ⟨ ∂ s w , w ⟩ . See [17 §5.9.2, Theorem 3 and its proof] , where the case w ~ = w \widetilde w = w w = w is stated; the general case follows by polarization. The approximations w k w_k w k are obtained there by extending w w w to a larger interval and mollifying in time; extending by even reflection about T 0 T_0 T 0 and T 1 T_1 T 1 preserves the hypotheses, which gives the stated convergence up to the endpoints. The proof is written for H 0 1 H_0^1 H 0 1 of a bounded domain and applies verbatim on T n \mathbb{T}^n T n .
(E2) Existence. For every w 0 ∈ L 2 w_0 \in L^2 w 0 ∈ L 2 there is an energy solution of (87) on I I I with w ( T 0 ) = w 0 w(T_0) = w_0 w ( T 0 ) = w 0 . This is proved by Galerkin approximation, as in [17 §7.1.2, Theorems 1–3] , with the spaces spanned by { e 2 π i k ⋅ Y : ∣ k ∣ ≤ N } \{e^{2\pi i k\cdot Y}: |k| \le N\} { e 2 π ik ⋅ Y : ∣ k ∣ ≤ N } in place of those spanned by the Dirichlet eigenfunctions (the orthogonal projections onto these spaces are contractions in H 1 H^{1} H 1 ). The energy estimate holds because ( b ⋅ ∇ v , v ) = 0 (\mathbf{b}\cdot\nabla v,v)=0 ( b ⋅ ∇ v , v ) = 0 by (86) , and only the boundedness of b \mathbf{b} b is used. The weak formulation obtained there (for each φ ∈ H 1 \varphi\in H^{1} φ ∈ H 1 , for a.e. s s s ) implies (88) since H 1 H^{1} H 1 is separable.
We always use the representative in C ( I ; L 2 ) C(I;L^{2}) C ( I ; L 2 ) . The proofs of the five lemmas of this section are standard; they are collected in Appendix A.
Lemma 8.2 (Basic properties). Let b \mathbf{b} b satisfy (D) and let w w w be an energy solution of (87) on I I I .
(a) ∥ w ( s ) ∥ 2 2 + 2 μ ∫ T 0 s ∥ ∇ w ∥ 2 2 = ∥ w ( T 0 ) ∥ 2 2 \lVert w(s)\rVert_{2}^{2}+2\mu\int_{T_{0}}^{s}\lVert\nabla w\rVert_{2}^{2}=\lVert w(T_{0})\rVert_{2}^{2} ∥ w ( s ) ∥ 2 2 + 2 μ ∫ T 0 s ∥ ∇ w ∥ 2 2 = ∥ w ( T 0 ) ∥ 2 2 for s ∈ I s\in I s ∈ I . Consequently the energy solution with a given initial value is unique, and it depends linearly on the initial value.
(b) ∫ w ( s ) = ∫ w ( T 0 ) \int w(s)=\int w(T_{0}) ∫ w ( s ) = ∫ w ( T 0 ) for s ∈ I s\in I s ∈ I , and if w ( T 0 ) ≥ 0 w(T_{0})\geq0 w ( T 0 ) ≥ 0 a.e., then w ( s ) ≥ 0 w(s)\geq0 w ( s ) ≥ 0 a.e. for every s ∈ I s\in I s ∈ I .
(c) (Evolution identity.) Let T 0 ≤ s 1 < s 2 ≤ T 1 T_{0}\leq s_{1}<s_{2}\leq T_{1} T 0 ≤ s 1 < s 2 ≤ T 1 and let G : [ s 1 , s 2 ] × T n → R G:[s_{1},s_{2}]\times\mathbb{T}^{n}\to\mathbb{R} G : [ s 1 , s 2 ] × T n → R be such that G G G , ∂ s G \partial_{s}G ∂ s G , ∇ G \nabla G ∇ G and ∇ 2 G \nabla^{2}G ∇ 2 G exist and are continuous on [ s 1 , s 2 ] × T n [s_{1},s_{2}]\times\mathbb{T}^{n} [ s 1 , s 2 ] × T n (one-sided in s s s at the endpoints). Then
∫ G ( s 2 ) w ( s 2 ) − ∫ G ( s 1 ) w ( s 1 ) = ∫ s 1 s 2 ∫ ( ∂ s G + b ⋅ ∇ G + μ Δ G ) w d Y d s \int G(s_{2})w(s_{2})-\int G(s_{1})w(s_{1})=\int_{s_{1}}^{s_{2}}\int\left(\partial_{s}G+\mathbf{b}\cdot\nabla G+\mu\Delta G\right)w\,dY\,ds ∫ G ( s 2 ) w ( s 2 ) − ∫ G ( s 1 ) w ( s 1 ) = ∫ s 1 s 2 ∫ ( ∂ s G + b ⋅ ∇ G + μ Δ G ) w d Y d s For an energy solution w w w with w ( T 0 ) ≥ 0 w(T_{0})\geq0 w ( T 0 ) ≥ 0 and ∫ w ( T 0 ) = 1 \int w(T_{0})=1 ∫ w ( T 0 ) = 1 , each w ( s ) w(s) w ( s ) is therefore a probability density. For T 0 ≤ s 1 < s 2 ≤ T 1 T_{0}\leq s_{1}<s_{2}\leq T_{1} T 0 ≤ s 1 < s 2 ≤ T 1 and Borel sets A , B ⊂ T n A,B\subset\mathbb{T}^{n} A , B ⊂ T n we put
Π s 1 , s 2 ( A ; B ) : = sup { ∫ B w ( s 2 ) } (89) \Pi_{s_{1},s_{2}}(A;B):=\sup\left\{\int_{B}w(s_{2})\right\} \tag*{(89)} Π s 1 , s 2 ( A ; B ) := sup { ∫ B w ( s 2 ) } (89) the supremum over all energy solutions w w w on [ s 1 , s 2 ] [s_{1},s_{2}] [ s 1 , s 2 ] whose initial value w ( s 1 ) ∈ L 2 w(s_{1})\in L^{2} w ( s 1 ) ∈ L 2 is a probability density vanishing a.e. outside A A A (and Π s 1 , s 2 ( A ; B ) : = 0 \Pi_{s_{1},s_{2}}(A;B):=0 Π s 1 , s 2 ( A ; B ) := 0 if A A A is a null set).
Lemma 8.3 (Transfer). Let w w w be an energy solution on [ s 1 , s 2 ] [s_{1},s_{2}] [ s 1 , s 2 ] with w ( s 1 ) ≥ 0 w(s_{1})\geq0 w ( s 1 ) ≥ 0 . Then for Borel sets A , B ⊂ T n A,B\subset\mathbb{T}^{n} A , B ⊂ T n , with A c = T n ∖ A A^{c}=\mathbb{T}^{n}\setminus A A c = T n ∖ A ,
∫ B w ( s 2 ) ≤ ∫ A w ( s 1 ) + Π s 1 , s 2 ( A c ; B ) ∫ A c w ( s 1 ) . \int_{B}w(s_{2})\leq\int_{A}w(s_{1})+\Pi_{s_{1},s_{2}}(A^{c};B)\int_{A^{c}}w(s_{1}). ∫ B w ( s 2 ) ≤ ∫ A w ( s 1 ) + Π s 1 , s 2 ( A c ; B ) ∫ A c w ( s 1 ) .
Lemma 8.4 (Duality). Let b \mathbf{b} b satisfy (D) on [ 0 , T ] [0,T] [ 0 , T ] and put b ← ( s , Y ) : = − b ( T − s , Y ) \overleftarrow{\mathbf{b}}(s,Y):=-\mathbf{b}(T-s,Y) b ( s , Y ) := − b ( T − s , Y ) . Let w w w be an energy solution of (87) with drift b \mathbf{b} b on [ 0 , T ] [0,T] [ 0 , T ] , and ϱ \varrho ϱ an energy solution with drift b ← \overleftarrow{\mathbf{b}} b on [ 0 , T ] [0,T] [ 0 , T ] . Then ( w ( T ) , ϱ ( 0 ) ) = ( w ( 0 ) , ϱ ( T ) ) (w(T),\varrho(0))=(w(0),\varrho(T)) ( w ( T ) , ϱ ( 0 )) = ( w ( 0 ) , ϱ ( T )) .
Lemma 8.5 (Stability). Let b k , b \mathbf{b}_{k},\mathbf{b} b k , b satisfy (D) on I I I with the same B 0 B_{0} B 0 , and b k → b \mathbf{b}_{k}\to\mathbf{b} b k → b a.e. on I × T n I\times\mathbb{T}^{n} I × T n . Let w 0 ∈ L 2 w_{0}\in L^{2} w 0 ∈ L 2 and let w k , w w_{k},w w k , w be the energy solutions with drifts b k , b \mathbf{b}_{k},\mathbf{b} b k , b and initial value w 0 w_{0} w 0 . Then w k ( s ) → w ( s ) w_{k}(s)\to w(s) w k ( s ) → w ( s ) weakly in L 2 L^{2} L 2 for every s ∈ I s\in I s ∈ I .
Lemma 8.6 (Products and isometries). (a) Let b \mathbf{b} b satisfy (D) on I × T d I\times\mathbb{T}^{d} I × T d , and let θ 1 , θ 2 \theta_{1},\theta_{2} θ 1 , θ 2 be energy solutions of (87) on T d \mathbb{T}^{d} T d with drift b \mathbf{b} b . Then b ( s , y , y ′ ) : = ( b ( s , y ) , b ( s , y ′ ) ) \mathbf{b}(s,y,y'):=\left(\mathbf{b}(s,y),\mathbf{b}(s,y')\right) b ( s , y , y ′ ) := ( b ( s , y ) , b ( s , y ′ ) ) satisfies (D) on I × T 2 d I\times\mathbb{T}^{2d} I × T 2 d , with B 0 B_{0} B 0 replaced by 2 B 0 \sqrt{2}B_{0} 2 B 0 (the gradient in (D) and (87) being that of T 2 d \mathbb{T}^{2d} T 2 d , so that b ⋅ ∇ = b ( s , y ) ⋅ ∇ y + b ( s , y ′ ) ⋅ ∇ y ′ \mathbf{b}\cdot\nabla=\mathbf{b}(s,y)\cdot\nabla_{y}+\mathbf{b}(s,y')\cdot\nabla_{y'} b ⋅ ∇ = b ( s , y ) ⋅ ∇ y + b ( s , y ′ ) ⋅ ∇ y ′ ), and ( θ 1 ⊗ θ 2 ) ( s , y , y ′ ) : = θ 1 ( s , y ) θ 2 ( s , y ′ ) (\theta_{1}\otimes\theta_{2})(s,y,y'):=\theta_{1}(s,y)\theta_{2}(s,y') ( θ 1 ⊗ θ 2 ) ( s , y , y ′ ) := θ 1 ( s , y ) θ 2 ( s , y ′ ) is the energy solution on T 2 d \mathbb{T}^{2d} T 2 d with drift
b b b and initial value θ 1 ( T 0 ) ⊗ θ 2 ( T 0 ) \theta_1(T_0) \otimes\theta_2(T_0) θ 1 ( T 0 ) ⊗ θ 2 ( T 0 ) . The constant 1 1 1 is an energy solution for every drift.
(b) Let ι ( Y ) = g Y + a \iota(Y)=gY+a ι ( Y ) = g Y + a with g ∈ O ( n ) g\in O(n) g ∈ O ( n ) , g Z n = Z n g\mathbb{Z}^n=\mathbb{Z}^n g Z n = Z n , and a ∈ R n a\in\mathbb{R}^n a ∈ R n , an isometry of T n \mathbb{T}^n T n . If w w w is an energy solution with drift b b b , then w ∘ ι − 1 w\circ\iota^{-1} w ∘ ι − 1 is the energy solution with drift ( ι ∗ b ) ( s , Y ) : = g b ( s , ι − 1 Y ) (\iota_*b)(s,Y):=g\,b(s,\iota^{-1}Y) ( ι ∗ b ) ( s , Y ) := g b ( s , ι − 1 Y ) , which satisfies (D), and initial value w ( T 0 ) ∘ ι − 1 w(T_0)\circ\iota^{-1} w ( T 0 ) ∘ ι − 1 .
Dispersion of diffusing pairs
For anomalous dissipation we need a separation estimate for solutions of the two-point advection–diffusion equation, in both directions of time. We use the same corrector and jump identities as in § 7 ; exponential moment bounds replace the deterministic control of separation within a window.
The two-point equation and separation growth
Fix J ≥ j 0 J\ge j_0 J ≥ j 0 , t 0 ∈ ( 0 , 1 ] t_0\in(0,1] t 0 ∈ ( 0 , 1 ] and κ ≤ κ ∗ \kappa\le\kappa_* κ ≤ κ ∗ , and let μ = μ J \mu=\mu_J μ = μ J be as in (38) . Let v v v denote either
(F) the forward field u = u ( J ) u=u^{(J)} u = u ( J ) on [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] , with the renewal times k τ j k\tau_j k τ j and the phases h j h_j h j of Section 5; or
(B) the reversed field u ← \overleftarrow{u} u on [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] of Definition 6.3, with its renewal times and phases.
We use the same notation for renewal times in the two cases. Let Z \mathcal{Z} Z be the (finite, deterministic) set of renewal times of all scales in ( 0 , t 0 ) (0,t_0) ( 0 , t 0 ) . No renewal is performed at t 0 t_0 t 0 : the window data and phases at t 0 t_0 t 0 are their left limits. For s ∈ ( 0 , t 0 ) s\in(0,t_0) s ∈ ( 0 , t 0 ) let s j ( s ) s_j(s) s j ( s ) be the largest renewal time of scale j j j in ( 0 , s ] (0,s] ( 0 , s ] , or 0 0 0 if there is none, and s j ( t 0 ) : = s j ( t 0 − ) s_j(t_0):=s_j(t_0-) s j ( t 0 ) := s j ( t 0 − ) ; in both cases 0 ≤ s − s j ( s ) ≤ τ j 0\le s-s_j(s)\le\tau_j 0 ≤ s − s j ( s ) ≤ τ j . All renewal times lie in τ J N \tau_J\mathbb{N} τ J N in case (F) and in t 0 − τ J N t_0-\tau_J\mathbb{N} t 0 − τ J N in case (B) (Lemma 2.1 ); we call the open intervals between consecutive points of Z ∪ { 0 , t 0 } \mathcal{Z}\cup\{0,t_0\} Z ∪ { 0 , t 0 } slabs. On each slab, v v v is C 3 C^3 C 3 in x x x and C 1 C^1 C 1 in time, with one-sided limits at the endpoints (Lemma 2.3 and Definition 6.3).
We write Y = ( y , y ′ ) ∈ T d × T d = T 2 d Y=(y,y')\in\mathbb{T}^d\times\mathbb{T}^d=\mathbb{T}^{2d} Y = ( y , y ′ ) ∈ T d × T d = T 2 d , R ( Y ) = dist ( y , y ′ ) R(Y)=\operatorname{dist}(y,y') R ( Y ) = dist ( y , y ′ ) , and, when R ( Y ) < 1 / 2 R(Y)<1/2 R ( Y ) < 1/2 , z = z ( Y ) z=z(Y) z = z ( Y ) for the shortest representative of y − y ′ y-y' y − y ′ , so ∣ z ∣ = R |z|=R ∣ z ∣ = R . Put
v ( s , Y ) = ( v ( y , s ) , v ( y ′ , s ) ) , ∇ Y = ( ∇ y , ∇ y ′ ) , Δ Y = Δ y + Δ y ′ , v(s,Y)=(v(y,s),v(y',s)),\qquad\nabla_Y=(\nabla_y,\nabla_{y'}),\qquad\Delta_Y=\Delta_y+\Delta_{y'}, v ( s , Y ) = ( v ( y , s ) , v ( y ′ , s )) , ∇ Y = ( ∇ y , ∇ y ′ ) , Δ Y = Δ y + Δ y ′ ,
T s G = v ( s ) ⋅ ∇ Y G + μ Δ Y G . \mathcal{T}_sG=v(s)\cdot\nabla_YG+\mu\Delta_YG. T s G = v ( s ) ⋅ ∇ Y G + μ Δ Y G .
Thus ∇ Y \nabla_Y ∇ Y and Δ Y \Delta_Y Δ Y are the gradient and the Laplacian of T 2 d \mathbb{T}^{2d} T 2 d , so that v ⋅ ∇ Y = v ( y , s ) ⋅ ∇ y + v ( y ′ , s ) ⋅ ∇ y ′ v\cdot\nabla_Y=v(y,s)\cdot\nabla_y+v(y',s)\cdot\nabla_{y'} v ⋅ ∇ Y = v ( y , s ) ⋅ ∇ y + v ( y ′ , s ) ⋅ ∇ y ′ , and v v v satisfies (D) on T 2 d \mathbb{T}^{2d} T 2 d (Lemmas 2.3 and 8.6 (a)). We study energy solutions of the two-point equation
∂ s ϱ + v ⋅ ∇ Y ϱ = μ Δ Y ϱ on [ 0 , t 0 ] × T 2 d , (90) \partial_s\varrho+v\cdot\nabla_Y\varrho=\mu\Delta_Y\varrho\qquad\text{on }[0,t_0]\times\mathbb{T}^{2d}, \tag*{(90)} ∂ s ϱ + v ⋅ ∇ Y ϱ = μ Δ Y ϱ on [ 0 , t 0 ] × T 2 d , (90)
which is the tracer equation with diffusion in the 2 d 2d 2 d variables Y Y Y , the field acting on y y y and on y ′ y' y ′ separately. For a function G G G of z z z alone, T s G = δ v ⋅ ∇ z G + 2 μ Δ z G \mathcal{T}_sG=\delta v\cdot\nabla_zG+2\mu\Delta_zG T s G = δ v ⋅ ∇ z G + 2 μ Δ z G with δ v = v ( y , s ) − v ( y ′ , s ) \delta v=v(y,s)-v(y',s) δ v = v ( y , s ) − v ( y ′ , s ) , because ∇ y ′ z = − I \nabla_{y'}z=-I ∇ y ′ z = − I .
For 0 < ς ≤ 1 / 4 0<\varsigma\le1/4 0 < ς ≤ 1/4 let D ς \mathcal{D}_{\varsigma} D ς be the set of probability densities ϱ 0 ∈ L 2 ( T 2 d ) \varrho_0\in L^2(\mathbb{T}^{2d}) ϱ 0 ∈ L 2 ( T 2 d ) that vanish a.e. on { R > ς } \{R>\varsigma\} { R > ς } . Thus these densities describe initial pairs of points whose separation is at most ς \varsigma ς : the two particles start close together. For ϱ 0 ∈ D ς \varrho_0\in\mathcal{D}_{\varsigma} ϱ 0 ∈ D ς we denote by ϱ \varrho ϱ the energy solution of (90) on [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] with ϱ ( 0 ) = ϱ 0 \varrho(0)=\varrho_0 ϱ ( 0 ) = ϱ 0 , and write ϱ s = ϱ ( s ) \varrho_s=\varrho(s) ϱ s = ϱ ( s ) and ϱ s ( B ) = ∫ B ϱ s \varrho_s(B)=\int_B\varrho_s ϱ s ( B ) = ∫ B ϱ s . By Lemma 8.2 , each ϱ s \varrho_s ϱ s is a probability density, and Π s , ς s \Pi^s_{s,\varsigma} Π s , ς s of (89) refers to (90) .
Before taking expectations, we record the dependence of the solutions on the random data. Besides measurability, we shall need the fact that the density at a renewal time does not depend on the data newly introduced at that time.
Lemma 9.1 (Dependence on the data). Fix J J J , t 0 t_0 t 0 , case ( F ) (F) ( F ) or ( B ) (B) ( B ) , and ϱ 0 ∈ L 2 ( T 2 d ) \varrho_0 \in L^2(\mathbb{T}^{2d}) ϱ 0 ∈ L 2 ( T 2 d ) . For each s ∈ [ 0 , t 0 ] s \in[0,t_0] s ∈ [ 0 , t 0 ] , ϱ s \varrho_s ϱ s is determined by the restriction of v v v to [ 0 , s ) [0,s) [ 0 , s ) , and for every G ∈ L 2 ( T 2 d ) G \in L^2(\mathbb{T}^{2d}) G ∈ L 2 ( T 2 d ) , ∫ G ϱ s \int G\varrho_s ∫ G ϱ s is a continuous function of the finitely many data variables of the windows meeting [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] . In particular, the function is Borel. The same holds for the energy solutions of (12) , and D J ( θ 0 ) D_J(\theta_0) D J ( θ 0 ) is a Borel function of the data. All of this remains true, with the same proof, when the diffusivity μ J \mu_J μ J is replaced by an arbitrary μ > 0 \mu>0 μ > 0 .
Proof. The restriction of ϱ \varrho ϱ to [ 0 , s ] [0,s] [ 0 , s ] is the energy solution on [ 0 , s ] [0,s] [ 0 , s ] (Lemma 8.2 (a)); hence two drifts that agree a.e. on [ 0 , s ) × T 2 d [0,s)\times\mathbb{T}^{2d} [ 0 , s ) × T 2 d give the same ϱ s \varrho_s ϱ s . The data space of these windows is a compact metric space, a finite product of cubes Q j Q_j Q j and intervals $[-M,M] ({xref}Section 1.2), so sequential continuity suffices, and by {xref}Lemma 8.5 it suffices to show that $v(s,Y)$ is a continuous function of the data for every $Y$ and every $s \notin \mathcal{Z}\cup\{0,t_0\}$; the bound in (D) is uniform by {xref}Lemma 2.3. On each window of scale $j$ the centers solve {xref}(1.5), whose right-hand side $u_{<j}$ is Lipschitz in $x$ with a constant valid for every datum ({xref}Lemma 2.2) and, by induction on $j$, continuous in the data uniformly in $(x,t)$. By Gronwall’s inequality the centers depend continuously on the data, uniformly in time, and since the $W_j^\nu$ are uniformly continuous, $u_j(x,t)$ depends continuously on the data, uniformly in $(x,t)$. In case (B), $\overleftarrow{u}(x,s)=-u(x,t_0-s)$ for $s\notin\mathcal{Z}\cup\{0,t_0\}$. The same argument applies to {xref}(1.12); then $\theta^{(J)}(t_0)$ depends weakly continuously on the data, so $\|\theta^{(J)}(t_0)\|^2$ is lower semicontinuous and, by {xref}(1.13)`, D J ( θ 0 ) D_J(\theta_0) D J ( θ 0 ) is Borel. □
Theorem 9.2 (Richardson growth near the diagonal). *Let d ≥ 2 d\ge2 d ≥ 2 and assume (ND) and (4) . There exist t 1 ∈ ( 0 , 1 ] t_1\in(0,1] t 1 ∈ ( 0 , 1 ] , c E > 0 c_E>0 c E > 0 , κ 1 ∈ ( 0 , κ ∗ ) \kappa_1\in(0,\kappa_*) κ 1 ∈ ( 0 , κ ∗ ) and c 5 , C , C δ > 0 c_5,C,C_\delta>0 c 5 , C , C δ > 0 , depending only on d , α , ψ , R , M , σ d,\alpha,\psi,R,M,\sigma d , α , ψ , R , M , σ , and for each κ ∈ ( 0 , κ 1 ] \kappa\in(0,\kappa_1] κ ∈ ( 0 , κ 1 ] a number ε 1 ( κ ) ∈ ( 0 , 1 ] \varepsilon_1(\kappa)\in(0,1] ε 1 ( κ ) ∈ ( 0 , 1 ] , such that the following holds for all κ ∈ ( 0 , κ 1 ] \kappa\in(0,\kappa_1] κ ∈ ( 0 , κ 1 ] , ε 0 ∈ ( 0 , ε 1 ( κ ) ] \varepsilon_0\in(0,\varepsilon_1(\kappa)] ε 0 ∈ ( 0 , ε 1 ( κ )] , J ≥ J 0 : = max ( j 0 , 5 ) J\ge J_0:=\max(j_0,5) J ≥ J 0 := max ( j 0 , 5 ) and t 0 ∈ ( 0 , 1 ] t_0\in(0,1] t 0 ∈ ( 0 , 1 ] , in both cases ( F ) (F) ( F ) and ( B ) (B) ( B ) . If ϱ 0 ∈ D ς \varrho_0\in\mathcal{D}_\varsigma ϱ 0 ∈ D ς with ς ≤ r ♯ / 8 \varsigma\le r_\sharp/8 ς ≤ r ♯ /8 does not depend on the data, then for every t ∈ ( 0 , min ( t 0 , t 1 ) ] t\in(0,\min(t_0,t_1)] t ∈ ( 0 , min ( t 0 , t 1 )]
( 1 + δ κ ) E ∫ R β ϱ t ≥ c 5 κ t − Err J , (91) (1+\delta_\kappa)\mathbb{E}\int R^\beta\varrho_t \ge c_5\kappa t-\operatorname{Err}_J, \tag*{(91)} ( 1 + δ κ ) E ∫ R β ϱ t ≥ c 5 κ t − Err J , (91) where
Err J : = C ( 1 + log 2 ε r ♯ ) ε J β + C e − c E / μ , \operatorname{Err}_J:=C\left(1+\log\frac{2\varepsilon}{r_\sharp}\right)\varepsilon_J^\beta+Ce^{-c_E/\mu}, Err J := C ( 1 + log r ♯ 2 ε ) ε J β + C e − c E / μ , where δ κ : = C δ κ ( 1 + ∣ log κ ∣ ) ≤ 1 4 \delta_\kappa:=C_\delta\kappa(1+|\log\kappa|)\le\frac14 δ κ := C δ κ ( 1 + ∣ log κ ∣ ) ≤ 4 1 . Since ε / r ♯ \varepsilon/r_\sharp ε / r ♯ does not depend on J J J , Err J → 0 \operatorname{Err}_J\to0 Err J → 0 as J → ∞ J\to\infty J → ∞ , uniformly in t 0 t_0 t 0 , t t t and ϱ 0 \varrho_0 ϱ 0 . Whenever Err J ≤ c 5 κ t / 2 \operatorname{Err}_J\le c_5\kappa t/2 Err J ≤ c 5 κ t /2 ,
E ∫ R β ϱ t ≥ 2 5 c 5 κ t , E ∫ R 2 ϱ t ≥ ( 2 5 c 5 κ t ) 2 / β . (92) \mathbb{E}\int R^\beta\varrho_t\ge\frac25c_5\kappa t,\qquad
\mathbb{E}\int R^2\varrho_t\ge\left(\frac25c_5\kappa t\right)^{2/\beta}. \tag*{(92)} E ∫ R β ϱ t ≥ 5 2 c 5 κ t , E ∫ R 2 ϱ t ≥ ( 5 2 c 5 κ t ) 2/ β . (92)
We prove the theorem in the remainder of this section. First we derive the evolution identity and estimate the errors introduced by the spatial cutoff and by diffusion. We then compare the mass in each separation range with its mass at the preceding renewal, and sum the resulting gains. Throughout, the cutoff parameter from § 3 is L = ( K / κ ) 1 / β L=(K/\kappa)^{1/\beta} L = ( K / κ ) 1/ β and N L = 1 + log L N_L=1+\log L N L = 1 + log L ; the constant K K K will be fixed in the proof.
The cut-off observable and its evolution
The corrected observable is defined near the diagonal. To use it as a test function on the torus, we first localize it in that region. Fix a nonincreasing χ ∘ ∈ C ∞ ( [ 0 , ∞ ) ) \chi_{\circ}\in C^{\infty}([0,\infty)) χ ∘ ∈ C ∞ ([ 0 , ∞ )) with χ ∘ = 1 \chi_{\circ}=1 χ ∘ = 1 on [ 0 , 1 8 ] [0,\frac{1}{8}] [ 0 , 8 1 ] and χ ∘ = 0 \chi_{\circ}=0 χ ∘ = 0 on [ 3 16 , ∞ ) [\frac{3}{16},\infty) [ 16 3 , ∞ ) ; then Y ↦ χ ∘ ( R ( Y ) ) Y\mapsto\chi_{\circ}(R(Y)) Y ↦ χ ∘ ( R ( Y )) is smooth on T 2 d \mathbb{T}^{2d} T 2 d . For s ∈ [ 0 , t 0 ] s\in[0,t_0] s ∈ [ 0 , t 0 ] let Θ s \Theta_s Θ s and h ( s ) h(s) h ( s ) be the current window data and phases, and Θ s − , h ( s − ) \Theta_{s-},h(s-) Θ s − , h ( s − ) their left limits. Identifying the pair Y = ( y , y ′ ) Y=(y,y') Y = ( y , y ′ ) with the particles ( x , x ~ ) (x,\tilde{x}) ( x , x ~ ) of a configuration, put
Q s ♯ ( Y ) : = { χ ∘ ( R ( Y ) ) Q ♯ ( ( Y , Θ s ) , h ( s ) ) , R ( Y ) ≤ 1 4 , 0 , R ( Y ) > 1 4 . Q_s^{\sharp}(Y):=
\begin{cases}
\chi_{\circ}(R(Y))\,Q^{\sharp}((Y,\Theta_s),h(s)), & R(Y)\leq\frac{1}{4},\\
0, & R(Y)>\frac{1}{4}.
\end{cases} Q s ♯ ( Y ) := { χ ∘ ( R ( Y )) Q ♯ (( Y , Θ s ) , h ( s )) , 0 , R ( Y ) ≤ 4 1 , R ( Y ) > 4 1 .
and define Q s − Q_s^{-} Q s − in the same way with ( Θ s − , h ( s − ) ) (\Theta_{s-},h(s-)) ( Θ s − , h ( s − )) . By Corollary 5.3, Q ♯ Q^{\sharp} Q ♯ is defined and continuous up to the diagonal y = y ′ y=y' y = y ′ . The functions Q ♯ Q^{\sharp} Q ♯ , Φ ♯ ( z ) \Phi^{\sharp}(z) Φ ♯ ( z ) , E 0 E_0 E 0 , E 3 E_3 E 3 and their derivatives are defined only on { R < 1 4 } \{R<\frac{1}{4}\} { R < 4 1 } . Every product of one of them with χ ∘ ∘ R \chi_{\circ}\circ R χ ∘ ∘ R or with a derivative of χ ∘ ∘ R \chi_{\circ}\circ R χ ∘ ∘ R (as in Q s ♯ Q_s^{\sharp} Q s ♯ and K s \mathcal{K}_s K s below) is extended by 0 0 0 to { R ≥ 1 4 } \{R\geq\frac{1}{4}\} { R ≥ 4 1 } ; since χ ∘ ∘ R \chi_{\circ}\circ R χ ∘ ∘ R vanishes on { R ≥ 3 16 } \{R\geq\frac{3}{16}\} { R ≥ 16 3 } , this does not affect smoothness.
Lemma 9.3 (Evolution of the cut-off observable). Assume (39) . For fixed data, every ϱ 0 ∈ L 2 ( T 2 d ) \varrho_0\in L^2(\mathbb{T}^{2d}) ϱ 0 ∈ L 2 ( T 2 d ) and t ∈ ( 0 , t 0 ] t\in(0,t_0] t ∈ ( 0 , t 0 ] ,
∫ Q t ♯ ϱ t = ∫ Q 0 ♯ ϱ 0 + ∫ 0 t ∫ K s ϱ s d s + ∑ ζ ~ ∈ Z , ζ ~ ≤ t ∫ ( Q ζ ~ ♯ − Q ζ ~ − ) d ϱ ζ ~ . (93) \int Q_t^{\sharp}\varrho_t=\int Q_0^{\sharp}\varrho_0+\int_0^t\int\mathcal{K}_s\varrho_s\,ds+\sum_{\widetilde{\zeta}\in\mathcal{Z},\,\widetilde{\zeta}\leq t}\int\left(Q_{\widetilde{\zeta}}^{\sharp}-Q_{\widetilde{\zeta}}^{-}\right)d\varrho_{\widetilde{\zeta}}. \tag*{(93)} ∫ Q t ♯ ϱ t = ∫ Q 0 ♯ ϱ 0 + ∫ 0 t ∫ K s ϱ s d s + ζ ∈ Z , ζ ≤ t ∑ ∫ ( Q ζ ♯ − Q ζ − ) d ϱ ζ . (93) where, with everything evaluated at ( ( Y , Θ s ) , h ( s ) ) ((Y,\Theta_s),h(s)) (( Y , Θ s ) , h ( s )) ,
K s : = χ ∘ ( R ) [ E 0 + E 3 + 2 μ Δ Φ ♯ ( z ) + μ Δ Y ( Q ♯ − Φ ♯ ) ] + K s c u t , \mathcal{K}_s:=\chi_{\circ}(R)\left[E_0+E_3+2\mu\Delta\Phi^{\sharp}(z)+\mu\Delta_Y\left(Q^{\sharp}-\Phi^{\sharp}\right)\right]+\mathcal{K}_s^{\mathrm{cut}}, K s := χ ∘ ( R ) [ E 0 + E 3 + 2 μ Δ Φ ♯ ( z ) + μ Δ Y ( Q ♯ − Φ ♯ ) ] + K s cut , K s c u t : = Q ♯ T s ( χ ∘ ∘ R ) + 2 μ ( ∇ y ( χ ∘ ∘ R ) ⋅ ∇ y Q ♯ + ∇ y ′ ( χ ∘ ∘ R ) ⋅ ∇ y ′ Q ♯ ) , \mathcal{K}_s^{\mathrm{cut}}:=Q^{\sharp}\mathcal{T}_s(\chi_{\circ}\circ R)+2\mu\left(\nabla_y(\chi_{\circ}\circ R)\cdot\nabla_yQ^{\sharp}+\nabla_{y'}(\chi_{\circ}\circ R)\cdot\nabla_{y'}Q^{\sharp}\right), K s cut := Q ♯ T s ( χ ∘ ∘ R ) + 2 μ ( ∇ y ( χ ∘ ∘ R ) ⋅ ∇ y Q ♯ + ∇ y ′ ( χ ∘ ∘ R ) ⋅ ∇ y ′ Q ♯ ) , E 0 , E 3 E_0,E_3 E 0 , E 3 are those of (73) for ( Φ ♯ , Φ j ♯ ) (\Phi^{\sharp},\Phi_j^{\sharp}) ( Φ ♯ , Φ j ♯ ) , and K s c u t = 0 \mathcal{K}_s^{\mathrm{cut}}=0 K s cut = 0 unless 1 8 ≤ R ≤ 3 16 \frac{1}{8}\leq R\leq\frac{3}{16} 8 1 ≤ R ≤ 16 3 . Moreover, if ϱ 0 \varrho_0 ϱ 0 does not depend on the data, then for every ζ ~ ∈ Z \widetilde{\zeta}\in\mathcal{Z} ζ ∈ Z
E ∫ ( Q ζ ~ ♯ − Q ζ ~ − ) d ϱ ζ ~ = E ∫ χ ∘ ( R ) ∑ j ∈ J ( ζ ~ ) τ j 2 Γ j ♯ ( R ) d ϱ ζ ~ . (94) \mathbb{E}\int\left(Q_{\widetilde{\zeta}}^{\sharp}-Q_{\widetilde{\zeta}}^{-}\right)d\varrho_{\widetilde{\zeta}}
=
\mathbb{E}\int\chi_{\circ}(R)\sum_{j\in J(\widetilde{\zeta})}\tau_j^2\Gamma_j^{\sharp}(R)d\varrho_{\widetilde{\zeta}}. \tag*{(94)} E ∫ ( Q ζ ♯ − Q ζ − ) d ϱ ζ = E ∫ χ ∘ ( R ) j ∈ J ( ζ ) ∑ τ j 2 Γ j ♯ ( R ) d ϱ ζ . (94)
Proof. We apply the weak evolution identity separately between successive renewals and then sum. We first justify this operation, compute the continuous contribution, and finally average the jumps.
On { R < 1 4 } \{R<\frac{1}{4}\} { R < 4 1 } , Q ♯ Q^{\sharp} Q ♯ is C 2 C^2 C 2 in Y Y Y and C 1 C^1 C 1 in the centers and phases, and these derivatives are jointly continuous: Φ b ♯ ( z ) \Phi_b^{\sharp}(z) Φ b ♯ ( z ) is smooth (Lemma 3.3 (b)), u a ∈ C 3 u_a\in C^3 u a ∈ C 3 , B a u b ∈ C 2 ( E ) \mathcal{B}_{a u_b}\in C^2(\mathcal{E}) B a u b ∈ C 2 ( E ) (Definition 4.2 ), and q a b q_{ab} q ab is C 1 C^1 C 1 . On the closure of a slab, the coefficients and shifts are constant, the centers are C 1 C^1 C 1 in time with one-sided limits, and the phases are affine. Since χ ∘ ∘ R \chi_{\circ}\circ R χ ∘ ∘ R is smooth and vanishes for R ≥ 3 16 R\geq\frac{3}{16} R ≥ 16 3 , the function equal to Q ∘ Q^{\circ} Q ∘ with the data of a slab, extended to the closed slab by one-sided limits, satisfies the hypotheses of Lemma 8.2 (c).
Inside a slab, write
c ( Y ) : = χ ∘ ( R ( Y ) ) , q s ( Y ) : = Q ♯ ( ( Y , Θ s ) , h ( s ) ) , Q s ∘ = c q s . c(Y):=\chi_{\circ}(R(Y)), \qquad q_s(Y):=Q^{\sharp}((Y,\Theta_s),h(s)), \qquad Q_s^{\circ}=cq_s. c ( Y ) := χ ∘ ( R ( Y )) , q s ( Y ) := Q ♯ (( Y , Θ s ) , h ( s )) , Q s ∘ = c q s .
At fixed Y Y Y , only the centers and phases vary with s s s ; the coefficients are constant. Since h ˙ j = − 1 \dot{h}_j=-1 h ˙ j = − 1 , the chain rule gives
∂ s q s = ∑ m , i p ˙ i m ( s ) ⋅ ∇ p i m Q ♯ − ∑ j ∂ h j Q ♯ . \partial_s q_s=\sum_{m,i}\dot{p}^{m}_i(s)\cdot\nabla_{p^m_i}Q^{\sharp}-\sum_j\partial_{h_j}Q^{\sharp}. ∂ s q s = m , i ∑ p ˙ i m ( s ) ⋅ ∇ p i m Q ♯ − j ∑ ∂ h j Q ♯ .
Here and below the derivatives of Q ♯ Q^{\sharp} Q ♯ are evaluated at ( ( Y , Θ s ) , h ( s ) ) ((Y,\Theta_s),h(s)) (( Y , Θ s ) , h ( s )) . The phase term is D h Q ♯ D_hQ^{\sharp} D h Q ♯ . The center-motion term, together with the motion of the pair, is the configuration generator:
B Q ♯ = v ⋅ ∇ Y Q ♯ + ∑ m , i p ˙ i m ( s ) ⋅ ∇ p i m Q ♯ . \mathcal{B}Q^{\sharp}=v\cdot\nabla_YQ^{\sharp}+\sum_{m,i}\dot{p}^{m}_i(s)\cdot\nabla_{p^m_i}Q^{\sharp}. B Q ♯ = v ⋅ ∇ Y Q ♯ + m , i ∑ p ˙ i m ( s ) ⋅ ∇ p i m Q ♯ .
In case (B), the centers and generator in these formulas are the reversed ones of Lemma 6.4 (b). Thus, exactly as in the proof of (76) ,
( ∂ s + v ⋅ ∇ Y ) q s = ( D h + B ) Q ♯ = E 0 + E 3 . (\partial_s+v\cdot\nabla_Y)q_s=(D_h+\mathcal{B})Q^{\sharp}=E_0+E_3. ( ∂ s + v ⋅ ∇ Y ) q s = ( D h + B ) Q ♯ = E 0 + E 3 .
The last equality follows from (73) and Corollary 5.3 for y ≠ y ′ y\ne y' y = y ′ . Both sides are continuous up to y = y ′ y=y' y = y ′ , by the regularity just established and Corollary 5.3, so the identity extends to the diagonal.
Since c c c is independent of time, the product rule yields
( ∂ s + T s ) ( c q s ) = c ( ∂ s + v ⋅ ∇ Y ) q s + q s v ⋅ ∇ Y c + μ c Δ Y q s + μ q s Δ Y c + 2 μ ∇ Y c ⋅ ∇ Y q s = c ( E 0 + E 3 + μ Δ Y q s ) + q s T s c + 2 μ ∇ Y c ⋅ ∇ Y q s . \begin{aligned}
(\partial_s+\mathcal{T}_s)(cq_s)=c(\partial_s+v\cdot\nabla_Y)q_s+q_s\,v\cdot\nabla_Yc
+\mu c\mathop{\Delta}\nolimits_Yq_s+\mu q_s\mathop{\Delta}\nolimits_Yc+2\mu\nabla_Yc\cdot\nabla_Yq_s \\
&=c(E_0+E_3+\mu\mathop{\Delta}\nolimits_Yq_s)+q_s\mathcal{T}_sc+2\mu\nabla_Yc\cdot\nabla_Yq_s.
\end{aligned} ( ∂ s + T s ) ( c q s ) = c ( ∂ s + v ⋅ ∇ Y ) q s + q s v ⋅ ∇ Y c + μ c Δ Y q s + μ q s Δ Y c + 2 μ ∇ Y c ⋅ ∇ Y q s = c ( E 0 + E 3 + μ Δ Y q s ) + q s T s c + 2 μ ∇ Y c ⋅ ∇ Y q s .
To separate the main observable from its correctors, write
Δ Y q s = Δ Y Φ ♯ ( z ) + Δ Y ( q s − Φ ♯ ( z ) ) . \mathop{\Delta}\nolimits_Yq_s=\mathop{\Delta}\nolimits_Y\Phi^{\sharp}(z)+\mathop{\Delta}\nolimits_Y(q_s-\Phi^{\sharp}(z)). Δ Y q s = Δ Y Φ ♯ ( z ) + Δ Y ( q s − Φ ♯ ( z )) .
On functions of z = y − y ′ z=y-y' z = y − y ′ , ∇ y = ∇ z \nabla_y=\nabla_z ∇ y = ∇ z and ∇ y ′ = − ∇ z \nabla_{y'}=-\nabla_z ∇ y ′ = − ∇ z , so each Laplacian contributes the same term:
Δ Y Φ ♯ ( z ) = ( Δ y + Δ y ′ ) Φ ♯ ( z ) = 2 Δ z Φ ♯ ( z ) . \mathop{\Delta}\nolimits_Y\Phi^{\sharp}(z)=(\mathop{\Delta}\nolimits_y+\mathop{\Delta}\nolimits_{y'})\Phi^{\sharp}(z)=2\mathop{\Delta}\nolimits_z\Phi^{\sharp}(z). Δ Y Φ ♯ ( z ) = ( Δ y + Δ y ′ ) Φ ♯ ( z ) = 2 Δ z Φ ♯ ( z ) .
Consequently,
( ∂ s + T s ) Q s ∘ = c [ E 0 + E 3 + 2 μ Δ z Φ ♯ ( z ) + μ Δ Y ( q s − Φ ♯ ( z ) ) ] + q s T s c + 2 μ ∇ Y c ⋅ ∇ Y q s = K s . (\partial_s+\mathcal{T}_s)Q_s^{\circ}
=c\left[E_0+E_3+2\mu\mathop{\Delta}\nolimits_z\Phi^{\sharp}(z)+\mu\mathop{\Delta}\nolimits_Y(q_s-\Phi^{\sharp}(z))\right]
+q_s\mathcal{T}_sc+2\mu\nabla_Yc\cdot\nabla_Yq_s=\mathcal{K}_s. ( ∂ s + T s ) Q s ∘ = c [ E 0 + E 3 + 2 μ Δ z Φ ♯ ( z ) + μ Δ Y ( q s − Φ ♯ ( z )) ] + q s T s c + 2 μ ∇ Y c ⋅ ∇ Y q s = K s .
Indeed, the last two terms are precisely K s c u t \mathcal{K}^{\mathrm{cut}}_s K s cut , since
∇ Y c ⋅ ∇ Y q s = ∇ y c ⋅ ∇ y q s + ∇ y ′ c ⋅ ∇ y ′ q s . \nabla_Yc\cdot\nabla_Yq_s=\nabla_yc\cdot\nabla_yq_s+\nabla_{y'}c\cdot\nabla_{y'}q_s. ∇ Y c ⋅ ∇ Y q s = ∇ y c ⋅ ∇ y q s + ∇ y ′ c ⋅ ∇ y ′ q s .
These are exactly the terms in which a spatial derivative falls on the cutoff. All products are extended by zero as specified above.
We now sum over the slabs. Let 0 = ζ ~ 0 < ζ ~ 1 < ⋯ < ζ ~ n < ζ ~ n + 1 = t 0=\widetilde{\zeta}_0<\widetilde{\zeta}_1<\cdots<\widetilde{\zeta}_n<\widetilde{\zeta}_{n+1}=t 0 = ζ 0 < ζ 1 < ⋯ < ζ n < ζ n + 1 = t be the points of Z ∩ ( 0 , t ) \mathcal{Z}\cap(0,t) Z ∩ ( 0 , t ) together with 0 0 0 and t t t . Apply Lemma 8.2 (c) on each [ ζ ~ k , ζ ~ k + 1 ] [\widetilde{\zeta}_k,\widetilde{\zeta}_{k+1}] [ ζ k , ζ k + 1 ] to the function equal to Q ∘ Q^\circ Q ∘ with the data of ( ζ ~ k , ζ ~ k + 1 ) (\widetilde{\zeta}_k,\widetilde{\zeta}_{k+1}) ( ζ k , ζ k + 1 ) , which takes the value Q ζ ~ k ∘ Q^\circ_{\widetilde{\zeta}_k} Q ζ k ∘ at ζ ~ k \widetilde{\zeta}_k ζ k and Q ζ ~ k + 1 − ∘ Q^\circ_{\widetilde{\zeta}_{k+1}-} Q ζ k + 1 − ∘ at ζ ~ k + 1 \widetilde{\zeta}_{k+1} ζ k + 1 . Summing, the boundary terms telescope up to the jumps at the points of Z \mathcal{Z} Z ; if t ∈ Z t \in\mathcal{Z} t ∈ Z , its jump is added, and if t ∉ Z t \notin\mathcal{Z} t ∈ / Z , then Q t − ∘ = Q t ∘ Q_{t^-}^{\circ}=Q_t^{\circ} Q t − ∘ = Q t ∘ (also for t = t 0 t=t_0 t = t 0 , by convention).
To obtain (94) , we condition on all data except those refreshed at ζ ~ \widetilde{\zeta} ζ . By Lemma 9.1, ϱ ζ ~ ♯ \varrho_{\widetilde{\zeta}}^{\sharp} ϱ ζ ♯ is determined by v \mathbf{v} v on [ 0 , ζ ~ ) [0,\widetilde{\zeta}) [ 0 , ζ ) : in case (F) by the data of windows starting before ζ ~ \widetilde{\zeta} ζ , in case (B) by ω ζ ~ c \omega_{\widetilde{\zeta}}^{c} ω ζ c (Lemma 6.4(c1)). In both cases it does not depend on the fresh data ω ζ ~ \omega_{\widetilde{\zeta}} ω ζ . At a fixed Y Y Y with R ( Y ) ≤ 3 16 R(Y)\leq\frac{3}{16} R ( Y ) ≤ 16 3 , Proposition 6.5 gives
E ζ ~ [ Q ζ ~ ♯ − Q ζ ~ − ♯ ] ( Y ) = ∑ j ∈ J ( ζ ~ ) τ j 2 F j ♯ ( R ( Y ) ) . \mathbb{E}_{\widetilde{\zeta}}\left[Q_{\widetilde{\zeta}}^{\sharp}-Q_{\widetilde{\zeta}^{-}}^{\sharp}\right](Y)=\sum_{j\in\mathcal{J}(\widetilde{\zeta})}\tau_j^2\mathbb{F}_j^{\sharp}(R(Y)). E ζ [ Q ζ ♯ − Q ζ − ♯ ] ( Y ) = j ∈ J ( ζ ) ∑ τ j 2 F j ♯ ( R ( Y )) .
Multiply by χ ∘ ( R ) ϱ ζ ~ ♯ ( Y ) \chi_{\circ}(R)\varrho_{\widetilde{\zeta}}^{\sharp}(Y) χ ∘ ( R ) ϱ ζ ♯ ( Y ) , integrate in Y Y Y , and use Fubini’s theorem and E = E E ζ ~ \mathbb{E}=\mathbb{E}\mathbb{E}_{\widetilde{\zeta}} E = E E ζ . This is justified because the jump is continuous in Y Y Y and in the data and bounded uniformly in both (by (75) , for fixed κ \kappa κ and L L L ), because ∥ ϱ ζ ~ ♯ ∥ L 1 ≤ ∥ ϱ ζ ~ ♯ ∥ ≤ ∥ ϱ 0 ∥ \|\varrho_{\widetilde{\zeta}}^{\sharp}\|_{L^1}\leq\|\varrho_{\widetilde{\zeta}}^{\sharp}\|\leq\|\varrho_0\| ∥ ϱ ζ ♯ ∥ L 1 ≤ ∥ ϱ ζ ♯ ∥ ≤ ∥ ϱ 0 ∥ (Lemma 8.2(a)), and because the map from the data to ϱ ζ ~ ♯ ∈ L 2 ( T 2 ) \varrho_{\widetilde{\zeta}}^{\sharp}\in L^2(\mathbb{T}^2) ϱ ζ ♯ ∈ L 2 ( T 2 ) is Borel (Lemma 9.1), so that ( ω , Y ) ↦ ϱ ζ ~ ♯ [ ω ] ( Y ) (\omega,Y)\mapsto\varrho_{\widetilde{\zeta}}^{\sharp}[\omega](Y) ( ω , Y ) ↦ ϱ ζ ♯ [ ω ] ( Y ) has a jointly measurable version. □ \square □
The cutoff contributes only on an annulus separated from the diagonal. On this annulus the singular derivative bounds become uniform, and the diffusivity absorbs the growth in the number of active scales.
Lemma 9.4 (Cut-off error). Assume (39) , κ N L ≤ 1 \kappa N_L\leq1 κ N L ≤ 1 and J ≥ J 0 J\geq J_0 J ≥ J 0 . There is C B C_B C B , independent of J J J , L L L , κ \kappa κ , ε 0 \varepsilon_0 ε 0 , with ∣ K s c u t ∣ ≤ C B 1 { 1 / 8 ≤ R ≤ 3 / 16 } |\mathcal{K}_{s}^{\mathrm{cut}}|\leq C_B\mathbf{1}_{\{1/8\leq R\leq3/16\}} ∣ K s cut ∣ ≤ C B 1 { 1/8 ≤ R ≤ 3/16 } .
Proof. Let 1 8 ≤ R ≤ 3 16 \frac{1}{8}\leq R\leq\frac{3}{16} 8 1 ≤ R ≤ 16 3 . Since J ≥ J 0 J\geq J_0 J ≥ J 0 , ℓ J ≤ 1 32 \ell_J\leq\frac{1}{32} ℓ J ≤ 32 1 , so S = R S=R S = R , m ♯ = R m_{\sharp}=R m ♯ = R (as r ♯ ≤ ℓ J / 16 r_{\sharp}\leq\ell_J/16 r ♯ ≤ ℓ J /16 ) and R ≥ 2 ε R\geq2\varepsilon R ≥ 2 ε (as 2 ε < ℓ J / 4 2\varepsilon<\ell_J/4 2 ε < ℓ J /4 ); thus S S S and m ♯ m_{\sharp} m ♯ are comparable to 1 1 1 . By Lemma 3.3(a) and (29) , Φ ♯ = Φ = R β \Phi^{\sharp}=\Phi=R^{\beta} Φ ♯ = Φ = R β there, and by (75) ∣ Q ♯ ∣ ≤ C |Q^{\sharp}|\leq C ∣ Q ♯ ∣ ≤ C . Since ∣ v ∣ ≤ C |v|\leq C ∣ v ∣ ≤ C (Lemma 2.3), ∣ T s ( χ ∘ ∘ R ) ∣ ≤ C |\mathcal{T}_s(\chi_{\circ}\circ R)|\leq C ∣ T s ( χ ∘ ∘ R ) ∣ ≤ C . For the gradient, ∣ ∇ y Φ ♯ ∣ ≤ C |\nabla_y\Phi^{\sharp}|\leq C ∣ ∇ y Φ ♯ ∣ ≤ C by (3.14); for b ∈ I = { b : ℓ b ≤ 3 L S } b\in I=\{b:\ell_b\leq3LS\} b ∈ I = { b : ℓ b ≤ 3 L S } (the other terms vanish), by (3.14), ∣ δ u b ∣ ≤ C A b ≤ C λ b |\delta u_b|\leq CA_b\leq C\lambda_b ∣ δ u b ∣ ≤ C A b ≤ C λ b and ∥ ∇ u b ∥ ≤ C λ b \|\nabla u_b\|\leq C\lambda_b ∥∇ u b ∥ ≤ C λ b ,
∣ ∇ y ( h b ∇ Φ b ♯ ( z ) ⋅ δ u b ) ∣ ≤ C h b ( ∣ δ u b ∣ + ∥ ∇ u b ∥ ) ≤ C κ , \left|\nabla_y\left(h_b\nabla\Phi_b^{\sharp}(z)\cdot\delta u_b\right)\right|\leq Ch_b\left(|\delta u_b|+\|\nabla u_b\|\right)\leq C\kappa, ∇ y ( h b ∇ Φ b ♯ ( z ) ⋅ δ u b ) ≤ C h b ( ∣ δ u b ∣ + ∥∇ u b ∥ ) ≤ C κ ,
and, using also (53) and ∥ ∇ F a b ∥ ≤ C λ a λ b \|\nabla F_{ab}\|\leq C\lambda_a\lambda_b ∥∇ F ab ∥ ≤ C λ a λ b (Step 2 of the proof of Lemma 5.4),
∣ ∇ y ( q a b B a B b Φ ♯ ) ∣ ≤ C q a b λ a λ b ≤ C κ 2 λ a max ( λ a , λ b ) . \left|\nabla_y\left(q_{ab}B_aB_b\Phi^{\sharp}\right)\right|\leq Cq_{ab}\lambda_a\lambda_b\leq C\kappa^2\frac{\lambda_a}{\max(\lambda_a,\lambda_b)}. ∇ y ( q ab B a B b Φ ♯ ) ≤ C q ab λ a λ b ≤ C κ 2 max ( λ a , λ b ) λ a .
With N = N L + log ( S / ℓ J ) ≤ N L + J N=N_L+\log(S/\ell_J)\leq N_L+J N = N L + log ( S / ℓ J ) ≤ N L + J (as S < 1 S<1 S < 1 and ℓ J = 2 − J \ell_J=2^{-J} ℓ J = 2 − J , so log ( S / ℓ J ) ≤ J log 2 ≤ J \log(S/\ell_J)\leq J\log2\leq J log ( S / ℓ J ) ≤ J log 2 ≤ J ), # I ≤ C N \#I\leq CN # I ≤ C N and ∑ a λ a / max ( λ a , λ b ) ≤ C N \sum_a\lambda_a/\max(\lambda_a,\lambda_b)\leq CN ∑ a λ a / max ( λ a , λ b ) ≤ C N , so ∣ ∇ y Q ♯ ∣ ≤ C ( 1 + κ N + κ 2 N 2 ) ≤ C ( 1 + J ) 2 |\nabla_yQ^{\sharp}|\leq C(1+\kappa N+\kappa^2N^2)\leq C(1+J)^2 ∣ ∇ y Q ♯ ∣ ≤ C ( 1 + κ N + κ 2 N 2 ) ≤ C ( 1 + J ) 2 , using κ N L ≤ 1 \kappa N_L\leq1 κ N L ≤ 1 . The same holds for ∇ y ϱ ′ \nabla_y\varrho' ∇ y ϱ ′ . Hence μ ∣ ∇ Y Q ♯ ∣ ≤ C ε 0 ℓ J 1 + α ( 1 + J ) 2 ≤ C \mu|\nabla_YQ^{\sharp}|\leq C\varepsilon_0\ell_J^{1+\alpha}(1+J)^2\leq C μ ∣ ∇ Y Q ♯ ∣ ≤ C ε 0 ℓ J 1 + α ( 1 + J ) 2 ≤ C . □ \square □
Exponential moment bounds
We next control the transfer of mass between separation ranges. This is needed because the gain is evaluated at renewal times, whereas the density evolves throughout each window. By Lemma 3.3(f), the rate τ j F j ♯ \tau_j\mathbb{F}_j^{\sharp} τ j F j ♯ is bounded below by a multiple of κ \kappa κ on the appropriate range of separations. To use this bound at later times, we must exclude rapid changes of separation, apart from a small exceptional mass.
The estimates below are parabolic off-diagonal bounds. We test (90) against an exponential whose spatial part varies only across the annulus being crossed. The Hölder bound controls the drift term; a suitable exponential parameter and a linear correction in time control the diffusion terms. This produces a supersolution and bounds the mass that crosses the annulus in either direction. Fix λ ↑ , λ ↓ , 2 , λ ↓ , 3 / 2 ∈ C ∞ ( [ 0 , ∞ ) ) \lambda_{\uparrow},\lambda_{\downarrow,2},\lambda_{\downarrow,3/2}\in C^{\infty}([0,\infty)) λ ↑ , λ ↓ , 2 , λ ↓ , 3/2 ∈ C ∞ ([ 0 , ∞ )) with λ ↑ \lambda_{\uparrow} λ ↑ nondecreasing, λ ↑ = 1 \lambda_{\uparrow}=1 λ ↑ = 1 on [ 0 , 1 ] [0,1] [ 0 , 1 ] , λ ↑ = 2 β \lambda_{\uparrow}=2^{\beta} λ ↑ = 2 β on [ 2 , ∞ ) [2,\infty) [ 2 , ∞ ) , and, for θ ∈ { 3 2 , 2 } \theta\in\left\{\frac{3}{2},2\right\} θ ∈ { 2 3 , 2 } , λ ↓ , θ \lambda_{\downarrow,\theta} λ ↓ , θ nonincreasing, λ ↓ , θ = 1 \lambda_{\downarrow,\theta}=1 λ ↓ , θ = 1 on [ 0 , 1 ] [0,1] [ 0 , 1 ] , λ ↓ , θ = 0 \lambda_{\downarrow,\theta}=0 λ ↓ , θ = 0 on [ θ , ∞ ) [\theta,\infty) [ θ , ∞ ) . Let C λ ≥ 1 C_{\lambda}\ge1 C λ ≥ 1 bound the first two derivatives of these three functions; once the profiles are fixed as functions of α \alpha α (through β = 1 − α \beta=1-\alpha β = 1 − α ), this constant depends only on α \alpha α . For a > 0 a>0 a > 0 put
m a : = min ( 1 , ( a / ℓ J ) β ) , k a : = 2 C ∗ C λ m a . m_a:=\min\left(1,(a/\ell_J)^\beta\right),\qquad k_a:=2C_*C_{\lambda}m_a. m a := min ( 1 , ( a / ℓ J ) β ) , k a := 2 C ∗ C λ m a .
Lemma 9.5 (Exponential moment bounds). There is a constant c e m > 0 c_{\mathrm{em}}>0 c em > 0 , depending only on d d d , α \alpha α , the drift bound C ∗ C_* C ∗ and the fixed cutoff profiles, with the following property. Let 0 ≤ s 0 < s 1 ≤ t 0 0\le s_0<s_1\le t_0 0 ≤ s 0 < s 1 ≤ t 0 , and let ϱ \varrho ϱ be an energy solution of (90) on [ s 0 , s 1 ] [s_0,s_1] [ s 0 , s 1 ] such that ϱ ( s 0 ) \varrho(s_0) ϱ ( s 0 ) is a probability density. The same constant works in both cases ( F ) (F) ( F ) and ( B ) (B) ( B ) , independently of J J J , κ \kappa κ , μ \mu μ , the times and the data.
(a) (Growth.) Let 0 < a ≤ 1 8 0<a\le\frac{1}{8} 0 < a ≤ 8 1 satisfy
m a ( s 1 − s 0 ) ≤ c e m a β , μ a β − 2 ≤ c e m m a . m_a(s_1-s_0)\le c_{\mathrm{em}}a^\beta,\qquad\mu a^{\beta-2}\le c_{\mathrm{em}}m_a. m a ( s 1 − s 0 ) ≤ c em a β , μ a β − 2 ≤ c em m a . If ϱ ( s 0 ) \varrho(s_0) ϱ ( s 0 ) vanishes a.e. on { R > a } \{R>a\} { R > a } , then
ϱ s 1 ( { R ≥ 2 a } ) ≤ exp ( − c e m m a a 2 − β μ ) . \varrho_{s_1}(\{R\ge2a\})\le\exp\left(-c_{\mathrm{em}}\frac{m_a a^{2-\beta}}{\mu}\right). ϱ s 1 ({ R ≥ 2 a }) ≤ exp ( − c em μ m a a 2 − β ) . (b) (Shrinking.) Let θ ∈ { 3 2 , 2 } \theta\in\left\{\frac{3}{2},2\right\} θ ∈ { 2 3 , 2 } and 0 < b ≤ 1 8 0<b\le\frac{1}{8} 0 < b ≤ 8 1 satisfy
m b ( s 1 − s 0 ) ≤ c e m b β , μ b β − 2 ≤ c e m m b . m_b(s_1-s_0)\le c_{\mathrm{em}}b^\beta,\qquad\mu b^{\beta-2}\le c_{\mathrm{em}}m_b. m b ( s 1 − s 0 ) ≤ c em b β , μ b β − 2 ≤ c em m b . If ϱ ( s 0 ) \varrho(s_0) ϱ ( s 0 ) vanishes a.e. on { R < θ b } \{R<\theta b\} { R < θ b } , then
ϱ s 1 ( { R ≤ b } ) ≤ exp ( − c e m m b b 2 − β μ ) . \varrho_{s_1}(\{R\le b\})\le\exp\left(-c_{\mathrm{em}}\frac{m_b b^{2-\beta}}{\mu}\right). ϱ s 1 ({ R ≤ b }) ≤ exp ( − c em μ m b b 2 − β ) .
Proof. We prove (a) by constructing an exponential test function which is larger on the destination region than on the initial support. Choose c e m c_{\mathrm{em}} c em sufficiently small, in terms of the stated parameters, that the hypotheses in (a) imply
6 k a ( s 1 − s 0 ) ≤ ( 2 β − 1 ) a β , 2 d μ C λ a β − 2 ≤ k a . 6k_a(s_1-s_0)\le(2^\beta-1)a^\beta,\qquad2d\mu C_{\lambda}a^{\beta-2}\le k_a. 6 k a ( s 1 − s 0 ) ≤ ( 2 β − 1 ) a β , 2 d μ C λ a β − 2 ≤ k a .
The same choice makes the hypotheses in (b) imply the corresponding inequalities with a a a replaced by b b b and ( 2 β − 1 ) a β (2^\beta-1)a^\beta ( 2 β − 1 ) a β replaced by b β b^\beta b β . Let f ( Y ) = a β λ ↑ ( R ( Y ) / a ) f(Y)=a^\beta\lambda_{\uparrow}(R(Y)/a) f ( Y ) = a β λ ↑ ( R ( Y ) / a ) . Since f f f is constant where R ≤ a R\le a R ≤ a and where R ≥ 2 a R\ge2a R ≥ 2 a , and 2 a ≤ 1 4 2a\le\frac{1}{4} 2 a ≤ 4 1 , f f f is smooth on T 2 d \mathbb{T}^{2d} T 2 d and is a function of z z z where it is not constant. On { a ≤ R ≤ 2 a } \{a\le R\le2a\} { a ≤ R ≤ 2 a } , radial differentiation in z ∈ R d z\in\mathbb{R}^d z ∈ R d gives
∣ ∇ z f ∣ ≤ C λ a β − 1 , ∣ Δ z f ∣ ≤ C λ a β − 2 + ( d − 1 ) C λ a β − 1 / R ≤ d C λ a β − 2 . |\nabla_zf|\le C_{\lambda}a^{\beta-1},
\qquad
|\Delta_zf|\le C_{\lambda}a^{\beta-2}+(d-1)C_{\lambda}a^{\beta-1}/R\le dC_{\lambda}a^{\beta-2}. ∣ ∇ z f ∣ ≤ C λ a β − 1 , ∣ Δ z f ∣ ≤ C λ a β − 2 + ( d − 1 ) C λ a β − 1 / R ≤ d C λ a β − 2 .
The relative drift satisfies (17) , also in case (B) by Lemma 6.4(b). On the same annulus,
∣ δ v ∣ ≤ C ∗ min ( R α , λ J R ) ≤ C ∗ min ( ( 2 a ) α , 2 λ J a ) ≤ 2 C ∗ a α m a , |\delta v|\le C_*\min(R^\alpha,\lambda_JR)\le C_*\min((2a)^\alpha,2\lambda_Ja)\le2C_*a^\alpha m_a, ∣ δ v ∣ ≤ C ∗ min ( R α , λ J R ) ≤ C ∗ min (( 2 a ) α , 2 λ J a ) ≤ 2 C ∗ a α m a ,
where we used λ J a = a α ( a / ℓ J ) β \lambda_J a = a^{\alpha}(a/\ell_J)^{\beta} λ J a = a α ( a / ℓ J ) β . The choice α + β = 1 \alpha+\beta=1 α + β = 1 therefore yields ∣ δ v ⋅ ∇ z f ∣ ≤ k a |\delta v\cdot\nabla_z f|\le k_a ∣ δ v ⋅ ∇ z f ∣ ≤ k a . We compensate for this term and the two diffusion terms by setting
M : = k a a 2 − 2 β 2 C λ 2 μ , K : = 3 k a , U ( s , Y ) : = exp ( M ( f ( Y ) + K ( s 1 − s ) ) ) . M:=\frac{k_a a^{2-2\beta}}{2C_\lambda^2\mu},\qquad K:=3k_a,\qquad U(s,Y):=\exp(M(f(Y)+K(s_1-s))). M := 2 C λ 2 μ k a a 2 − 2 β , K := 3 k a , U ( s , Y ) := exp ( M ( f ( Y ) + K ( s 1 − s ))) .
Since T s \mathcal{T}_s T s acts on functions of z z z as δ v ⋅ ∇ z + 2 μ Δ z \delta v\cdot\nabla_z+2\mu\Delta_z δ v ⋅ ∇ z + 2 μ Δ z ,
( ∂ s + T s ) U = M U [ − K + δ v ⋅ ∇ z f + 2 μ Δ z f + 2 μ M ∣ ∇ z f ∣ 2 ] ≤ M U [ − 3 k a + k a + k a + k a ] = 0 \begin{aligned}
(\partial_s+\mathcal{T}_s)U=MU[-K+\delta v\cdot\nabla_zf+2\mu\Delta_zf+2\mu M|\nabla_zf|^2] \\
&\le MU[-3k_a+k_a+k_a+k_a]=0
\end{aligned} ( ∂ s + T s ) U = M U [ − K + δ v ⋅ ∇ z f + 2 μ Δ z f + 2 μ M ∣ ∇ z f ∣ 2 ] ≤ M U [ − 3 k a + k a + k a + k a ] = 0
for a.e. s s s , using 2 d μ C λ a β − 2 ≤ k a 2d\mu C_\lambda a^{\beta-2}\le k_a 2 d μ C λ a β − 2 ≤ k a and 2 μ M C 2 2 a 2 β − 2 = k a 2\mu M C_2^2 a^{2\beta-2}=k_a 2 μ M C 2 2 a 2 β − 2 = k a on { a ≤ R ≤ 2 a } \{a\le R\le2a\} { a ≤ R ≤ 2 a } ; elsewhere the bracket equals − K -K − K . By Lemma 8.2 (c) and ϱ ≥ 0 \varrho\ge0 ϱ ≥ 0 ,
∫ U ( s 1 ) ϱ s 1 ≤ ∫ U ( s 0 ) ϱ s 0 = e M ( a β + K ( s 1 − s 0 ) ) . \int U(s_1)\varrho_{s_1}\le\int U(s_0)\varrho_{s_0}=e^{M(a^\beta+K(s_1-s_0))}. ∫ U ( s 1 ) ϱ s 1 ≤ ∫ U ( s 0 ) ϱ s 0 = e M ( a β + K ( s 1 − s 0 )) .
because f = a β f=a^\beta f = a β on { R ≤ a } \{R\le a\} { R ≤ a } . On { R ≥ 2 a } \{R\ge2a\} { R ≥ 2 a } , U ( s 1 ) = e M ( 2 a ) β U(s_1)=e^{M(2a)^\beta} U ( s 1 ) = e M ( 2 a ) β , and U > 0 U>0 U > 0 . Therefore
ϱ s 1 ( { R ≥ 2 a } ) ≤ exp ( − M [ ( 2 β − 1 ) a β − 3 k a ( s 1 − s 0 ) ] ) ≤ exp ( − M ( 2 β − 1 ) a β / 2 ) . \begin{aligned}
\varrho_{s_1}(\{R\ge2a\})\le\exp\left(-M\left[(2^\beta-1)a^\beta-3k_a(s_1-s_0)\right]\right) \\
&\le\exp\left(-M(2^\beta-1)a^\beta/2\right).
\end{aligned} ϱ s 1 ({ R ≥ 2 a }) ≤ exp ( − M [ ( 2 β − 1 ) a β − 3 k a ( s 1 − s 0 ) ] ) ≤ exp ( − M ( 2 β − 1 ) a β /2 ) .
Since k a = 2 C ∗ C λ m a k_a=2C_*C_\lambda m_a k a = 2 C ∗ C λ m a , the last exponent is a fixed positive multiple of m a a 2 − β / μ m_a a^{2-\beta}/\mu m a a 2 − β / μ . Decreasing c e m c_{\rm em} c em if necessary gives (a).
For (b), reverse the spatial cutoff while keeping the same time correction. Let g ( Y ) = b β λ ↓ , θ ( R ( Y ) / b ) g(Y)=b^\beta\lambda_{\downarrow,\theta}(R(Y)/b) g ( Y ) = b β λ ↓ , θ ( R ( Y ) / b ) , which is smooth since θ b ≤ 1 4 \theta b\le\frac14 θ b ≤ 4 1 ; ∇ g \nabla g ∇ g is supported in { b ≤ R ≤ θ b } \{b\le R\le\theta b\} { b ≤ R ≤ θ b } , where the same bounds hold with a a a replaced by b b b (as θ ≤ 2 \theta\le2 θ ≤ 2 ). With
M = k b b 2 − 2 β 2 C λ 2 μ , K = 3 k b M=\frac{k_b b^{2-2\beta}}{2C_\lambda^2\mu},\qquad K=3k_b M = 2 C λ 2 μ k b b 2 − 2 β , K = 3 k b
and U = exp ( M ( g + K ( s 1 − s ) ) ) U=\exp(M(g+K(s_1-s))) U = exp ( M ( g + K ( s 1 − s ))) , again ( ∂ s + T s ) U ≤ 0 (\partial_s+\mathcal{T}_s)U\le0 ( ∂ s + T s ) U ≤ 0 . Now g = 0 g=0 g = 0 on { R ≥ θ b } \{R\ge\theta b\} { R ≥ θ b } and g = b β g=b^\beta g = b β on { R ≤ b } \{R\le b\} { R ≤ b } , so
ϱ s 1 ( { R ≤ b } ) ≤ exp ( − M [ b β − 3 k b ( s 1 − s 0 ) ] ) ≤ exp ( − M b β / 2 ) . \varrho_{s_1}(\{R\le b\})\le\exp(-M[b^\beta-3k_b(s_1-s_0)])\le\exp(-Mb^\beta/2). ϱ s 1 ({ R ≤ b }) ≤ exp ( − M [ b β − 3 k b ( s 1 − s 0 )]) ≤ exp ( − M b β /2 ) .
Again the exponent is a fixed positive multiple of m b b 2 − β / μ m_b b^{2-\beta}/\mu m b b 2 − β / μ . Choose c e m c_{\rm em} c em small enough for both parts and fix it for the rest of the paper. □ \square □
Exit and transfer bounds
We use the preceding lemma in two ways: to keep most of the mass near the diagonal, and to compare the mass in a given separation range with its mass at the preceding renewal. For the second application, we partition the separations into target sets and enlarge each of them to a set on which the corresponding gain has a lower bound.
Let C 1 ′ : = max ( 2 , ( 16 ℓ j 0 ) − 1 ) C'_1:=\max(2,(16\ell_{j_0})^{-1}) C 1 ′ := max ( 2 , ( 16 ℓ j 0 ) − 1 ) . For R ∈ [ ℓ J , 1 16 ) R\in[\ell_J,\frac{1}{16}) R ∈ [ ℓ J , 16 1 ) let j ( R ) j(R) j ( R ) be the index with ℓ j ≤ R < 2 ℓ j \ell_j\le R<2\ell_j ℓ j ≤ R < 2 ℓ j if R ≤ ℓ j 0 R\le\ell_{j_0} R ≤ ℓ j 0 , and j ( R ) = j 0 j(R)=j_0 j ( R ) = j 0 otherwise; then 1 ≤ R / ℓ j ( R ) ≤ C 1 ′ 1\le R/\ell_{j(R)}\le C'_1 1 ≤ R / ℓ j ( R ) ≤ C 1 ′ . Define the target sets
Σ c o r e = { R ≤ r ♯ / 2 } , \Sigma_{\mathrm{core}}=\{R\le r_\sharp/2\}, Σ core = { R ≤ r ♯ /2 } ,
Σ b = { r ♯ / 2 < R < ℓ J } , \Sigma_b=\{r_\sharp/2<R<\ell_J\}, Σ b = { r ♯ /2 < R < ℓ J } ,
Σ j = { ℓ J ≤ R < 1 / 16 , j ( R ) = j } , \Sigma_j=\{\ell_J\le R<1/16,\ j(R)=j\}, Σ j = { ℓ J ≤ R < 1/16 , j ( R ) = j } ,
which partition { R < 1 / 16 } \{R<1/16\} { R < 1/16 } (for J ≥ J 0 J\ge J_0 J ≥ J 0 , ℓ J ≤ 1 / 32 \ell_J\le1/32 ℓ J ≤ 1/32 ), and, with b j : = min ( C 1 ′ ℓ j , 1 / 16 ) b_j:=\min(C'_1\ell_j,1/16) b j := min ( C 1 ′ ℓ j , 1/16 ) , the gain sets
Σ j + = { 1 2 ℓ j ≤ R ≤ 2 b j } , Σ b + = { 1 4 r ♯ ≤ R ≤ 3 2 ℓ J } . \Sigma_j^+=\{\tfrac12\ell_j\le R\le2b_j\},\qquad\Sigma_b^+=\{\tfrac14r_\sharp\le R\le\tfrac32\ell_J\}. Σ j + = { 2 1 ℓ j ≤ R ≤ 2 b j } , Σ b + = { 4 1 r ♯ ≤ R ≤ 2 3 ℓ J } .
Then Σ j ⊂ { ℓ j ≤ R ≤ b j } ⊂ Σ j + \Sigma_j \subset\{\ell_j \le R \le b_j\} \subset\Sigma_j^+ Σ j ⊂ { ℓ j ≤ R ≤ b j } ⊂ Σ j + , Σ b ⊂ Σ b + \Sigma_b \subset\Sigma_b^+ Σ b ⊂ Σ b + , Σ b + ⊂ Σ b + \Sigma_b^+ \subset\Sigma_b^+ Σ b + ⊂ Σ b + , and Σ j ≠ ∅ \Sigma_j \ne\varnothing Σ j = ∅ only if ℓ j < 1 16 \ell_j < \frac{1}{16} ℓ j < 16 1 , in which case b j ≥ ℓ j b_j \ge\ell_j b j ≥ ℓ j .
Corollary 9.6 (Exit and transfer). There exist t 1 ∈ ( 0 , 1 ] t_1 \in(0,1] t 1 ∈ ( 0 , 1 ] and positive constants c E , c T , c b , κ e x , ε e x c_E,c_T,c_b,\kappa_{\mathrm{ex}},\varepsilon_{\mathrm{ex}} c E , c T , c b , κ ex , ε ex , depending only on d , α , ψ , R , M , σ d,\alpha,\psi,R,M,\sigma d , α , ψ , R , M , σ , such that the following holds. Let J ≥ J 0 J \ge J_0 J ≥ J 0 , assume (39) , and suppose
0 < κ ≤ κ e x , 0 < ε 0 ≤ ε e x . 0 < \kappa\le\kappa_{\mathrm{ex}}, \qquad0 < \varepsilon_0 \le\varepsilon_{\mathrm{ex}}. 0 < κ ≤ κ ex , 0 < ε 0 ≤ ε ex . These constants are independent of J J J , t 0 t_0 t 0 , the realization and the initial probability density.
(a) (Exit.) If ϱ 0 \varrho_0 ϱ 0 vanishes a.e. on { R > 1 32 } \{R > \frac{1}{32}\} { R > 32 1 } , then
ϱ s ( { R ≥ 1 16 } ) ≤ e − c E / μ , s ∈ ( 0 , min ( t 0 , t 1 ) ] . \varrho_s(\{R \ge\tfrac{1}{16}\}) \le e^{-c_E/\mu}, \qquad s \in(0,\min(t_0,t_1)]. ϱ s ({ R ≥ 16 1 }) ≤ e − c E / μ , s ∈ ( 0 , min ( t 0 , t 1 )] . (b) (Transfer.) Let 0 ≤ s ′ < s ≤ t 0 0 \le s' < s \le t_0 0 ≤ s ′ < s ≤ t 0 . For each j j j with Σ j ≠ ∅ \Sigma_j \ne\varnothing Σ j = ∅ , if s − s ′ ≤ τ j s-s' \le\tau_j s − s ′ ≤ τ j , then
Π s ′ , s ( ( Σ j + ) c ; Σ j ) ≤ ϵ j : = exp ( − c T ( ℓ j / ℓ J ) 1 + α / ε 0 ) . \Pi_{s',s}((\Sigma_j^+)^c;\Sigma_j) \le\epsilon_j := \exp\left(-c_T(\ell_j/\ell_J)^{1+\alpha}/\varepsilon_0\right). Π s ′ , s (( Σ j + ) c ; Σ j ) ≤ ϵ j := exp ( − c T ( ℓ j / ℓ J ) 1 + α / ε 0 ) . If s − s ′ ≤ τ J s-s' \le\tau_J s − s ′ ≤ τ J , then
Π s ′ , s ( ( Σ b + ) c ; Σ b ) ≤ ϵ b : = e − c b / κ + e − c T / ε 0 . \Pi_{s',s}((\Sigma_b^+)^c;\Sigma_b) \le\epsilon_b := e^{-c_b/\kappa}+e^{-c_T/\varepsilon_0}. Π s ′ , s (( Σ b + ) c ; Σ b ) ≤ ϵ b := e − c b / κ + e − c T / ε 0 .
Proof. Write c e m c_{\mathrm{em}} c em for the constant in Lemma 9.5 . All constants below depend only on the parameters in the statement. Choose t 1 ∈ ( 0 , 1 ] t_1 \in(0,1] t 1 ∈ ( 0 , 1 ] so that
t 1 ≤ c e m 32 − β . t_1 \le c_{\mathrm{em}}32^{-\beta}. t 1 ≤ c em 3 2 − β .
For (a), apply Lemma 9.5 (a) with a = 1 / 32 a=1/32 a = 1/32 , s 0 = 0 s_0=0 s 0 = 0 and s 1 = s s_1=s s 1 = s . Here m a = 1 m_a=1 m a = 1 . The time condition follows from s ≤ t 1 s \le t_1 s ≤ t 1 , and the diffusion condition follows from μ ≤ ε 0 \mu\le\varepsilon_0 μ ≤ ε 0 once ε e x \varepsilon_{\mathrm{ex}} ε ex is sufficiently small. The resulting exponent is a fixed positive multiple of 1 / μ 1/\mu 1/ μ , giving the claimed estimate for a suitable c E > 0 c_E>0 c E > 0 .
For (b), split an initial probability density supported in ( Σ j + ) c (\Sigma_j^+)^c ( Σ j + ) c into its parts on { R < ℓ j / 2 } \{R<\ell_j/2\} { R < ℓ j /2 } and { R > 2 b j } \{R>2b_j\} { R > 2 b j } . By linearity and positivity, it suffices to bound the mass reaching Σ j \Sigma_j Σ j from either part after normalization. The two bounds are then weighted by initial masses whose sum is one.
For the first part, use Lemma 9.5 (a) with a = ℓ j / 2 a=\ell_j/2 a = ℓ j /2 ; the target lies in { R ≥ 2 a } \{R\ge2a\} { R ≥ 2 a } . Since a ≥ ℓ J / 2 a\ge\ell_J/2 a ≥ ℓ J /2 , we have
m a ≥ 2 − β , m a ( s − s ′ ) a β ≤ 2 β κ , μ a β − 2 m a ≤ 4 ε 0 m_a \ge2^{-\beta}, \qquad\frac{m_a(s-s')}{a^\beta} \le2^\beta\kappa, \qquad\frac{\mu a^{\beta-2}}{m_a} \le4\varepsilon_0 m a ≥ 2 − β , a β m a ( s − s ′ ) ≤ 2 β κ , m a μ a β − 2 ≤ 4 ε 0
(using s − s ′ ≤ τ j ≤ κ ℓ j β s-s' \le\tau_j \le\kappa\ell_j^\beta s − s ′ ≤ τ j ≤ κ ℓ j β and μ a β − 2 / m a ≤ 4 ε 0 ( ℓ J / ℓ j ) 1 + α \mu a^{\beta-2}/m_a \le4\varepsilon_0(\ell_J/\ell_j)^{1+\alpha} μ a β − 2 / m a ≤ 4 ε 0 ( ℓ J / ℓ j ) 1 + α ). Thus the time and diffusion conditions hold for sufficiently small κ e x \kappa_{\mathrm{ex}} κ ex and ε e x \varepsilon_{\mathrm{ex}} ε ex . Moreover,
m a a 2 − β μ ≥ c ( ℓ j / ℓ J ) 1 + α ε 0 , \frac{m_a a^{2-\beta}}{\mu} \ge c\frac{(\ell_j/\ell_J)^{1+\alpha}}{\varepsilon_0}, μ m a a 2 − β ≥ c ε 0 ( ℓ j / ℓ J ) 1 + α ,
using 2 − β = 1 + α 2-\beta=1+\alpha 2 − β = 1 + α and μ = ε 0 ℓ J 1 + α \mu=\varepsilon_0\ell_J^{1+\alpha} μ = ε 0 ℓ J 1 + α . For the second part, use Lemma 9.5 (b) with θ = 2 \theta=2 θ = 2 and b = b j b=b_j b = b j . Here b ≥ ℓ j ≥ ℓ J b\ge\ell_j\ge\ell_J b ≥ ℓ j ≥ ℓ J , so m b = 1 m_b=1 m b = 1 ,
s − s ′ b β ≤ κ , μ b β − 2 ≤ ε 0 , b 2 − β μ ≥ ( ℓ j / ℓ J ) 1 + α ε 0 . \frac{s-s'}{b^\beta}\le\kappa,\qquad\mu b^{\beta-2}\le\varepsilon_0,\qquad\frac{b^{2-\beta}}{\mu}\ge\frac{(\ell_j/\ell_J)^{1+\alpha}}{\varepsilon_0}. b β s − s ′ ≤ κ , μ b β − 2 ≤ ε 0 , μ b 2 − β ≥ ε 0 ( ℓ j / ℓ J ) 1 + α .
The two applications therefore give the stated ϵ j \epsilon_j ϵ j with one constant c T > 0 c_T > 0 c T > 0 , decreased if necessary.
For Σ ♭ \Sigma_{\flat} Σ ♭ , split instead at r ♯ / 4 r_{\sharp}/4 r ♯ /4 and 3 ℓ J / 2 3\ell_J/2 3 ℓ J /2 . For the lower part apply Lemma 9.5 (a) with a = r ♯ / 4 a = r_{\sharp}/4 a = r ♯ /4 . Now m a = ( a / ℓ J ) β m_a = (a/\ell_J)^\beta m a = ( a / ℓ J ) β , and (38) gives
m a ( s − s ′ ) a β ≤ κ , μ a β − 2 m a = μ ℓ J β a 2 = 4 κ β ρ β , m a a 2 − β μ = β ρ β 4 κ . \frac{m_a(s-s')}{a^\beta}\leq\kappa,\qquad\frac{\mu a^{\beta-2}}{m_a}=\frac{\mu\ell_J^\beta}{a^2}=\frac{4\kappa}{\beta\rho^\beta},\qquad\frac{m_a a^{2-\beta}}{\mu}=\frac{\beta\rho^\beta}{4\kappa}. a β m a ( s − s ′ ) ≤ κ , m a μ a β − 2 = a 2 μ ℓ J β = β ρ β 4 κ , μ m a a 2 − β = 4 κ β ρ β .
Thus sufficiently small κ e x \kappa_{\mathrm{ex}} κ ex gives the hypotheses and a bound e − c ♭ / κ e^{-c_{\flat}/\kappa} e − c ♭ / κ , with c ♭ > 0 c_{\flat}>0 c ♭ > 0 fixed. For the upper part use Lemma 9.5 (b) with θ = 3 / 2 \theta=3/2 θ = 3/2 and b = ℓ J b=\ell_J b = ℓ J . The same estimates as above give e − c T / ϵ 0 e^{-c_T/\epsilon_0} e − c T / ϵ 0 , after decreasing c T c_T c T if necessary. Adding the bounds proves the claim. The conditions used above are all implied by
2 β κ e x ≤ c e m , 4 κ e x β ρ β ≤ c e m , 32 2 − β ϵ e x ≤ c e m . 2^\beta\kappa_{\mathrm{ex}}\leq c_{\mathrm{em}},\qquad\frac{4\kappa_{\mathrm{ex}}}{\beta\rho^\beta}\leq c_{\mathrm{em}},\qquad32^{2-\beta}\epsilon_{\mathrm{ex}}\leq c_{\mathrm{em}}. 2 β κ ex ≤ c em , β ρ β 4 κ ex ≤ c em , 3 2 2 − β ϵ ex ≤ c em .
(the first implies κ e x ≤ c e m \kappa_{\mathrm{ex}}\leq c_{\mathrm{em}} κ ex ≤ c em ; the last is the diffusion condition in (a), where μ a β − 2 ≤ 32 2 − β ϵ 0 \mu a^{\beta-2}\leq32^{2-\beta}\epsilon_0 μ a β − 2 ≤ 3 2 2 − β ϵ 0 for a = 1 / 32 a=1/32 a = 1/32 , and it implies 4 ϵ e x ≤ c e m 4\epsilon_{\mathrm{ex}}\leq c_{\mathrm{em}} 4 ϵ ex ≤ c em ); we fix κ e x \kappa_{\mathrm{ex}} κ ex and ϵ e x \epsilon_{\mathrm{ex}} ϵ ex accordingly. □
The transfer errors are summable uniformly in J J J . Indeed, since ( ℓ j / ℓ J ) 1 + α ≥ 2 J − j ≥ 1 + J − j (\ell_j/\ell_J)^{1+\alpha}\geq2^{J-j}\geq1+J-j ( ℓ j / ℓ J ) 1 + α ≥ 2 J − j ≥ 1 + J − j , if e − c T / ϵ 0 ≤ 1 2 e^{-c_T/\epsilon_0}\leq\frac12 e − c T / ϵ 0 ≤ 2 1 , then
ϵ t o t : = ∑ j ϵ j + ϵ ♭ ≤ ∑ m ≥ 1 e − m c T / ϵ 0 + ϵ ♭ ≤ 3 e − c T / ϵ 0 + e − c ♭ / κ . (95) \epsilon_{\mathrm{tot}}:=\sum_j\epsilon_j+\epsilon_{\flat}\leq\sum_{m\geq1}e^{-mc_T/\epsilon_0}+\epsilon_{\flat}\leq3e^{-c_T/\epsilon_0}+e^{-c_{\flat}/\kappa}. \tag*{(95)} ϵ tot := j ∑ ϵ j + ϵ ♭ ≤ m ≥ 1 ∑ e − m c T / ϵ 0 + ϵ ♭ ≤ 3 e − c T / ϵ 0 + e − c ♭ / κ . (95)
Accumulation of the gains
We now distribute each renewal contribution over the following window. The resulting quantity G ( s ) G(s) G ( s ) can be integrated in time, while the transfer estimates compare it with the density at time s s s . Diffusion supplies the gain on the quadratic core, where no renewal contribution is needed.
Let c 3 : = min ( c 1 2 , c 2 2 , 1 4 c 1 , 2 c 1 ′ ) c_3:=\min\left(\frac{c_1}{2},\frac{c_2}{2},\frac14c_1,2c_1'\right) c 3 := min ( 2 c 1 , 2 c 2 , 4 1 c 1 , 2 c 1 ′ ) , with c 2 c_2 c 2 from Lemma 3.3 (f) and c a , b c_{a,b} c a , b from (35) .
The functions Γ j ♯ ( R ) \Gamma_j^{\sharp}(R) Γ j ♯ ( R ) are defined for R ≤ 1 4 R\leq\frac14 R ≤ 4 1 ; products with χ ∘ ( R ) \chi_\circ(R) χ ∘ ( R ) are extended by 0 0 0 .
Recall that s j ( s ) s_j(s) s j ( s ) is the most recent renewal time of scale j j j at or before s s s , or 0 0 0 if there has been no renewal since time 0 0 0 , with the left-limit convention at t 0 t_0 t 0 . Thus ϱ s j ( s ) \varrho_{s_j(s)} ϱ s j ( s ) is the density at the start of the current window, and the indicator 1 { s j ( s ) > 0 } \mathbf{1}_{\{s_j(s)>0\}} 1 { s j ( s ) > 0 } excludes the initial window. For s ∈ ( 0 , t 0 ] s\in(0,t_0] s ∈ ( 0 , t 0 ] put
G ( s ) : = ∑ j τ j 1 { s j ( s ) > 0 } ∫ { R ≤ 1 / 8 } Γ j ♯ ϱ s j ( s ) . G(s):=\sum_j\tau_j\mathbf{1}_{\{s_j(s)>0\}}\int_{\{R\leq1/8\}}\Gamma_j^{\sharp}\varrho_{s_j(s)}. G ( s ) := j ∑ τ j 1 { s j ( s ) > 0 } ∫ { R ≤ 1/8 } Γ j ♯ ϱ s j ( s ) .
Lemma 9.7 (Gains). Assume (39) , L ≥ 2 L\geq2 L ≥ 2 , J ≥ J 0 J\geq J_0 J ≥ J 0 and the hypotheses of Corollary 9.6 , and let ϱ 0 ∈ D ς \varrho_0\in\mathcal{D}_{\varsigma} ϱ 0 ∈ D ς with ς ≤ r ♯ / 8 \varsigma\leq r_{\sharp}/8 ς ≤ r ♯ /8 . For fixed data:
*(a) for t ∈ ( 0 , t 0 ] t\in(0,t_0] t ∈ ( 0 , t 0 ] ,
∑ ζ ~ ∈ Z , ζ ~ ≤ t ∑ j ∈ J ( ζ ~ ) τ j 2 ∫ χ ∘ Γ j ♯ ϱ ζ ~ ≥ ∫ 0 t G ( s ) d s ; \sum_{\widetilde{\zeta}\in\mathcal{Z},\,\widetilde{\zeta}\leq t}\ \sum_{j\in\mathcal{J}(\widetilde{\zeta})}\tau_j^2\int\chi_\circ\Gamma_j^{\sharp}\varrho_{\widetilde{\zeta}}\geq\int_0^tG(s)\,ds; ζ ∈ Z , ζ ≤ t ∑ j ∈ J ( ζ ) ∑ τ j 2 ∫ χ ∘ Γ j ♯ ϱ ζ ≥ ∫ 0 t G ( s ) d s ; :::`
(b) for s ∈ ( 0 , t 0 ] s \in(0,t_0] s ∈ ( 0 , t 0 ] ,
G ( s ) + ∫ χ ∘ 2 μ Δ Φ ♯ ϱ s ≥ c 3 κ ( 1 − ϱ s ( { R ≥ 1 16 } ) − ϵ t o t ) . G(s)+\int\chi_\circ2\mu\Delta\Phi^\sharp\varrho_s \ge c_3\kappa\left(1-\varrho_s(\{R\ge\tfrac{1}{16}\})-\epsilon_{\mathrm{tot}}\right). G ( s ) + ∫ χ ∘ 2 μ Δ Φ ♯ ϱ s ≥ c 3 κ ( 1 − ϱ s ({ R ≥ 16 1 }) − ϵ tot ) .
Proof. All integrands are nonnegative (Lemma 3.3 (e),(f)), and χ ∘ = 1 \chi_\circ=1 χ ∘ = 1 on { R ≤ 1 8 } \{R\le\tfrac{1}{8}\} { R ≤ 8 1 } . For (a), fix j j j . The set of s ∈ ( 0 , t ] s\in(0,t] s ∈ ( 0 , t ] with s j ( s ) > 0 s_j(s)>0 s j ( s ) > 0 is the disjoint union, over the renewal times ζ ~ ∈ ( 0 , t ] \widetilde{\zeta}\in(0,t] ζ ∈ ( 0 , t ] of scale j j j , of the sets of s ∈ ( 0 , t ] s\in(0,t] s ∈ ( 0 , t ] in the window of scale j j j starting at ζ ~ \widetilde{\zeta} ζ ; each has length at most τ j \tau_j τ j , and s j ( s ) = ζ ~ s_j(s)=\widetilde{\zeta} s j ( s ) = ζ on it. Hence the contribution of j j j to the right side is at most
∑ ζ ~ τ j 2 ∫ { R ≤ 1 / 8 } Γ j ♯ ϱ ζ ~ ≤ ∑ ζ ~ τ j 2 ∫ χ ∘ Γ j ♯ ϱ ζ ~ , \sum_{\widetilde{\zeta}}\tau_j^2\int_{\{R\le1/8\}}\Gamma_j^\sharp\varrho_{\widetilde{\zeta}}\le\sum_{\widetilde{\zeta}}\tau_j^2\int\chi_\circ\Gamma_j^\sharp\varrho_{\widetilde{\zeta}}, ζ ∑ τ j 2 ∫ { R ≤ 1/8 } Γ j ♯ ϱ ζ ≤ ζ ∑ τ j 2 ∫ χ ∘ Γ j ♯ ϱ ζ ,
and a renewal time ζ ~ ≤ t \widetilde{\zeta}\le t ζ ≤ t of scale j j j is a point of Z \mathcal{Z} Z with j ∈ J ( ζ ~ ) j\in J(\widetilde{\zeta}) j ∈ J ( ζ ) .
To prove (b), we first apply the lower bound for each gain on its enlarged set. On Σ j + \Sigma_j^+ Σ j + with j < J j<J j < J , τ j Γ j ♯ ≥ 1 2 c 1 / 2 , 2 C 1 ′ κ \tau_j\Gamma_j^\sharp\ge\tfrac{1}{2}c_{1/2,2C_1'}\kappa τ j Γ j ♯ ≥ 2 1 c 1/2 , 2 C 1 ′ κ by Lemma 3.3 (f) and (35) : there R / ℓ j ∈ [ 1 2 , 2 C 1 ′ ] R/\ell_j\in[\tfrac{1}{2},2C_1'] R / ℓ j ∈ [ 2 1 , 2 C 1 ′ ] , R ≤ 1 8 R\le\tfrac{1}{8} R ≤ 8 1 , R ≥ 1 2 ℓ J > 2 ϵ R\ge\tfrac{1}{2}\ell_J>2\epsilon R ≥ 2 1 ℓ J > 2 ϵ , and w j ( R ) = 1 w_j(R)=1 w j ( R ) = 1 since ℓ j ≤ 2 R ≤ L R \ell_j\le2R\le LR ℓ j ≤ 2 R ≤ L R . On Σ J + ∪ Σ b + \Sigma_J^+\cup\Sigma_b^+ Σ J + ∪ Σ b + , τ J Γ J ♯ ≥ min ( 1 2 c 1 / 2 , 2 C 1 ′ , c 2 ) κ \tau_J\Gamma_J^\sharp\ge\min(\tfrac{1}{2}c_{1/2,2C_1'},c_2)\kappa τ J Γ J ♯ ≥ min ( 2 1 c 1/2 , 2 C 1 ′ , c 2 ) κ , and 1 Σ J + ∪ Σ b + ≥ 1 2 ( 1 Σ J + + 1 Σ b + ) \mathbf{1}_{\Sigma_J^+\cup\Sigma_b^+}\ge\tfrac{1}{2}(\mathbf{1}_{\Sigma_J^+}+\mathbf{1}_{\Sigma_b^+}) 1 Σ J + ∪ Σ b + ≥ 2 1 ( 1 Σ J + + 1 Σ b + ) . On Σ c o r e \Sigma_{\mathrm{core}} Σ core , 2 μ Δ Φ ♯ = d κ 2 2\mu\Delta\Phi^\sharp=\frac{d\kappa}{2} 2 μ Δ Φ ♯ = 2 d κ by Lemma 3.3 (e). Hence the left side is at least
c 3 κ [ ∑ j 1 { s j ( s ) > 0 } ϱ s j ( s ) ( Σ j + ) + 1 { s J ( s ) > 0 } ϱ s J ( s ) ( Σ b + ) + ϱ s ( Σ c o r e ) ] . c_3\kappa\left[\sum_j\mathbf{1}_{\{s_j(s)>0\}}\varrho_{s_j(s)}(\Sigma_j^+)+\mathbf{1}_{\{s_J(s)>0\}}\varrho_{s_J(s)}(\Sigma_b^+)+\varrho_s(\Sigma_{\mathrm{core}})\right]. c 3 κ [ j ∑ 1 { s j ( s ) > 0 } ϱ s j ( s ) ( Σ j + ) + 1 { s J ( s ) > 0 } ϱ s J ( s ) ( Σ b + ) + ϱ s ( Σ core ) ] .
Terms with Σ j = ∅ \Sigma_j=\varnothing Σ j = ∅ are nonnegative and may be discarded. It remains to compare these masses at renewal times with the masses of the target sets at time s s s . More precisely, we claim that
1 { s j ( s ) > 0 } ϱ s j ( s ) ( Σ j + ) ≥ ϱ s ( Σ j ) − ϵ j \mathbf{1}_{\{s_j(s)>0\}}\varrho_{s_j(s)}(\Sigma_j^+)\ge\varrho_s(\Sigma_j)-\epsilon_j 1 { s j ( s ) > 0 } ϱ s j ( s ) ( Σ j + ) ≥ ϱ s ( Σ j ) − ϵ j
for every j j j with Σ j ≠ ∅ \Sigma_j\ne\varnothing Σ j = ∅ , and
1 { s J ( s ) > 0 } ϱ s J ( s ) ( Σ b + ) ≥ ϱ s ( Σ b ) − ϵ b . \mathbf{1}_{\{s_J(s)>0\}}\varrho_{s_J(s)}(\Sigma_b^+)\ge\varrho_s(\Sigma_b)-\epsilon_b. 1 { s J ( s ) > 0 } ϱ s J ( s ) ( Σ b + ) ≥ ϱ s ( Σ b ) − ϵ b .
If s j ( s ) = s s_j(s)=s s j ( s ) = s , this holds because Σ j ⊂ Σ j + \Sigma_j\subset\Sigma_j^+ Σ j ⊂ Σ j + . If s j ( s ) < s s_j(s)<s s j ( s ) < s , then s − s j ( s ) ≤ τ j s-s_j(s)\le\tau_j s − s j ( s ) ≤ τ j , and Lemma 8.3 with Corollary 9.6 (b) gives ϱ s ( Σ j ) ≤ ϱ s j ( s ) ( Σ j + ) + ϵ j \varrho_s(\Sigma_j)\le\varrho_{s_j(s)}(\Sigma_j^+)+\epsilon_j ϱ s ( Σ j ) ≤ ϱ s j ( s ) ( Σ j + ) + ϵ j ; if moreover s j ( s ) = 0 s_j(s)=0 s j ( s ) = 0 , then ϱ 0 ( Σ j + ) = 0 \varrho_0(\Sigma_j^+)=0 ϱ 0 ( Σ j + ) = 0 , because ϱ 0 \varrho_0 ϱ 0 vanishes on { R > r ♯ / 8 } \{R>r_\sharp/8\} { R > r ♯ /8 } and Σ j + ⊂ { R ≥ ℓ J / 2 } \Sigma_j^+\subset\{R\ge\ell_J/2\} Σ j + ⊂ { R ≥ ℓ J /2 } . The same argument applies to Σ b + \Sigma_b^+ Σ b + , since Σ b + ⊂ { R ≥ r ♯ / 4 } \Sigma_b^+\subset\{R\ge r_\sharp/4\} Σ b + ⊂ { R ≥ r ♯ /4 } . Finally, the sets Σ c o r e \Sigma_{\mathrm{core}} Σ core , Σ b \Sigma_b Σ b and Σ j \Sigma_j Σ j partition { R < 1 16 } \{R<\tfrac{1}{16}\} { R < 16 1 } . □
Proof of Theorem 9.2
We choose the parameters so that the errors in the evolution identity are smaller than the accumulated gain. We then compare the corrected observable with R β R^\beta R β and obtain the second moment by Jensen’s inequality.
Choice of parameters. Let C 0 , C 3 C_0,C_3 C 0 , C 3 be the constants in ∣ E 0 ∣ ≤ C 0 L − β |E_0|\le C_0L^{-\beta} ∣ E 0 ∣ ≤ C 0 L − β and ∣ E 3 ∣ ≤ C 3 κ 2 N L 2 |E_3|\le C_3\kappa^2N_L^2 ∣ E 3 ∣ ≤ C 3 κ 2 N L 2 (Proposition 5.2 and Corollary 5.3), C 9 C_9 C 9 that of Lemma 5.4, C B C_B C B that of Lemma 9.4, and C ′ C' C ′ that of (5.15). Fix K ≥ 1 K\ge1 K ≥ 1 (not necessarily the K K K of Subsection 7.1) with C 0 / K ≤ c 3 / 8 C_0/K\le c_3/8 C 0 / K ≤ c 3 /8 , and L = ( K / κ ) 1 / β L=(K/\kappa)^{1/\beta} L = ( K / κ ) 1/ β , so that ∣ E 0 ∣ ≤ c 3 κ / 8 |E_0|\le c_3\kappa/8 ∣ E 0 ∣ ≤ c 3 κ /8 . Since N L ≤ C ( 1 + ∣ log κ ∣ ) N_L \le C(1+|\log\kappa|) N L ≤ C ( 1 + ∣ log κ ∣ ) , we can choose κ 1 ≤ min ( κ ∗ , κ e x ) \kappa_1 \le\min(\kappa_*,\kappa_{\mathrm{ex}}) κ 1 ≤ min ( κ ∗ , κ ex ) , with κ e x \kappa_{\mathrm{ex}} κ ex from Corollary 9.6 , such that, for κ ≤ κ 1 \kappa\le\kappa_1 κ ≤ κ 1 ,
L ≥ 2 , κ N L ≤ 1 , c 3 κ ≤ 1 , L \ge2,\qquad\kappa N_L \le1,\qquad c_3\kappa\le1, L ≥ 2 , κ N L ≤ 1 , c 3 κ ≤ 1 ,
C 3 κ N L 2 ≤ c 3 / 8 , C 9 κ N L ≤ c 3 / 8 , e − c b / κ ≤ 1 16 . C_3\kappa N_L^2 \le c_3/8,\qquad C_9\kappa N_L \le c_3/8,\qquad e^{-c_b/\kappa} \le\frac{1}{16}. C 3 κ N L 2 ≤ c 3 /8 , C 9 κ N L ≤ c 3 /8 , e − c b / κ ≤ 16 1 .
We also require
δ κ : = C δ κ ( 1 + ∣ log κ ∣ ) ≤ 1 4 , C ′ ( κ N L + κ 2 N L 2 ) ≤ δ κ , \delta_\kappa:= C_\delta\kappa(1+|\log\kappa|) \le\frac{1}{4},\qquad C'( \kappa N_L+\kappa^2N_L^2) \le\delta_\kappa, δ κ := C δ κ ( 1 + ∣ log κ ∣ ) ≤ 4 1 , C ′ ( κ N L + κ 2 N L 2 ) ≤ δ κ ,
where C δ C_\delta C δ is chosen to ensure the second inequality. Let ε 1 ( κ ) > 0 \varepsilon_1(\kappa)>0 ε 1 ( κ ) > 0 be such that ε 0 ≤ ε 1 ( κ ) \varepsilon_0\le\varepsilon_1(\kappa) ε 0 ≤ ε 1 ( κ ) implies (39) , ε 0 ≤ κ \varepsilon_0\le\kappa ε 0 ≤ κ , ε 0 ≤ ε e x \varepsilon_0\le\varepsilon_{\mathrm{ex}} ε 0 ≤ ε ex and 3 e − c T / ε 0 ≤ 1 16 3e^{-c_T/\varepsilon_0}\le\frac{1}{16} 3 e − c T / ε 0 ≤ 16 1 . Then ε t o t ≤ 1 8 \varepsilon_{\mathrm{tot}}\le\frac{1}{8} ε tot ≤ 8 1 by (95) .
The evolution estimate. Let ϱ 0 ∈ D ς \varrho_0\in\mathcal{D}_{\varsigma} ϱ 0 ∈ D ς with ς ≤ r ♯ / 8 \varsigma\le r_\sharp/8 ς ≤ r ♯ /8 be independent of the data, and fix t ∈ ( 0 , min ( t 0 , t 1 ) ] t\in(0,\min(t_0,t_1)] t ∈ ( 0 , min ( t 0 , t 1 )] . Lemma 9.1 gives measurability in the data. Together with continuity of ϱ s \varrho_s ϱ s in L 2 L^2 L 2 and piecewise continuity of K s K_s K s and G ( s ) \mathcal{G}(s) G ( s ) , this gives joint measurability in ( s , data ) (s,\text{data}) ( s , data ) and allows us to interchange expectations and time integrals.
First, bound the initial value. On the support of ϱ 0 \varrho_0 ϱ 0 , χ ∘ = 1 \chi_\circ=1 χ ∘ = 1 and Φ ♯ ≥ Φ ♯ ( 0 ) \Phi^\sharp\ge\Phi^\sharp(0) Φ ♯ ≥ Φ ♯ ( 0 ) by Lemma 3.3 (a). The endpoint estimate (75) , extended to the diagonal by continuity, gives
∣ Q ♯ − Φ ♯ ∣ ≤ δ κ ( R β + ℓ J β ) ≤ ℓ J β , ∫ Q 0 ∘ ϱ 0 ≥ Φ ♯ ( 0 ) − ℓ J β . |Q^\sharp-\Phi^\sharp|\le\delta_\kappa(R^\beta+\ell_J^\beta)\le\ell_J^\beta,\qquad\int Q_0^\circ\varrho_0\ge\Phi^\sharp(0)-\ell_J^\beta. ∣ Q ♯ − Φ ♯ ∣ ≤ δ κ ( R β + ℓ J β ) ≤ ℓ J β , ∫ Q 0 ∘ ϱ 0 ≥ Φ ♯ ( 0 ) − ℓ J β .
Next, separate the positive diffusion term from the errors in K s \mathcal{K}_s K s . Corollary 5.3 and Lemma 5.4 give
χ ∘ ∣ E 0 + E 3 + μ Δ Y ( Q ♯ − Φ ♯ ) ∣ ≤ 3 8 c 3 κ , ∣ K s c u t ∣ ≤ C B 1 { R ≥ 1 / 16 } . \chi_\circ|E_0+E_3+\mu\Delta_Y(Q^\sharp-\Phi^\sharp)|\le\frac{3}{8}c_3\kappa,\qquad|\mathcal{K}^{\mathrm{cut}}_s|\le C_B\mathbf{1}_{\{R\ge1/16\}}. χ ∘ ∣ E 0 + E 3 + μ Δ Y ( Q ♯ − Φ ♯ ) ∣ ≤ 8 3 c 3 κ , ∣ K s cut ∣ ≤ C B 1 { R ≥ 1/16 } .
Since ∫ ϱ s = 1 \int\varrho_s=1 ∫ ϱ s = 1 , these cost at most 3 8 c 3 κ t \frac{3}{8}c_3\kappa t 8 3 c 3 κ t and C B ∫ 0 t E ϱ s ( { R ≥ 1 / 16 } ) d s C_B\int_0^t\mathbb{E}_{\varrho_s}(\{R\ge1/16\})\,ds C B ∫ 0 t E ϱ s ({ R ≥ 1/16 }) d s , respectively. Take expectations in (93) , use (94) , and apply Lemma 9.7(a) to replace the sum of renewal gains by ∫ 0 t G ( s ) d s \int_0^t\mathcal{G}(s)\,ds ∫ 0 t G ( s ) d s . We obtain
E ∫ Q t ∘ ϱ t ≥ Φ ♯ ( 0 ) − ℓ J β + ∫ 0 t E [ G ( s ) + ∫ χ ∘ 2 μ Δ Φ ♯ ϱ s ] d s − 3 8 c 3 κ t − C B ∫ 0 t E ϱ s ( { R ≥ 1 16 } ) d s . \mathbb{E}\int Q_t^\circ\varrho_t\ge\Phi^\sharp(0)-\ell_J^\beta+\int_0^t\mathbb{E}\left[\mathcal{G}(s)+\int\chi_\circ\,2\mu\Delta\Phi^\sharp\varrho_s\right]ds-\frac{3}{8}c_3\kappa t-C_B\int_0^t\mathbb{E}_{\varrho_s}(\{R\ge\tfrac{1}{16}\})\,ds. E ∫ Q t ∘ ϱ t ≥ Φ ♯ ( 0 ) − ℓ J β + ∫ 0 t E [ G ( s ) + ∫ χ ∘ 2 μ Δ Φ ♯ ϱ s ] d s − 8 3 c 3 κ t − C B ∫ 0 t E ϱ s ({ R ≥ 16 1 }) d s .
Now bound the combined renewal and diffusion gains. Lemma 9.7(b), ε t o t ≤ 1 / 8 \varepsilon_{\mathrm{tot}}\le1/8 ε tot ≤ 1/8 , and the exit estimate in Corollary 9.6 (a) give, for every s ≤ t s\le t s ≤ t ,
E [ G ( s ) + ∫ χ ∘ 2 μ Δ Φ ♯ ϱ s ] ≥ c 3 κ [ 1 − E ϱ s ( R ≥ 1 16 ) − ε t o t ] ≥ c 3 κ ( 7 8 − e − c E / μ ) . \begin{aligned}
\mathbb{E}\left[\mathcal{G}(s)+\int\chi_\circ\,2\mu\Delta\Phi^\sharp\varrho_s\right]\ge c_3\kappa\left[1-\mathbb{E}_{\varrho_s}\left(R\ge\frac{1}{16}\right)-\varepsilon_{\mathrm{tot}}\right] \\
&\ge c_3\kappa\left(\frac{7}{8}-e^{-c_E/\mu}\right).
\end{aligned} E [ G ( s ) + ∫ χ ∘ 2 μ Δ Φ ♯ ϱ s ] ≥ c 3 κ [ 1 − E ϱ s ( R ≥ 16 1 ) − ε tot ] ≥ c 3 κ ( 8 7 − e − c E / μ ) .
The same exit estimate bounds the cutoff error by C B t e − c E / μ C_Bte^{-c_E/\mu} C B t e − c E / μ . Substituting these two bounds above leaves
E ∫ Q t ∘ ϱ t ≥ Φ ♯ ( 0 ) − ℓ J β + ( 7 8 − 3 8 ) c 3 κ t − ( c 3 κ + C B ) t e − c E / μ ≥ Φ ♯ ( 0 ) − ℓ J β + 1 2 c 3 κ t − ( 1 + C B ) e − c E / μ , \begin{aligned}
\mathbb{E}\int Q_t^\circ\varrho_t\ge\Phi^\sharp(0)-\ell_J^\beta+\left(\frac{7}{8}-\frac{3}{8}\right)c_3\kappa t-(c_3\kappa+C_B)te^{-c_E/\mu} \\
&\ge\Phi^\sharp(0)-\ell_J^\beta+\frac{1}{2}c_3\kappa t-(1+C_B)e^{-c_E/\mu},
\end{aligned} E ∫ Q t ∘ ϱ t ≥ Φ ♯ ( 0 ) − ℓ J β + ( 8 7 − 8 3 ) c 3 κ t − ( c 3 κ + C B ) t e − c E / μ ≥ Φ ♯ ( 0 ) − ℓ J β + 2 1 c 3 κ t − ( 1 + C B ) e − c E / μ ,
where the last inequality uses t ≤ 1 t \le1 t ≤ 1 and c 3 κ ≤ 1 c_3\kappa\le1 c 3 κ ≤ 1 .
Return to the separation moments. We now compare the corrected observable with R β R^\beta R β . Where Q ∘ > 0 Q^\circ> 0 Q ∘ > 0 , we have Q ♯ > 0 Q^\sharp> 0 Q ♯ > 0 and 0 ≤ χ ∘ ≤ 1 0 \le\chi^\circ\le1 0 ≤ χ ∘ ≤ 1 , so Lemma 3.3 (a) and (75) give
Q ∘ ≤ Q ♯ ≤ Φ ♯ + ∣ Q ♯ − Φ ♯ ∣ ≤ R β + ℓ J β + δ κ ( R β + ℓ J β ) ≤ ( 1 + δ κ ) R β + 2 ℓ J β . \begin{aligned}
Q^\circ\le Q^\sharp\le\Phi^\sharp+ \lvert Q^\sharp- \Phi^\sharp\rvert\\
&\le R^\beta+ \ell_J^\beta+ \delta_\kappa(R^\beta+ \ell_J^\beta) \\
&\le(1+\delta_\kappa)R^\beta+ 2\ell_J^\beta.
\end{aligned} Q ∘ ≤ Q ♯ ≤ Φ ♯ + ∣ Q ♯ − Φ ♯ ∣ ≤ R β + ℓ J β + δ κ ( R β + ℓ J β ) ≤ ( 1 + δ κ ) R β + 2 ℓ J β .
Where Q ∘ ≤ 0 Q^\circ\le0 Q ∘ ≤ 0 , the last bound is immediate. Integrating it against the probability density ϱ t \varrho_t ϱ t and using the evolution estimate therefore yields
( 1 + δ κ ) E ∫ R β ϱ t ≥ E ∫ Q ∘ ϱ t − 2 ℓ J β ≥ 1 2 c 3 κ t + Φ ♯ ( 0 ) − 3 ℓ J β − ( 1 + C B ) e − c E / μ . \begin{aligned}
(1+\delta_\kappa)\mathbb{E}\int R^\beta\varrho_t \ge\mathbb{E}\int Q^\circ\varrho_t - 2\ell_J^\beta\\
&\ge\frac{1}{2}c_3\kappa t+\Phi^\sharp(0)-3\ell_J^\beta-(1+C_B)e^{-cE/\mu}.
\end{aligned} ( 1 + δ κ ) E ∫ R β ϱ t ≥ E ∫ Q ∘ ϱ t − 2 ℓ J β ≥ 2 1 c 3 κ t + Φ ♯ ( 0 ) − 3 ℓ J β − ( 1 + C B ) e − c E / μ .
Thus the remaining endpoint error comes from Φ ♯ ( 0 ) \Phi^\sharp(0) Φ ♯ ( 0 ) . Lemma 3.3 (c) and ε ≤ ℓ J \varepsilon\le\ell_J ε ≤ ℓ J give
Φ ♯ ( 0 ) ≥ − β ε β ( 1 + 4 β log ( 2 ε / r ♯ ) ) ≥ − ( β + β 4 β log ( 2 ε / r ♯ ) ) ℓ J β . \Phi^\sharp(0)\ge-\beta\varepsilon^\beta(1+4^\beta\log(2\varepsilon/r_\sharp))\ge-(\beta+\beta4^\beta\log(2\varepsilon/r_\sharp))\ell_J^\beta. Φ ♯ ( 0 ) ≥ − β ε β ( 1 + 4 β log ( 2 ε / r ♯ )) ≥ − ( β + β 4 β log ( 2 ε / r ♯ )) ℓ J β .
Substituting this bound above, we obtain
( 1 + δ κ ) E ∫ R β ϱ t ≥ 1 2 c 3 κ t − ( 1 + C B ) e − c E / μ (1+\delta_\kappa)\mathbb{E}\int R^\beta\varrho_t \ge\frac{1}{2}c_3\kappa t-(1+C_B)e^{-cE/\mu} ( 1 + δ κ ) E ∫ R β ϱ t ≥ 2 1 c 3 κ t − ( 1 + C B ) e − c E / μ
− ( 3 + β + β 4 β log ( 2 ε / r ♯ ) ) ℓ J β . -(3+\beta+\beta4^\beta\log(2\varepsilon/r_\sharp))\ell_J^\beta. − ( 3 + β + β 4 β log ( 2 ε / r ♯ )) ℓ J β .
This is (91) with c 5 = c 3 / 2 c_5=c_3/2 c 5 = c 3 /2 , after choosing the constant in Err J \operatorname{Err}_J Err J large enough. If Err J ≤ c 5 κ t / 2 \operatorname{Err}_J\le c_5\kappa t/2 Err J ≤ c 5 κ t /2 , then δ κ ≤ 1 / 4 \delta_\kappa\le1/4 δ κ ≤ 1/4 gives
E ∫ R β ϱ t ≥ c 5 κ t − Err J 1 + δ κ ≥ 4 5 c 5 κ t 2 = 2 5 c 5 κ t . \mathbb{E}\int R^\beta\varrho_t\ge\frac{c_5\kappa t-\operatorname{Err}_J}{1+\delta_\kappa}\ge\frac{4}{5}\frac{c_5\kappa t}{2}=\frac{2}{5}c_5\kappa t. E ∫ R β ϱ t ≥ 1 + δ κ c 5 κ t − Err J ≥ 5 4 2 c 5 κ t = 5 2 c 5 κ t .
Finally, R 2 = ( R β ) 2 / β R^2=(R^\beta)^{2/\beta} R 2 = ( R β ) 2/ β and 2 / β > 1 2/\beta>1 2/ β > 1 . Jensen’s inequality, first against ϱ t \varrho_t ϱ t for each realization and then against the probability law of the field, gives
E ∫ R 2 ϱ t ≥ E [ ( ∫ R β ϱ t ) 2 / β ] ≥ ( E ∫ R β ϱ t ) 2 / β ≥ ( 2 5 c 5 κ t ) 2 / β . \begin{aligned}
\mathbb{E}\int R^2\varrho_t\ge\mathbb{E}\left[\left(\int R^\beta\varrho_t\right)^{2/\beta}\right] \\
&\ge\left(\mathbb{E}\int R^\beta\varrho_t\right)^{2/\beta}\ge\left(\frac{2}{5}c_5\kappa t\right)^{2/\beta}.
\end{aligned} E ∫ R 2 ϱ t ≥ E [ ( ∫ R β ϱ t ) 2/ β ] ≥ ( E ∫ R β ϱ t ) 2/ β ≥ ( 5 2 c 5 κ t ) 2/ β .
This proves both bounds in (92) . □
Anomalous dissipation
We convert the separation estimate into a lower bound for scalar dissipation. The first step is a duality identity expressing the dissipation through the reversed two-point equation. Translation invariance then diagonalizes the mean dissipation in Fourier space, and the remaining symmetries turn the separation bound into a lower bound for each nonzero mode. A Fourier truncation gives the result for general initial data.
The last subsection passes from the truncated fields to the single limiting field. This uses a deterministic energy comparison and does not require a new dispersion argument.
The dissipation as a two-point quantity
To approximate two particles starting at the same point, we take their initial separation to be less than ς \varsigma ς , while keeping each marginal equal to Lebesgue measure. For 0 < ς ≤ 1 4 0 < \varsigma\le\frac{1}{4} 0 < ς ≤ 4 1 let
ϱ 0 ς ( y , y ′ ) : = ∣ B ς ∣ − 1 1 { dist ( y , y ′ ) < ς } , M ς f ( y ) : = ∫ T d f ( y ′ ) ϱ 0 ς ( y , y ′ ) d y ′ , \varrho_0^\varsigma(y,y') := |B_\varsigma|^{-1}\mathbf{1}_{\{\operatorname{dist}(y,y')<\varsigma\}}, \qquad M_\varsigma f(y) := \int_{\mathbb{T}^d} f(y')\varrho_0^\varsigma(y,y')\,\mathrm{d}y', ϱ 0 ς ( y , y ′ ) := ∣ B ς ∣ − 1 1 { dist ( y , y ′ ) < ς } , M ς f ( y ) := ∫ T d f ( y ′ ) ϱ 0 ς ( y , y ′ ) d y ′ ,
where ∣ B ς ∣ |B_\varsigma| ∣ B ς ∣ is the volume of the Euclidean ball of radius ς \varsigma ς in R d \mathbb{R}^d R d . Since ς ≤ 1 4 < 1 2 \varsigma\le\frac{1}{4}<\frac{1}{2} ς ≤ 4 1 < 2 1 , the set { y ′ : dist ( y , y ′ ) < ς } \{y':\operatorname{dist}(y,y')<\varsigma\} { y ′ : dist ( y , y ′ ) < ς } is isometric to that ball for every y y y , so both marginals of ϱ 0 ς \varrho_0^\varsigma ϱ 0 ς are Lebesgue measure on T d \mathbb{T}^d T d ; thus ϱ 0 ς ∈ D ς \varrho_0^\varsigma\in\mathcal{D}_\varsigma ϱ 0 ς ∈ D ς , and M ς M_\varsigma M ς is the average over balls of radius ς \varsigma ς . As ς ↓ 0 \varsigma\downarrow0 ς ↓ 0 , M ς f → f M_\varsigma f \to f M ς f → f in L 2 ( T d ) L^2(\mathbb{T}^d) L 2 ( T d ) for every f ∈ L 2 ( T d ) f \in L^2(\mathbb{T}^d) f ∈ L 2 ( T d ) .
Lemma 10.1 (Dissipation identity). Fix J J J , t 0 ∈ ( 0 , 1 ] t_0 \in(0,1] t 0 ∈ ( 0 , 1 ] and the data. For 0 < ς ≤ 1 4 0 < \varsigma\le\frac{1}{4} 0 < ς ≤ 4 1 let ϱ ς \varrho^\varsigma ϱ ς be the energy solution of (90) in case (B) (the reversed field with final time t 0 t_0 t 0 ) with ϱ ς ( 0 ) = ϱ 0 ς \varrho^\varsigma(0)=\varrho_0^\varsigma ϱ ς ( 0 ) = ϱ 0 ς . Then for every real θ 0 ∈ L ∞ ( T d ) \theta_0 \in L^\infty(\mathbb{T}^d) θ 0 ∈ L ∞ ( T d ) , with θ = θ ( J ) \theta=\theta^{(J)} θ = θ ( J ) ,
1 2 ∬ ( θ 0 ( y ) − θ 0 ( y ′ ) ) 2 ϱ t 0 ς ( y , y ′ ) d y d y ′ = ∥ θ 0 ∥ 2 2 − ( θ ( t 0 ) , M ς θ ( t 0 ) ) , (96) \frac{1}{2}\iint\bigl(\theta_0(y)-\theta_0(y')\bigr)^2\varrho^\varsigma_{t_0}(y,y')\,\mathrm{d}y\,\mathrm{d}y'=\|\theta_0\|_2^2-\bigl(\theta(t_0),M_\varsigma\theta(t_0)\bigr), \tag*{(96)} 2 1 ∬ ( θ 0 ( y ) − θ 0 ( y ′ ) ) 2 ϱ t 0 ς ( y , y ′ ) d y d y ′ = ∥ θ 0 ∥ 2 2 − ( θ ( t 0 ) , M ς θ ( t 0 ) ) , (96)
the left side lies in [ 0 , 2 ∥ θ 0 ∥ ∞ 2 ] [0,2\|\theta_0\|_\infty^2] [ 0 , 2∥ θ 0 ∥ ∞ 2 ] , and consequently
D J ( θ 0 ) = lim ς ↓ 0 1 2 ∬ ( θ 0 ( y ) − θ 0 ( y ′ ) ) 2 ϱ t 0 ς ( y , y ′ ) d y d y ′ . D_J(\theta_0)=\lim_{\varsigma\downarrow0}\frac{1}{2}\iint\bigl(\theta_0(y)-\theta_0(y')\bigr)^2\varrho^\varsigma_{t_0}(y,y')\,\mathrm{d}y\,\mathrm{d}y'. D J ( θ 0 ) = ς ↓ 0 lim 2 1 ∬ ( θ 0 ( y ) − θ 0 ( y ′ ) ) 2 ϱ t 0 ς ( y , y ′ ) d y d y ′ .
Proof. Duality moves the reversed evolution from the density to the scalar increments. The product formula then reduces the result to the one-point scalar equation. Let
F ( y , y ′ ) : = ( θ 0 ( y ) − θ 0 ( y ′ ) ) 2 = θ 0 2 ⊗ 1 + 1 ⊗ θ 0 2 − 2 θ 0 ⊗ θ 0 ∈ L 2 ( T 2 d ) , F(y,y'):=\bigl(\theta_0(y)-\theta_0(y')\bigr)^2
=\theta_0^2\otimes1+1\otimes\theta_0^2-2\theta_0\otimes\theta_0\in L^2(\mathbb{T}^{2d}), F ( y , y ′ ) := ( θ 0 ( y ) − θ 0 ( y ′ ) ) 2 = θ 0 2 ⊗ 1 + 1 ⊗ θ 0 2 − 2 θ 0 ⊗ θ 0 ∈ L 2 ( T 2 d ) ,
and let w w w be the energy solution of (90) in case (F) on [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] with w ( 0 ) = F w(0)=F w ( 0 ) = F . The drift of case (B) is − u ( t 0 − s ) -\mathbf{u}(t_0-s) − u ( t 0 − s ) for all but finitely many s s s , where u \mathbf{u} u is the drift of case (F) (Definition 6.3 ). By Lemma 8.4 ,
∫ F ϱ t 0 ς = ∫ w ( t 0 ) ϱ 0 ς . \int F\varrho^\varsigma_{t_0}=\int w(t_0)\varrho_0^\varsigma. ∫ F ϱ t 0 ς = ∫ w ( t 0 ) ϱ 0 ς .
For g ∈ L 2 ( T d ) g\in L^2(\mathbb{T}^d) g ∈ L 2 ( T d ) let P 0 , t 0 g P_{0,t_0}g P 0 , t 0 g be the value at time t 0 t_0 t 0 of the energy solution of (12) with initial value g g g at time 0 0 0 . Lemma 8.6 (a) and linearity give
w ( t 0 , y , y ′ ) = P 0 , t 0 ( θ 0 2 ) ( y ) + P 0 , t 0 ( θ 0 2 ) ( y ′ ) − 2 P 0 , t 0 θ 0 ( y ) P 0 , t 0 θ 0 ( y ′ ) . w(t_0,y,y')=P_{0,t_0}(\theta_0^2)(y)+P_{0,t_0}(\theta_0^2)(y')-2P_{0,t_0}\theta_0(y)P_{0,t_0}\theta_0(y'). w ( t 0 , y , y ′ ) = P 0 , t 0 ( θ 0 2 ) ( y ) + P 0 , t 0 ( θ 0 2 ) ( y ′ ) − 2 P 0 , t 0 θ 0 ( y ) P 0 , t 0 θ 0 ( y ′ ) .
We integrate these three terms against ϱ 0 s \varrho_0^s ϱ 0 s separately. For the first term, the marginal identity and conservation of the integral (Lemma 8.2 (b)) give
∬ P 0 , t 0 ( θ 0 2 ) ( y ) ϱ 0 s ( y , y ′ ) d y ′ d y = ∫ P 0 , t 0 ( θ 0 2 ) ( y ) ( ∫ ϱ 0 s ( y , y ′ ) d y ′ ) d y = ∫ P 0 , t 0 ( θ 0 2 ) ( y ) d y = ∫ θ 0 2 ( y ) d y = ∥ θ 0 ∥ 2 2 . \begin{aligned}
\iint\mathcal{P}_{0,t_0}(\theta_0^2)(y)\varrho_0^s(y,y')\,dy'\,dy
&= \int\mathcal{P}_{0,t_0}(\theta_0^2)(y)\left(\int\varrho_0^s(y,y')\,dy'\right)\,dy \\
&= \int\mathcal{P}_{0,t_0}(\theta_0^2)(y)\,dy
= \int\theta_0^2(y)\,dy
= \|\theta_0\|_2^2.
\end{aligned} ∬ P 0 , t 0 ( θ 0 2 ) ( y ) ϱ 0 s ( y , y ′ ) d y ′ d y = ∫ P 0 , t 0 ( θ 0 2 ) ( y ) ( ∫ ϱ 0 s ( y , y ′ ) d y ′ ) d y = ∫ P 0 , t 0 ( θ 0 2 ) ( y ) d y = ∫ θ 0 2 ( y ) d y = ∥ θ 0 ∥ 2 2 .
The second term gives the same quantity, using the other marginal. Here P 0 , t 0 ( θ 0 2 ) \mathcal{P}_{0,t_0}(\theta_0^2) P 0 , t 0 ( θ 0 2 ) denotes the evolution of the initial square, whose integral is conserved. For the cross term, the definition of M ς \mathcal{M}_\varsigma M ς yields
∬ P 0 , t 0 θ 0 ( y ) P 0 , t 0 θ 0 ( y ′ ) ϱ 0 s ( y , y ′ ) d y ′ d y = ∫ P 0 , t 0 θ 0 ( y ) ( ∫ P 0 , t 0 θ 0 ( y ′ ) ϱ 0 s ( y , y ′ ) d y ′ ) d y = ( P 0 , t 0 θ 0 , M ς P 0 , t 0 θ 0 ) . \begin{aligned}
\iint\mathcal{P}_{0,t_0}\theta_0(y)\mathcal{P}_{0,t_0}\theta_0(y')\varrho_0^s(y,y')\,dy'\,dy
&= \int\mathcal{P}_{0,t_0}\theta_0(y)\left(\int\mathcal{P}_{0,t_0}\theta_0(y')\varrho_0^s(y,y')\,dy'\right)\,dy \\
&= \left(\mathcal{P}_{0,t_0}\theta_0,\mathcal{M}_\varsigma\mathcal{P}_{0,t_0}\theta_0\right).
\end{aligned} ∬ P 0 , t 0 θ 0 ( y ) P 0 , t 0 θ 0 ( y ′ ) ϱ 0 s ( y , y ′ ) d y ′ d y = ∫ P 0 , t 0 θ 0 ( y ) ( ∫ P 0 , t 0 θ 0 ( y ′ ) ϱ 0 s ( y , y ′ ) d y ′ ) d y = ( P 0 , t 0 θ 0 , M ς P 0 , t 0 θ 0 ) .
Adding the three contributions, we obtain
∫ w ( t 0 ) ϱ 0 s = 2 ∥ θ 0 ∥ 2 2 − 2 ( P 0 , t 0 θ 0 , M ς P 0 , t 0 θ 0 ) . \int w(t_0)\varrho_0^s
=2\|\theta_0\|_2^2-2\left(\mathcal{P}_{0,t_0}\theta_0,\mathcal{M}_\varsigma\mathcal{P}_{0,t_0}\theta_0\right). ∫ w ( t 0 ) ϱ 0 s = 2∥ θ 0 ∥ 2 2 − 2 ( P 0 , t 0 θ 0 , M ς P 0 , t 0 θ 0 ) .
which is ((96) ), as P 0 , t 0 θ 0 = θ ( t 0 ) \mathcal{P}_{0,t_0}\theta_0=\theta(t_0) P 0 , t 0 θ 0 = θ ( t 0 ) . The bounds follow from 0 ≤ F ≤ 4 ∥ θ 0 ∥ ∞ 2 0\leq F\leq4\|\theta_0\|_\infty^2 0 ≤ F ≤ 4∥ θ 0 ∥ ∞ 2 and the fact that ϱ 0 s \varrho_0^s ϱ 0 s is a probability density. Finally, M ς θ ( t 0 ) → θ ( t 0 ) \mathcal{M}_\varsigma\theta(t_0)\to\theta(t_0) M ς θ ( t 0 ) → θ ( t 0 ) in L 2 L^2 L 2 as ς ↓ 0 \varsigma\downarrow0 ς ↓ 0 , so
lim ς ↓ 0 1 2 ∬ ( θ 0 ( y ) − θ 0 ( y ′ ) ) 2 ϱ t 0 s ( y , y ′ ) d y d y ′ = ∥ θ 0 ∥ 2 2 − ∥ θ ( t 0 ) ∥ 2 2 = 2 μ J ∫ 0 t 0 ∥ ∇ θ ( t ) ∥ 2 d t = D J ( θ 0 ) , \begin{aligned}
\lim_{\varsigma\downarrow0}\frac{1}{2}\iint(\theta_0(y)-\theta_0(y'))^2\varrho_{t_0}^s(y,y')\,dy\,dy'
=\|\theta_0\|_2^2-\|\theta(t_0)\|_2^2 \\
&=2\mu_J\int_0^{t_0}\|\nabla\theta(t)\|^2\,dt=D_J(\theta_0),
\end{aligned} ς ↓ 0 lim 2 1 ∬ ( θ 0 ( y ) − θ 0 ( y ′ ) ) 2 ϱ t 0 s ( y , y ′ ) d y d y ′ = ∥ θ 0 ∥ 2 2 − ∥ θ ( t 0 ) ∥ 2 2 = 2 μ J ∫ 0 t 0 ∥∇ θ ( t ) ∥ 2 d t = D J ( θ 0 ) ,
where the second equality is the energy identity ((13) ).
Symmetries and the mean dissipation of single modes
We distinguish two uses of symmetry. Translations eliminate the interaction between distinct Fourier modes. Signed permutations of the coordinates will subsequently allow us to estimate a mode in terms of the full separation, rather than its projection onto a single direction.
Let G d G_d G d be the group of signed permutation matrices. For g ∈ G d g\in G_d g ∈ G d and a ∈ R d a\in\mathbb{R}^d a ∈ R d let ι ( x ) = g x + a \iota(x)=gx+a ι ( x ) = g x + a , an isometry of T d \mathbb{T}^d T d (since g Z d = Z d g\mathbb{Z}^d=\mathbb{Z}^d g Z d = Z d ), and ι ( y , y ′ ) = ( ι y , ι y ′ ) \iota(y,y')=(\iota y,\iota y') ι ( y , y ′ ) = ( ιy , ι y ′ ) . For a vector field v v v put ( ι ∗ v ) ( x , t ) : = g v ( ι − 1 x , t ) (\iota_*v)(x,t):=g\,v(\iota^{-1}x,t) ( ι ∗ v ) ( x , t ) := g v ( ι − 1 x , t ) . Write e k ( x ) = e 2 π i k ⋅ x e_k(x)=e^{2\pi i k\cdot x} e k ( x ) = e 2 π ik ⋅ x and θ 0 ^ ( k ) = ∫ θ 0 e k ‾ \widehat{\theta_0}(k)=\int\theta_0\overline{e_k} θ 0 ( k ) = ∫ θ 0 e k .
Lemma 10.2 (Symmetries of the law). Assume (S). For every ι \iota ι as above there is a bijection T ι T_\iota T ι of the data space, preserving its law, such that u T ι ω = ι ∗ u ω u_{T_\iota\omega}=\iota_*u_\omega u T ι ω = ι ∗ u ω for every J J J ; then also u ← T ι ω = ι ∗ u ← ω \overleftarrow{u}_{T_\iota\omega}=\iota_*\overleftarrow{u}_\omega u T ι ω = ι ∗ u ω for every final time. Consequently the solutions ϱ s \varrho^s ϱ s of Lemma 10.1, for any final time, satisfy ϱ s [ T ι ω ] = ϱ s [ ω ] ∘ ι − 1 \varrho^s[T_\iota\omega]=\varrho^s[\omega]\circ\iota^{-1} ϱ s [ T ι ω ] = ϱ s [ ω ] ∘ ι − 1 for all s s s .
Proof. We construct the transformation of the random data explicitly, first in one renewal window and then across all scales. Its action on the wavelets determines the necessary relabelling and signs. Since ψ \psi ψ is radial, ∇ ( ψ ∘ g − 1 ) = g ∇ ψ ∘ g − 1 \nabla(\psi\circ g^{-1}) = g\nabla\psi\circ g^{-1} ∇ ( ψ ∘ g − 1 ) = g ∇ ψ ∘ g − 1 , so g W j p q ( g − 1 y ) = ( g J p q g T ) ∇ ψ ( y ) gW^{pq}_j(g^{-1}y) = (gJ^{pq}g^{\mathsf T})\nabla\psi(y) g W j pq ( g − 1 y ) = ( g J pq g T ) ∇ ψ ( y ) . If g e p = s p e π ( p ) ge_p = s_pe_{\pi(p)} g e p = s p e π ( p ) with s p = ± 1 s_p = \pm1 s p = ± 1 , then
g J p q g T = s p s q ( e π ( q ) e π ( p ) T − e π ( p ) e π ( q ) T ) = ϵ ( g , ν ) J π g ν gJ^{pq}g^{\mathsf T} = s_ps_q(e_{\pi(q)}e_{\pi(p)}^{\mathsf T} - e_{\pi(p)}e_{\pi(q)}^{\mathsf T})
= \epsilon(g,\nu)J^{\pi_g\nu} g J pq g T = s p s q ( e π ( q ) e π ( p ) T − e π ( p ) e π ( q ) T ) = ϵ ( g , ν ) J π g ν
with ϵ ( g , ν ) = ± 1 \epsilon(g,\nu) = \pm1 ϵ ( g , ν ) = ± 1 and π g \pi_g π g a permutation of A \mathcal{A} A . Periodization commutes with g g g , so g W j ν ( g − 1 y ) = ϵ ( g , ν ) W j π g ν ( y ) gW_j^\nu(g^{-1}y) = \epsilon(g,\nu)W_j^{\pi_g\nu}(y) g W j ν ( g − 1 y ) = ϵ ( g , ν ) W j π g ν ( y ) .
Consider one window of one scale j j j , with shift ξ \xi ξ and coefficient family η = ( η i ν ) i ∈ I j , ν ∈ A \eta= (\eta_i^\nu)_{i\in I_j,\nu\in\mathcal{A}} η = ( η i ν ) i ∈ I j , ν ∈ A . Define ξ ′ ∈ Q j \xi' \in Q_j ξ ′ ∈ Q j and n ( ξ ) ∈ Z d n(\xi) \in\mathbb{Z}^d n ( ξ ) ∈ Z d by g ξ + a = ξ ′ + ℓ j n ( ξ ) g\xi+a = \xi' + \ell_j n(\xi) g ξ + a = ξ ′ + ℓ j n ( ξ ) . Since g g g preserves the lattice ℓ j Z d \ell_j\mathbb{Z}^d ℓ j Z d , there is a permutation σ ξ \sigma_\xi σ ξ of I j I_j I j with g ( y i j + ξ ) + a = y σ ξ ( i ) j + ξ ′ g(y_i^j+\xi)+a = y_{\sigma_\xi(i)}^j+\xi' g ( y i j + ξ ) + a = y σ ξ ( i ) j + ξ ′ in T d \mathbb{T}^d T d ; it depends on ξ \xi ξ only through n ( ξ ) n(\xi) n ( ξ ) , so it is piecewise constant and Borel in ξ \xi ξ . Set η σ ξ ( i ) , π g ν ′ : = ϵ ( g , ν ) η i ν \eta'_{\sigma_\xi(i),\pi_g\nu} := \epsilon(g,\nu)\eta_i^\nu η σ ξ ( i ) , π g ν ′ := ϵ ( g , ν ) η i ν . The map ξ ↦ ξ ′ \xi\mapsto\xi' ξ ↦ ξ ′ is induced by an affine isometry of R d / ℓ j Z d \mathbb{R}^d/\ell_j\mathbb{Z}^d R d / ℓ j Z d , so it preserves the uniform law on Q j Q_j Q j ; for fixed ξ \xi ξ , the map η ↦ η ′ \eta\mapsto\eta' η ↦ η ′ is a signed permutation of an independent, identically distributed and symmetric family, so it preserves the law of η \eta η . Since ξ \xi ξ and η \eta η are independent, Fubini’s theorem gives, for bounded Borel f f f ,
E f ( ξ ′ , η ′ ) = E ξ E η f ( ξ ′ , η ) = E f ( ξ , η ) . \mathbb{E}f(\xi',\eta')=\mathbb{E}_{\xi}\mathbb{E}_{\eta}f(\xi',\eta)=\mathbb{E}f(\xi,\eta). E f ( ξ ′ , η ′ ) = E ξ E η f ( ξ ′ , η ) = E f ( ξ , η ) .
Doing this independently in every window defines T ι T_\iota T ι , a law-preserving bijection (its inverse is constructed from ι − 1 \iota^{-1} ι − 1 in the same way).
We next verify that this transformation has the asserted effect on the velocity. Proceed by induction on j j j . If ι ∗ u < j = u < j ′ \iota_*u_{<j}=u'_{<j} ι ∗ u < j = u < j ′ (primes refer to T ι ω T_\iota\omega T ι ω ), then ι P \iota P ι P solves d d t ι P = ( ι ∗ u < j ) ( ι P ) \frac{d}{dt}\iota P=(\iota_*u_{<j})(\iota P) d t d ι P = ( ι ∗ u < j ) ( ι P ) and starts at ι ( y i j + ξ ) = y σ ξ ( i ) j + ξ ′ \iota(y_i^j+\xi)=y_{\sigma_\xi(i)}^j+\xi' ι ( y i j + ξ ) = y σ ξ ( i ) j + ξ ′ , so ι P i j = P σ ξ ( i ) ′ j \iota P_i^j=P_{\sigma_\xi(i)}^{\prime j} ι P i j = P σ ξ ( i ) ′ j ; since ι − 1 x − P i j = g − 1 ( x − ι P i j ) \iota^{-1}x-P_i^j=g^{-1}(x-\iota P_i^j) ι − 1 x − P i j = g − 1 ( x − ι P i j ) , the mode identity gives g u j ( ι − 1 x ) = u j ′ ( x ) gu_j(\iota^{-1}x)=u'_j(x) g u j ( ι − 1 x ) = u j ′ ( x ) . The reversed field is obtained from u u u by a fixed operation in time, which commutes with ι ∗ \iota_* ι ∗ .
Finally, the transformation passes to the solutions by uniqueness. The two-point drift of ι ∗ v \iota_*v ι ∗ v is ι ∗ v \iota_*v ι ∗ v , with the orthogonal matrix diag ( g , g ) \operatorname{diag}(g,g) diag ( g , g ) , which preserves Z 2 d \mathbb{Z}^{2d} Z 2 d . By Lemma 8.6(b), ϱ s [ ω ] ∘ ι − 1 \varrho^s[\omega]\circ\iota^{-1} ϱ s [ ω ] ∘ ι − 1 is the energy solution with drift v T ι ω v_{T_\iota\omega} v T ι ω and initial value ϱ 0 s ∘ ι − 1 = ϱ 0 s \varrho_0^s\circ\iota^{-1}=\varrho_0^s ϱ 0 s ∘ ι − 1 = ϱ 0 s ; by uniqueness it equals ϱ s [ T ι ω ] \varrho^s[T_\iota\omega] ϱ s [ T ι ω ] . □
For t ∈ ( 0 , 1 ] t\in(0,1] t ∈ ( 0 , 1 ] let D ( t ) ( θ 0 ) : = 2 μ J ∫ 0 t ∥ ∇ ϑ ( J ) ( s ) ∥ 2 d s D^{(t)}(\theta_0):=2\mu_J\int_0^t\lVert\nabla\vartheta^{(J)}(s)\rVert^2\,\mathrm{d}s D ( t ) ( θ 0 ) := 2 μ J ∫ 0 t ∥ ∇ ϑ ( J ) ( s ) ∥ 2 d s be the dissipation up to time t t t . For 0 < ς ≤ 1 / 4 0<\varsigma\leq1/4 0 < ς ≤ 1/4 , let ϱ ς , t \varrho^{\varsigma,t} ϱ ς , t denote the solution of Lemma 10.1 with final time t t t . More explicitly, ϱ s ς , t ( y , y ′ ) \varrho_s^{\varsigma,t}(y,y') ϱ s ς , t ( y , y ′ ) , 0 ≤ s ≤ t 0\leq s\leq t 0 ≤ s ≤ t , is the unique energy solution of the two-point advection–diffusion equation with the velocity reversed from time t t t :
{ ∂ s ϱ s ς , t − u ( J ) ( y , t − s ) ⋅ ∇ y ϱ s ς , t − u ( J ) ( y ′ , t − s ) ⋅ ∇ y ′ ϱ s ς , t = μ J ( Δ y + Δ y ′ ) ϱ s ς , t , ϱ 0 ς , t ( y , y ′ ) = ϱ 0 ς ( y , y ′ ) = ∣ B ς ∣ − 1 1 { dist ( y , y ′ ) < ς } . \begin{cases}
\partial_s\varrho_s^{\varsigma,t}-u^{(J)}(y,t-s)\cdot\nabla_y\varrho_s^{\varsigma,t}-u^{(J)}(y',t-s)\cdot\nabla_{y'}\varrho_s^{\varsigma,t}=\mu_J(\Delta_y+\Delta_{y'})\varrho_s^{\varsigma,t},\\
\varrho_0^{\varsigma,t}(y,y')=\varrho_0^{\varsigma}(y,y')=|B_\varsigma|^{-1}\mathbf{1}_{\{\operatorname{dist}(y,y')<\varsigma\}}.
\end{cases} { ∂ s ϱ s ς , t − u ( J ) ( y , t − s ) ⋅ ∇ y ϱ s ς , t − u ( J ) ( y ′ , t − s ) ⋅ ∇ y ′ ϱ s ς , t = μ J ( Δ y + Δ y ′ ) ϱ s ς , t , ϱ 0 ς , t ( y , y ′ ) = ϱ 0 ς ( y , y ′ ) = ∣ B ς ∣ − 1 1 { dist ( y , y ′ ) < ς } .
The equation holds on ( 0 , t ) × T 2 d (0,t)\times\mathbb{T}^{2d} ( 0 , t ) × T 2 d ; the reversed drift is understood with the convention of Definition 6.3 at renewal times. Thus the superscript t t t specifies the final time used to reverse the velocity, while the subscript s s s is the elapsed time in this reversed evolution.
The following proposition is a standard consequence of translation invariance. By linearity of the scalar equation, the expected dissipation is a quadratic form in the initial datum. Translation invariance of the velocity law makes this form diagonal in the Fourier basis: each mode contributes its initial energy multiplied by its expected dissipated fraction, with no cross terms between distinct modes. The energy identity places these fractions in [ 0 , 1 ] [0,1] [ 0 , 1 ] , and Lemma 10.1 expresses them through the two-point density. Thus the result uses our construction only through translation invariance of its law; neither coordinate symmetries nor the Richardson estimates are needed here. The field-specific estimates enter afterwards, when we prove positive lower bounds for the dissipated fractions.
Proposition 10.3 (Mean dissipation mode by mode). Assume (S). Fix J J J and t ∈ ( 0 , 1 ] t \in(0,1] t ∈ ( 0 , 1 ] . For every k ∈ Z d ∖ { 0 } k \in\mathbb{Z}^{d} \setminus\{0\} k ∈ Z d ∖ { 0 } the limit
m J ( t ) ( k ) : = lim ς ↓ 0 2 E ∫ sin 2 ( π k ⋅ ( y − y ′ ) ) ϱ t S , t [ ω ] ( y , y ′ ) d y d y ′ m_J^{(t)}(k):=\lim_{\varsigma\downarrow0}2\mathbb{E}\int\sin^2\bigl(\pi k\cdot(y-y')\bigr)\varrho_t^{S,t}[\omega](y,y')\,dy\,dy' m J ( t ) ( k ) := ς ↓ 0 lim 2 E ∫ sin 2 ( π k ⋅ ( y − y ′ ) ) ϱ t S , t [ ω ] ( y , y ′ ) d y d y ′ exists, m J ( t ) ( − k ) = m J ( t ) ( k ) ∈ [ 0 , 1 ] m_J^{(t)}(-k)=m_J^{(t)}(k)\in[0,1] m J ( t ) ( − k ) = m J ( t ) ( k ) ∈ [ 0 , 1 ] , and for every real θ 0 ∈ L 2 ( T d ) \theta_0\in L^2(\mathbb{T}^d) θ 0 ∈ L 2 ( T d ) with ∫ θ 0 = 0 \int\theta_0=0 ∫ θ 0 = 0 ,
E D J ( t ) ( θ 0 ) = ∑ k ≠ 0 ∣ θ 0 ^ ( k ) ∣ 2 m J ( t ) ( k ) . \mathbb{E}\mathbf{D}_J^{(t)}(\theta_0)=\sum_{k\ne0}|\widehat{\theta_0}(k)|^2m_J^{(t)}(k). E D J ( t ) ( θ 0 ) = k = 0 ∑ ∣ θ 0 ( k ) ∣ 2 m J ( t ) ( k ) . Proof. We first diagonalize the dissipation for trigonometric polynomials. The existence of the coefficient for each mode then follows by testing a single cosine, and L 2 L^2 L 2 contraction allows passage to general data.
Let θ 0 = ∑ 0 < ∣ k ∣ ≤ K a k e k \theta_0=\sum_{0<|k|\le K}a_ke_k θ 0 = ∑ 0 < ∣ k ∣ ≤ K a k e k be a real trigonometric polynomial. For fixed data ω \omega ω , put
I ς ( ω ) : = 1 2 ∬ ∣ θ 0 ( y ) − θ 0 ( y ′ ) ∣ 2 ϱ t S , t [ ω ] ( y , y ′ ) d y d y ′ . I_{\varsigma}(\omega):=\frac{1}{2}\iint|\theta_0(y)-\theta_0(y')|^2\varrho_t^{S,t}[\omega](y,y')\,dy\,dy'. I ς ( ω ) := 2 1 ∬ ∣ θ 0 ( y ) − θ 0 ( y ′ ) ∣ 2 ϱ t S , t [ ω ] ( y , y ′ ) d y d y ′ . Lemma 10.1, with final time t t t , gives I ς ( ω ) → D J ( t ) ( θ 0 ) ( ω ) I_{\varsigma}(\omega)\to\mathbf{D}_J^{(t)}(\theta_0)(\omega) I ς ( ω ) → D J ( t ) ( θ 0 ) ( ω ) as ς ↓ 0 \varsigma\downarrow0 ς ↓ 0 . Since ϱ t S , t \varrho_t^{S,t} ϱ t S , t is nonnegative and has integral 1 1 1 ,
0 ≤ I ς ( ω ) ≤ 2 ∥ θ 0 ∥ ∞ 2 for every ς and ω . 0\le I_{\varsigma}(\omega)\le2\|\theta_0\|_{\infty}^{2}\qquad\text{for every }\varsigma\text{ and }\omega. 0 ≤ I ς ( ω ) ≤ 2∥ θ 0 ∥ ∞ 2 for every ς and ω . Since ω ↦ I ς ( ω ) \omega\mapsto I_{\varsigma}(\omega) ω ↦ I ς ( ω ) is Borel (Lemma 9.1 ), dominated convergence with respect to the law of the data therefore allows us to interchange the limit and expectation:
E D J ( t ) ( θ 0 ) = lim ς ↓ 0 E I ς . \mathbb{E}\mathbf{D}_J^{(t)}(\theta_0)=\lim_{\varsigma\downarrow0}\mathbb{E}I_{\varsigma}. E D J ( t ) ( θ 0 ) = ς ↓ 0 lim E I ς . Next expand the finite Fourier sum. Writing
F k k ′ ( y , y ′ ) : = ( e k ( y ) − e k ( y ′ ) ) ( e k ′ ( y ) ‾ − e k ′ ( y ′ ) ‾ ) , F_{kk'}(y,y'):=(e_k(y)-e_k(y'))(\overline{e_{k'}(y)}-\overline{e_{k'}(y')}), F k k ′ ( y , y ′ ) := ( e k ( y ) − e k ( y ′ )) ( e k ′ ( y ) − e k ′ ( y ′ ) ) , we have ∣ F k k ′ ∣ ≤ 4 |F_{kk'}|\le4 ∣ F k k ′ ∣ ≤ 4 and
∣ θ 0 ( y ) − θ 0 ( y ′ ) ∣ 2 = ∣ ∑ k a k ( e k ( y ) − e k ( y ′ ) ) ∣ 2 = ∑ k , k ′ a k a k ′ ‾ F k k ′ ( y , y ′ ) . |\theta_0(y)-\theta_0(y')|^2
=\left|\sum_k a_k(e_k(y)-e_k(y'))\right|^2
=\sum_{k,k'}a_k\overline{a_{k'}}F_{kk'}(y,y'). ∣ θ 0 ( y ) − θ 0 ( y ′ ) ∣ 2 = k ∑ a k ( e k ( y ) − e k ( y ′ )) 2 = k , k ′ ∑ a k a k ′ F k k ′ ( y , y ′ ) .
The sum is finite, so integration and expectation can be taken term by term. Thus, setting G k k ′ ς , t : = E ∫ F k k ′ Q s ς , t d s \mathcal{G}^{\varsigma,t}_{kk'} := \mathbb{E}\int F_{kk'}Q^{\varsigma,t}_{s}\,\mathrm{d}s G k k ′ ς , t := E ∫ F k k ′ Q s ς , t d s , we obtain
E I ς = 1 2 ∑ k , k ′ a k a k ′ ‾ G k k ′ ς , \mathbb{E}I_{\varsigma}=\frac{1}{2}\sum_{k,k'}a_k\overline{a_{k'}}\mathcal{G}^{\varsigma}_{kk'}, E I ς = 2 1 k , k ′ ∑ a k a k ′ G k k ′ ς ,
E D J ( t ) ( θ 0 ) = lim ς ↓ 0 1 2 ∑ k , k ′ a k a k ′ ‾ G k k ′ ς . \mathbb{E}D_J^{(t)}(\theta_0)=\lim_{\varsigma\downarrow0}\frac{1}{2}\sum_{k,k'}a_k\overline{a_{k'}}\mathcal{G}^{\varsigma}_{kk'}. E D J ( t ) ( θ 0 ) = ς ↓ 0 lim 2 1 k , k ′ ∑ a k a k ′ G k k ′ ς .
At this stage we use convergence of the whole sum, not of its individual terms.
By Lemma 10.2 with g = I g=I g = I and ι ( x ) = x + a \iota(x)=x+a ι ( x ) = x + a , and since T ι T_\iota T ι preserves the law,
E ∫ F Q s ς , t d s = E ∫ ( F ∘ ι ) Q s ς , t d s \mathbb{E}\int FQ^{\varsigma,t}_{s}\,\mathrm{d}s=\mathbb{E}\int(F\circ\iota)Q^{\varsigma,t}_{s}\,\mathrm{d}s E ∫ F Q s ς , t d s = E ∫ ( F ∘ ι ) Q s ς , t d s
for bounded Borel F F F . As F k k ′ ∘ ι = e 2 π i ( k − k ′ ) ⋅ a F k k ′ F_{kk'}\circ\iota=e^{2\pi i(k-k')\cdot a}F_{kk'} F k k ′ ∘ ι = e 2 π i ( k − k ′ ) ⋅ a F k k ′ , it follows that G k k ′ ς = 0 \mathcal{G}^{\varsigma}_{kk'}=0 G k k ′ ς = 0 for k ≠ k ′ k\ne k' k = k ′ . Since ∣ e k ( y ) − e k ( y ′ ) ∣ = 2 ∣ sin ( π k ⋅ ( y − y ′ ) ) ∣ |e_k(y)-e_k(y')|=2|\sin(\pi k\cdot(y-y'))| ∣ e k ( y ) − e k ( y ′ ) ∣ = 2∣ sin ( π k ⋅ ( y − y ′ )) ∣ , 1 2 G k k ς = 2 E ∫ sin 2 ( π k ⋅ ( y − y ′ ) ) Q s ς , t d s = : M ς ( k ) \frac{1}{2}\mathcal{G}^{\varsigma}_{kk}=2\mathbb{E}\int\sin^2(\pi k\cdot(y-y'))Q^{\varsigma,t}_{s}\,\mathrm{d}s=:M^{\varsigma}(k) 2 1 G k k ς = 2 E ∫ sin 2 ( π k ⋅ ( y − y ′ )) Q s ς , t d s =: M ς ( k ) , and M ς ( − k ) = M ς ( k ) M^{\varsigma}(-k)=M^{\varsigma}(k) M ς ( − k ) = M ς ( k ) . Hence E D J ( t ) ( θ 0 ) = lim ς ↓ 0 ∑ k ∣ a k ∣ 2 M ς ( k ) \mathbb{E}D_J^{(t)}(\theta_0)=\lim_{\varsigma\downarrow0}\sum_k|a_k|^2M^{\varsigma}(k) E D J ( t ) ( θ 0 ) = lim ς ↓ 0 ∑ k ∣ a k ∣ 2 M ς ( k ) . For θ 0 = cos ( 2 π k ⋅ x ) \theta_0=\cos(2\pi k\cdot x) θ 0 = cos ( 2 π k ⋅ x ) this reads 2 E D J ( t ) ( θ 0 ) = lim ς ↓ 0 M ς ( k ) 2\mathbb{E}D_J^{(t)}(\theta_0)=\lim_{\varsigma\downarrow0}M^{\varsigma}(k) 2 E D J ( t ) ( θ 0 ) = lim ς ↓ 0 M ς ( k ) , so the limit m J ( t ) ( k ) m_J^{(t)}(k) m J ( t ) ( k ) exists and lies in [ 0 , 1 ] [0,1] [ 0 , 1 ] (as D J ( t ) ( θ 0 ) ≤ 1 2 D_J^{(t)}(\theta_0)\le\frac{1}{2} D J ( t ) ( θ 0 ) ≤ 2 1 ), and the formula holds for trigonometric polynomials.
For general θ 0 \theta_0 θ 0 let θ K \theta_K θ K be its Fourier truncations. Since the solution map of (12) is linear and contracts L 2 L^2 L 2 (Lemma 8.2 (a)), (13) gives ∣ D J ( t ) ( θ K ) − D J ( t ) ( θ 0 ) ∣ ≤ 4 ∥ θ K − θ 0 ∥ ∥ θ 0 ∥ |D_J^{(t)}(\theta_K)-D_J^{(t)}(\theta_0)|\le4\|\theta_K-\theta_0\|\|\theta_0\| ∣ D J ( t ) ( θ K ) − D J ( t ) ( θ 0 ) ∣ ≤ 4∥ θ K − θ 0 ∥∥ θ 0 ∥ for every realization; let K → ∞ K\to\infty K → ∞ and use monotone convergence on the right side. □
The preceding formula reduces the problem to a lower bound for m J ( t ) ( k ) m_J^{(t)}(k) m J ( t ) ( k ) . For modes up to frequency R \mathcal{R} R , we work at a time when the separation is still at most a small multiple of R − 1 \mathcal{R}^{-1} R − 1 , except for an exponentially small tail. In this region the sine in the dissipation formula can be bounded below by its argument.
Proposition 10.4 (The modes up to a given frequency). Let d ≥ 2 d\ge2 d ≥ 2 and assume (ND), (4) and (S). Let κ ∈ ( 0 , κ 1 ] \kappa\in(0,\kappa_1] κ ∈ ( 0 , κ 1 ] and ε 0 ∈ ( 0 , ε 1 ( κ ) ] \varepsilon_0\in(0,\varepsilon_1(\kappa)] ε 0 ∈ ( 0 , ε 1 ( κ )] , with κ 1 , ε 1 , c 5 \kappa_1,\varepsilon_1,c_5 κ 1 , ε 1 , c 5 from Theorem 9.2 and t 1 t_1 t 1 from Corollary 9.6 . Retain c e m c_{\mathrm{em}} c em from Lemma 9.5 , and let t 0 ∈ ( 0 , 1 ] t_0\in(0,1] t 0 ∈ ( 0 , 1 ] .
For R ≥ 1 \mathcal{R}\ge1 R ≥ 1 put r R : = 1 / ( 8 R ) r_{\mathcal{R}}:=1/(8\mathcal{R}) r R := 1/ ( 8 R ) and
t R : = min ( t 0 , t 1 , c e m ( r R / 2 ) β ) . t_{\mathcal{R}}:=\min\left(t_0,t_1,c_{\mathrm{em}}(r_{\mathcal{R}}/2)^\beta\right). t R := min ( t 0 , t 1 , c em ( r R /2 ) β ) . Let J 1 ( R ) > J 0 : = max ( j 0 , 5 ) J_1(\mathcal{R})>J_0:=\max(j_0,5) J 1 ( R ) > J 0 := max ( j 0 , 5 ) be such that for all J ≥ J 1 ( R ) J\ge J_1(\mathcal{R}) J ≥ J 1 ( R ) : ℓ J ≤ r R / 2 \ell_J\le r_{\mathcal{R}}/2 ℓ J ≤ r R /2 , μ J ≤ c e m ( r R / 2 ) 2 − β \mu_J\le c_{\mathrm{em}}(r_{\mathcal{R}}/2)^{2-\beta} μ J ≤ c em ( r R /2 ) 2 − β , Err J ≤ c 5 κ t R / 2 \operatorname{Err}_J\le c_5\kappa t_{\mathcal{R}}/2 Err J ≤ c 5 κ t R /2 , and
( d 2 ) β exp ( − c e m ( r R / 2 ) 2 − β μ J ) ≤ c 5 κ t R 5 \left(\frac{\sqrt{d}}{2}\right)^\beta\exp\left(-c_{\mathrm{em}}\frac{(r_{\mathcal{R}}/2)^{2-\beta}}{\mu_J}\right)\le\frac{c_5\kappa t_{\mathcal{R}}}{5} ( 2 d ) β exp ( − c em μ J ( r R /2 ) 2 − β ) ≤ 5 c 5 κ t R (all these quantities tend to 0 as J → ∞ J\to\infty J → ∞ ). Then for J ≥ J 1 ( R ) J\ge J_1(\mathcal{R}) J ≥ J 1 ( R ) and 0 < ∣ k ∣ ≤ R 0<|k|\le\mathcal{R} 0 < ∣ k ∣ ≤ R ,
m J ( t R ) ( k ) ≥ 8 ∣ k ∣ 2 d ( c 5 κ t R 5 ) 2 / β . m_J^{(t_{\mathcal{R}})}(k)\ge\frac{8|k|^2}{d}\left(\frac{c_5\kappa t_{\mathcal{R}}}{5}\right)^{2/\beta}. m J ( t R ) ( k ) ≥ d 8∣ k ∣ 2 ( 5 c 5 κ t R ) 2/ β .
Proof. Let 0 < ς ≤ r ♯ / 8 0<\varsigma\le r_{\sharp}/8 0 < ς ≤ r ♯ /8 , let ϱ : = ϱ ς , t R \varrho:=\varrho_{\varsigma,t_{\mathcal{R}}} ϱ := ϱ ς , t R (reversed field with final time t R t_{\mathcal{R}} t R ), and V : = { R ≤ r R } \mathcal{V}:=\{R\le r_{\mathcal{R}}\} V := { R ≤ r R } .
We first localize the separation moment to V \mathcal{V} V . Apply Lemma 9.5 (a) with
a = r R / 2 ≤ 1 16 , s 0 = 0 , s 1 = t R . a=r_{\mathcal{R}}/2\le\frac{1}{16},\qquad s_0=0,\qquad s_1=t_{\mathcal{R}}. a = r R /2 ≤ 16 1 , s 0 = 0 , s 1 = t R .
The initial density is supported in { R ≤ a } \{R \le a\} { R ≤ a } , since ς ≤ ℓ J ≤ a \varsigma\le\ell_J \le a ς ≤ ℓ J ≤ a . Moreover, m a = 1 m_a=1 m a = 1 .
The required time and diffusion conditions are precisely
t K ≤ c e m ( r K / 2 ) β , μ J ≤ c e m a 2 − β . t_{\mathcal K} \le c_{\mathrm{em}}(r_{\mathcal K}/2)^\beta,\qquad\mu_J \le c_{\mathrm{em}}a^{2-\beta}. t K ≤ c em ( r K /2 ) β , μ J ≤ c em a 2 − β .
which hold by our choices of t K t_{\mathcal K} t K and J 1 ( K ) J_1(\mathcal K) J 1 ( K ) . We obtain
ϱ ( V c ) ≤ exp ( − c e m a 2 − β / μ J ) . \varrho(\mathcal V^c) \le\exp\left(-c_{\mathrm{em}}a^{2-\beta}/\mu_J\right). ϱ ( V c ) ≤ exp ( − c em a 2 − β / μ J ) .
On the other hand, (92) in case (B), with final time t K ≤ t 1 t_{\mathcal K} \le t_1 t K ≤ t 1 , gives
E ∫ R β ϱ ≥ 2 5 c 5 κ t K . \mathbb{E}\int R^\beta\varrho\ge\frac{2}{5}c_5\kappa t_{\mathcal K}. E ∫ R β ϱ ≥ 5 2 c 5 κ t K .
Since R ≤ d / 2 R \le\sqrt{d}/2 R ≤ d /2 , the tail condition in the proposition implies
E ∫ V R β ϱ ≥ 1 5 c 5 κ t K . \mathbb{E}\int_{\mathcal V}R^\beta\varrho\ge\frac{1}{5}c_5\kappa t_{\mathcal K}. E ∫ V R β ϱ ≥ 5 1 c 5 κ t K .
Apply Jensen’s inequality to R β 1 V R^\beta1_{\mathcal V} R β 1 V , first with respect to the probability density ϱ \varrho ϱ and then with respect to E \mathbb{E} E . It follows that
E ∫ V R 2 ϱ ≥ ( E ∫ V R β ϱ ) 2 / β . \mathbb{E}\int_{\mathcal V}R^2\varrho\ge\left(\mathbb{E}\int_{\mathcal V}R^\beta\varrho\right)^{2/\beta}. E ∫ V R 2 ϱ ≥ ( E ∫ V R β ϱ ) 2/ β .
We now use the signed permutation symmetry to recover the full second moment from its projection in the direction of k k k . For Y ∈ V Y\in\mathcal V Y ∈ V write z = z ( Y ) z=z(Y) z = z ( Y ) , so ∣ z ∣ = R ≤ 1 8 |z|=R\le\frac{1}{8} ∣ z ∣ = R ≤ 8 1 , and note z ( ι Y ) = g z ( Y ) z(\iota Y)=gz(Y) z ( ι Y ) = g z ( Y ) when ι ( x ) = g x \iota(x)=gx ι ( x ) = g x . By Lemma 10.2 with ι ( x ) = g x \iota(x)=gx ι ( x ) = g x , since T ι T_\iota T ι preserves the law and V \mathcal V V is invariant under ι \iota ι , for every bounded Borel ϕ \phi ϕ and g ∈ G d g\in G_d g ∈ G d ,
E ∫ V ϕ ( z ) ϱ = E ∫ V ϕ ( g z ) ϱ . \mathbb{E}\int_{\mathcal V}\phi(z)\varrho=\mathbb{E}\int_{\mathcal V}\phi(gz)\varrho. E ∫ V ϕ ( z ) ϱ = E ∫ V ϕ ( g z ) ϱ .
With ⟨ ⋅ ⟩ \langle\cdot\rangle ⟨ ⋅ ⟩ the average over G d G_d G d ,
E ∫ sin 2 ( π k ⋅ ( y − y ′ ) ) ϱ ≥ E ∫ V ⟨ sin 2 ( π ( g T k ) ⋅ z ) ⟩ ϱ . \mathbb{E}\int\sin^2(\pi k\cdot(y-y'))\varrho\ge\mathbb{E}\int_{\mathcal V}\left\langle\sin^2(\pi(g^Tk)\cdot z)\right\rangle\varrho. E ∫ sin 2 ( π k ⋅ ( y − y ′ )) ϱ ≥ E ∫ V ⟨ sin 2 ( π ( g T k ) ⋅ z ) ⟩ ϱ .
On V \mathcal V V , ∣ ( g T k ) ⋅ z ∣ ≤ κ r K = 1 8 |(g^Tk)\cdot z|\le\kappa r_{\mathcal K}=\frac{1}{8} ∣ ( g T k ) ⋅ z ∣ ≤ κ r K = 8 1 , and sin 2 ( π s ) ≥ 4 s 2 \sin^2(\pi s)\ge4s^2 sin 2 ( π s ) ≥ 4 s 2 for ∣ s ∣ ≤ 1 2 |s|\le\frac{1}{2} ∣ s ∣ ≤ 2 1 , so the average is at least
4 z T ⟨ g T k k T g ⟩ z = 4 ∣ k ∣ 2 ∣ z ∣ 2 / d . 4z^T\left\langle g^Tkk^Tg\right\rangle z=4|k|^2|z|^2/d. 4 z T ⟨ g T k k T g ⟩ z = 4∣ k ∣ 2 ∣ z ∣ 2 / d .
Indeed, sign changes remove the off-diagonal entries of the averaged matrix, and permutations equalize its diagonal entries. Its trace is ∣ k ∣ 2 |k|^2 ∣ k ∣ 2 , so the matrix is ∣ k ∣ 2 I / d |k|^2I/d ∣ k ∣ 2 I / d . Consequently,
2 E ∫ sin 2 ( π k ⋅ ( y − y ′ ) ) ϱ ≥ 8 ∣ k ∣ 2 d E ∫ V R 2 ϱ . 2\mathbb{E}\int\sin^2(\pi k\cdot(y-y'))\varrho\ge\frac{8|k|^2}{d}\mathbb{E}\int_{\mathcal V}R^2\varrho. 2 E ∫ sin 2 ( π k ⋅ ( y − y ′ )) ϱ ≥ d 8∣ k ∣ 2 E ∫ V R 2 ϱ .
This holds for every ς ≤ r ♯ / 8 \varsigma\le r_\sharp/8 ς ≤ r ♯ /8 . Letting ς ↓ 0 \varsigma\downarrow0 ς ↓ 0 and using the localized second moment proves the proposition. □
Corollary 10.5 (Every mode at the final time). Under the hypotheses of Proposition 10.4, for every k ∈ Z d ∖ { 0 } k\in\mathbb{Z}^d\setminus\{0\} k ∈ Z d ∖ { 0 } and every J ≥ J 1 ( ∣ k ∣ ) J\ge J_1(|k|) J ≥ J 1 ( ∣ k ∣ ) ,
m J ( t 0 ) ( k ) ≥ c ( κ min ( t 0 , t 1 ) ) 2 / β , c = 8 d ( c 5 5 ) 2 / β . m_J^{(t_0)}(k)\ge c\bigl(\kappa\min(t_0,t_1)\bigr)^{2/\beta},\qquad c=\frac{8}{d}\left(\frac{c_5}{5}\right)^{2/\beta}. m J ( t 0 ) ( k ) ≥ c ( κ min ( t 0 , t 1 ) ) 2/ β , c = d 8 ( 5 c 5 ) 2/ β .
Proof. Dissipation is nondecreasing in time, so it suffices to use the estimate at t ∣ k ∣ t_{|k|} t ∣ k ∣ . Indeed, the solution of (12) on [ 0 , t ] [0,t] [ 0 , t ] is the restriction of the solution on [ 0 , t 0 ] [0,t_{0}] [ 0 , t 0 ] , so t ↦ D J ( t ) ( θ 0 ) t \mapsto D_{J}^{(t)}(\theta_{0}) t ↦ D J ( t ) ( θ 0 ) is nondecreasing. By Proposition 10.3 with θ 0 = cos ( 2 π k ⋅ x ) \theta_{0}=\cos(2\pi k\cdot x) θ 0 = cos ( 2 π k ⋅ x ) , m J ( t ) ( k ) = 2 E D J ( t ) ( θ 0 ) m_{J}^{(t)}(k)=2\mathbb{E}D_{J}^{(t)}(\theta_{0}) m J ( t ) ( k ) = 2 E D J ( t ) ( θ 0 ) ; hence m J ( t 0 ) ( k ) ≥ m J ( t ∣ k ∣ ) ( k ) m_{J}^{(t_{0})}(k)\ge m_{J}^{(t_{|k|})}(k) m J ( t 0 ) ( k ) ≥ m J ( t ∣ k ∣ ) ( k ) .
Apply Proposition 10.4 with R = ∣ k ∣ \mathfrak{R}=|k| R = ∣ k ∣ . If t ∣ k ∣ = min ( t 0 , t 1 ) t_{|k|}=\min(t_{0},t_{1}) t ∣ k ∣ = min ( t 0 , t 1 ) , use ∣ k ∣ 2 ≥ 1 |k|^{2}\ge1 ∣ k ∣ 2 ≥ 1 . Otherwise t ∣ k ∣ = c e m ( 16 ∣ k ∣ ) − β t_{|k|}=c_{\mathrm{em}}(16|k|)^{-\beta} t ∣ k ∣ = c em ( 16∣ k ∣ ) − β , so
∣ k ∣ 2 t ∣ k ∣ 2 / β = ( c e m 16 − β ) 2 / β ≥ t 1 2 / β , \frac{|k|^{2}}{t_{|k|}^{2/\beta}}=(c_{\mathrm{em}}16^{-\beta})^{2/\beta}\ge t_{1}^{2/\beta}, t ∣ k ∣ 2/ β ∣ k ∣ 2 = ( c em 1 6 − β ) 2/ β ≥ t 1 2/ β ,
by the choice t 1 ≤ c e m 32 − β t_{1}\le c_{\mathrm{em}}32^{-\beta} t 1 ≤ c em 3 2 − β in the proof of Corollary 9.6 . □
Proof of Theorem 1.4
The lower bound is now uniform over any fixed finite set of nonzero modes. We first pass to arbitrary L 2 L^{2} L 2 data, then obtain the probability statement, and finally make the cutoff uniform under a Sobolev bound.
Let κ 1 \kappa_{1} κ 1 , ε 1 ( κ ) \varepsilon_{1}(\kappa) ε 1 ( κ ) and c 5 c_{5} c 5 be as in Theorem 9.2 , t 1 t_{1} t 1 as in Corollary 9.6 , and c m o d e : = 8 d ( c 5 5 ) 2 / β c_{\mathrm{mode}}:=\frac{8}{d}\left(\frac{c_{5}}{5}\right)^{2/\beta} c mode := d 8 ( 5 c 5 ) 2/ β , the constant in Corollary 10.5. Set c : = c m o d e t 1 2 / β c:=c_{\mathrm{mode}}t_{1}^{2/\beta} c := c mode t 1 2/ β . Since t 0 , t 1 ∈ ( 0 , 1 ] t_{0},t_{1}\in(0,1] t 0 , t 1 ∈ ( 0 , 1 ] , we have min ( t 0 , t 1 ) ≥ t 1 t 0 \min(t_{0},t_{1})\ge t_{1}t_{0} min ( t 0 , t 1 ) ≥ t 1 t 0 , and hence
c m o d e ( κ min ( t 0 , t 1 ) ) 2 / β ≥ c ( κ t 0 ) 2 / β = c A D . c_{\mathrm{mode}}(\kappa\min(t_{0},t_{1}))^{2/\beta}\ge c(\kappa t_{0})^{2/\beta}=c_{\mathrm{AD}}. c mode ( κ min ( t 0 , t 1 ) ) 2/ β ≥ c ( κ t 0 ) 2/ β = c AD .
Thus the auxiliary time t 1 t_{1} t 1 is absorbed into the model-dependent constant c c c . All constants have the dependencies listed in the theorem. Fix κ ∈ ( 0 , κ 1 ] \kappa\in(0,\kappa_{1}] κ ∈ ( 0 , κ 1 ] , ε 0 ∈ ( 0 , ε 1 ( κ ) ] \varepsilon_{0}\in(0,\varepsilon_{1}(\kappa)] ε 0 ∈ ( 0 , ε 1 ( κ )] , t 0 ∈ ( 0 , 1 ] t_{0}\in(0,1] t 0 ∈ ( 0 , 1 ] and a real θ 0 ∈ L 2 ( T d ) \theta_{0}\in L^{2}(\mathbb{T}^{d}) θ 0 ∈ L 2 ( T d ) with ∫ θ 0 = 0 \int\theta_{0}=0 ∫ θ 0 = 0 .
Part (a): general initial data. Let J 2 ( R ) : = max { J 1 ( ∣ k ∣ ) : 0 < ∣ k ∣ ≤ R } J_{2}(\mathfrak{R}):=\max\{J_{1}(|k|):0<|k|\le\mathfrak{R}\} J 2 ( R ) := max { J 1 ( ∣ k ∣ ) : 0 < ∣ k ∣ ≤ R } . By Proposition 10.3 at t = t 0 t=t_{0} t = t 0 , m J ( t 0 ) ≥ 0 m_{J}^{(t_{0})}\ge0 m J ( t 0 ) ≥ 0 and Corollary 10.5, for every R ≥ 1 \mathfrak{R}\ge1 R ≥ 1 and J ≥ J 2 ( R ) J\ge J_{2}(\mathfrak{R}) J ≥ J 2 ( R ) ,
E D J ( θ 0 ) ≥ c A D ∑ 0 < ∣ k ∣ ≤ R ∣ θ 0 ^ ( k ) ∣ 2 . \mathbb{E}D_{J}(\theta_{0})\ge c_{\mathrm{AD}}\sum_{0<|k|\le\mathfrak{R}}|\widehat{\theta_{0}}(k)|^{2}. E D J ( θ 0 ) ≥ c AD 0 < ∣ k ∣ ≤ R ∑ ∣ θ 0 ( k ) ∣ 2 .
Let J → ∞ J\to\infty J → ∞ and then R → ∞ \mathfrak{R}\to\infty R → ∞ ; recall θ 0 ^ ( 0 ) = 0 \widehat{\theta_{0}}(0)=0 θ 0 ( 0 ) = 0 .
Part (b): the probability bound. All u ( J ) u^{(J)} u ( J ) are defined on one probability space (§ 1.2 ), and X J : = D J ( θ 0 ) / ∥ θ 0 ∥ 2 ∈ [ 0 , 1 ] X_{J}:=D_{J}(\theta_{0})/\|\theta_{0}\|^{2}\in[0,1] X J := D J ( θ 0 ) /∥ θ 0 ∥ 2 ∈ [ 0 , 1 ] by (13) . Fatou’s lemma applied to 1 − X J 1-X_{J} 1 − X J gives
E lim sup J X J ≥ lim sup J E X J ≥ c A D , \mathbb{E}\limsup_{J}X_{J}\ge\limsup_{J}\mathbb{E}X_{J}\ge c_{\mathrm{AD}}, E J lim sup X J ≥ J lim sup E X J ≥ c AD ,
where the last inequality is (a). Since lim sup J X J ≤ 1 \limsup_{J}X_{J}\le1 lim sup J X J ≤ 1 , this gives
P ( lim sup J X J > 0 ) ≥ c A D . \mathbb{P}(\limsup_{J}X_{J}>0)\ge c_{\mathrm{AD}}. P ( J lim sup X J > 0 ) ≥ c AD .
Part (c): uniformity under a Sobolev bound. Fix γ > 0 \gamma>0 γ > 0 and A ≥ 1 A\ge1 A ≥ 1 , and put R : = max ( 1 , ( 2 A 2 ) 1 / ( 2 γ ) / ( 2 π ) ) \mathfrak{R}:=\max(1,(2A^{2})^{1/(2\gamma)}/(2\pi)) R := max ( 1 , ( 2 A 2 ) 1/ ( 2 γ ) / ( 2 π )) . For every real mean-zero θ 0 ∈ H γ ( T d ) \theta_{0}\in H^{\gamma}(\mathbb{T}^{d}) θ 0 ∈ H γ ( T d ) with ∥ θ 0 ∥ H γ ≤ A ∥ θ 0 ∥ \|\theta_{0}\|_{H^{\gamma}}\le A\|\theta_{0}\| ∥ θ 0 ∥ H γ ≤ A ∥ θ 0 ∥ ,
∑ ∣ k ∣ > R ∣ θ 0 ^ ( k ) ∣ 2 ≤ ( 2 π R ) − 2 γ ∥ θ 0 ∥ H γ 2 ≤ ( 2 π R ) − 2 γ A 2 ∥ θ 0 ∥ 2 ≤ 1 2 ∥ θ 0 ∥ 2 . \sum_{|k|>\mathfrak{R}}|\widehat{\theta_{0}}(k)|^{2}\le(2\pi\mathfrak{R})^{-2\gamma}\|\theta_{0}\|_{H^{\gamma}}^{2}\le(2\pi\mathfrak{R})^{-2\gamma}A^{2}\|\theta_{0}\|^{2}\le\frac{1}{2}\|\theta_{0}\|^{2}. ∣ k ∣ > R ∑ ∣ θ 0 ( k ) ∣ 2 ≤ ( 2 π R ) − 2 γ ∥ θ 0 ∥ H γ 2 ≤ ( 2 π R ) − 2 γ A 2 ∥ θ 0 ∥ 2 ≤ 2 1 ∥ θ 0 ∥ 2 .
The finite-frequency bound in (a), with J ∗ : = J 2 ( R ) J_{*}:=J_{2}(\mathfrak{R}) J ∗ := J 2 ( R ) , therefore gives the claim, uniformly over all such data. □
Remarks
Remark 10.6 (Fixed cutoff). For each fixed J J J , the field u ( J ) u^{(J)} u ( J ) is spatially Lipschitz, so its scalar dissipation vanishes as μ ↓ 0 \mu\downarrow0 μ ↓ 0 : for θ 0 ∈ H 1 ( T d ) \theta_0 \in H^1(\mathbb{T}^d) θ 0 ∈ H 1 ( T d ) , the standard energy estimate for ∇ θ u \nabla\theta^u ∇ θ u (differentiate the equation, use ∇ ⋅ u ( J ) = 0 \nabla\cdot u^{(J)} = 0 ∇ ⋅ u ( J ) = 0 and (17) , and apply Gronwall’s inequality) gives ∥ ∇ θ u ( t ) ∥ ≤ e c ∗ λ J t ∥ ∇ θ 0 ∥ \lVert\nabla\theta^u(t)\rVert\le e^{c_* \lambda_J t}\lVert\nabla\theta_0\rVert ∥ ∇ θ u ( t )∥ ≤ e c ∗ λ J t ∥ ∇ θ 0 ∥ , hence 2 μ ∫ 0 t 0 ∥ ∇ θ u ∥ 2 d t ≤ 2 μ t 0 e 2 c ∗ λ J t 0 ∥ ∇ θ 0 ∥ 2 → 0 2\mu\int_0^{t_0}\lVert\nabla\theta^u\rVert^2\,dt \le2\mu t_0 e^{2c_*\lambda_Jt_0}\lVert\nabla\theta_0\rVert^2 \to0 2 μ ∫ 0 t 0 ∥ ∇ θ u ∥ 2 d t ≤ 2 μ t 0 e 2 c ∗ λ J t 0 ∥ ∇ θ 0 ∥ 2 → 0 as μ ↓ 0 \mu\downarrow0 μ ↓ 0 ; anomalous dissipation arises here as the velocity cutoff is removed simultaneously with the diffusivity.
Remark 10.7. (1) For α = 1 / 3 \alpha= 1/3 α = 1/3 the constant in Theorem 1.4 is c A D = c κ 3 t 0 3 c_{\mathrm{AD}}=c\kappa^3t_0^3 c AD = c κ 3 t 0 3 , with c c c independent of κ \kappa κ and t 0 t_0 t 0 ; this reflects the Richardson law (11) .
(2) The probability bound in Theorem 1.4(b) is for each fixed datum, not simultaneous over all data; we do not claim that the dissipation event has probability one.
Anomalous dissipation for the fixed limiting field
We now compare the dissipation for u ( ∞ ) u^{(\infty)} u ( ∞ ) with that of a truncation whose matched diffusivity is comparable to μ \mu μ . The velocity-tail estimate of Lemma 2.4 supplies the drift perturbation in the following comparison.
Comparison under a perturbation of drift and diffusivity
Lemma 10.8 (Comparison of dissipations). ** Let u , v u,v u , v be drifts satisfying (D) on [ 0 , t 0 ] × T d [0,t_0]\times\mathbb{T}^d [ 0 , t 0 ] × T d , and suppose u − v = div B u-v=\operatorname{div}B u − v = div B in the sense of distributions on T d \mathbb{T}^d T d for a.e. t t t , where B ∈ L ∞ ( [ 0 , t 0 ] × T d ) B\in L^\infty([0,t_0]\times\mathbb{T}^d) B ∈ L ∞ ([ 0 , t 0 ] × T d ) is skew-symmetric. Write b : = ∥ B ∥ ∞ b:=\lVert B\rVert_\infty b := ∥ B ∥ ∞ . For μ , μ ′ > 0 \mu,\mu'>0 μ , μ ′ > 0 let θ , ϕ \theta,\phi θ , ϕ be the energy solutions of (87) with drifts u , v u,v u , v and diffusivities μ , μ ′ \mu,\mu' μ , μ ′ , respectively, and common initial value θ 0 ∈ L 2 \theta_0\in L^2 θ 0 ∈ L 2 . Set
D u , μ : = 2 μ ∫ 0 t 0 ∥ ∇ θ ∥ 2 d t , D v , μ ′ : = 2 μ ′ ∫ 0 t 0 ∥ ∇ ϕ ∥ 2 d t . D_{u,\mu}:=2\mu\int_0^{t_0}\lVert\nabla\theta\rVert^2\,dt,\qquad D_{v,\mu'}:=2\mu'\int_0^{t_0}\lVert\nabla\phi\rVert^2\,dt. D u , μ := 2 μ ∫ 0 t 0 ∥ ∇ θ ∥ 2 d t , D v , μ ′ := 2 μ ′ ∫ 0 t 0 ∥ ∇ ϕ ∥ 2 d t . Then
D u , μ ≥ μ ′ μ ( μ ′ + ∣ μ − μ ′ ∣ + b ) 2 D v , μ ′ . (97) D_{u,\mu}\geq\frac{\mu'\mu}{(\mu'+\lvert\mu-\mu'\rvert+b)^2}D_{v,\mu'}. \tag*{(97)} D u , μ ≥ ( μ ′ + ∣ μ − μ ′ ∣ + b ) 2 μ ′ μ D v , μ ′ . (97) No smallness of b / μ b/\mu b / μ is required. Exchanging the roles of ( u , μ ) (u,\mu) ( u , μ ) and ( v , μ ′ ) (v,\mu') ( v , μ ′ ) , with B B B replaced by − B -B − B , gives the reverse inequality; in particular, for μ = μ ′ \mu=\mu' μ = μ ′ ,
( 1 + b / μ ) − 2 D v , μ ≤ D u , μ ≤ ( 1 + b / μ ) 2 D v , μ . (98) (1+b/\mu)^{-2}D_{v,\mu}\leq D_{u,\mu}\leq(1+b/\mu)^2D_{v,\mu}. \tag*{(98)} ( 1 + b / μ ) − 2 D v , μ ≤ D u , μ ≤ ( 1 + b / μ ) 2 D v , μ . (98)
Proof. The skew-symmetry of B B B lets us express the drift perturbation as a bilinear form involving only first derivatives. We then estimate the difference of the two scalar solutions in the energy norm.
Step 1: the skew-symmetry identity. For θ , ϕ ∈ H 1 ( T d ) \theta,\phi\in H^1(\mathbb{T}^d) θ , ϕ ∈ H 1 ( T d ) and a.e. t t t ,
( ( div B ) ⋅ ∇ θ , ϕ ) = ( B ∇ θ , ∇ ϕ ) . (99) ((\operatorname{div}B)\cdot\nabla\theta,\phi)=(B\nabla\theta,\nabla\phi). \tag*{(99)} (( div B ) ⋅ ∇ θ , ϕ ) = ( B ∇ θ , ∇ ϕ ) . (99)
Indeed, for θ , ϕ ∈ C ∞ \theta,\phi\in C^\infty θ , ϕ ∈ C ∞ , the definition of the distributional divergence gives
( ( div B ) ⋅ ∇ θ , ϕ ) = − ∑ i , m ( B i m , ∂ m ( ϕ ∂ i θ ) ) = − ∑ i , m ( B i m ∂ i θ , ∂ m ϕ ) − ∑ i , m ( B i m , ϕ ∂ i m θ ) . \begin{aligned}
((\operatorname{div}B)\cdot\nabla\theta,\phi)
=-\sum_{i,m}(B_{im},\partial_m(\phi\,\partial_i\theta)) \\
&=-\sum_{i,m}(B_{im}\partial_i\theta,\partial_m\phi)-\sum_{i,m}(B_{im},\phi\,\partial_{im}\theta).
\end{aligned} (( div B ) ⋅ ∇ θ , ϕ ) = − i , m ∑ ( B im , ∂ m ( ϕ ∂ i θ )) = − i , m ∑ ( B im ∂ i θ , ∂ m ϕ ) − i , m ∑ ( B im , ϕ ∂ im θ ) .
The last sum vanishes, because B B B is skew-symmetric and ∂ i m θ \partial_{im}\theta ∂ im θ is symmetric in ( i , m ) (i,m) ( i , m ) , and the first is − ( B T ∇ θ , ∇ φ ) = ( B ∇ θ , ∇ φ ) -(B^{\mathsf T}\nabla\theta,\nabla\varphi)=(B\nabla\theta,\nabla\varphi) − ( B T ∇ θ , ∇ φ ) = ( B ∇ θ , ∇ φ ) . Both sides of (99) are continuous in ( θ , φ ) ∈ H 1 × H 1 (\theta,\varphi)\in H^1\times H^1 ( θ , φ ) ∈ H 1 × H 1 , because div B = u − v \operatorname{div}B=u-v div B = u − v and B B B are bounded, and C ∞ C^\infty C ∞ is dense in H 1 H^1 H 1 .
Step 2: the equation for the difference. Put w : = ϕ − θ w:=\phi-\theta w := ϕ − θ . Then w ∈ L 2 ( 0 , t 0 ; H 1 ) w\in L^2(0,t_0;H^1) w ∈ L 2 ( 0 , t 0 ; H 1 ) , ∂ t w ∈ L 2 ( 0 , t 0 ; H − 1 ) \partial_t w\in L^2(0,t_0;H^{-1}) ∂ t w ∈ L 2 ( 0 , t 0 ; H − 1 ) and w ( 0 ) = 0 w(0)=0 w ( 0 ) = 0 . Subtracting the identities (88) for ϕ \phi ϕ (drift v v v , diffusivity μ ′ \mu' μ ′ ) and for θ \theta θ (drift u u u , diffusivity μ \mu μ ), and using v ⋅ ∇ ϕ − u ⋅ ∇ θ = v ⋅ ∇ w − ( u − v ) ⋅ ∇ θ v\cdot\nabla\phi-u\cdot\nabla\theta=v\cdot\nabla w-(u-v)\cdot\nabla\theta v ⋅ ∇ ϕ − u ⋅ ∇ θ = v ⋅ ∇ w − ( u − v ) ⋅ ∇ θ , u − v = div B u-v=\operatorname{div}B u − v = div B and (99) , we get, for a.e. t t t and all φ ∈ H 1 \varphi\in H^1 φ ∈ H 1 ,
⟨ ∂ t w , φ ⟩ + ( v ⋅ ∇ w , φ ) + μ ′ ( ∇ w , ∇ φ ) = ( ( ( μ − μ ′ ) I + B ) ∇ θ , ∇ φ ) . (100) \langle\partial_t w,\varphi\rangle+(v\cdot\nabla w,\varphi)+\mu'(\nabla w,\nabla\varphi)=(((\mu-\mu')I+B)\nabla\theta,\nabla\varphi). \tag*{(100)} ⟨ ∂ t w , φ ⟩ + ( v ⋅ ∇ w , φ ) + μ ′ ( ∇ w , ∇ φ ) = ((( μ − μ ′ ) I + B ) ∇ θ , ∇ φ ) . (100)
Step 3: energy estimate. Take φ = w ( t ) \varphi=w(t) φ = w ( t ) in (100) . By (E1) and (86) , and since ∣ B ∇ θ ∣ ≤ b ∣ ∇ θ ∣ |B\nabla\theta|\le b|\nabla\theta| ∣ B ∇ θ ∣ ≤ b ∣∇ θ ∣ pointwise, for a.e. t t t
1 2 d d t ∥ w ∥ 2 + μ ′ ∥ ∇ w ∥ 2 ≤ ( ∣ μ − μ ′ ∣ + b ) ∥ ∇ θ ∥ ∥ ∇ w ∥ . \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\|w\|^2+\mu'\|\nabla w\|^2\le(|\mu-\mu'|+b)\|\nabla\theta\|\|\nabla w\|. 2 1 d t d ∥ w ∥ 2 + μ ′ ∥∇ w ∥ 2 ≤ ( ∣ μ − μ ′ ∣ + b ) ∥∇ θ ∥∥∇ w ∥.
Integrating over [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] , using w ( 0 ) = 0 w(0)=0 w ( 0 ) = 0 and the Cauchy–Schwarz inequality in t t t ,
1 2 ∥ w ( t 0 ) ∥ 2 + μ ′ ∥ ∇ w ∥ L t , x 2 2 ≤ ( ∣ μ − μ ′ ∣ + b ) ∥ ∇ θ ∥ L t , x 2 ∥ ∇ w ∥ L t , x 2 . \frac{1}{2}\|w(t_0)\|^2+\mu'\|\nabla w\|_{L^2_{t,x}}^2\le(|\mu-\mu'|+b)\|\nabla\theta\|_{L^2_{t,x}}\|\nabla w\|_{L^2_{t,x}}. 2 1 ∥ w ( t 0 ) ∥ 2 + μ ′ ∥∇ w ∥ L t , x 2 2 ≤ ( ∣ μ − μ ′ ∣ + b ) ∥∇ θ ∥ L t , x 2 ∥∇ w ∥ L t , x 2 .
so
∥ ∇ w ∥ L t , x 2 ≤ ∣ μ − μ ′ ∣ + b μ ′ ∥ ∇ θ ∥ L t , x 2 . \|\nabla w\|_{L^2_{t,x}}\le\frac{|\mu-\mu'|+b}{\mu'}\|\nabla\theta\|_{L^2_{t,x}}. ∥∇ w ∥ L t , x 2 ≤ μ ′ ∣ μ − μ ′ ∣ + b ∥∇ θ ∥ L t , x 2 .
Since ϕ = θ + w \phi=\theta+w ϕ = θ + w ,
∥ ∇ ϕ ∥ L t , x 2 ≤ ∥ ∇ θ ∥ L t , x 2 + ∥ ∇ w ∥ L t , x 2 ≤ μ ′ + ∣ μ − μ ′ ∣ + b μ ′ ∥ ∇ θ ∥ L t , x 2 . \|\nabla\phi\|_{L^2_{t,x}}\le\|\nabla\theta\|_{L^2_{t,x}}+\|\nabla w\|_{L^2_{t,x}}\le\frac{\mu'+|\mu-\mu'|+b}{\mu'}\|\nabla\theta\|_{L^2_{t,x}}. ∥∇ ϕ ∥ L t , x 2 ≤ ∥∇ θ ∥ L t , x 2 + ∥∇ w ∥ L t , x 2 ≤ μ ′ μ ′ + ∣ μ − μ ′ ∣ + b ∥∇ θ ∥ L t , x 2 .
Squaring and multiplying by 2 μ ′ 2\mu' 2 μ ′ gives D v , μ ′ ≤ ( μ ′ + ∣ μ − μ ′ ∣ + b ) 2 μ ′ μ D u , μ D_{v,\mu'}\le\frac{(\mu'+|\mu-\mu'|+b)^2}{\mu'\mu}D_{u,\mu} D v , μ ′ ≤ μ ′ μ ( μ ′ + ∣ μ − μ ′ ∣ + b ) 2 D u , μ , which is (97) . The hypotheses are symmetric in ( u , μ , B ) (u,\mu,B) ( u , μ , B ) and ( v , μ ′ , − B ) (v,\mu',-B) ( v , μ ′ , − B ) , which gives the last assertion. □
Remark 10.9. Lemma 10.8 is a time-dependent form of the classical bound μ e f f ≤ μ + ∥ B ∥ ∞ 2 / μ \mu_{\mathrm{eff}}\le\mu+\|B\|_\infty^2/\mu μ eff ≤ μ + ∥ B ∥ ∞ 2 / μ for the effective diffusivity of a periodic flow u = div B u=\operatorname{div}B u = div B in terms of its stream matrix B B B , compare [20 ] : a perturbation of the drift by div B \operatorname{div}B div B with ∥ B ∥ ∞ ≲ μ \|B\|_\infty\lesssim\mu ∥ B ∥ ∞ ≲ μ , that is, at bounded Péclet number, changes the dissipation by at most a constant factor, in either direction.
Proof of Theorem 1.5
Proof of Theorem 1.5. We first justify measurability for the limiting field and then apply the comparison lemma at a truncation adapted to μ \mu μ . Put
a ∗ : = ( 2 ( 1 + α ) / 2 + C p o t ε 0 ) − 2 ∈ ( 0 , 1 ) . (101) a_*:=\left(2^{(1+\alpha)/2}+\frac{C_{\mathrm{pot}}}{\varepsilon_0}\right)^{-2}\in(0,1). \tag*{(101)} a ∗ := ( 2 ( 1 + α ) /2 + ε 0 C pot ) − 2 ∈ ( 0 , 1 ) . (101)
Since u ( N ) → u ( ∞ ) u^{(N)}\to u^{(\infty)} u ( N ) → u ( ∞ ) uniformly and each u ( N ) u^{(N)} u ( N ) is bounded, Borel measurable, and divergence-free, the limiting field satisfies (D) on [ 0 , 1 ] × T d [0,1]\times\mathbb{T}^d [ 0 , 1 ] × T d . Hence Section 8 gives existence, uniqueness and the energy identity for (14) , and D ∞ μ ( θ 0 ) = ∥ θ 0 ∥ 2 − ∥ θ μ ( t 0 ) ∥ 2 D^\mu_\infty(\theta_0)=\|\theta_0\|^2-\|\theta^\mu(t_0)\|^2 D ∞ μ ( θ 0 ) = ∥ θ 0 ∥ 2 − ∥ θ μ ( t 0 ) ∥ 2 .
Measurability with respect to the random data defining the velocity. To take the expectations in (a) and (c), we first show that, for each fixed μ > 0 \mu>0 μ > 0 , the dissipation D ∞ μ ( θ 0 ) D^\mu_\infty(\theta_0) D ∞ μ ( θ 0 ) is a measurable function of the random data. For the probability statement in (b), we also show continuity in μ > 0 \mu>0 μ > 0 for each realization; this allows the lim sup \limsup lim sup to be taken over rational diffusivities and ensures that the event in (b) is measurable.
Fix μ > 0 \mu>0 μ > 0 and let θ μ , N \theta^{\mu,N} θ μ , N be the energy solution with drift u ( N ) u^{(N)} u ( N ) and diffusivity μ \mu μ . Subtracting the identities (88) for θ μ , N \theta^{\mu,N} θ μ , N and θ μ \theta^\mu θ μ and using (86) , w : = θ μ , N − θ μ w:=\theta^{\mu,N}-\theta^\mu w := θ μ , N − θ μ satisfies, for a.e. t t t and all φ ∈ H 1 \varphi\in H^1 φ ∈ H 1 ,
⟨ ∂ t w , φ ⟩ + ( u ( N ) ⋅ ∇ w , φ ) + μ ( ∇ w , ∇ φ ) = ( ( u ( ∞ ) − u ( N ) ) ⋅ ∇ θ μ , φ ) = − ( ( u ( ∞ ) − u ( N ) ) θ μ , ∇ φ ) . \langle\partial_t w,\varphi\rangle+(u^{(N)}\cdot\nabla w,\varphi)+\mu(\nabla w,\nabla\varphi)
=((u^{(\infty)}-u^{(N)})\cdot\nabla\theta^\mu,\varphi)
=-((u^{(\infty)}-u^{(N)})\theta^\mu,\nabla\varphi). ⟨ ∂ t w , φ ⟩ + ( u ( N ) ⋅ ∇ w , φ ) + μ ( ∇ w , ∇ φ ) = (( u ( ∞ ) − u ( N ) ) ⋅ ∇ θ μ , φ ) = − (( u ( ∞ ) − u ( N ) ) θ μ , ∇ φ ) .
With φ = w ( t ) \varphi=w(t) φ = w ( t ) , (E1), Young’s inequality and ∥ θ μ ( t ) ∥ ≤ ∥ θ 0 ∥ \|\theta^\mu(t)\|\le\|\theta_0\| ∥ θ μ ( t ) ∥ ≤ ∥ θ 0 ∥ ,
sup t ≤ t 0 ∥ θ μ , N ( t ) − θ μ ( t ) ∥ 2 ≤ t 0 μ ∥ u ( ∞ ) − u ( N ) ∥ ∞ 2 ∥ θ 0 ∥ 2 ⟶ 0 ( N → ∞ ) . \sup_{t\le t_0}\|\theta^{\mu,N}(t)-\theta^\mu(t)\|^2\le\frac{t_0}{\mu}\|u^{(\infty)}-u^{(N)}\|_\infty^2\|\theta_0\|^2\longrightarrow0\qquad(N\to\infty). t ≤ t 0 sup ∥ θ μ , N ( t ) − θ μ ( t ) ∥ 2 ≤ μ t 0 ∥ u ( ∞ ) − u ( N ) ∥ ∞ 2 ∥ θ 0 ∥ 2 ⟶ 0 ( N → ∞ ) .
Hence D ∞ μ ( θ 0 ) D^\mu_\infty(\theta_0) D ∞ μ ( θ 0 ) is the pointwise limit of ∥ θ 0 ∥ 2 − ∥ θ μ , N ( t 0 ) ∥ 2 \|\theta_0\|^2-\|\theta^{\mu,N}(t_0)\|^2 ∥ θ 0 ∥ 2 − ∥ θ μ , N ( t 0 ) ∥ 2 , each of which is a Borel function of the data by Lemma 9.1 (applied with J = N J=N J = N and diffusivity μ \mu μ ); so it is Borel. Next, for a fixed realization and μ , μ ′ > 0 \mu,\mu'>0 μ , μ ′ > 0 , the same computation for w : = θ μ ′ − θ μ w:=\theta^{\mu'}-\theta^\mu w := θ μ ′ − θ μ , whose equation has drift u ( ∞ ) u^{(\infty)} u ( ∞ ) , diffusivity μ ′ \mu' μ ′ and right-hand side ( μ − μ ′ ) ( ∇ θ μ , ∇ φ ) (\mu-\mu')(\nabla\theta^\mu,\nabla\varphi) ( μ − μ ′ ) ( ∇ θ μ , ∇ φ ) , gives
sup t ≤ t 0 ∥ θ μ ′ ( t ) − θ μ ( t ) ∥ 2 ≤ ∣ μ ′ − μ ∣ 2 μ ′ ∫ 0 t 0 ∥ ∇ θ μ ∥ 2 d t ≤ ∣ μ ′ − μ ∣ 2 2 μ μ ′ ∥ θ 0 ∥ 2 . \sup_{t\le t_0}\|\theta^{\mu'}(t)-\theta^\mu(t)\|^2\le\frac{|\mu'-\mu|^2}{\mu'}\int_0^{t_0}\|\nabla\theta^\mu\|^2\,dt\le\frac{|\mu'-\mu|^2}{2\mu\mu'}\|\theta_0\|^2. t ≤ t 0 sup ∥ θ μ ′ ( t ) − θ μ ( t ) ∥ 2 ≤ μ ′ ∣ μ ′ − μ ∣ 2 ∫ 0 t 0 ∥∇ θ μ ∥ 2 d t ≤ 2 μ μ ′ ∣ μ ′ − μ ∣ 2 ∥ θ 0 ∥ 2 .
Thus μ ↦ D ∞ μ ( θ 0 ) \mu\mapsto D^\mu_\infty(\theta_0) μ ↦ D ∞ μ ( θ 0 ) is continuous on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , the lim sup in (b) may be taken over rational μ \mu μ , and the event in (b) is measurable.
Comparison with a truncation. Let μ J = ε 0 ℓ J 1 + α \mu_J=\varepsilon_0\ell_J^{1+\alpha} μ J = ε 0 ℓ J 1 + α and q : = 2 1 + α = μ J / μ J + 1 q:=2^{1+\alpha}=\mu_J/\mu_{J+1} q := 2 1 + α = μ J / μ J + 1 . For 0 < μ < q μ j 0 0<\mu<q\mu_{j_0} 0 < μ < q μ j 0 let J = J ( μ ) ≥ j 0 J=J(\mu)\ge j_0 J = J ( μ ) ≥ j 0 be the unique index with μ J ≤ μ < q μ J \mu_J\le\mu<q\mu_J μ J ≤ μ < q μ J ; then J ( μ ) → ∞ J(\mu)\to\infty J ( μ ) → ∞ as μ ↓ 0 \mu\downarrow0 μ ↓ 0 . Lemmas 2.4 and 10.8, applied for each realization with ( u , μ ) = ( u ( ∞ ) , μ ) (u,\mu)=(u^{(\infty)},\mu) ( u , μ ) = ( u ( ∞ ) , μ ) and ( v , μ ′ ) = ( u ( N ) , μ J ) (v,\mu')=(u^{(N)},\mu_J) ( v , μ ′ ) = ( u ( N ) , μ J ) , so that D u , μ = D ∞ μ ( θ 0 ) D_{u,\mu}=D^\mu_\infty(\theta_0) D u , μ = D ∞ μ ( θ 0 ) and D v , μ ′ = D J ( θ 0 ) D_{v,\mu'}=D_J(\theta_0) D v , μ ′ = D J ( θ 0 ) , with B = B > J B=B_{>J} B = B > J (so that u − v = div B > J u-v=\operatorname{div}B_{>J} u − v = div B > J ) and b ≤ C p o t ℓ J 1 + α = ( C p o t / ε 0 ) μ J b\le C_{\mathrm{pot}}\ell_J^{1+\alpha}=(C_{\mathrm{pot}}/\varepsilon_0)\mu_J b ≤ C pot ℓ J 1 + α = ( C pot / ε 0 ) μ J , give
D ∞ μ ( θ 0 ) ≥ μ J μ ( μ + C p o t ℓ J 1 + α ) 2 D J ( θ 0 ) ≥ a ∗ D J ( μ ) ( θ 0 ) . D^\mu_\infty(\theta_0)\ge\frac{\mu_J\mu}{(\mu+C_{\mathrm{pot}}\ell_J^{1+\alpha})^2}D_J(\theta_0)\ge a_*D_{J(\mu)}(\theta_0). D ∞ μ ( θ 0 ) ≥ ( μ + C pot ℓ J 1 + α ) 2 μ J μ D J ( θ 0 ) ≥ a ∗ D J ( μ ) ( θ 0 ) .
The first inequality uses μ ≥ μ ′ = μ J \mu\ge\mu'=\mu_J μ ≥ μ ′ = μ J , so that μ ′ + ∣ μ − μ ′ ∣ = μ \mu'+|\mu-\mu'|=\mu μ ′ + ∣ μ − μ ′ ∣ = μ in (97) . For the last inequality write c ′ : = μ / μ J ∈ [ 1 , q ) c':=\mu/\mu_J\in[1,q) c ′ := μ / μ J ∈ [ 1 , q ) : the middle factor is
c ′ ( c ′ + C p o t / ε 0 ) 2 = ( c ′ + C p o t ε 0 c ′ ) − 2 ≥ ( q + C p o t ε 0 ) − 2 = a ∗ . \frac{c'}{(c'+C_{\mathrm{pot}}/\varepsilon_0)^2}
=\left(\sqrt{c'}+\frac{C_{\mathrm{pot}}}{\varepsilon_0\sqrt{c'}}\right)^{-2}
\ge\left(\sqrt q+\frac{C_{\mathrm{pot}}}{\varepsilon_0}\right)^{-2}=a_*. ( c ′ + C pot / ε 0 ) 2 c ′ = ( c ′ + ε 0 c ′ C pot ) − 2 ≥ ( q + ε 0 C pot ) − 2 = a ∗ .
Taking expectations and letting μ ↓ 0 \mu\downarrow0 μ ↓ 0 , so that J ( μ ) → ∞ J(\mu)\to\infty J ( μ ) → ∞ , Theorem 1.4(a) gives (a). For (c), fix γ > 0 \gamma>0 γ > 0 and A ≥ 1 A\ge1 A ≥ 1 , let J ∗ = J ∗ ( γ , A , κ , ε 0 , t 0 ) J_*=J_*(\gamma,A,\kappa,\varepsilon_0,t_0) J ∗ = J ∗ ( γ , A , κ , ε 0 , t 0 ) be as in Theorem 1.4(c), uniformly for real mean-zero θ 0 ∈ H γ \theta_0\in H^\gamma θ 0 ∈ H γ with ∥ θ 0 ∥ H γ ≤ A ∥ θ 0 ∥ \|\theta_0\|_{H^\gamma}\le A\|\theta_0\| ∥ θ 0 ∥ H γ ≤ A ∥ θ 0 ∥ , and set μ ∗ : = μ J ∗ \mu_*:=\mu_{J_*} μ ∗ := μ J ∗ ; then μ ≤ μ ∗ \mu\le\mu_* μ ≤ μ ∗ implies J ( μ ) ≥ J ∗ J(\mu)\ge J_* J ( μ ) ≥ J ∗ , and (c) follows.
The probability statement. For (b), apply (98) at the matched diffusivities μ = μ ′ = μ J \mu=\mu'=\mu_J μ = μ ′ = μ J :
( 1 + C p o t / ε 0 ) − 2 D J ( θ 0 ) ≤ D ∞ μ ( θ 0 ) ≤ ( 1 + C p o t / ε 0 ) 2 D J ( θ 0 ) . (102) (1+C_{\mathrm{pot}}/\varepsilon_0)^{-2}D_J(\theta_0)\le D^\mu_\infty(\theta_0)\le(1+C_{\mathrm{pot}}/\varepsilon_0)^2D_J(\theta_0). \tag*{(102)} ( 1 + C pot / ε 0 ) − 2 D J ( θ 0 ) ≤ D ∞ μ ( θ 0 ) ≤ ( 1 + C pot / ε 0 ) 2 D J ( θ 0 ) . (102)
valid for every realization and every J ≥ j 0 J\ge j_0 J ≥ j 0 . Hence the event { lim sup J D J ( θ 0 ) > 0 } \{\limsup_JD_J(\theta_0)>0\} { lim sup J D J ( θ 0 ) > 0 } is contained in { lim sup μ ↓ 0 D ∞ μ ( θ 0 ) > 0 } \{\limsup_{\mu\downarrow0}D^\mu_\infty(\theta_0)>0\} { lim sup μ ↓ 0 D ∞ μ ( θ 0 ) > 0 } , and Theorem 1.4(b) gives the probability bound. □
Part (b) says that, on an event of probability at least c A D c_{\mathrm{AD}} c AD , a single realization of u ( ∞ ) u^{(\infty)} u ( ∞ ) dissipates the variance of θ 0 \theta_0 θ 0 anomalously. For the explicit model of Proposition 2.6 this gives a deterministic statement.
Corollary 10.10. Let d ≥ 2 d \ge2 d ≥ 2 , 0 < α < 1 0 < \alpha< 1 0 < α < 1 and t 0 ∈ ( 0 , 1 ] t_0 \in(0,1] t 0 ∈ ( 0 , 1 ] . There is a divergence-free vector field u ∈ L ∞ ( [ 0 , 1 ] ; C α ( T d ) ) u \in L^\infty([0,1];C^\alpha(\mathbb{T}^d)) u ∈ L ∞ ([ 0 , 1 ] ; C α ( T d )) such that, for θ 0 ( x ) = cos ( 2 π x 1 ) \theta_0(x) = \cos(2\pi x_1) θ 0 ( x ) = cos ( 2 π x 1 ) , the energy solutions θ μ \theta^\mu θ μ of ∂ t θ μ + u ⋅ ∇ θ μ = μ Δ θ μ \partial_t\theta^\mu+ u \cdot\nabla\theta^\mu= \mu\Delta\theta^\mu ∂ t θ μ + u ⋅ ∇ θ μ = μ Δ θ μ , θ μ ( 0 ) = θ 0 \theta^\mu(0) = \theta_0 θ μ ( 0 ) = θ 0 , satisfy
lim sup μ ↓ 0 μ ∫ 0 t 0 ∥ ∇ θ μ ( t ) ∥ 2 d t > 0. \limsup_{\mu\downarrow0}\mu\int_0^{t_0}\|\nabla\theta^\mu(t)\|^2\,dt > 0. μ ↓ 0 lim sup μ ∫ 0 t 0 ∥∇ θ μ ( t ) ∥ 2 d t > 0.
Proof. Take ψ ( x ) = ( 1 − ∣ x ∣ 2 ) + 5 \psi(x)=(1-|x|^2)_+^5 ψ ( x ) = ( 1 − ∣ x ∣ 2 ) + 5 and independent coefficients ± 1 \pm1 ± 1 with probability 1 2 \frac12 2 1 . This model satisfies (ND) (Proposition 2.6), (1.4) with M = σ = 1 M=\sigma=1 M = σ = 1 , and (S). Fix κ = κ 1 \kappa=\kappa_1 κ = κ 1 and ε 0 = ε 1 ( κ 1 ) \varepsilon_0=\varepsilon_1(\kappa_1) ε 0 = ε 1 ( κ 1 ) . By Theorem 1.5 (b) with this t 0 t_0 t 0 and θ 0 = cos ( 2 π x 1 ) \theta_0=\cos(2\pi x_1) θ 0 = cos ( 2 π x 1 ) , the event { lim sup μ ↓ 0 D ∞ μ ( θ 0 ) > 0 } \{\limsup_{\mu\downarrow0}D^\mu_\infty(\theta_0)>0\} { lim sup μ ↓ 0 D ∞ μ ( θ 0 ) > 0 } has probability at least c ( κ 1 t 0 ) 2 / β > 0 c(\kappa_1t_0)^{2/\beta}>0 c ( κ 1 t 0 ) 2/ β > 0 ; let u u u be the field u ( ∞ ) u^{(\infty)} u ( ∞ ) of any realization in this event. □
Remark 10.11. (1) By (10.7), at the diffusivity μ J \mu_J μ J the dissipation for the fixed field and the dissipation for the truncation u ( J ) u^{(J)} u ( J ) are comparable, with the constant ( 1 + C p o t / ε 0 ) 2 (1+C_{\mathrm{pot}}/\varepsilon_0)^2 ( 1 + C pot / ε 0 ) 2 , for every realization: the scales below the diffusive scale change the dissipation by at most a constant factor. The same comparison, with the constant ( 2 2 + α + C p o t / ε 0 ) 2 (2^{2+\alpha}+C_{\mathrm{pot}}/\varepsilon_0)^2 ( 2 2 + α + C pot / ε 0 ) 2 , holds between D ∞ μ ( θ 0 ) D^\mu_\infty(\theta_0) D ∞ μ ( θ 0 ) and D J ( μ ) ( θ 0 ) D_{J(\mu)}(\theta_0) D J ( μ ) ( θ 0 ) for every μ < 2 1 + α μ J 0 \mu<2^{1+\alpha}\mu_{J_0} μ < 2 1 + α μ J 0 , by Lemma 10.8 and μ J ( μ ) ≤ μ < 2 1 + α μ J ( μ ) \mu_{J(\mu)}\le\mu<2^{1+\alpha}\mu_{J(\mu)} μ J ( μ ) ≤ μ < 2 1 + α μ J ( μ ) .
(2) Part (a) is a lim inf of expectations over all diffusivities tending to zero. In part (b), the lim sup is taken for each fixed realization of the velocity; it is positive with probability at least c A D c_{\mathrm{AD}} c AD for each fixed nonzero datum. We do not claim a positive lim inf for individual realizations, probability one, or a single realization working simultaneously for all data: from E X J ≥ c \mathbb{E}X_J\ge c E X J ≥ c and X J ∈ [ 0 , 1 ] X_J\in[0,1] X J ∈ [ 0 , 1 ] one cannot conclude anything about lim inf J X J \liminf_J X_J lim inf J X J without information on the joint law of the X J X_J X J .
Appendix A. Proofs of the lemmas of Section 8
The notation, the facts (E1) and (E2), and the identities (8.1)–(8.3) are those of Section 8.
Proof of Lemma 8.2. (a) By (E1) with w ~ = w \widetilde w=w w = w , (8.3) with φ = w ( s ) \varphi=w(s) φ = w ( s ) and (8.1),
d d s ∥ w ∥ 2 = 2 ⟨ ∂ s w , w ⟩ = − 2 μ ∥ ∇ w ∥ 2 \frac{d}{ds}\|w\|^2=2\langle\partial_s w,w\rangle=-2\mu\|\nabla w\|^2 d s d ∥ w ∥ 2 = 2 ⟨ ∂ s w , w ⟩ = − 2 μ ∥∇ w ∥ 2
for a.e. s s s . The difference of two energy solutions with the same initial value is an energy solution with initial value 0 0 0 , hence vanishes.
(b) The constant 1 1 1 is in L 2 ( I ; H 1 ) L^2(I;H^1) L 2 ( I ; H 1 ) with derivative 0 0 0 , so by (E1) and (8.3) with φ = 1 \varphi=1 φ = 1 ,
d d s ∫ w = ⟨ ∂ s w , 1 ⟩ = − ( b ⋅ ∇ w , 1 ) = 0 \frac{d}{ds}\int w=\langle\partial_s w,1\rangle=-(b\cdot\nabla w,1)=0 d s d ∫ w = ⟨ ∂ s w , 1 ⟩ = − ( b ⋅ ∇ w , 1 ) = 0
by (8.1). For positivity write v − = max ( − v , 0 ) v_-=\max(-v,0) v − = max ( − v , 0 ) and let w k w_k w k be as in (E1). The functional v ↦ 1 2 ∥ v − ∥ 2 v\mapsto\frac12\|v_-\|^2 v ↦ 2 1 ∥ v − ∥ 2 is continuously differentiable on L 2 L^2 L 2 with derivative − v − -v_- − v − , since r ↦ 1 2 ( r − ) 2 r\mapsto\frac12(r_-)^2 r ↦ 2 1 ( r − ) 2 has the 1-Lipschitz derivative − r − -r_- − r − . As s ↦ w k ( s ) ∈ L 2 s\mapsto w_k(s)\in L^2 s ↦ w k ( s ) ∈ L 2 is C 1 C^1 C 1 ,
1 2 ∥ ( w k ) − ( s ) ∥ 2 − 1 2 ∥ ( w k ) − ( T 0 ) ∥ 2 = − ∫ T 0 s ⟨ ∂ s w k , ( w k ) − ⟩ . (103) \frac12\|(w_k)_-(s)\|^2-\frac12\|(w_k)_-(T_0)\|^2=-\int_{T_0}^s\langle\partial_s w_k,(w_k)_-\rangle. \tag*{(103)} 2 1 ∥ ( w k ) − ( s ) ∥ 2 − 2 1 ∥ ( w k ) − ( T 0 ) ∥ 2 = − ∫ T 0 s ⟨ ∂ s w k , ( w k ) − ⟩ . (103)
As k → ∞ k \to\infty k → ∞ the left side converges to the same expression for w w w , because ∣ a − − b − ∣ ≤ ∣ a − b ∣ |a_{-}-b_{-}| \le|a-b| ∣ a − − b − ∣ ≤ ∣ a − b ∣ and w k → w w_k \to w w k → w in C ( I ; L 2 ) C(I;L^2) C ( I ; L 2 ) . On the right, ∂ s w k → ∂ s w \partial_s w_k \to\partial_s w ∂ s w k → ∂ s w in L 2 ( I ; H − 1 ) L^2(I;H^{-1}) L 2 ( I ; H − 1 ) , while ( w k ) − → w − (w_k)_{-} \to w_{-} ( w k ) − → w − in L 2 ( I ; L 2 ) L^2(I;L^2) L 2 ( I ; L 2 ) and ∥ ∇ ( w k ) − ∥ ≤ ∥ ∇ w k ∥ \|\nabla(w_k)_{-}\| \le\|\nabla w_k\| ∥∇ ( w k ) − ∥ ≤ ∥∇ w k ∥ , so ( w k ) − (w_k)_{-} ( w k ) − is bounded in L 2 ( I ; H 1 ) L^2(I;H^1) L 2 ( I ; H 1 ) and converges to w − w_{-} w − weakly there. Hence 1 2 ∥ w − ( s ) ∥ 2 − 1 2 ∥ w − ( T 0 ) ∥ 2 = − ∫ T 0 s ⟨ ∂ s w , w − ⟩ \frac{1}{2}\|w_{-}(s)\|^2-\frac{1}{2}\|w_{-}(T_0)\|^2=-\int_{T_0}^{s}\langle\partial_s w,w_{-}\rangle 2 1 ∥ w − ( s ) ∥ 2 − 2 1 ∥ w − ( T 0 ) ∥ 2 = − ∫ T 0 s ⟨ ∂ s w , w − ⟩ . Here we used that v − ∈ H 1 v_{-}\in H^1 v − ∈ H 1 with ∇ v − = − 1 { v < 0 } ∇ v \nabla v_{-}=-\mathbf{1}_{\{v<0\}}\nabla v ∇ v − = − 1 { v < 0 } ∇ v for v ∈ H 1 v\in H^1 v ∈ H 1 [[23 Lemma 7.6] ]. By (88) , this formula, the pointwise identity ( b ⋅ ∇ w ) w − = − ( b ⋅ ∇ w − ) w − (b\cdot\nabla w)w_{-}=-(b\cdot\nabla w_{-})w_{-} ( b ⋅ ∇ w ) w − = − ( b ⋅ ∇ w − ) w − and (86) ,
− ⟨ ∂ s w , w − ⟩ = ( b ⋅ ∇ w , w − ) + μ ( ∇ w , ∇ w − ) = − ( b ⋅ ∇ w − , w − ) − μ ∥ ∇ w − ∥ 2 ≤ 0. -\langle\partial_s w,w_{-}\rangle=(b\cdot\nabla w,w_{-})+\mu(\nabla w,\nabla w_{-})=-(b\cdot\nabla w_{-},w_{-})-\mu\|\nabla w_{-}\|^2\le0. − ⟨ ∂ s w , w − ⟩ = ( b ⋅ ∇ w , w − ) + μ ( ∇ w , ∇ w − ) = − ( b ⋅ ∇ w − , w − ) − μ ∥∇ w − ∥ 2 ≤ 0.
Thus ∥ w − ( s ) ∥ ≤ ∥ w − ( T 0 ) ∥ = 0 \|w_{-}(s)\|\le\|w_{-}(T_0)\|=0 ∥ w − ( s ) ∥ ≤ ∥ w − ( T 0 ) ∥ = 0 . (c) The function G G G belongs to C ( [ s 1 , s 2 ] ; H 1 ) C([s_1,s_2];H^1) C ([ s 1 , s 2 ] ; H 1 ) , and its classical derivative ∂ s G \partial_sG ∂ s G , which is continuous, is its distributional derivative. By (E1) on [ s 1 , s 2 ] [s_1,s_2] [ s 1 , s 2 ] , s ↦ ∫ G w s\mapsto\int Gw s ↦ ∫ Gw is absolutely continuous with derivative ⟨ ∂ s w , G ⟩ + ( ∂ s G , w ) \langle\partial_s w,G\rangle+(\partial_sG,w) ⟨ ∂ s w , G ⟩ + ( ∂ s G , w ) . By (88) , (86) and an integration by parts, ⟨ ∂ s w , G ⟩ = − ( b ⋅ ∇ w , G ) − μ ( ∇ w , ∇ G ) = ( b ⋅ ∇ G , w ) + μ ( w , Δ G ) \langle\partial_s w,G\rangle=-(b\cdot\nabla w,G)-\mu(\nabla w,\nabla G)=(b\cdot\nabla G,w)+\mu(w,\Delta G) ⟨ ∂ s w , G ⟩ = − ( b ⋅ ∇ w , G ) − μ ( ∇ w , ∇ G ) = ( b ⋅ ∇ G , w ) + μ ( w , Δ G ) . Integrate in s s s . □
Proof of Lemma 8.3. Let w ( 1 ) , w ( 2 ) w^{(1)},w^{(2)} w ( 1 ) , w ( 2 ) be the energy solutions with initial values 1 A w ( s 1 ) \mathbf{1}_A w(s_1) 1 A w ( s 1 ) and 1 A c w ( s 1 ) \mathbf{1}_{A^c}w(s_1) 1 A c w ( s 1 ) . By Lemma 8.2(a), w = w ( 1 ) + w ( 2 ) w=w^{(1)}+w^{(2)} w = w ( 1 ) + w ( 2 ) . By Lemma 8.2(b), w ( 1 ) ≥ 0 w^{(1)}\ge0 w ( 1 ) ≥ 0 and ∫ w ( 1 ) ( s 2 ) = ∫ A w ( s 1 ) \int w^{(1)}(s_2)=\int_A w(s_1) ∫ w ( 1 ) ( s 2 ) = ∫ A w ( s 1 ) , so ∫ B w ( 1 ) ( s 2 ) ≤ ∫ A w ( s 1 ) \int_B w^{(1)}(s_2)\le\int_A w(s_1) ∫ B w ( 1 ) ( s 2 ) ≤ ∫ A w ( s 1 ) . If m : = ∫ A c w ( s 1 ) > 0 m:=\int_{A^c}w(s_1)>0 m := ∫ A c w ( s 1 ) > 0 , then w ( 2 ) / m w^{(2)}/m w ( 2 ) / m is admissible in (89) for the set A c A^c A c , so ∫ B w ( 2 ) ( s 2 ) ≤ m Π s 1 , s 2 ( A c , B ) \int_B w^{(2)}(s_2)\le m\Pi_{s_1,s_2}(A^c,B) ∫ B w ( 2 ) ( s 2 ) ≤ m Π s 1 , s 2 ( A c , B ) ; if m = 0 m=0 m = 0 , then w ( 2 ) = 0 w^{(2)}=0 w ( 2 ) = 0 . □
Proof of Lemma 8.4. The function w ~ ( s ) : = w ( T − s ) \widetilde w(s):=w(T-s) w ( s ) := w ( T − s ) lies in L 2 ( 0 , T ; H 1 ) L^2(0,T;H^1) L 2 ( 0 , T ; H 1 ) , with ∂ s w ~ ( s ) = − ( ∂ s w ) ( T − s ) ∈ L 2 ( 0 , T ; H − 1 ) \partial_s\widetilde w(s)=-(\partial_s w)(T-s)\in L^2(0,T;H^{-1}) ∂ s w ( s ) = − ( ∂ s w ) ( T − s ) ∈ L 2 ( 0 , T ; H − 1 ) . By (E1), I ( s ) : = ( w ~ ( s ) , ϱ ( s ) ) I(s):=(\widetilde w(s),\varrho(s)) I ( s ) := ( w ( s ) , ϱ ( s )) is absolutely continuous, and for a.e. s s s , by (88) for w w w at time T − s T-s T − s and for ϱ \varrho ϱ at time s s s , with b = b ( T − s ) b=b(T-s) b = b ( T − s ) ,
I ′ ( s ) = [ ( b ⋅ ∇ w ~ , ϱ ) + μ ( ∇ w ~ , ∇ ϱ ) ] + [ ( b ⋅ ∇ ϱ , w ~ ) − μ ( ∇ ϱ , ∇ w ~ ) ] = 0 , I'(s)=[(b\cdot\nabla\widetilde w,\varrho)+\mu(\nabla\widetilde w,\nabla\varrho)]+[(b\cdot\nabla\varrho,\widetilde w)-\mu(\nabla\varrho,\nabla\widetilde w)]=0, I ′ ( s ) = [( b ⋅ ∇ w , ϱ ) + μ ( ∇ w , ∇ ϱ )] + [( b ⋅ ∇ ϱ , w ) − μ ( ∇ ϱ , ∇ w )] = 0 ,
by (86) . Hence I ( 0 ) = I ( T ) I(0)=I(T) I ( 0 ) = I ( T ) . □
Proof of Lemma 8.5. By Lemma 8.2(a), ∥ w k ( s ) ∥ ≤ ∥ w 0 ∥ \|w_k(s)\|\le\|w_0\| ∥ w k ( s ) ∥ ≤ ∥ w 0 ∥ and 2 μ ∫ I ∥ ∇ w k ∥ 2 ≤ ∥ w 0 ∥ 2 2\mu\int_I\|\nabla w_k\|^2\le\|w_0\|^2 2 μ ∫ I ∥∇ w k ∥ 2 ≤ ∥ w 0 ∥ 2 . It suffices to show that every subsequence has a further subsequence along which w k ( s ) ⇀ w ( s ) w_k(s)\rightharpoonup w(s) w k ( s ) ⇀ w ( s ) for every s s s . Pass to a subsequence along which w k ⇀ w ‾ w_k\rightharpoonup\overline w w k ⇀ w weakly in L 2 ( I ; H 1 ) L^2(I;H^1) L 2 ( I ; H 1 ) . For φ ∈ C ∞ ( T n ) \varphi\in C^\infty(\mathbb{T}^n) φ ∈ C ∞ ( T n ) , Lemma 8.2(c) with G ( s , Y ) = φ ( Y ) G(s,Y)=\varphi(Y) G ( s , Y ) = φ ( Y ) gives
( w k ( s ) , φ ) = ( w 0 , φ ) + ∫ T 0 s ( w k , b k ⋅ ∇ φ + μ Δ φ ) d r . (w_k(s),\varphi)=(w_0,\varphi)+\int_{T_0}^{s}(w_k,b_k\cdot\nabla\varphi+\mu\Delta\varphi)\,dr. ( w k ( s ) , φ ) = ( w 0 , φ ) + ∫ T 0 s ( w k , b k ⋅ ∇ φ + μ Δ φ ) d r .
Since b k ⋅ ∇ φ → b ⋅ ∇ φ b_k\cdot\nabla\varphi\to b\cdot\nabla\varphi b k ⋅ ∇ φ → b ⋅ ∇ φ in L 2 ( I × T n ) L^2(I\times\mathbb{T}^n) L 2 ( I × T n ) by dominated convergence, the right side converges for every s s s to ℓ s ( φ ) : = ( w 0 , φ ) + ∫ T 0 s ( w ‾ , b ⋅ ∇ φ + μ Δ φ ) d r \ell_s(\varphi):=(w_0,\varphi)+\int_{T_0}^{s}(\overline w,b\cdot\nabla\varphi+\mu\Delta\varphi)\,dr ℓ s ( φ ) := ( w 0 , φ ) + ∫ T 0 s ( w , b ⋅ ∇ φ + μ Δ φ ) d r . As ∥ w k ( s ) ∥ ≤ ∥ w 0 ∥ \|w_k(s)\|\le\|w_0\| ∥ w k ( s ) ∥ ≤ ∥ w 0 ∥ , it follows that for every s s s , w k ( s ) w_k(s) w k ( s ) converges weakly in L 2 L^2 L 2 to some λ s \lambda_s λ s with ( λ s , φ ) = ℓ s ( φ ) (\lambda_s,\varphi)=\ell_s(\varphi) ( λ s , φ ) = ℓ s ( φ ) . Testing against χ ∈ L 2 ( I ) \chi\in L^2(I) χ ∈ L 2 ( I ) and using dominated convergence shows ( w ‾ ( s ) , φ ) = ℓ s ( φ ) (\overline w(s),\varphi)=\ell_s(\varphi) ( w ( s ) , φ ) = ℓ s ( φ ) for a.e. s s s . By (86) , ( w ‾ ( r ) , b ⋅ ∇ φ + μ Δ φ ) = ⟨ a r , φ ⟩ (\overline w(r),b\cdot\nabla\varphi+\mu\Delta\varphi)=\langle a_r,\varphi\rangle ( w ( r ) , b ⋅ ∇ φ + μ Δ φ ) = ⟨ a r , φ ⟩ , where a r : = − b ( r ) ⋅ ∇ w ‾ ( r ) + μ Δ w ‾ ( r ) ∈ H − 1 a_r:=-b(r)\cdot\nabla\overline w(r)+\mu\Delta\overline w(r)\in H^{-1} a r := − b ( r ) ⋅ ∇ w ( r ) + μ Δ w ( r ) ∈ H − 1 has norm at most ( B 0 + μ ) ∥ ∇ w ‾ ( r ) ∥ (B_0+\mu)\|\nabla\overline w(r)\| ( B 0 + μ ) ∥∇ w ( r ) ∥ . The map r ↦ a r r\mapsto a_r r ↦ a r is weakly measurable, hence strongly measurable by Pettis’ theorem, and lies in L 2 ( I ; H − 1 ) L^2(I;H^{-1}) L 2 ( I ; H − 1 ) . Thus w ‾ = w 0 + ∫ T 0 s a r d r \overline w=w_0+\int_{T_0}^{s}a_r\,dr w = w 0 + ∫ T 0 s a r d r in H − 1 H^{-1} H − 1 for a.e. s s s , so ∂ s w ‾ = a ∈ L 2 ( I ; H − 1 ) \partial_s\overline w=a\in L^2(I;H^{-1}) ∂ s w = a ∈ L 2 ( I ; H − 1 ) , and (88) holds: w ‾ \overline w w is an energy solution. Its continuous representative satisfies ( w ˉ ( s ) , φ ) = ℓ s ( φ ) (\bar{w}(s), \varphi) = \ell_s(\varphi) ( w ˉ ( s ) , φ ) = ℓ s ( φ ) for all s s s , in particular w ˉ ( T 0 ) = w 0 \bar{w}(T_0) = w_0 w ˉ ( T 0 ) = w 0 . By uniqueness w ˉ = w \bar{w} = w w ˉ = w , so λ s = w ( s ) \lambda_s = w(s) λ s = w ( s ) for every s s s . □ \square □
Proof of Lemma 8.6. (a) Let N N N be the null set of (D) for b b b . For s ∉ N s \notin N s ∈ / N and φ ∈ C ∞ ( T 2 d ) \varphi\in C^\infty(\mathbb{T}^{2d}) φ ∈ C ∞ ( T 2 d ) ,
∬ b ( s , y ) ⋅ ∇ y φ d y d y ′ = 0 , \iint b(s,y)\cdot\nabla_y\varphi\,\mathrm{d}y\,\mathrm{d}y' = 0, ∬ b ( s , y ) ⋅ ∇ y φ d y d y ′ = 0 ,
since the inner integral vanishes for each y ′ y' y ′ , and similarly in y ′ y' y ′ ; so b b b satisfies (D). Since sup s ∥ θ i ( s ) ∥ < ∞ \sup_s\|\theta_i(s)\|<\infty sup s ∥ θ i ( s ) ∥ < ∞ and θ i ∈ L 2 ( I ; H 1 ) \theta_i\in L^2(I;H^1) θ i ∈ L 2 ( I ; H 1 ) , θ 1 ⊗ θ 2 ∈ L 2 ( I ; H 1 ( T 2 d ) ) \theta_1\otimes\theta_2\in L^2(I;H^1(\mathbb{T}^{2d})) θ 1 ⊗ θ 2 ∈ L 2 ( I ; H 1 ( T 2 d )) and θ 1 ⊗ θ 2 ∈ C ( I ; L 2 ( T 2 d ) ) \theta_1\otimes\theta_2\in C(I;L^2(\mathbb{T}^{2d})) θ 1 ⊗ θ 2 ∈ C ( I ; L 2 ( T 2 d )) . For φ 1 , φ 2 ∈ C ∞ ( T d ) \varphi_1,\varphi_2\in C^\infty(\mathbb{T}^d) φ 1 , φ 2 ∈ C ∞ ( T d ) , by (E1) and (88) for θ 1 , θ 2 \theta_1,\theta_2 θ 1 , θ 2 , the function ( ( θ 1 ⊗ θ 2 ) ( s ) , φ 1 ⊗ φ 2 ) = ( θ 1 ( s ) , φ 1 ) ( θ 2 ( s ) , φ 2 ) ((\theta_1\otimes\theta_2)(s),\varphi_1\otimes\varphi_2)=(\theta_1(s),\varphi_1)(\theta_2(s),\varphi_2) (( θ 1 ⊗ θ 2 ) ( s ) , φ 1 ⊗ φ 2 ) = ( θ 1 ( s ) , φ 1 ) ( θ 2 ( s ) , φ 2 ) is absolutely continuous with derivative − a s ( φ 1 ⊗ φ 2 ) -a_s(\varphi_1\otimes\varphi_2) − a s ( φ 1 ⊗ φ 2 ) , where, for φ ∈ H 1 ( T 2 d ) \varphi\in H^1(\mathbb{T}^{2d}) φ ∈ H 1 ( T 2 d ) ,
a s ( φ ) : = ( b ( s ) ⋅ ∇ ( θ 1 ⊗ θ 2 ) ( s ) , φ ) + μ ( ∇ ( θ 1 ⊗ θ 2 ) ( s ) , ∇ φ ) , a_s(\varphi):=(b(s)\cdot\nabla(\theta_1\otimes\theta_2)(s),\varphi)+\mu(\nabla(\theta_1\otimes\theta_2)(s),\nabla\varphi), a s ( φ ) := ( b ( s ) ⋅ ∇ ( θ 1 ⊗ θ 2 ) ( s ) , φ ) + μ ( ∇ ( θ 1 ⊗ θ 2 ) ( s ) , ∇ φ ) ,
∣ a s ( φ ) ∣ ≤ ( 2 B 0 + μ ) ∥ ∇ ( θ 1 ⊗ θ 2 ) ( s ) ∥ ∥ φ ∥ H 1 . |a_s(\varphi)|\leq(\sqrt{2}B_0+\mu)\|\nabla(\theta_1\otimes\theta_2)(s)\|\|\varphi\|_{H^1}. ∣ a s ( φ ) ∣ ≤ ( 2 B 0 + μ ) ∥∇ ( θ 1 ⊗ θ 2 ) ( s ) ∥∥ φ ∥ H 1 .
Finite sums of such products are dense in H 1 ( T 2 d ) H^1(\mathbb{T}^{2d}) H 1 ( T 2 d ) (trigonometric polynomials), so ∂ s ( θ 1 ⊗ θ 2 ) = − a s ∈ L 2 ( I ; H − 1 ( T 2 d ) ) \partial_s(\theta_1\otimes\theta_2)=-a_s\in L^2(I;H^{-1}(\mathbb{T}^{2d})) ∂ s ( θ 1 ⊗ θ 2 ) = − a s ∈ L 2 ( I ; H − 1 ( T 2 d )) and (88) holds. The claim follows by uniqueness; the last statement is obvious.
(b) The map ι \iota ι preserves Lebesgue measure, ∇ ( v ∘ ι − 1 ) = g ( ∇ v ) ∘ ι − 1 \nabla(v\circ\iota^{-1})=g(\nabla v)\circ\iota^{-1} ∇ ( v ∘ ι − 1 ) = g ( ∇ v ) ∘ ι − 1 , and g g g is orthogonal. Hence ( ι ∗ b ⋅ ∇ ( w ∘ ι − 1 ) , φ ) = ( b ⋅ ∇ w , φ ∘ ι ) (\iota_*b\cdot\nabla(w\circ\iota^{-1}),\varphi)=(b\cdot\nabla w,\varphi\circ\iota) ( ι ∗ b ⋅ ∇ ( w ∘ ι − 1 ) , φ ) = ( b ⋅ ∇ w , φ ∘ ι ) , ( ∇ ( w ∘ ι − 1 ) , ∇ φ ) = ( ∇ w , ∇ ( φ ∘ ι ) ) (\nabla(w\circ\iota^{-1}),\nabla\varphi)=(\nabla w,\nabla(\varphi\circ\iota)) ( ∇ ( w ∘ ι − 1 ) , ∇ φ ) = ( ∇ w , ∇ ( φ ∘ ι )) and ⟨ ∂ s ( w ∘ ι − 1 ) , φ ⟩ = ⟨ ∂ s w , φ ∘ ι ⟩ \langle\partial_s(w\circ\iota^{-1}),\varphi\rangle=\langle\partial_sw,\varphi\circ\iota\rangle ⟨ ∂ s ( w ∘ ι − 1 ) , φ ⟩ = ⟨ ∂ s w , φ ∘ ι ⟩ , so (88) for w ∘ ι − 1 w\circ\iota^{-1} w ∘ ι − 1 and φ \varphi φ is (88) for w w w and φ ∘ ι \varphi\circ\iota φ ∘ ι . The same change of variables gives (D) for ι ∗ b \iota_*b ι ∗ b . □ \square □
References
Department of Mathematics, Princeton University, Fine Hall, Washington Road, Princeton, NJ 08544, USA
Email address: cf@math.princeton.edu
Department of Mathematics, ETH Zürich, Rämistrasse 101, 8092 Zürich, Switzerland
Email address: svitlana.mayboroda@math.ethz.ch