Introduction
Let ( Ω , F , P ) (\Omega,\mathcal{F},\mathbb{P}) ( Ω , F , P ) be a probability space, and let V ( y , ω ) V(y,\omega) V ( y , ω ) be a real valued stationary ergodic random field on R d \mathbb{R}^d R d with mean
V ˉ : = E [ V ( 0 , ω ) ] . (1) \bar{V} := \mathbb{E}[V(0,\omega)]. \tag*{(1)} V ˉ := E [ V ( 0 , ω )] . (1)
We study the random Schrödinger equation with rapidly oscillating stationary random potential. Define the microscopic and macroscopic Schrödinger operators
H ε : = − Δ + V ( x ε , ω ) , H 0 : = − Δ + V ˉ . (2) H_\varepsilon:= -\Delta+ V\left(\frac{x}{\varepsilon},\omega\right), \qquad H_0 := -\Delta+ \bar{V}. \tag*{(2)} H ε := − Δ + V ( ε x , ω ) , H 0 := − Δ + V ˉ . (2)
The microscopic equation is
{ i ∂ t u ε = H ε u ε , u ε ( 0 , x ) = u 0 ( x ) , ( t , x ) ∈ R × R d (LSE ε ) \left\{\begin{aligned} i\partial_tu_\varepsilon&= H_\varepsilon u_\varepsilon,\\ u_\varepsilon(0,x) &= u_0(x),\end{aligned}\right. \qquad(t,x)\in\mathbb{R}\times\mathbb{R}^d \qquad\text{(LSE}_{\varepsilon}\text{)} { i ∂ t u ε u ε ( 0 , x ) = H ε u ε , = u 0 ( x ) , ( t , x ) ∈ R × R d (LSE ε )
with a real potential V V V and initial data u 0 ∈ L 2 ( R d ) u_0\in L^2(\mathbb{R}^d) u 0 ∈ L 2 ( R d ) . As ε → 0 \varepsilon\to0 ε → 0 , the microscopic model (LSEε _\varepsilon ε ) formally converges to the deterministic effective equation
{ i ∂ t u ˉ = H 0 u ˉ , u ˉ ( 0 , x ) = u 0 ( x ) , ( t , x ) ∈ R × R d (LSE 0 ) \left\{\begin{aligned} i\partial_t\bar{u} &= H_0\bar{u},\\ \bar{u}(0,x) &= u_0(x),\end{aligned}\right. \qquad(t,x)\in\mathbb{R}\times\mathbb{R}^d \qquad\text{(LSE}_{0}\text{)} { i ∂ t u ˉ u ˉ ( 0 , x ) = H 0 u ˉ , = u 0 ( x ) , ( t , x ) ∈ R × R d (LSE 0 )
Define the fluctuation field
w ε : = u ε − u ˉ . (3) w_\varepsilon:= u_\varepsilon-\bar{u}. \tag*{(3)} w ε := u ε − u ˉ . (3)
The core mathematical problems addressed in this work are the decay of w ε w_\varepsilon w ε as ε → 0 \varepsilon\to0 ε → 0 , the explicit convergence rate, and the statistical limiting behavior of the rescaled fluctuation ε − d 2 w ε \varepsilon^{-\frac{d}{2}}w_\varepsilon ε − 2 d w ε .
F. Meng was supported by the National Natural Science Foundation of China (Grant No. 12601455) and partially supported by the National Natural Science Foundation of China (Grant No. 12671284).
The stochastic fluctuation behavior is entirely governed by the covariance decay of the centered field
a ( y ) : = V ( y ) − V ‾ , (4) a(y) := V(y) - \overline{V}, \tag*{(4)} a ( y ) := V ( y ) − V , (4)
whose covariance is
R ( y ) : = E [ a ( y ) a ( 0 ) ] , R ^ ( ξ ) = ∫ R d e − i ξ ⋅ y R ( y ) d y . (5) R(y) := \mathbb{E}[a(y)a(0)], \qquad\widehat{R}(\xi) = \int_{\mathbb{R}^{d}} e^{-i\xi\cdot y}R(y)\,dy. \tag*{(5)} R ( y ) := E [ a ( y ) a ( 0 )] , R ( ξ ) = ∫ R d e − i ξ ⋅ y R ( y ) d y . (5)
Under exponential mixing, R R R decays exponentially and R ^ \widehat{R} R is continuous. The critical parameter is R ^ ( 0 ) = ∫ R d R ( y ) d y \widehat{R}(0) = \int_{\mathbb{R}^{d}} R(y)\,dy R ( 0 ) = ∫ R d R ( y ) d y , the variance of the spatial block average.
The main results of this paper are threefold. First, we prove the optimal algebraic rate
sup 0 ≤ t ≤ T ∥ w ε ( t ) ∥ L 2 ( Ω ; L 2 ( R d ) ) ≤ C T , u 0 ε d 2 . (6) \sup_{0\leq t\leq T} \|w_{\varepsilon}(t)\|_{L^{2}(\Omega;L^{2}(\mathbb{R}^{d}))} \leq C_{T,u_{0}}\varepsilon^{\frac{d}{2}}. \tag*{(6)} 0 ≤ t ≤ T sup ∥ w ε ( t ) ∥ L 2 ( Ω ; L 2 ( R d )) ≤ C T , u 0 ε 2 d . (6)
Second, we prove almost sure convergence,
lim ε → 0 sup 0 ≤ t ≤ T ∥ w ε ( t ) ∥ L 2 ( R d ) = 0 a.s. (7) \lim_{\varepsilon\to0}\sup_{0\leq t\leq T}\|w_{\varepsilon}(t)\|_{L^{2}(\mathbb{R}^{d})}=0 \qquad\text{a.s.} \tag*{(7)} ε → 0 lim 0 ≤ t ≤ T sup ∥ w ε ( t ) ∥ L 2 ( R d ) = 0 a.s. (7)
Third, we prove a fluctuation central limit theorem: the rescaled field ε − d 2 w ε \varepsilon^{-\frac{d}{2}}w_{\varepsilon} ε − 2 d w ε converges in distribution to the Gaussian solution G \mathcal{G} G of the linear random Schrödinger equation
{ i ∂ t G = H 0 G + R ^ ( 0 ) W ˙ ( t , x ) , G ( 0 , x ) = 0 , ( t , x ) ∈ R × R d . \begin{cases}
i\partial_t\mathcal{G}=H_{0}\mathcal{G}+\sqrt{\widehat{R}(0)}\,\dot{W}(t,x),\\
\mathcal{G}(0,x)=0,
\end{cases}
\qquad(t,x)\in\mathbb{R}\times\mathbb{R}^{d}. { i ∂ t G = H 0 G + R ( 0 ) W ˙ ( t , x ) , G ( 0 , x ) = 0 , ( t , x ) ∈ R × R d .
where W ˙ \dot{W} W ˙ is space time white noise. The additive white noise enters because the fluctuation is driven by the centered potential a ( x ε ) a\left(\frac{x}{\varepsilon}\right) a ( ε x ) at the first Duhamel order.
We assume that V V V has almost surely continuous sample paths on R d \mathbb{R}^{d} R d , hence is bounded on compact sets. Under the sub-Gaussian and exponential mixing assumptions, V V V has at most Gaussian tails and is in particular in the Kato class, so − Δ + V -\Delta+V − Δ + V is self-adjoint on H 2 ( R d ) H^{2}(\mathbb{R}^{d}) H 2 ( R d ) and ( L S E ε ) (\mathrm{LSE}_{\varepsilon}) ( LSE ε ) is well-posed in L 2 ( R d ) L^{2}(\mathbb{R}^{d}) L 2 ( R d ) by the standard Schrödinger theory [22 Chapter 1] . The estimates below hold in every dimension d ≥ 1 d\geq1 d ≥ 1 , with constants depending only on d d d , and are uniform in V ‾ \overline{V} V . The transparent Fourier computation of the second moment is written first for d = 1 , 2 , 3 d=1,2,3 d = 1 , 2 , 3 : the far region dyadic shell weight 2 k ( d − 4 ) 2^{k(d-4)} 2 k ( d − 4 ) , which comes from Lemma 3.3 and Proposition 3.4, converges geometrically precisely when d < 4 d<4 d < 4 . Although exponential mixing makes R R R decay exponentially and hence R ^ \widehat{R} R smooth, the dyadic sum only uses the L ∞ L^{\infty} L ∞ bound on R ^ \widehat{R} R , and the radial tail ∫ ε ∞ r d − 5 d r \int_{\varepsilon}^{\infty}r^{d-5}\,dr ∫ ε ∞ r d − 5 d r diverges at d ≥ 4 d\geq4 d ≥ 4 . The same upper bound in d ≥ 4 d\geq4 d ≥ 4 uses a physical space oscillatory integral argument, with M = ⌈ d / 4 ⌉ + 1 M=\lceil d/4\rceil+1 M = ⌈ d /4 ⌉ + 1 time integrations by parts, developed in Appendix A. The matching lower bound, the all moment estimate, the almost sure convergence and the fluctuation CLT use only the near region ∣ ξ ′ ∣ ≤ ε |\xi'|\leq\varepsilon ∣ ξ ′ ∣ ≤ ε , which is pure volume and is dimension independent.
Background and setting
Quantitative stochastic homogenization for divergence form elliptic operators traces back to the foundational work of Papanicolaou–Varadhan [21 ] , who laid the framework for studying partial differential equations with stationary ergodic random coefficients. Substantial progress was later achieved by Gloria–Neukamm–Otto in their influential quantitative program, yielding sharp variance estimates and optimal L 2 L^{2} L 2 homogenization rates under spectral-gap assumptions [8 , 10 , 11 ] . The spectral gap condition encodes a strong decorrelation mechanism; further quantitative ergodicity analysis under this condition can be found in [9 ] . Related variance bounds for correlated random fields are developed in Mourrat [20 ] .
For parabolic and wave equations in random media, homogenization theory and associated fluctuation central limit theorem results have been extensively investigated by Bal [2 , 3 ] , Gu–Bal [13 ] , and Bal–Jing [4 ] . In the parabolic setting, the heat semigroup possesses intrinsic smoothing which damps high frequency random oscillations automatically, making many homogenization and fluctuation arguments comparatively tractable [5 ] . Central limit theorem phenomena for homogenization fluctuations in scalar stochastic PDE are further studied in Gu [12 ] , and the martingale decomposition underlying our fluctuation analysis dates back to Kipnis–Varadhan [19 ] .
The dispersive Schrödinger equation exhibits fundamentally different behavior: the propagator e − i t H e^{-itH} e − i t H is unitary and lacks regularizing smoothing effects, so random potential oscillations are not naturally suppressed. Deterministic periodic homogenization for Schrödinger type systems originates from the Bloch wave analysis of Allaire–Piatnitski [1 ] . For stationary random potentials, the primary prior quantitative contribution is due to Zhang–Bal [24 ] , who established convergence in probability under a spectral-gap hypothesis. Further results for time dependent random Schrödinger homogenization appear in Gu–Ryzhik [14 ] . Recent work by Harrop–Griffiths–Killip–Visan [17 , 18 ] and Harrop–Griffiths–Ntekoume [16 ] studies the Schrödinger equation with sprinkled random nonlinearity.
The unitarity of the Schrödinger propagator changes the fluctuation computation in a specific way. In the parabolic case, the heat semigroup damps the off-diagonal frequencies, and the corrector solves an elliptic cell problem. Here the time double integral in the second moment is
∫ 0 t ∫ 0 t e i ( r − r ′ ) σ d r d r ′ = 2 ( 1 − cos ( t σ ) ) σ 2 , σ = ∣ ξ ∣ 2 + ∣ η ∣ 2 − 2 ξ ⋅ η + V ˉ ≥ V ˉ > 0. \int_0^t \int_0^t e^{i(r-r')\sigma}\,dr\,dr' = \frac{2(1-\cos(t\sigma))}{\sigma^2}, \qquad\sigma= |\xi|^2 + |\eta|^2 - 2\xi\cdot\eta+ \bar V \ge\bar V > 0. ∫ 0 t ∫ 0 t e i ( r − r ′ ) σ d r d r ′ = σ 2 2 ( 1 − cos ( t σ )) , σ = ∣ ξ ∣ 2 + ∣ η ∣ 2 − 2 ξ ⋅ η + V ˉ ≥ V ˉ > 0.
The positivity σ ≥ V ˉ \sigma\ge\bar V σ ≥ V ˉ is what localizes the spatial integral to the diagonal ξ ′ = 0 \xi' = 0 ξ ′ = 0 and closes the estimate. There is no smoothing to rely on. Because the propagator is unitary and the effective potential is just V ˉ = E [ V ] \bar V = \mathbb{E}[V] V ˉ = E [ V ] , no L 2 L^2 L 2 corrector appears in the expansion. The first Duhamel term already carries the full fluctuation, and the remainder is of higher order. This paper is the Schrödinger analogue of the Gloria–Neukamm–Otto quantitative program, with the spectral gap replaced by exponential mixing plus sub-Gaussian tails.
We recall the standard definitions. Let τ z \tau_z τ z denote the shift on Ω \Omega Ω , defined by
V ( τ z ω , y ) : = V ( ω , y + z ) . V(\tau_z\omega,y) := V(\omega,y+z). V ( τ z ω , y ) := V ( ω , y + z ) .
Definition 1.1 (Stationarity). The field V V V is stationary if the shift τ z \tau_z τ z preserves P \mathbb{P} P for every z ∈ R d z \in\mathbb{R}^d z ∈ R d , that is, for any k ≥ 1 k \ge1 k ≥ 1 , any y 1 , … , y k ∈ R d y_1,\ldots,y_k \in\mathbb{R}^d y 1 , … , y k ∈ R d , and any Borel set A ⊂ R k A \subset\mathbb{R}^k A ⊂ R k ,
P ( ( V ( y 1 ) , … , V ( y k ) ) ∈ A ) = P ( ( V ( y 1 + z ) , … , V ( y k + z ) ) ∈ A ) ∀ z ∈ R d . \mathbb{P}\big((V(y_1),\ldots,V(y_k)) \in A\big) = \mathbb{P}\big((V(y_1+z),\ldots,V(y_k+z)) \in A\big) \qquad\forall z \in\mathbb{R}^d. P ( ( V ( y 1 ) , … , V ( y k )) ∈ A ) = P ( ( V ( y 1 + z ) , … , V ( y k + z )) ∈ A ) ∀ z ∈ R d .
Definition 1.2 (Ergodicity). The stationary field V V V is ergodic if the shift group { τ z } z ∈ R d \{\tau_z\}_{z\in\mathbb{R}^d} { τ z } z ∈ R d is ergodic, that is, any event B ∈ F B \in\mathcal{F} B ∈ F satisfying τ z − 1 B = B \tau_z^{-1}B = B τ z − 1 B = B for all z ∈ R d z \in\mathbb{R}^d z ∈ R d has P ( B ) ∈ { 0 , 1 } \mathbb{P}(B) \in\{0,1\} P ( B ) ∈ { 0 , 1 } .
The only prior homogenization result for the dispersive Schrödinger equation with random stationary potentials is due to Zhang–Bal [24 ] . Their result relies on the spectral gap assumption.
Theorem 1.3 (Zhang–Bal [24 ] , Theorem 1.1). Let V V V be a bounded stationary ergodic field. Assume the centered field a ( y ) = V ( y ) − V ˉ a(y)=V(y)-\bar V a ( y ) = V ( y ) − V ˉ satisfies the spectral gap condition: there exist constants C , κ > 0 C,\kappa>0 C , κ > 0 such that for every f ∈ L 2 ( Ω ) f\in L^2(\Omega) f ∈ L 2 ( Ω ) measurable with respect to σ { a ( y ) : ∣ y − z ∣ ≤ R } \sigma\{a(y):|y-z|\le R\} σ { a ( y ) : ∣ y − z ∣ ≤ R } ,
E [ ( f − E [ f ] ) 2 ] ≤ C e − κ R E [ ( f − E [ f ] ) 2 ∣ σ { a ( y ) : ∣ y − z ∣ > 2 R } ] . \mathbb{E}\left[(f-\mathbb{E}[f])^2\right] \le Ce^{-\kappa R}\,\mathbb{E}\left[(f-\mathbb{E}[f])^2 \mid\sigma\{a(y):|y-z|>2R\}\right]. E [ ( f − E [ f ] ) 2 ] ≤ C e − κ R E [ ( f − E [ f ] ) 2 ∣ σ { a ( y ) : ∣ y − z ∣ > 2 R } ] . Then, for every δ > 0 \delta>0 δ > 0 ,
lim ε → 0 P ( sup 0 ≤ t ≤ T ∥ u ε ( t ) − u ˉ ( t ) ∥ L 2 ( R d ) > δ ) = 0. (8) \lim_{\varepsilon\to0}\mathbb{P}\left(\sup_{0\le t\le T}\|u_\varepsilon(t)-\bar u(t)\|_{L^2(\mathbb{R}^d)}>\delta\right)=0. \tag*{(8)} ε → 0 lim P ( 0 ≤ t ≤ T sup ∥ u ε ( t ) − u ˉ ( t ) ∥ L 2 ( R d ) > δ ) = 0. (8)
In this paper, we go beyond the qualitative convergence in probability above. We establish the optimal algebraic convergence rate, the almost sure strong convergence, and the fluctuation central limit theorem, under exponential mixing plus sub-Gaussian tails.
We now state the standing assumption.
Assumption 1.4. The centered stationary ergodic random field a ( y , ω ) : = V ( y , ω ) − V ˉ a(y,\omega):=V(y,\omega)-\bar V a ( y , ω ) := V ( y , ω ) − V ˉ , with V ˉ = E [ V ( 0 ) ] \bar V=\mathbb{E}[V(0)] V ˉ = E [ V ( 0 )] , satisfies:
(i) (Sub-Gaussian marginals) There exists C 0 > 0 C_0>0 C 0 > 0 such that
E [ e λ a ( y ) ] ≤ e 1 2 C 0 λ 2 ∀ y ∈ R d , λ ∈ R . \mathbb{E}\left[e^{\lambda a(y)}\right]\le e^{\frac{1}{2}C_0\lambda^2}\qquad\forall y\in\mathbb{R}^d,\ \lambda\in\mathbb{R}. E [ e λa ( y ) ] ≤ e 2 1 C 0 λ 2 ∀ y ∈ R d , λ ∈ R .
(ii) (Exponential mixing) There exist κ , δ > 0 \kappa,\delta>0 κ , δ > 0 such that for all R > 0 R>0 R > 0 ,
∣ P ( A ∩ B ) − P ( A ) P ( B ) ∣ ≤ δ e − κ R , (9) \left|\mathbb{P}(A\cap B)-\mathbb{P}(A)\mathbb{P}(B)\right|\le\delta e^{-\kappa R}, \tag*{(9)} ∣ P ( A ∩ B ) − P ( A ) P ( B ) ∣ ≤ δ e − κ R , (9)
whenever A ∈ σ ( a ( y ) : ∣ y ∣ ≤ R ) A\in\sigma(a(y):|y|\le R) A ∈ σ ( a ( y ) : ∣ y ∣ ≤ R ) and B ∈ σ ( a ( y ) : ∣ y ∣ ≥ 2 R ) B\in\sigma(a(y):|y|\ge2R) B ∈ σ ( a ( y ) : ∣ y ∣ ≥ 2 R ) .
Condition (i) gives uniform exponential integrability and closes the cumulant expansion, while condition (ii) controls the off-diagonal terms by exponential decay of correlations. Neither implies the other. The class of media covered by Assumption 1.4 strictly contains the Zhang–Bal spectral gap class [24 ] .
Four concrete examples illustrate the relationship between the two assumptions and why both are needed.
(a) Let W W W be white noise on R d \mathbb{R}^d R d . The stationary Gaussian field
V 1 ( y ) : = ∫ R d e i y ⋅ ξ ( 1 + ∣ ξ ∣ 2 ) − d + α 4 d W ( ξ ) , α > d , V_1(y):=\int_{\mathbb{R}^d}e^{\mathrm{i}y\cdot\xi}(1+|\xi|^2)^{-\frac{d+\alpha}{4}}\,\mathrm{d}W(\xi),\qquad\alpha>d, V 1 ( y ) := ∫ R d e i y ⋅ ξ ( 1 + ∣ ξ ∣ 2 ) − 4 d + α d W ( ξ ) , α > d ,
has covariance decaying as ∣ y ∣ − α |y|^{-\alpha} ∣ y ∣ − α , which is polynomially rather than exponentially decaying. It is sub-Gaussian but fails exponential mixing, and since α > d \alpha>d α > d the covariance is integrable and R ^ ( 0 ) < ∞ \widehat{R}(0)<\infty R ( 0 ) < ∞ . This shows that sub-Gaussianity alone does not give quantitative homogenization.
(b) Let { η k } k ∈ Z d \{\eta_k\}_{k\in\mathbb{Z}^d} { η k } k ∈ Z d be i.i.d. Cauchy variables and χ ∈ C c ∞ ( ( − 1 2 , 1 2 ) d ) \chi\in C^\infty_c\left(\left(-\frac{1}{2},\frac{1}{2}\right)^d\right) χ ∈ C c ∞ ( ( − 2 1 , 2 1 ) d ) a bump. The field
V 2 ( y ) : = ∑ k ∈ Z d η k χ ( y − k ) V_2(y):=\sum_{k\in\mathbb{Z}^d}\eta_k\chi(y-k) V 2 ( y ) := k ∈ Z d ∑ η k χ ( y − k )
is perfectly mixing (disjoint supports) but has no exponential moment. This shows that mixing alone does not control higher cumulants.
(c) The Gaussian field
V 3 ( y ) : = ∫ R d e i y ⋅ ξ ( 1 + ∣ ξ ∣ 2 ) − 1 2 d W ( ξ ) V_3(y):=\int_{\mathbb{R}^d}e^{\mathrm{i}y\cdot\xi}(1+|\xi|^2)^{-\frac{1}{2}}\,\mathrm{d}W(\xi) V 3 ( y ) := ∫ R d e i y ⋅ ξ ( 1 + ∣ ξ ∣ 2 ) − 2 1 d W ( ξ )
has covariance given by the Bessel kernel, decaying as e − c ∣ y ∣ e^{-c|y|} e − c ∣ y ∣ . It satisfies both assumptions and is the canonical model covered by our results, with R ^ ( 0 ) = ( 2 π ) d > 0 \widehat{R}(0)=(2\pi)^d>0 R ( 0 ) = ( 2 π ) d > 0 .
(d) Let η \eta η be Gaussian with exponentially decaying covariance and f = tanh f=\tanh f = tanh . The bounded non-Gaussian field
V 4 ( y ) : = tanh ( η ( y ) ) V_4(y):=\tanh(\eta(y)) V 4 ( y ) := tanh ( η ( y ))
satisfies both assumptions by Lipschitz transfer of mixing
∣ Cov ( g ( V 4 ) , h ( V 4 ) ) ∣ ≤ ∥ g ∥ ∞ ∥ h ∥ ∞ e − κ R . \left|\operatorname{Cov}(g(V_4),h(V_4))\right|\le\|g\|_\infty\|h\|_\infty e^{-\kappa R}. ∣ Cov ( g ( V 4 ) , h ( V 4 )) ∣ ≤ ∥ g ∥ ∞ ∥ h ∥ ∞ e − κ R .
This is the canonical non-Gaussian sub-Gaussian example.
We write a ( y ) = V ( y ) − V ‾ a(y)=V(y)-\overline{V} a ( y ) = V ( y ) − V for the centered field, and
R ( y ) : = E [ a ( y ) a ( 0 ) ] , R ^ ( ξ ) = ∫ R d e − i ξ ⋅ y R ( y ) d y . R(y):=\mathbb{E}[a(y)a(0)],\qquad\widehat{R}(\xi)=\int_{\mathbb{R}^d}e^{-\mathrm{i}\xi\cdot y}R(y)\,\mathrm{d}y. R ( y ) := E [ a ( y ) a ( 0 )] , R ( ξ ) = ∫ R d e − i ξ ⋅ y R ( y ) d y .
Assumption 1.4 implies R ^ ∈ L 1 ( R d ) \widehat{R}\in L^1(\mathbb{R}^d) R ∈ L 1 ( R d ) and R ^ ( 0 ) ≥ 0 \widehat{R}(0)\ge0 R ( 0 ) ≥ 0 .
Methods and results
The Schrödinger group generated by H 0 H_0 H 0 is
S ( t , r ) : = e − i H 0 ( t − r ) , F ( S ( t , r ) f ) ( ξ ) = e − i ( ∣ ξ ∣ 2 + V ‾ ) ( t − r ) f ^ ( ξ ) . S(t,r):=e^{-iH_0(t-r)},\qquad\mathcal{F}(S(t,r)f)(\xi)=e^{-i(|\xi|^2+\overline{V})(t-r)}\widehat{f}(\xi). S ( t , r ) := e − i H 0 ( t − r ) , F ( S ( t , r ) f ) ( ξ ) = e − i ( ∣ ξ ∣ 2 + V ) ( t − r ) f ( ξ ) .
Subtracting (LSE0 _0 0 ) from (LSEε _\varepsilon ε ), the fluctuation equation is
i ∂ t w ε = H 0 w ε + a ε ( u ‾ + w ε ) , w ε ( 0 ) = 0 , i\partial_t w_\varepsilon=H_0w_\varepsilon+a_\varepsilon(\overline{u}+w_\varepsilon),\qquad w_\varepsilon(0)=0, i ∂ t w ε = H 0 w ε + a ε ( u + w ε ) , w ε ( 0 ) = 0 ,
where a ε ( x ) : = a ( x ε ) a_\varepsilon(x):=a\left(\frac{x}{\varepsilon}\right) a ε ( x ) := a ( ε x ) . The Duhamel formula shows that
w ε ( t ) = − i ∫ 0 t S ( t , r ) a ε ( r ) ( u ‾ ( r ) + w ε ( r ) ) d r . w_\varepsilon(t)=-i\int_0^t S(t,r)a_\varepsilon(r)\bigl(\overline{u}(r)+w_\varepsilon(r)\bigr)\,dr. w ε ( t ) = − i ∫ 0 t S ( t , r ) a ε ( r ) ( u ( r ) + w ε ( r ) ) d r .
We split w ε = w ε ( 1 ) + w ε ( 2 ) w_\varepsilon=w_\varepsilon^{(1)}+w_\varepsilon^{(2)} w ε = w ε ( 1 ) + w ε ( 2 ) by the first Duhamel iteration,
w ε ( 1 ) ( t ) : = − i ∫ 0 t S ( t , r ) a ε ( r ) u ‾ ( r ) d r , w ε ( 2 ) ( t ) : = − i ∫ 0 t S ( t , r ) a ε ( r ) w ε ( r ) d r . w_\varepsilon^{(1)}(t):=-i\int_0^t S(t,r)a_\varepsilon(r)\overline{u}(r)\,dr,\qquad
w_\varepsilon^{(2)}(t):=-i\int_0^t S(t,r)a_\varepsilon(r)w_\varepsilon(r)\,dr. w ε ( 1 ) ( t ) := − i ∫ 0 t S ( t , r ) a ε ( r ) u ( r ) d r , w ε ( 2 ) ( t ) := − i ∫ 0 t S ( t , r ) a ε ( r ) w ε ( r ) d r .
For w ε ( 1 ) w_\varepsilon^{(1)} w ε ( 1 ) , the second moment is the double integral of R ^ ( ε η ) \widehat{R}(\varepsilon\eta) R ( ε η ) against the oscillatory time factor 2 ( 1 − cos ( t σ ) ) / σ 2 2(1-\cos(t\sigma))/\sigma^2 2 ( 1 − cos ( t σ )) / σ 2 and ∣ u ^ 0 ( ξ − η ) ∣ 2 |\widehat{u}_0(\xi-\eta)|^2 ∣ u 0 ( ξ − η ) ∣ 2 . After the change of variables ξ ′ = ε ( ξ − η ) \xi'=\varepsilon(\xi-\eta) ξ ′ = ε ( ξ − η ) , the remaining ξ ′ \xi' ξ ′ -integral is
∫ R d R ^ ( ξ ′ ) 2 ( 1 − cos ( t ϕ ) ) ϕ 2 d ξ ′ , ϕ : = 2 ζ ⋅ ξ ′ ε + ∣ ξ ′ ∣ 2 ε 2 . \int_{\mathbb{R}^d}\widehat{R}(\xi')\frac{2(1-\cos(t\phi))}{\phi^2}\,d\xi',
\qquad
\phi:=\frac{2\zeta\cdot\xi'}{\varepsilon}+\frac{|\xi'|^2}{\varepsilon^2}. ∫ R d R ( ξ ′ ) ϕ 2 2 ( 1 − cos ( tϕ )) d ξ ′ , ϕ := ε 2 ζ ⋅ ξ ′ + ε 2 ∣ ξ ′ ∣ 2 .
This integral is split into the near region ∣ ξ ′ ∣ ≤ ε |\xi'|\leq\varepsilon ∣ ξ ′ ∣ ≤ ε and the far region ∣ ξ ′ ∣ > ε |\xi'|>\varepsilon ∣ ξ ′ ∣ > ε (see Lemma 3.2 and Lemma 3.3). On the near region, the time factor is bounded by t 2 t^2 t 2 and the volume is O ( ε d ) O(\varepsilon^d) O ( ε d ) . On the far region, the quadratic term ∣ ξ ′ ∣ 2 ε 2 \frac{|\xi'|^2}{\varepsilon^2} ε 2 ∣ ξ ′ ∣ 2 dominates the phase, and a dyadic shell decomposition gives a geometrically convergent series for d ≤ 3 d\leq3 d ≤ 3 . Combining,
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≤ C T , u 0 ε d . (10) \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^2\right]\leq C_{T,u_0}\varepsilon^d. \tag*{(10)} E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≤ C T , u 0 ε d . (10)
The Fourier representation (31) isolates the near region ∣ ξ ′ ∣ ≤ ε |\xi'|\leq\varepsilon ∣ ξ ′ ∣ ≤ ε , which identifies the diagonal leading term and gives the matching lower bound and the fluctuation covariance.
For w ε ( 2 ) w_\varepsilon^{(2)} w ε ( 2 ) , the fluctuation w ε w_\varepsilon w ε inside the integral is itself expanded by the Duhamel formula, producing a second iteration. Repeating this self-similar substitution yields a hierarchy of iterates w ε ( k ) w_\varepsilon^{(k)} w ε ( k ) , k ≥ 2 k\geq2 k ≥ 2 , each carrying one additional power of the random potential. The same moment estimate as above gives
E [ ∥ w ε ( k ) ( t ) ∥ L x 2 2 ] ≤ C k ε k d , k ≥ 2 , \mathbb{E}\left[\|w_\varepsilon^{(k)}(t)\|_{L_x^2}^2\right]\leq C^k\varepsilon^{kd},\qquad k\geq2, E [ ∥ w ε ( k ) ( t ) ∥ L x 2 2 ] ≤ C k ε k d , k ≥ 2 ,
and summing the geometric series over k ≥ 2 k\geq2 k ≥ 2 yields
∥ w ε ( 2 ) ( t ) ∥ L 2 ( Ω ; L x 2 ) 2 ≤ C ε d + 1 . (11) \|w_\varepsilon^{(2)}(t)\|_{L^2(\Omega;L_x^2)}^2\leq C\varepsilon^{d+1}. \tag*{(11)} ∥ w ε ( 2 ) ( t ) ∥ L 2 ( Ω ; L x 2 ) 2 ≤ C ε d + 1 . (11)
Combining (10) and (11) and applying Gronwall to the Duhamel formula yields the first main result, the optimal algebraic convergence rate ε d / 2 \varepsilon^{d/2} ε d /2 in L 2 ( Ω ; L x 2 ) L^2(\Omega;L_x^2) L 2 ( Ω ; L x 2 ) .
Theorem 1.5. Under Assumption 1.4 , the following hold for every d ≥ 1 d\geq1 d ≥ 1 .
(i) For all 0 < ε ≤ 1 0<\varepsilon\leq1 0 < ε ≤ 1 ,
sup 0 ≤ t ≤ T ∥ u ε ( t ) − u ‾ ( t ) ∥ L 2 ( Ω ; L 2 ( R d ) ) ≤ C T , u 0 ε d / 2 . (12) \sup_{0\leq t\leq T}\|u_\varepsilon(t)-\overline{u}(t)\|_{L^2(\Omega;L^2(\mathbb{R}^d))}\leq C_{T,u_0}\varepsilon^{d/2}. \tag*{(12)} 0 ≤ t ≤ T sup ∥ u ε ( t ) − u ( t ) ∥ L 2 ( Ω ; L 2 ( R d )) ≤ C T , u 0 ε d /2 . (12) If additionally R ^ ( 0 ) > 0 \widehat{R}(0)>0 R ( 0 ) > 0 and u ^ 0 ( 0 ) ≠ 0 \widehat{u}_0(0)\neq0 u 0 ( 0 ) = 0 , then there exist c t , u 0 > 0 c_{t,u_0}>0 c t , u 0 > 0 and ε 0 > 0 \varepsilon_0>0 ε 0 > 0 such that
E [ ∥ u ε ( t ) − u ‾ ( t ) ∥ L 2 ( R d ) 2 ] ≥ c t , u 0 ε d for all 0 < ε ≤ ε 0 . (13) \mathbb{E}\left[\|u_\varepsilon(t)-\overline{u}(t)\|_{L^2(\mathbb{R}^d)}^2\right]\geq c_{t,u_0}\varepsilon^d\qquad\text{for all }0<\varepsilon\leq\varepsilon_0. \tag*{(13)} E [ ∥ u ε ( t ) − u ( t ) ∥ L 2 ( R d ) 2 ] ≥ c t , u 0 ε d for all 0 < ε ≤ ε 0 . (13) The rate ε d / 2 \varepsilon^{d/2} ε d /2 is sharp.
(ii) For every finite m ≥ 1 m \ge1 m ≥ 1 ,
E [ sup 0 ≤ t ≤ T ∥ u ε ( t ) − u ˉ ( t ) ∥ L 2 ( R d ) m ] ≤ C m , T , u 0 ε m d 2 . (14) \mathbb{E}\left[\sup_{0\le t\le T}\|u_\varepsilon(t)-\bar{u}(t)\|_{L^2(\mathbb{R}^d)}^m\right]\le C_{m,T,u_0}\varepsilon^{\frac{md}{2}}. \tag*{(14)} E [ 0 ≤ t ≤ T sup ∥ u ε ( t ) − u ˉ ( t ) ∥ L 2 ( R d ) m ] ≤ C m , T , u 0 ε 2 m d . (14)
Remark 1.6. The lower bound (13) requires R ^ ( 0 ) > 0 \widehat{R}(0)>0 R ( 0 ) > 0 . If R ^ ( 0 ) = 0 \widehat{R}(0)=0 R ( 0 ) = 0 , the block-average variance is governed by the next spectral moment, the ε d 2 \varepsilon^{\frac{d}{2}} ε 2 d fluctuations are suppressed, and the rate and CLT scale must be recomputed from the local regularity of R ^ \widehat{R} R near the origin. We do not treat this case.
Remark 1.7. The compact support of u ^ 0 \widehat{u}_0 u 0 is technical and localizes the ζ \zeta ζ -integral only. For d ≤ 3 d\le3 d ≤ 3 , the constants depend only on ∥ u 0 ∥ L 2 \|u_0\|_{L^2} ∥ u 0 ∥ L 2 , so the estimates extend to u 0 ∈ L 2 ( R d ) u_0\in L^2(\mathbb{R}^d) u 0 ∈ L 2 ( R d ) by density directly. For d ≥ 4 d\ge4 d ≥ 4 , the physical space bound of Lemma A.1 is an operator norm estimate uniform in ε \varepsilon ε , and the ε d \varepsilon^d ε d prefactor absorbs the Schwartz norm dependence. Approximating u 0 ∈ L 2 u_0\in L^2 u 0 ∈ L 2 by C c ∞ C^\infty_c C c ∞ converges in L 2 ( Ω ; L x 2 ) L^2(\Omega;L^2_x) L 2 ( Ω ; L x 2 ) by the unitarity of S ( t , r ) S(t,r) S ( t , r ) .
Remark 1.8. The sub-Gaussian assumption 1.4(i) is used at two levels: the L 2 L^2 L 2 rate requires only a moment of order p > 2 p>2 p > 2 together with exponential mixing, while the fluctuation CLT and the all moment estimate require the full sub-Gaussian proxy to close the cumulant expansion and apply Lemma 2.13.
The rate estimate above is in L 2 ( Ω ; L x 2 ) L^2(\Omega;L^2_x) L 2 ( Ω ; L x 2 ) . To upgrade it to almost sure convergence, we apply the dyadic Borel–Cantelli argument. At the dyadic sequence ε n = 2 − n \varepsilon_n=2^{-n} ε n = 2 − n ,
E [ sup 0 ≤ t ≤ T ∥ w ε n ( t ) ∥ L x 2 2 q ] ≤ C ε n d q . \mathbb{E}\left[\sup_{0\le t\le T}\|w_{\varepsilon_n}(t)\|_{L^2_x}^{2q}\right]\le C\varepsilon_n^{dq}. E [ 0 ≤ t ≤ T sup ∥ w ε n ( t ) ∥ L x 2 2 q ] ≤ C ε n d q .
By Markov’s inequality,
P ( sup t ∥ w ε n ( t ) ∥ > ε n d 2 − η ) ≤ C ε n d q η , \mathbb{P}\left(\sup_t\|w_{\varepsilon_n}(t)\|>\varepsilon_n^{\frac{d}{2}-\eta}\right)\le C\varepsilon_n^{dq\eta}, P ( t sup ∥ w ε n ( t ) ∥ > ε n 2 d − η ) ≤ C ε n d q η ,
summable for q q q large. The time increment satisfies
E [ ∥ w ε ( 1 ) ( t ) − w ε ( 1 ) ( s ) ∥ L x 2 2 q ] ≤ C ∣ t − s ∣ 2 q , \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)-w_\varepsilon^{(1)}(s)\|_{L^2_x}^{2q}\right]\le C|t-s|^{2q}, E [ ∥ w ε ( 1 ) ( t ) − w ε ( 1 ) ( s ) ∥ L x 2 2 q ] ≤ C ∣ t − s ∣ 2 q ,
so Kolmogorov–Chentsov gives equicontinuity in C ( [ 0 , T ] ; L x 2 ) C([0,T];L^2_x) C ([ 0 , T ] ; L x 2 ) uniformly in ε \varepsilon ε , filling the gaps between dyadic points. This yields the almost sure strong convergence.
Theorem 1.9. Under Assumption 1.4, for every d ≥ 1 d\ge1 d ≥ 1 , every fixed T > 0 T>0 T > 0 and u 0 ∈ L 2 ( R d ) u_0\in L^2(\mathbb{R}^d) u 0 ∈ L 2 ( R d ) ,
lim ε → 0 sup 0 ≤ t ≤ T ∥ u ε ( t ) − u ˉ ( t ) ∥ L 2 ( R d ) = 0 a.s. (15) \lim_{\varepsilon\to0}\sup_{0\le t\le T}\|u_\varepsilon(t)-\bar{u}(t)\|_{L^2(\mathbb{R}^d)}=0 \qquad\text{a.s.} \tag*{(15)} ε → 0 lim 0 ≤ t ≤ T sup ∥ u ε ( t ) − u ˉ ( t ) ∥ L 2 ( R d ) = 0 a.s. (15)
Beyond the convergence, we study the limiting distribution of the rescaled fluctuation. Define
F ε ( t , x ) : = ε − d 2 w ε ( t , x ) . \mathcal{F}_\varepsilon(t,x):=\varepsilon^{-\frac{d}{2}}w_\varepsilon(t,x). F ε ( t , x ) := ε − 2 d w ε ( t , x ) .
Partition space into blocks of side L ε = 2 N ( − log ε ) κ L_\varepsilon=\frac{2N(-\log\varepsilon)}{\kappa} L ε = κ 2 N ( − l o g ε ) :
F ε ( t ) = ∑ j = 1 N ε M j ε ( t ) + ρ ε ( t ) , \mathcal{F}_\varepsilon(t)=\sum_{j=1}^{N_\varepsilon}M_j^\varepsilon(t)+\rho_\varepsilon(t), F ε ( t ) = j = 1 ∑ N ε M j ε ( t ) + ρ ε ( t ) ,
where the remainder satisfies
E [ ∥ ρ ε ( t ) ∥ L x 2 2 ] ≤ C L ε − d → 0. \mathbb{E}\left[\|\rho_\varepsilon(t)\|_{L^2_x}^2\right]\le CL_\varepsilon^{-d}\to0. E [ ∥ ρ ε ( t ) ∥ L x 2 2 ] ≤ C L ε − d → 0.
The block array { M j ε } \{M_j^\varepsilon\} { M j ε } is an approximate martingale by exponential mixing beyond distance L ε L_\varepsilon L ε . The Lindeberg condition (see Lemma 2.3 below) follows from
∑ j E [ ∣ M j ε ∣ 4 ] ≤ C L ε − d → 0. \sum_j\mathbb{E}\left[|M_j^\varepsilon|^4\right]\le CL_\varepsilon^{-d}\to0. j ∑ E [ ∣ M j ε ∣ 4 ] ≤ C L ε − d → 0.
Tightness follows from
E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 2 ] ≤ C ∣ t − s ∣ , \mathbb{E}\left[\|\mathcal{F}_{\varepsilon}(t)-\mathcal{F}_{\varepsilon}(s)\|_{L_x^2}^2\right] \le C|t-s|, E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 2 ] ≤ C ∣ t − s ∣ ,
and, by sub-Gaussian concentration (Lemma 2.13 ),
E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 4 ] ≤ C ∣ t − s ∣ 2 . \mathbb{E}\left[\|\mathcal{F}_{\varepsilon}(t)-\mathcal{F}_{\varepsilon}(s)\|_{L_x^2}^4\right] \le C|t-s|^2. E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 4 ] ≤ C ∣ t − s ∣ 2 .
The covariance converges to
lim ε → 0 E [ ⟨ F ε ( t ) , φ ⟩ ⟨ F ε ( s ) , ψ ⟩ ‾ ] = ∫ 0 t ∧ s ⟨ S ( t , r ) ∗ φ , R ^ ( 0 ) S ( s , r ) ∗ ψ ⟩ d r . \lim_{\varepsilon\to0}\mathbb{E}\left[\langle\mathcal{F}_{\varepsilon}(t),\varphi\rangle\overline{\langle\mathcal{F}_{\varepsilon}(s),\psi\rangle}\right]=\int_0^{t\wedge s}\langle S(t,r)^*\varphi,\widehat{R}(0)S(s,r)^*\psi\rangle\,dr. ε → 0 lim E [ ⟨ F ε ( t ) , φ ⟩ ⟨ F ε ( s ) , ψ ⟩ ] = ∫ 0 t ∧ s ⟨ S ( t , r ) ∗ φ , R ( 0 ) S ( s , r ) ∗ ψ ⟩ d r .
This gives our final main result: the Gaussian fluctuation limit.
Theorem 1.10. Assume d ≥ 1 d\ge1 d ≥ 1 . As ε → 0 \varepsilon\to0 ε → 0 , the family { F ε } ε > 0 \{\mathcal{F}_{\varepsilon}\}_{\varepsilon>0} { F ε } ε > 0 converges in distribution in C ( [ 0 , T ] ; L 2 ( R d ) ) C([0,T];L^2(\mathbb{R}^d)) C ([ 0 , T ] ; L 2 ( R d )) to a centered Gaussian field G \mathcal{G} G with covariance
Cov ( ⟨ G ( t ) , φ ⟩ , ⟨ G ( s ) , ψ ⟩ ) = ∫ 0 t ∧ s ⟨ S ( t , r ) ∗ φ , R ^ ( 0 ) S ( s , r ) ∗ ψ ⟩ L 2 d r . (16) \operatorname{Cov}(\langle\mathcal{G}(t),\varphi\rangle,\langle\mathcal{G}(s),\psi\rangle)=\int_0^{t\wedge s}\langle S(t,r)^*\varphi,\widehat{R}(0)S(s,r)^*\psi\rangle_{L^2}\,dr. \tag*{(16)} Cov (⟨ G ( t ) , φ ⟩ , ⟨ G ( s ) , ψ ⟩) = ∫ 0 t ∧ s ⟨ S ( t , r ) ∗ φ , R ( 0 ) S ( s , r ) ∗ ψ ⟩ L 2 d r . (16)
Remark 1.11. The covariance (1.16) is exactly that of the mild solution to (SPDE). The limit G \mathcal{G} G is the Gaussian field governed by the linear random Schrödinger equation with additive white noise, the dispersive analogue of [13 ] . The noise is interpreted through the mild formulation
G ( t ) = − i R ^ ( 0 ) ∫ 0 t S ( t , r ) d W ( r ) , \mathcal{G}(t)=-i\sqrt{\widehat{R}(0)}\int_0^t S(t,r)\,dW(r), G ( t ) = − i R ( 0 ) ∫ 0 t S ( t , r ) d W ( r ) ,
where W W W is an L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) -valued cylindrical Brownian motion. The Itô isometry and the unitarity of S ( t , r ) S(t,r) S ( t , r ) ensure that G ( t ) \mathcal{G}(t) G ( t ) is a well-defined L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) -valued Gaussian process for all d ≥ 1 d\ge1 d ≥ 1 , without spatial smoothing of the noise.
The rest of the paper is organized as follows. Section 2 collects the analytic tools and the probabilistic lemmas used throughout. Section 3 develops the Duhamel decomposition: the second moment estimates in Subsection 3.1, the second order correction and sharpness in Subsection 3.2, and the all moment estimates in Subsection 3.3, which together prove Theorem 1.5. Section 4 proves the almost sure convergence, Theorem 1.9. Section 5 proves the fluctuation central limit theorem, Theorem 1.10. The Fourier near/far second moment bound in Subsection 3.1 is written for d ≤ 3 d\le3 d ≤ 3 . Appendix A supplies the physical space oscillatory integral argument that extends the same upper bound to every d ≥ 4 d\ge4 d ≥ 4 , and the remaining estimates are dimension independent.
Preliminaries
We write F ( g ) ( ξ ) \mathcal{F}(g)(\xi) F ( g ) ( ξ ) for the Fourier transform of a compound expression g g g , while g ^ \widehat{g} g denotes the Fourier transform of a single symbol. We use the convention
F ( f ) ( ξ ) = ∫ R d e − i x ⋅ ξ f ( x ) d x , f ( x ) = ( 2 π ) − d ∫ R d e i x ⋅ ξ F ( f ) ( ξ ) d ξ , \mathcal{F}(f)(\xi)=\int_{\mathbb{R}^d}e^{-ix\cdot\xi}f(x)\,dx,\qquad f(x)=(2\pi)^{-d}\int_{\mathbb{R}^d}e^{ix\cdot\xi}\mathcal{F}(f)(\xi)\,d\xi, F ( f ) ( ξ ) = ∫ R d e − i x ⋅ ξ f ( x ) d x , f ( x ) = ( 2 π ) − d ∫ R d e i x ⋅ ξ F ( f ) ( ξ ) d ξ ,
so Plancherel identity is
∥ f ∥ L 2 2 = ( 2 π ) − d ∥ F ( f ) ∥ L 2 2 . \|f\|_{L^2}^2=(2\pi)^{-d}\|\mathcal{F}(f)\|_{L^2}^2. ∥ f ∥ L 2 2 = ( 2 π ) − d ∥ F ( f ) ∥ L 2 2 .
The Schrödinger group generated by H 0 = − Δ + V ‾ H_0=-\Delta+\overline{V} H 0 = − Δ + V is
F ( S ( t , r ) f ) ( ξ ) = e − i ( ∣ ξ ∣ 2 + V ‾ ) ( t − r ) f ^ ( ξ ) , (17) \mathcal{F}(S(t,r)f)(\xi)=e^{-i(|\xi|^2+\overline{V})(t-r)}\widehat{f}(\xi), \qquad \tag*{(17)} F ( S ( t , r ) f ) ( ξ ) = e − i ( ∣ ξ ∣ 2 + V ) ( t − r ) f ( ξ ) , (17)
and its kernel is the Fresnel kernel
K t ( x ) = ( 4 π i t ) − d 2 e i ∣ x ∣ 2 4 t , t ≠ 0. (18) K_t(x)=(4\pi i t)^{-\frac{d}{2}}e^{i\frac{|x|^2}{4t}},\qquad t\ne0. \tag*{(18)} K t ( x ) = ( 4 π i t ) − 2 d e i 4 t ∣ x ∣ 2 , t = 0. (18)
The Schrödinger group S ( t , r ) = e − i H 0 ( t − r ) S(t,r)=e^{-iH_0(t-r)} S ( t , r ) = e − i H 0 ( t − r ) is unitary on L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) [[22 Chapter VIII] ], that is,
∥ S ( t , r ) f ∥ L x 2 = ∥ f ∥ L x 2 . \|S(t,r)f\|_{L_x^2}=\|f\|_{L_x^2}. ∥ S ( t , r ) f ∥ L x 2 = ∥ f ∥ L x 2 .
for all t , r ∈ R t,r\in\mathbb{R} t , r ∈ R and f ∈ L 2 ( R d ) f\in L^2(\mathbb{R}^d) f ∈ L 2 ( R d ) .
Lemma 2.1 (Gronwall’s inequality, [7 Chapter 1] ). If f : [ 0 , T ] → R + f:[0,T]\to\mathbb{R}_{+} f : [ 0 , T ] → R + is continuous and
f ( t ) ≤ A ( t ) + ∫ 0 t B ( s ) f ( s ) d s . f(t)\le A(t)+\int_{0}^{t}B(s)f(s)\,\mathrm{d}s. f ( t ) ≤ A ( t ) + ∫ 0 t B ( s ) f ( s ) d s . then,
f ( t ) ≤ A ( t ) + ∫ 0 t A ( s ) B ( s ) exp ( ∫ s t B ( r ) d r ) d s . f(t)\le A(t)+\int_{0}^{t}A(s)B(s)\exp\left(\int_{s}^{t}B(r)\,\mathrm{d}r\right)\,\mathrm{d}s. f ( t ) ≤ A ( t ) + ∫ 0 t A ( s ) B ( s ) exp ( ∫ s t B ( r ) d r ) d s . In particular, if A ( t ) ≤ A 0 A(t)\le A_{0} A ( t ) ≤ A 0 and B ( s ) ≤ B 0 B(s)\le B_{0} B ( s ) ≤ B 0 , then f ( t ) ≤ A 0 e B 0 t f(t)\le A_{0}e^{B_{0}t} f ( t ) ≤ A 0 e B 0 t .
Martingale limit theorems
Lemma 2.2 (Burkholder–Davis–Gundy, [23 Theorem 1.1] ). Let M t M_t M t be a continuous martingale and p ≥ 1 p\ge1 p ≥ 1 . There exists a constant C p C_p C p depending only on p p p such that
E [ sup 0 ≤ t ≤ T ∣ M t ∣ p ] ≤ C p E [ ⟨ M ⟩ T p / 2 ] . \mathbb{E}\left[\sup_{0\le t\le T}|M_t|^p\right]\le C_p\mathbb{E}\left[\langle M\rangle_T^{p/2}\right]. E [ 0 ≤ t ≤ T sup ∣ M t ∣ p ] ≤ C p E [ ⟨ M ⟩ T p /2 ] .
Lemma 2.3 (Martingale central limit theorem, [15 Theorem 3.2] ). Let { M j ε } \{M_j^\varepsilon\} { M j ε } be a triangular array of martingale differences satisfying the Lindeberg condition, namely
∑ j E [ ( M j ε ) 2 1 { ∣ M j ε ∣ > η } ] → 0 for every η > 0 , (19) \sum_j\mathbb{E}\left[(M_j^\varepsilon)^2\mathbf{1}_{\{|M_j^\varepsilon|>\eta\}}\right]\to0\qquad\text{for every }\eta>0, \tag*{(19)} j ∑ E [ ( M j ε ) 2 1 { ∣ M j ε ∣ > η } ] → 0 for every η > 0 , (19) and
∑ j E [ ( M j ε ) 2 ] → σ 2 . \sum_j\mathbb{E}\left[(M_j^\varepsilon)^2\right]\to\sigma^2. j ∑ E [ ( M j ε ) 2 ] → σ 2 . Then ∑ j M j ε \sum_j M_j^\varepsilon ∑ j M j ε converges in distribution to a centered Gaussian random variable with variance σ 2 \sigma^2 σ 2 .
Lemma 2.4 (Lévy continuity theorem, [6 Theorem 3.3] ). Let { μ n } \{\mu_n\} { μ n } be probability measures on R k \mathbb{R}^k R k with characteristic functions ϕ n ( θ ) = ∫ R k e i θ ⋅ x d μ n ( x ) \phi_n(\theta)=\int_{\mathbb{R}^k}e^{i\theta\cdot x}\,\mathrm{d}\mu_n(x) ϕ n ( θ ) = ∫ R k e i θ ⋅ x d μ n ( x ) . If ϕ n ( θ ) \phi_n(\theta) ϕ n ( θ ) converges pointwise to a function ϕ ( θ ) \phi(\theta) ϕ ( θ ) that is continuous at θ = 0 \theta=0 θ = 0 , then ϕ \phi ϕ is the characteristic function of a probability measure μ \mu μ on R k \mathbb{R}^k R k , and μ n \mu_n μ n converges weakly to μ \mu μ .
Lemma 2.5 (Slutsky, [6 Theorem 3.1] ). If X ε X_\varepsilon X ε converges in distribution to X X X and Y ε Y_\varepsilon Y ε converges in probability to a constant c c c , then ( X ε , Y ε ) (X_\varepsilon,Y_\varepsilon) ( X ε , Y ε ) converges in distribution to ( X , c ) (X,c) ( X , c ) . In particular, if ρ ε → 0 \rho_\varepsilon\to0 ρ ε → 0 in probability, then X ε + ρ ε X_\varepsilon+\rho_\varepsilon X ε + ρ ε converges in distribution to X X X .
Corollary 2.6. Let { M j ε ∈ R k } \{M_j^\varepsilon\in\mathbb{R}^k\} { M j ε ∈ R k } be a k k k -dimensional triangular array of martingale differences adapted to the same filtration { G j } \{G_j\} { G j } . Assume that in every component l = 1 , … , k l=1,\ldots,k l = 1 , … , k ,
lim ε → 0 ∑ j E [ ( M j , l ε ) 2 1 { ∣ M j , l ε ∣ > η } ] = 0 for every η > 0 , (20) \lim_{\varepsilon\to0}\sum_j\mathbb{E}\left[(M_{j,l}^\varepsilon)^2\mathbf{1}_{\{|M_{j,l}^\varepsilon|>\eta\}}\right]=0\qquad\text{for every }\eta>0, \tag*{(20)} ε → 0 lim j ∑ E [ ( M j , l ε ) 2 1 { ∣ M j , l ε ∣ > η } ] = 0 for every η > 0 , (20) and the covariance matrix converges entrywise:
lim ε → 0 ∑ j E [ M j , l ε M j , m ε ] = Σ l m for every l , m = 1 , … , k . (21) \lim_{\varepsilon\to0}\sum_j\mathbb{E}\left[M_{j,l}^\varepsilon M_{j,m}^\varepsilon\right]=\Sigma_{lm}\qquad\text{for every }l,m=1,\ldots,k. \tag*{(21)} ε → 0 lim j ∑ E [ M j , l ε M j , m ε ] = Σ l m for every l , m = 1 , … , k . (21) Then ∑ j M j ε \sum_j M_j^\varepsilon ∑ j M j ε converges in distribution on R k \mathbb{R}^k R k to the centered Gaussian measure N ( 0 , Σ ) \mathcal{N}(0,\Sigma) N ( 0 , Σ ) .
Proof. Fix θ ∈ R k \theta\in\mathbb{R}^k θ ∈ R k and define
X j ε : = θ ⋅ M j ε = ∑ l = 1 k θ l M j , l ε . X_j^\varepsilon:=\theta\cdot M_j^\varepsilon=\sum_{l=1}^{k}\theta_lM_{j,l}^\varepsilon. X j ε := θ ⋅ M j ε = l = 1 ∑ k θ l M j , l ε .
Since each M j , l ε M_{j,l}^{\varepsilon} M j , l ε is G j \mathcal{G}_j G j -measurable and E [ M j , l ε ∣ G j − 1 ] = 0 \mathbb{E}[M_{j,l}^{\varepsilon}\mid\mathcal{G}_{j-1}]=0 E [ M j , l ε ∣ G j − 1 ] = 0 , taking the linear combination gives
E [ X j ε ∣ G j − 1 ] = ∑ l = 1 k θ l E [ M j , l ε ∣ G j − 1 ] = 0. \mathbb{E}[X_j^{\varepsilon}\mid\mathcal{G}_{j-1}]=\sum_{l=1}^{k}\theta_l\mathbb{E}[M_{j,l}^{\varepsilon}\mid\mathcal{G}_{j-1}]=0. E [ X j ε ∣ G j − 1 ] = l = 1 ∑ k θ l E [ M j , l ε ∣ G j − 1 ] = 0.
For any η > 0 \eta>0 η > 0 , the bound
∣ X j ε ∣ ≤ ∑ l = 1 k ∣ θ l ∣ ∣ M j , l ε ∣ ≤ ∣ θ ∣ max 1 ≤ l ≤ k ∣ M j , l ε ∣ |X_j^{\varepsilon}|\leq\sum_{l=1}^{k}|\theta_l||M_{j,l}^{\varepsilon}|\leq|\theta|\max_{1\leq l\leq k}|M_{j,l}^{\varepsilon}| ∣ X j ε ∣ ≤ l = 1 ∑ k ∣ θ l ∣∣ M j , l ε ∣ ≤ ∣ θ ∣ 1 ≤ l ≤ k max ∣ M j , l ε ∣
implies the event containment
{ ∣ X j ε ∣ > η } ⊂ ⋃ l = 1 k { ∣ M j , l ε ∣ > η ∣ θ ∣ } . (22) \{|X_j^{\varepsilon}|>\eta\}\subset\bigcup_{l=1}^{k}\left\{|M_{j,l}^{\varepsilon}|>\frac{\eta}{|\theta|}\right\}. \tag*{(22)} { ∣ X j ε ∣ > η } ⊂ l = 1 ⋃ k { ∣ M j , l ε ∣ > ∣ θ ∣ η } . (22)
for θ ≠ 0 \theta\neq0 θ = 0 . The case θ = 0 \theta=0 θ = 0 is trivial.
By the union bound and (2.4) applied with threshold η ∣ θ ∣ \frac{\eta}{|\theta|} ∣ θ ∣ η ,
∑ j E [ ( X j ε ) 2 1 { ∣ X j ε ∣ > η } ] ≤ k ∣ θ ∣ 2 ∑ l = 1 k ∑ j E [ ( M j , l ε ) 2 1 { ∣ M j , l ε ∣ > η ∣ θ ∣ } ] ⟶ 0 \sum_j\mathbb{E}\left[(X_j^{\varepsilon})^2\mathbf{1}_{\{|X_j^{\varepsilon}|>\eta\}}\right]\leq k|\theta|^2\sum_{l=1}^{k}\sum_j\mathbb{E}\left[(M_{j,l}^{\varepsilon})^2\mathbf{1}_{\{|M_{j,l}^{\varepsilon}|>\frac{\eta}{|\theta|}\}}\right]\longrightarrow0 j ∑ E [ ( X j ε ) 2 1 { ∣ X j ε ∣ > η } ] ≤ k ∣ θ ∣ 2 l = 1 ∑ k j ∑ E [ ( M j , l ε ) 2 1 { ∣ M j , l ε ∣ > ∣ θ ∣ η } ] ⟶ 0
as ε → 0 \varepsilon\to0 ε → 0 . The scalar variance expands as
∑ j E [ ( X j ε ) 2 ] = ∑ l , m = 1 k θ l θ m ∑ j E [ M j , l ε M j , m ε ] . \sum_j\mathbb{E}[(X_j^{\varepsilon})^2]=\sum_{l,m=1}^{k}\theta_l\theta_m\sum_j\mathbb{E}[M_{j,l}^{\varepsilon}M_{j,m}^{\varepsilon}]. j ∑ E [( X j ε ) 2 ] = l , m = 1 ∑ k θ l θ m j ∑ E [ M j , l ε M j , m ε ] .
By (2.5), for each l , m l,m l , m , the sum converges to Σ l m \Sigma_{lm} Σ l m . Therefore,
lim ε → 0 ∑ j E [ ( X j ε ) 2 ] = ∑ l , m = 1 k θ l θ m Σ l m = θ ⊤ Σ θ . \lim_{\varepsilon\to0}\sum_j\mathbb{E}[(X_j^{\varepsilon})^2]=\sum_{l,m=1}^{k}\theta_l\theta_m\Sigma_{lm}=\theta^{\top}\Sigma\theta. ε → 0 lim j ∑ E [( X j ε ) 2 ] = l , m = 1 ∑ k θ l θ m Σ l m = θ ⊤ Σ θ .
The scalar array { X j ε } \{X_j^{\varepsilon}\} { X j ε } satisfies the Lindeberg condition (2.3) and its variance converges to θ ⊤ Σ θ \theta^{\top}\Sigma\theta θ ⊤ Σ θ . By Lemma 2.3, for every bounded continuous f : R → R f:\mathbb{R}\to\mathbb{R} f : R → R ,
lim ε → 0 E [ f ( ∑ j X j ε ) ] = 1 2 π θ ⊤ Σ θ ∫ R f ( x ) e − x 2 2 θ ⊤ Σ θ d x . \lim_{\varepsilon\to0}\mathbb{E}\left[f\left(\sum_jX_j^{\varepsilon}\right)\right]=\frac{1}{\sqrt{2\pi\theta^{\top}\Sigma\theta}}\int_{\mathbb{R}}f(x)e^{-\frac{x^2}{2\theta^{\top}\Sigma\theta}}\,dx. ε → 0 lim E [ f ( j ∑ X j ε ) ] = 2 π θ ⊤ Σ θ 1 ∫ R f ( x ) e − 2 θ ⊤ Σ θ x 2 d x .
Equivalently, the characteristic function satisfies
lim ε → 0 E [ e i t θ ⋅ ∑ j M j ε ] = e − t 2 2 θ ⊤ Σ θ for every t ∈ R , θ ∈ R k . \lim_{\varepsilon\to0}\mathbb{E}\left[e^{it\theta\cdot\sum_jM_j^{\varepsilon}}\right]=e^{-\frac{t^2}{2}\theta^{\top}\Sigma\theta}\qquad\text{for every }t\in\mathbb{R},\ \theta\in\mathbb{R}^k. ε → 0 lim E [ e i tθ ⋅ ∑ j M j ε ] = e − 2 t 2 θ ⊤ Σ θ for every t ∈ R , θ ∈ R k .
Taking t = 1 t=1 t = 1 , this is the characteristic function of N ( 0 , Σ ) \mathcal{N}(0,\Sigma) N ( 0 , Σ ) evaluated at θ \theta θ . Since this holds for every θ ∈ R k \theta\in\mathbb{R}^k θ ∈ R k , the family of distributions of ∑ j M j ε \sum_jM_j^{\varepsilon} ∑ j M j ε has characteristic functions converging pointwise to the characteristic function of N ( 0 , Σ ) \mathcal{N}(0,\Sigma) N ( 0 , Σ ) , which is continuous at θ = 0 \theta=0 θ = 0 . By the Lévy continuity theorem (Lemma 2.4), ∑ j M j ε \sum_jM_j^{\varepsilon} ∑ j M j ε converges in distribution on R k \mathbb{R}^k R k to N ( 0 , Σ ) \mathcal{N}(0,\Sigma) N ( 0 , Σ ) . □ \square □
Tightness and moment bounds.
Lemma 2.7 (Mean ergodic theorem, [23 Theorem 2.3] ). *For any f ∈ L 1 ( Ω ; R ) f\in L^1(\Omega;\mathbb{R}) f ∈ L 1 ( Ω ; R ) ,
lim λ → ∞ 1 λ d ∫ [ 0 , λ ] d f ( τ y ω ) d y = E [ f ] a.s. \lim_{\lambda\to\infty}\frac{1}{\lambda^d}\int_{[0,\lambda]^d}f(\tau_y\omega)\,dy=\mathbb{E}[f]\qquad\text{a.s.} λ → ∞ lim λ d 1 ∫ [ 0 , λ ] d f ( τ y ω ) d y = E [ f ] a.s.
Lemma 2.8 (Borel–Cantelli lemma, [6 Chapter 12] ). If
∑ n = 1 ∞ P ( A n ) < ∞ , \sum_{n=1}^{\infty} \mathbb{P}(A_n) < \infty, n = 1 ∑ ∞ P ( A n ) < ∞ , then,
P ( lim sup n → ∞ A n ) = 0. \mathbb{P}\left(\limsup_{n\to\infty} A_n\right)=0. P ( n → ∞ lim sup A n ) = 0.
Lemma 2.9 (Isserlis’ theorem, [6 Chapter 1] ). If X 1 , … , X 2 q X_1,\ldots,X_{2q} X 1 , … , X 2 q are jointly centered Gaussian random variables, then,
E [ ∏ j = 1 2 q X j ] = ∑ π ∈ P ( 2 q ) ∏ { j , k } ∈ π E [ X j X k ] , \mathbb{E}\left[\prod_{j=1}^{2q}X_j\right]
=
\sum_{\pi\in\mathcal{P}(2q)}
\prod_{\{j,k\}\in\pi}\mathbb{E}[X_jX_k], E [ j = 1 ∏ 2 q X j ] = π ∈ P ( 2 q ) ∑ { j , k } ∈ π ∏ E [ X j X k ] , where P ( 2 q ) \mathcal{P}(2q) P ( 2 q ) is the set of all pairings of { 1 , … , 2 q } \{1,\ldots,2q\} { 1 , … , 2 q } .
Lemma 2.10 (Kolmogorov–Chentsov compactness criterion, [23 Chapter 3] ). Let { X ε ( t ) } ε > 0 \{X_\varepsilon(t)\}_{\varepsilon>0} { X ε ( t ) } ε > 0 be a family of processes taking values in a Banach space B B B . If there exist constants p > 0 p>0 p > 0 , α > 1 p \alpha>\frac{1}{p} α > p 1 , and C > 0 C>0 C > 0 such that
E [ ∥ X ε ( t ) − X ε ( s ) ∥ B p ] ≤ C ∣ t − s ∣ α \mathbb{E}\left[\lVert X_\varepsilon(t)-X_\varepsilon(s)\rVert_B^p\right]\le C|t-s|^\alpha E [ ∥ X ε ( t ) − X ε ( s ) ∥ B p ] ≤ C ∣ t − s ∣ α uniformly in ε \varepsilon ε , then the family { X ε } \{X_\varepsilon\} { X ε } is relatively compact in C ( [ 0 , T ] ; B ) C([0,T];B) C ([ 0 , T ] ; B ) .
Lemma 2.11 (Prokhorov, [23 Theorem 1.3] ). A family Π \Pi Π of probability measures on a complete metric space E E E is relatively compact in the weak topology if and only if it is tight: for every δ > 0 \delta>0 δ > 0 there exists a compact K δ ⊂ E K_\delta\subset E K δ ⊂ E such that
sup μ ∈ Π μ ( E ∖ K δ ) < δ . \sup_{\mu\in\Pi}\mu(E\setminus K_\delta)<\delta. μ ∈ Π sup μ ( E ∖ K δ ) < δ .
Lemma 2.12. If X 1 , … , X N X_1,\ldots,X_N X 1 , … , X N are independent centered random variables satisfying
E [ e λ X j ] ≤ e 1 2 σ j 2 λ 2 ∀ λ ∈ R , \mathbb{E}\left[e^{\lambda X_j}\right]\le e^{\frac{1}{2}\sigma_j^2\lambda^2}\qquad\forall\lambda\in\mathbb{R}, E [ e λ X j ] ≤ e 2 1 σ j 2 λ 2 ∀ λ ∈ R , then, for any coefficients c j ∈ R c_j\in\mathbb{R} c j ∈ R ,
E [ e λ ∑ j c j X j ] ≤ exp ( λ 2 2 ∑ j c j 2 σ j 2 ) ∀ λ ∈ R . (23) \mathbb{E}\left[e^{\lambda\sum_j c_jX_j}\right]\le\exp\left(\frac{\lambda^2}{2}\sum_j c_j^2\sigma_j^2\right)\qquad\forall\lambda\in\mathbb{R}. \tag*{(23)} E [ e λ ∑ j c j X j ] ≤ exp ( 2 λ 2 j ∑ c j 2 σ j 2 ) ∀ λ ∈ R . (23) In particular, if a centered random variable X X X satisfies E [ e λ X ] ≤ e 1 2 σ 2 λ 2 \mathbb{E}[e^{\lambda X}]\le e^{\frac{1}{2}\sigma^2\lambda^2} E [ e λ X ] ≤ e 2 1 σ 2 λ 2 , then
E [ ∣ X ∣ p ] ≤ ( C p ) p 2 σ p ∀ p ≥ 2 , (24) \mathbb{E}[|X|^p]\le(Cp)^{\frac{p}{2}}\sigma^p\qquad\forall p\ge2, \tag*{(24)} E [ ∣ X ∣ p ] ≤ ( C p ) 2 p σ p ∀ p ≥ 2 , (24) where C C C is an absolute constant.
Proof. By independence,
E [ e λ ∑ j c j X j ] = ∏ j = 1 N E [ e λ c j X j ] . \mathbb{E}\left[e^{\lambda\sum_j c_jX_j}\right]
=
\prod_{j=1}^{N}\mathbb{E}\left[e^{\lambda c_jX_j}\right]. E [ e λ ∑ j c j X j ] = j = 1 ∏ N E [ e λ c j X j ] . Using the hypothesis with λ \lambda λ replaced by λ c j \lambda c_j λ c j ,
E [ e λ c j X j ] ≤ e 1 2 c j 2 σ j 2 λ 2 . \mathbb{E}\left[e^{\lambda c_jX_j}\right]\le e^{\frac{1}{2}c_j^2\sigma_j^2\lambda^2}. E [ e λ c j X j ] ≤ e 2 1 c j 2 σ j 2 λ 2 . Multiplying the inequalities over j j j gives
E [ e λ ∑ j c j X j ] ≤ ∏ j = 1 N e 1 2 c j 2 σ j 2 λ 2 = exp ( λ 2 2 ∑ j = 1 N c j 2 σ j 2 ) , \mathbb{E}\left[e^{\lambda\sum_j c_jX_j}\right]
\le
\prod_{j=1}^{N}e^{\frac{1}{2}c_j^2\sigma_j^2\lambda^2}
=
\exp\left(\frac{\lambda^2}{2}\sum_{j=1}^{N}c_j^2\sigma_j^2\right), E [ e λ ∑ j c j X j ] ≤ j = 1 ∏ N e 2 1 c j 2 σ j 2 λ 2 = exp ( 2 λ 2 j = 1 ∑ N c j 2 σ j 2 ) , which is (2.7).
We now prove (2.8). Taking N = 1 N=1 N = 1 in (2.7), the centered variable X X X satisfies
E [ e λ X ] ≤ e 1 2 σ 2 λ 2 ∀ λ ∈ R . \mathbb{E}\left[e^{\lambda X}\right] \le e^{\frac{1}{2}\sigma^2\lambda^2} \qquad\forall\ \lambda\in\mathbb{R}. E [ e λ X ] ≤ e 2 1 σ 2 λ 2 ∀ λ ∈ R .
For any t > 0 t>0 t > 0 and λ > 0 \lambda>0 λ > 0 , Markov’s inequality applied to e λ X e^{\lambda X} e λ X yields
P ( X > t ) ≤ e − λ t E [ e λ X ] ≤ exp ( − λ t + σ 2 λ 2 2 ) . \mathbb{P}(X>t) \le e^{-\lambda t}\mathbb{E}\left[e^{\lambda X}\right] \le\exp\left(-\lambda t+\frac{\sigma^2\lambda^2}{2}\right). P ( X > t ) ≤ e − λ t E [ e λ X ] ≤ exp ( − λ t + 2 σ 2 λ 2 ) .
Optimizing the right hand side over λ > 0 \lambda>0 λ > 0 , namely λ = t σ 2 \lambda=\frac{t}{\sigma^2} λ = σ 2 t , gives
P ( X > t ) ≤ exp ( − t 2 2 σ 2 ) . \mathbb{P}(X>t) \le\exp\left(-\frac{t^2}{2\sigma^2}\right). P ( X > t ) ≤ exp ( − 2 σ 2 t 2 ) .
Applying the same argument to − X -X − X and combining,
P ( ∣ X ∣ > t ) ≤ 2 exp ( − t 2 2 σ 2 ) . \mathbb{P}(|X|>t) \le2\exp\left(-\frac{t^2}{2\sigma^2}\right). P ( ∣ X ∣ > t ) ≤ 2 exp ( − 2 σ 2 t 2 ) .
For any p ≥ 2 p\ge2 p ≥ 2 , layer cake integration gives
E [ ∣ X ∣ p ] = ∫ 0 ∞ p t p − 1 P ( ∣ X ∣ > t ) d t ≤ 2 p ∫ 0 ∞ t p − 1 exp ( − t 2 2 σ 2 ) d t . \mathbb{E}[|X|^p]=\int_0^\infty pt^{p-1}\mathbb{P}(|X|>t)\,dt\le2p\int_0^\infty t^{p-1}\exp\left(-\frac{t^2}{2\sigma^2}\right)\,dt. E [ ∣ X ∣ p ] = ∫ 0 ∞ p t p − 1 P ( ∣ X ∣ > t ) d t ≤ 2 p ∫ 0 ∞ t p − 1 exp ( − 2 σ 2 t 2 ) d t .
With the change of variables u = t 2 2 σ 2 u=\frac{t^2}{2\sigma^2} u = 2 σ 2 t 2 , so that t = ( 2 σ 2 u ) 1 2 t=(2\sigma^2u)^{\frac{1}{2}} t = ( 2 σ 2 u ) 2 1 and d t = σ ( 2 u ) − 1 2 d u dt=\sigma(2u)^{-\frac{1}{2}}du d t = σ ( 2 u ) − 2 1 d u , the integral becomes
E [ ∣ X ∣ p ] ≤ 2 p σ p 2 p 2 − 1 ∫ 0 ∞ u p 2 − 1 e − u d u = 2 p 2 p 2 σ p Γ ( p 2 ) . \mathbb{E}[|X|^p]\le2p\sigma^p2^{\frac{p}{2}-1}\int_0^\infty u^{\frac{p}{2}-1}e^{-u}\,du=2p2^{\frac{p}{2}}\sigma^p\Gamma\left(\frac{p}{2}\right). E [ ∣ X ∣ p ] ≤ 2 p σ p 2 2 p − 1 ∫ 0 ∞ u 2 p − 1 e − u d u = 2 p 2 2 p σ p Γ ( 2 p ) .
Using Γ ( z ) ≤ C z z \Gamma(z)\le Cz^z Γ ( z ) ≤ C z z for z ≥ 1 z\ge1 z ≥ 1 with z = p 2 z=\frac{p}{2} z = 2 p , there is an absolute constant C C C such that
E [ ∣ X ∣ p ] ≤ ( C p ) p 2 σ p , \mathbb{E}[|X|^p]\le(Cp)^{\frac{p}{2}}\sigma^p, E [ ∣ X ∣ p ] ≤ ( C p ) 2 p σ p ,
which is (2.8).
Lemma 2.13. Let a a a be a centered stationary sub-Gaussian field satisfying Assumption 1.4(i). For any bounded measurable compactly supported test function g ∈ L 2 ( R d ) g\in L^2(\mathbb{R}^d) g ∈ L 2 ( R d ) , define the linear functional
X ( g ) : = ∫ R d g ( y ) a ( y ) d y . X(g):=\int_{\mathbb{R}^d}g(y)a(y)\,dy. X ( g ) := ∫ R d g ( y ) a ( y ) d y . There exists a constant C s g > 0 C_{\mathrm{sg}}>0 C sg > 0 depending on the sub-Gaussian constant C 0 C_0 C 0 and the mixing constants κ , δ \kappa,\delta κ , δ , such that for every p ≥ 2 p\ge2 p ≥ 2 ,
( E [ ∣ X ( g ) ∣ p ] ) 1 p ≤ C s g p ( E [ ∣ X ( g ) ∣ 2 ] ) 1 2 . (25) \left(\mathbb{E}\left[|X(g)|^p\right]\right)^{\frac{1}{p}}\le C_{\mathrm{sg}}\sqrt{p}\left(\mathbb{E}\left[|X(g)|^2\right]\right)^{\frac{1}{2}}. \tag*{(25)} ( E [ ∣ X ( g ) ∣ p ] ) p 1 ≤ C sg p ( E [ ∣ X ( g ) ∣ 2 ] ) 2 1 . (25) In particular, for p = 4 p=4 p = 4 ,
( E [ ∣ X ( g ) ∣ 4 ] ) 1 4 ≤ C s g 2 ( E [ ∣ X ( g ) ∣ 2 ] ) 1 2 . (26) \left(\mathbb{E}\left[|X(g)|^4\right]\right)^{\frac{1}{4}}\le C_{\mathrm{sg}}\sqrt{2}\left(\mathbb{E}\left[|X(g)|^2\right]\right)^{\frac{1}{2}}. \tag*{(26)} ( E [ ∣ X ( g ) ∣ 4 ] ) 4 1 ≤ C sg 2 ( E [ ∣ X ( g ) ∣ 2 ] ) 2 1 . (26)
Proof. By Assumption 1.4(i), for every y ∈ R d y\in\mathbb{R}^d y ∈ R d ,
E [ e λ a ( y ) ] ≤ e 1 2 C 0 λ 2 ∀ λ ∈ R . \mathbb{E}\left[e^{\lambda a(y)}\right]\le e^{\frac{1}{2}C_0\lambda^2}\qquad\forall\ \lambda\in\mathbb{R}. E [ e λa ( y ) ] ≤ e 2 1 C 0 λ 2 ∀ λ ∈ R .
The pointwise sub-Gaussian bound does not by itself control the integrated variable X ( g ) X(g) X ( g ) , because the samples a ( y ) a(y) a ( y ) at distinct points are correlated. We use the exponential mixing of Assumption 1.4(ii) to decorrelate.
We recall the standard decoupling consequence of (9) . If F 1 , … , F N \mathcal{F}_1,\ldots,\mathcal{F}_N F 1 , … , F N are σ \sigma σ -algebras generated by the field on sets pairwise separated by distance R R R , then for bounded measurable functions F j F_j F j of F j \mathcal{F}_j F j with ∥ F j ∥ ∞ ≤ M \lVert F_j\rVert_\infty\le M ∥ F j ∥ ∞ ≤ M , the exponential moments satisfy
∣ E [ e ∑ j λ F j ] − ∏ j E [ e λ F j ] ∣ ≤ C N M 2 e − κ R , (27) \left\lvert\mathbb{E}\left[e^{\sum_j \lambda F_j}\right]-\prod_j\mathbb{E}\left[e^{\lambda F_j}\right]\right\rvert\le C N M^2 e^{-\kappa R}, \tag*{(27)} E [ e ∑ j λ F j ] − j ∏ E [ e λ F j ] ≤ C N M 2 e − κ R , (27)
where C C C depends on δ \delta δ but not on R R R or N N N . Indeed, this follows by the standard mixing coupling: successively replace the law of F j F_j F j by an independent copy conditioned on the past. Each replacement changes the exponential moment by at most C M 2 e − κ R C M^2 e^{-\kappa R} C M 2 e − κ R by (9) , and summing over the N N N steps gives the bound (27) . The N N N factor appears because each step is a single conditional expectation perturbation, see e.g. [23 Chapter 4] for a standard coupling inequality.
Fix ℓ > 0 \ell> 0 ℓ > 0 and partition the support of g g g into cubes Q k Q_k Q k of side L ≫ ℓ L \gg\ell L ≫ ℓ . On each cube, write g ( y ) = g k + O ( L ) g(y)=g_k+O(L) g ( y ) = g k + O ( L ) with g k g_k g k the average of g g g on Q k Q_k Q k . Define the block integrals
Z k : = ∫ Q k a ( y ) d y , X L ( g ) : = ∑ k g k Z k . Z_k:=\int_{Q_k}a(y)\,\mathrm{d}y,\qquad X_L(g):=\sum_k g_k Z_k. Z k := ∫ Q k a ( y ) d y , X L ( g ) := k ∑ g k Z k .
First, each Z k Z_k Z k is sub-Gaussian. Subdivide Q k Q_k Q k into sub cubes of side ℓ \ell ℓ , and replace Z k Z_k Z k by the Riemann sum ∑ m a ( y k m ) ∣ Q k m ∣ \sum_m a(y_{km})\lvert Q_{km}\rvert ∑ m a ( y k m ) ∣ Q k m ∣ . The sample points y k m y_{km} y k m are pairwise separated by ℓ \ell ℓ , so by (27) with R = ℓ R=\ell R = ℓ ,
E [ e λ Z k ] ≤ exp ( λ 2 2 E [ Z k 2 ] ) e C k ( ℓ ) , C k ( ℓ ) = O ( ∣ Q k ∣ 2 e − κ ℓ / 2 ) , \mathbb{E}\left[e^{\lambda Z_k}\right]\le\exp\left(\frac{\lambda^2}{2}\mathbb{E}[Z_k^2]\right)e^{C_k(\ell)},\qquad C_k(\ell)=O\left(\lvert Q_k\rvert^2e^{-\kappa\ell/2}\right), E [ e λ Z k ] ≤ exp ( 2 λ 2 E [ Z k 2 ] ) e C k ( ℓ ) , C k ( ℓ ) = O ( ∣ Q k ∣ 2 e − κ ℓ /2 ) ,
where the O ( ∣ Q k ∣ 2 ) O(\lvert Q_k\rvert^2) O (∣ Q k ∣ 2 ) comes from the number of sub cubes. Since L ≫ ℓ L\gg\ell L ≫ ℓ , the sub leading correction C k ( ℓ ) C_k(\ell) C k ( ℓ ) is absorbed into the variance constant, which is allowed to depend on κ , δ , L / ℓ \kappa,\delta,L/\ell κ , δ , L / ℓ . Passing to the Riemann limit in ℓ \ell ℓ gives
E [ e λ Z k ] ≤ exp ( λ 2 2 C m i x E [ Z k 2 ] ) , E [ Z k 2 ] = ∫ Q k ∫ Q k R ( y − z ) d y d z ≤ C L d , \mathbb{E}\left[e^{\lambda Z_k}\right]\le\exp\left(\frac{\lambda^2}{2}C_{\mathrm{mix}}\mathbb{E}[Z_k^2]\right),\qquad\mathbb{E}[Z_k^2]=\int_{Q_k}\int_{Q_k}R(y-z)\,\mathrm{d}y\,\mathrm{d}z\le C L^d, E [ e λ Z k ] ≤ exp ( 2 λ 2 C mix E [ Z k 2 ] ) , E [ Z k 2 ] = ∫ Q k ∫ Q k R ( y − z ) d y d z ≤ C L d ,
with C m i x ≥ 1 C_{\mathrm{mix}}\ge1 C mix ≥ 1 depending only on κ , δ \kappa,\delta κ , δ .
Second, the block integrals Z k Z_k Z k for distinct cubes are separated by L ≫ ℓ L\gg\ell L ≫ ℓ , so applying (27) once more with R = L R=L R = L and F k = λ g k Z k F_k=\lambda g_kZ_k F k = λ g k Z k ,
E [ e λ X L ( g ) ] ≤ ∏ k E [ e λ g k Z k ] e C ( κ , δ ) N M 2 e − κ L ≤ exp ( λ 2 2 C m i x ∑ k g k 2 E [ Z k 2 ] ) , \mathbb{E}\left[e^{\lambda X_L(g)}\right]\le\prod_k\mathbb{E}\left[e^{\lambda g_kZ_k}\right]e^{C(\kappa,\delta)NM^2e^{-\kappa L}}\le\exp\left(\frac{\lambda^2}{2}C_{\mathrm{mix}}\sum_k g_k^2\mathbb{E}[Z_k^2]\right), E [ e λ X L ( g ) ] ≤ k ∏ E [ e λ g k Z k ] e C ( κ , δ ) N M 2 e − κ L ≤ exp ( 2 λ 2 C mix k ∑ g k 2 E [ Z k 2 ] ) ,
where the second step absorbs the e − κ L / 2 e^{-\kappa L/2} e − κ L /2 error into the constant C m i x C_{\mathrm{mix}} C mix by choosing L / ℓ L/\ell L / ℓ large enough. Since
∑ k g k 2 E [ Z k 2 ] ⟶ E [ X ( g ) 2 ] = ∫ R d ∫ R d g ( y ) g ( z ) R ( y − z ) d y d z \sum_k g_k^2\mathbb{E}[Z_k^2]\longrightarrow\mathbb{E}[X(g)^2]=\int_{\mathbb{R}^d}\int_{\mathbb{R}^d}g(y)g(z)R(y-z)\,\mathrm{d}y\,\mathrm{d}z k ∑ g k 2 E [ Z k 2 ] ⟶ E [ X ( g ) 2 ] = ∫ R d ∫ R d g ( y ) g ( z ) R ( y − z ) d y d z
as ℓ → 0 \ell\to0 ℓ → 0 , passing to the limit gives
E [ e λ X ( g ) ] ≤ exp ( λ 2 2 C m i x E [ X ( g ) 2 ] ) . \mathbb{E}\left[e^{\lambda X(g)}\right]\le\exp\left(\frac{\lambda^2}{2}C_{\mathrm{mix}}\mathbb{E}[X(g)^2]\right). E [ e λ X ( g ) ] ≤ exp ( 2 λ 2 C mix E [ X ( g ) 2 ] ) .
By the moment estimate (24) in Lemma 2.12,
E [ ∣ X ( g ) ∣ p ] ≤ ( C p ) p / 2 ( C m i x E [ X ( g ) 2 ] ) p / 2 . \mathbb{E}\left[\lvert X(g)\rvert^p\right]\le(C p)^{p/2}\left(C_{\mathrm{mix}}\mathbb{E}[X(g)^2]\right)^{p/2}. E [ ∣ X ( g ) ∣ p ] ≤ ( C p ) p /2 ( C mix E [ X ( g ) 2 ] ) p /2 .
Taking the p p p -th root and setting C s g = C C m i x C_{\mathrm{sg}}=C\sqrt{C_{\mathrm{mix}}} C sg = C C mix proves (25) . The estimate extends to general g ∈ L 2 ( R d ) g\in L^2(\mathbb{R}^d) g ∈ L 2 ( R d ) by approximation. The constant depends only on C 0 C_0 C 0 and on the mixing constants κ , δ \kappa,\delta κ , δ . □
Duhamel decomposition and iteration
Write u ε = u ˉ + w ε u_\varepsilon=\bar{u}+w_\varepsilon u ε = u ˉ + w ε . Subtracting ( L S E 0 ) (\mathrm{LSE}_0) ( LSE 0 ) from ( L S E ε ) (\mathrm{LSE}_\varepsilon) ( LSE ε ) ,
i ∂ t w ε = H 0 w ε + a ε ( u ˉ + w ε ) , w ε ( 0 ) = 0 , i\partial_t w_\varepsilon=H_0w_\varepsilon+a_\varepsilon(\bar{u}+w_\varepsilon),\qquad w_\varepsilon(0)=0, i ∂ t w ε = H 0 w ε + a ε ( u ˉ + w ε ) , w ε ( 0 ) = 0 ,
where we recall a ε ( x ) = a ( x ε ) a_\varepsilon(x)=a\left(\frac{x}{\varepsilon}\right) a ε ( x ) = a ( ε x ) . The Duhamel form is
w ε ( t ) = − i ∫ 0 t S ( t , r ) a ε ( r ) ( u ˉ ( r ) + w ε ( r ) ) d r . (28) w_\varepsilon(t)=-i\int_0^t S(t,r)a_\varepsilon(r)\bigl(\bar{u}(r)+w_\varepsilon(r)\bigr)\,dr. \tag*{(28)} w ε ( t ) = − i ∫ 0 t S ( t , r ) a ε ( r ) ( u ˉ ( r ) + w ε ( r ) ) d r . (28)
Iterating once,
w ε = w ε ( 1 ) + w ε ( 2 ) , w ε ( 1 ) ( t ) : = − i ∫ 0 t S ( t , r ) a ε ( r ) u ˉ ( r ) d r , w_\varepsilon=w_\varepsilon^{(1)}+w_\varepsilon^{(2)},\qquad w_\varepsilon^{(1)}(t):=-i\int_0^t S(t,r)a_\varepsilon(r)\bar{u}(r)\,dr, w ε = w ε ( 1 ) + w ε ( 2 ) , w ε ( 1 ) ( t ) := − i ∫ 0 t S ( t , r ) a ε ( r ) u ˉ ( r ) d r ,
and the remainder is
w ε ( 2 ) ( t ) : = − i ∫ 0 t S ( t , r ) a ε ( r ) w ε ( r ) d r . w_\varepsilon^{(2)}(t):=-i\int_0^t S(t,r)a_\varepsilon(r)w_\varepsilon(r)\,dr. w ε ( 2 ) ( t ) := − i ∫ 0 t S ( t , r ) a ε ( r ) w ε ( r ) d r .
Second moment estimates
The Schrödinger propagator e i ( t − s ) Δ e^{i(t-s)\Delta} e i ( t − s ) Δ has infinite propagation speed: its Fresnel kernel is nonzero for every x ≠ y x\ne y x = y with only algebraic decay. A direct physical space Gronwall estimate would therefore accumulate contributions from the entire space at every time step. We instead work in Fourier space, where the off diagonal terms are controlled by the rapid oscillation of the phase factor e i ( t − s ) ( ∣ ξ + ξ ′ ε ∣ 2 − ∣ ξ ∣ 2 ) e^{i(t-s)\left(\left|\xi+\frac{\xi'}{\varepsilon}\right|^2-|\xi|^2\right)} e i ( t − s ) ( ξ + ε ξ ′ 2 − ∣ ξ ∣ 2 ) combined with the exponential decay of the spectral density R ^ ( ξ ′ ) \widehat{R}(\xi') R ( ξ ′ ) .
By (17) and the convolution theorem, the Fourier transform of the first order correction has the following representation.
Proposition 3.1. Under Assumption 1.4, for the first order Duhamel term w ε ( 1 ) w_\varepsilon^{(1)} w ε ( 1 ) ,
w ε ( 1 ) ^ ( t , ξ ) = − i ε d e − i ( ∣ ξ ∣ 2 + V ˉ ) t ( 2 π ) − d ∫ R d a ^ ( ε η ) u ^ 0 ( ξ − η ) [ ∫ 0 t e i r σ ( ξ , η ) d r ] d η , (29) \widehat{w_\varepsilon^{(1)}}(t,\xi)=-i\varepsilon^d e^{-i(|\xi|^2+\bar{V})t}(2\pi)^{-d}\int_{\mathbb{R}^d}\widehat{a}(\varepsilon\eta)\widehat{u}_0(\xi-\eta)\left[\int_0^t e^{ir\sigma(\xi,\eta)}\,dr\right]d\eta, \tag*{(29)} w ε ( 1 ) ( t , ξ ) = − i ε d e − i ( ∣ ξ ∣ 2 + V ˉ ) t ( 2 π ) − d ∫ R d a ( ε η ) u 0 ( ξ − η ) [ ∫ 0 t e i r σ ( ξ , η ) d r ] d η , (29) where
σ ( ξ , η ) : = ∣ ξ ∣ 2 − ∣ ξ − η ∣ 2 = 2 ξ ⋅ η − ∣ η ∣ 2 . \sigma(\xi,\eta):=|\xi|^2-|\xi-\eta|^2=2\xi\cdot\eta-|\eta|^2. σ ( ξ , η ) := ∣ ξ ∣ 2 − ∣ ξ − η ∣ 2 = 2 ξ ⋅ η − ∣ η ∣ 2 . Proof. By (17) and the convolution theorem F ( f g ) = ( 2 π ) − d f ^ ∗ g ^ \mathcal{F}(fg)=(2\pi)^{-d}\widehat{f}*\widehat{g} F ( f g ) = ( 2 π ) − d f ∗ g ,
w ε ( 1 ) ^ ( t , ξ ) = − i ε d ∫ 0 t ∫ R d e − i ( ∣ ξ ∣ 2 + V ˉ ) ( t − r ) a ^ ( ε η ) F ( u ˉ ( r ) ) ( ξ − η ) ( 2 π ) − d d η d r . \widehat{w_\varepsilon^{(1)}}(t,\xi)=-i\varepsilon^d\int_0^t\int_{\mathbb{R}^d}e^{-i(|\xi|^2+\bar{V})(t-r)}\widehat{a}(\varepsilon\eta)\mathcal{F}(\bar{u}(r))(\xi-\eta)(2\pi)^{-d}\,d\eta\,dr. w ε ( 1 ) ( t , ξ ) = − i ε d ∫ 0 t ∫ R d e − i ( ∣ ξ ∣ 2 + V ˉ ) ( t − r ) a ( ε η ) F ( u ˉ ( r )) ( ξ − η ) ( 2 π ) − d d η d r . Since u ˉ \bar{u} u ˉ solves ( L S E 0 ) (\mathrm{LSE}_0) ( LSE 0 ) ,
F ( u ˉ ( r ) ) ( ξ ′ ) = e − i ( ∣ ξ ′ ∣ 2 + V ˉ ) r u ^ 0 ( ξ ′ ) . \mathcal{F}(\bar{u}(r))(\xi')=e^{-i(|\xi'|^2+\bar{V})r}\widehat{u}_0(\xi'). F ( u ˉ ( r )) ( ξ ′ ) = e − i ( ∣ ξ ′ ∣ 2 + V ˉ ) r u 0 ( ξ ′ ) . Substituting and integrating the time exponential,
∫ 0 t e − i ( ∣ ξ ∣ 2 + V ˉ ) ( t − r ) e − i ( ∣ ξ − η ∣ 2 + V ˉ ) r d r = e − i ( ∣ ξ ∣ 2 + V ˉ ) t ∫ 0 t e i r ( ∣ ξ ∣ 2 − ∣ ξ − η ∣ 2 ) d r . \int_0^t e^{-i(|\xi|^2+\bar{V})(t-r)}e^{-i(|\xi-\eta|^2+\bar{V})r}\,dr=e^{-i(|\xi|^2+\bar{V})t}\int_0^t e^{ir(|\xi|^2-|\xi-\eta|^2)}\,dr. ∫ 0 t e − i ( ∣ ξ ∣ 2 + V ˉ ) ( t − r ) e − i ( ∣ ξ − η ∣ 2 + V ˉ ) r d r = e − i ( ∣ ξ ∣ 2 + V ˉ ) t ∫ 0 t e i r ( ∣ ξ ∣ 2 − ∣ ξ − η ∣ 2 ) d r . This gives (29) . □
For the generalized Fourier transform of the stationary field,
E [ a ^ ( α ) a ^ ( β ) ] = ( 2 π ) d δ ( α − β ) R ^ ( α ) , \mathbb{E}\bigl[\widehat{a}(\alpha)\widehat{a}(\beta)\bigr]=(2\pi)^d\delta(\alpha-\beta)\widehat{R}(\alpha), E [ a ( α ) a ( β ) ] = ( 2 π ) d δ ( α − β ) R ( α ) ,
and with α = ε η \alpha=\varepsilon\eta α = ε η , β = ε η ′ \beta=\varepsilon\eta' β = ε η ′ ,
E [ a ^ ( ε η ) a ^ ( ε η ′ ) ] = ( 2 π ) d ε − d δ ( η − η ′ ) R ^ ( ε η ) . (30) \mathbb{E}\bigl[\widehat{a}(\varepsilon\eta)\widehat{a}(\varepsilon\eta')\bigr]=(2\pi)^d\varepsilon^{-d}\delta(\eta-\eta')\widehat{R}(\varepsilon\eta). \tag*{(30)} E [ a ( ε η ) a ( ε η ′ ) ] = ( 2 π ) d ε − d δ ( η − η ′ ) R ( ε η ) . (30)
We now compute the second moment in Fourier space. Multiplying (3.2) by its conjugate and taking expectation,
E [ ∣ w ^ ε ( 1 ) ( t , ξ ) ∣ 2 ] = ε d ( 2 π ) − d ∫ R d R ^ ( ε η ) ∣ ∫ 0 t e i r σ ( ξ , η ) d r ∣ 2 ∣ u ^ 0 ( ξ − η ) ∣ 2 d η . \mathbb{E}\left[\left|\widehat{w}_{\varepsilon}^{(1)}(t,\xi)\right|^2\right]=\varepsilon^d(2\pi)^{-d}\int_{\mathbb{R}^d}\widehat{R}(\varepsilon\eta)\left|\int_0^t e^{ir\sigma(\xi,\eta)}\,\mathrm{d}r\right|^2\left|\widehat{u}_0(\xi-\eta)\right|^2\mathrm{d}\eta. E [ w ε ( 1 ) ( t , ξ ) 2 ] = ε d ( 2 π ) − d ∫ R d R ( ε η ) ∫ 0 t e i r σ ( ξ , η ) d r 2 ∣ u 0 ( ξ − η ) ∣ 2 d η .
By Plancherel,
E [ ∥ w ε ( 1 ) ( t ) ∥ L 2 x 2 ] = ε d ( 2 π ) − 2 d ∫ R d ∫ R d R ^ ( ε η ) ∣ ∫ 0 t e i r σ ( ξ , η ) d r ∣ 2 ∣ u ^ 0 ( ξ − η ) ∣ 2 d η d ξ . \mathbb{E}\left[\left\|w_{\varepsilon}^{(1)}(t)\right\|_{L^x_2}^2\right]=\varepsilon^d(2\pi)^{-2d}\int_{\mathbb{R}^d}\int_{\mathbb{R}^d}\widehat{R}(\varepsilon\eta)\left|\int_0^t e^{ir\sigma(\xi,\eta)}\,\mathrm{d}r\right|^2\left|\widehat{u}_0(\xi-\eta)\right|^2\mathrm{d}\eta\mathrm{d}\xi. E [ w ε ( 1 ) ( t ) L 2 x 2 ] = ε d ( 2 π ) − 2 d ∫ R d ∫ R d R ( ε η ) ∫ 0 t e i r σ ( ξ , η ) d r 2 ∣ u 0 ( ξ − η ) ∣ 2 d η d ξ .
The time double integral evaluates to
∣ ∫ 0 t e i r σ d r ∣ 2 = ∫ 0 t ∫ 0 t e i ( r − r ′ ) σ d r d r ′ = 2 ( 1 − cos ( t σ ) ) σ 2 . \left|\int_0^t e^{ir\sigma}\,\mathrm{d}r\right|^2=\int_0^t\int_0^t e^{i(r-r')\sigma}\,\mathrm{d}r\mathrm{d}r'=\frac{2(1-\cos(t\sigma))}{\sigma^2}. ∫ 0 t e i r σ d r 2 = ∫ 0 t ∫ 0 t e i ( r − r ′ ) σ d r d r ′ = σ 2 2 ( 1 − cos ( t σ )) .
Change variables ζ = ξ − η \zeta=\xi-\eta ζ = ξ − η , so that σ ( ξ , ξ − ζ ) = ∣ ξ ∣ 2 − ∣ ζ ∣ 2 \sigma(\xi,\xi-\zeta)=|\xi|^2-|\zeta|^2 σ ( ξ , ξ − ζ ) = ∣ ξ ∣ 2 − ∣ ζ ∣ 2 , and then set ξ ′ = ε ( ξ − ζ ) \xi'=\varepsilon(\xi-\zeta) ξ ′ = ε ( ξ − ζ ) , so ξ = ζ + ξ ′ ε \xi=\zeta+\frac{\xi'}{\varepsilon} ξ = ζ + ε ξ ′ and d ξ = ε − d d ξ ′ \mathrm{d}\xi=\varepsilon^{-d}\mathrm{d}\xi' d ξ = ε − d d ξ ′ . Then
∣ ξ ∣ 2 − ∣ ζ ∣ 2 = 2 ζ ⋅ ξ ′ ε + ∣ ξ ′ ∣ 2 ε 2 , |\xi|^2-|\zeta|^2=\frac{2\zeta\cdot\xi'}{\varepsilon}+\frac{|\xi'|^2}{\varepsilon^2}, ∣ ξ ∣ 2 − ∣ ζ ∣ 2 = ε 2 ζ ⋅ ξ ′ + ε 2 ∣ ξ ′ ∣ 2 ,
and the integral becomes
E [ ∥ w ε ( 1 ) ( t ) ∥ L 2 x 2 ] = ( 2 π ) − 2 d ∫ R d ∣ u ^ 0 ( ζ ) ∣ 2 ∫ R d R ^ ( ξ ′ ) 2 ( 1 − cos ( t ϕ ) ) ϕ 2 d ξ ′ d ζ , ϕ : = 2 ζ ⋅ ξ ′ ε + ∣ ξ ′ ∣ 2 ε 2 . (31) \mathbb{E}\left[\left\|w_{\varepsilon}^{(1)}(t)\right\|_{L^x_2}^2\right]=(2\pi)^{-2d}\int_{\mathbb{R}^d}\left|\widehat{u}_0(\zeta)\right|^2\int_{\mathbb{R}^d}\widehat{R}(\xi')\frac{2(1-\cos(t\phi))}{\phi^2}\,\mathrm{d}\xi'\mathrm{d}\zeta,\qquad\phi:=\frac{2\zeta\cdot\xi'}{\varepsilon}+\frac{|\xi'|^2}{\varepsilon^2}. \tag*{(31)} E [ w ε ( 1 ) ( t ) L 2 x 2 ] = ( 2 π ) − 2 d ∫ R d ∣ u 0 ( ζ ) ∣ 2 ∫ R d R ( ξ ′ ) ϕ 2 2 ( 1 − cos ( tϕ )) d ξ ′ d ζ , ϕ := ε 2 ζ ⋅ ξ ′ + ε 2 ∣ ξ ′ ∣ 2 . (31)
Since u ^ 0 \widehat{u}_0 u 0 is compactly supported, write supp ( u ^ 0 ) ⊂ B R \operatorname{supp}(\widehat{u}_0)\subset B_R supp ( u 0 ) ⊂ B R , so that ∣ ζ ∣ ≤ R |\zeta|\le R ∣ ζ ∣ ≤ R uniformly. The estimate of the ξ ′ \xi' ξ ′ -integral in (31) is split into the near region ∣ ξ ′ ∣ ≤ ε |\xi'|\le\varepsilon ∣ ξ ′ ∣ ≤ ε and the far region ∣ ξ ′ ∣ > ε |\xi'|>\varepsilon ∣ ξ ′ ∣ > ε .
Lemma 3.2 (Near region). On ∣ ξ ′ ∣ ≤ ε |\xi'|\le\varepsilon ∣ ξ ′ ∣ ≤ ε ,
∫ ∣ ξ ′ ∣ ≤ ε R ^ ( ξ ′ ) 2 ( 1 − cos ( t ϕ ) ) ϕ 2 d ξ ′ ≤ C T , u 0 ε d . (32) \int_{|\xi'|\le\varepsilon}\widehat{R}(\xi')\frac{2(1-\cos(t\phi))}{\phi^2}\,\mathrm{d}\xi'\le C_{T,u_0}\varepsilon^d. \tag*{(32)} ∫ ∣ ξ ′ ∣ ≤ ε R ( ξ ′ ) ϕ 2 2 ( 1 − cos ( tϕ )) d ξ ′ ≤ C T , u 0 ε d . (32) Proof. Since ∣ ξ ′ ∣ ≤ ε |\xi'|\le\varepsilon ∣ ξ ′ ∣ ≤ ε ,
∣ ϕ ∣ ≤ 2 ∣ ζ ∣ ∣ ξ ′ ∣ ε + ∣ ξ ′ ∣ 2 ε 2 ≤ 2 ∣ ζ ∣ + 1. |\phi|\le\frac{2|\zeta||\xi'|}{\varepsilon}+\frac{|\xi'|^2}{\varepsilon^2}\le2|\zeta|+1. ∣ ϕ ∣ ≤ ε 2∣ ζ ∣∣ ξ ′ ∣ + ε 2 ∣ ξ ′ ∣ 2 ≤ 2∣ ζ ∣ + 1. With ∣ ζ ∣ ≤ R |\zeta|\le R ∣ ζ ∣ ≤ R uniformly, ∣ ϕ ∣ ≤ 2 R + 1 = : C R |\phi|\le2R+1=:C_R ∣ ϕ ∣ ≤ 2 R + 1 =: C R . Using 1 − cos u ≤ u 2 2 1-\cos u\le\frac{u^2}{2} 1 − cos u ≤ 2 u 2 ,
2 ( 1 − cos ( t ϕ ) ) ϕ 2 ≤ t 2 , \frac{2(1-\cos(t\phi))}{\phi^2}\le t^2, ϕ 2 2 ( 1 − cos ( tϕ )) ≤ t 2 , and hence,
∫ ∣ ξ ′ ∣ ≤ ε R ^ ( ξ ′ ) 2 ( 1 − cos ( t ϕ ) ) ϕ 2 d ξ ′ ≤ t 2 ∥ R ^ ∥ L ∞ ∫ ∣ ξ ′ ∣ ≤ ε d ξ ′ = C T , u 0 ε d . \int_{|\xi'|\le\varepsilon}\widehat{R}(\xi')\frac{2(1-\cos(t\phi))}{\phi^2}\,\mathrm{d}\xi'\le t^2\|\widehat{R}\|_{L^\infty}\int_{|\xi'|\le\varepsilon}\mathrm{d}\xi'=C_{T,u_0}\varepsilon^d. ∫ ∣ ξ ′ ∣ ≤ ε R ( ξ ′ ) ϕ 2 2 ( 1 − cos ( tϕ )) d ξ ′ ≤ t 2 ∥ R ∥ L ∞ ∫ ∣ ξ ′ ∣ ≤ ε d ξ ′ = C T , u 0 ε d .
Lemma 3.3 (Far region). For d ≤ 3 d\le3 d ≤ 3 , on ∣ ξ ′ ∣ > ε |\xi'|>\varepsilon ∣ ξ ′ ∣ > ε ,
∫ ∣ ξ ′ ∣ > ε R ^ ( ξ ′ ) 2 ( 1 − cos ( t ϕ ) ) ϕ 2 d ξ ′ ≤ C T , u 0 ε d . (33) \int_{|\xi'|>\varepsilon}\widehat{R}(\xi')\frac{2(1-\cos(t\phi))}{\phi^2}\,\mathrm{d}\xi'\le C_{T,u_0}\varepsilon^d. \tag*{(33)} ∫ ∣ ξ ′ ∣ > ε R ( ξ ′ ) ϕ 2 2 ( 1 − cos ( tϕ )) d ξ ′ ≤ C T , u 0 ε d . (33) Proof. For ∣ ξ ′ ∣ > ε |\xi'|>\varepsilon ∣ ξ ′ ∣ > ε , the quadratic term dominates,
∣ ϕ ∣ = ∣ 2 ζ ⋅ ξ ′ ε + ∣ ξ ′ ∣ 2 ε 2 ∣ ≥ ∣ ξ ′ ∣ 2 ε 2 − 2 ∣ ζ ∣ ∣ ξ ′ ∣ ε ≥ ∣ ξ ′ ∣ 2 2 ε 2 , |\phi|=\left|\frac{2\zeta\cdot\xi'}{\varepsilon}+\frac{|\xi'|^2}{\varepsilon^2}\right|\ge\frac{|\xi'|^2}{\varepsilon^2}-\frac{2|\zeta||\xi'|}{\varepsilon}\ge\frac{|\xi'|^2}{2\varepsilon^2}, ∣ ϕ ∣ = ε 2 ζ ⋅ ξ ′ + ε 2 ∣ ξ ′ ∣ 2 ≥ ε 2 ∣ ξ ′ ∣ 2 − ε 2∣ ζ ∣∣ ξ ′ ∣ ≥ 2 ε 2 ∣ ξ ′ ∣ 2 ,
where the last inequality holds for ε \varepsilon ε sufficiently small so that 4 ∣ ζ ∣ ε ≤ ∣ ξ ′ ∣ 4|\zeta|\varepsilon\le|\xi'| 4∣ ζ ∣ ε ≤ ∣ ξ ′ ∣ on ∣ ξ ′ ∣ > ε |\xi'|>\varepsilon ∣ ξ ′ ∣ > ε , using ∣ ζ ∣ ≤ R |\zeta|\le R ∣ ζ ∣ ≤ R uniformly. Using 1 − cos u ≤ 2 1-\cos u\le2 1 − cos u ≤ 2 ,
2 ( 1 − cos ( t ϕ ) ) ϕ 2 ≤ 4 ϕ 2 ≤ 16 ε 4 ∣ ξ ′ ∣ 4 . \frac{2(1-\cos(t\phi))}{\phi^2}\le\frac{4}{\phi^2}\le\frac{16\varepsilon^4}{|\xi'|^4}. ϕ 2 2 ( 1 − cos ( tϕ )) ≤ ϕ 2 4 ≤ ∣ ξ ′ ∣ 4 16 ε 4 .
Dyadically decompose ∣ ξ ′ ∣ ∈ [ 2 k ε , 2 k + 1 ε ] |\xi'|\in[2^k\varepsilon,2^{k+1}\varepsilon] ∣ ξ ′ ∣ ∈ [ 2 k ε , 2 k + 1 ε ] , k ≥ 0 k\ge0 k ≥ 0 . On the k k k -th shell, ∣ ξ ′ ∣ − 4 ∼ 2 − 4 k ε − 4 |\xi'|^{-4}\sim2^{-4k}\varepsilon^{-4} ∣ ξ ′ ∣ − 4 ∼ 2 − 4 k ε − 4 , the shell volume is O ( 2 k d ε d ) O(2^{kd}\varepsilon^d) O ( 2 k d ε d ) , and the integrand bound becomes O ( 2 − 4 k ) O(2^{-4k}) O ( 2 − 4 k ) . The shell contribution is
O ( 2 − 4 k ) ⋅ O ( 2 k d ε d ) = O ( ε d 2 k ( d − 4 ) ) . O(2^{-4k})\cdot O(2^{kd}\varepsilon^d)=O(\varepsilon^d2^{k(d-4)}). O ( 2 − 4 k ) ⋅ O ( 2 k d ε d ) = O ( ε d 2 k ( d − 4 ) ) .
For d ≤ 3 d\le3 d ≤ 3 , d − 4 ≤ − 1 d-4\le-1 d − 4 ≤ − 1 , so the series over k ≥ 0 k\ge0 k ≥ 0 converges geometrically,
∫ ∣ ξ ′ ∣ > ε R ^ ( ξ ′ ) 2 ( 1 − cos ( t ϕ ) ) ϕ 2 d ξ ′ ≤ C ε d ∑ k = 0 ∞ 2 k ( d − 4 ) ≤ C T , u 0 ε d . \int_{|\xi'|>\varepsilon}\widehat R(\xi')\frac{2(1-\cos(t\phi))}{\phi^2}\,d\xi'
\le C\varepsilon^d\sum_{k=0}^{\infty}2^{k(d-4)}
\le C_{T,u_0}\varepsilon^d. ∫ ∣ ξ ′ ∣ > ε R ( ξ ′ ) ϕ 2 2 ( 1 − cos ( tϕ )) d ξ ′ ≤ C ε d k = 0 ∑ ∞ 2 k ( d − 4 ) ≤ C T , u 0 ε d .
Proposition 3.4. Under Assumption 1.4, for every d ≥ 1 d\ge1 d ≥ 1 , uniformly in t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≤ C T , u 0 ε d . (34) \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^2\right]\le C_{T,u_0}\varepsilon^d. \tag*{(34)} E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≤ C T , u 0 ε d . (34)
Proof. For d ≤ 3 d\le3 d ≤ 3 , by (31) , the ξ ′ \xi' ξ ′ -integral is the sum of the near and far regions, bounded by Lemma 3.2 and Lemma 3.3 respectively, each of order C T , u 0 ε d C_{T,u_0}\varepsilon^d C T , u 0 ε d . The prefactor
( 2 π ) − 2 d ∫ ∣ u ^ 0 ( ζ ) ∣ 2 d ζ (2\pi)^{-2d}\int|\widehat u_0(\zeta)|^2\,d\zeta ( 2 π ) − 2 d ∫ ∣ u 0 ( ζ ) ∣ 2 d ζ
is a constant, and the Jacobian cancellation d ξ = ε − d d ξ ′ d\xi=\varepsilon^{-d}d\xi' d ξ = ε − d d ξ ′ exactly cancels the prefactor ε d \varepsilon^d ε d from the Fourier transform of a ε a_\varepsilon a ε .
For d ≥ 4 d\ge4 d ≥ 4 , the same bound follows from the physical space oscillatory integral estimate of Lemma A.1 in Appendix A.
Hence, (3.7) holds for every d ≥ 1 d\ge1 d ≥ 1 .
Proposition 3.5. Assume R ^ ( 0 ) > 0 \widehat R(0)>0 R ( 0 ) > 0 . For any fixed t > 0 t>0 t > 0 and any u 0 ∈ L 2 ( R d ) u_0\in L^2(\mathbb{R}^d) u 0 ∈ L 2 ( R d ) with u ^ 0 ( 0 ) ≠ 0 \widehat u_0(0)\ne0 u 0 ( 0 ) = 0 , there exist constants c t , u 0 > 0 c_{t,u_0}>0 c t , u 0 > 0 and ε 0 > 0 \varepsilon_0>0 ε 0 > 0 such that
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≥ c t , u 0 ε d for all 0 < ε ≤ ε 0 . (35) \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^2\right]\ge c_{t,u_0}\varepsilon^d\qquad\text{for all }0<\varepsilon\le\varepsilon_0. \tag*{(35)} E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≥ c t , u 0 ε d for all 0 < ε ≤ ε 0 . (35)
Proof. From the diagonal contribution in the second moment expansion,
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≥ ( 2 π ) − 2 d ∫ ∣ ξ ′ ∣ ≤ ε R ^ ( ξ ′ ) 2 ( 1 − cos ( t ϕ ) ) ϕ 2 ∣ u ^ 0 ( ζ + ξ ′ ε ) ∣ 2 d ξ ′ d ζ , \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^2\right]\ge(2\pi)^{-2d}\int_{|\xi'|\le\varepsilon}\widehat R(\xi')\frac{2(1-\cos(t\phi))}{\phi^2}\left|\widehat u_0\left(\zeta+\frac{\xi'}{\varepsilon}\right)\right|^2\,d\xi'\,d\zeta, E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≥ ( 2 π ) − 2 d ∫ ∣ ξ ′ ∣ ≤ ε R ( ξ ′ ) ϕ 2 2 ( 1 − cos ( tϕ )) u 0 ( ζ + ε ξ ′ ) 2 d ξ ′ d ζ ,
where
ϕ = 2 ζ ⋅ ξ ′ ε + ∣ ξ ′ ∣ 2 ε 2 . \phi=\frac{2\zeta\cdot\xi'}{\varepsilon}+\frac{|\xi'|^2}{\varepsilon^2}. ϕ = ε 2 ζ ⋅ ξ ′ + ε 2 ∣ ξ ′ ∣ 2 .
On the near region ∣ ξ ′ ∣ ≤ ε |\xi'|\le\varepsilon ∣ ξ ′ ∣ ≤ ε ,
∣ ϕ ∣ ≤ 2 ∣ ζ ∣ + 1. |\phi|\le2|\zeta|+1. ∣ ϕ ∣ ≤ 2∣ ζ ∣ + 1.
Since u ^ 0 \widehat u_0 u 0 is continuous and u ^ 0 ( 0 ) ≠ 0 \widehat u_0(0)\ne0 u 0 ( 0 ) = 0 , there exists r > 0 r>0 r > 0 such that ∣ u ^ 0 ( ξ ) ∣ ≥ c u 0 > 0 |\widehat u_0(\xi)|\ge c_{u_0}>0 ∣ u 0 ( ξ ) ∣ ≥ c u 0 > 0 for all ∣ ξ ∣ ≤ r |\xi|\le r ∣ ξ ∣ ≤ r . Restrict the ζ \zeta ζ -integral to ∣ ζ ∣ ≤ r / 2 |\zeta|\le r/2 ∣ ζ ∣ ≤ r /2 , which is nonempty and has positive volume for any r > 0 r>0 r > 0 . Then, for ∣ ξ ′ ∣ ≤ ε |\xi'|\le\varepsilon ∣ ξ ′ ∣ ≤ ε and ε \varepsilon ε sufficiently small,
∣ ζ + ξ ′ ε ∣ ≤ ∣ ζ ∣ + ∣ ξ ′ ∣ ε ≤ r 2 + 1 ≤ r , \left|\zeta+\frac{\xi'}{\varepsilon}\right|\le|\zeta|+\frac{|\xi'|}{\varepsilon}\le\frac r2+1\le r, ζ + ε ξ ′ ≤ ∣ ζ ∣ + ε ∣ ξ ′ ∣ ≤ 2 r + 1 ≤ r ,
so,
∣ u ^ 0 ( ζ + ξ ′ ε ) ∣ 2 ≥ c u 0 2 , ∣ ϕ ∣ ≤ 2 ∣ ζ ∣ + 1 ≤ r + 1 = : C r . \left|\widehat u_0\left(\zeta+\frac{\xi'}{\varepsilon}\right)\right|^2\ge c_{u_0}^2,\qquad|\phi|\le2|\zeta|+1\le r+1=:C_r. u 0 ( ζ + ε ξ ′ ) 2 ≥ c u 0 2 , ∣ ϕ ∣ ≤ 2∣ ζ ∣ + 1 ≤ r + 1 =: C r .
Therefore,
2 ( 1 − cos ( t ϕ ) ) ϕ 2 ≥ c t > 0 \frac{2(1-\cos(t\phi))}{\phi^2} \ge c_t>0 ϕ 2 2 ( 1 − cos ( tϕ )) ≥ c t > 0
uniformly for ∣ ϕ ∣ ≤ C r |\phi|\le C_r ∣ ϕ ∣ ≤ C r (using 1 − cos u ≥ u 2 4 1-\cos u\ge\frac{u^2}{4} 1 − cos u ≥ 4 u 2 for ∣ u ∣ ≤ C r |u|\le C_r ∣ u ∣ ≤ C r ). Since R ^ \widehat{R} R is continuous at 0 0 0 and R ^ ( 0 ) > 0 \widehat{R}(0)>0 R ( 0 ) > 0 , there exists ε 0 > 0 \varepsilon_0>0 ε 0 > 0 such that
R ^ ( ξ ′ ) ≥ R ^ ( 0 ) 2 for ∣ ξ ′ ∣ ≤ ε 0 . \widehat{R}(\xi')\ge\frac{\widehat{R}(0)}{2}\qquad\text{for }|\xi'|\le\varepsilon_0. R ( ξ ′ ) ≥ 2 R ( 0 ) for ∣ ξ ′ ∣ ≤ ε 0 .
Shrinking ε 0 \varepsilon_0 ε 0 further so that ε 0 ≤ ε \varepsilon_0\le\varepsilon ε 0 ≤ ε , the near region ∣ ξ ′ ∣ ≤ ε |\xi'|\le\varepsilon ∣ ξ ′ ∣ ≤ ε is contained in ∣ ξ ′ ∣ ≤ ε 0 |\xi'|\le\varepsilon_0 ∣ ξ ′ ∣ ≤ ε 0 . Therefore,
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≥ c t c u 0 2 R ^ ( 0 ) 2 ∫ ∣ ξ ∣ ≤ r / 2 d ξ ∫ ∣ ξ ′ ∣ ≤ ε d ξ ′ = c t , u 0 ω d ( r 2 ) d ε d , \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^2\right]\ge c_t c_{u_0}^2\frac{\widehat{R}(0)}{2}\int_{|\xi|\le r/2}d\xi\int_{|\xi'|\le\varepsilon}d\xi'=c_{t,u_0}\omega_d\left(\frac{r}{2}\right)^d\varepsilon^d, E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≥ c t c u 0 2 2 R ( 0 ) ∫ ∣ ξ ∣ ≤ r /2 d ξ ∫ ∣ ξ ′ ∣ ≤ ε d ξ ′ = c t , u 0 ω d ( 2 r ) d ε d ,
where ω d \omega_d ω d is the volume of the unit ball.
Second order correction and sharpness
The Fourier computation in the proof of Proposition 3.4 identifies the leading diagonal contribution. Iterating the Duhamel formula once more,
w ε ( 2 ) ( t ) = ( − i ) 2 ∫ 0 t ∫ 0 r S ( t , r ) a ε ( r ) S ( r , s ) a ε ( s ) ( u ˉ ( s ) + w ε ( s ) ) d s d r . w_\varepsilon^{(2)}(t)=(-i)^2\int_0^t\int_0^r S(t,r)a_\varepsilon(r)S(r,s)a_\varepsilon(s)\bigl(\bar{u}(s)+w_\varepsilon(s)\bigr)\,ds\,dr. w ε ( 2 ) ( t ) = ( − i ) 2 ∫ 0 t ∫ 0 r S ( t , r ) a ε ( r ) S ( r , s ) a ε ( s ) ( u ˉ ( s ) + w ε ( s ) ) d s d r .
Taking expectation, the leading term is the deterministic mean
E [ w ε ( 2 ) ( t ) ] = − ∫ 0 t ∫ 0 r R ( r − s ε ) S ( t , r ) S ( r , s ) u ˉ ( s ) d s d r . \mathbb{E}[w_\varepsilon^{(2)}(t)]=-\int_0^t\int_0^r R\left(\frac{r-s}{\varepsilon}\right)S(t,r)S(r,s)\bar{u}(s)\,ds\,dr. E [ w ε ( 2 ) ( t )] = − ∫ 0 t ∫ 0 r R ( ε r − s ) S ( t , r ) S ( r , s ) u ˉ ( s ) d s d r .
This is the first moment (mean) of the second order term, not its second moment. Rescaling u = r − s ε u=\frac{r-s}{\varepsilon} u = ε r − s ,
E [ w ε ( 2 ) ( t ) ] = − ε ∫ 0 t / ε ∫ 0 r R ( u ) S ( t , r ) S ( r , r − ε u ) u ˉ ( r − ε u ) d u d r . \mathbb{E}[w_\varepsilon^{(2)}(t)]=-\varepsilon\int_0^{t/\varepsilon}\int_0^r R(u)S(t,r)S(r,r-\varepsilon u)\bar{u}(r-\varepsilon u)\,du\,dr. E [ w ε ( 2 ) ( t )] = − ε ∫ 0 t / ε ∫ 0 r R ( u ) S ( t , r ) S ( r , r − ε u ) u ˉ ( r − ε u ) d u d r .
The factor ε \varepsilon ε comes from the time-rescaling r − s = ε u r-s=\varepsilon u r − s = ε u and shows that the deterministic mean of the second order correction is O ( ε ) O(\varepsilon) O ( ε ) . The second moment E ∥ w ε ( 2 ) ∥ L x 2 2 \mathbb{E}\|w_\varepsilon^{(2)}\|_{L_x^2}^2 E ∥ w ε ( 2 ) ∥ L x 2 2 , which is the quantity entering the remainder estimate (36) , is smaller by an additional factor ε d \varepsilon^d ε d from the spatial covariance and is treated separately in the proof of Proposition 3.7 below.
Proposition 3.7. Write w ε = w ε ( 1 ) + r ε w_\varepsilon=w_\varepsilon^{(1)}+r_\varepsilon w ε = w ε ( 1 ) + r ε with r ε r_\varepsilon r ε satisfying (28) . Then,
∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C T , u 0 ε d / 2 , uniformly in t ∈ [ 0 , T ] . \|r_\varepsilon(t)\|_{L^2(\Omega;L_x^2)}\le C_{T,u_0}\varepsilon^{d/2},\qquad\text{uniformly in }t\in[0,T]. ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C T , u 0 ε d /2 , uniformly in t ∈ [ 0 , T ] .
Proof. From (28) ,
r ε ( t ) = − i ∫ 0 t S ( t , r ) a ε ( r ) ( w ε ( 1 ) ( r ) + r ε ( r ) ) d r . r_\varepsilon(t)=-i\int_0^t S(t,r)a_\varepsilon(r)\left(w_\varepsilon^{(1)}(r)+r_\varepsilon(r)\right)\,dr. r ε ( t ) = − i ∫ 0 t S ( t , r ) a ε ( r ) ( w ε ( 1 ) ( r ) + r ε ( r ) ) d r .
Taking L 2 ( Ω ; L x 2 ) L^2(\Omega;L_x^2) L 2 ( Ω ; L x 2 ) norms and using the unitarity of S ( t , r ) S(t,r) S ( t , r ) ,
∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) 2 ≤ 2 ∫ 0 t E [ ∥ a ε ( r ) w ε ( 1 ) ( r ) ∥ L x 2 2 ] d r + 2 ∫ 0 t E [ ∥ a ε ( r ) r ε ( r ) ∥ L x 2 2 ] d r . \|r_\varepsilon(t)\|_{L^2(\Omega;L_x^2)}^2\le2\int_0^t\mathbb{E}\left[\|a_\varepsilon(r)w_\varepsilon^{(1)}(r)\|_{L_x^2}^2\right]\,dr+2\int_0^t\mathbb{E}\left[\|a_\varepsilon(r)r_\varepsilon(r)\|_{L_x^2}^2\right]\,dr. ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) 2 ≤ 2 ∫ 0 t E [ ∥ a ε ( r ) w ε ( 1 ) ( r ) ∥ L x 2 2 ] d r + 2 ∫ 0 t E [ ∥ a ε ( r ) r ε ( r ) ∥ L x 2 2 ] d r .
By the sub-Gaussian moment bound in Assumption 1.4(i), the field a ε ( x ) = a ( x ε ) a_\varepsilon(x)=a\left(\frac{x}{\varepsilon}\right) a ε ( x ) = a ( ε x ) has uniform spatial moments. For any x ∈ R d x\in\mathbb{R}^d x ∈ R d and any p < ∞ p<\infty p < ∞ , E [ ∣ a ε ( x ) ∣ p ] ≤ C p \mathbb{E}[|a_\varepsilon(x)|^p]\le C_p E [ ∣ a ε ( x ) ∣ p ] ≤ C p uniformly in ε \varepsilon ε . Applying Hölder’s inequality in Ω \Omega Ω and then Cauchy–Schwarz in x x x ,
E [ ∥ a ε ( r ) f ( r ) ∥ L x 2 2 ] = ∫ R d E [ ∣ a ε ( x ) ∣ 2 ∣ f ( r , x ) ∣ 2 ] d x ≤ C ∥ f ( r ) ∥ L 4 ( Ω ; L x 2 ) 2 , \mathbb{E}\left[\|a_\varepsilon(r)f(r)\|_{L_x^2}^2\right]=\int_{\mathbb{R}^d}\mathbb{E}[|a_\varepsilon(x)|^2|f(r,x)|^2]\,dx\le C\|f(r)\|_{L^4(\Omega;L^2_x)}^2, E [ ∥ a ε ( r ) f ( r ) ∥ L x 2 2 ] = ∫ R d E [ ∣ a ε ( x ) ∣ 2 ∣ f ( r , x ) ∣ 2 ] d x ≤ C ∥ f ( r ) ∥ L 4 ( Ω ; L x 2 ) 2 ,
where the constant absorbs the fixed spatial support of u ˉ \bar u u ˉ and the uniform L 4 ( Ω ) L^4(\Omega) L 4 ( Ω ) bound on a ε ( x ) a_\varepsilon(x) a ε ( x ) . The L 4 ( Ω ; L x 2 ) L^4(\Omega;L_x^2) L 4 ( Ω ; L x 2 ) control of f f f follows from the Wick cumulant estimates in Lemma 3.10 (whose q = 2 q=2 q = 2 case is used here and is proved in Subsection 3.3 below) and the moment bound in Lemma 2.13. For f = w ε ( 1 ) ( r ) f=w_\varepsilon^{(1)}(r) f = w ε ( 1 ) ( r ) , by Lemma 3.10,
E [ ∥ w ε ( 1 ) ( r ) ∥ L x 2 2 ] ≤ C ε d , \mathbb{E}\left[\|w_\varepsilon^{(1)}(r)\|_{L_x^2}^2\right]\le C\varepsilon^d, E [ ∥ w ε ( 1 ) ( r ) ∥ L x 2 2 ] ≤ C ε d ,
and the L 4 L^4 L 4 bound is of the same order by the moment bound, so the first integral is O ( ε d ) O(\varepsilon^d) O ( ε d ) . For f = r ε ( r ) f=r_\varepsilon(r) f = r ε ( r ) ,
∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) 2 ≤ C T , u 0 ε d + C ∫ 0 t ∥ r ε ( r ) ∥ L 2 ( Ω ; L x 2 ) 2 d r . \|r_\varepsilon(t)\|_{L^2(\Omega;L_x^2)}^2\le C_{T,u_0}\varepsilon^d+C\int_0^t\|r_\varepsilon(r)\|_{L^2(\Omega;L_x^2)}^2\,dr. ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) 2 ≤ C T , u 0 ε d + C ∫ 0 t ∥ r ε ( r ) ∥ L 2 ( Ω ; L x 2 ) 2 d r .
By Gronwall’s inequality (Lemma 2.1),
∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) 2 ≤ C T , u 0 ε d e C t ≤ C T , u 0 ε d . \|r_\varepsilon(t)\|_{L^2(\Omega;L_x^2)}^2\le C_{T,u_0}\varepsilon^d e^{Ct}\le C_{T,u_0}\varepsilon^d. ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) 2 ≤ C T , u 0 ε d e C t ≤ C T , u 0 ε d .
Taking square roots gives the result. □
The lower bound (35) and the upper bound from Proposition 3.4 together give
c t , u 0 ε d ≤ E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≤ C T , u 0 ε d . c_{t,u_0}\varepsilon^d\le\mathbb{E}\left[\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^2\right]\le C_{T,u_0}\varepsilon^d. c t , u 0 ε d ≤ E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≤ C T , u 0 ε d .
Remark 3.8. The diagonal lower bound has a direct block average interpretation. Splitting R d \mathbb{R}^d R d into cubes Q k Q_k Q k of side ε \varepsilon ε ,
w ε ( 1 ) ( t ) ≈ − i ∑ k ε d S ( t , r ) ( ε − d ∫ Q k a ( y ε ) d y ) u ˉ ( r ) d r . w_\varepsilon^{(1)}(t)\approx-i\sum_k\varepsilon^dS(t,r)\left(\varepsilon^{-d}\int_{Q_k}a\left(\frac{y}{\varepsilon}\right)\,dy\right)\bar u(r)\,dr. w ε ( 1 ) ( t ) ≈ − i k ∑ ε d S ( t , r ) ( ε − d ∫ Q k a ( ε y ) d y ) u ˉ ( r ) d r . Each block average has variance ε − d R ^ ( 0 ) + o ( ε − d ) \varepsilon^{-d}\widehat R(0)+o(\varepsilon^{-d}) ε − d R ( 0 ) + o ( ε − d ) , and there are ε − d \varepsilon^{-d} ε − d independent blocks in the support of u ˉ \bar u u ˉ . Therefore, the total variance is
ε − d × ( ε d ) 2 × ε − d R ^ ( 0 ) = R ^ ( 0 ) ε d , \varepsilon^{-d}\times(\varepsilon^d)^2\times\varepsilon^{-d}\widehat R(0)=\widehat R(0)\varepsilon^d, ε − d × ( ε d ) 2 × ε − d R ( 0 ) = R ( 0 ) ε d , which is exactly the diagonal contribution of Proposition 3.5. The off diagonal terms ∣ ξ ′ ∣ > 0 |\xi'|>0 ∣ ξ ′ ∣ > 0 contribute O ( ε d + δ ) O(\varepsilon^{d+\delta}) O ( ε d + δ ) by the exponential decay of R ^ \widehat R R , and are negligible.
Proposition 3.9. Assume R ^ ( 0 ) > 0 \widehat R(0)>0 R ( 0 ) > 0 and u ^ 0 ( 0 ) ≠ 0 \widehat u_0(0)\ne0 u 0 ( 0 ) = 0 . There exist c t , u 0 > 0 c_{t,u_0}>0 c t , u 0 > 0 and ε 0 > 0 \varepsilon_0>0 ε 0 > 0 such that
E [ ∥ w ε ( t ) ∥ L x 2 2 ] ≥ c t , u 0 ε d for all 0 < ε ≤ ε 0 . \mathbb{E}\left[\|w_\varepsilon(t)\|_{L_x^2}^2\right]\ge c_{t,u_0}\varepsilon^d\qquad\text{for all }0<\varepsilon\le\varepsilon_0. E [ ∥ w ε ( t ) ∥ L x 2 2 ] ≥ c t , u 0 ε d for all 0 < ε ≤ ε 0 .
Proof. We begin with the Duhamel iteration r ε = w ε ( 2 ) + w ε ( 3 ) + ⋯ r_\varepsilon= w_\varepsilon^{(2)} + w_\varepsilon^{(3)} + \cdots r ε = w ε ( 2 ) + w ε ( 3 ) + ⋯ and aim to prove that
E [ ∥ w ε ( 2 ) ( t ) ∥ L x 2 2 ] ≤ C T , u 0 ε d + 1 . (36) \mathbb{E}\left[\left\|w_\varepsilon^{(2)}(t)\right\|_{L_x^2}^2\right] \le C_{T,u_0}\varepsilon^{d+1}. \tag*{(36)} E [ w ε ( 2 ) ( t ) L x 2 2 ] ≤ C T , u 0 ε d + 1 . (36)
The second order term is
w ε ( 2 ) ( t ) = ( − i ) 2 ∫ 0 t ∫ 0 r S ( t , r ) a ε ( r ) S ( r , s ) a ε ( s ) u ‾ ( s ) d s d r . w_\varepsilon^{(2)}(t)=(-i)^2\int_0^t\int_0^r S(t,r)a_\varepsilon(r)S(r,s)a_\varepsilon(s)\overline{u}(s)\,ds\,dr. w ε ( 2 ) ( t ) = ( − i ) 2 ∫ 0 t ∫ 0 r S ( t , r ) a ε ( r ) S ( r , s ) a ε ( s ) u ( s ) d s d r .
Taking the Fourier transform and using S ( t , r ) ξ = e − i ( ∣ ξ ∣ 2 + V ‾ ) ( t − r ) S(t,r)\xi=e^{-i(|\xi|^2+\overline{V})(t-r)} S ( t , r ) ξ = e − i ( ∣ ξ ∣ 2 + V ) ( t − r ) , and noting that a ε ^ ( k ) = ε d a ^ ( ε k ) \widehat{a_\varepsilon}(k)=\varepsilon^d\widehat{a}(\varepsilon k) a ε ( k ) = ε d a ( ε k ) for a ε ( x ) = a ( x ε ) a_\varepsilon(x)=a\left(\frac{x}{\varepsilon}\right) a ε ( x ) = a ( ε x ) ,
w ε ( 2 ) ^ ( t , ξ ) = − ( 2 π ) − 2 d ε 2 d ∫ 0 t ∫ 0 r ∫ R 2 d e − i ( ∣ ξ ∣ 2 + V ‾ ) ( t − r ) a ^ ( ε ( ξ − η ) ) e − i ( ∣ η ∣ 2 + V ‾ ) ( r − s ) a ^ ( ε ( η − ζ ) ) e − i ( ∣ ζ ∣ 2 + V ‾ ) s u ^ 0 ( ζ ) d ζ d η d s d r . (37) \widehat{w_\varepsilon^{(2)}}(t,\xi)=-(2\pi)^{-2d}\varepsilon^{2d}\int_0^t\int_0^r\int_{\mathbb{R}^{2d}}e^{-i(|\xi|^2+\overline{V})(t-r)}\widehat{a}(\varepsilon(\xi-\eta))
e^{-i(|\eta|^2+\overline{V})(r-s)}\widehat{a}(\varepsilon(\eta-\zeta))e^{-i(|\zeta|^2+\overline{V})s}\widehat{u}_0(\zeta)\,d\zeta\,d\eta\,ds\,dr. \tag*{(37)} w ε ( 2 ) ( t , ξ ) = − ( 2 π ) − 2 d ε 2 d ∫ 0 t ∫ 0 r ∫ R 2 d e − i ( ∣ ξ ∣ 2 + V ) ( t − r ) a ( ε ( ξ − η )) e − i ( ∣ η ∣ 2 + V ) ( r − s ) a ( ε ( η − ζ )) e − i ( ∣ ζ ∣ 2 + V ) s u 0 ( ζ ) d ζ d η d s d r . (37)
By Plancherel,
E [ ∥ w ε ( 2 ) ( t ) ∥ L x 2 2 ] = ( 2 π ) − d ∫ R d E [ ∣ w ε ( 2 ) ^ ( t , ξ ) ∣ 2 ] d ξ . \mathbb{E}\left[\left\|w_\varepsilon^{(2)}(t)\right\|_{L_x^2}^2\right]=(2\pi)^{-d}\int_{\mathbb{R}^d}\mathbb{E}\left[\left|\widehat{w_\varepsilon^{(2)}}(t,\xi)\right|^2\right]\,d\xi. E [ w ε ( 2 ) ( t ) L x 2 2 ] = ( 2 π ) − d ∫ R d E [ w ε ( 2 ) ( t , ξ ) 2 ] d ξ .
Inserting (37) and its complex conjugate, the expectation contains four centered field factors a ^ ( ε ( ξ − η ) ) \widehat{a}(\varepsilon(\xi-\eta)) a ( ε ( ξ − η )) , a ^ ( ε ( η − ζ ) ) ‾ \overline{\widehat{a}(\varepsilon(\eta-\zeta))} a ( ε ( η − ζ )) , a ^ ( ε ( ξ − η ′ ) ) ‾ \overline{\widehat{a}(\varepsilon(\xi-\eta'))} a ( ε ( ξ − η ′ )) , a ^ ( ε ( η ′ − ζ ′ ) ) \widehat{a}(\varepsilon(\eta'-\zeta')) a ( ε ( η ′ − ζ ′ )) , integrated over 0 ≤ s ≤ r ≤ t 0\le s\le r\le t 0 ≤ s ≤ r ≤ t , 0 ≤ s ′ ≤ r ′ ≤ t 0\le s'\le r'\le t 0 ≤ s ′ ≤ r ′ ≤ t , and ζ , ζ ′ , η , η ′ ∈ R d \zeta,\zeta',\eta,\eta'\in\mathbb{R}^d ζ , ζ ′ , η , η ′ ∈ R d . By the Isserlis theorem (Lemma 2.9 ), the fourth moment expands over the three pairings. The leading contribution is the diagonal pairing
E [ a ^ ( ε ( ξ − η ) ) a ^ ( ε ( ξ − η ′ ) ) ‾ ] E [ a ^ ( ε ( η − ζ ) ) ‾ a ^ ( ε ( η ′ − ζ ′ ) ) ] , (38) \mathbb{E}\left[\widehat{a}(\varepsilon(\xi-\eta))\overline{\widehat{a}(\varepsilon(\xi-\eta'))}\right]\mathbb{E}\left[\overline{\widehat{a}(\varepsilon(\eta-\zeta))}\widehat{a}(\varepsilon(\eta'-\zeta'))\right], \tag*{(38)} E [ a ( ε ( ξ − η )) a ( ε ( ξ − η ′ )) ] E [ a ( ε ( η − ζ )) a ( ε ( η ′ − ζ ′ )) ] , (38)
which by (30) equals
( 2 π ) d ε − d δ ( η − η ′ ) R ^ ( ε ( ξ − η ) ) ( 2 π ) d ε − d δ ( ζ − ζ ′ ) R ^ ( ε ( η − ζ ′ ) ) . (2\pi)^d\varepsilon^{-d}\delta(\eta-\eta')\widehat{R}(\varepsilon(\xi-\eta))(2\pi)^d\varepsilon^{-d}\delta(\zeta-\zeta')\widehat{R}(\varepsilon(\eta-\zeta')). ( 2 π ) d ε − d δ ( η − η ′ ) R ( ε ( ξ − η )) ( 2 π ) d ε − d δ ( ζ − ζ ′ ) R ( ε ( η − ζ ′ )) .
Together with the prefactor ε 4 d \varepsilon^{4d} ε 4 d from two factors of ε 2 d \varepsilon^{2d} ε 2 d in (37) and its conjugate, the net spatial prefactor before the change of variables is ε 4 d ⋅ ε − 2 d = ε 2 d \varepsilon^{4d}\cdot\varepsilon^{-2d}=\varepsilon^{2d} ε 4 d ⋅ ε − 2 d = ε 2 d .
By δ ( η − η ′ ) δ ( ζ − ζ ′ ) \delta(\eta-\eta')\delta(\zeta-\zeta') δ ( η − η ′ ) δ ( ζ − ζ ′ ) , we have η = η ′ \eta=\eta' η = η ′ , ζ = ζ ′ \zeta=\zeta' ζ = ζ ′ . Rescale α = ε ( ξ − η ) \alpha=\varepsilon(\xi-\eta) α = ε ( ξ − η ) , β = ε ( η − ζ ) \beta=\varepsilon(\eta-\zeta) β = ε ( η − ζ ) , so that d η d ζ = ε − 2 d d α d β d\eta\,d\zeta=\varepsilon^{-2d}d\alpha\,d\beta d η d ζ = ε − 2 d d α d β . The net prefactor is therefore ε 2 d ⋅ ε − 2 d = O ( 1 ) \varepsilon^{2d}\cdot\varepsilon^{-2d}=O(1) ε 2 d ⋅ ε − 2 d = O ( 1 ) . The remaining integral is
∫ R d ∣ u ^ 0 ( ζ ) ∣ 2 d ζ ∫ R d ∫ R d R ^ ( α ) R ^ ( β ) 2 ( 1 − cos ( t ω 1 ) ) ω 1 2 2 ( 1 − cos ( t ω 2 ) ) ω 2 2 d α d β , \int_{\mathbb{R}^d}|\widehat{u}_0(\zeta)|^2\,d\zeta\int_{\mathbb{R}^d}\int_{\mathbb{R}^d}\widehat{R}(\alpha)\widehat{R}(\beta)\frac{2(1-\cos(t\omega_1))}{\omega_1^2}\frac{2(1-\cos(t\omega_2))}{\omega_2^2}\,d\alpha\,d\beta, ∫ R d ∣ u 0 ( ζ ) ∣ 2 d ζ ∫ R d ∫ R d R ( α ) R ( β ) ω 1 2 2 ( 1 − cos ( t ω 1 )) ω 2 2 2 ( 1 − cos ( t ω 2 )) d α d β ,
where
ω 1 = 2 η ⋅ α ε + ∣ α ∣ 2 ε 2 , ω 2 = 2 ζ ⋅ β ε + ∣ β ∣ 2 ε 2 . \omega_1=\frac{2\eta\cdot\alpha}{\varepsilon}+\frac{|\alpha|^2}{\varepsilon^2},\qquad\omega_2=\frac{2\zeta\cdot\beta}{\varepsilon}+\frac{|\beta|^2}{\varepsilon^2}. ω 1 = ε 2 η ⋅ α + ε 2 ∣ α ∣ 2 , ω 2 = ε 2 ζ ⋅ β + ε 2 ∣ β ∣ 2 .
On the near region ∣ α ∣ ≤ ε |\alpha|\le\varepsilon ∣ α ∣ ≤ ε , ∣ β ∣ ≤ ε |\beta|\le\varepsilon ∣ β ∣ ≤ ε , both time factors are O ( 1 ) O(1) O ( 1 ) and the volume is O ( ε 2 d ) O(\varepsilon^{2d}) O ( ε 2 d ) . On the far region ∣ α ∣ > ε |\alpha|>\varepsilon ∣ α ∣ > ε or ∣ β ∣ > ε |\beta|>\varepsilon ∣ β ∣ > ε , the bound 1 − cos u ≤ 2 1-\cos u\le2 1 − cos u ≤ 2 gives
2 ( 1 − cos ( t ω ) ) ω 2 ≤ 4 ε 4 ∣ α ∣ 4 , \frac{2(1-\cos(t\omega))}{\omega^2}\le\frac{4\varepsilon^4}{|\alpha|^4}, ω 2 2 ( 1 − cos ( t ω )) ≤ ∣ α ∣ 4 4 ε 4 ,
and the dyadic sum converges geometrically for d ≤ 3 d\le3 d ≤ 3 , giving the same order O ( ε 2 d ) O(\varepsilon^{2d}) O ( ε 2 d ) . The d ≥ 4 d\ge4 d ≥ 4 case is handled by the same physical space argument as in Appendix § 8 . Hence,
diagonal pairing ≤ C T , u 0 ε 2 d ≤ C T , u 0 ε d + 1 , \text{diagonal pairing}\le C_{T,u_0}\varepsilon^{2d}\le C_{T,u_0}\varepsilon^{d+1}, diagonal pairing ≤ C T , u 0 ε 2 d ≤ C T , u 0 ε d + 1 ,
where the last inequality uses 2 d ≥ d + 1 2d\ge d+1 2 d ≥ d + 1 for d ≥ 1 d\ge1 d ≥ 1 . The cross pairings are controlled by (1.9). This proves (3.9).
The k k k -th iterate w ε ( k ) w_\varepsilon^{(k)} w ε ( k ) contains k k k insertions of a ε a_\varepsilon a ε , giving
E [ ∥ w ε ( k ) ( t ) ∥ L x 2 2 ] ≤ C k ε k d . \mathbb{E}\left[\left\|w_\varepsilon^{(k)}(t)\right\|_{L_x^2}^2\right]\le C^k\varepsilon^{kd}. E [ w ε ( k ) ( t ) L x 2 2 ] ≤ C k ε k d .
Since k d ≥ d + k − 1 kd \ge d+k-1 k d ≥ d + k − 1 for k ≥ 2 k \ge2 k ≥ 2 ,
E [ ∥ w ε ( k ) ( t ) ∥ L x 2 2 ] ≤ C k ε d + k − 1 . \mathbb{E}\left[\|w_\varepsilon^{(k)}(t)\|_{L_x^2}^2\right] \le C^k\varepsilon^{d+k-1}. E [ ∥ w ε ( k ) ( t ) ∥ L x 2 2 ] ≤ C k ε d + k − 1 .
Summing over k ≥ 2 k \ge2 k ≥ 2 ,
∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) 2 ≤ ∑ k ≥ 2 C k ε d + k − 1 ≤ C T , u 0 ε d + 1 . (39) \|r_\varepsilon(t)\|_{L^2(\Omega;L_x^2)}^2 \le\sum_{k\ge2} C^k\varepsilon^{d+k-1} \le C_{T,u_0}\varepsilon^{d+1}. \tag*{(39)} ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) 2 ≤ k ≥ 2 ∑ C k ε d + k − 1 ≤ C T , u 0 ε d + 1 . (39)
By the reverse triangle inequality,
∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≥ ∥ w ε ( 1 ) ( t ) ∥ L 2 ( Ω ; L x 2 ) − ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) . \|w_\varepsilon(t)\|_{L^2(\Omega;L_x^2)} \ge\|w_\varepsilon^{(1)}(t)\|_{L^2(\Omega;L_x^2)}-\|r_\varepsilon(t)\|_{L^2(\Omega;L_x^2)}. ∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≥ ∥ w ε ( 1 ) ( t ) ∥ L 2 ( Ω ; L x 2 ) − ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) .
Combining this with the lower bound (35) and the remainder estimate (39) ,
∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≥ c t , u 0 ε d / 2 − C T , u 0 ε ( d + 1 ) / 2 ≥ 1 2 c t , u 0 ε d / 2 . \|w_\varepsilon(t)\|_{L^2(\Omega;L_x^2)} \ge c_{t,u_0}\varepsilon^{d/2}-C_{T,u_0}\varepsilon^{(d+1)/2} \ge\frac{1}{2}c_{t,u_0}\varepsilon^{d/2}. ∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≥ c t , u 0 ε d /2 − C T , u 0 ε ( d + 1 ) /2 ≥ 2 1 c t , u 0 ε d /2 .
Squaring and taking expectation gives
E [ ∥ w ε ( t ) ∥ L x 2 2 ] ≥ 1 4 c t , u 0 2 ε d . (40) \mathbb{E}\left[\|w_\varepsilon(t)\|_{L_x^2}^2\right] \ge\frac{1}{4}c_{t,u_0}^2\varepsilon^d. \tag*{(40)} E [ ∥ w ε ( t ) ∥ L x 2 2 ] ≥ 4 1 c t , u 0 2 ε d . (40)
which completes the proof. □
Combining the upper bound from Proposition 3.4 and Proposition 3.7 via the triangle inequality,
∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ ∥ w ε ( 1 ) ( t ) ∥ L 2 ( Ω ; L x 2 ) + ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C T , u 0 ε d / 2 , \|w_\varepsilon(t)\|_{L^2(\Omega;L_x^2)} \le\|w_\varepsilon^{(1)}(t)\|_{L^2(\Omega;L_x^2)}+\|r_\varepsilon(t)\|_{L^2(\Omega;L_x^2)} \le C_{T,u_0}\varepsilon^{d/2}, ∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ ∥ w ε ( 1 ) ( t ) ∥ L 2 ( Ω ; L x 2 ) + ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C T , u 0 ε d /2 ,
with the matching lower bound from Proposition 3.9, we have established
c t , u 0 ε d ≤ E [ ∥ w ε ( t ) ∥ L x 2 2 ] ≤ C T , u 0 ε d , c_{t,u_0}\varepsilon^d \le\mathbb{E}\left[\|w_\varepsilon(t)\|_{L_x^2}^2\right] \le C_{T,u_0}\varepsilon^d, c t , u 0 ε d ≤ E [ ∥ w ε ( t ) ∥ L x 2 2 ] ≤ C T , u 0 ε d ,
which proves part (i) of Theorem 1.5: the L 2 ( Ω ; L 2 ) L^2(\Omega;L^2) L 2 ( Ω ; L 2 ) convergence rate ε d / 2 \varepsilon^{d/2} ε d /2 and its sharpness.
All moments and the rate
The second moment bound extends to all moments via a cumulant expansion. For Gaussian fields, this reduces to the exact Wick pairing. For general sub-Gaussian fields, the higher cumulants are controlled by the exponential mixing assumption.
Lemma 3.10. For every integer q ≥ 1 q \ge1 q ≥ 1 ,
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] ≤ ( C q , T , u 0 ) q ε d q . (41) \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^{2q}\right] \le(C_{q,T,u_0})^q\varepsilon^{dq}. \tag*{(41)} E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] ≤ ( C q , T , u 0 ) q ε d q . (41)
Proof. By Plancherel,
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] = ( 2 π ) − d q ∫ R d q [ E ∏ j = 1 q w ^ ε ( 1 ) ( t , ξ j ) w ^ ε ( 1 ) ( t , ξ j ) ] d ξ 1 ⋯ d ξ q . \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^{2q}\right] = (2\pi)^{-dq}\int_{\mathbb{R}^{dq}}\left[\mathbb{E}\prod_{j=1}^{q}\widehat{w}_\varepsilon^{(1)}(t,\xi_j)\widehat{w}_\varepsilon^{(1)}(t,\xi_j)\right]\,d\xi_1\cdots d\xi_q. E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] = ( 2 π ) − d q ∫ R d q [ E j = 1 ∏ q w ε ( 1 ) ( t , ξ j ) w ε ( 1 ) ( t , ξ j ) ] d ξ 1 ⋯ d ξ q .
By the cumulant expansion,
E [ ∏ j = 1 2 q X j ] = ∑ π ∈ P 2 ( 2 q ) ∏ { j , k } ∈ π E [ X j X k ] + ∑ m ≥ 3 ∑ π m ∏ blocks κ m ( ⋅ ) . \mathbb{E}\left[\prod_{j=1}^{2q}X_j\right]
=
\sum_{\pi\in P_2(2q)}\prod_{\{j,k\}\in\pi}\mathbb{E}[X_jX_k]
+
\sum_{m\ge3}\sum_{\pi_m}\prod_{\text{blocks}}\kappa_m(\cdot). E [ j = 1 ∏ 2 q X j ] = π ∈ P 2 ( 2 q ) ∑ { j , k } ∈ π ∏ E [ X j X k ] + m ≥ 3 ∑ π m ∑ blocks ∏ κ m ( ⋅ ) .
The number of pairings is
∣ P 2 ( 2 q ) ∣ = ( 2 q − 1 ) ! ! = ( 2 q ) ! 2 q q ! ≤ ( C q ) q . |P_2(2q)|=(2q-1)!!=\frac{(2q)!}{2^q q!}\le(Cq)^q. ∣ P 2 ( 2 q ) ∣ = ( 2 q − 1 )!! = 2 q q ! ( 2 q )! ≤ ( C q ) q .
For each pair { j , k } \{j,k\} { j , k } ,
ε d ( 2 π ) − d ∫ R d ∣ R ^ ( ε η ) ∣ ∣ ∫ 0 t e i r σ ( ξ j , η ) d r ∣ 2 ∣ u ^ 0 ( ξ j − η ) ∣ 2 d η ≤ C T , u 0 ε d . \varepsilon^d(2\pi)^{-d}\int_{\mathbb{R}^d}\left|\widehat{R}(\varepsilon\eta)\right|\left|\int_0^t e^{ir\sigma(\xi_j,\eta)}\,dr\right|^2|\widehat{u}_0(\xi_j-\eta)|^2\,d\eta\le C_{T,u_0}\varepsilon^d. ε d ( 2 π ) − d ∫ R d R ( ε η ) ∫ 0 t e i r σ ( ξ j , η ) d r 2 ∣ u 0 ( ξ j − η ) ∣ 2 d η ≤ C T , u 0 ε d .
Thus the Gaussian Wick sum contributes
∑ π ∈ P 2 ( 2 q ) ∏ { j , k } ∈ π E [ X j X k ] ≤ ( 2 q − 1 ) ! ! ( C T , u 0 ε d ) q ≤ ( C q , T , u 0 ) q ε d q . \sum_{\pi\in P_2(2q)}\prod_{\{j,k\}\in\pi}\mathbb{E}[X_jX_k]\le(2q-1)!!\left(C_{T,u_0}\varepsilon^{d}\right)^q\le(C_{q,T,u_0})^q\varepsilon^{dq}. π ∈ P 2 ( 2 q ) ∑ { j , k } ∈ π ∏ E [ X j X k ] ≤ ( 2 q − 1 )!! ( C T , u 0 ε d ) q ≤ ( C q , T , u 0 ) q ε d q .
For non-Gaussian cumulants of order m ≥ 3 m\ge3 m ≥ 3 , the standard cluster expansion under exponential mixing [10 Sec. 4] gives
∣ κ m ( X 1 , … , X m ) ∣ ≤ C m ∏ j = 1 m ∥ X j ∥ L m e − κ sep ( D 1 , … , D m ) . \left|\kappa_m(X_1,\ldots,X_m)\right|\le C_m\prod_{j=1}^{m}\|X_j\|_{L^m}e^{-\kappa\operatorname{sep}(D_1,\ldots,D_m)}. ∣ κ m ( X 1 , … , X m ) ∣ ≤ C m j = 1 ∏ m ∥ X j ∥ L m e − κ sep ( D 1 , … , D m ) .
where sep ( D 1 , … , D m ) \operatorname{sep}(D_1,\ldots,D_m) sep ( D 1 , … , D m ) is the minimal separation between the supports of the variables. This bound follows by iterating the mixing inequality (9) over the dependency graph of the cumulant. Integrating over separations,
∫ 0 ∞ e − κ R R d − 1 d R = ( d − 1 ) ! κ d < ∞ . \int_0^\infty e^{-\kappa R}R^{d-1}\,\mathrm{d}R=\frac{(d-1)!}{\kappa^d}<\infty. ∫ 0 ∞ e − κ R R d − 1 d R = κ d ( d − 1 )! < ∞.
Choose R ε = ∣ log ε ∣ κ R_\varepsilon=\frac{|\log\varepsilon|}{\kappa} R ε = κ ∣ l o g ε ∣ . Then,
∫ R ε ∞ e − κ R R d − 1 d R = O ( ε ∣ log ε ∣ d − 1 ) = O ( ε ) . \int_{R_\varepsilon}^\infty e^{-\kappa R}R^{d-1}\,\mathrm{d}R=O\left(\varepsilon|\log\varepsilon|^{d-1}\right)=O(\varepsilon). ∫ R ε ∞ e − κ R R d − 1 d R = O ( ε ∣ log ε ∣ d − 1 ) = O ( ε ) .
Each non-pairing contraction introduces one additional mixing factor O ( ε ) O(\varepsilon) O ( ε ) and reduces the pairing count by one, so
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] ≤ ( C q , T , u 0 ) q ε d q + O ( ε d q + d ) ≤ ( C q , T , u 0 ) q ε d q . \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^{2q}\right]\le(C_{q,T,u_0})^q\varepsilon^{dq}+O\left(\varepsilon^{dq+d}\right)\le(C_{q,T,u_0})^q\varepsilon^{dq}. E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] ≤ ( C q , T , u 0 ) q ε d q + O ( ε d q + d ) ≤ ( C q , T , u 0 ) q ε d q .
This proves Lemma 3.10 . □
Combining the Wick estimate with the remainder estimate gives the convergence rate.
Proposition 3.11. Under Assumption 1.4 ,
sup 0 ≤ t ≤ T ∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C T , u 0 ε d 2 . (42) \sup_{0\le t\le T}\|w_\varepsilon(t)\|_{L^2(\Omega;L_x^2)}\le C_{T,u_0}\varepsilon^{\frac{d}{2}}. \tag*{(42)} 0 ≤ t ≤ T sup ∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C T , u 0 ε 2 d . (42) Moreover, for every finite m ≥ 1 m\ge1 m ≥ 1 ,
lim ε → 0 E [ sup 0 ≤ t ≤ T ∥ w ε ( t ) ∥ L x 2 m ] = 0. \lim_{\varepsilon\to0}\mathbb{E}\left[\sup_{0\le t\le T}\|w_\varepsilon(t)\|_{L_x^2}^{m}\right]=0. ε → 0 lim E [ 0 ≤ t ≤ T sup ∥ w ε ( t ) ∥ L x 2 m ] = 0.
Proof. By the Duhamel formula,
w ε ( t ) = − i ∫ 0 t S ( t , r ) a ε ( r ) u ‾ ( r ) d r − i ∫ 0 t S ( t , r ) a ε ( r ) w ε ( r ) d r . w_\varepsilon(t)=-i\int_0^t S(t,r)a_\varepsilon(r)\overline{u}(r)\,\mathrm{d}r-i\int_0^t S(t,r)a_\varepsilon(r)w_\varepsilon(r)\,\mathrm{d}r. w ε ( t ) = − i ∫ 0 t S ( t , r ) a ε ( r ) u ( r ) d r − i ∫ 0 t S ( t , r ) a ε ( r ) w ε ( r ) d r .
Taking L 2 ( Ω ; L x 2 ) L^2(\Omega;L_x^2) L 2 ( Ω ; L x 2 ) norms and using the unitarity of S ( t , r ) S(t,r) S ( t , r ) ,
∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ ∥ w ε ( 1 ) ( t ) ∥ L 2 ( Ω ; L x 2 ) + ∫ 0 t ∥ a ε ( r ) w ε ( r ) ∥ L 2 ( Ω ; L x 2 ) d r . \|w_\varepsilon(t)\|_{L^2(\Omega;L_x^2)}\le\|w_\varepsilon^{(1)}(t)\|_{L^2(\Omega;L_x^2)}+\int_0^t\|a_\varepsilon(r)w_\varepsilon(r)\|_{L^2(\Omega;L_x^2)}\,\mathrm{d}r. ∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ ∥ w ε ( 1 ) ( t ) ∥ L 2 ( Ω ; L x 2 ) + ∫ 0 t ∥ a ε ( r ) w ε ( r ) ∥ L 2 ( Ω ; L x 2 ) d r .
By Lemma 2.13 , ∥ a ε ( r ) w ε ( r ) ∥ L 2 ( Ω ; L x 2 ) ≤ C ∥ w ε ( r ) ∥ L 2 ( Ω ; L x 2 ) \|a_\varepsilon(r)w_\varepsilon(r)\|_{L^2(\Omega;L_x^2)}\le C\|w_\varepsilon(r)\|_{L^2(\Omega;L_x^2)} ∥ a ε ( r ) w ε ( r ) ∥ L 2 ( Ω ; L x 2 ) ≤ C ∥ w ε ( r ) ∥ L 2 ( Ω ; L x 2 ) . Hence,
∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C T , u 0 ε d 2 + C ∫ 0 t ∥ w ε ( r ) ∥ L 2 ( Ω ; L x 2 ) d r . \|w_\varepsilon(t)\|_{L^2(\Omega;L_x^2)}\le C_{T,u_0}\varepsilon^{\frac{d}{2}}+C\int_0^t\|w_\varepsilon(r)\|_{L^2(\Omega;L_x^2)}\,\mathrm{d}r. ∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C T , u 0 ε 2 d + C ∫ 0 t ∥ w ε ( r ) ∥ L 2 ( Ω ; L x 2 ) d r .
Applying Gronwall’s inequality (Lemma 2.1 ),
∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C T , u 0 ε d 2 e C t ≤ C T , u 0 ε d 2 . \|w_\varepsilon(t)\|_{L^2(\Omega;L_x^2)}\le C_{T,u_0}\varepsilon^{\frac{d}{2}}e^{Ct}\le C_{T,u_0}\varepsilon^{\frac{d}{2}}. ∥ w ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C T , u 0 ε 2 d e C t ≤ C T , u 0 ε 2 d .
The L m ( Ω ) L^m(\Omega) L m ( Ω ) estimate follows by applying Lemma 3.10 with 2 q = m 2q=m 2 q = m and the same Gronwall argument in L m ( Ω ; L x 2 ) L^m(\Omega;L_x^2) L m ( Ω ; L x 2 ) ,
E [ sup 0 ≤ t ≤ T ∥ w ε ( t ) ∥ L x 2 m ] ≤ C m , T , u 0 ε m d 2 . \mathbb{E}\left[\sup_{0\le t\le T}\|w_\varepsilon(t)\|_{L_x^2}^m\right]\le C_{m,T,u_0}\varepsilon^{\frac{md}{2}}. E [ 0 ≤ t ≤ T sup ∥ w ε ( t ) ∥ L x 2 m ] ≤ C m , T , u 0 ε 2 m d .
This proves Proposition 3.11.
This proves part (ii) of Theorem 1.5.
Almost sure convergence
The pointwise in time rate upgrades to a supremum estimate via the Kolmogorov maximal inequality.
Proposition 4.1. For every integer q ≥ 1 q\ge1 q ≥ 1 ,
E [ sup 0 ≤ t ≤ T ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] ≤ ( C q , T , u 0 ) q ε d q . \mathbb{E}\left[\sup_{0\le t\le T}\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^{2q}\right]\le(C_{q,T,u_0})^q\varepsilon^{dq}. E [ 0 ≤ t ≤ T sup ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] ≤ ( C q , T , u 0 ) q ε d q .
Proof. The process w ε ( 1 ) ( t ) w_\varepsilon^{(1)}(t) w ε ( 1 ) ( t ) is continuous in t t t by dominated convergence. For s < t s<t s < t ,
w ε ( 1 ) ( t ) − w ε ( 1 ) ( s ) = − i ∫ s t S ( t , r ) a ε ( r ) u ˉ ( r ) d r − i ∫ 0 s ( S ( t , r ) − S ( s , r ) ) a ε ( r ) u ˉ ( r ) d r w_\varepsilon^{(1)}(t)-w_\varepsilon^{(1)}(s)
=-\mathrm{i}\int_s^t S(t,r)a_\varepsilon(r)\bar{u}(r)\,\mathrm{d}r
-\mathrm{i}\int_0^s\left(S(t,r)-S(s,r)\right)a_\varepsilon(r)\bar{u}(r)\,\mathrm{d}r w ε ( 1 ) ( t ) − w ε ( 1 ) ( s ) = − i ∫ s t S ( t , r ) a ε ( r ) u ˉ ( r ) d r − i ∫ 0 s ( S ( t , r ) − S ( s , r ) ) a ε ( r ) u ˉ ( r ) d r
where the first term is = : I ε ( t , s ) =:I_\varepsilon(t,s) =: I ε ( t , s ) and the second term is = : I I ε ( t , s ) =:II_\varepsilon(t,s) =: I I ε ( t , s ) .
By the sub-Gaussian moment bound of Assumption 1.4(i), uniformly in r r r ,
∥ a ε ( r ) u ˉ ( r ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C q ∥ u ˉ ( r ) ∥ L x 2 2 q ≤ C q , T , u 0 . \|a_\varepsilon(r)\bar{u}(r)\|_{L^{2q}(\Omega;L_x^2)}
\le C_q\|\bar{u}(r)\|_{L_x^2}^{2q}
\le C_{q,T,u_0}. ∥ a ε ( r ) u ˉ ( r ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C q ∥ u ˉ ( r ) ∥ L x 2 2 q ≤ C q , T , u 0 .
For I ε ( t , s ) I_\varepsilon(t,s) I ε ( t , s ) , the time interval has length t − s t-s t − s , so by the triangle inequality in L 2 q ( Ω ; L x 2 ) L^{2q}(\Omega;L_x^2) L 2 q ( Ω ; L x 2 ) and unitarity of S ( t , r ) S(t,r) S ( t , r ) ,
∥ I ε ( t , s ) ∥ L 2 q ( Ω ; L x 2 ) ≤ ( t − s ) sup r ∥ a ε ( r ) u ˉ ( r ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C q , T , u 0 ( t − s ) . (43) \|I_\varepsilon(t,s)\|_{L^{2q}(\Omega;L_x^2)}
\le(t-s)\sup_r\|a_\varepsilon(r)\bar{u}(r)\|_{L^{2q}(\Omega;L_x^2)}
\le C_{q,T,u_0}(t-s). \tag*{(43)} ∥ I ε ( t , s ) ∥ L 2 q ( Ω ; L x 2 ) ≤ ( t − s ) r sup ∥ a ε ( r ) u ˉ ( r ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C q , T , u 0 ( t − s ) . (43)
For I I ε ( t , s ) II_\varepsilon(t,s) I I ε ( t , s ) , since S ( t , r ) − S ( s , r ) = S ( s , r ) ( S ( t − s ) − I ) S(t,r)-S(s,r)=S(s,r)(S(t-s)-I) S ( t , r ) − S ( s , r ) = S ( s , r ) ( S ( t − s ) − I ) , and in Fourier space the multiplier satisfies
∣ e − i ( ∣ ξ ∣ 2 + V ˉ ) ( t − s ) − 1 ∣ ≤ ∣ t − s ∣ ( ∣ ξ ∣ 2 + V ˉ ) , \left|e^{-\mathrm{i}(|\xi|^2+\bar{V})(t-s)}-1\right|
\le|t-s|(|\xi|^2+\bar{V}), e − i ( ∣ ξ ∣ 2 + V ˉ ) ( t − s ) − 1 ≤ ∣ t − s ∣ ( ∣ ξ ∣ 2 + V ˉ ) ,
and u ^ 0 \hat{u}_0 u ^ 0 is compactly supported so ∣ ξ ∣ |\xi| ∣ ξ ∣ is bounded,
∥ S ( t − s ) − I ∥ L x 2 → L x 2 ≤ C u 0 ( t − s ) . \|S(t-s)-I\|_{L_x^2\to L_x^2}\le C_{u_0}(t-s). ∥ S ( t − s ) − I ∥ L x 2 → L x 2 ≤ C u 0 ( t − s ) .
Therefore, by unitarity of S ( s , r ) S(s,r) S ( s , r ) ,
∥ I I ε ( t , s ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C u 0 ( t − s ) ∫ 0 s ∥ a ε ( r ) u ˉ ( r ) ∥ L 2 q ( Ω ; L x 2 ) d r ≤ C q , T , u 0 ( t − s ) . (44) \|II_\varepsilon(t,s)\|_{L^{2q}(\Omega;L_x^2)}
\le C_{u_0}(t-s)\int_0^s\|a_\varepsilon(r)\bar{u}(r)\|_{L^{2q}(\Omega;L_x^2)}\,\mathrm{d}r
\le C_{q,T,u_0}(t-s). \tag*{(44)} ∥ I I ε ( t , s ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C u 0 ( t − s ) ∫ 0 s ∥ a ε ( r ) u ˉ ( r ) ∥ L 2 q ( Ω ; L x 2 ) d r ≤ C q , T , u 0 ( t − s ) . (44)
By (4.1) and (4.2),
∥ w ε ( 1 ) ( t ) − w ε ( 1 ) ( s ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C q , T , u 0 ( t − s ) , \|w_\varepsilon^{(1)}(t)-w_\varepsilon^{(1)}(s)\|_{L^{2q}(\Omega;L_x^2)}
\le C_{q,T,u_0}(t-s), ∥ w ε ( 1 ) ( t ) − w ε ( 1 ) ( s ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C q , T , u 0 ( t − s ) ,
which raised to the 2 q 2q 2 q -th power gives
E [ ∥ w ε ( 1 ) ( t ) − w ε ( 1 ) ( s ) ∥ L x 2 2 q ] ≤ C q , T , u 0 ∣ t − s ∣ 2 q . \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)-w_\varepsilon^{(1)}(s)\|_{L_x^2}^{2q}\right]
\le C_{q,T,u_0}|t-s|^{2q}. E [ ∥ w ε ( 1 ) ( t ) − w ε ( 1 ) ( s ) ∥ L x 2 2 q ] ≤ C q , T , u 0 ∣ t − s ∣ 2 q .
Applying the Kolmogorov–Chentsov maximal inequality (Lemma 2.10) with p = 2 q p=2q p = 2 q and α = 2 q > 1 2 q \alpha=2q>\frac{1}{2q} α = 2 q > 2 q 1 , and using (3.14),
E [ sup 0 ≤ t ≤ T ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] ≤ C ( E [ ∥ w ε ( 1 ) ( 0 ) ∥ L x 2 2 q ] + ∫ 0 T E [ ∥ w ε ( 1 ) ( r ) ∥ L x 2 2 q ] d r ] ≤ C q , T , u 0 ε d q . \mathbb{E}\left[\sup_{0\le t\le T}\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^{2q}\right]
\le C\left(\mathbb{E}\left[\|w_\varepsilon^{(1)}(0)\|_{L_x^2}^{2q}\right]
+\int_0^T\mathbb{E}\left[\|w_\varepsilon^{(1)}(r)\|_{L_x^2}^{2q}\right]\,\mathrm{d}r\right]
\le C_{q,T,u_0}\varepsilon^{dq}. E [ 0 ≤ t ≤ T sup ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] ≤ C ( E [ ∥ w ε ( 1 ) ( 0 ) ∥ L x 2 2 q ] + ∫ 0 T E [ ∥ w ε ( 1 ) ( r ) ∥ L x 2 2 q ] d r ] ≤ C q , T , u 0 ε d q .
This proves Proposition 4.1.
The rate estimate combined with the dyadic Borel–Cantelli argument gives almost sure convergence.
Proposition 4.2. Under Assumption 1.4,
lim ε → 0 sup 0 ≤ t ≤ T ∥ u ε ( t ) − u ˉ ( t ) ∥ L 2 ( R d ) = 0 a.s. \lim_{\varepsilon\to0}\sup_{0\le t\le T}\|u_{\varepsilon}(t)-\bar{u}(t)\|_{L^2(\mathbb{R}^d)}=0 \qquad\text{a.s.} ε → 0 lim 0 ≤ t ≤ T sup ∥ u ε ( t ) − u ˉ ( t ) ∥ L 2 ( R d ) = 0 a.s.
Proof. Fix the dyadic sequence ε n = 2 − n \varepsilon_n=2^{-n} ε n = 2 − n . By Markov’s inequality applied to the 2 q 2q 2 q -th moment estimate of Proposition 4.1, for any η > 0 \eta>0 η > 0 and any integer q > 1 d η q>\frac{1}{d\eta} q > d η 1 ,
P ( sup 0 ≤ t ≤ T ∥ w ε n ( t ) ∥ L x 2 2 ≥ ε n d 2 − η ) ≤ C q ε n q d η . \mathbb{P}\left(\sup_{0\le t\le T}\|w_{\varepsilon_n}(t)\|_{L_x^2}^2\ge\varepsilon_n^{\frac d2-\eta}\right)\le C_q\varepsilon_n^{qd\eta}. P ( 0 ≤ t ≤ T sup ∥ w ε n ( t ) ∥ L x 2 2 ≥ ε n 2 d − η ) ≤ C q ε n q d η .
Choosing q q q large enough that q d η > 1 qd\eta>1 q d η > 1 , the series ∑ n ε n q d η = ∑ n 2 − n q d η < ∞ \sum_n\varepsilon_n^{qd\eta}=\sum_n2^{-nqd\eta}<\infty ∑ n ε n q d η = ∑ n 2 − n q d η < ∞ . By Lemma 2.8 ,
sup 0 ≤ t ≤ T ∥ w ε n ( t ) ∥ L x 2 → 0 a.s. (45) \sup_{0\le t\le T}\|w_{\varepsilon_n}(t)\|_{L_x^2}\to0 \qquad\text{a.s.} \tag*{(45)} 0 ≤ t ≤ T sup ∥ w ε n ( t ) ∥ L x 2 → 0 a.s. (45)
It remains to upgrade the dyadic convergence to full convergence. The shell ε ∈ [ ε n + 1 , ε n ] \varepsilon\in[\varepsilon_{n+1},\varepsilon_n] ε ∈ [ ε n + 1 , ε n ] has relative width 1 2 \frac12 2 1 . We control the shell supremum in two steps.
Step 1. The Schrödinger group S ( t , r ) S(t,r) S ( t , r ) is unitary and the Duhamel integrand a ε ( r ) u ˉ ( r ) a_\varepsilon(r)\bar{u}(r) a ε ( r ) u ˉ ( r ) is uniformly bounded in L x 2 L_x^2 L x 2 by sub-Gaussian concentration (Lemma 2.13 ). Hence, for any s < t s<t s < t ,
∥ w ε ( 1 ) ( t ) − w ε ( 1 ) ( s ) ∥ L x 2 ≤ ∫ s t ∥ a ε ( r ) u ˉ ( r ) ∥ L x 2 d r ≤ C ∣ t − s ∣ a.s. \|w_\varepsilon^{(1)}(t)-w_\varepsilon^{(1)}(s)\|_{L_x^2}\le\int_s^t\|a_\varepsilon(r)\bar{u}(r)\|_{L_x^2}\,dr\le C|t-s| \qquad\text{a.s.} ∥ w ε ( 1 ) ( t ) − w ε ( 1 ) ( s ) ∥ L x 2 ≤ ∫ s t ∥ a ε ( r ) u ˉ ( r ) ∥ L x 2 d r ≤ C ∣ t − s ∣ a.s.
on compact time intervals, by Fubini and the uniform spatial moment bound. Choose the time grid t j = j δ t_j=j\delta t j = j δ , j = 0 , … , ⌊ T δ ⌋ j=0,\ldots,\left\lfloor\frac{T}{\delta}\right\rfloor j = 0 , … , ⌊ δ T ⌋ , with δ = ε n d 2 − η \delta=\varepsilon_n^{\frac d2-\eta} δ = ε n 2 d − η . For any t ∈ [ t j , t j + 1 ] t\in[t_j,t_{j+1}] t ∈ [ t j , t j + 1 ] ,
∥ w ε ( t ) ∥ L x 2 ≤ ∥ w ε ( t j ) ∥ L x 2 + C δ . \|w_\varepsilon(t)\|_{L_x^2}\le\|w_\varepsilon(t_j)\|_{L_x^2}+C\delta. ∥ w ε ( t ) ∥ L x 2 ≤ ∥ w ε ( t j ) ∥ L x 2 + C δ .
Hence,
sup 0 ≤ t ≤ T ∥ w ε ( t ) ∥ L x 2 ≤ max 0 ≤ j ≤ ⌊ T δ ⌋ ∥ w ε ( t j ) ∥ L x 2 + C δ . \sup_{0\le t\le T}\|w_\varepsilon(t)\|_{L_x^2}\le\max_{0\le j\le\left\lfloor\frac{T}{\delta}\right\rfloor}\|w_\varepsilon(t_j)\|_{L_x^2}+C\delta. 0 ≤ t ≤ T sup ∥ w ε ( t ) ∥ L x 2 ≤ 0 ≤ j ≤ ⌊ δ T ⌋ max ∥ w ε ( t j ) ∥ L x 2 + C δ .
The number of grid points is O ( T δ ) = O ( ε n − d 2 + η ) O\left(\frac{T}{\delta}\right)=O\left(\varepsilon_n^{-\frac d2+\eta}\right) O ( δ T ) = O ( ε n − 2 d + η ) , and the factor C δ = O ( ε n d 2 − η ) C\delta=O\left(\varepsilon_n^{\frac d2-\eta}\right) C δ = O ( ε n 2 d − η ) is negligible compared with the threshold ε n d 2 − η \varepsilon_n^{\frac d2-\eta} ε n 2 d − η .
Step 2. For ε ∈ [ ε n + 1 , ε n ] \varepsilon\in[\varepsilon_{n+1},\varepsilon_n] ε ∈ [ ε n + 1 , ε n ] , we compare w ε ( t ) w_\varepsilon(t) w ε ( t ) to w ε n ( t ) w_{\varepsilon_n}(t) w ε n ( t ) . Subtracting the two Duhamel formulas and inserting the identity
a ε w ε − a ε n w ε n = ( a ε − a ε n ) u ˉ + a ε ( w ε − w ε n ) + a ε n w ε n , a_\varepsilon w_\varepsilon-a_{\varepsilon_n}w_{\varepsilon_n}=(a_\varepsilon-a_{\varepsilon_n})\bar{u}+a_\varepsilon(w_\varepsilon-w_{\varepsilon_n})+a_{\varepsilon_n}w_{\varepsilon_n}, a ε w ε − a ε n w ε n = ( a ε − a ε n ) u ˉ + a ε ( w ε − w ε n ) + a ε n w ε n ,
we have
w ε ( t ) − w ε n ( t ) = − i ∫ 0 t S ( t , r ) ( a ε ( r ) − a ε n ( r ) ) u ˉ ( r ) d r (46) w_\varepsilon(t)-w_{\varepsilon_n}(t)=-i\int_0^t S(t,r)(a_\varepsilon(r)-a_{\varepsilon_n}(r))\bar{u}(r)\,dr \tag*{(46)} w ε ( t ) − w ε n ( t ) = − i ∫ 0 t S ( t , r ) ( a ε ( r ) − a ε n ( r )) u ˉ ( r ) d r (46)
− i ∫ 0 t S ( t , r ) a ε ( r ) ( w ε ( r ) − w ε n ( r ) ) d r (47) -i\int_0^t S(t,r)a_\varepsilon(r)(w_\varepsilon(r)-w_{\varepsilon_n}(r))\,dr \tag*{(47)} − i ∫ 0 t S ( t , r ) a ε ( r ) ( w ε ( r ) − w ε n ( r )) d r (47)
− i ∫ 0 t S ( t , r ) a ε n ( r ) w ε n ( r ) d r . (48) -i\int_0^t S(t,r)a_{\varepsilon_n}(r)w_{\varepsilon_n}(r)\,dr. \tag*{(48)} − i ∫ 0 t S ( t , r ) a ε n ( r ) w ε n ( r ) d r . (48)
We treat the three terms on the right hand side in turn.
For (46) , the difference field a ε − a ε n a_\varepsilon-a_{\varepsilon_n} a ε − a ε n is not stationary, but its second moment can be bounded directly in Fourier space. The Fourier transform of a ε ( x ) = a ( x ε ) a_\varepsilon(x)=a\left(\frac{x}{\varepsilon}\right) a ε ( x ) = a ( ε x ) is
F ( a ε ) ( ξ ) = ε d a ^ ( ε ξ ) , \mathcal{F}(a_\varepsilon)(\xi)=\varepsilon^d\widehat{a}(\varepsilon\xi), F ( a ε ) ( ξ ) = ε d a ( ε ξ ) ,
so the difference satisfies
F ( a ε − a ε n ) ( ξ ) = ε d a ^ ( ε ξ ) − ε n d a ^ ( ε n ξ ) . \mathcal{F}(a_\varepsilon-a_{\varepsilon_n})(\xi)=\varepsilon^d\widehat{a}(\varepsilon\xi)-\varepsilon_n^d\widehat{a}(\varepsilon_n\xi). F ( a ε − a ε n ) ( ξ ) = ε d a ( ε ξ ) − ε n d a ( ε n ξ ) .
The covariance of this non stationary field is computed by writing
E [ ∣ F ( a ε − a ε n ) ( ξ ) ∣ 2 ] = ε 2 d E [ ∣ a ^ ( ε ξ ) ∣ 2 ] + ε n 2 d E [ ∣ a ^ ( ε n ξ ) ∣ 2 ] − 2 ε d ε n d ℜ E [ a ^ ( ε ξ ) a ^ ( ε n ξ ) ‾ ] . \mathbb{E}\left[\left|\mathcal{F}(a_{\varepsilon}-a_{\varepsilon_n})(\xi)\right|^2\right]=\varepsilon^{2d}\mathbb{E}\left[\left|\widehat{a}(\varepsilon\xi)\right|^2\right]+\varepsilon_n^{2d}\mathbb{E}\left[\left|\widehat{a}(\varepsilon_n\xi)\right|^2\right]-2\varepsilon^d\varepsilon_n^d\Re\mathbb{E}\left[\widehat{a}(\varepsilon\xi)\overline{\widehat{a}(\varepsilon_n\xi)}\right]. E [ ∣ F ( a ε − a ε n ) ( ξ ) ∣ 2 ] = ε 2 d E [ ∣ a ( ε ξ ) ∣ 2 ] + ε n 2 d E [ ∣ a ( ε n ξ ) ∣ 2 ] − 2 ε d ε n d ℜ E [ a ( ε ξ ) a ( ε n ξ ) ] .
Using (3.3), the diagonal terms give
ε 2 d ( 2 π ) d ε − d R ^ ( ε ξ ) + ε n 2 d ( 2 π ) d ε n − d R ^ ( ε n ξ ) = ( 2 π ) d [ ε d R ^ ( ε ξ ) + ε n d R ^ ( ε n ξ ) ] . \varepsilon^{2d}(2\pi)^d\varepsilon^{-d}\widehat{R}(\varepsilon\xi)+\varepsilon_n^{2d}(2\pi)^d\varepsilon_n^{-d}\widehat{R}(\varepsilon_n\xi)=(2\pi)^d\left[\varepsilon^d\widehat{R}(\varepsilon\xi)+\varepsilon_n^d\widehat{R}(\varepsilon_n\xi)\right]. ε 2 d ( 2 π ) d ε − d R ( ε ξ ) + ε n 2 d ( 2 π ) d ε n − d R ( ε n ξ ) = ( 2 π ) d [ ε d R ( ε ξ ) + ε n d R ( ε n ξ ) ] .
The cross covariance is, by stationarity and the Fourier covariance identity,
E [ a ^ ( ε ξ ) a ^ ( ε n ξ ) ‾ ] = ( 2 π ) d δ ( ε ξ − ε n ξ ) R ^ ( ε ξ ) , \mathbb{E}\left[\widehat{a}(\varepsilon\xi)\overline{\widehat{a}(\varepsilon_n\xi)}\right]=(2\pi)^d\delta(\varepsilon\xi-\varepsilon_n\xi)\widehat{R}(\varepsilon\xi), E [ a ( ε ξ ) a ( ε n ξ ) ] = ( 2 π ) d δ ( ε ξ − ε n ξ ) R ( ε ξ ) ,
which for ε ≠ ε n \varepsilon\ne\varepsilon_n ε = ε n is supported on the hyperplane ξ = 0 \xi=0 ξ = 0 , a measure zero set in R d \mathbb{R}^d R d . The cross term therefore contributes only through the near region ∣ ξ ∣ ≲ ε n − 1 |\xi|\lesssim\varepsilon_n^{-1} ∣ ξ ∣ ≲ ε n − 1 , where the delta is regularized by the spectral smoothness of R ^ \widehat{R} R . It is controlled by the same near region volume estimate as in Lemma 3.2. The leading L 2 q L^{2q} L 2 q norm of the difference integral is therefore controlled by the near region volume O ( ε n d ) O(\varepsilon_n^d) O ( ε n d ) and the exponential decay of R ^ \widehat{R} R . Hence,
∥ − i ∫ 0 t S ( t , r ) ( a ε ( r ) − a ε n ( r ) ) u ‾ ( r ) d r ∥ L 2 q ( Ω ; L x 2 ) ≤ C q , T , u 0 ε n d 2 . (49) \left\|-i\int_0^t S(t,r)(a_\varepsilon(r)-a_{\varepsilon_n}(r))\overline{u}(r)\,dr\right\|_{L^{2q}(\Omega;L_x^2)}\le C_{q,T,u_0}\varepsilon_n^{\frac{d}{2}}. \tag*{(49)} − i ∫ 0 t S ( t , r ) ( a ε ( r ) − a ε n ( r )) u ( r ) d r L 2 q ( Ω ; L x 2 ) ≤ C q , T , u 0 ε n 2 d . (49)
The prefactor ∣ ε − ε n ∣ ε n ≤ 1 \frac{|\varepsilon-\varepsilon_n|}{\varepsilon_n}\le1 ε n ∣ ε − ε n ∣ ≤ 1 is absorbed into the constant.
For (4.4b), taking L 2 q ( Ω ; L x 2 ) L^{2q}(\Omega;L_x^2) L 2 q ( Ω ; L x 2 ) norms and using the unitarity of S ( t , r ) S(t,r) S ( t , r ) ,
∥ − i ∫ 0 t S ( t , r ) a ε ( r ) ( w ε ( r ) − w ε n ( r ) ) d r ∥ L 2 q ( Ω ; L x 2 ) ≤ C ∫ 0 t ∥ w ε ( r ) − w ε n ( r ) ∥ L 2 q ( Ω ; L x 2 ) d r , \left\|-i\int_0^t S(t,r)a_\varepsilon(r)(w_\varepsilon(r)-w_{\varepsilon_n}(r))\,dr\right\|_{L^{2q}(\Omega;L_x^2)}
\le C\int_0^t\|w_\varepsilon(r)-w_{\varepsilon_n}(r)\|_{L^{2q}(\Omega;L_x^2)}\,dr, − i ∫ 0 t S ( t , r ) a ε ( r ) ( w ε ( r ) − w ε n ( r )) d r L 2 q ( Ω ; L x 2 ) ≤ C ∫ 0 t ∥ w ε ( r ) − w ε n ( r ) ∥ L 2 q ( Ω ; L x 2 ) d r ,
where Lemma 2.13 bounds ∥ a ε ( r ) ( w ε ( r ) − w ε n ( r ) ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C ∥ w ε ( r ) − w ε n ( r ) ∥ L 2 q ( Ω ; L x 2 ) \|a_\varepsilon(r)(w_\varepsilon(r)-w_{\varepsilon_n}(r))\|_{L^{2q}(\Omega;L_x^2)}\le C\|w_\varepsilon(r)-w_{\varepsilon_n}(r)\|_{L^{2q}(\Omega;L_x^2)} ∥ a ε ( r ) ( w ε ( r ) − w ε n ( r )) ∥ L 2 q ( Ω ; L x 2 ) ≤ C ∥ w ε ( r ) − w ε n ( r ) ∥ L 2 q ( Ω ; L x 2 ) .
For (4.4c), by the same moment bound,
∥ − i ∫ 0 t S ( t , r ) a ε n ( r ) w ε n ( r ) d r ∥ L 2 q ( Ω ; L x 2 ) ≤ C ∥ w ε n ( t ) ∥ L 2 q ( Ω ; L x 2 ) , (50) \left\|-i\int_0^t S(t,r)a_{\varepsilon_n}(r)w_{\varepsilon_n}(r)\,dr\right\|_{L^{2q}(\Omega;L_x^2)}
\le C\|w_{\varepsilon_n}(t)\|_{L^{2q}(\Omega;L_x^2)}, \tag*{(50)} − i ∫ 0 t S ( t , r ) a ε n ( r ) w ε n ( r ) d r L 2 q ( Ω ; L x 2 ) ≤ C ∥ w ε n ( t ) ∥ L 2 q ( Ω ; L x 2 ) , (50)
which is controlled by the dyadic estimate (4.3).
Combining the three estimates and applying Gronwall’s inequality (Lemma 2.1) to the integral inequality for w ε − w ε n w_\varepsilon-w_{\varepsilon_n} w ε − w ε n ,
∥ w ε ( t ) − w ε n ( t ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C q , T , u 0 ε n d 2 + C q , T , u 0 ∥ w ε n ( t ) ∥ L 2 q ( Ω ; L x 2 ) . (51) \|w_\varepsilon(t)-w_{\varepsilon_n}(t)\|_{L^{2q}(\Omega;L_x^2)}
\le C_{q,T,u_0}\varepsilon_n^{\frac{d}{2}}+C_{q,T,u_0}\|w_{\varepsilon_n}(t)\|_{L^{2q}(\Omega;L_x^2)}. \tag*{(51)} ∥ w ε ( t ) − w ε n ( t ) ∥ L 2 q ( Ω ; L x 2 ) ≤ C q , T , u 0 ε n 2 d + C q , T , u 0 ∥ w ε n ( t ) ∥ L 2 q ( Ω ; L x 2 ) . (51)
Step 3. By Proposition 4.1 and Lemma 2.13, for any integer q ≥ 1 q\ge1 q ≥ 1 ,
∥ ∥ w ε ( t ) − w ε n ( t ) ∥ L x 2 ∥ L 2 q ( Ω ) ≤ C q ε n d 2 . \left\|\|w_\varepsilon(t)-w_{\varepsilon_n}(t)\|_{L_x^2}\right\|_{L^{2q}(\Omega)}\le C_q\varepsilon_n^{\frac{d}{2}}. ∥ w ε ( t ) − w ε n ( t ) ∥ L x 2 L 2 q ( Ω ) ≤ C q ε n 2 d .
Applying Markov’s inequality to the 2 q 2q 2 q -th moment and union bound over the time grid (O ( ε n − d 2 + η ) O(\varepsilon_n^{-\frac{d}{2}+\eta}) O ( ε n − 2 d + η ) points) and the dyadic shell,
P ( sup ε ∈ [ ε n + 1 , ε n ] sup 0 ≤ t ≤ T ∥ w ε ( t ) − w ε n ( t ) ∥ L x 2 > 2 C ε n d 2 − η ) ≤ C q ε n − d 2 + η ε n q d η . \mathbb{P}\left(\sup_{\varepsilon\in[\varepsilon_{n+1},\varepsilon_n]}\sup_{0\le t\le T}\|w_\varepsilon(t)-w_{\varepsilon_n}(t)\|_{L_x^2}>2C\varepsilon_n^{\frac{d}{2}-\eta}\right)\le C_q\varepsilon_n^{-\frac{d}{2}+\eta}\varepsilon_n^{qd\eta}. P ( ε ∈ [ ε n + 1 , ε n ] sup 0 ≤ t ≤ T sup ∥ w ε ( t ) − w ε n ( t ) ∥ L x 2 > 2 C ε n 2 d − η ) ≤ C q ε n − 2 d + η ε n q d η .
Choosing q q q large enough that q d η > 1 + d 2 − η qd\eta>1+\frac{d}{2}-\eta q d η > 1 + 2 d − η , the exponent is positive and the series is summable. By Lemma 2.8,
sup ε ∈ [ ε n + 1 , ε n ] sup 0 ≤ t ≤ T ∥ w ε ( t ) − w ε n ( t ) ∥ L x 2 ≤ 2 C ε n d 2 − η \sup_{\varepsilon\in[\varepsilon_{n+1},\varepsilon_n]}\sup_{0\le t\le T}\|w_\varepsilon(t)-w_{\varepsilon_n}(t)\|_{L_x^2}\le2C\varepsilon_n^{\frac{d}{2}-\eta} ε ∈ [ ε n + 1 , ε n ] sup 0 ≤ t ≤ T sup ∥ w ε ( t ) − w ε n ( t ) ∥ L x 2 ≤ 2 C ε n 2 d − η
for all sufficiently large n n n , almost surely. Combining this with (4.3),
sup ε ∈ [ ε n + 1 , ε n ] sup 0 ≤ t ≤ T ∥ w ε ( t ) ∥ L x 2 ≤ ∥ w ε n ( t ) ∥ + C ε n d 2 − η → 0 a.s. \sup_{\varepsilon\in[\varepsilon_{n+1},\varepsilon_n]}\sup_{0\le t\le T}\|w_\varepsilon(t)\|_{L_x^2}\le\|w_{\varepsilon_n}(t)\|+C\varepsilon_n^{\frac{d}{2}-\eta}\to0\qquad\text{a.s.} ε ∈ [ ε n + 1 , ε n ] sup 0 ≤ t ≤ T sup ∥ w ε ( t ) ∥ L x 2 ≤ ∥ w ε n ( t ) ∥ + C ε n 2 d − η → 0 a.s.
This proves the result. □
Theorem 1.9 follows directly from Proposition 4.2 , which proves the almost sure convergence by the dyadic Borel–Cantelli argument.
Fluctuation CLT
Having established the convergence rate, we now turn to the fluctuations at the critical scale ε d 2 \varepsilon^{\frac{d}{2}} ε 2 d . Recall the rescaled fluctuation field
F ε ( t , x ) : = ε − d 2 ( u ε ( t , x ) − u ‾ ( t , x ) ) . \mathcal{F}_{\varepsilon}(t,x) := \varepsilon^{-\frac{d}{2}}(u_{\varepsilon}(t,x)-\overline{u}(t,x)). F ε ( t , x ) := ε − 2 d ( u ε ( t , x ) − u ( t , x )) .
By Proposition 3.11 ,
F ε ( t ) = − i ε − d 2 ∫ 0 t S ( t , r ) a ε ( r ) u ‾ ( r ) d r + o L 2 ( Ω ) ( 1 ) . \mathcal{F}_{\varepsilon}(t)=-i\varepsilon^{-\frac{d}{2}}\int_{0}^{t}S(t,r)a_{\varepsilon}(r)\overline{u}(r)\,dr+o_{L^{2}(\Omega)}(1). F ε ( t ) = − i ε − 2 d ∫ 0 t S ( t , r ) a ε ( r ) u ( r ) d r + o L 2 ( Ω ) ( 1 ) .
The leading term is a spatial average of the centered field a ε a_{\varepsilon} a ε against a deterministic profile. To identify its limiting law we decompose this average into a martingale array on spatial blocks, using the standard martingale approximation of Kipnis–Varadhan [19 ] , and invoke the martingale central limit theorem of Hall–Heyde [15 ] .
Fix N > d N>d N > d and set the block size
L ε : = 2 N ( − log ε ) κ . (52) L_{\varepsilon}:=\frac{2N(-\log\varepsilon)}{\kappa}. \tag*{(52)} L ε := κ 2 N ( − log ε ) . (52)
Partition R d \mathbb{R}^{d} R d into cubes Q j Q_{j} Q j of side L ε L_{\varepsilon} L ε , ordered lexicographically, and define the filtration
G j : = σ ( a ( y ′ ) : y ′ ∈ Q 1 ∪ ⋯ ∪ Q j ) . \mathcal{G}_{j}:=\sigma\left(a(y^{\prime}):y^{\prime}\in Q_{1}\cup\cdots\cup Q_{j}\right). G j := σ ( a ( y ′ ) : y ′ ∈ Q 1 ∪ ⋯ ∪ Q j ) .
For φ ∈ L 2 ( R d ) \varphi\in L^{2}(\mathbb{R}^{d}) φ ∈ L 2 ( R d ) , define the block contributions
M j ε ( t ) : = − i ε − d 2 ∫ 0 t ∫ Q j a ( y ′ ) S ( t , r ) ∗ φ ( ε y ′ ) ‾ u ‾ ( r , ε y ′ ) d y ′ d r . (53) M_{j}^{\varepsilon}(t):=-i\varepsilon^{-\frac{d}{2}}\int_{0}^{t}\int_{Q_{j}}a(y^{\prime})\overline{S(t,r)^{*}\varphi(\varepsilon y^{\prime})}\,\overline{u}(r,\varepsilon y^{\prime})\,dy^{\prime}\,dr. \tag*{(53)} M j ε ( t ) := − i ε − 2 d ∫ 0 t ∫ Q j a ( y ′ ) S ( t , r ) ∗ φ ( ε y ′ ) u ( r , ε y ′ ) d y ′ d r . (53)
Proposition 5.1. With the block size (52) , filtration G j \mathcal{G}_{j} G j , and block contributions (53) ,
⟨ F ε ( t ) , φ ⟩ = ∑ j M j ε ( t ) + o L 2 ( Ω ) ( 1 ) , \langle\mathcal{F}_{\varepsilon}(t),\varphi\rangle=\sum_{j}M_{j}^{\varepsilon}(t)+o_{L^{2}(\Omega)}(1), ⟨ F ε ( t ) , φ ⟩ = j ∑ M j ε ( t ) + o L 2 ( Ω ) ( 1 ) , and the centered array { M j ε − E [ M j ε ∣ G j − 1 ] } j \{M_{j}^{\varepsilon}-\mathbb{E}[M_{j}^{\varepsilon}\mid\mathcal{G}_{j-1}]\}_{j} { M j ε − E [ M j ε ∣ G j − 1 ] } j satisfies the Lindeberg condition (19) .
Proof. After rescaling y ′ = y ε y^{\prime}=\frac{y}{\varepsilon} y ′ = ε y ,
⟨ F ε ( t ) , φ ⟩ = − i ε d 2 ∫ 0 t ∫ R d a ( y ′ ) S ( t , r ) ∗ φ ( ε y ′ ) ‾ u ‾ ( r , ε y ′ ) d y ′ d r = ∑ j M j ε ( t ) + boundary . \langle\mathcal{F}_{\varepsilon}(t),\varphi\rangle=-i\varepsilon^{\frac{d}{2}}\int_{0}^{t}\int_{\mathbb{R}^{d}}a(y^{\prime})\overline{S(t,r)^{*}\varphi(\varepsilon y^{\prime})}\overline{u}(r,\varepsilon y^{\prime})\,dy^{\prime}\,dr=\sum_{j}M_{j}^{\varepsilon}(t)+\text{boundary}. ⟨ F ε ( t ) , φ ⟩ = − i ε 2 d ∫ 0 t ∫ R d a ( y ′ ) S ( t , r ) ∗ φ ( ε y ′ ) u ( r , ε y ′ ) d y ′ d r = j ∑ M j ε ( t ) + boundary .
For dist ( Q j , Q k ) ≥ L ε \operatorname{dist}(Q_{j},Q_{k})\ge L_{\varepsilon} dist ( Q j , Q k ) ≥ L ε , by (9) with R = L ε 2 R=\frac{L_{\varepsilon}}{2} R = 2 L ε ,
∣ E [ f ( M j ε ) g ( M k ε ) ] − E [ f ( M j ε ) ] E [ g ( M k ε ) ] ∣ ≤ C ∥ f ∥ ∞ ∥ g ∥ ∞ exp ( − κ L ε 2 ) ≤ C ε N . \left|\mathbb{E}[f(M_{j}^{\varepsilon})g(M_{k}^{\varepsilon})]-\mathbb{E}[f(M_{j}^{\varepsilon})]\mathbb{E}[g(M_{k}^{\varepsilon})]\right|\le C\|f\|_{\infty}\|g\|_{\infty}\exp\left(-\frac{\kappa L_{\varepsilon}}{2}\right)\le C\varepsilon^{N}. E [ f ( M j ε ) g ( M k ε )] − E [ f ( M j ε )] E [ g ( M k ε )] ≤ C ∥ f ∥ ∞ ∥ g ∥ ∞ exp ( − 2 κ L ε ) ≤ C ε N .
The number of blocks is N ε = O ( ε − d L ε − d ) N_{\varepsilon}=O(\varepsilon^{-d}L_{\varepsilon}^{-d}) N ε = O ( ε − d L ε − d ) , so
N ε ⋅ ε N = O ( ε N − d L ε − d ) = o ( 1 ) . N_{\varepsilon}\cdot\varepsilon^{N}=O(\varepsilon^{N-d}L_{\varepsilon}^{-d})=o(1). N ε ⋅ ε N = O ( ε N − d L ε − d ) = o ( 1 ) .
By the mixing condition,
E [ M j ε ∣ G j − 1 ] = E [ M j ε ] + O ( ε N ) , \mathbb{E}[M_{j}^{\varepsilon}\mid\mathcal{G}_{j-1}]=\mathbb{E}[M_{j}^{\varepsilon}]+O(\varepsilon^{N}), E [ M j ε ∣ G j − 1 ] = E [ M j ε ] + O ( ε N ) ,
and
∑ j = 1 N ε ∣ E [ M j ε ∣ G j − 1 ] − E [ M j ε ] ∣ ≤ N ε ⋅ C ε N = C ε N − d ( − log ε ) d ⟶ 0. \sum_{j=1}^{N_{\varepsilon}}\left|\mathbb{E}[M_{j}^{\varepsilon}\mid\mathcal{G}_{j-1}]-\mathbb{E}[M_{j}^{\varepsilon}]\right|\le N_{\varepsilon}\cdot C\varepsilon^{N}=C\frac{\varepsilon^{N-d}}{(-\log\varepsilon)^{d}}\longrightarrow0. j = 1 ∑ N ε E [ M j ε ∣ G j − 1 ] − E [ M j ε ] ≤ N ε ⋅ C ε N = C ( − log ε ) d ε N − d ⟶ 0.
By Chebyshev,
∑ j E [ ∣ M j ε ∣ 2 1 { ∣ M j ε ∣ > η } ] ≤ η − 2 ∑ j E [ ∣ M j ε ∣ 4 ] . \sum_j \mathbb{E}\left[|M_j^\varepsilon|^2\mathbf{1}_{\{|M_j^\varepsilon|>\eta\}}\right]\leq\eta^{-2}\sum_j\mathbb{E}\left[|M_j^\varepsilon|^4\right]. j ∑ E [ ∣ M j ε ∣ 2 1 { ∣ M j ε ∣ > η } ] ≤ η − 2 j ∑ E [ ∣ M j ε ∣ 4 ] .
Let g ( y ′ , r ) : = S ‾ ( t , r ) ∗ φ ( ε y ′ ) u ‾ ( r , ε y ′ ) g(y',r):=\overline{S}(t,r)*\varphi(\varepsilon y')\overline{u}(r,\varepsilon y') g ( y ′ , r ) := S ( t , r ) ∗ φ ( ε y ′ ) u ( r , ε y ′ ) . By Lemma 2.13,
E [ ∣ M j ε ∣ 4 ] ≤ C s g 2 ( E [ ∣ M j ε ∣ 2 ] ) 2 . \mathbb{E}\left[|M_j^\varepsilon|^4\right]\leq C_{\mathrm{sg}}^2\left(\mathbb{E}\left[|M_j^\varepsilon|^2\right]\right)^2. E [ ∣ M j ε ∣ 4 ] ≤ C sg 2 ( E [ ∣ M j ε ∣ 2 ] ) 2 .
The local second moment is
E [ ∣ M j ε ∣ 2 ] = ε d ∫ Q j ∫ Q j R ( y ′ − y ′ ′ ) ∣ g ( y ′ ) ∣ 2 d y ′ d y ′ ′ = O ( ε d L ε d ) ∣ g j ∣ 2 . \mathbb{E}\left[|M_j^\varepsilon|^2\right]=\varepsilon^d\int_{Q_j}\int_{Q_j}R(y'-y'')|g(y')|^2\,\mathrm{d}y'\,\mathrm{d}y''=O\left(\varepsilon^dL_\varepsilon^d\right)|g_j|^2. E [ ∣ M j ε ∣ 2 ] = ε d ∫ Q j ∫ Q j R ( y ′ − y ′′ ) ∣ g ( y ′ ) ∣ 2 d y ′ d y ′′ = O ( ε d L ε d ) ∣ g j ∣ 2 .
The fourth moment is
E [ ∣ M j ε ∣ 4 ] ≤ C ε 2 d L ε 2 d ∣ g j ∣ 4 . \mathbb{E}\left[|M_j^\varepsilon|^4\right]\leq C\varepsilon^{2d}L_\varepsilon^{2d}|g_j|^4. E [ ∣ M j ε ∣ 4 ] ≤ C ε 2 d L ε 2 d ∣ g j ∣ 4 .
Using ∑ j ∣ g j ∣ 4 L ε d = O ( ε − d ) \sum_j|g_j|^4L_\varepsilon^d=O(\varepsilon^{-d}) ∑ j ∣ g j ∣ 4 L ε d = O ( ε − d ) ,
∑ j E [ ∣ M j ε ∣ 4 ] = O ( ε 2 d L ε 2 d ∑ j ∣ g j ∣ 4 ) = O ( ε 2 d L ε 2 d ⋅ ε − d L ε − d ) = O ( ε d L ε d ) ⟶ 0. (54) \sum_j\mathbb{E}\left[|M_j^\varepsilon|^4\right]=O\left(\varepsilon^{2d}L_\varepsilon^{2d}\sum_j|g_j|^4\right)=O\left(\varepsilon^{2d}L_\varepsilon^{2d}\cdot\varepsilon^{-d}L_\varepsilon^{-d}\right)=O\left(\varepsilon^dL_\varepsilon^d\right)\longrightarrow0. \tag*{(54)} j ∑ E [ ∣ M j ε ∣ 4 ] = O ( ε 2 d L ε 2 d j ∑ ∣ g j ∣ 4 ) = O ( ε 2 d L ε 2 d ⋅ ε − d L ε − d ) = O ( ε d L ε d ) ⟶ 0. (54)
The variance sum is
∑ j E [ ∣ M j ε ∣ 2 ] = ε d ∑ j ∣ g j ∣ 2 L ε d R ^ ( 0 ) ⟶ σ 2 : = ∫ 0 t ∥ S ( t , r ) ∗ φ ∥ L 2 2 R ^ ( 0 ) d r . \sum_j\mathbb{E}\left[|M_j^\varepsilon|^2\right]=\varepsilon^d\sum_j|g_j|^2L_\varepsilon^d\widehat{R}(0)\longrightarrow\sigma^2:=\int_0^t\|S(t,r)*\varphi\|_{L^2}^2\widehat{R}(0)\,\mathrm{d}r. j ∑ E [ ∣ M j ε ∣ 2 ] = ε d j ∑ ∣ g j ∣ 2 L ε d R ( 0 ) ⟶ σ 2 := ∫ 0 t ∥ S ( t , r ) ∗ φ ∥ L 2 2 R ( 0 ) d r .
Thus ∑ j E [ ∣ M j ε ∣ 4 ] → 0 \sum_j\mathbb{E}[|M_j^\varepsilon|^4]\to0 ∑ j E [ ∣ M j ε ∣ 4 ] → 0 , and the Lindeberg condition follows. □
With the martingale array in hand, we compute its limiting covariance.
Proposition 5.2. For every t , s ∈ [ 0 , T ] t,s\in[0,T] t , s ∈ [ 0 , T ] and φ , ψ ∈ L 2 ( R d ) \varphi,\psi\in L^2(\mathbb{R}^d) φ , ψ ∈ L 2 ( R d ) ,
lim ε → 0 E [ ⟨ F ε ( t ) , φ ⟩ ⟨ F ε ( s ) , ψ ⟩ ] = ∫ 0 t ∧ s ⟨ S ( t , r ) ∗ φ , R ^ ( 0 ) S ( s , r ) ∗ ψ ⟩ L 2 d r . (55) \lim_{\varepsilon\to0}\mathbb{E}\left[\langle\mathcal{F}_\varepsilon(t),\varphi\rangle\langle\mathcal{F}_\varepsilon(s),\psi\rangle\right]=\int_0^{t\wedge s}\langle S(t,r)*\varphi,\widehat{R}(0)S(s,r)*\psi\rangle_{L^2}\,\mathrm{d}r. \tag*{(55)} ε → 0 lim E [ ⟨ F ε ( t ) , φ ⟩ ⟨ F ε ( s ) , ψ ⟩ ] = ∫ 0 t ∧ s ⟨ S ( t , r ) ∗ φ , R ( 0 ) S ( s , r ) ∗ ψ ⟩ L 2 d r . (55)
Proof. Expanding the inner product,
E [ ⟨ F ε ( t ) , φ ⟩ ⟨ F ε ( s ) , ψ ⟩ ] = ε − d ∫ 0 t ∫ 0 s ∫ R 2 d S ‾ ( t , r ) ∗ φ ( x ) S ( s , r ′ ) ∗ ψ ( x ) R ( x − y ε ) u ‾ ( r , x ) u ‾ ( r ′ , y ) d x d y d r d r ′ . \mathbb{E}\left[\langle\mathcal{F}_\varepsilon(t),\varphi\rangle\langle\mathcal{F}_\varepsilon(s),\psi\rangle\right]=\varepsilon^{-d}\int_0^t\int_0^s\int_{\mathbb{R}^{2d}}\overline{S}(t,r)*\varphi(x)S(s,r')*\psi(x)R\left(\frac{x-y}{\varepsilon}\right)\overline{u}(r,x)\overline{u}(r',y)\,\mathrm{d}x\,\mathrm{d}y\,\mathrm{d}r\,\mathrm{d}r'. E [ ⟨ F ε ( t ) , φ ⟩ ⟨ F ε ( s ) , ψ ⟩ ] = ε − d ∫ 0 t ∫ 0 s ∫ R 2 d S ( t , r ) ∗ φ ( x ) S ( s , r ′ ) ∗ ψ ( x ) R ( ε x − y ) u ( r , x ) u ( r ′ , y ) d x d y d r d r ′ .
Rescaling x − y = ε z x-y=\varepsilon z x − y = ε z ,
∫ R d R ( x − y ε ) d y = ε d ∫ R d R ( z ) d z = ε d R ^ ( 0 ) . \int_{\mathbb{R}^d}R\left(\frac{x-y}{\varepsilon}\right)\,\mathrm{d}y=\varepsilon^d\int_{\mathbb{R}^d}R(z)\,\mathrm{d}z=\varepsilon^d\widehat{R}(0). ∫ R d R ( ε x − y ) d y = ε d ∫ R d R ( z ) d z = ε d R ( 0 ) .
Passing to Fourier space,
e − i ( ∣ ξ ∣ 2 + V ‾ ) ( t − r ) e i ( ∣ ξ ∣ 2 + V ‾ ) ( s − r ′ ) = e − i ( ∣ ξ ∣ 2 + V ‾ ) ( t − s ) e i ( ∣ ξ ∣ 2 + V ‾ ) ( r − r ′ ) . e^{-i(|\xi|^2+\overline{V})(t-r)}e^{i(|\xi|^2+\overline{V})(s-r')}=e^{-i(|\xi|^2+\overline{V})(t-s)}e^{i(|\xi|^2+\overline{V})(r-r')}. e − i ( ∣ ξ ∣ 2 + V ) ( t − r ) e i ( ∣ ξ ∣ 2 + V ) ( s − r ′ ) = e − i ( ∣ ξ ∣ 2 + V ) ( t − s ) e i ( ∣ ξ ∣ 2 + V ) ( r − r ′ ) .
The double time integral is
∫ 0 t ∫ 0 s e i λ ( r − r ′ ) d r d r ′ = ∫ − s t ( t ∧ s − ∣ u ∣ ) + e i λ u d u , λ : = ∣ ξ ∣ 2 + V ‾ . (56) \int_0^t\int_0^s e^{i\lambda(r-r')}\,\mathrm{d}r\,\mathrm{d}r'=\int_{-s}^{t}(t\wedge s-|u|)_+e^{i\lambda u}\,\mathrm{d}u,\qquad\lambda:=|\xi|^2+\overline{V}. \tag*{(56)} ∫ 0 t ∫ 0 s e iλ ( r − r ′ ) d r d r ′ = ∫ − s t ( t ∧ s − ∣ u ∣ ) + e iλ u d u , λ := ∣ ξ ∣ 2 + V . (56)
The triangular kernel ( t ∧ s − ∣ u ∣ ) + (t\wedge s-|u|)_+ ( t ∧ s − ∣ u ∣ ) + has Fourier transform 2 ( 1 − cos ( L λ ) ) λ 2 \frac{2(1-\cos(L\lambda))}{\lambda^2} λ 2 2 ( 1 − c o s ( L λ )) with L = t ∧ s L=t\wedge s L = t ∧ s . After the change of variables ξ ′ = ε ( ξ − ζ ) \xi'=\varepsilon(\xi-\zeta) ξ ′ = ε ( ξ − ζ ) , the ξ ′ \xi' ξ ′ -integral is
∫ R d R ^ ( ξ ′ ) 2 ( 1 − cos ( t ϕ ( ξ ′ ) ) ) ϕ ( ξ ′ ) 2 ∣ u ^ ( t , ξ ) ∣ 2 d ξ ′ , \int_{\mathbb{R}^d}\widehat{R}(\xi')\frac{2(1-\cos(t\phi(\xi')))}{\phi(\xi')^2}|\widehat{u}(t,\xi)|^2\,\mathrm{d}\xi', ∫ R d R ( ξ ′ ) ϕ ( ξ ′ ) 2 2 ( 1 − cos ( tϕ ( ξ ′ ))) ∣ u ( t , ξ ) ∣ 2 d ξ ′ ,
where ϕ ( ξ ′ ) = 2 ζ ⋅ ξ ′ ε + ∣ ξ ′ ∣ 2 ε 2 \phi(\xi') = \frac{2\zeta\cdot\xi'}{\varepsilon} + \frac{|\xi'|^2}{\varepsilon^2} ϕ ( ξ ′ ) = ε 2 ζ ⋅ ξ ′ + ε 2 ∣ ξ ′ ∣ 2 is the phase in (3.4). This is exactly the integral estimated in Lemma 3.2 and Lemma 3.3: the near region ∣ ξ ′ ∣ ≤ ε |\xi'| \le\varepsilon ∣ ξ ′ ∣ ≤ ε contributes O ( ε d ) O(\varepsilon^d) O ( ε d ) and identifies the diagonal R ^ ( 0 ) \widehat{R}(0) R ( 0 ) term, while the far region ∣ ξ ′ ∣ > ε |\xi'| > \varepsilon ∣ ξ ′ ∣ > ε contributes O ( ε d ) O(\varepsilon^d) O ( ε d ) for d ≤ 3 d \le3 d ≤ 3 by the dyadic shell bound 2 k ( d − 4 ) 2^{k(d-4)} 2 k ( d − 4 ) , and O ( ε d ) O(\varepsilon^d) O ( ε d ) for d ≥ 4 d \ge4 d ≥ 4 by the physical space argument of Lemma A.1. In particular, the off diagonal u ≠ 0 u \ne0 u = 0 contribution is O ( ε d ) O(\varepsilon^d) O ( ε d ) and vanishes after the prefactor ε − d \varepsilon^{-d} ε − d is applied. The diagonal u = 0 u = 0 u = 0 gives
∫ 0 t ∧ s e − i ( t − r ) λ e + i ( s − r ) λ d r = ( t ∧ s ) e − i ( t − s ) λ . \int_0^{t\wedge s} e^{-i(t-r)\lambda}e^{+i(s-r)\lambda}\,\mathrm{d}r = (t\wedge s)e^{-i(t-s)\lambda}. ∫ 0 t ∧ s e − i ( t − r ) λ e + i ( s − r ) λ d r = ( t ∧ s ) e − i ( t − s ) λ .
Therefore,
lim ε → 0 E [ ⟨ F ε ( t ) , ϕ ⟩ ⟨ F ε ( s ) , ψ ⟩ ‾ ] = ∫ 0 t ∧ s ⟨ S ( t , r ) ∗ ϕ , R ^ ( 0 ) S ( s , r ) ∗ ψ ⟩ L 2 d r . \lim_{\varepsilon\to0}\mathbb{E}\left[\langle\mathcal{F}_{\varepsilon}(t),\phi\rangle\overline{\langle\mathcal{F}_{\varepsilon}(s),\psi\rangle}\right] = \int_0^{t\wedge s}\langle S(t,r)^*\phi,\widehat{R}(0)S(s,r)^*\psi\rangle_{L^2}\,\mathrm{d}r. ε → 0 lim E [ ⟨ F ε ( t ) , ϕ ⟩ ⟨ F ε ( s ) , ψ ⟩ ] = ∫ 0 t ∧ s ⟨ S ( t , r ) ∗ ϕ , R ( 0 ) S ( s , r ) ∗ ψ ⟩ L 2 d r .
which proves Proposition 5.2. □
It remains to establish tightness in the path space.
Proposition 5.3. The family { F ε } ε > 0 \{\mathcal{F}_{\varepsilon}\}_{\varepsilon>0} { F ε } ε > 0 is tight in C ( [ 0 , T ] ; L 2 ( R d ) ) C([0,T];L^2(\mathbb{R}^d)) C ([ 0 , T ] ; L 2 ( R d )) .
Proof. By Lemma 2.10, it suffices to prove
E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 2 ] ≤ C ∣ t − s ∣ , E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 4 ] ≤ C ∣ t − s ∣ 2 . \mathbb{E}\left[\|\mathcal{F}_{\varepsilon}(t)-\mathcal{F}_{\varepsilon}(s)\|_{L_x^2}^2\right]\le C|t-s|,\qquad\mathbb{E}\left[\|\mathcal{F}_{\varepsilon}(t)-\mathcal{F}_{\varepsilon}(s)\|_{L_x^2}^4\right]\le C|t-s|^2. E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 2 ] ≤ C ∣ t − s ∣ , E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 4 ] ≤ C ∣ t − s ∣ 2 .
For s < t s<t s < t , decompose the increment as
F ε ( t ) − F ε ( s ) = − i ε − d 2 ∫ s t S ( t , r ) a ε ( r ) u ‾ ( r ) d r − i ε − d 2 ∫ 0 s ( S ( t , r ) − S ( s , r ) ) a ε ( r ) u ‾ ( r ) d r . \mathcal{F}_{\varepsilon}(t)-\mathcal{F}_{\varepsilon}(s)=-i\varepsilon^{-\frac d2}\int_s^t S(t,r)a_{\varepsilon}(r)\overline{u}(r)\,\mathrm{d}r-i\varepsilon^{-\frac d2}\int_0^s\bigl(S(t,r)-S(s,r)\bigr)a_{\varepsilon}(r)\overline{u}(r)\,\mathrm{d}r. F ε ( t ) − F ε ( s ) = − i ε − 2 d ∫ s t S ( t , r ) a ε ( r ) u ( r ) d r − i ε − 2 d ∫ 0 s ( S ( t , r ) − S ( s , r ) ) a ε ( r ) u ( r ) d r .
where the first term is denoted A ε ( t , s ) A_{\varepsilon}(t,s) A ε ( t , s ) and the second term is denoted B ε ( t , s ) B_{\varepsilon}(t,s) B ε ( t , s ) .
By Proposition 3.4 restricted to [ s , t ] [s,t] [ s , t ] ,
E [ ∥ A ε ( t , s ) ∥ L x 2 2 ] = ε − d E [ ∥ ∫ s t S ( t , r ) a ε ( r ) u ‾ ( r ) d r ∥ L x 2 2 ] ≤ ε − d ⋅ C ε d ∣ t − s ∣ ≤ C ∣ t − s ∣ . (57) \mathbb{E}\left[\|A_{\varepsilon}(t,s)\|_{L_x^2}^2\right]=\varepsilon^{-d}\mathbb{E}\left[\left\|\int_s^t S(t,r)a_{\varepsilon}(r)\overline{u}(r)\,\mathrm{d}r\right\|_{L_x^2}^2\right]\le\varepsilon^{-d}\cdot C\varepsilon^d|t-s|\le C|t-s|. \tag*{(57)} E [ ∥ A ε ( t , s ) ∥ L x 2 2 ] = ε − d E [ ∫ s t S ( t , r ) a ε ( r ) u ( r ) d r L x 2 2 ] ≤ ε − d ⋅ C ε d ∣ t − s ∣ ≤ C ∣ t − s ∣. (57)
Since S ( t , r ) − S ( s , r ) = S ( s , r ) ( S ( t − s ) − I ) S(t,r)-S(s,r)=S(s,r)(S(t-s)-I) S ( t , r ) − S ( s , r ) = S ( s , r ) ( S ( t − s ) − I ) , by unitarity,
∥ B ε ( t , s ) ∥ L x 2 2 = ε − d ∥ ∫ 0 s ( S ( t − s ) − I ) a ε ( r ) u ‾ ( r ) d r ∥ L x 2 2 . \|B_{\varepsilon}(t,s)\|_{L_x^2}^2=\varepsilon^{-d}\left\|\int_0^s(S(t-s)-I)a_{\varepsilon}(r)\overline{u}(r)\,\mathrm{d}r\right\|_{L_x^2}^2. ∥ B ε ( t , s ) ∥ L x 2 2 = ε − d ∫ 0 s ( S ( t − s ) − I ) a ε ( r ) u ( r ) d r L x 2 2 .
The multiplier is
e − i ϕ h − 1 , ϕ = ∣ ξ ∣ 2 + V ‾ , h = t − s , e^{-i\phi h}-1,\qquad\phi=|\xi|^2+\overline{V},\qquad h=t-s, e − i ϕ h − 1 , ϕ = ∣ ξ ∣ 2 + V , h = t − s ,
∣ e − i ϕ h − 1 ∣ 2 ≤ min ( 4 , ϕ 2 h 2 ) . |e^{-i\phi h}-1|^2\le\min(4,\phi^2h^2). ∣ e − i ϕ h − 1 ∣ 2 ≤ min ( 4 , ϕ 2 h 2 ) .
Near region ∣ ξ ′ ∣ ≤ ε |\xi'|\le\varepsilon ∣ ξ ′ ∣ ≤ ε :
∣ e − i ϕ h − 1 ∣ 2 ≤ ϕ 2 h 2 = O ( h 2 ) , contribution ≤ C ε d h 2 . |e^{-i\phi h}-1|^2\le\phi^2h^2=O(h^2),\qquad\text{contribution}\le C\varepsilon^d h^2. ∣ e − i ϕ h − 1 ∣ 2 ≤ ϕ 2 h 2 = O ( h 2 ) , contribution ≤ C ε d h 2 .
Far region ∣ ξ ′ ∣ > ε |\xi'|>\varepsilon ∣ ξ ′ ∣ > ε : using 2 ( 1 − cos ( t ϕ ) ) ϕ 2 ≤ 4 ϕ 2 \frac{2(1-\cos(t\phi))}{\phi^2}\le\frac{4}{\phi^2} ϕ 2 2 ( 1 − c o s ( tϕ )) ≤ ϕ 2 4 ,
∣ ∫ 0 s e i r σ d r ∣ 2 ≤ C ϕ 2 , ϕ 2 h 2 ⋅ ϕ − 2 = h 2 , contribution ≤ C ε d h 2 . \left|\int_0^s e^{ir\sigma}\,\mathrm{d}r\right|^2\le\frac{C}{\phi^2},\qquad\phi^2h^2\cdot\phi^{-2}=h^2,\qquad\text{contribution}\le C\varepsilon^d h^2. ∫ 0 s e i r σ d r 2 ≤ ϕ 2 C , ϕ 2 h 2 ⋅ ϕ − 2 = h 2 , contribution ≤ C ε d h 2 .
Therefore,
E [ ∥ B ε ( t , s ) ∥ L x 2 2 ] = ε − d E [ ∥ ∫ 0 s ( S ( t − s ) − I ) a ε ( r ) u ‾ ( r ) d r ∥ L x 2 2 ] ≤ ε − d ⋅ C ε d ∣ t − s ∣ 2 ≤ C ∣ t − s ∣ 2 . (58) \mathbb{E}\left[\|B_{\varepsilon}(t,s)\|_{L_x^2}^2\right]=\varepsilon^{-d}\mathbb{E}\left[\left\|\int_0^s(S(t-s)-I)a_{\varepsilon}(r)\overline{u}(r)\,\mathrm{d}r\right\|_{L_x^2}^2\right]\le\varepsilon^{-d}\cdot C\varepsilon^d|t-s|^2\le C|t-s|^2. \tag*{(58)} E [ ∥ B ε ( t , s ) ∥ L x 2 2 ] = ε − d E [ ∫ 0 s ( S ( t − s ) − I ) a ε ( r ) u ( r ) d r L x 2 2 ] ≤ ε − d ⋅ C ε d ∣ t − s ∣ 2 ≤ C ∣ t − s ∣ 2 . (58)
By (5.6) and (5.7),
E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 2 ] ≤ 2 E [ ∥ A ε ∥ L x 2 2 ] + 2 E [ ∥ B ε ∥ L x 2 2 ] ≤ C ∣ t − s ∣ + C ∣ t − s ∣ 2 ≤ C ∣ t − s ∣ . \mathbb{E}\left[\|\mathcal{F}_{\varepsilon}(t)-\mathcal{F}_{\varepsilon}(s)\|_{L_x^2}^2\right]\le2\mathbb{E}\left[\|A_{\varepsilon}\|_{L_x^2}^2\right]+2\mathbb{E}\left[\|B_{\varepsilon}\|_{L_x^2}^2\right]\le C|t-s|+C|t-s|^2\le C|t-s|. E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 2 ] ≤ 2 E [ ∥ A ε ∥ L x 2 2 ] + 2 E [ ∥ B ε ∥ L x 2 2 ] ≤ C ∣ t − s ∣ + C ∣ t − s ∣ 2 ≤ C ∣ t − s ∣.
By Lemma 2.13,
E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 2 ] ≤ C ∣ t − s ∣ 2 . \mathbb{E}\left[\left\|\mathcal{F}_{\varepsilon}(t)-\mathcal{F}_{\varepsilon}(s)\right\|_{L_x^2}^2\right]\le C|t-s|^2. E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 2 ] ≤ C ∣ t − s ∣ 2 .
By Lemma 2.10 with p = 4 p=4 p = 4 , α = 1 2 \alpha=\frac{1}{2} α = 2 1 , the family is relatively compact in C 0 , 1 2 − η ( [ 0 , T ] ; L 2 ( R d ) ) C^{0,\frac{1}{2}-\eta}([0,T];L^2(\mathbb{R}^d)) C 0 , 2 1 − η ([ 0 , T ] ; L 2 ( R d )) for any η > 0 \eta>0 η > 0 , hence in C ( [ 0 , T ] ; L 2 ( R d ) ) C([0,T];L^2(\mathbb{R}^d)) C ([ 0 , T ] ; L 2 ( R d )) .
Proof of Theorem 1.10. Step 1. Fix 0 ≤ t 1 < ⋯ < t k ≤ T 0\le t_1<\cdots<t_k\le T 0 ≤ t 1 < ⋯ < t k ≤ T and φ 1 , … , φ k ∈ L 2 ( R d ) \varphi_1,\ldots,\varphi_k\in L^2(\mathbb{R}^d) φ 1 , … , φ k ∈ L 2 ( R d ) . The block decomposition is
F ε ( t ) = ∑ j = 1 N ε M j ε ( t ) + ρ ε ( t ) , E [ ∥ ρ ε ( t ) ∥ L x 2 2 ] ≤ C L ε − d . \mathcal{F}_{\varepsilon}(t)=\sum_{j=1}^{N_{\varepsilon}}M_j^{\varepsilon}(t)+\rho_{\varepsilon}(t),\qquad
\mathbb{E}\left[\left\|\rho_{\varepsilon}(t)\right\|_{L_x^2}^2\right]\le CL_{\varepsilon}^{-d}. F ε ( t ) = j = 1 ∑ N ε M j ε ( t ) + ρ ε ( t ) , E [ ∥ ρ ε ( t ) ∥ L x 2 2 ] ≤ C L ε − d .
Thus L ε − d 2 = ( κ 2 N ( − log ε ) ) d 2 → 0 L_{\varepsilon}^{-\frac{d}{2}}=\left(\frac{\kappa}{2N(-\log\varepsilon)}\right)^{\frac{d}{2}}\to0 L ε − 2 d = ( 2 N ( − l o g ε ) κ ) 2 d → 0 , and
E [ ∥ ρ ε ( t ) ∥ L x 2 ] ≤ C L ε − d 2 → 0. \mathbb{E}\left[\left\|\rho_{\varepsilon}(t)\right\|_{L_x^2}\right]\le CL_{\varepsilon}^{-\frac{d}{2}}\to0. E [ ∥ ρ ε ( t ) ∥ L x 2 ] ≤ C L ε − 2 d → 0.
By Lemma 2.5, it suffices to prove the convergence of
( ⟨ ∑ j M j ε ( t 1 ) , φ 1 ⟩ , … , ⟨ ∑ j M j ε ( t k ) , φ k ⟩ ) . \left(\left\langle\sum_j M_j^{\varepsilon}(t_1),\varphi_1\right\rangle,\ldots,\left\langle\sum_j M_j^{\varepsilon}(t_k),\varphi_k\right\rangle\right). ( ⟨ j ∑ M j ε ( t 1 ) , φ 1 ⟩ , … , ⟨ j ∑ M j ε ( t k ) , φ k ⟩ ) .
The array { M j ε ( t ) } j = 1 N ε \{M_j^{\varepsilon}(t)\}_{j=1}^{N_{\varepsilon}} { M j ε ( t ) } j = 1 N ε is adapted to { G j } \{\mathcal{G}_j\} { G j } . By Proposition 5.1,
∑ j = 1 N ε E [ ∣ M j ε ( t ) ∣ 2 1 { ∣ M j ε ( t ) ∣ > η } ] ⟶ 0 for every η > 0. (59) \sum_{j=1}^{N_{\varepsilon}}\mathbb{E}\left[\left|M_j^{\varepsilon}(t)\right|^2\mathbf{1}_{\{|M_j^{\varepsilon}(t)|>\eta\}}\right]\longrightarrow0\qquad\text{for every }\eta>0. \tag*{(59)} j = 1 ∑ N ε E [ M j ε ( t ) 2 1 { ∣ M j ε ( t ) ∣ > η } ] ⟶ 0 for every η > 0. (59)
By Proposition 5.2, the quadratic variation converges in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) to its limit,
lim ε → 0 E [ ∣ ⟨ ∑ j = 1 N ε M j ε ( t ) ⟩ − ∫ 0 t ∧ T ⟨ S ( t − r ) ∗ φ , R ^ ( 0 ) S ( s − r ) ∗ ψ ⟩ L x 2 d r ∣ 2 ] = 0. \lim_{\varepsilon\to0}\mathbb{E}\left[\left|\left\langle\sum_{j=1}^{N_{\varepsilon}}M_j^{\varepsilon}(t)\right\rangle-\int_0^{t\wedge T}\left\langle S(t-r)^*\varphi,\widehat{R}(0)S(s-r)^*\psi\right\rangle_{L_x^2}\,\mathrm{d}r\right|^2\right]=0. ε → 0 lim E ⟨ j = 1 ∑ N ε M j ε ( t ) ⟩ − ∫ 0 t ∧ T ⟨ S ( t − r ) ∗ φ , R ( 0 ) S ( s − r ) ∗ ψ ⟩ L x 2 d r 2 = 0.
By Chebyshev’s inequality, for every δ > 0 \delta>0 δ > 0 ,
lim ε → 0 P ( ∣ ⟨ ∑ j = 1 N ε M j ε ( t ) ⟩ − ∫ 0 t ∧ T ⟨ S ( t − r ) ∗ φ , R ^ ( 0 ) S ( s − r ) ∗ ψ ⟩ L x 2 d r ∣ > δ ) = 0. (60) \lim_{\varepsilon\to0}\mathbb{P}\left(\left|\left\langle\sum_{j=1}^{N_{\varepsilon}}M_j^{\varepsilon}(t)\right\rangle-\int_0^{t\wedge T}\left\langle S(t-r)^*\varphi,\widehat{R}(0)S(s-r)^*\psi\right\rangle_{L_x^2}\,\mathrm{d}r\right|>\delta\right)=0. \tag*{(60)} ε → 0 lim P ( ⟨ j = 1 ∑ N ε M j ε ( t ) ⟩ − ∫ 0 t ∧ T ⟨ S ( t − r ) ∗ φ , R ( 0 ) S ( s − r ) ∗ ψ ⟩ L x 2 d r > δ ) = 0. (60)
The vector valued array ( M j ε ( t 1 ) , … , M j ε ( t k ) ) (M_j^{\varepsilon}(t_1),\ldots,M_j^{\varepsilon}(t_k)) ( M j ε ( t 1 ) , … , M j ε ( t k )) satisfies (5.8) componentwise, and its covariance matrix converges to
Σ l m ( t ) = ∫ 0 t l ∧ t m ⟨ S ( t l − r ) ∗ φ l , R ^ ( 0 ) S ( t m − r ) ∗ φ m ⟩ L x 2 d r . (61) \Sigma_{lm}(t)=\int_0^{t_l\wedge t_m}\left\langle S(t_l-r)^*\varphi_l,\widehat{R}(0)S(t_m-r)^*\varphi_m\right\rangle_{L_x^2}\,\mathrm{d}r. \tag*{(61)} Σ l m ( t ) = ∫ 0 t l ∧ t m ⟨ S ( t l − r ) ∗ φ l , R ( 0 ) S ( t m − r ) ∗ φ m ⟩ L x 2 d r . (61)
Applying Corollary 2.6, for every bounded continuous F : R k → R F:\mathbb{R}^k\to\mathbb{R} F : R k → R ,
lim ε → 0 E [ F ( ⟨ F ε ( t 1 ) , φ 1 ⟩ , … , ⟨ F ε ( t k ) , φ k ⟩ ) ] = E [ F ( Z ) ] , Z ∼ N ( 0 , Σ ( t ) ) , (62) \lim_{\varepsilon\to0}\mathbb{E}\left[F\left(\langle\mathcal{F}_{\varepsilon}(t_1),\varphi_1\rangle,\ldots,\langle\mathcal{F}_{\varepsilon}(t_k),\varphi_k\rangle\right)\right]=\mathbb{E}[F(Z)],\qquad Z\sim\mathcal{N}(0,\Sigma(t)), \tag*{(62)} ε → 0 lim E [ F ( ⟨ F ε ( t 1 ) , φ 1 ⟩ , … , ⟨ F ε ( t k ) , φ k ⟩ ) ] = E [ F ( Z )] , Z ∼ N ( 0 , Σ ( t )) , (62)
where Σ ( t ) \Sigma(t) Σ ( t ) is the matrix (5.10), i.e. the finite dimensional distribution of G \mathcal{G} G .
Step 2. By Proposition 5.3,
E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 4 ] ≤ C ∣ t − s ∣ 2 , 0 ≤ s , t ≤ T . (63) \mathbb{E}\left[\left\|\mathcal{F}_{\varepsilon}(t)-\mathcal{F}_{\varepsilon}(s)\right\|_{L_x^2}^4\right]\le C|t-s|^2,\qquad0\le s,t\le T. \tag*{(63)} E [ ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 4 ] ≤ C ∣ t − s ∣ 2 , 0 ≤ s , t ≤ T . (63)
By the Kolmogorov–Chentsov criterion (Lemma 2.10) with p = 4 p=4 p = 4 , α = 1 2 \alpha=\frac{1}{2} α = 2 1 , the paths satisfy
E [ sup s ≠ t ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 ∣ t − s ∣ 1 2 − η ] < ∞ for any η > 0 , \mathbb{E}\left[\sup_{s\ne t}\frac{\left\|\mathcal{F}_{\varepsilon}(t)-\mathcal{F}_{\varepsilon}(s)\right\|_{L_x^2}}{|t-s|^{\frac{1}{2}-\eta}}\right]<\infty\qquad\text{for any }\eta>0, E [ s = t sup ∣ t − s ∣ 2 1 − η ∥ F ε ( t ) − F ε ( s ) ∥ L x 2 ] < ∞ for any η > 0 ,
uniformly in ε \varepsilon ε . Hence { F ε } \{F_\varepsilon\} { F ε } is tight in C ( [ 0 , T ] ; L 2 ( R d ) ) C([0,T];L^2(\mathbb{R}^d)) C ([ 0 , T ] ; L 2 ( R d )) .
Step 3. By Prokhorov’s theorem, for any sequence ε n → 0 \varepsilon_n \to0 ε n → 0 there exists a subsequence ε n ′ \varepsilon_{n'} ε n ′ and a random field G ′ G' G ′ such that for every bounded continuous functional Φ : C ( [ 0 , T ] ; L 2 ( R d ) ) → R \Phi: C([0,T];L^2(\mathbb{R}^d)) \to\mathbb{R} Φ : C ([ 0 , T ] ; L 2 ( R d )) → R ,
lim n ′ → ∞ E [ Φ ( F ε n ′ ) ] = E [ Φ ( G ′ ) ] . \lim_{n' \to\infty} \mathbb{E}\left[\Phi(F_{\varepsilon_{n'}})\right] = \mathbb{E}\left[\Phi(G')\right]. n ′ → ∞ lim E [ Φ ( F ε n ′ ) ] = E [ Φ ( G ′ ) ] .
By Step 1, for every k k k , t 1 , … , t k t_1,\ldots,t_k t 1 , … , t k , φ 1 , … , φ k \varphi_1,\ldots,\varphi_k φ 1 , … , φ k , the finite dimensional distributions of G ′ G' G ′ and G G G agree, i.e.,
( ⟨ G ′ ( t 1 ) , φ 1 ⟩ , … , ⟨ G ′ ( t k ) , φ k ⟩ ) \left(\langle G'(t_1),\varphi_1\rangle,\ldots,\langle G'(t_k),\varphi_k\rangle\right) ( ⟨ G ′ ( t 1 ) , φ 1 ⟩ , … , ⟨ G ′ ( t k ) , φ k ⟩ )
and
( ⟨ G ( t 1 ) , φ 1 ⟩ , … , ⟨ G ( t k ) , φ k ⟩ ) \left(\langle G(t_1),\varphi_1\rangle,\ldots,\langle G(t_k),\varphi_k\rangle\right) ( ⟨ G ( t 1 ) , φ 1 ⟩ , … , ⟨ G ( t k ) , φ k ⟩ )
have the same distribution on R k \mathbb{R}^k R k . The Borel σ \sigma σ -algebra of C ( [ 0 , T ] ; L 2 ( R d ) ) C([0,T];L^2(\mathbb{R}^d)) C ([ 0 , T ] ; L 2 ( R d )) is generated by the finite dimensional projections. Hence G ′ G' G ′ and G G G have the same law on C ( [ 0 , T ] ; L 2 ( R d ) ) C([0,T];L^2(\mathbb{R}^d)) C ([ 0 , T ] ; L 2 ( R d )) . Every convergent subsequence has the same limit, so for every bounded continuous Φ : C ( [ 0 , T ] ; L 2 ( R d ) ) → R \Phi: C([0,T];L^2(\mathbb{R}^d)) \to\mathbb{R} Φ : C ([ 0 , T ] ; L 2 ( R d )) → R ,
lim ε → 0 E [ Φ ( F ε ) ] = E [ Φ ( G ) ] . \lim_{\varepsilon\to0} \mathbb{E}\left[\Phi(F_\varepsilon)\right] = \mathbb{E}\left[\Phi(G)\right]. ε → 0 lim E [ Φ ( F ε ) ] = E [ Φ ( G ) ] .
This proves Theorem 1.10.
The second moment in dimensions d ≥ 4 d \ge4 d ≥ 4 : a physical space argument
In the main text the near/far split in Fourier space is used only for d ≤ 3 d \le3 d ≤ 3 , where the radial integral ∫ ε ∞ r d − 5 d r \int_\varepsilon^\infty r^{d-5}\,dr ∫ ε ∞ r d − 5 d r converges. For d ≥ 4 d \ge4 d ≥ 4 , the crude bound 1 − cos u ≤ 2 1-\cos u \le2 1 − cos u ≤ 2 discards the oscillation and the dyadic sum diverges, because R ^ \widehat{R} R is only known to be bounded. This appendix records the physical space oscillatory integral argument that, under the smoothness supplied by the homogenized solution, yields the same upper bound (34) in all dimensions. We state explicitly which steps rely on standard oscillatory integral estimates and which are left as regularity consequences of u 0 ∈ C c ∞ u_0 \in C^\infty_c u 0 ∈ C c ∞ .
Physical space second moment
The first order term is
w ε ( 1 ) ( t , x ) = − i ∫ 0 t ∫ R d K t − r ( x − y ) a ε ( y ) u ‾ ( r , y ) d y d r , a ε ( y ) : = a ( y ε ) . (64) w_\varepsilon^{(1)}(t,x)=-i\int_0^t\int_{\mathbb{R}^d}K_{t-r}(x-y)a_\varepsilon(y)\overline{u}(r,y)\,dy\,dr,\qquad a_\varepsilon(y):=a\left(\frac{y}{\varepsilon}\right). \tag*{(64)} w ε ( 1 ) ( t , x ) = − i ∫ 0 t ∫ R d K t − r ( x − y ) a ε ( y ) u ( r , y ) d y d r , a ε ( y ) := a ( ε y ) . (64)
with the Fresnel kernel
K τ ( y ) = ( 4 π i τ ) − d 2 exp ( i ∣ y ∣ 2 4 τ ) , F ( K τ ) ( ξ ) = e − i τ ∣ ξ ∣ 2 . (65) K_\tau(y)=(4\pi i\tau)^{-\frac{d}{2}}\exp\left(\frac{i|y|^2}{4\tau}\right),\qquad\mathcal{F}(K_\tau)(\xi)=e^{-i\tau|\xi|^2}. \tag*{(65)} K τ ( y ) = ( 4 π i τ ) − 2 d exp ( 4 τ i ∣ y ∣ 2 ) , F ( K τ ) ( ξ ) = e − i τ ∣ ξ ∣ 2 . (65)
Multiplying (A.1) by its conjugate, integrating in x x x , and using
E [ a ε ( y ) a ε ( z ) ] = R ( y − z ε ) , \mathbb{E}[a_\varepsilon(y)a_\varepsilon(z)]=R\left(\frac{y-z}{\varepsilon}\right), E [ a ε ( y ) a ε ( z )] = R ( ε y − z ) ,
the x x x -integral is evaluated with the unitarity and semigroup property of K K K :
∫ R d K t − r ( x − y ) K t − s ( x − z ) ‾ d x = K r − s ( z − y ) . (66) \int_{\mathbb{R}^d}K_{t-r}(x-y)\overline{K_{t-s}(x-z)}\,dx=K_{r-s}(z-y). \tag*{(66)} ∫ R d K t − r ( x − y ) K t − s ( x − z ) d x = K r − s ( z − y ) . (66)
Indeed, inserting (A.2) and integrating in x x x produces the factor ( 2 π ) d δ ( ξ − η ) (2\pi)^d\delta(\xi-\eta) ( 2 π ) d δ ( ξ − η ) , and the remaining phase is e i ( r − s ) ∣ ξ ∣ 2 e^{i(r-s)|\xi|^2} e i ( r − s ) ∣ ξ ∣ 2 . Setting y ′ = y − z ε y'=\frac{y-z}{\varepsilon} y ′ = ε y − z and y = z + ε y ′ y=z+\varepsilon y' y = z + ε y ′ , so that d y = ε d d y ′ dy=\varepsilon^d dy' d y = ε d d y ′ , gives
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] = ε d ∫ 0 t ∫ 0 t ∫ R d K r − s ( ε y ′ ) ‾ R ( y ′ ) C r , s ( y ′ ) d y ′ d s d r . (67) \mathbb{E}\left[\left\|w_\varepsilon^{(1)}(t)\right\|_{L_x^2}^2\right]=\varepsilon^d\int_0^t\int_0^t\int_{\mathbb{R}^d}\overline{K_{r-s}(\varepsilon y')}R(y')C_{r,s}(y')\,dy'\,ds\,dr. \tag*{(67)} E [ w ε ( 1 ) ( t ) L x 2 2 ] = ε d ∫ 0 t ∫ 0 t ∫ R d K r − s ( ε y ′ ) R ( y ′ ) C r , s ( y ′ ) d y ′ d s d r . (67)
where the inner z z z -integral is the space auto-correlation
C r , s ( y ′ ) : = ∫ R d u ‾ ( r , z + ε y ′ ) u ( s , z ) d z = ⟨ u ‾ ( r , ⋅ + ε y ′ ) , u ‾ ( s , ⋅ ) ⟩ L x 2 . (68) C_{r,s}(y'):=\int_{\mathbb{R}^d}\overline{u}(r,z+\varepsilon y')u(s,z)\,dz=\langle\overline{u}(r,\cdot+\varepsilon y'),\overline{u}(s,\cdot)\rangle_{L_x^2}. \tag*{(68)} C r , s ( y ′ ) := ∫ R d u ( r , z + ε y ′ ) u ( s , z ) d z = ⟨ u ( r , ⋅ + ε y ′ ) , u ( s , ⋅ ) ⟩ L x 2 . (68)
We now justify the regularity of the family { C r , s } \{C_{r,s}\} { C r , s } . Since u 0 ∈ C c ∞ u_0 \in C^\infty_c u 0 ∈ C c ∞ , the free Schrödinger flow u ˉ ( r ) = e i r Δ u 0 \bar{u}(r)=e^{ir\Delta}u_0 u ˉ ( r ) = e i r Δ u 0 preserves the Schwartz class on compact time intervals, together with all its time derivatives. For fixed r , s ∈ [ 0 , t ] r,s \in[0,t] r , s ∈ [ 0 , t ] , C r , s ( y ′ ) C_{r,s}(y') C r , s ( y ′ ) is the translation autocorrelation of two Schwartz functions, hence is Schwartz in y ′ y' y ′ . Moreover, every y ′ y' y ′ -derivative and every ( r , s ) (r,s) ( r , s ) -derivative is bounded uniformly in r , s ∈ [ 0 , t ] r,s \in[0,t] r , s ∈ [ 0 , t ] and in ε \varepsilon ε . By the exponential mixing in Assumption 1.4 (ii), R R R decays exponentially and is smooth. The amplitude
b r , s ( y ′ ) : = R ( y ′ ) C r , s ( y ′ ) (69) b_{r,s}(y') := R(y')C_{r,s}(y') \tag*{(69)} b r , s ( y ′ ) := R ( y ′ ) C r , s ( y ′ ) (69)
is therefore uniformly Schwartz, and in particular belongs to W m , 1 ( R d ) W^{m,1}(\mathbb{R}^d) W m , 1 ( R d ) for every m m m , with norm bounded uniformly in r , s r,s r , s and ε \varepsilon ε .
A uniform oscillatory integral bound
Define the operator obtained from (A.4) by removing the prefactor ε d \varepsilon^d ε d :
T ε b : = ∫ 0 t ∫ 0 t ∫ R d b r , s ( y ′ ) K r − s ( ε y ′ ) ‾ d y ′ d s d r . (70) T_\varepsilon b := \int_0^t \int_0^t \int_{\mathbb{R}^d} b_{r,s}(y')\overline{K_{r-s}(\varepsilon y')} \,dy'\,ds\,dr. \tag*{(70)} T ε b := ∫ 0 t ∫ 0 t ∫ R d b r , s ( y ′ ) K r − s ( ε y ′ ) d y ′ d s d r . (70)
The following is the physical space analogue of the dyadic near/far estimate. We record it as a lemma, but the reader should note that it is a soft operator bound: it controls the double time integral without identifying the leading constant.
Lemma A.1. Fix an integer M ≥ 1 M \ge1 M ≥ 1 . Let { b r , s } r , s ∈ [ 0 , t ] \{b_{r,s}\}_{r,s\in[0,t]} { b r , s } r , s ∈ [ 0 , t ] be a family uniformly bounded, together with all mixed derivatives ∂ r j ∂ s k b r , s \partial_r^j\partial_s^k b_{r,s} ∂ r j ∂ s k b r , s , for 0 ≤ j , k ≤ M 0 \le j,k \le M 0 ≤ j , k ≤ M , in W m , 1 ( R d ) W^{m,1}(\mathbb{R}^d) W m , 1 ( R d ) with m = 2 M + d + 2 m=2M+d+2 m = 2 M + d + 2 . Then,
∣ T ε b ∣ ≤ C d , t , M sup 0 ≤ j , k ≤ M sup r , s ∥ ∂ r j ∂ s k b r , s ∥ W m , 1 , |T_\varepsilon b| \le C_{d,t,M}\sup_{0\le j,k\le M}\sup_{r,s}\|\partial_r^j\partial_s^k b_{r,s}\|_{W^{m,1}}, ∣ T ε b ∣ ≤ C d , t , M 0 ≤ j , k ≤ M sup r , s sup ∥ ∂ r j ∂ s k b r , s ∥ W m , 1 , uniformly in ε > 0 \varepsilon>0 ε > 0 . The estimate is uniform in ε \varepsilon ε precisely when 4 M > d 4M>d 4 M > d .
Proof. By Plancherel in y ′ y' y ′ for each fixed ( r , s ) (r,s) ( r , s ) , with τ = r − s \tau=r-s τ = r − s ,
∫ R d b r , s ( y ′ ) K τ ( ε y ′ ) d y ′ = ( 2 π ) − d ∫ R d b ^ r , s ( − ε ξ ) e i τ ∣ ξ ∣ 2 d ξ . \int_{\mathbb{R}^d} b_{r,s}(y')K_\tau(\varepsilon y')\,dy'=(2\pi)^{-d}\int_{\mathbb{R}^d}\widehat{b}_{r,s}(-\varepsilon\xi)e^{i\tau|\xi|^2}\,d\xi. ∫ R d b r , s ( y ′ ) K τ ( ε y ′ ) d y ′ = ( 2 π ) − d ∫ R d b r , s ( − ε ξ ) e i τ ∣ ξ ∣ 2 d ξ .
We keep the amplitude inside the time integral, since b ^ r , s \widehat{b}_{r,s} b r , s depends on ( r , s ) (r,s) ( r , s ) . Define
F ε ( ξ ) : = ∫ 0 t ∫ 0 t b ^ r , s ( − ε ξ ) e i ( r − s ) ∣ ξ ∣ 2 d s d r , T ε b = ( 2 π ) − d ∫ R d F ε ( ξ ) d ξ . F_\varepsilon(\xi):=\int_0^t\int_0^t\widehat{b}_{r,s}(-\varepsilon\xi)e^{i(r-s)|\xi|^2}\,ds\,dr,\qquad T_\varepsilon b=(2\pi)^{-d}\int_{\mathbb{R}^d}F_\varepsilon(\xi)\,d\xi. F ε ( ξ ) := ∫ 0 t ∫ 0 t b r , s ( − ε ξ ) e i ( r − s ) ∣ ξ ∣ 2 d s d r , T ε b = ( 2 π ) − d ∫ R d F ε ( ξ ) d ξ .
We bound F ε ( ξ ) F_\varepsilon(\xi) F ε ( ξ ) by integration by parts in the time variables. Write λ = ∣ ξ ∣ 2 \lambda=|\xi|^2 λ = ∣ ξ ∣ 2 and B r , s ( ξ ) : = b ^ r , s ( − ε ξ ) B_{r,s}(\xi):=\widehat{b}_{r,s}(-\varepsilon\xi) B r , s ( ξ ) := b r , s ( − ε ξ ) . We integrate by parts M M M times in r r r and M M M times in s s s , using ∂ r e i r λ = i λ e i r λ \partial_r e^{ir\lambda}=i\lambda e^{ir\lambda} ∂ r e i r λ = iλ e i r λ and ∂ s e − i s λ = − i λ e − i s λ \partial_s e^{-is\lambda}=-i\lambda e^{-is\lambda} ∂ s e − i s λ = − iλ e − i s λ . Each integration by parts brings down one factor λ − 1 \lambda^{-1} λ − 1 and transfers one derivative onto B r , s B_{r,s} B r , s . The boundary evaluations on the square [ 0 , t ] 2 [0,t]^2 [ 0 , t ] 2 and its edges are themselves oscillatory one dimensional integrals and are integrated by parts further, at the cost of one more λ − 1 \lambda^{-1} λ − 1 . After M M M integrations in each variable, collecting all boundary terms, we obtain
∣ F ε ( ξ ) ∣ ≤ C λ 2 M ∑ 0 ≤ j , k ≤ M t j + k sup [ 0 , t ] 2 ∣ ∂ r j ∂ s k B r , s ( ξ ) ∣ . (71) |F_\varepsilon(\xi)|\le\frac{C}{\lambda^{2M}}\sum_{0\le j,k\le M}t^{j+k}\sup_{[0,t]^2}|\partial_r^j\partial_s^k B_{r,s}(\xi)|. \tag*{(71)} ∣ F ε ( ξ ) ∣ ≤ λ 2 M C 0 ≤ j , k ≤ M ∑ t j + k [ 0 , t ] 2 sup ∣ ∂ r j ∂ s k B r , s ( ξ ) ∣. (71)
where every term is evaluated at ( − ε ξ ) (-\varepsilon\xi) ( − ε ξ ) and C C C depends only on t t t and M M M . Since λ = ∣ ξ ∣ 2 \lambda=|\xi|^2 λ = ∣ ξ ∣ 2 , the prefactor in (A.8) is ∣ ξ ∣ − 4 M |\xi|^{-4M} ∣ ξ ∣ − 4 M . We now split the ξ \xi ξ -integral. On ∣ ξ ∣ ≤ 1 |\xi|\le1 ∣ ξ ∣ ≤ 1 , we do not integrate by parts. Directly,
∣ F ε ( ξ ) ∣ ≤ t 2 sup r , s ∥ B r , s ∥ L ∞ , |F_\varepsilon(\xi)|\le t^2\sup_{r,s}\|B_{r,s}\|_{L^\infty}, ∣ F ε ( ξ ) ∣ ≤ t 2 r , s sup ∥ B r , s ∥ L ∞ ,
which is integrable on the unit ball. On ∣ ξ ∣ > 1 |\xi|>1 ∣ ξ ∣ > 1 , we use (A.8), namely ∣ F ε ( ξ ) ∣ ≤ C ∣ ξ ∣ − 4 M M ( ξ ) |F_\varepsilon(\xi)|\le C|\xi|^{-4M}M(\xi) ∣ F ε ( ξ ) ∣ ≤ C ∣ ξ ∣ − 4 M M ( ξ ) with
M ( ξ ) : = ∑ 0 ≤ j , k ≤ M sup r , s ∣ ∂ r j ∂ s k B r , s ( ξ ) ∣ . M(\xi):=\sum_{0\le j,k\le M}\sup_{r,s}|\partial_r^j\partial_s^k B_{r,s}(\xi)|. M ( ξ ) := 0 ≤ j , k ≤ M ∑ r , s sup ∣ ∂ r j ∂ s k B r , s ( ξ ) ∣.
Because b r , s b_{r,s} b r , s is uniformly Schwartz, each ∂ r j ∂ s k b r , s \partial_r^j\partial_s^k b_{r,s} ∂ r j ∂ s k b r , s decays faster than any power of ∣ ξ ∣ |\xi| ∣ ξ ∣ , uniformly in ( r , s ) (r,s) ( r , s ) and in ε \varepsilon ε . Hence M ( ξ ) M(\xi) M ( ξ ) decays faster than any power. On the intermediate range 1 < ∣ ξ ∣ < 1 / ε 1<|\xi|<1/\varepsilon 1 < ∣ ξ ∣ < 1/ ε , we use M ( ξ ) ≤ C M(\xi)\le C M ( ξ ) ≤ C and obtain
∫ 1 < ∣ ξ ∣ < 1 / ε ∣ ξ ∣ − 4 M d ξ ≲ ∫ 1 1 / ε ρ d − 1 − 4 M d ρ , \int_{1<|\xi|<1/\varepsilon}|\xi|^{-4M}\,\mathrm{d}\xi\lesssim\int_1^{1/\varepsilon}\rho^{d-1-4M}\,\mathrm{d}\rho, ∫ 1 < ∣ ξ ∣ < 1/ ε ∣ ξ ∣ − 4 M d ξ ≲ ∫ 1 1/ ε ρ d − 1 − 4 M d ρ ,
which is bounded uniformly in ε \varepsilon ε exactly when d − 4 M < 0 d-4M<0 d − 4 M < 0 , i.e. 4 M > d 4M>d 4 M > d . The tail ∣ ξ ∣ > 1 / ε |\xi|>1/\varepsilon ∣ ξ ∣ > 1/ ε is controlled by the rapid decay of M ( ξ ) M(\xi) M ( ξ ) . Choosing M = ⌈ d 4 ⌉ + 1 M=\left\lceil\frac{d}{4}\right\rceil+1 M = ⌈ 4 d ⌉ + 1 gives the claim. □
Remark A.2. The order M = ⌈ d 4 ⌉ + 1 M=\left\lceil\frac{d}{4}\right\rceil+1 M = ⌈ 4 d ⌉ + 1 is the only dimension dependent choice: each pair of time integrations by parts gains four powers of ∣ ξ ∣ − 1 |\xi|^{-1} ∣ ξ ∣ − 1 , so M M M pairs win over the radial growth ρ d − 1 \rho^{d-1} ρ d − 1 precisely when 4 M > d 4M>d 4 M > d . The required time derivatives ∂ r j ∂ s k b r , s \partial_r^j\partial_s^k b_{r,s} ∂ r j ∂ s k b r , s exist uniformly because u ‾ ( r ) = e i r Δ u 0 \overline{u}(r)=e^{ir\Delta}u_0 u ( r ) = e i r Δ u 0 with u 0 ∈ C c ∞ u_0\in C^\infty_c u 0 ∈ C c ∞ , and the free Schrödinger flow preserves the Schwartz class together with all its time derivatives. We have not written out the Sobolev embedding that converts these uniform bounds into the rapid decay of b ^ r , s \widehat{b}_{r,s} b r , s , which is a standard consequence. The integration by parts (A.8) is the one dimensional van der Corput lemma applied 2 M 2M 2 M times. The standard oscillatory integral estimates used here may be found in Reed–Simon [22 Chapter 8] and Sogge [24 ] . The conclusion is an upper bound only.
The diagonal contribution and the crude stationary phase bound
For r ≠ s r\ne s r = s the phase in the physical space integral is
ϕ r , s ( y ′ ) : = ε 2 ∣ y ′ ∣ 2 4 ( r − s ) , \phi_{r,s}(y'):=\frac{\varepsilon^2|y'|^2}{4(r-s)}, ϕ r , s ( y ′ ) := 4 ( r − s ) ε 2 ∣ y ′ ∣ 2 ,
whose Hessian equals ε 2 2 ( r − s ) I \frac{\varepsilon^2}{2(r-s)}I 2 ( r − s ) ε 2 I . A non-degenerate stationary phase estimate gives
∣ ∫ R d b r , s ( y ′ ) e i ϕ r , s ( y ′ ) d y ′ ∣ ≤ C ∣ ε 2 r − s ∣ − d 2 ∥ b ∥ H d + 2 . (72) \left|\int_{\mathbb{R}^d}b_{r,s}(y')e^{i\phi_{r,s}(y')}\,\mathrm{d}y'\right|\le C\left|\frac{\varepsilon^2}{r-s}\right|^{-\frac{d}{2}}\|b\|_{H^{d+2}}. \tag*{(72)} ∫ R d b r , s ( y ′ ) e i ϕ r , s ( y ′ ) d y ′ ≤ C r − s ε 2 − 2 d ∥ b ∥ H d + 2 . (72)
After multiplication by the prefactor ( 4 π i ( r − s ) ) − d 2 (4\pi i(r-s))^{-\frac{d}{2}} ( 4 π i ( r − s ) ) − 2 d in K r − s K_{r-s} K r − s ,
∣ ∫ R d b r , s ( y ′ ) K ‾ r − s ( ε y ′ ) d y ′ ∣ ≤ C min ( ∣ r − s ∣ − d 2 , ε − d ) ∥ b ∥ W d + 2 , 1 . (73) \left|\int_{\mathbb{R}^d}b_{r,s}(y')\overline{K}_{r-s}(\varepsilon y')\,\mathrm{d}y'\right|\le C\min\left(|r-s|^{-\frac{d}{2}},\varepsilon^{-d}\right)\|b\|_{W^{d+2,1}}. \tag*{(73)} ∫ R d b r , s ( y ′ ) K r − s ( ε y ′ ) d y ′ ≤ C min ( ∣ r − s ∣ − 2 d , ε − d ) ∥ b ∥ W d + 2 , 1 . (73)
We caution that this pointwise bound, integrated over the time square, is not sufficient. For d ≥ 2 d\ge2 d ≥ 2 , the integral ∫ δ t τ − d / 2 d τ \int_\delta^t\tau^{-d/2}\,\mathrm{d}\tau ∫ δ t τ − d /2 d τ diverges as δ → 0 \delta\to0 δ → 0 (logarithmically in d = 2 d=2 d = 2 , by a power in d ≥ 3 d\ge3 d ≥ 3 ). The uniform estimate in Lemma A.1 works precisely because it retains the joint ( r , s ) (r,s) ( r , s ) oscillation in (A.8) rather than bounding each ( r , s ) (r,s) ( r , s ) -slice separately. The strip ∣ r − s ∣ ≤ δ |r-s|\le\delta ∣ r − s ∣ ≤ δ , where K τ → δ 0 K_\tau\to\delta_0 K τ → δ 0 as τ → 0 \tau\to0 τ → 0 , contributes at most
ε d ∫ 0 t ∫ ∣ r − s ∣ ≤ δ ∣ ∫ b r , s K ‾ r − s d y ′ ∣ d s d r ≤ C ε d t δ . (74) \varepsilon^d\int_0^t\int_{|r-s|\le\delta}\left|\int b_{r,s}\overline{K}_{r-s}\,\mathrm{d}y'\right|\,\mathrm{d}s\,\mathrm{d}r\le C\varepsilon^d t\delta. \tag*{(74)} ε d ∫ 0 t ∫ ∣ r − s ∣ ≤ δ ∫ b r , s K r − s d y ′ d s d r ≤ C ε d t δ . (74)
which is o ( ε d ) o(\varepsilon^d) o ( ε d ) for any fixed δ → 0 \delta\to0 δ → 0 . It is in this diagonal strip that the leading constant R ^ ( 0 ) \widehat{R}(0) R ( 0 ) resides. The off-diagonal contribution is controlled by Lemma A.1. Combining the two,
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≤ C d , t , u 0 ε d , (75) \mathbb{E}\left[\|w_\varepsilon^{(1)}(t)\|_{L_x^2}^2\right]\le C_{d,t,u_0}\varepsilon^d, \tag*{(75)} E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 ] ≤ C d , t , u 0 ε d , (75)
in every dimension d ≥ 1 d\ge1 d ≥ 1 , with the caveats of Remark A.2.
What extends to d ≥ 4 d \ge4 d ≥ 4 , and what does not.
The lower bound, the all moments estimate, the almost sure convergence, and the fluctuation CLT use only the near region ∣ ξ ′ ∣ ≤ ε |\xi'| \le\varepsilon ∣ ξ ′ ∣ ≤ ε and the cumulant expansion, neither of which depends on the dimension. The near region volume estimate of Lemma 3.2 is unchanged. The cluster expansion of Lemma 3.10 under exponential mixing, with the higher cumulants bounded as in [10 Sec. 4] , relies on the integrability of the mixing factor and not on the Fourier dyadic sum. Consequently the estimates
E [ ∥ w ε ( 1 ) ( t ) ∥ L x 2 2 q ] ≤ C q ε d q , ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C ε d 2 , \mathbb{E}\left[\left\|w_{\varepsilon}^{(1)}(t)\right\|_{L_x^2}^{2q}\right] \le C_q \varepsilon^{dq}, \qquad\left\|r_\varepsilon(t)\right\|_{L^2(\Omega;L_x^2)} \le C \varepsilon^{\frac d2}, E [ w ε ( 1 ) ( t ) L x 2 2 q ] ≤ C q ε d q , ∥ r ε ( t ) ∥ L 2 ( Ω ; L x 2 ) ≤ C ε 2 d ,
and the covariance limit (55) are compatible with every dimension, the second moment upper bound being supplied by (A.12) when d ≥ 4 d \ge4 d ≥ 4 and by the Fourier near/far split of § 5.1 when d ≤ 3 d \le3 d ≤ 3 .
We record honestly the limits of this appendix. Lemma A.1 is an upper bound operator estimate. It does not by itself identify the leading constant R ^ ( 0 ) \widehat{R}(0) R ( 0 ) , which is instead picked out by the near region computation in the main text. The lower bound, the all moment estimate, the almost sure convergence and the fluctuation CLT use only the near region ∣ ξ ′ ∣ ≤ ε |\xi'| \le\varepsilon ∣ ξ ′ ∣ ≤ ε and the cumulant expansion, neither of which depends on the dimension, so they require no d ≥ 4 d \ge4 d ≥ 4 argument. The distinction between the two mechanisms for the upper bound is that the Fourier argument uses the boundedness of R ^ \widehat{R} R and converges geometrically on dyadic shells only for d < 4 d < 4 d < 4 , whereas the physical space argument uses the integrability and smoothness of b r , s b_{r,s} b r , s and the non degenerate Fresnel phase, retaining the oscillation through M = ⌈ d 4 ⌉ + 1 M = \lceil\frac d4 \rceil+ 1 M = ⌈ 4 d ⌉ + 1 time integrations by parts. Together the two arguments cover every d ≥ 1 d \ge1 d ≥ 1 , and the main statements in the introduction are formulated accordingly.
College of Science, China Agricultural University, Beijing, China
Email address: meng-fanfei@cau.edu.cn