Introduction
Let O ⊂ R d \mathcal{O} \subset\mathbb{R}^{d} O ⊂ R d be bounded, connected, with C 2 C^{2} C 2 boundary. The deterministic problem
{ ∂ t u ( t , x ) = Δ ( ∣ u ( t , x ) ∣ m − 1 u ( t , x ) ) , t > 0 , x ∈ O , u ( t , x ) = 0 , x ∈ ∂ O , u ( 0 , x ) = u 0 , x ∈ O , (1) \begin{cases}
\partial_t u(t,x)=\Delta\left(|u(t,x)|^{m-1}u(t,x)\right), & t>0,\ x\in\mathcal{O},\\
u(t,x)=0, & x\in\partial\mathcal{O},\\
u(0,x)=u_0, & x\in\mathcal{O}, \tag*{(1)}
\end{cases} ⎩ ⎨ ⎧ ∂ t u ( t , x ) = Δ ( ∣ u ( t , x ) ∣ m − 1 u ( t , x ) ) , u ( t , x ) = 0 , u ( 0 , x ) = u 0 , t > 0 , x ∈ O , x ∈ ∂ O , x ∈ O , (1)
is degenerate at u = 0 u=0 u = 0 and has finite propagation when m > 1 m>1 m > 1 . Its diffusion coefficient vanishes at the vacuum, and this degeneracy produces a free boundary and finite speed of propagation: bounded nonnegative initial data supported strictly inside the domain leave an interior vacuum for sufficiently small positive times [1 ] . Carrillo et al. [8 ] used mass transportation techniques to give a simple proof of the finite speed of propagation of the solution to the one-dimensional porous medium equation (1) .
A natural question is whether stochastic perturbations destroy this localization phenomenon. More precisely, which noise structures can make the porous medium equation (PME) lose finite propagation? The answer depends on how the noise acts on the vacuum. We first recall the relevant results.
Barbu and Röckner [4 ] studied stochastic porous medium equations with finite-dimensional linear multiplicative noise
{ d u ( t , x ) − Δ ( ∣ u ( t , x ) ∣ m − 1 u ( t , x ) ) d t = ∑ k = 1 N μ k e k u ( t , x ) d β k ( t ) , t > 0 , x ∈ O , u ( t , x ) = 0 , x ∈ ∂ O , u ( 0 , x ) = u 0 ( x ) , x ∈ O , (2) \left\{
\begin{aligned}
\mathrm{d}u(t,x)-\Delta\left(|u(t,x)|^{m-1}u(t,x)\right)\,\mathrm{d}t&=\sum_{k=1}^{N}\mu_k e_k u(t,x)\,\mathrm{d}\beta_k(t), &&t>0,\ x\in\mathcal{O},\\
u(t,x)&=0, &&x\in\partial\mathcal{O},\\
u(0,x)&=u_0(x), &&x\in\mathcal{O},
\end{aligned}
\right. \tag*{(2)} ⎩ ⎨ ⎧ d u ( t , x ) − Δ ( ∣ u ( t , x ) ∣ m − 1 u ( t , x ) ) d t u ( t , x ) u ( 0 , x ) = k = 1 ∑ N μ k e k u ( t , x ) d β k ( t ) , = 0 , = u 0 ( x ) , t > 0 , x ∈ O , x ∈ ∂ O , x ∈ O , (2)
where β k \beta_k β k is a sequence of independent Brownian motions on a filtered probability space ( Ω , F , F , P ) (\Omega,\mathcal{F},\mathbb{F},\mathbb{P}) ( Ω , F , F , P ) , e k e_k e k is an orthonormal system in L 2 ( O ) L^2(\mathcal{O}) L 2 ( O ) , and { μ k } \{\mu_k\} { μ k } is a sequence of nonnegative numbers. Under their assumptions, including 1 < m ≤ 5 1<m\leq5 1 < m ≤ 5 and d ≤ 3 d\leq3 d ≤ 3 , they proved almost-sure localization on a sufficiently small random time interval. Their proof used a random exponential transformation that converts the stochastic partial differential equation (SPDE) into a degenerate deterministic partial differential equation (PDE) with random coefficients, followed by a local energy method, also see [2 , 3 ] . More precisely, let
v ( t , x ) = e μ ( t ) u ( t , x ) , μ ( t ) = − ∑ k = 1 N μ k e k β k ( t ) , v(t,x)=e^{\mu(t)}u(t,x),\qquad\mu(t)=-\sum_{k=1}^{N}\mu_k e_k\beta_k(t), v ( t , x ) = e μ ( t ) u ( t , x ) , μ ( t ) = − k = 1 ∑ N μ k e k β k ( t ) ,
then the stochastic partial differential equation (2) becomes a deterministic PDE with random coefficients.
Gess [9 ] proved finite speed of propagation for porous medium equations perturbed by linear multiplicative space-time rough signals,
{ d u ( t , x ) = Δ ( ∣ u ( t , x ) ∣ m sgn u ( t , x ) ) d t + ∑ k f k u ( t , x ) ∘ d z k ( t ) , t > 0 , x ∈ O , u ( t , x ) = 0 , x ∈ ∂ O , u ( 0 , x ) = u 0 ( x ) , x ∈ O , \left\{
\begin{aligned}
\mathrm{d}u(t,x)&=\Delta\left(|u(t,x)|^m\operatorname{sgn}u(t,x)\right)\,\mathrm{d}t+\sum_k f_k u(t,x)\circ\mathrm{d}z^k(t), &&t>0,\ x\in\mathcal{O},\\
u(t,x)&=0, &&x\in\partial\mathcal{O},\\
u(0,x)&=u_0(x), &&x\in\mathcal{O},
\end{aligned}
\right. ⎩ ⎨ ⎧ d u ( t , x ) u ( t , x ) u ( 0 , x ) = Δ ( ∣ u ( t , x ) ∣ m sgn u ( t , x ) ) d t + k ∑ f k u ( t , x ) ∘ d z k ( t ) , = 0 , = u 0 ( x ) , t > 0 , x ∈ O , x ∈ ∂ O , x ∈ O ,
where rough driving signals z k ∈ C ( [ 0 , T ] ; R ) z^k\in C([0,T];\mathbb{R}) z k ∈ C ([ 0 , T ] ; R ) and diffusion coefficients f k ∈ C ∞ ( O ) f_k\in C^\infty(\mathcal{O}) f k ∈ C ∞ ( O ) . Moreover, Gess obtained the explicit and optimal estimates for the speed of propagation. Hensel [11 ] established finite time extinction with probability one for weak solutions of the Cauchy-Dirichlet problem for the one-dimensional stochastic porous medium equation with Stratonovich transport noise and compactly supported smooth initial datum. Fischer and Grün developed a stochastic energy method for finite propagation and waiting-time phenomena under linear multiplicative noise in dimensions up to three [10 ] . They obtained waiting-time criteria and adapted the method to finite speed of propagation. This gives further evidence that stochasticity alone does not remove the free boundary when the noise remains local and vacuum preserving. These results establish preservation of localization for the local coefficients and hypotheses considered in those works.
The common structural feature is
u = 0 ⟹ f k u = 0 , or μ k e k u = 0. u=0\quad\Longrightarrow\quad f_k u=0,\ \text{or}\ \mu_k e_k u=0. u = 0 ⟹ f k u = 0 , or μ k e k u = 0.
Thus the stochastic coefficient itself vanishes in a vacuum region. The random forcing may move the interface or alter its speed, but it does not directly inject a disturbance into a region where the state vanishes. Thus the noise is vacuum preserving. These results are therefore not in conflict with the proposed infinite-propagation mechanism.
On the other hand, Stan, del Teso and Vázquez [17 , 18 ] studied the porous medium equations with fractional pressure or nonlocal pressure
{ ∂ t u ( t , x ) = ∇ ⋅ ( u m − 1 ( t , x ) ∇ ( − Δ ) − s u ( t , x ) ) , t > 0 , u ( t , x ) ∣ t = 0 = u 0 ( x ) , x ∈ R d . \left\{
\begin{aligned}
\partial_tu(t,x)&=\nabla\cdot\left(u^{m-1}(t,x)\nabla(-\Delta)^{-s}u(t,x)\right), &&t>0,\\
u(t,x)\big|_{t=0}&=u_0(x), &&x\in\mathbb{R}^d.
\end{aligned}
\right. { ∂ t u ( t , x ) u ( t , x ) t = 0 = ∇ ⋅ ( u m − 1 ( t , x ) ∇ ( − Δ ) − s u ( t , x ) ) , = u 0 ( x ) , t > 0 , x ∈ R d .
where m > 1 m > 1 m > 1 , 0 < s < 1 0 < s < 1 0 < s < 1 . The main results of [17 , 18 ] show that spatial nonlocality in the deterministic flux can change the finite speed propagation regime. For 1 < m < 2 1 < m < 2 1 < m < 2 , the infinite-propagation result in [18 Theorem 1.4 (b)] is one-dimensional and is stated first as a positive-measure tail property; pointwise positivity requires additional symmetry and monotonicity assumptions. The case m = 1 m = 1 m = 1 reduces to a fractional heat equation. Further existence results are given in [19 ] . Our equation keeps the classical local porous medium drift and moves the nonlocal mechanism into the stochastic state-to-noise operator.
We investigate two noise structures that can activate an interior vacuum. One is additive noise ; the other is nonlocal multiplicative noise . Additive noise generally does not preserve nonnegativity. We therefore use distributional spatial support; when the state is a function, this agrees with the support of its absolute value. Brownian nonlocal multiplicative noise need not preserve nonnegativity either. For nonlocal multiplicative noise, we are motivated by [17 , 18 ] . The main difference between local and nonlocal multiplicative noise can be interpreted as: for an open bounded domain O \mathcal O O ,
B k ( u ) ( t , x ) : = g k ( x ) u ( t , x ) ⟹ u ∣ O = 0 ⟹ B k ( u ) ∣ O = 0 B_k(u)(t,x) := g_k(x)u(t,x) \quad\Longrightarrow\quad u|_O = 0 \Longrightarrow B_k(u)|_O = 0 B k ( u ) ( t , x ) := g k ( x ) u ( t , x ) ⟹ u ∣ O = 0 ⟹ B k ( u ) ∣ O = 0
for local multiplicative noise, where O ⋐ O O \Subset\mathcal O O ⋐ O is a nonempty open set, whereas
B ( u ) ( t , x ) : = ∫ O K ( x , y ) u ( t , y ) d y B(u)(t,x) := \int_{\mathcal O} K(x,y)u(t,y)\,\mathrm{d}y B ( u ) ( t , x ) := ∫ O K ( x , y ) u ( t , y ) d y
may satisfy
u ∣ O = 0 , u ≢ 0 ⟹ B ( u ) ∣ O ≠ 0. u|_O = 0,\quad u \not\equiv0 \quad\Longrightarrow\quad B(u)|_O \ne0. u ∣ O = 0 , u ≡ 0 ⟹ B ( u ) ∣ O = 0.
For a strictly positive convolution kernel and a nonnegative nontrivial state,
supp u ⋐ O ⟹ supp O B ( u ) = O ‾ . \operatorname{supp} u \Subset\mathcal O \quad\Longrightarrow\quad\operatorname{supp}_{\mathcal O} B(u) = \overline{\mathcal O}. supp u ⋐ O ⟹ supp O B ( u ) = O .
This is the mechanism absent from the known local multiplicative finite speed propagation theory.
The above discussions motivate the following questions:
Question 1: Can additive noise destroy the compact support in the following sense, for every fixed t > 0 t > 0 t > 0 ,
P ( supp u ( t , ⋅ ) = O ‾ ) = 1 ? \mathbb{P}\bigl(\operatorname{supp}u(t,\cdot) = \overline{\mathcal O}\bigr) = 1? P ( supp u ( t , ⋅ ) = O ) = 1 ?
Question 2: Can one construct a multiplicative stochastic perturbation for which the deterministic porous medium equation has finite speed of propagation, whereas the stochastic equation has instantaneous full spatial support?
The central idea of this paper is that the relevant distinction is not merely between deterministic and stochastic equations, nor between finite- and infinite-dimensional Wiener processes. Rather, one should distinguish between vacuum-preserving and vacuum-destroying noise coefficients. Theorem 2.15 makes this mechanism rigorous for additive noise. For multiplicative noise, we propose the genuinely nonlocal multiplicative model
d u ( t , x ) = Δ Φ ( u ( t , x ) ) d t + ( K ∗ u ) ( t , x ) d β t , Φ ( r ) = ∣ r ∣ m − 1 r , m > 1 , (3) \mathrm{d}u(t,x) = \Delta\Phi(u(t,x))\,\mathrm{d}t + (K * u)(t,x)\,\mathrm{d}\beta_t,\qquad\Phi(r) = |r|^{m-1}r,\qquad m > 1, \tag*{(3)} d u ( t , x ) = ΔΦ ( u ( t , x )) d t + ( K ∗ u ) ( t , x ) d β t , Φ ( r ) = ∣ r ∣ m − 1 r , m > 1 , (3)
where
( K ∗ u ) ( t , x ) = ∫ O K ( x − y ) u ( t , y ) d y . (K * u)(t,x) = \int_{\mathcal O} K(x-y)u(t,y)\,\mathrm{d}y. ( K ∗ u ) ( t , x ) = ∫ O K ( x − y ) u ( t , y ) d y .
If K > 0 K > 0 K > 0 almost everywhere, and u ≥ 0 u \ge0 u ≥ 0 and u ≢ 0 u \not\equiv0 u ≡ 0 , then
( K ∗ u ) ( t , x ) > 0 for a.e. x ∈ O , (K * u)(t,x) > 0 \quad\text{for a.e. } x \in\mathcal O, ( K ∗ u ) ( t , x ) > 0 for a.e. x ∈ O ,
even when u ( t , x ) = 0 u(t,x)=0 u ( t , x ) = 0 on a large open set in spatial variable. Hence the stochastic coefficient directly sees every vacuum region. This is stronger than merely choosing a nonlocal covariance operator. For example, u Q 1 / 2 d W uQ^{1/2}\,\mathrm{d}W u Q 1/2 d W may have a spatially nonlocal covariance, but the pointwise factor u ( x ) u(x) u ( x ) still forces the noise to vanish at every vacuum point. In (3) , by contrast, the state-to-noise map itself is nonlocal.
The propagation mechanisms and the conclusions proved here are summarized in Table 1.1.
Table 1. Local and nonlocal propagation mechanisms.
The main contributions, in this paper, are
The full-domain support is obtained for additive noise. The additive-noise Malliavin method is used and comparing with [6 ] , our differentiation step instead applies a finite-dimensional monotone-SDE theorem before taking a weak Galerkin limit.
Dirichlet full-support or zero-state dichotomy is given for Dirichlet boundary problem. In order to get the Dirichlet full-support, the backward non-hitting property should be added.
Different from Dirichlet boundary problem, the full support for the Neumann mass model with nonzero initial mass is obtained.
The rest of the paper is organized as follows. Section 2 proves the additive result and develops the common approximation framework. Section 3 establishes the mixed-norm terminal estimate and the Dirichlet dichotomy. Section 4 proves the Neumann result by conditioning on the mass noise. A simple numerical example is given is Section 5. Conclusion is given in Section 6.
Additive noise and the approximation framework
Throughout the paper, m > 1 m>1 m > 1 , p = m + 1 p=m+1 p = m + 1 , and O ⊂ R d \mathcal{O}\subset\mathbb{R}^{d} O ⊂ R d is bounded and connected with C 2 C^{2} C 2 boundary. All initial conditions in the Malliavin statements are deterministic. We write ⟨ v , φ ⟩ \langle v,\varphi\rangle ⟨ v , φ ⟩ for distributional duality and ∥ v ∥ q \lVert v\rVert_{q} ∥ v ∥ q for the L q ( O ) L^{q}(\mathcal{O}) L q ( O ) norm. We first introduce the definition of spatial support.
Definition 2.1. For v ∈ D ′ ( O ) v \in\mathcal{D}'(\mathcal{O}) v ∈ D ′ ( O ) , let
S O ( v ) = O ∖ ⋃ { O ′ ⊂ O : O ′ is open and v ∣ O ′ = 0 } . S_{\mathcal{O}}(v)=\mathcal{O}\setminus\bigcup\{\mathcal{O}'\subset\mathcal{O}:\mathcal{O}'\text{ is open and }v|_{\mathcal{O}'}=0\}. S O ( v ) = O ∖ ⋃ { O ′ ⊂ O : O ′ is open and v ∣ O ′ = 0 } . Define
supp O v = S O ( v ) ‾ R d . \operatorname{supp}_{\mathcal{O}}v=\overline{S_{\mathcal{O}}(v)}^{\mathbb{R}^{d}}. supp O v = S O ( v ) R d .
In particular, supp O 0 = ∅ \operatorname{supp}_{\mathcal{O}}0=\varnothing supp O 0 = ∅ , where 0 0 0 means the zero function. Full-domain support means supp O v = O ‾ \operatorname{supp}_{\mathcal{O}}v=\overline{\mathcal{O}} supp O v = O .
Every support assertion below uses Definition 2.1 . If v v v is locally integrable, it vanishes as a distribution on an open set exactly when it vanishes almost everywhere there. Thus this convention agrees with the closed support of ∣ v ∣ |v| ∣ v ∣ , and allows signed states without imposing a positivity property on the stochastic equation.
In this paper, we want to show the full-support of stochastic porous medium equation. In order that, we need the instantaneous full-domain propagation for any prescribed time.
Definition 2.2. A stochastic solution has instantaneous full-domain propagation at fixed positive times if, for every fixed t > 0 t>0 t > 0 ,
P ( supp O u ( t , ⋅ ) = O ‾ ) = 1. \mathbb{P}\left(\operatorname{supp}_{\mathcal{O}}u(t,\cdot)=\overline{\mathcal{O}}\right)=1. P ( supp O u ( t , ⋅ ) = O ) = 1.
Definition 2.2 concerns each prescribed time, so it does not assert a single probability-one event valid for all positive real times. It is also compatible with the homogeneous Dirichlet boundary condition: a function can have zero boundary trace while being nonzero on every interior open set. No pointwise strict positivity is included in the definition. We recall the deterministic benchmark before introducing noise. Let B R ( x 0 ) B_R(x_0) B R ( x 0 ) be the open Euclidean ball with radius R R R centred x 0 x_0 x 0 .
Theorem 2.3 (Persistence of a deterministic interior vacuum). Let 0 ≤ v 0 ∈ L ∞ ( O ) 0\leq v_0\in L^\infty(\mathcal{O}) 0 ≤ v 0 ∈ L ∞ ( O ) , and let v v v be the nonnegative weak solution of (1) . Suppose v 0 = 0 v_0=0 v 0 = 0 on B R ( x 0 ) ⋐ O B_R(x_0)\Subset\mathcal{O} B R ( x 0 ) ⋐ O . For every 0 < R 1 < R 0<R_1<R 0 < R 1 < R , there is T ∗ > 0 T_*>0 T ∗ > 0 such that v ( t ) = 0 v(t)=0 v ( t ) = 0 on B R 1 ( x 0 ) B_{R_1}(x_0) B R 1 ( x 0 ) for 0 < t < T ∗ 0<t<T_* 0 < t < T ∗ . In particular, bounded nonnegative initial data supported strictly inside O \mathcal{O} O do not produce instantaneous full-domain propagation.
Proof. The result is known and we only give the outline of proof. The classical finite-propagation, or hole-filling, estimate for the porous medium equation bounds the advance of a free boundary into an initial vacuum by a function η ( t ) \eta(t) η ( t ) tending to zero as t ↓ 0 t\downarrow0 t ↓ 0 . It follows either from the local energy method or from comparison with porous-medium barriers; see [1 , 20 ] . Applied to the initially empty ball, it gives v ( t ) = 0 v(t)=0 v ( t ) = 0 on B R − η ( t ) ( x 0 ) B_{R-\eta(t)}(x_0) B R − η ( t ) ( x 0 ) as long as this radius is positive. The bound depends on the initial L ∞ L^\infty L ∞ norm and the local geometry, and the zero Dirichlet condition supplies no incoming boundary source.
Choose T ∗ T_* T ∗ so that η ( t ) < R − R 1 \eta(t)<R-R_1 η ( t ) < R − R 1 for 0 < t < T ∗ 0<t<T_* 0 < t < T ∗ . Then the smaller ball remains empty throughout that interval. If supp O v 0 ⋐ O \operatorname{supp}_{\mathcal{O}}v_0\Subset\mathcal{O} supp O v 0 ⋐ O , its complement contains such an interior ball, so the first assertion applies. This argument concerns small positive times; it does not say that the support remains strictly inside the domain for every finite time.
The contrast with Theorem 2.3 will be that sufficiently nondegenerate additive noise removes every interior vacuum at any fixed positive time, including when the initial datum is compactly supported.
Additive noise and well-posedness
Let
H D = H − 1 ( O ) = ( H 0 1 ( O ) ) ∗ with ∥ v ∥ H D 2 = ⟨ v , ( − Δ D ) − 1 v ⟩ . H_D = H^{-1}(\mathcal{O}) = (H^1_0(\mathcal{O}))^* \quad\text{with}\quad\|v\|_{H_D}^2 = \langle v,(-\Delta_D)^{-1}v\rangle. H D = H − 1 ( O ) = ( H 0 1 ( O ) ) ∗ with ∥ v ∥ H D 2 = ⟨ v , ( − Δ D ) − 1 v ⟩ .
The variational triple is V = L p ( O ) ⊂ H D ⊂ V ∗ V = L^p(\mathcal{O}) \subset H_D \subset V^* V = L p ( O ) ⊂ H D ⊂ V ∗ . Define the positive drift A : V → V ∗ A: V \to V^* A : V → V ∗ by
V ∗ ⟨ A ( v ) , w ⟩ V = ∫ O Φ ( v ( x ) ) w ( x ) d x , Φ ( r ) = ∣ r ∣ m − 1 r . (4) {}_{V^*}\langle A(v),w\rangle_V = \int_{\mathcal{O}} \Phi(v(x))w(x)\,\mathrm{d}x,\qquad\Phi(r)=|r|^{m-1}r. \tag*{(4)} V ∗ ⟨ A ( v ) , w ⟩ V = ∫ O Φ ( v ( x )) w ( x ) d x , Φ ( r ) = ∣ r ∣ m − 1 r . (4)
Thus − A -A − A is the variational realization of Δ Φ \Delta\Phi ΔΦ with homogeneous Dirichlet boundary conditions. An H D H_D H D -continuous solution need not be an L p L^p L p function at every fixed time; this is one reason to use distributional support.
We establish the estimates for affine noise, so that they can be reused. Let W W W be cylindrical on a separable Hilbert space U U U , let g ∈ C ( [ 0 , T ] ; L 2 ( U , H D ) ) g\in C([0,T];L_2(U,H_D)) g ∈ C ([ 0 , T ] ; L 2 ( U , H D )) be deterministic, and let C ∈ L ( H D , L 2 ( U , H D ) ) C\in\mathcal{L}(H_D,L_2(U,H_D)) C ∈ L ( H D , L 2 ( U , H D )) . Write C k v = ( C v ) e k C_kv=(Cv)e_k C k v = ( C v ) e k for an orthonormal basis ( e k ) (e_k) ( e k ) of U U U , and assume
∑ k ≥ 1 ∥ C k v ∥ H D 2 ≤ L ∥ v ∥ H D 2 , v ∈ H D . (5) \sum_{k\geq1}\|C_kv\|_{H_D}^2\leq L\|v\|_{H_D}^2,\qquad v\in H_D. \tag*{(5)} k ≥ 1 ∑ ∥ C k v ∥ H D 2 ≤ L ∥ v ∥ H D 2 , v ∈ H D . (5)
Set Σ ( s , v ) = g ( s ) + C v \Sigma(s,v)=g(s)+Cv Σ ( s , v ) = g ( s ) + C v and consider
d X ( s ) + A ( X ) d s = Σ ( s , X ) d W s , X ( 0 ) = x ∈ H D . (6) \mathrm{d}X(s)+A(X)\,\mathrm{d}s=\Sigma(s,X)\,\mathrm{d}W_s,\qquad X(0)=x\in H_D. \tag*{(6)} d X ( s ) + A ( X ) d s = Σ ( s , X ) d W s , X ( 0 ) = x ∈ H D . (6)
Here we let u u u replaced by X X X in order to consider the abstract SDEs. The diffusion term Σ ( s , X ) \Sigma(s,X) Σ ( s , X ) contains the additive noise and linear multiplicative noise. For a smooth test φ \varphi φ , the notation Σ ( s , v ) ∗ φ ∈ U \Sigma(s,v)^*\varphi\in U Σ ( s , v ) ∗ φ ∈ U means the vector with coordinates ⟨ Σ ( s , v ) e k , φ ⟩ \langle\Sigma(s,v)e_k,\varphi\rangle ⟨ Σ ( s , v ) e k , φ ⟩ . It is defined by duality with H 0 1 H^1_0 H 0 1 , rather than by identifying that test with an element of H D H_D H D . If the noise takes values in L 2 L^2 L 2 , this agrees with the usual L 2 L^2 L 2 adjoint notation.
The additive equation is the special case g ( s ) = G g(s)=G g ( s ) = G and C = 0 C=0 C = 0 , where G ∈ L 2 ( U , H D ) G\in L_2(U,H_D) G ∈ L 2 ( U , H D ) is deterministic.
Assumption 2.4 (Additive local nondegeneracy). For every nonzero φ ∈ C c ∞ ( O ) \varphi\in C^\infty_c(\mathcal{O}) φ ∈ C c ∞ ( O ) , ∥ G ∗ φ ∥ U > 0 \|G^*\varphi\|_U>0 ∥ G ∗ φ ∥ U > 0 .
Assumption 2.4 asks that every localized scalar observation see at least one noise direction. It does not require a uniform lower bound over all tests. The following familiar trace-class example shows that the assumption is compatible with spatially correlated noise.
Proposition 2.5 (A spectral sufficient condition). Let ( e k ) (e_k) ( e k ) be a complete orthonormal basis of L 2 ( O ) L^2(\mathcal{O}) L 2 ( O ) and let Q e k = q k e k Qe_k=q_ke_k Q e k = q k e k , where q k > 0 q_k>0 q k > 0 and ∑ k q k < ∞ \sum_kq_k<\infty ∑ k q k < ∞ . Then
G = Q 1 / 2 : L 2 ( O ) → L 2 ( O ) , G=Q^{1/2}:L^2(\mathcal{O})\to L^2(\mathcal{O}), G = Q 1/2 : L 2 ( O ) → L 2 ( O ) , also viewed as an H D H_D H D -valued operator, satisfies Assumption 2.4.
Proof. The Hilbert–Schmidt norm into L 2 L^2 L 2 is
∥ G ∥ L 2 2 = ∑ k q k < ∞ . \|G\|_{L_2}^2=\sum_kq_k<\infty. ∥ G ∥ L 2 2 = k ∑ q k < ∞. The continuous embedding L 2 ⊂ H D L^2\subset H_D L 2 ⊂ H D also makes G G G Hilbert–Schmidt into H D H_D H D . For a nonzero test φ \varphi φ , completeness supplies a k 0 k_0 k 0 with ( φ , e k 0 ) L 2 ≠ 0 (\varphi,e_{k_0})_{L^2}\ne0 ( φ , e k 0 ) L 2 = 0 . Consequently
∥ G ∗ φ ∥ U 2 = ∑ k q k ∣ ( φ , e k ) L 2 ∣ 2 ≥ q k 0 ∣ ( φ , e k 0 ) L 2 ∣ 2 > 0. \|G^*\varphi\|_U^2=\sum_kq_k|(\varphi,e_k)_{L^2}|^2\geq q_{k_0}|(\varphi,e_{k_0})_{L^2}|^2>0. ∥ G ∗ φ ∥ U 2 = k ∑ q k ∣ ( φ , e k ) L 2 ∣ 2 ≥ q k 0 ∣ ( φ , e k 0 ) L 2 ∣ 2 > 0. Here the adjoint is used in the duality sense specified above, which gives exactly the displayed coordinates.
For F t , φ = ⟨ u ( t ) , φ ⟩ F_{t,\varphi}=\langle u(t),\varphi\rangle F t , φ = ⟨ u ( t ) , φ ⟩ , the scalar weak equation is
F t , φ = ⟨ u 0 , φ ⟩ + ∫ 0 t ∫ O Φ ( u ( s ) ) Δ φ d x d s + ∫ 0 t ( G ∗ φ , d W s ) U . (7) F_{t,\varphi}=\langle u_0,\varphi\rangle+\int_0^t\int_{\mathcal{O}}\Phi(u(s))\Delta\varphi\,\mathrm{d}x\,\mathrm{d}s+\int_0^t(G^*\varphi,\mathrm{d}W_s)_U. \tag*{(7)} F t , φ = ⟨ u 0 , φ ⟩ + ∫ 0 t ∫ O Φ ( u ( s )) Δ φ d x d s + ∫ 0 t ( G ∗ φ , d W s ) U . (7)
Applying Itô’s isometry to the deterministic integrand G ∗ φ G^*\varphi G ∗ φ , we get
E [ ∣ ∫ 0 t ( G ∗ φ , d W s ) U ∣ 2 ] = t ∥ G ∗ φ ∥ U 2 > 0. \mathbb{E}\left[\left\lvert\int_0^t(G^*\varphi,\mathrm{d}W_s)_U\right\rvert^2\right]=t\lVert G^*\varphi\rVert_U^2>0. E [ ∫ 0 t ( G ∗ φ , d W s ) U 2 ] = t ∥ G ∗ φ ∥ U 2 > 0.
The above equality explains why the noise can act inside a vacuum: its scalar coefficient is independent of the state. It does not yet prove P ( F t , φ = 0 ) = 0 \mathbb{P}(F_{t,\varphi}=0)=0 P ( F t , φ = 0 ) = 0 . The drift in (7) depends on the same Wiener path, so the observable is not an independent random shift of the displayed Gaussian term. The approximation and Malliavin arguments below establish the needed nondegeneracy of the complete observable.
To establish well-posedness, we use a Galerkin approximation based on the eigenfunctions of the Dirichlet Laplacian. Let ( ρ j , λ j ) j ≥ 1 (\rho_j,\lambda_j)_{j\geq1} ( ρ j , λ j ) j ≥ 1 satisfy
− Δ ρ j = λ j ρ j in O , ρ j = 0 on ∂ O , -\Delta\rho_j=\lambda_j\rho_j\quad\text{in }\mathcal{O},\qquad\rho_j=0\quad\text{on }\partial\mathcal{O}, − Δ ρ j = λ j ρ j in O , ρ j = 0 on ∂ O ,
where ( ρ j ) j ≥ 1 (\rho_j)_{j\geq1} ( ρ j ) j ≥ 1 is an orthonormal basis of L 2 ( O ) L^2(\mathcal{O}) L 2 ( O ) and the eigenvalues are listed in nondecreasing order. Let H n = span { ρ 1 , … , ρ n } H_n=\operatorname{span}\{\rho_1,\ldots,\rho_n\} H n = span { ρ 1 , … , ρ n } and let P n P_n P n be the spectral projection. It is also orthogonal in H D H_D^{} H D and contractive on H 0 1 H_0^1 H 0 1 . Let Π n \Pi_n Π n project U U U onto its first n n n coordinates. In this paper, we do not approximate the noise [14 ] and only discretized the space.
Lemma 2.6 (Projected test functions). For φ ∈ C c ∞ ( O ) \varphi\in C^\infty_c(\mathcal{O}) φ ∈ C c ∞ ( O ) , set φ n = P n φ \varphi_n=P_n\varphi φ n = P n φ . Then φ n → φ \varphi_n\to\varphi φ n → φ in H 0 1 H_0^1 H 0 1 and
sup n ∥ Δ φ n ∥ ∞ < ∞ , Δ φ n ⟶ Δ φ uniformly on O . (8) \sup_n\lVert\Delta\varphi_n\rVert_\infty<\infty,\qquad\Delta\varphi_n\longrightarrow\Delta\varphi\quad\text{uniformly on }\mathcal{O}. \tag*{(8)} n sup ∥ Δ φ n ∥ ∞ < ∞ , Δ φ n ⟶ Δ φ uniformly on O . (8) The corresponding statement holds for the positive Neumann modes, with P n φ → φ − φ ‾ P_n\varphi\to\varphi-\overline{\varphi} P n φ → φ − φ in H 1 H^1 H 1 .
Proof. A compactly supported smooth test belongs to the domain of every integer power of the Dirichlet Laplacian. For the Neumann Laplacian this is true of φ − φ ‾ \varphi-\overline{\varphi} φ − φ . The standard heat-kernel and eigenvalue estimates on a bounded C 2 C^2 C 2 domain give
∥ ρ j ∥ ∞ ≤ C ( 1 + λ j ) d / 4 \lVert\rho_j\rVert_\infty\leq C(1+\lambda_j)^{d/4} ∥ ρ j ∥ ∞ ≤ C ( 1 + λ j ) d /4
and λ j ≥ c j 2 / d \lambda_j\geq cj^{2/d} λ j ≥ c j 2/ d for the positive modes. If a j = ( φ , ρ j ) L 2 a_j=(\varphi,\rho_j)_{L^2} a j = ( φ , ρ j ) L 2 , then
∑ j ( 1 + λ j ) 2 q ∣ a j ∣ 2 < ∞ \sum_j(1+\lambda_j)^{2q}|a_j|^2<\infty j ∑ ( 1 + λ j ) 2 q ∣ a j ∣ 2 < ∞
for every integer q q q . For q > 1 + d / 2 q>1+d/2 q > 1 + d /2 , Cauchy–Schwarz gives
∑ j λ j ∣ a j ∣ ∥ ρ j ∥ ∞ < ∞ . \sum_j\lambda_j|a_j|\lVert\rho_j\rVert_\infty<\infty. j ∑ λ j ∣ a j ∣ ∥ ρ j ∥ ∞ < ∞.
More explicitly, the second factor in that inequality is bounded by
C ( ∑ j ( 1 + λ j ) 2 + d / 2 − 2 q ) 1 / 2 < ∞ , C\left(\sum_j(1+\lambda_j)^{2+d/2-2q}\right)^{1/2}<\infty, C ( j ∑ ( 1 + λ j ) 2 + d /2 − 2 q ) 1/2 < ∞ ,
while the first is ( ∑ j ( 1 + λ j ) 2 q ∣ a j ∣ 2 ) 1 / 2 (\sum_j(1+\lambda_j)^{2q}|a_j|^2)^{1/2} ( ∑ j ( 1 + λ j ) 2 q ∣ a j ∣ 2 ) 1/2 . The series for Δ φ n \Delta\varphi_n Δ φ n therefore converges absolutely and uniformly, with limit Δ φ \Delta\varphi Δ φ . Spectral convergence in the energy norm proves the other assertions. ▫
Although φ \varphi φ is compactly supported, φ n \varphi_n φ n generally is not. We will therefore estimate the linearized drift over the whole domain O \mathcal{O} O . The uniform Laplacian bound in Lemma 2.6 is what makes that estimate independent of the Galerkin dimension.
Define A n : H n → H n A_n : H_n \to H_n A n : H n → H n by
( A n ( v ) , w ) H D = ∫ O Φ ( v ) w d x for w ∈ H n . (A_n(v),w)_{H_D} = \int_{\mathcal{O}} \Phi(v)w\,\mathrm{d}x \quad\text{for } w \in H_n. ( A n ( v ) , w ) H D = ∫ O Φ ( v ) w d x for w ∈ H n .
We use the actual porous-medium nonlinearity in the finite-dimensional SDE
d X n ( s ) + A n ( X n ) d s = P n Σ ( s , X n ) Π n d W s , X n ( 0 ) = P n x . (9) \mathrm{d}X_n(s) + A_n(X_n)\,\mathrm{d}s = P_n\Sigma(s,X_n)\Pi_n\,\mathrm{d}W_s,\qquad X_n(0)=P_nx. \tag*{(9)} d X n ( s ) + A n ( X n ) d s = P n Σ ( s , X n ) Π n d W s , X n ( 0 ) = P n x . (9)
No globally bounded derivative is imposed on Φ \Phi Φ . Since m > 1 m>1 m > 1 , Φ \Phi Φ is C 1 C^1 C 1 , which suffices for the finite-dimensional differentiability result used below. In particular, no incompatible combination of bounded derivative and superquadratic coercivity enters the approximation.
Proposition 2.7 (Well-posedness). Under (5) , equation (6) has a unique solution X X X in L 2 ( Ω ; C ( [ 0 , T ] ; H D ) ) ∩ L p ( Ω × ( 0 , T ) × O ) L^2(\Omega;C([0,T];H_D))\cap L^p(\Omega\times(0,T)\times\mathcal{O}) L 2 ( Ω ; C ([ 0 , T ] ; H D )) ∩ L p ( Ω × ( 0 , T ) × O ) . The solutions of (9) satisfy
sup n { E sup s ≤ T ∥ X n ( s ) ∥ H D 2 + E ∫ 0 T ∥ X n ( s ) ∥ p p d s } < ∞ . (10) \sup_n\left\{\mathbb{E}\sup_{s\leq T}\lVert X_n(s)\rVert_{H_D}^2+\mathbb{E}\int_0^T\lVert X_n(s)\rVert_p^p\,\mathrm{d}s\right\}<\infty. \tag*{(10)} n sup { E s ≤ T sup ∥ X n ( s ) ∥ H D 2 + E ∫ 0 T ∥ X n ( s ) ∥ p p d s } < ∞. (10) They converge weakly to X X X in L p ( Ω × ( 0 , T ) × O ) ∩ L 2 ( Ω × ( 0 , T ) ; H D ) L^p(\Omega\times(0,T)\times\mathcal{O})\cap L^2(\Omega\times(0,T);H_D) L p ( Ω × ( 0 , T ) × O ) ∩ L 2 ( Ω × ( 0 , T ) ; H D ) . For every fixed t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] and φ ∈ C ∞ ( O ) \varphi\in C^\infty(\mathcal{O}) φ ∈ C ∞ ( O ) ,
⟨ X n ( t ) , φ ⟩ ⇀ ⟨ X ( t ) , φ ⟩ in L 2 ( Ω ) . (11) \langle X_n(t),\varphi\rangle\rightharpoonup\langle X(t),\varphi\rangle\quad\text{in }L^2(\Omega). \tag*{(11)} ⟨ X n ( t ) , φ ⟩ ⇀ ⟨ X ( t ) , φ ⟩ in L 2 ( Ω ) . (11)
Proof. The proof is divided into three steps.
Step 1: the variational hypotheses and uniform energy. For (4) , semi-continuity follows from continuity of Φ \Phi Φ , while
⟨ A ( v ) − A ( w ) , v − w ⟩ = ∫ O [ Φ ( v ) − Φ ( w ) ] ( v − w ) d x ≥ 0 , ∥ A ( v ) ∥ V ∗ ≤ ∥ v ∥ p p − 1 . \langle A(v)-A(w),v-w\rangle=\int_{\mathcal{O}}[\Phi(v)-\Phi(w)](v-w)\,\mathrm{d}x\geq0,\qquad\lVert A(v)\rVert_{V^*}\leq\lVert v\rVert_p^{p-1}. ⟨ A ( v ) − A ( w ) , v − w ⟩ = ∫ O [ Φ ( v ) − Φ ( w )] ( v − w ) d x ≥ 0 , ∥ A ( v ) ∥ V ∗ ≤ ∥ v ∥ p p − 1 .
Moreover ⟨ A ( v ) , v ⟩ = ∥ v ∥ p p \langle A(v),v\rangle=\lVert v\rVert_p^p ⟨ A ( v ) , v ⟩ = ∥ v ∥ p p and
∥ Σ ( s , v ) ∥ L 2 2 ≤ 2 ∥ g ( s ) ∥ L 2 2 + 2 L ∥ v ∥ H D 2 . \lVert\Sigma(s,v)\rVert_{L_2}^2\leq2\lVert g(s)\rVert_{L_2}^2+2L\lVert v\rVert_{H_D}^2. ∥ Σ ( s , v ) ∥ L 2 2 ≤ 2 ∥ g ( s ) ∥ L 2 2 + 2 L ∥ v ∥ H D 2 .
These are the semi-continuity, monotonicity, coercivity and growth hypotheses of the variational theorem [13 ] . In particular, we use its polynomial variational setting, without attributing arbitrary H D H_D H D -valued noise to a theorem that imposes spatial smoothing on the noise.
The Galerkin drift is C 1 C^1 C 1 and dissipative, and the noise is affine and globally Lipschitz. Itô’s formula gives exactly
∥ X n ( t ) ∥ H D 2 + 2 ∫ 0 t ∥ X n ( s ) ∥ p p d s = ∥ P n x ∥ H D 2 + ∫ 0 t ∥ P n Σ ( s , X n ) Π n ∥ L 2 2 d s + 2 ∫ 0 t ( X n , P n Σ ( s , X n ) Π n d W s ) H D . \lVert X_n(t)\rVert_{H_D}^2+2\int_0^t\lVert X_n(s)\rVert_p^p\,\mathrm{d}s=\lVert P_nx\rVert_{H_D}^2+\int_0^t\lVert P_n\Sigma(s,X_n)\Pi_n\rVert_{L_2}^2\,\mathrm{d}s+2\int_0^t(X_n,P_n\Sigma(s,X_n)\Pi_n\,\mathrm{d}W_s)_{H_D}. ∥ X n ( t ) ∥ H D 2 + 2 ∫ 0 t ∥ X n ( s ) ∥ p p d s = ∥ P n x ∥ H D 2 + ∫ 0 t ∥ P n Σ ( s , X n ) Π n ∥ L 2 2 d s + 2 ∫ 0 t ( X n , P n Σ ( s , X n ) Π n d W s ) H D .
Note that the uniformity in n n n is used repeatedly below. Set
Σ n ( s ) = P n Σ ( s , X n ( s ) ) Π n , M n ( t ) = 2 ∫ 0 t ( X n ( s ) , Σ n ( s ) d W s ) H D . \Sigma_n(s)=P_n\Sigma(s,X_n(s))\Pi_n,\qquad M_n(t)=2\int_0^t(X_n(s),\Sigma_n(s)\,\mathrm{d}W_s)_{H_D}. Σ n ( s ) = P n Σ ( s , X n ( s )) Π n , M n ( t ) = 2 ∫ 0 t ( X n ( s ) , Σ n ( s ) d W s ) H D .
For R > 0 R>0 R > 0 , let
τ R = inf { s ≥ 0 : ∥ X n ( s ) ∥ H D ≥ R } ∧ T . \tau_R=\inf\{s\geq0:\lVert X_n(s)\rVert_{H_D}\geq R\}\wedge T. τ R = inf { s ≥ 0 : ∥ X n ( s ) ∥ H D ≥ R } ∧ T .
Taking the supremum in the preceding Itô identity up to t ∧ τ R t \wedge\tau_R t ∧ τ R and then expectation gives
E sup θ ≤ t ∧ τ R ∥ X n ( θ ) ∥ H D 2 + 2 E ∫ 0 t ∧ τ R ∥ X n ( s ) ∥ p p d s ≤ ∥ P n x ∥ H D 2 + E ∫ 0 t ∧ τ R ∥ Σ n ( s ) ∥ L 2 2 d s + E sup θ ≤ t ∧ τ R ∣ M n ( θ ) ∣ . \mathbb{E}\sup_{\theta\leq t\wedge\tau_R}\|X_n(\theta)\|_{H_D}^2+2\mathbb{E}\int_0^{t\wedge\tau_R}\|X_n(s)\|_p^p\,\mathrm{d}s
\leq\|P_nx\|_{H_D}^2+\mathbb{E}\int_0^{t\wedge\tau_R}\|\Sigma_n(s)\|_{\mathcal{L}_2}^2\,\mathrm{d}s+\mathbb{E}\sup_{\theta\leq t\wedge\tau_R}|M_n(\theta)|. E θ ≤ t ∧ τ R sup ∥ X n ( θ ) ∥ H D 2 + 2 E ∫ 0 t ∧ τ R ∥ X n ( s ) ∥ p p d s ≤ ∥ P n x ∥ H D 2 + E ∫ 0 t ∧ τ R ∥ Σ n ( s ) ∥ L 2 2 d s + E θ ≤ t ∧ τ R sup ∣ M n ( θ ) ∣.
Since P n P_n P n and Π n \Pi_n Π n are contractions,
∥ Σ n ( s ) ∥ L 2 2 ≤ ∥ Σ ( s , X n ( s ) ) ∥ L 2 2 ≤ 2 ∥ g ( s ) ∥ L 2 2 + 2 L ∥ X n ( s ) ∥ H D 2 . \|\Sigma_n(s)\|_{\mathcal{L}_2}^2\leq\|\Sigma(s,X_n(s))\|_{\mathcal{L}_2}^2\leq2\|g(s)\|_{\mathcal{L}_2}^2+2L\|X_n(s)\|_{H_D}^2. ∥ Σ n ( s ) ∥ L 2 2 ≤ ∥Σ ( s , X n ( s )) ∥ L 2 2 ≤ 2∥ g ( s ) ∥ L 2 2 + 2 L ∥ X n ( s ) ∥ H D 2 .
Moreover, the BDG inequality followed by Young’s inequality yields
E sup θ ≤ t ∧ τ R ∣ M n ( θ ) ∣ ≤ C E [ sup θ ≤ t ∧ τ R ∥ X n ( θ ) ∥ H D ( ∫ 0 t ∧ τ R ∥ Σ n ( s ) ∥ L 2 2 d s ) 1 / 2 ] ≤ 1 2 E sup θ ≤ t ∧ τ R ∥ X n ( θ ) ∥ H D 2 + C E ∫ 0 t ∧ τ R ∥ Σ n ( s ) ∥ L 2 2 d s . \begin{aligned}
\mathbb{E}\sup_{\theta\leq t\wedge\tau_R}|M_n(\theta)|
\leq C\mathbb{E}\left[\sup_{\theta\leq t\wedge\tau_R}\|X_n(\theta)\|_{H_D}\left(\int_0^{t\wedge\tau_R}\|\Sigma_n(s)\|_{\mathcal{L}_2}^2\,\mathrm{d}s\right)^{1/2}\right] \\
&\leq\frac{1}{2}\mathbb{E}\sup_{\theta\leq t\wedge\tau_R}\|X_n(\theta)\|_{H_D}^2+CE\int_0^{t\wedge\tau_R}\|\Sigma_n(s)\|_{\mathcal{L}_2}^2\,\mathrm{d}s.
\end{aligned} E θ ≤ t ∧ τ R sup ∣ M n ( θ ) ∣ ≤ C E [ θ ≤ t ∧ τ R sup ∥ X n ( θ ) ∥ H D ( ∫ 0 t ∧ τ R ∥ Σ n ( s ) ∥ L 2 2 d s ) 1/2 ] ≤ 2 1 E θ ≤ t ∧ τ R sup ∥ X n ( θ ) ∥ H D 2 + C E ∫ 0 t ∧ τ R ∥ Σ n ( s ) ∥ L 2 2 d s .
Consequently,
1 2 E sup θ ≤ t ∧ τ R ∥ X n ( θ ) ∥ H D 2 + 2 E ∫ 0 t ∧ τ R ∥ X n ( s ) ∥ p p d s ≤ ∥ x ∥ H D 2 + C ∫ 0 t ∥ g ( s ) ∥ L 2 2 d s + C L ∫ 0 t E sup θ ≤ s ∧ τ R ∥ X n ( θ ) ∥ H D 2 d s . \frac{1}{2}\mathbb{E}\sup_{\theta\leq t\wedge\tau_R}\|X_n(\theta)\|_{H_D}^2+2\mathbb{E}\int_0^{t\wedge\tau_R}\|X_n(s)\|_p^p\,\mathrm{d}s
\leq\|x\|_{H_D}^2+C\int_0^t\|g(s)\|_{\mathcal{L}_2}^2\,\mathrm{d}s+CL\int_0^t\mathbb{E}\sup_{\theta\leq s\wedge\tau_R}\|X_n(\theta)\|_{H_D}^2\,\mathrm{d}s. 2 1 E θ ≤ t ∧ τ R sup ∥ X n ( θ ) ∥ H D 2 + 2 E ∫ 0 t ∧ τ R ∥ X n ( s ) ∥ p p d s ≤ ∥ x ∥ H D 2 + C ∫ 0 t ∥ g ( s ) ∥ L 2 2 d s + C L ∫ 0 t E θ ≤ s ∧ τ R sup ∥ X n ( θ ) ∥ H D 2 d s .
Gronwall’s inequality gives
sup n E sup θ ≤ T ∧ τ R ∥ X n ( θ ) ∥ H D 2 ≤ C T ( ∥ x ∥ H D 2 + ∫ 0 T ∥ g ( s ) ∥ L 2 2 d s ) . \sup_n\mathbb{E}\sup_{\theta\leq T\wedge\tau_R}\|X_n(\theta)\|_{H_D}^2\leq C_T\left(\|x\|_{H_D}^2+\int_0^T\|g(s)\|_{\mathcal{L}_2}^2\,\mathrm{d}s\right). n sup E θ ≤ T ∧ τ R sup ∥ X n ( θ ) ∥ H D 2 ≤ C T ( ∥ x ∥ H D 2 + ∫ 0 T ∥ g ( s ) ∥ L 2 2 d s ) .
with a constant independent of n n n and R R R . Substitution into the preceding energy inequality gives the same uniform control of E ∫ 0 T ∧ τ R ∥ X n ( s ) ∥ p p d s \mathbb{E}\int_0^{T\wedge\tau_R}\|X_n(s)\|_p^p\,\mathrm{d}s E ∫ 0 T ∧ τ R ∥ X n ( s ) ∥ p p d s . Letting R ↑ ∞ R\uparrow\infty R ↑ ∞ and using Fatou’s lemma for the supremum term and monotone convergence for the nonnegative integral proves (10) ; in particular, the stopped solutions cannot explode.
For later use we also record why higher H D H_D H D moments are available. Applying Itô’s formula to
Ψ q ( y ) = ( 1 + ∥ y ∥ H D 2 ) q / 2 , q ≥ 2 , \Psi_q(y)=(1+\|y\|_{H_D}^2)^{q/2},\qquad q\geq2, Ψ q ( y ) = ( 1 + ∥ y ∥ H D 2 ) q /2 , q ≥ 2 ,
the dissipative drift contributes a nonpositive term, while the first- and second-order diffusion terms are bounded, by the linear-growth estimate above, by C q ( 1 + ∥ g ( s ) ∥ L 2 q + Ψ q ( X n ( s ) ) ) C_q(1+\|g(s)\|_{\mathcal{L}_2}^q+\Psi_q(X_n(s))) C q ( 1 + ∥ g ( s ) ∥ L 2 q + Ψ q ( X n ( s ))) . The same stopped BDG–Young–Gronwall argument therefore gives
sup n E sup s ≤ T ( 1 + ∥ X n ( s ) ∥ H D 2 ) q / 2 < ∞ \sup_n\mathbb{E}\sup_{s\leq T}(1+\|X_n(s)\|_{H_D}^2)^{q/2}<\infty n sup E s ≤ T sup ( 1 + ∥ X n ( s ) ∥ H D 2 ) q /2 < ∞
for every fixed $q\geq 2.
Step 2: identification of the weak Galerkin limit. The energy bound gives a weakly convergent subsequence of X n X_n X n in L p L^p L p and in L 2 ( Ω × ( 0 , T ) ; H D ) L^2(\Omega\times(0,T);H_D) L 2 ( Ω × ( 0 , T ) ; H D ) . The drifts are bounded in L p / ( p − 1 ) ( Ω × ( 0 , T ) ; V ∗ ) L^{p/(p-1)}(\Omega\times(0,T);V^*) L p / ( p − 1 ) ( Ω × ( 0 , T ) ; V ∗ ) , so denote their weak limit by A ‾ \overline{A} A . Affine linearity identifies the weak noise limit with Σ ( ⋅ , X ) \Sigma(\cdot,X) Σ ( ⋅ , X ) : in a pairing with a fixed Hilbert–Schmidt test, move P n P_n P n and Π n \Pi_n Π n to that test and use their strong convergence. Testing the Galerkin equations against each fixed finite span gives
X ( t ) = x − ∫ 0 t A ‾ ( s ) d s + ∫ 0 t Σ ( s , X ( s ) ) d W s in V ∗ . X(t)=x-\int_0^t\overline{A}(s)\,\mathrm{d}s+\int_0^t\Sigma(s,X(s))\,\mathrm{d}W_s\quad\text{in }V^*. X ( t ) = x − ∫ 0 t A ( s ) d s + ∫ 0 t Σ ( s , X ( s )) d W s in V ∗ .
By using the variational Itô formula, we can get an H D H_D H D -continuous version. We next justify the fixed-time weak convergence needed in the Minty argument, without assuming any path-space strong convergence. The estimate (10) makes { X n ( t ) } n \{X_n(t)\}_n { X n ( t ) } n bounded in L 2 ( Ω ; H D ) L^2(\Omega;H_D) L 2 ( Ω ; H D ) for every fixed t t t . Hence a subsequence converges weakly there to some Y ( t ) Y(t) Y ( t ) . Let h h h belong to a fixed finite eigenspace. For all sufficiently large n n n , testing the Galerkin equation with h h h gives
( X n ( t ) , h ) H D = ( P n x , h ) H D − ∫ 0 t V ∗ ⟨ A ( X n ( s ) ) , h ⟩ V d s + ( ∫ 0 t Σ n ( s ) d W s , h ) H D . (X_n(t),h)_{H_D}=(P_nx,h)_{H_D}-\int_0^t {}_{V^*}\langle A(X_n(s)),h\rangle_V\,\mathrm{d}s+\left(\int_0^t\Sigma_n(s)\,\mathrm{d}W_s,h\right)_{H_D}. ( X n ( t ) , h ) H D = ( P n x , h ) H D − ∫ 0 t V ∗ ⟨ A ( X n ( s )) , h ⟩ V d s + ( ∫ 0 t Σ n ( s ) d W s , h ) H D .
The time-integration map sends L p ′ ( Ω × ( 0 , t ) ; V ∗ ) L^{p'}(\Omega\times(0,t);V^*) L p ′ ( Ω × ( 0 , t ) ; V ∗ ) continuously into the corresponding scalar L p ′ ( Ω ) L^{p'}(\Omega) L p ′ ( Ω ) pairing with h h h , while stochastic integration is a bounded linear map from the predictable L 2 ( Ω × ( 0 , t ) ; L 2 ( U , H D ) ) L^2(\Omega\times(0,t);L_2(U,H_D)) L 2 ( Ω × ( 0 , t ) ; L 2 ( U , H D )) space into L 2 ( Ω ; H D ) L^2(\Omega;H_D) L 2 ( Ω ; H D ) . Thus, after multiplying by an arbitrary bounded random variable Z Z Z , the weak convergences of A ( X n ) A(X_n) A ( X n ) and Σ n \Sigma_n Σ n pass through the two integrals. The limit of the scalar pairing is
( x , h ) H D − ∫ 0 t V ∗ ⟨ A ‾ ( s ) , h ⟩ V d s + ( ∫ 0 t Σ ( s , X ( s ) ) d W s , h ) H D = ( X ( t ) , h ) H D . (x,h)_{H_D}-\int_0^t {}_{V^*}\langle\overline{A}(s),h\rangle_V\,\mathrm{d}s+\left(\int_0^t\Sigma(s,X(s))\,\mathrm{d}W_s,h\right)_{H_D}=(X(t),h)_{H_D}. ( x , h ) H D − ∫ 0 t V ∗ ⟨ A ( s ) , h ⟩ V d s + ( ∫ 0 t Σ ( s , X ( s )) d W s , h ) H D = ( X ( t ) , h ) H D .
Finite eigenspaces are dense in H D H_D H D , bounded random variables are dense in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) , and the variables X n ( t ) X_n(t) X n ( t ) are uniformly bounded in L 2 ( Ω ; H D ) L^2(\Omega;H_D) L 2 ( Ω ; H D ) . Finite sums of simple tensors Z h Zh Z h are therefore dense in L 2 ( Ω ; H D ) L^2(\Omega;H_D) L 2 ( Ω ; H D ) , so Y ( t ) = X ( t ) Y(t)=X(t) Y ( t ) = X ( t ) . Consequently
X n ( t ) ⇀ X ( t ) in L 2 ( Ω ; H D ) X_n(t)\rightharpoonup X(t)\qquad\text{in }L^2(\Omega;H_D) X n ( t ) ⇀ X ( t ) in L 2 ( Ω ; H D )
for every fixed t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] .
We spell out the monotonicity argument because weak convergence by itself does not identify a nonlinear drift. Choose κ > L \kappa>L κ > L . The weighted Galerkin energy identity and contractivity of the noise projections give
J n : = E ∫ 0 T e − κ s ( 2 ⟨ A ( X n ) , X n ⟩ + κ ∥ X n ∥ H D 2 − ∥ Σ ( s , X n ) ∥ L 2 2 ) d s ≤ ∥ P n x ∥ H D 2 − E e − κ T ∥ X n ( T ) ∥ H D 2 . J_n:=\mathbb{E}\int_0^T e^{-\kappa s}\left(2\langle A(X_n),X_n\rangle+\kappa\|X_n\|_{H_D}^2-\|\Sigma(s,X_n)\|_{L_2}^2\right)\,\mathrm{d}s
\leq\|P_nx\|_{H_D}^2-\mathbb{E}e^{-\kappa T}\|X_n(T)\|_{H_D}^2. J n := E ∫ 0 T e − κ s ( 2 ⟨ A ( X n ) , X n ⟩ + κ ∥ X n ∥ H D 2 − ∥Σ ( s , X n ) ∥ L 2 2 ) d s ≤ ∥ P n x ∥ H D 2 − E e − κ T ∥ X n ( T ) ∥ H D 2 .
Lower semicontinuity of the terminal norm and the energy identity for the candidate show that
lim sup n J n ≤ E ∫ 0 T e − κ s ( 2 ⟨ A ‾ , X ⟩ + κ ∥ X ∥ H D 2 − ∥ Σ ( s , X ) ∥ L 2 2 ) d s . \limsup_{n}J_n\leq\mathbb{E}\int_0^T e^{-\kappa s}\left(2\langle\overline{A},X\rangle+\kappa\|X\|_{H_D}^2-\|\Sigma(s,X)\|_{L_2}^2\right)\,\mathrm{d}s. n lim sup J n ≤ E ∫ 0 T e − κ s ( 2 ⟨ A , X ⟩ + κ ∥ X ∥ H D 2 − ∥Σ ( s , X ) ∥ L 2 2 ) d s .
For a predictable v ∈ L p ( Ω × ( 0 , T ) ; V ) v\in L^p(\Omega\times(0,T);V) v ∈ L p ( Ω × ( 0 , T ) ; V ) , monotonicity and the Lipschitz noise bound imply
0 ≤ E ∫ 0 T e − κ s ( 2 ⟨ A ( X n ) − A ( v ) , X n − v ⟩ + κ ∥ X n − v ∥ H D 2 − ∥ Σ ( s , X n ) − Σ ( s , v ) ∥ L 2 2 ) d s . 0\leq\mathbb{E}\int_0^T e^{-\kappa s}\left(2\langle A(X_n)-A(v),X_n-v\rangle+\kappa\|X_n-v\|_{H_D}^2-\|\Sigma(s,X_n)-\Sigma(s,v)\|_{L_2}^2\right)\,\mathrm{d}s. 0 ≤ E ∫ 0 T e − κ s ( 2 ⟨ A ( X n ) − A ( v ) , X n − v ⟩ + κ ∥ X n − v ∥ H D 2 − ∥Σ ( s , X n ) − Σ ( s , v ) ∥ L 2 2 ) d s .
Expand this expression. Its quadratic part is J n J_n J n , and all remaining terms pass to the limit by weak convergence. The preceding limsup bound therefore yields
0 ≤ E ∫ 0 T e − κ s ( 2 ⟨ A ‾ − A ( v ) , X − v ⟩ + κ ∥ X − v ∥ H D 2 − ∥ Σ ( s , X ) − Σ ( s , v ) ∥ L 2 2 ) d s . 0\leq\mathbb{E}\int_0^T e^{-\kappa s}\left(2\langle\overline{A}-A(v),X-v\rangle+\kappa\|X-v\|_{H_D}^2-\|\Sigma(s,X)-\Sigma(s,v)\|_{L_2}^2\right)\,\mathrm{d}s. 0 ≤ E ∫ 0 T e − κ s ( 2 ⟨ A − A ( v ) , X − v ⟩ + κ ∥ X − v ∥ H D 2 − ∥Σ ( s , X ) − Σ ( s , v ) ∥ L 2 2 ) d s .
Take v = X − ε w v=X-\varepsilon w v = X − εw and v = X + ε w v=X+\varepsilon w v = X + εw , first with bounded predictable V V V -valued w w w . Divide by ε \varepsilon ε and let ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 . Hemicontinuity and the polynomial growth bound give A ‾ = A ( X ) \overline{A}=A(X) A = A ( X ) by the two opposite inequalities and density. This is the required Minty identification.
For two solutions driven by the same noise and with the same initial condition, Itô’s formula and monotonicity give
E ∥ X ( t ) − X ~ ( t ) ∥ H D 2 ≤ L ∫ 0 t E ∥ X ( s ) − X ~ ( s ) ∥ H D 2 d s . \mathbb{E}\lVert X(t)-\widetilde{X}(t)\rVert_{H_D}^{2}\le L\int_{0}^{t}\mathbb{E}\lVert X(s)-\widetilde{X}(s)\rVert_{H_D}^{2}\,\mathrm{d}s. E ∥ X ( t ) − X ( t ) ∥ H D 2 ≤ L ∫ 0 t E ∥ X ( s ) − X ( s ) ∥ H D 2 d s .
Gronwall proves uniqueness. Thus every weakly convergent subsequence has the same limit, and the whole Galerkin sequence converges in the asserted weak topologies.
Step 3: the scalar observation at a prescribed time. One still needs a fixed-time assertion for Malliavin calculus. Let F n ( s ) = ⟨ X n ( s ) , φ ⟩ F_n(s)=\langle X_n(s),\varphi\rangle F n ( s ) = ⟨ X n ( s ) , φ ⟩ . Its scalar equation has drift
∫ O Φ ( X n ) Δ φ n d x \int_{\mathcal{O}}\Phi(X_n)\Delta\varphi_n\,\mathrm{d}x ∫ O Φ ( X n ) Δ φ n d x
and noise Π n Σ ( s , X n ) ∗ φ n \Pi_n\Sigma(s,X_n)*\varphi_n Π n Σ ( s , X n ) ∗ φ n . By Lemma 2.6, Hölder’s inequality and (10) ,
sup n E ∣ F n ( t ) − F n ( s ) ∣ ≤ C φ ( ∣ t − s ∣ 1 p + ∣ t − s ∣ 1 2 ) . (12) \sup_n\mathbb{E}|F_n(t)-F_n(s)|\le C_\varphi\left(|t-s|^{\frac{1}{p}}+|t-s|^{\frac{1}{2}}\right). \tag*{(12)} n sup E ∣ F n ( t ) − F n ( s ) ∣ ≤ C φ ( ∣ t − s ∣ p 1 + ∣ t − s ∣ 2 1 ) . (12)
For completeness, we indicate the estimate behind (12) . The drift increment satisfies, uniformly in n n n ,
E ∣ ∫ s t ∫ O Φ ( X n ( q ) ) Δ φ n d x d q ∣ ≤ C φ E ∫ s t ∥ X n ( q ) ∥ p p − 1 d q ≤ C φ ∣ t − s ∣ 1 / p ( E ∫ s t ∥ X n ( q ) ∥ p p d q ) ( p − 1 ) / p ≤ C φ ∣ t − s ∣ 1 / p , \begin{aligned}
\mathbb{E}\left|\int_s^t\int_{\mathcal{O}}\Phi(X_n(q))\Delta\varphi_n\,\mathrm{d}x\,\mathrm{d}q\right|
\le C_\varphi\mathbb{E}\int_s^t\lVert X_n(q)\rVert_p^{p-1}\,\mathrm{d}q \\
&\le C_\varphi|t-s|^{1/p}\left(\mathbb{E}\int_s^t\lVert X_n(q)\rVert_p^p\,\mathrm{d}q\right)^{(p-1)/p}
\le C_\varphi|t-s|^{1/p},
\end{aligned} E ∫ s t ∫ O Φ ( X n ( q )) Δ φ n d x d q ≤ C φ E ∫ s t ∥ X n ( q ) ∥ p p − 1 d q ≤ C φ ∣ t − s ∣ 1/ p ( E ∫ s t ∥ X n ( q ) ∥ p p d q ) ( p − 1 ) / p ≤ C φ ∣ t − s ∣ 1/ p ,
where (8) and (10) were used. The stochastic increment is bounded by Itô’s isometry and the linear-growth estimate:
E ∣ ∫ s t ( Π n Σ ( q , X n ( q ) ) ∗ φ n , d W q ) U ∣ ≤ C φ ∣ t − s ∣ 1 / 2 . \mathbb{E}\left|\int_s^t(\Pi_n\Sigma(q,X_n(q))*\varphi_n,\mathrm{d}W_q)_U\right|\le C_\varphi|t-s|^{1/2}. E ∫ s t ( Π n Σ ( q , X n ( q )) ∗ φ n , d W q ) U ≤ C φ ∣ t − s ∣ 1/2 .
This proves (12) . It also yields the one-sided averaging bound
∣ E [ Z F n ( t ) ] − 1 h ∫ t − h t E [ Z F n ( s ) ] d s ∣ ≤ C φ ∥ Z ∥ ∞ ( h 1 / p + h 1 / 2 ) , 0 < h < t , \left|\mathbb{E}[ZF_n(t)]-\frac{1}{h}\int_{t-h}^{t}\mathbb{E}[ZF_n(s)]\,\mathrm{d}s\right|
\le C_\varphi\lVert Z\rVert_\infty\left(h^{1/p}+h^{1/2}\right),\qquad0<h<t, E [ Z F n ( t )] − h 1 ∫ t − h t E [ Z F n ( s )] d s ≤ C φ ∥ Z ∥ ∞ ( h 1/ p + h 1/2 ) , 0 < h < t ,
for every bounded random variable Z Z Z . The weak convergence in L 2 ( Ω × ( 0 , T ) ; H D ) L^2(\Omega\times(0,T);H_D) L 2 ( Ω × ( 0 , T ) ; H D ) identifies the limit of the time average, whereas the same estimate applies to the limiting variational equation. Letting h ↓ 0 h\downarrow0 h ↓ 0 gives convergence against bounded Z Z Z . Alternatively, and more directly, the fixed-time weak convergence proved in Step 2 implies
⟨ X n ( t ) , φ ⟩ ⇀ ⟨ X ( t ) , φ ⟩ in L 2 ( Ω ) , \langle X_n(t),\varphi\rangle\rightharpoonup\langle X(t),\varphi\rangle\quad\text{in }L^2(\Omega), ⟨ X n ( t ) , φ ⟩ ⇀ ⟨ X ( t ) , φ ⟩ in L 2 ( Ω ) ,
because φ ∈ H 0 1 ( O ) \varphi\in H_0^1(\mathcal{O}) φ ∈ H 0 1 ( O ) defines a continuous functional on H D H_D H D . This is precisely (11) . No strong convergence in L 2 ( Ω ; C ( [ 0 , T ] ; H D ) ) L^2(\Omega;C([0,T];H_D)) L 2 ( Ω ; C ([ 0 , T ] ; H D )) is used.
We note that the regularity of time and space are not optimal, see [7 ] .
Remark 2.8 (Optional scalar smoothing). If a smooth scalar nonlinearity is desired, one may use
Φ ε ( r ) = r ( r 2 + ε 2 ) ( m − 1 ) / 2 + ε r . \Phi_\varepsilon(r)=r(r^2+\varepsilon^2)^{(m-1)/2}+\varepsilon r. Φ ε ( r ) = r ( r 2 + ε 2 ) ( m − 1 ) /2 + ε r .
Then
r Φ ε ( r ) ≥ ∣ r ∣ m + 1 and 0 < Φ ε ′ ( r ) ≤ C m ( 1 + ∣ r ∣ m − 1 ) for 0 < ε ≤ 1. r\Phi_\varepsilon(r)\ge|r|^{m+1}\qquad\text{and}\qquad0<\Phi_\varepsilon'(r)\le C_m(1+|r|^{m-1})\qquad\text{for }0<\varepsilon\le1. r Φ ε ( r ) ≥ ∣ r ∣ m + 1 and 0 < Φ ε ′ ( r ) ≤ C m ( 1 + ∣ r ∣ m − 1 ) for 0 < ε ≤ 1.
Its derivative is not globally bounded. The Galerkin construction and energy estimates remain compatible, and monotone variational convergence removes ε \varepsilon ε . The direct C 1 C^1 C 1 Galerkin construction above makes this additional parameter unnecessary. In particular, the proof below does not use an ε \varepsilon ε -regularization. If this optional route is adopted, one must additionally pass ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 in the Malliavin derivatives by a uniform derivative bound and the weak closedness of the Malliavin derivative.
Malliavin differentiation before passage to the limit
Lemma 2.9 (Finite-dimensional derivative and its energy). For fixed n n n , X n ( t ) X_n(t) X n ( t ) is Malliavin differentiable. For k ≤ n k \le n k ≤ n and almost every r ≤ t r \le t r ≤ t , set Y r , k n ( s ) = D r k X n ( s ) Y^{n}_{r,k}(s) = D^k_r X_n(s) Y r , k n ( s ) = D r k X n ( s ) , s ≥ r s \ge r s ≥ r . It satisfies
d Y r , k n = P n Δ ( a n Y r , k n ) d s + ∑ ℓ ≤ n P n C ℓ Y r , k n d β ℓ ( s ) , \mathrm{d}Y^{n}_{r,k} = P_n\Delta(a_nY^{n}_{r,k})\,\mathrm{d}s + \sum_{\ell\le n}P_nC_\ell Y^{n}_{r,k}\,\mathrm{d}\beta_\ell(s), d Y r , k n = P n Δ ( a n Y r , k n ) d s + ℓ ≤ n ∑ P n C ℓ Y r , k n d β ℓ ( s ) , {#eq:2.10}
Y^{n}_{r,k}(r) = P_n\Sigma(r,X_n(r))e_k,\qquad a_n(s,x)=m|X_n(s,x)|^{m-1}.
The drift projection here is understood in the finite-dimensional variational sense. With η r , k n = Y r , k n ( r ) \eta^n_{r,k}=Y^n_{r,k}(r) η r , k n = Y r , k n ( r ) ,
sup r ≤ s ≤ T E ∥ Y r , k n ( s ) ∥ H D 2 + E ∫ r T ∫ O a n ∣ Y r , k n ∣ 2 d x d s ≤ C T E ∥ η r , k n ∥ H D 2 . (13) \sup_{r\le s\le T}\mathbb{E}\|Y^n_{r,k}(s)\|_{H_D}^2+\mathbb{E}\int_r^T\int_{\mathcal O}a_n|Y^n_{r,k}|^2\,\mathrm{d}x\,\mathrm{d}s\le C_T\mathbb{E}\|\eta^n_{r,k}\|_{H_D}^2. \tag*{(13)} r ≤ s ≤ T sup E ∥ Y r , k n ( s ) ∥ H D 2 + E ∫ r T ∫ O a n ∣ Y r , k n ∣ 2 d x d s ≤ C T E ∥ η r , k n ∥ H D 2 . (13) Consequently, for F n = ⟨ X n ( t ) , φ ⟩ F_n=\langle X_n(t),\varphi\rangle F n = ⟨ X n ( t ) , φ ⟩ ,
sup n E ∫ 0 t ∥ D r F n ∥ U 2 d r < ∞ . (14) \sup_n\mathbb{E}\int_0^t\|D_rF_n\|_U^2\,\mathrm{d}r<\infty. \tag*{(14)} n sup E ∫ 0 t ∥ D r F n ∥ U 2 d r < ∞. (14) If C = 0 C=0 C = 0 , the stronger pathwise inequality holds:
1 2 ∥ Y r , k n ( s ) ∥ H D 2 + ∫ r s ∫ O a n ∣ Y r , k n ∣ 2 d x d q ≤ 1 2 ∥ P n g ( r ) e k ∥ H D 2 . (15) \frac{1}{2}\|Y^n_{r,k}(s)\|_{H_D}^2+\int_r^s\int_{\mathcal O}a_n|Y^n_{r,k}|^2\,\mathrm{d}x\,\mathrm{d}q\le\frac{1}{2}\|P_ng(r)e_k\|_{H_D}^2. \tag*{(15)} 2 1 ∥ Y r , k n ( s ) ∥ H D 2 + ∫ r s ∫ O a n ∣ Y r , k n ∣ 2 d x d q ≤ 2 1 ∥ P n g ( r ) e k ∥ H D 2 . (15)
Proof. Step 1: verification of the finite-dimensional Malliavin theorem. Use H D H_D H D -orthonormal coordinates on H n H_n H n and write
b n ( v ) = − A n ( v ) , σ n ( s , v ) = P n ( g ( s ) + C v ) Π n . b_n(v)=-A_n(v),\qquad\sigma_n(s,v)=P_n(g(s)+Cv)\Pi_n. b n ( v ) = − A n ( v ) , σ n ( s , v ) = P n ( g ( s ) + C v ) Π n .
We verify the hypotheses of the deterministic-coefficient result [12 Corollary 3.5] , rather than obtaining Malliavin differentiability by a state truncation. Since m > 1 m>1 m > 1 , the scalar map Φ ( r ) = ∣ r ∣ m − 1 r \Phi(r)=|r|^{m-1}r Φ ( r ) = ∣ r ∣ m − 1 r is C 1 C^1 C 1 . In fixed dimension this gives A n ∈ C 1 ( H n ; H n ) A_n\in C^1(H_n;H_n) A n ∈ C 1 ( H n ; H n ) and local Lipschitz continuity. Moreover,
( b n ( v ) − b n ( w ) , v − w ) H D = − ∫ O [ Φ ( v ) − Φ ( w ) ] ( v − w ) d x ≤ 0 , (b_n(v)-b_n(w),v-w)_{H_D}=-\int_{\mathcal O}[\Phi(v)-\Phi(w)](v-w)\,\mathrm{d}x\le0, ( b n ( v ) − b n ( w ) , v − w ) H D = − ∫ O [ Φ ( v ) − Φ ( w )] ( v − w ) d x ≤ 0 ,
( D A n ( v ) y , y ) H D = ∫ O m ∣ v ∣ m − 1 y 2 d x ≥ 0. (D A_n(v)y,y)_{H_D}=\int_{\mathcal O}m|v|^{m-1}y^2\,\mathrm{d}x\ge0. ( D A n ( v ) y , y ) H D = ∫ O m ∣ v ∣ m − 1 y 2 d x ≥ 0.
Thus the drift is one-sided Lipschitz with constant zero. The diffusion is C 1 C^1 C 1 in the state variable, with
D v σ n ( s , v ) y = P n C y Π n , ∥ σ n ( s , v ) − σ n ( s , w ) ∥ L 2 ≤ L ∥ v − w ∥ H D 2 D_v\sigma_n(s,v)y=P_nCy\Pi_n,\qquad\|\sigma_n(s,v)-\sigma_n(s,w)\|_{L^2}\le L\|v-w\|_{H_D}^2 D v σ n ( s , v ) y = P n C y Π n , ∥ σ n ( s , v ) − σ n ( s , w ) ∥ L 2 ≤ L ∥ v − w ∥ H D 2
by using (5) . Moreover, we have
∫ 0 T ∥ σ n ( s , 0 ) ∥ L 2 2 d s < ∞ \int_0^T\|\sigma_n(s,0)\|_{L^2}^2\,\mathrm{d}s<\infty ∫ 0 T ∥ σ n ( s , 0 ) ∥ L 2 2 d s < ∞
because g g g is deterministic and continuous. The initial value P n x P_nx P n x is deterministic. Finally, the last assertion in Step 1 of Proposition 2.7 gives, for every fixed r ≥ 2 r\ge2 r ≥ 2 ,
sup n E sup s ≤ T ( 1 + ∥ X n ( s ) ∥ H D 2 ) r / 2 < ∞ . \sup_n\mathbb{E}\sup_{s\le T}(1+\|X_n(s)\|_{H_D}^2)^{r/2}<\infty. n sup E s ≤ T sup ( 1 + ∥ X n ( s ) ∥ H D 2 ) r /2 < ∞.
In particular the moment requirement with some exponent r > 2 r>2 r > 2 is satisfied. Therefore all hypotheses of the cited finite-dimensional theorem hold and X n X_n X n is Malliavin differentiable. Its derivative is exactly (2.10). This argument is performed at fixed n n n , before any SPDE limit, and does not require either a globally bounded Φ ′ \Phi' Φ ′ or a state cutoff.
Step 2: relation to Cameron–Martin perturbations. For clarity, consider a deterministic control h ∈ L 2 ( 0 , T ; U ) h \in L^2(0,T;U) h ∈ L 2 ( 0 , T ; U ) and the shift
W s θ = W s + θ ∫ 0 s h ( r ) d r . W_s^\theta= W_s + \theta\int_0^s h(r)\,dr. W s θ = W s + θ ∫ 0 s h ( r ) d r .
The shifted finite-dimensional equation acquires the extra drift θ P n Σ ( s , X n θ ) Π n h ( s ) \theta P_n\Sigma(s,X_n^\theta)\Pi_nh(s) θ P n Σ ( s , X n θ ) Π n h ( s ) . On a bounded set of controls, monotonicity, the Lipschitz noise estimate, BDG, and Gronwall give
E sup s ≤ T ∥ X n θ ( s ) − X n ( s ) ∥ H D 2 ≤ C T , R θ 2 ∥ h ∥ L 2 ( 0 , T ; U ) 2 ( 1 + ∥ x ∥ H D 2 + sup s ≤ T ∥ g ( s ) ∥ L 2 2 ) . \mathbb{E}\sup_{s\leq T}\lVert X_n^\theta(s)-X_n(s)\rVert_{H_D}^2 \leq C_{T,R}\theta^2\lVert h\rVert_{L^2(0,T;U)}^2\left(1+\lVert x\rVert_{H_D}^2+\sup_{s\leq T}\lVert g(s)\rVert_{L_2}^2\right). E s ≤ T sup ∥ X n θ ( s ) − X n ( s ) ∥ H D 2 ≤ C T , R θ 2 ∥ h ∥ L 2 ( 0 , T ; U ) 2 ( 1 + ∥ x ∥ H D 2 + s ≤ T sup ∥ g ( s ) ∥ L 2 2 ) .
when ∥ h ∥ L 2 ≤ R \lVert h\rVert_{L^2}\leq R ∥ h ∥ L 2 ≤ R and ∣ θ ∣ ≤ 1 |\theta|\leq1 ∣ θ ∣ ≤ 1 . The drift in the difference quotient contains the averaged Jacobian
∫ 0 1 D A n ( X n + λ ( X n θ − X n ) ) d λ . \int_0^1 DA_n\left(X_n+\lambda\left(X_n^\theta-X_n\right)\right)\,d\lambda. ∫ 0 1 D A n ( X n + λ ( X n θ − X n ) ) d λ .
In fixed dimension this converges locally to D A n ( X n ) DA_n(X_n) D A n ( X n ) , since A n A_n A n is C 1 C^1 C 1 . Localization and the moment bounds identify the directional derivative with the solution Z h Z_h Z h of
d Z h = − D A n ( X n ) Z h d s + ∑ ℓ ≤ n P n C ℓ Z h d β ℓ ( s ) + P n Σ ( s , X n ) Π n h ( s ) d s , Z h ( 0 ) = 0. dZ_h=-DA_n(X_n)Z_h\,ds+\sum_{\ell\leq n}P_nC_\ell Z_h\,d\beta_\ell(s)+P_n\Sigma(s,X_n)\Pi_nh(s)\,ds,\qquad Z_h(0)=0. d Z h = − D A n ( X n ) Z h d s + ℓ ≤ n ∑ P n C ℓ Z h d β ℓ ( s ) + P n Σ ( s , X n ) Π n h ( s ) d s , Z h ( 0 ) = 0.
Variation of constants, or stochastic Fubini followed by uniqueness of this linear equation, then gives
D h X n ( t ) = Z h ( t ) = ∑ k ≤ n ∫ 0 t Y r , k n ( t ) h k ( r ) d r . (16) D_hX_n(t)=Z_h(t)=\sum_{k\leq n}\int_0^t Y_{r,k}^n(t)h_k(r)\,dr. \tag*{(16)} D h X n ( t ) = Z h ( t ) = k ≤ n ∑ ∫ 0 t Y r , k n ( t ) h k ( r ) d r . (16)
In particular, an indicator control gives an integral of the kernel over an interval. It does not give a value of D r X n ( t ) D_rX_n(t) D r X n ( t ) at one time. Equation (16) explains the derivative equation, while the finite-dimensional theorem used in Step 1 supplies membership in the Malliavin Sobolev space. Directional differentiability alone would not supply that membership.
Step 3: estimates independent of dimension. Applying Itô’s formula to (2.10) gives
d ∥ Y ∥ H D 2 + 2 ∫ O a n Y 2 d x d s = ∑ ℓ ≤ n ∥ P n C ℓ Y ∥ H D 2 d s + 2 ∑ ℓ ≤ n ( Y , P n C ℓ Y ) H D d β ℓ . d\lVert Y\rVert_{H_D}^2+2\int_O a_nY^2\,dx\,ds=\sum_{\ell\leq n}\lVert P_nC_\ell Y\rVert_{H_D}^2\,ds+2\sum_{\ell\leq n}(Y,P_nC_\ell Y)_{H_D}\,d\beta_\ell. d ∥ Y ∥ H D 2 + 2 ∫ O a n Y 2 d x d s = ℓ ≤ n ∑ ∥ P n C ℓ Y ∥ H D 2 d s + 2 ℓ ≤ n ∑ ( Y , P n C ℓ Y ) H D d β ℓ .
For simplicity, we write Y = Y r , k n Y=Y_{r,k}^n Y = Y r , k n and η = η r , k n \eta=\eta_{r,k}^n η = η r , k n . We give the complete argument because the uniformity in n n n , r r r and k k k is essential later. For R > 0 R>0 R > 0 , set
τ R = inf { s ≥ r : ∥ Y ( s ) ∥ H D ≥ R } ∧ T . \tau_R=\inf\{s\geq r:\lVert Y(s)\rVert_{H_D}\geq R\}\wedge T. τ R = inf { s ≥ r : ∥ Y ( s ) ∥ H D ≥ R } ∧ T .
Integrating the preceding Itô identity up to s ∧ τ R s\wedge\tau_R s ∧ τ R and taking expectation eliminates the stopped local martingale and gives, by (5) ,
E ∥ Y ( s ∧ τ R ) ∥ H D 2 + 2 E ∫ r s ∧ τ R ∫ O a n ∣ Y ∣ 2 d x d q ≤ E ∥ η ∥ H D 2 + L E ∫ r s ∧ τ R ∥ Y ( q ) ∥ H D 2 d q . \mathbb{E}\lVert Y(s\wedge\tau_R)\rVert_{H_D}^2+2\mathbb{E}\int_r^{s\wedge\tau_R}\int_O a_n|Y|^2\,dx\,dq
\leq\mathbb{E}\lVert\eta\rVert_{H_D}^2+L\mathbb{E}\int_r^{s\wedge\tau_R}\lVert Y(q)\rVert_{H_D}^2\,dq. E ∥ Y ( s ∧ τ R ) ∥ H D 2 + 2 E ∫ r s ∧ τ R ∫ O a n ∣ Y ∣ 2 d x d q ≤ E ∥ η ∥ H D 2 + L E ∫ r s ∧ τ R ∥ Y ( q ) ∥ H D 2 d q .
Dropping the nonnegative space–time integral and applying Gronwall yields
sup r ≤ s ≤ T E ∥ Y ( s ∧ τ R ) ∥ H D 2 ≤ e L ( T − r ) E ∥ η ∥ H D 2 ≤ e L T E ∥ η ∥ H D 2 . \sup_{r\leq s\leq T}\mathbb{E}\lVert Y(s\wedge\tau_R)\rVert_{H_D}^2\leq e^{L(T-r)}\mathbb{E}\lVert\eta\rVert_{H_D}^2\leq e^{LT}\mathbb{E}\lVert\eta\rVert_{H_D}^2. r ≤ s ≤ T sup E ∥ Y ( s ∧ τ R ) ∥ H D 2 ≤ e L ( T − r ) E ∥ η ∥ H D 2 ≤ e L T E ∥ η ∥ H D 2 .
Returning to the preceding inequality with s = T s=T s = T then gives
2 E ∫ r T ∧ τ R ∫ O a n ∣ Y ∣ 2 d x d q ≤ E ∥ η ∥ H D 2 + L ∫ r T E ∥ Y ( q ∧ τ R ) ∥ H D 2 d q ≤ ( 1 + L T e L T ) E ∥ η ∥ H D 2 . 2\mathbb{E}\int_{r}^{T\wedge\tau_R}\int_{\mathcal{O}}a_n|Y|^2\,\mathrm{d}x\,\mathrm{d}q\leq\mathbb{E}\|\eta\|_{H_D}^2+L\int_r^T\mathbb{E}\|Y(q\wedge\tau_R)\|_{H_D}^2\,\mathrm{d}q
\leq(1+LT e^{LT})\mathbb{E}\|\eta\|_{H_D}^2. 2 E ∫ r T ∧ τ R ∫ O a n ∣ Y ∣ 2 d x d q ≤ E ∥ η ∥ H D 2 + L ∫ r T E ∥ Y ( q ∧ τ R ) ∥ H D 2 d q ≤ ( 1 + L T e L T ) E ∥ η ∥ H D 2 .
The constants are independent of n , r , k n,r,k n , r , k and R R R . Letting R ↑ ∞ R\uparrow\infty R ↑ ∞ , Fatou’s lemma and monotone convergence give (13) , with for instance C T = e L T + 1 2 ( 1 + L T e L T ) C_T=e^{LT}+\frac{1}{2}(1+LT e^{LT}) C T = e L T + 2 1 ( 1 + L T e L T ) after enlarging the constant.
The summation over the initial noise direction is legitimate because, by orthogonality of Π n \Pi_n Π n and contractivity of P n P_n P n ,
∑ k ≤ n ∥ η r , k n ∥ H D 2 = ∑ k ≤ n ∥ P n Σ ( r , X n ( r ) ) e k ∥ H D 2 ≤ ∥ Σ ( r , X n ( r ) ) ∥ L 2 ( U , H D ) 2 . \sum_{k\leq n}\|\eta_{r,k}^n\|_{H_D}^2=\sum_{k\leq n}\|P_n\Sigma(r,X_n(r))e_k\|_{H_D}^2\leq\|\Sigma(r,X_n(r))\|_{\mathcal{L}_2(U,H_D)}^2. k ≤ n ∑ ∥ η r , k n ∥ H D 2 = k ≤ n ∑ ∥ P n Σ ( r , X n ( r )) e k ∥ H D 2 ≤ ∥Σ ( r , X n ( r )) ∥ L 2 ( U , H D ) 2 .
Since D r k F n = ⟨ Y n r , k ( t ) , φ ⟩ D_r^kF_n=\langle Y^{r,k}_n(t),\varphi\rangle D r k F n = ⟨ Y n r , k ( t ) , φ ⟩ , duality and (10) give the explicit bound
E ∫ 0 t ∥ D r F n ∥ U 2 d r ≤ C T ∥ φ ∥ H 0 1 2 ∫ 0 t E ∥ Σ ( r , X n ( r ) ) ∥ L 2 2 d r . \mathbb{E}\int_0^t\|D_rF_n\|_U^2\,\mathrm{d}r\leq C_T\|\varphi\|_{H_0^1}^2\int_0^t\mathbb{E}\|\Sigma(r,X_n(r))\|_{\mathcal{L}_2}^2\,\mathrm{d}r. E ∫ 0 t ∥ D r F n ∥ U 2 d r ≤ C T ∥ φ ∥ H 0 1 2 ∫ 0 t E ∥Σ ( r , X n ( r )) ∥ L 2 2 d r .
The right-hand side is uniform in n n n , proving (14) . When C = 0 C=0 C = 0 , both stochastic terms in the energy equation disappear; integration gives (15) pathwise.
Lemma 2.10 (Weak closure at a fixed time). For every t > 0 t>0 t > 0 and φ ∈ C ∞ ( O ) \varphi\in C^\infty(\mathcal{O}) φ ∈ C ∞ ( O ) , F = ⟨ X ( t ) , φ ⟩ F=\langle X(t),\varphi\rangle F = ⟨ X ( t ) , φ ⟩ belongs to D 1 , 2 \mathbb{D}^{1,2} D 1 , 2 . Along any weakly convergent subsequence of the derivatives in (14) , the weak limit is D F DF D F .
Proof. The Malliavin derivative is a closed linear operator from L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) to L 2 ( Ω ; L 2 ( 0 , t ; U ) ) L^2(\Omega;L^2(0,t;U)) L 2 ( Ω ; L 2 ( 0 , t ; U )) ; see [15 ] . Its graph is a closed linear subspace of the product Hilbert space and therefore is weakly closed. Now apply (11) and (14) : along a subsequence, ( F n , D F n ) ⇀ ( F , V ) (F_n,DF_n)\rightharpoonup(F,V) ( F n , D F n ) ⇀ ( F , V ) in that product space. Weak closedness implies F ∈ D 1 , 2 F\in\mathbb{D}^{1,2} F ∈ D 1 , 2 and V = D F V=DF V = D F .
The identification can also be seen directly from Malliavin integration by parts. For a smooth cylindrical random variable Z Z Z and a deterministic control h h h , write δ W \delta_W δ W for the divergence operator. Then
E [ Z ⟨ D F n , h ⟩ L 2 ( 0 , t ; U ) ] = E [ F n δ W ( Z h ) ] . \mathbb{E}[Z\langle DF_n,h\rangle_{L^2(0,t;U)}]=\mathbb{E}[F_n\delta_W(Zh)]. E [ Z ⟨ D F n , h ⟩ L 2 ( 0 , t ; U ) ] = E [ F n δ W ( Z h )] .
Both test quantities on the two sides are in L 2 L^2 L 2 , so both pairings pass under weak convergence. The resulting identity identifies V V V as the Sobolev derivative of F F F . Uniqueness of this derivative identifies every weak subsequential limit. This step uses the scalar fixed-time convergence, not an unproved strong path-space convergence of the approximate states.
The additive terminal estimate and full support
Take g ( s ) = G ∈ L 2 ( U , H D ) g(s)=G\in L_2(U,H_D) g ( s ) = G ∈ L 2 ( U , H D ) and C = 0 C=0 C = 0 . The original L 2 L^2 L 2 -valued additive noise is included. Let u = X u=X u = X and F t , φ = ⟨ u ( t ) , φ ⟩ F_{t,\varphi}=\langle u(t),\varphi\rangle F t , φ = ⟨ u ( t ) , φ ⟩ .
Proposition 2.11 (Additive terminal convergence). For 0 < δ ≤ min ( t , 1 ) 0<\delta\leq\min(t,1) 0 < δ ≤ min ( t , 1 ) ,
E [ 1 δ ∫ t − δ t ∥ D r F t , φ − G ∗ φ ∥ U 2 d r ] ≤ C t , φ δ 2 m + 1 . (17) \mathbb{E}\left[\frac{1}{\delta}\int_{t-\delta}^{t}\|D_rF_{t,\varphi}-G^*\varphi\|_U^2\,\mathrm{d}r\right]\leq C_{t,\varphi}\delta^{\frac{2}{m+1}}. \tag*{(17)} E [ δ 1 ∫ t − δ t ∥ D r F t , φ − G ∗ φ ∥ U 2 d r ] ≤ C t , φ δ m + 1 2 . (17)
Proof. Step 1: the finite-dimensional remainder. At the Galerkin level, pair (2.10) with φ \varphi φ to obtain
D r F n = Π n G ∗ φ n + R r n D_rF_n=\Pi_nG^*\varphi_n+R_r^n D r F n = Π n G ∗ φ n + R r n
with
R r , k n = ∫ r t ∫ O a n Y r , k n Δ φ n d x d s . R_{r,k}^{n}=\int_r^t\int_{\mathcal O}a_nY_{r,k}^{n}\Delta\varphi_n\,dx\,ds. R r , k n = ∫ r t ∫ O a n Y r , k n Δ φ n d x d s .
Weighted Cauchy–Schwarz, followed by (15) summed over k k k , yields
∥ R r n ∥ U 2 ≤ 1 2 ∥ G ∥ L 2 2 C n , δ , C n , δ = ∫ t − δ t ∫ O a n ∣ Δ φ n ∣ 2 d x d s . \lVert R_r^n\rVert_{\mathcal U}^2\le\frac{1}{2}\lVert G\rVert_{\mathcal L_2}^2 C_{n,\delta},\qquad C_{n,\delta}=\int_{t-\delta}^t\int_{\mathcal O}a_n\lvert\Delta\varphi_n\rvert^2\,dx\,ds. ∥ R r n ∥ U 2 ≤ 2 1 ∥ G ∥ L 2 2 C n , δ , C n , δ = ∫ t − δ t ∫ O a n ∣ Δ φ n ∣ 2 d x d s .
Indeed, the two factors before summing in k k k are
∫ r t ∫ O a n ∣ Y r , k n ∣ 2 \int_r^t\int_{\mathcal O}a_n\lvert Y_{r,k}^n\rvert^2 ∫ r t ∫ O a n ∣ Y r , k n ∣ 2
and
∫ r t ∫ O a n ∣ Δ φ n ∣ 2 . \int_r^t\int_{\mathcal O}a_n\lvert\Delta\varphi_n\rvert^2. ∫ r t ∫ O a n ∣ Δ φ n ∣ 2 .
The first has a pathwise bound from the additive derivative energy, and the second is at most C n , δ C_{n,\delta} C n , δ for r ∈ [ t − δ , t ] r\in[t-\delta,t] r ∈ [ t − δ , t ] . The integral is over all of O \mathcal O O because the projected test need not retain the support of φ \varphi φ .
Step 2: smallness of the terminal strip. Set α = ( m − 1 ) / ( m + 1 ) \alpha=(m-1)/(m+1) α = ( m − 1 ) / ( m + 1 ) and
Z n = ∫ 0 T ∥ X n ( s ) ∥ p p d s . Z_n=\int_0^T\lVert X_n(s)\rVert_p^p\,ds. Z n = ∫ 0 T ∥ X n ( s ) ∥ p p d s .
By (8) ,
C n , δ ≤ C φ δ 1 − α Z n α , sup n E Z n < ∞ . (18) C_{n,\delta}\le C_\varphi\delta^{1-\alpha}Z_n^\alpha,\qquad\sup_n\mathbb E Z_n<\infty. \tag*{(18)} C n , δ ≤ C φ δ 1 − α Z n α , n sup E Z n < ∞. (18)
Thus
sup n E C n , δ ≤ C δ 1 − α . \sup_n\mathbb E C_{n,\delta}\le C\delta^{1-\alpha}. n sup E C n , δ ≤ C δ 1 − α .
To see the power of δ \delta δ , use Hölder on [ t − δ , t ] × O [t-\delta,t]\times\mathcal O [ t − δ , t ] × O :
∫ t − δ t ∫ O ∣ X n ∣ m − 1 d x d s ≤ ( δ ∣ O ∣ ) 1 − α ( ∫ t − δ t ∫ O ∣ X n ∣ m + 1 d x d s ) α . \int_{t-\delta}^t\int_{\mathcal O}\lvert X_n\rvert^{m-1}\,dx\,ds
\le(\delta\lvert\mathcal O\rvert)^{1-\alpha}
\left(\int_{t-\delta}^t\int_{\mathcal O}\lvert X_n\rvert^{m+1}\,dx\,ds\right)^\alpha. ∫ t − δ t ∫ O ∣ X n ∣ m − 1 d x d s ≤ ( δ ∣ O ∣ ) 1 − α ( ∫ t − δ t ∫ O ∣ X n ∣ m + 1 d x d s ) α .
Jensen’s inequality then uses only the first moment of Z n Z_n Z n . Notice that 1 − α = 2 / ( m + 1 ) > 0 1-\alpha=2/(m+1)>0 1 − α = 2/ ( m + 1 ) > 0 for every m > 1 m>1 m > 1 ; no nondegeneracy of the diffusion coefficient a n a_n a n is needed.
Step 3: passage to the degenerate equation. We finally justify the limit passage explicitly. Fix δ > 0 \delta>0 δ > 0 and work in the Hilbert space
K δ = L 2 ( Ω × ( t − δ , t ) ; U ) . \mathcal K_\delta=L^2(\Omega\times(t-\delta,t);\mathcal U). K δ = L 2 ( Ω × ( t − δ , t ) ; U ) .
Lemma 2.10 gives, after extraction, D F n ⇀ D F DF_n\rightharpoonup DF D F n ⇀ D F in K δ \mathcal K_\delta K δ . Since φ n → φ \varphi_n\to\varphi φ n → φ in H 0 1 H_0^1 H 0 1 and G ∈ L 2 ( U , H D ) G\in L_2(\mathcal U,H_D) G ∈ L 2 ( U , H D ) ,
Π n G ∗ φ n ⟶ G ∗ φ strongly in U . \Pi_nG^*\varphi_n\longrightarrow G^*\varphi\qquad\text{strongly in }\mathcal U. Π n G ∗ φ n ⟶ G ∗ φ strongly in U .
Hence
D F n − Π n G ∗ φ n ⇀ D F − G ∗ φ in K δ . DF_n-\Pi_nG^*\varphi_n\rightharpoonup DF-G^*\varphi\qquad\text{in }\mathcal K_\delta. D F n − Π n G ∗ φ n ⇀ D F − G ∗ φ in K δ .
Weak lower semicontinuity of the squared Hilbert norm therefore gives
E 1 δ ∫ t − δ t ∥ D r F − G ∗ φ ∥ U 2 d r ≤ lim inf n → ∞ E 1 δ ∫ t − δ t ∥ D r F n − Π n G ∗ φ n ∥ U 2 d r ≤ C t , φ δ 1 − α = C t , φ δ 2 / ( m + 1 ) . \begin{aligned}
\mathbb E\frac{1}{\delta}\int_{t-\delta}^t\lVert D_rF-G^*\varphi\rVert_{\mathcal U}^2\,dr
\le\liminf_{n\to\infty}\mathbb E\frac{1}{\delta}\int_{t-\delta}^t\lVert D_rF_n-\Pi_nG^*\varphi_n\rVert_{\mathcal U}^2\,dr \\
&\le C_{t,\varphi}\delta^{1-\alpha}=C_{t,\varphi}\delta^{2/(m+1)}.
\end{aligned} E δ 1 ∫ t − δ t ∥ D r F − G ∗ φ ∥ U 2 d r ≤ n → ∞ lim inf E δ 1 ∫ t − δ t ∥ D r F n − Π n G ∗ φ n ∥ U 2 d r ≤ C t , φ δ 1 − α = C t , φ δ 2/ ( m + 1 ) .
This is (17) . The order of limits is important: first n → ∞ n\to\infty n → ∞ for fixed δ \delta δ , and only afterwards δ ↓ 0 \delta\downarrow0 δ ↓ 0 .
For an adapted scalar observation at time t t t , define its Malliavin covariance by
Γ F = ∫ 0 t ∥ D r F ∥ U 2 d r . \Gamma_F = \int_0^t \lVert D_rF\rVert_{\mathcal{U}}^2\,dr. Γ F = ∫ 0 t ∥ D r F ∥ U 2 d r .
We need only positivity of this quantity, not an inverse-moment bound. The following averaged criterion is useful because a Malliavin derivative is an equivalence class in the time variable; pointwise convergence as r ↑ t r \uparrow t r ↑ t is unnecessary.
Lemma 2.12 (Terminal covariance criterion). Suppose F ∈ D 1 , 2 F \in\mathbb{D}^{1,2} F ∈ D 1 , 2 and Θ ∈ L 2 ( Ω ; U ) \Theta\in L^2(\Omega;\mathcal{U}) Θ ∈ L 2 ( Ω ; U ) satisfy
1 δ ∫ t − δ t ∥ D r F − Θ ∥ U 2 d r ⟶ 0 in probability as δ ↓ 0. (19) \frac{1}{\delta}\int_{t-\delta}^t \lVert D_rF-\Theta\rVert_{\mathcal{U}}^2\,dr \longrightarrow0 \quad\text{in probability as }\delta\downarrow0. \tag*{(19)} δ 1 ∫ t − δ t ∥ D r F − Θ ∥ U 2 d r ⟶ 0 in probability as δ ↓ 0. (19) Then
{ ∥ D F ∥ L 2 ( 0 , t ; U ) = 0 } ⊂ { Θ = 0 } up to a null set. \{\lVert DF\rVert_{L^2(0,t;\mathcal{U})}=0\}\subset\{\Theta=0\}\quad\text{up to a null set.} {∥ D F ∥ L 2 ( 0 , t ; U ) = 0 } ⊂ { Θ = 0 } up to a null set. In particular
P ( F = 0 , Θ ≠ 0 ) = 0. \mathbb{P}(F=0,\Theta\ne0)=0. P ( F = 0 , Θ = 0 ) = 0. If Θ ≠ 0 \Theta\ne0 Θ = 0 almost surely, F F F has an absolutely continuous law.
Proof. Select a deterministic sequence δ j ↓ 0 \delta_j\downarrow0 δ j ↓ 0 along which the nonnegative averages in (19) converge almost surely to zero. On { D F = 0 } \{DF=0\} { D F = 0 } every such average equals ∥ Θ ∥ 2 \lVert\Theta\rVert^2 ∥ Θ ∥ 2 . The first inclusion follows. Locality of the Malliavin derivative gives D F = 0 DF=0 D F = 0 almost surely on { F = 0 } \{F=0\} { F = 0 } , proving the second assertion. Bouleau–Hirsch gives the density assertion when the covariance is positive almost surely [15 ] .
The locality step in Lemma 2.12 is stronger than needed when Θ \Theta Θ is deterministic and nonzero. It is retained because the multiplicative terminal direction may vanish precisely on the global zero-state event.
Proposition 2.13 (Density of additive localized observations). Under Assumption 2.4, for every fixed t > 0 t>0 t > 0 and every nonzero φ ∈ C c ∞ ( O ) \varphi\in C^\infty_c(\mathcal{O}) φ ∈ C c ∞ ( O ) , the law of F t , φ F_{t,\varphi} F t , φ is absolutely continuous with respect to Lebesgue measure. In particular, P ( F t , φ = a ) = 0 \mathbb{P}(F_{t,\varphi}=a)=0 P ( F t , φ = a ) = 0 for every a ∈ R a\in\mathbb{R} a ∈ R .
Proof. Lemma 2.10 gives the Sobolev derivative. Proposition 2.11 implies (19) with Θ = G ∗ φ \Theta=G^*\varphi Θ = G ∗ φ by Markov’s inequality. This vector is nonzero by Assumption 2.4. The density conclusion follows from Lemma 2.12 . A singleton is a Lebesgue-null set, which gives the last assertion.
Absolute continuity in Proposition 2.13 does not assert that the density is strictly positive everywhere, or that the law has support all of R \mathbb{R} R . Only the absence of an atom at zero will be used to rule out an interior spatial vacuum.
Lemma 2.14 (Countable-basis support criterion). Let v v v be a random distribution on O \mathcal{O} O . Fix a countable basis ( O j ) (O_j) ( O j ) of interior balls and an event E E E . Suppose that each O j O_j O j admits a nonzero φ j ∈ C c ∞ ( O j ) \varphi_j\in C^\infty_c(O_j) φ j ∈ C c ∞ ( O j ) such that
P ( E ∩ { ⟨ v , φ j ⟩ = 0 } ) = 0. \mathbb{P}\bigl(E\cap\{\langle v,\varphi_j\rangle=0\}\bigr)=0. P ( E ∩ {⟨ v , φ j ⟩ = 0 } ) = 0. Then supp O v = O ‾ \operatorname{supp}_{\mathcal{O}}v=\overline{\mathcal{O}} supp O v = O almost surely on E E E .
Proof. If the support is not full, the distribution vanishes on a nonempty open subset of O \mathcal{O} O . That subset contains at least one basis ball O j O_j O j , and hence ⟨ v , φ j ⟩ = 0 \langle v,\varphi_j\rangle=0 ⟨ v , φ j ⟩ = 0 . Consequently
E ∩ { supp O v ≠ O ‾ } ⊂ ⋃ j ≥ 1 ( E ∩ { ⟨ v , φ j ⟩ = 0 } ) . E\cap\{\operatorname{supp}_{\mathcal{O}}v\ne\overline{\mathcal{O}}\}\subset\bigcup_{j\ge1}\left(E\cap\{\langle v,\varphi_j\rangle=0\}\right). E ∩ { supp O v = O } ⊂ j ≥ 1 ⋃ ( E ∩ {⟨ v , φ j ⟩ = 0 } ) .
The right-hand side is a countable union of null events. One can take balls with rational centers and radii whose closures lie in O \mathcal{O} O ; this explains why an uncountable intersection over all open sets is not required.
Theorem 2.15 (Additive full-domain support). Let u 0 ∈ H D u_0 \in H_D u 0 ∈ H D be deterministic and G ∈ L 2 ( U , H D ) G \in L_2(\mathcal{U},H_D) G ∈ L 2 ( U , H D ) . Assume Assumption 2.4. Then the solution of
d u = Δ Φ ( u ) d t + G d W t \mathrm{d}u = \Delta\Phi(u)\,\mathrm{d}t + G\,\mathrm{d}W_t d u = ΔΦ ( u ) d t + G d W t has, for every fixed t > 0 t>0 t > 0 ,
P ( supp O u ( t ) = O ‾ ) = 1. (20) \mathbb{P}\bigl(\operatorname{supp}_{\mathcal{O}} u(t)=\overline{\mathcal{O}}\bigr)=1. \tag*{(20)} P ( supp O u ( t ) = O ) = 1. (20) Every nonzero smooth localized scalar observation has a density.
Proof. Proposition 2.13 gives the density assertion and excludes a zero pairing for one chosen nonzero test in each basis ball. Apply Lemma 2.14 with E = Ω E=\Omega E = Ω to obtain (20) .
Proposition 2.5 supplies explicit trace-class coefficients for Theorem 2.15. The initial condition in this theorem may be signed, or even zero. In particular, for the nonnegative compactly supported data in Theorem 2.3, (20) removes every interior vacuum at a fixed positive time. This is instantaneous propagation in the precise sense of Definition 2.2.
Nonlocal multiplicative noise: the Dirichlet problem
Let H = H D H=H_D H = H D and let B ∈ L ( H , L 2 ( U , H ) ) B\in\mathcal{L}(H,L_2(\mathcal{U},H)) B ∈ L ( H , L 2 ( U , H )) . Write B ( v ) e k = K k v B(v)e_k=K_kv B ( v ) e k = K k v . We assume
∑ k ≥ 1 ∥ K k v ∥ H 2 ≤ C B ∥ v ∥ H 2 , v ∈ H , (21) \sum_{k\geq1}\lVert K_kv\rVert_H^2\leq C_B\lVert v\rVert_H^2,\qquad v\in H, \tag*{(21)} k ≥ 1 ∑ ∥ K k v ∥ H 2 ≤ C B ∥ v ∥ H 2 , v ∈ H , (21)
and consider
d u ( t , x ) = Δ Φ ( u ( t , x ) ) d t + B ( u ( t , x ) ) d W t , t > 0 , x ∈ O , u ( 0 ) = u 0 ∈ H . (22) \mathrm{d}u(t,x)=\Delta\Phi(u(t,x))\,\mathrm{d}t+B(u(t,x))\,\mathrm{d}W_t,\quad t>0,\ x\in\mathcal{O},\quad u(0)=u_0\in H. \tag*{(22)} d u ( t , x ) = ΔΦ ( u ( t , x )) d t + B ( u ( t , x )) d W t , t > 0 , x ∈ O , u ( 0 ) = u 0 ∈ H . (22)
This is the case g = 0 g=0 g = 0 , C = B C=B C = B of (6) ; its solution and Galerkin approximations are provided by Proposition 2.7. The deterministic initial condition may have either sign.
Well-posedness and nonlocal vacuum ellipticity
Similar to the additive noise, we first introduce the following definition.
Definition 3.1. A family ( K k ) (K_k) ( K k ) is vacuum preserving if, for every open O ⊂ O \mathcal{O}\subset\mathcal{O} O ⊂ O , v ∣ O = 0 v|_{\mathcal{O}}=0 v ∣ O = 0 implies ( K k v ) ∣ O = 0 (K_kv)|_{\mathcal{O}}=0 ( K k v ) ∣ O = 0 for all k k k . The restrictions are understood in the distributional sense.
Pointwise multiplication by sufficiently regular functions has the property in Definition 3.1. An integral operator need not: values of v v v outside O \mathcal{O} O can contribute to K k v K_kv K k v inside O \mathcal{O} O . This is the distinction relevant to activation of a spatial vacuum. It is separate from well-posedness, for which the Hilbert–Schmidt bound (21) is sufficient.
Proposition 3.2 (Well-posedness). Under (21) , equation (22) has a unique variational solution u u u in L 2 ( Ω ; C ( [ 0 , T ] ; H D ) ) ∩ L p ( Ω × ( 0 , T ) × O ) L^2(\Omega;C([0,T];H_D))\cap L^p(\Omega\times(0,T)\times\mathcal{O}) L 2 ( Ω ; C ([ 0 , T ] ; H D )) ∩ L p ( Ω × ( 0 , T ) × O ) . For each finite T T T ,
E sup s ≤ T ∥ u ( s ) ∥ H D 2 + E ∫ 0 T ∥ u ( s ) ∥ p p d s ≤ C T ∥ u 0 ∥ H D 2 . \mathbb{E}\sup_{s\leq T}\lVert u(s)\rVert_{H_D}^2+\mathbb{E}\int_0^T\lVert u(s)\rVert_p^p\,\mathrm{d}s\leq C_T\lVert u_0\rVert_{H_D}^2. E s ≤ T sup ∥ u ( s ) ∥ H D 2 + E ∫ 0 T ∥ u ( s ) ∥ p p d s ≤ C T ∥ u 0 ∥ H D 2 . The zero initial condition gives the identically zero solution.
Proof. Apply Proposition 2.7 with g = 0 g=0 g = 0 and C = B C=B C = B . For completeness, the energy balance now has no additive forcing:
d ∥ u ∥ H D 2 + 2 ∥ u ∥ p p d s = ∥ B ( u ) ∥ L 2 2 d s + 2 ( u , B ( u ) d W s ) H D . d\|u\|_{H_D}^{2}+2\|u\|_{p}^{p}\,ds=\|B(u)\|_{L_2}^{2}\,ds+2(u,B(u)\,dW_s)_{H_D}. d ∥ u ∥ H D 2 + 2∥ u ∥ p p d s = ∥ B ( u ) ∥ L 2 2 d s + 2 ( u , B ( u ) d W s ) H D .
The same BDG and Gronwall argument therefore gives the displayed bound with right-hand side proportional to ∥ u 0 ∥ H D 2 \|u_0\|_{H_D}^{2} ∥ u 0 ∥ H D 2 . For two solutions, monotonicity and (21) control the squared H D H_D H D norm of their difference, proving uniqueness. Since both drift and noise vanish at zero, uniqueness also identifies the zero solution.
The additional assumption below addresses which interior tests see the noise at a nonzero state.
Assumption 3.3 (Nonlocal vacuum ellipticity). For every v ∈ H ∖ { 0 } v\in H\setminus\{0\} v ∈ H ∖ { 0 } and every nonzero φ ∈ C c ∞ ( O ) \varphi\in C^\infty_c(O) φ ∈ C c ∞ ( O ) ,
∥ B ( v ) ∗ φ ∥ U 2 = ∑ k ≥ 1 ∣ ⟨ K k v , φ ⟩ ∣ 2 > 0. (23) \|B(v)^*\varphi\|_U^2=\sum_{k\geq1}|\langle K_kv,\varphi\rangle|^2>0. \tag*{(23)} ∥ B ( v ) ∗ φ ∥ U 2 = k ≥ 1 ∑ ∣ ⟨ K k v , φ ⟩ ∣ 2 > 0. (23)
Assumption 3.3 forces every nonzero state to be visible through every nonzero interior test. In particular, if v v v vanishes on O O O but is globally nonzero, no nonzero test supported in O O O is orthogonal to all the noise directions. Thus nonlocal vacuum ellipticity is incompatible with vacuum preservation for such states. It is a qualitative strict-positivity assumption, with no uniform ellipticity constant as v v v or the test varies.
Proposition 3.4 (An explicit coefficient). Let ( h j ) (h_j) ( h j ) be an orthonormal basis of H H H . On the noise space indexed by ( i , j ) ∈ N 2 (i,j)\in\mathbb{N}^2 ( i , j ) ∈ N 2 , define
K i j v = 2 − i − j ( v , h j ) H h i . K_{ij}v=2^{-i-j}(v,h_j)_Hh_i. K ij v = 2 − i − j ( v , h j ) H h i .
Then (3.1) and (3.3) hold.
Proof. Since ∑ i ≥ 1 4 − i = 1 / 3 \sum_{i\geq1}4^{-i}=1/3 ∑ i ≥ 1 4 − i = 1/3 ,
∑ i , j ∥ K i j v ∥ H 2 ≤ 1 12 ∥ v ∥ H 2 . \sum_{i,j}\|K_{ij}v\|_H^2\leq\frac{1}{12}\|v\|_H^2. i , j ∑ ∥ K ij v ∥ H 2 ≤ 12 1 ∥ v ∥ H 2 .
More explicitly,
∑ i , j ∥ K i j v ∥ H 2 = ( ∑ i 4 − i ) ( ∑ j 4 − j ∣ ( v , h j ) H ∣ 2 ) ≤ 1 12 ∑ j ∣ ( v , h j ) H ∣ 2 . \sum_{i,j}\|K_{ij}v\|_H^2=\left(\sum_i4^{-i}\right)\left(\sum_j4^{-j}|(v,h_j)_H|^2\right)\leq\frac{1}{12}\sum_j|(v,h_j)_H|^2. i , j ∑ ∥ K ij v ∥ H 2 = ( i ∑ 4 − i ) ( j ∑ 4 − j ∣ ( v , h j ) H ∣ 2 ) ≤ 12 1 j ∑ ∣ ( v , h j ) H ∣ 2 .
If v ≠ 0 v\ne0 v = 0 , some ( v , h j ) H (v,h_j)_H ( v , h j ) H is nonzero. If φ ≠ 0 \varphi\ne0 φ = 0 , its continuous functional on H H H is nonzero, so some ⟨ h i , φ ⟩ \langle h_i,\varphi\rangle ⟨ h i , φ ⟩ is nonzero. The corresponding term in (23) is positive. For the Dirichlet eigensystem one may take h j = λ j ρ j h_j=\sqrt{\lambda_j}\rho_j h j = λ j ρ j ; then for function-valued v v v ,
K i j v = 2 − i − j λ i / λ j ρ i ∫ O v ρ j d x , K_{ij}v=2^{-i-j}\sqrt{\lambda_i/\lambda_j}\rho_i\int_Ov\rho_j\,dx, K ij v = 2 − i − j λ i / λ j ρ i ∫ O v ρ j d x ,
an explicit rank-one integral operator.
The two indices in Proposition 3.4 have different roles: j j j detects the nonzero state and i i i distributes a noise direction visible to the chosen spatial test. One can more generally use K i j v = a i j ℓ j ( v ) f i K_{ij}v=a_{ij}\ell_j(v)f_i K ij v = a ij ℓ j ( v ) f i , where the bounded functionals ℓ j \ell_j ℓ j separate points, the profiles f i f_i f i separate nonzero interior tests, a i j > 0 a_{ij}>0 a ij > 0 , and
∑ i , j a i j 2 ∥ ℓ j ∥ 2 ∥ f i ∥ H 2 < ∞ . \sum_{i,j}a_{ij}^2\|\ell_j\|^2\|f_i\|_H^2<\infty. i , j ∑ a ij 2 ∥ ℓ j ∥ 2 ∥ f i ∥ H 2 < ∞.
The same calculation proves the Hilbert–Schmidt bound and nonlocal vacuum ellipticity. Neither construction by itself proves that a solution cannot reach zero.
Malliavin differentiation for the multiplicative equation
It is useful to make explicit which parts of the additive construction survive. Let u n u_n u n solve (9) with g = 0 g=0 g = 0 and C = B C=B C = B . For r ≤ s r\le s r ≤ s , its derivative has initial value P n K k u n ( r ) P_nK_k u_n(r) P n K k u n ( r ) and satisfies
d Y r , k n = P n Δ ( m ∣ u n ∣ m − 1 Y r , k n ) d s + ∑ ℓ ≤ n P n K ℓ Y r , k n d β ℓ ( s ) . \mathrm{d}Y_{r,k}^n=P_n\Delta\left(m|u_n|^{m-1}Y_{r,k}^n\right)\,\mathrm{d}s+\sum_{\ell\le n}P_nK_\ell Y_{r,k}^n\,\mathrm{d}\beta_\ell(s). d Y r , k n = P n Δ ( m ∣ u n ∣ m − 1 Y r , k n ) d s + ℓ ≤ n ∑ P n K ℓ Y r , k n d β ℓ ( s ) .
In particular, the stochastic term remains in the linearized equation.
Proposition 3.5 (Multiplicative scalar Malliavin bound). Under (21) , for each fixed t > 0 t>0 t > 0 and φ ∈ C c ∞ ( O ) \varphi\in C_c^\infty(\mathcal{O}) φ ∈ C c ∞ ( O ) , F t , φ = ⟨ u ( t ) , φ ⟩ F_{t,\varphi}=\langle u(t),\varphi\rangle F t , φ = ⟨ u ( t ) , φ ⟩ belongs to D 1 , 2 \mathbb{D}^{1,2} D 1 , 2 . The approximate derivatives satisfy
sup n E ∫ 0 t ∥ D r ⟨ u n ( t ) , φ ⟩ ∥ U 2 d r ≤ C T , φ C B ∫ 0 t sup n E ∥ u n ( r ) ∥ H D 2 d r < ∞ . \sup_n\mathbb{E}\int_0^t\|D_r\langle u_n(t),\varphi\rangle\|_U^2\,\mathrm{d}r\le C_{T,\varphi}C_B\int_0^t\sup_n\mathbb{E}\|u_n(r)\|_{H_D}^2\,\mathrm{d}r<\infty. n sup E ∫ 0 t ∥ D r ⟨ u n ( t ) , φ ⟩ ∥ U 2 d r ≤ C T , φ C B ∫ 0 t n sup E ∥ u n ( r ) ∥ H D 2 d r < ∞.
Proof. Lemma 2.9 applies to these finite-dimensional coefficients. Summing (13) over the initial noise directions gives
sup r ≤ s ≤ T ∑ k ≤ n E ∥ Y r , k n ( s ) ∥ H D 2 + ∑ k ≤ n E ∫ r T ∫ O m ∣ u n ∣ m − 1 ∣ Y r , k n ∣ 2 d x d s ≤ C T E ∥ B ( u n ( r ) ) ∥ L 2 2 ≤ C T C B E ∥ u n ( r ) ∥ H D 2 . \sup_{r\le s\le T}\sum_{k\le n}\mathbb{E}\|Y_{r,k}^n(s)\|_{H_D}^2+\sum_{k\le n}\mathbb{E}\int_r^T\int_{\mathcal{O}}m|u_n|^{m-1}|Y_{r,k}^n|^2\,\mathrm{d}x\,\mathrm{d}s
\le C_T\mathbb{E}\|B(u_n(r))\|_{L_2}^2\le C_TC_B\mathbb{E}\|u_n(r)\|_{H_D}^2. r ≤ s ≤ T sup k ≤ n ∑ E ∥ Y r , k n ( s ) ∥ H D 2 + k ≤ n ∑ E ∫ r T ∫ O m ∣ u n ∣ m − 1 ∣ Y r , k n ∣ 2 d x d s ≤ C T E ∥ B ( u n ( r )) ∥ L 2 2 ≤ C T C B E ∥ u n ( r ) ∥ H D 2 .
For completeness, the scalar estimate follows directly from duality. Since D r k ⟨ u n ( t ) , φ ⟩ = ⟨ Y r , k n ( t ) , φ ⟩ D_r^k\langle u_n(t),\varphi\rangle=\langle Y_{r,k}^n(t),\varphi\rangle D r k ⟨ u n ( t ) , φ ⟩ = ⟨ Y r , k n ( t ) , φ ⟩ ,
E ∥ D r ⟨ u n ( t ) , φ ⟩ ∥ U 2 = ∑ k ≤ n E ∣ ⟨ Y r , k n ( t ) , φ ⟩ ∣ 2 ≤ ∥ φ ∥ H 0 1 2 ∑ k ≤ n E ∥ Y r , k n ( t ) ∥ H D 2 ≤ C T ∥ φ ∥ H 0 1 2 E ∥ B ( u n ( r ) ) ∥ L 2 2 ≤ C T C B ∥ φ ∥ H 0 1 2 E ∥ u n ( r ) ∥ H D 2 . \begin{aligned}
\mathbb{E}\|D_r\langle u_n(t),\varphi\rangle\|_U^2=\sum_{k\le n}\mathbb{E}|\langle Y_{r,k}^n(t),\varphi\rangle|^2\le\|\varphi\|_{H_0^1}^2\sum_{k\le n}\mathbb{E}\|Y_{r,k}^n(t)\|_{H_D}^2 \\
&\le C_T\|\varphi\|_{H_0^1}^2\mathbb{E}\|B(u_n(r))\|_{L_2}^2 \\
&\le C_TC_B\|\varphi\|_{H_0^1}^2\mathbb{E}\|u_n(r)\|_{H_D}^2.
\end{aligned} E ∥ D r ⟨ u n ( t ) , φ ⟩ ∥ U 2 = k ≤ n ∑ E ∣ ⟨ Y r , k n ( t ) , φ ⟩ ∣ 2 ≤ ∥ φ ∥ H 0 1 2 k ≤ n ∑ E ∥ Y r , k n ( t ) ∥ H D 2 ≤ C T ∥ φ ∥ H 0 1 2 E ∥ B ( u n ( r )) ∥ L 2 2 ≤ C T C B ∥ φ ∥ H 0 1 2 E ∥ u n ( r ) ∥ H D 2 .
Integrating in r ∈ ( 0 , t ) r\in(0,t) r ∈ ( 0 , t ) and using the uniform state estimate (10) gives
sup n E ∫ 0 t ∥ D r ⟨ u n ( t ) , φ ⟩ ∥ U 2 d r ≤ C T , φ C B ∫ 0 t sup n E ∥ u n ( r ) ∥ H D 2 d r < ∞ . \sup_n\mathbb{E}\int_0^t\|D_r\langle u_n(t),\varphi\rangle\|_U^2\,\mathrm{d}r\le C_{T,\varphi}C_B\int_0^t\sup_n\mathbb{E}\|u_n(r)\|_{H_D}^2\,\mathrm{d}r<\infty. n sup E ∫ 0 t ∥ D r ⟨ u n ( t ) , φ ⟩ ∥ U 2 d r ≤ C T , φ C B ∫ 0 t n sup E ∥ u n ( r ) ∥ H D 2 d r < ∞.
The scalar weak convergence in Proposition 2.7 and the graph argument in Lemma 2.10 then identify the derivative of the limiting observable.
Proposition 3.5 provides the Sobolev regularity needed below. Its derivative energy is an expectation estimate. It cannot be inserted as a deterministic bound into a pathwise product with the random coefficient m ∣ u n ∣ m − 1 m|u_n|^{m-1} m ∣ u n ∣ m − 1 .
Next, we consider the random remainder in a mixed norm. Note that the pathwise contraction (15) is unavailable when the derivative equation contains stochastic terms. We instead use (13) in expectation and a mixed norm. Let
α = m − 1 m + 1 , q = m + 1 2 m = 1 1 + α , ρ = 2 q = m + 1 m ∈ ( 1 , 2 ) . (24) \alpha=\frac{m-1}{m+1},\qquad q=\frac{m+1}{2m}=\frac{1}{1+\alpha},\qquad\rho=2q=\frac{m+1}{m}\in(1,2). \tag*{(24)} α = m + 1 m − 1 , q = 2 m m + 1 = 1 + α 1 , ρ = 2 q = m m + 1 ∈ ( 1 , 2 ) . (24)
In particular, q ( 1 − α ) = 1 / m q(1-\alpha)=1/m q ( 1 − α ) = 1/ m .
Proposition 3.6 (Affine-noise terminal convergence). For the affine equation (6) , fix t > 0 t>0 t > 0 and φ ∈ C c ∞ ( O ) \varphi\in C_c^\infty(\mathcal{O}) φ ∈ C c ∞ ( O ) , and set F = ⟨ X ( t ) , φ ⟩ F=\langle X(t),\varphi\rangle F = ⟨ X ( t ) , φ ⟩ and Θ = Σ ( t , X ( t ) ) ∗ φ \Theta=\Sigma(t,X(t))^*\varphi Θ = Σ ( t , X ( t ) ) ∗ φ . Then F ∈ D 1 , 2 F\in\mathbb{D}^{1,2} F ∈ D 1 , 2 and
E [ ( 1 δ ∫ t − δ t ∥ D r F − Θ ∥ U 2 d r ) q ] ⟶ 0 , as δ ↓ 0. (25) \mathbb{E}\left[\left(\frac{1}{\delta}\int_{t-\delta}^t\|D_rF-\Theta\|_U^2\,\mathrm{d}r\right)^q\right]\longrightarrow0,\qquad\text{as }\delta\downarrow0. \tag*{(25)} E [ ( δ 1 ∫ t − δ t ∥ D r F − Θ ∥ U 2 d r ) q ] ⟶ 0 , as δ ↓ 0. (25)
Only the second-order state and linearized estimates are required.
Proof. We divide the proof into five steps.
Step 1: retain the approximate state at the differentiation time. At the Galerkin level write
D r F n − b n ( r ) = R r n + S r n , b n ( r ) = Π n Σ ( r , X n ( r ) ) ∗ φ n , (26) D_rF_n-b_n(r)=R_r^n+S_r^n,\qquad b_n(r)=\Pi_n\Sigma(r,X_n(r))^*\varphi_n, \tag*{(26)} D r F n − b n ( r ) = R r n + S r n , b n ( r ) = Π n Σ ( r , X n ( r ) ) ∗ φ n , (26)
where, for k ≤ n k\le n k ≤ n ,
R r , k n = ∫ r t ∫ O a n Y r , k n Δ φ n d x d s , S r , k n = ∑ ℓ ≤ n ∫ r t ⟨ C ℓ Y r , k n ( s ) , φ n ⟩ d β ℓ ( s ) . R_{r,k}^n=\int_r^t\int_{\mathcal{O}}a_nY_{r,k}^n\Delta\varphi_n\,dx\,ds,\qquad S_{r,k}^n=\sum_{\ell\le n}\int_r^t\langle C_\ell Y_{r,k}^n(s),\varphi_n\rangle\,d\beta_\ell(s). R r , k n = ∫ r t ∫ O a n Y r , k n Δ φ n d x d s , S r , k n = ℓ ≤ n ∑ ∫ r t ⟨ C ℓ Y r , k n ( s ) , φ n ⟩ d β ℓ ( s ) .
The center in (26) is evaluated at r r r , not at t t t . This choice removes any need for a uniform time modulus of the approximate states in H H H .
Step 2: control the product in the drift remainder. Define
A n ( r ) = ∑ k ≤ n ∫ r t ∫ O a n ∣ Y r , k n ∣ 2 d x d s and A ‾ n , δ = δ − 1 ∫ t − δ t A n ( r ) d r . A_n(r)=\sum_{k\le n}\int_r^t\int_{\mathcal{O}}a_n|Y_{r,k}^n|^2\,dx\,ds
\qquad\text{and}\qquad
\overline{A}_{n,\delta}=\delta^{-1}\int_{t-\delta}^t A_n(r)\,dr. A n ( r ) = k ≤ n ∑ ∫ r t ∫ O a n ∣ Y r , k n ∣ 2 d x d s and A n , δ = δ − 1 ∫ t − δ t A n ( r ) d r .
Let C n , δ C_{n,\delta} C n , δ and Z n Z_n Z n be as in (18) . Weighted Cauchy–Schwarz gives
1 δ ∫ t − δ t ∥ R r n ∥ U 2 d r ≤ A ‾ n , δ C n , δ . \frac{1}{\delta}\int_{t-\delta}^t\|R_r^n\|_U^2\,dr\le\overline{A}_{n,\delta}C_{n,\delta}. δ 1 ∫ t − δ t ∥ R r n ∥ U 2 d r ≤ A n , δ C n , δ .
Equations (13) and (10) yield sup n , δ E A ‾ n , δ < ∞ \sup_{n,\delta}\mathbb{E}\overline{A}_{n,\delta}<\infty sup n , δ E A n , δ < ∞ . By (18) ,
E C n , δ 1 / α ≤ C δ ( 1 − α ) / α E Z n . \mathbb{E}C_{n,\delta}^{1/\alpha}\le C\delta^{(1-\alpha)/\alpha}\mathbb{E}Z_n. E C n , δ 1/ α ≤ C δ ( 1 − α ) / α E Z n .
Hölder with exponents 1 / q 1/q 1/ q and 1 / ( 1 − q ) 1/(1-q) 1/ ( 1 − q ) therefore gives the valid random-product estimate
E ( A ‾ n , δ C n , δ ) q ≤ ( E A ‾ n , δ ) q ( E C n , δ 1 / α ) α q ≤ C δ q ( 1 − α ) = C δ 1 / m . (27) \mathbb{E}\left(\overline{A}_{n,\delta}C_{n,\delta}\right)^q
\le\left(\mathbb{E}\overline{A}_{n,\delta}\right)^q
\left(\mathbb{E}C_{n,\delta}^{1/\alpha}\right)^{\alpha q}
\le C\delta^{q(1-\alpha)}
=C\delta^{1/m}. \tag*{(27)} E ( A n , δ C n , δ ) q ≤ ( E A n , δ ) q ( E C n , δ 1/ α ) α q ≤ C δ q ( 1 − α ) = C δ 1/ m . (27)
In particular, no expectation bound has been substituted into a pathwise inequality, and no product of dependent random variables has been factored. The choice of q q q is exactly what makes this possible: q / ( 1 − q ) = 1 / α q/(1-q)=1/\alpha q / ( 1 − q ) = 1/ α . Thus the required moment of C n , δ C_{n,\delta} C n , δ reduces to the first moment of Z n Z_n Z n , already available from the state energy. An estimate of E ( A ‾ n , δ C n , δ ) \mathbb{E}(\overline{A}_{n,\delta}C_{n,\delta}) E ( A n , δ C n , δ ) would require additional integrability that has not been proved.
Step 3: control the stochastic remainder. For the stochastic remainder, Itô isometry, duality and (5) give
E ∥ S r n ∥ U 2 ≤ ∥ φ n ∥ H 0 1 2 ∑ k ≤ n ∫ r t E ∑ ℓ ≤ n ∥ C ℓ Y r , k n ( s ) ∥ H 2 d s ≤ C ∫ r t E ∥ Σ ( r , X n ( r ) ) ∥ L 2 2 d s ≤ C ( t − r ) . \begin{aligned}
\mathbb{E}\|S_r^n\|_U^2\le\|\varphi_n\|_{H_0^1}^2\sum_{k\le n}\int_r^t\mathbb{E}\sum_{\ell\le n}\|C_\ell Y_{r,k}^n(s)\|_H^2\,ds \\
&\le C\int_r^t\mathbb{E}\|\Sigma(r,X_n(r))\|_{L_2}^2\,ds\le C(t-r).
\end{aligned} E ∥ S r n ∥ U 2 ≤ ∥ φ n ∥ H 0 1 2 k ≤ n ∑ ∫ r t E ℓ ≤ n ∑ ∥ C ℓ Y r , k n ( s ) ∥ H 2 d s ≤ C ∫ r t E ∥Σ ( r , X n ( r )) ∥ L 2 2 d s ≤ C ( t − r ) .
Jensen’s inequality and (27) imply
sup n E [ ( 1 δ ∫ t − δ t ∥ D r F n − b n ( r ) ∥ U 2 d r ) q ] ≤ C ( δ 1 / m + δ q ) . (28) \sup_n\mathbb{E}\left[\left(\frac{1}{\delta}\int_{t-\delta}^t\|D_rF_n-b_n(r)\|_U^2\,dr\right)^q\right]\le C\left(\delta^{1/m}+\delta^q\right). \tag*{(28)} n sup E [ ( δ 1 ∫ t − δ t ∥ D r F n − b n ( r ) ∥ U 2 d r ) q ] ≤ C ( δ 1/ m + δ q ) . (28)
Here 0 < q < 1 0<q<1 0 < q < 1 , so the passage from the separate remainders uses ( a + b ) q ≤ a q + b q (a+b)^q\le a^q+b^q ( a + b ) q ≤ a q + b q for nonnegative a , b a,b a , b , after the squared triangle inequality. For the stochastic term,
E [ ( δ − 1 ∫ ∥ S r n ∥ 2 d r ) q ] ≤ ( δ − 1 ∫ E ∥ S r n ∥ 2 d r ) q ≤ C δ q . \mathbb{E}\left[\left(\delta^{-1}\int\|S_r^n\|^2\,dr\right)^q\right]
\le\left(\delta^{-1}\int\mathbb{E}\|S_r^n\|^2\,dr\right)^q
\le C\delta^q. E [ ( δ − 1 ∫ ∥ S r n ∥ 2 d r ) q ] ≤ ( δ − 1 ∫ E ∥ S r n ∥ 2 d r ) q ≤ C δ q .
Step 4: pass to the weak limit in the appropriate Banach space. Fix δ \delta δ before passing to the limit. Weak convergence of X n X_n X n in L 2 ( Ω × ( 0 , T ) ; H ) L^2(\Omega\times(0,T);H) L 2 ( Ω × ( 0 , T ) ; H ) and affine linearity of Σ \Sigma Σ imply b n ⇀ b b_n \rightharpoonup b b n ⇀ b , where b ( r ) = Σ ( r , X ( r ) ) ∗ φ b(r)=\Sigma(r,X(r))^*\varphi b ( r ) = Σ ( r , X ( r ) ) ∗ φ . The moving projections cause no difficulty: they are uniformly bounded and converge strongly when applied to each fixed dual test. By Lemma 2.10, D F n ⇀ D F DF_n \rightharpoonup DF D F n ⇀ D F in L 2 L^2 L 2 . Because ρ = 2 q > 1 \rho=2q>1 ρ = 2 q > 1 , the space L ρ ( Ω ; L 2 ( t − δ , t ; U ) ) L^\rho(\Omega;L^2(t-\delta,t;\mathcal{U})) L ρ ( Ω ; L 2 ( t − δ , t ; U )) is a Banach space and its norm is weakly lower semicontinuous. The continuous embedding of L 2 L^2 L 2 into this space permits passage to the limit in (28) . Thus the same estimate holds for D F − b DF-b D F − b . This step uses a norm with exponent ρ > 1 \rho>1 ρ > 1 , not convexity of x ↦ x q x\mapsto x^q x ↦ x q . Indeed the left-hand side of (28) is
δ − q ∥ D F n − b n ∥ L ρ ( Ω ; L 2 ( t − δ , t ; U ) ) ρ . \delta^{-q}\lVert DF_n-b_n\rVert_{L^\rho(\Omega;L^2(t-\delta,t;\mathcal{U}))}^{\rho}. δ − q ∥ D F n − b n ∥ L ρ ( Ω ; L 2 ( t − δ , t ; U )) ρ .
The inclusion from the ambient L 2 L^2 L 2 space is continuous because the probability measure has mass one and ρ < 2 \rho<2 ρ < 2 . It is therefore weakly continuous, and the usual norm lower semicontinuity applies.
Step 5: move the center only for the limiting solution. Finally, continuity of g g g and X ∈ L 2 ( Ω ; C ( [ 0 , T ] ; H ) ) X\in L^2(\Omega;C([0,T];H)) X ∈ L 2 ( Ω ; C ([ 0 , T ] ; H )) give
ω t ( δ ) : = 1 δ ∫ t − δ t E ∥ b ( r ) − b ( t ) ∥ U 2 d r ⟶ 0. \omega_t(\delta):=\frac{1}{\delta}\int_{t-\delta}^{t}\mathbb{E}\lVert b(r)-b(t)\rVert_{\mathcal{U}}^2\,\mathrm{d}r\longrightarrow0. ω t ( δ ) := δ 1 ∫ t − δ t E ∥ b ( r ) − b ( t ) ∥ U 2 d r ⟶ 0.
The averaged error with center Θ = b ( t ) \Theta=b(t) Θ = b ( t ) has q q q th moment at most
C ( δ 1 / m + δ q + ω t ( δ ) q ) . C(\delta^{1/m}+\delta^q+\omega_t(\delta)^q). C ( δ 1/ m + δ q + ω t ( δ ) q ) .
This proves (25) . The order of operations is: pass n → ∞ n\to\infty n → ∞ with center b n ( r ) b_n(r) b n ( r ) , replace b ( r ) b(r) b ( r ) by b ( t ) b(t) b ( t ) for the limiting solution, and then let δ ↓ 0 \delta\downarrow0 δ ↓ 0 .
Full support or the zero state
The terminal estimate has now been proved for the coefficient class (21) . The remaining issue is dynamical: multiplicative noise vanishes at the global zero state. Nonlocal vacuum ellipticity distinguishes every nonzero state from that state but does not itself establish a non-hitting theorem.
Theorem 3.7. Assume (21) and Assumption 3.3, and let u 0 ∈ H D u_0\in H_D u 0 ∈ H D be deterministic. For each fixed t > 0 t>0 t > 0 ,
P ( u ( t ) = 0 or supp O u ( t ) = O ‾ ) = 1. (29) \mathbb{P}\bigl(u(t)=0\ \text{or}\ \operatorname{supp}_{\mathcal O}u(t)=\overline{\mathcal O}\bigr)=1. \tag*{(29)} P ( u ( t ) = 0 or supp O u ( t ) = O ) = 1. (29) The two events are disjoint. In particular,
P ( supp O u ( t ) = O ‾ ) = P ( u ( t ) ≠ 0 ) . (30) \mathbb{P}\bigl(\operatorname{supp}_{\mathcal O}u(t)=\overline{\mathcal O}\bigr)=\mathbb{P}(u(t)\ne0). \tag*{(30)} P ( supp O u ( t ) = O ) = P ( u ( t ) = 0 ) . (30) For every nonzero φ ∈ C c ∞ ( O ) \varphi\in C^\infty_c(\mathcal O) φ ∈ C c ∞ ( O ) ,
P ( ⟨ u ( t ) , φ ⟩ = 0 , u ( t ) ≠ 0 ) = 0. \mathbb{P}\bigl(\langle u(t),\varphi\rangle=0,\ u(t)\ne0\bigr)=0. P ( ⟨ u ( t ) , φ ⟩ = 0 , u ( t ) = 0 ) = 0. Proof. Apply Proposition 3.6 with g = 0 g=0 g = 0 and C = B C=B C = B . Markov’s inequality yields (19) with Θ = B ( u ( t ) ) ∗ φ \Theta=B(u(t))^*\varphi Θ = B ( u ( t ) ) ∗ φ . By Assumption 3.3, Θ = 0 \Theta=0 Θ = 0 implies u ( t ) = 0 u(t)=0 u ( t ) = 0 . Lemma 2.12 and locality therefore give
P ( F t , φ = 0 , u ( t ) ≠ 0 ) = 0. \mathbb{P}(F_{t,\varphi}=0,\ u(t)\ne0)=0. P ( F t , φ = 0 , u ( t ) = 0 ) = 0. Apply Lemma 2.14 with E = { u ( t ) ≠ 0 } E=\{u(t)\ne0\} E = { u ( t ) = 0 } . This retains the same survival event for all the countably many tests and proves (29) . Since the zero distribution has empty support, the disjointness and (30) follow.
If we want to get the full-support, we need the following definition.
Definition 3.8 (Backward non-hitting for an initial state). For a deterministic u 0 ≠ 0 u_0 \ne0 u 0 = 0 , the solution has the backward non-hitting property, abbreviated BNH, if
P ( τ 0 = ∞ ) = 1 , τ 0 = inf { s ≥ 0 : u ( s ) = 0 in H D } . (31) \mathbb{P}(\tau_0=\infty)=1,\qquad\tau_0=\inf\{s\ge0:u(s)=0\text{ in }H_D\}. \tag*{(31)} P ( τ 0 = ∞ ) = 1 , τ 0 = inf { s ≥ 0 : u ( s ) = 0 in H D } . (31) The infimum of the empty set is ∞ \infty ∞ . A coefficient family has BNH for a class of initial data if this property holds for every nonzero initial datum in that class.
Continuity in H D H_D H D makes the hitting time in Definition 3.8 well defined. Since zero is absorbing, BNH rules out finite-time extinction from the chosen nonzero initial state. It is stronger than the nonvanishing assertion needed at one particular deterministic time.
Corollary 3.9 (Dirichlet full support under non-hitting). Under the hypotheses of Theorem 3.7 , suppose in addition that u 0 ≠ 0 u_0\ne0 u 0 = 0 and BNH holds in the sense of Definition 3.8 . Then u ( t ) u(t) u ( t ) has full-domain support almost surely at each fixed t > 0 t>0 t > 0 , and every nonzero localized scalar observation has a density. For a conclusion at one prescribed time, the weaker assumption P ( u ( t ) ≠ 0 ) = 1 \mathbb{P}(u(t)\ne0)=1 P ( u ( t ) = 0 ) = 1 suffices.
Proof. Use (30) . For the density assertion, nonlocal vacuum ellipticity makes the terminal direction nonzero almost surely; apply the density part of Lemma 2.12 .
Remark 3.10 (The remaining Dirichlet question). The present estimates do not establish (31) for every family satisfying (21) and nonlocal vacuum ellipticity, including the example in Proposition 3.4 . To see the issue in a logarithmic argument, put R = ∥ u ∥ H 2 R=\lVert u\rVert_H^2 R = ∥ u ∥ H 2 and stop before R R R reaches zero. Itô’s formula gives
d log R = ( − 2 ∥ u ∥ p p R + ∑ k ∥ K k u ∥ H 2 R − 2 ∑ k ( u , K k u ) H 2 R 2 ) d t + 2 R ∑ k ( u , K k u ) H d β k . \mathrm{d}\log R=\left(-\frac{2\lVert u\rVert_p^p}{R}+\frac{\sum_k\lVert K_k u\rVert_H^2}{R}-\frac{2\sum_k(u,K_k u)_H^2}{R^2}\right)\mathrm{d}t+\frac{2}{R}\sum_k(u,K_k u)_H\mathrm{d}\beta_k. d log R = ( − R 2 ∥ u ∥ p p + R ∑ k ∥ K k u ∥ H 2 − R 2 2 ∑ k ( u , K k u ) H 2 ) d t + R 2 k ∑ ( u , K k u ) H d β k .
The noise drift and the quadratic variation rate of the martingale are bounded by (21) . The negative drift − 2 ∥ u ∥ p p / R -2\lVert u\rVert_p^p/R − 2 ∥ u ∥ p p / R , however, is not bounded from below by that condition. The energy estimate controls ∫ ∥ u ∥ p p \int\lVert u\rVert_p^p ∫ ∥ u ∥ p p but not ∫ ∥ u ∥ p p / ∥ u ∥ H 2 \int\lVert u\rVert_p^p/\lVert u\rVert_H^2 ∫ ∥ u ∥ p p / ∥ u ∥ H 2 . It therefore does not exclude log R → − ∞ \log R\to-\infty log R → − ∞ in finite time. This is a missing non-hitting argument, not a counterexample to non-hitting. Corollary 3.9 explicitly keeps that extra input; the unconditional conclusion proved here is Theorem 3.7 .
An invertible flow for the linear noise equation would not alone settle Remark 3.10 either. Such a transformation removes the stochastic differential, but one still needs injectivity or an appropriate backward-uniqueness property for the transformed nonlinear evolution. Forward uniqueness and the absorbing nature of zero do not imply that a nonzero trajectory cannot arrive there.
The mass mechanism below applies to a different boundary problem. The constant function 1 1 1 does not belong to H 0 1 ( O ) H_0^1(\mathcal{O}) H 0 1 ( O ) , so mass is not a continuous functional on H D H_D H D and cannot simply be imported into the Dirichlet argument in Remark 3.10.
Nonlocal multiplicative noise: the Neumann problem
We now construct a conservative model for which the mass explicitly excludes the global zero state. Its Malliavin argument uses only the zero-mean noise coordinates. This avoids imposing a terminal estimate as an additional hypothesis.
The equation has the boundary interpretation
{ d u ( t , x ) = Δ Φ ( u ( t , x ) ) d t + B N ( u ( t , x ) ) d W t , ( t , x ) ∈ ( 0 , T ) × O , ∂ ν Φ ( u ( t , x ) ) = 0 , ( t , x ) ∈ ( 0 , T ) × ∂ O , u ( 0 , x ) = u 0 ( x ) , x ∈ O . (32) \begin{cases}
du(t,x) = \Delta\Phi(u(t,x))\,dt + B_N(u(t,x))\,dW_t, & (t,x) \in(0,T) \times\mathcal{O},\\
\partial_\nu\Phi(u(t,x)) = 0, & (t,x) \in(0,T) \times\partial\mathcal{O},\\
u(0,x) = u_0(x), & x \in\mathcal{O}. \tag*{(32)}
\end{cases} ⎩ ⎨ ⎧ d u ( t , x ) = ΔΦ ( u ( t , x )) d t + B N ( u ( t , x )) d W t , ∂ ν Φ ( u ( t , x )) = 0 , u ( 0 , x ) = u 0 ( x ) , ( t , x ) ∈ ( 0 , T ) × O , ( t , x ) ∈ ( 0 , T ) × ∂ O , x ∈ O . (32)
The Neumann boundary condition removes the diffusion flux through ∂ O \partial\mathcal{O} ∂ O . We specify below a coefficient B N B_N B N whose mass coordinate can be solved explicitly. In contrast with the Dirichlet state space, integration against the constant function is continuous in the Neumann state space we use.
Zero-mean equation and well-posedness
Let
V N 1 = { ψ ∈ H 1 ( O ) : ∫ O ψ d x = 0 } . V_N^1 = \left\{\psi\in H^1(\mathcal{O}) : \int_{\mathcal{O}} \psi\,dx = 0\right\}. V N 1 = { ψ ∈ H 1 ( O ) : ∫ O ψ d x = 0 } .
equipped with the gradient norm, and let H 0 = ( V N 1 ) ∗ H_0 = (V_N^1)^* H 0 = ( V N 1 ) ∗ . An element of H 0 H_0 H 0 acts on H 1 H^1 H 1 by first subtracting the mean, so its mass is zero. With A N , 0 = − Δ N A_{N,0} = -\Delta_N A N , 0 = − Δ N on positive Neumann modes, set
∥ z ∥ H 0 2 = ⟨ z , A N , 0 − 1 z ⟩ . \|z\|_{H_0}^2 = \langle z,A_{N,0}^{-1}z\rangle. ∥ z ∥ H 0 2 = ⟨ z , A N , 0 − 1 z ⟩ .
Use the full state space
H N = R ⊕ H 0 , v = M ( v ) f 0 + v 0 , f 0 = ∣ O ∣ − 1 1 O , ∥ v ∥ H N 2 = ∣ M ( v ) ∣ 2 + ∥ v 0 ∥ H 0 2 . \mathcal{H}_N = \mathbb{R} \oplus H_0,\qquad v = M(v)f_0 + v^0,\qquad f_0 = |\mathcal{O}|^{-1}\mathbf{1}_{\mathcal{O}},\qquad\|v\|_{\mathcal{H}_N}^2 = |M(v)|^2 + \|v^0\|_{H_0}^2. H N = R ⊕ H 0 , v = M ( v ) f 0 + v 0 , f 0 = ∣ O ∣ − 1 1 O , ∥ v ∥ H N 2 = ∣ M ( v ) ∣ 2 + ∥ v 0 ∥ H 0 2 .
Thus M M M is the mass coordinate, equal to ∫ v \int v ∫ v when v v v is a function. The boundary condition is ∂ ν Φ ( u ) = 0 \partial_\nu\Phi(u) = 0 ∂ ν Φ ( u ) = 0 in the Neumann variational sense.
Proposition 4.1 (Monotonicity on a fixed-mass space). If u = M f 0 + z u = Mf_0 + z u = M f 0 + z and v = M f 0 + w v = Mf_0 + w v = M f 0 + w have the same mass, then
⟨ A M ( z ) − A M ( w ) , z − w ⟩ = ∫ O [ Φ ( u ) − Φ ( v ) ] ( u − v ) d x ≥ 0 , \langle A_M(z) - A_M(w), z-w\rangle= \int_{\mathcal{O}} [\Phi(u)-\Phi(v)](u-v)\,dx \ge0, ⟨ A M ( z ) − A M ( w ) , z − w ⟩ = ∫ O [ Φ ( u ) − Φ ( v )] ( u − v ) d x ≥ 0 , where
⟨ A M ( z ) , ψ ⟩ = ∫ O Φ ( z + M / ∣ O ∣ ) ψ d x \langle A_M(z),\psi\rangle= \int_{\mathcal{O}} \Phi(z+M/|\mathcal{O}|)\psi\,dx ⟨ A M ( z ) , ψ ⟩ = ∫ O Φ ( z + M /∣ O ∣ ) ψ d x for zero-mean ψ \psi ψ . For two states with unequal masses the analogous expression need not be nonnegative.
Proof. The difference z − w = u − v z-w = u-v z − w = u − v has zero mass. If ξ \xi ξ is any zero-mean distribution, its squared H 0 H_0 H 0 norm is ⟨ ξ , A N , 0 − 1 ξ ⟩ \langle\xi,A_{N,0}^{-1}\xi\rangle ⟨ ξ , A N , 0 − 1 ξ ⟩ . Hence the drift paired with z − w z-w z − w in this norm becomes the displayed integral. Since Φ \Phi Φ is nondecreasing, the integrand is nonnegative pointwise for equal mass. For unequal masses the zero-mean projection of u − v u-v u − v replaces u − v u-v u − v in the integral. It is not pointwise aligned with Φ ( u ) − Φ ( v ) \Phi(u)-\Phi(v) Φ ( u ) − Φ ( v ) , as the calculation below shows.
For example, take m = 2 m=2 m = 2 on a domain of unit volume, split into two halves, with the constant values u = ( 10 , 2 ) u=(10,2) u = ( 10 , 2 ) and v = ( 9 , 0 ) v=(9,0) v = ( 9 , 0 ) on those halves. Then u − v = ( 1 , 2 ) u-v=(1,2) u − v = ( 1 , 2 ) has mean 3 / 2 3/2 3/2 and zero-mean part ( − 1 / 2 , 1 / 2 ) (-1/2,1/2) ( − 1/2 , 1/2 ) . Since Φ ( r ) = r ∣ r ∣ \Phi(r)=r|r| Φ ( r ) = r ∣ r ∣ , the putative monotonicity integral is
1 2 ( 100 − 81 ) × ( − 1 2 ) + 1 2 ( 4 − 0 ) × 1 2 = − 15 4 . \frac{1}{2}(100-81)\times\left(-\frac{1}{2}\right)+\frac{1}{2}(4-0)\times\frac{1}{2}=-\frac{15}{4}. 2 1 ( 100 − 81 ) × ( − 2 1 ) + 2 1 ( 4 − 0 ) × 2 1 = − 4 15 .
This explains why treating the full mass-plus-zero-mean product as a single monotone H − 1 H^{-1} H − 1 equation would be invalid. We solve the mass first and apply Proposition 4.1 only within its resulting fixed-mass equation.
Choose f i ∈ L 0 2 ( O ) f_i \in L_0^2(\mathcal{O}) f i ∈ L 0 2 ( O ) whose span is dense in L 0 2 L_0^2 L 0 2 , and bounded functionals ℓ j ∈ H 0 ∗ \ell_j \in H_0^* ℓ j ∈ H 0 ∗ that separate points of H 0 H_0 H 0 . For σ ∈ R \sigma\in\mathbb{R} σ ∈ R , b i > 0 b_i > 0 b i > 0 and a i j > 0 a_{ij} > 0 a ij > 0 , define
K 0 v = σ M ( v ) f 0 , K i 0 v = b i M ( v ) f i , K i j v = a i j ℓ j ( v ( 0 ) ) f i ( i , j ≥ 1 ) . (33) K_0v=\sigma M(v)f_0,\qquad K_{i0}v=b_iM(v)f_i,\qquad K_{ij}v=a_{ij}\ell_j(v^{(0)})f_i\quad(i,j\geq1). \tag*{(33)} K 0 v = σ M ( v ) f 0 , K i 0 v = b i M ( v ) f i , K ij v = a ij ℓ j ( v ( 0 ) ) f i ( i , j ≥ 1 ) . (33)
Assume
∑ i b i 2 ∥ f i ∥ H 0 2 < ∞ , ∑ i , j a i j 2 ∥ ℓ j ∥ H 0 ∗ 2 ∥ f i ∥ H 0 2 < ∞ . (34) \sum_i b_i^2\lVert f_i\rVert_{H_0}^2<\infty,\qquad\sum_{i,j}a_{ij}^2\lVert\ell_j\rVert_{H_0^*}^2\lVert f_i\rVert_{H_0}^2<\infty. \tag*{(34)} i ∑ b i 2 ∥ f i ∥ H 0 2 < ∞ , i , j ∑ a ij 2 ∥ ℓ j ∥ H 0 ∗ 2 ∥ f i ∥ H 0 2 < ∞. (34)
All driving Brownian motions β 0 \beta_0 β 0 , ( β i 0 ) (\beta_{i0}) ( β i 0 ) and ( β i j ) (\beta_{ij}) ( β ij ) are independent. Let W ′ W' W ′ denote the cylindrical noise formed by the last two families. The equation is
d u = Δ N Φ ( u ) d t + K 0 u d β 0 + ∑ i K i 0 u d β i 0 + ∑ i , j K i j u d β i j . (35) du=\Delta_N\Phi(u)\,dt+K_0u\,d\beta_0+\sum_i K_{i0}u\,d\beta_{i0}+\sum_{i,j}K_{ij}u\,d\beta_{ij}. \tag*{(35)} d u = Δ N Φ ( u ) d t + K 0 u d β 0 + i ∑ K i 0 u d β i 0 + i , j ∑ K ij u d β ij . (35)
The sums in (34) make the coefficient a bounded linear map from H N \mathcal{H}_N H N to the relevant Hilbert–Schmidt space. Indeed,
∥ B N ( v ) ∥ L 2 2 = σ 2 M ( v ) 2 + M ( v ) 2 ∑ i b i 2 ∥ f i ∥ H 0 2 + ∑ i , j a i j 2 ∣ ℓ j ( v ( 0 ) ) ∣ 2 ∥ f i ∥ H 0 2 , \lVert B_N(v)\rVert_{L_2}^2=\sigma^2M(v)^2+M(v)^2\sum_i b_i^2\lVert f_i\rVert_{H_0}^2+\sum_{i,j}a_{ij}^2\lvert\ell_j(v^{(0)})\rvert^2\lVert f_i\rVert_{H_0}^2, ∥ B N ( v ) ∥ L 2 2 = σ 2 M ( v ) 2 + M ( v ) 2 i ∑ b i 2 ∥ f i ∥ H 0 2 + i , j ∑ a ij 2 ∣ ℓ j ( v ( 0 ) ) ∣ 2 ∥ f i ∥ H 0 2 ,
which is at most a constant times ∥ v ∥ H N 2 \lVert v\rVert_{\mathcal{H}_N}^2 ∥ v ∥ H N 2 . All modes with indices i 0 i0 i 0 or i j ij ij have zero mass; only the mode indexed by 0 0 0 changes the mass.
Proposition 4.2 (A concrete Neumann family). The family (33) satisfies nonlocal vacuum ellipticity on H N \mathcal{H}_N H N against every nonzero φ ∈ C c ∞ ( O ) \varphi\in C^\infty_c(\mathcal{O}) φ ∈ C c ∞ ( O ) . An explicit choice is
f i = h i , ℓ j ( z ) = ( z , h j ) H 0 , b i = 2 − i and a i j = 2 − i − j , f_i=h_i,\qquad\ell_j(z)=(z,h_j)_{H_0},\qquad b_i=2^{-i}\qquad\text{and}\qquad a_{ij}=2^{-i-j}, f i = h i , ℓ j ( z ) = ( z , h j ) H 0 , b i = 2 − i and a ij = 2 − i − j , where h j = λ j ρ j h_j=\sqrt{\lambda_j}\rho_j h j = λ j ρ j for the positive L 2 L^2 L 2 -normalized Neumann eigenfunctions.
Proof. A nonzero compactly supported smooth test is not constant on O \mathcal{O} O . Density of ( f i ) (f_i) ( f i ) therefore supplies an i i i with ⟨ f i , φ ⟩ ≠ 0 \langle f_i,\varphi\rangle\ne0 ⟨ f i , φ ⟩ = 0 . If M ( v ) ≠ 0 M(v)\ne0 M ( v ) = 0 , then
⟨ K i 0 v , φ ⟩ ≠ 0. \langle K_{i0}v,\varphi\rangle\ne0. ⟨ K i 0 v , φ ⟩ = 0.
If M ( v ) = 0 M(v)=0 M ( v ) = 0 and v ≠ 0 v\ne0 v = 0 , choose j j j with ℓ j ( v ( 0 ) ) ≠ 0 \ell_j(v^{(0)})\ne0 ℓ j ( v ( 0 ) ) = 0 , then
⟨ K i j v , φ ⟩ ≠ 0. \langle K_{ij}v,\varphi\rangle\ne0. ⟨ K ij v , φ ⟩ = 0.
This proves nonlocal vacuum ellipticity. For the explicit choice, ( h i ) (h_i) ( h i ) is orthonormal in H 0 H_0 H 0 and its span is dense in L 0 2 L_0^2 L 0 2 . Each functional has norm one, and the two sums in (34) are 1 / 3 1/3 1/3 and 1 / 9 1/9 1/9 , respectively.
For the nonzero-mass class, the K i 0 K_{i0} K i 0 modes alone already make every nonzero interior test visible: their profiles are dense among zero-mean functions, and a nonzero compactly supported test cannot be constant on the connected domain. The K i j K_{ij} K ij modes extend nonlocal vacuum ellipticity to nonzero zero-mass states. This distinction will matter when we describe the scope of the support theorem.
Define
G 0 : ℓ 2 ( N ) → H 0 , G 0 e i = b i f i and C : H 0 → L 2 ( ℓ 2 ( N 2 ) , H 0 ) , ( C z ) e i j = a i j ℓ j ( z ) f i . G_0:\ell^2(\mathbb{N})\to H_0,\qquad G_0e_i=b_if_i\qquad\text{and}\qquad C:H_0\to L_2(\ell^2(\mathbb{N}^2),H_0),\qquad(Cz)e_{ij}=a_{ij}\ell_j(z)f_i. G 0 : ℓ 2 ( N ) → H 0 , G 0 e i = b i f i and C : H 0 → L 2 ( ℓ 2 ( N 2 ) , H 0 ) , ( C z ) e ij = a ij ℓ j ( z ) f i .
In formulas involving W ′ W' W ′ , extend these operators by zero to the other noise coordinates.
Proposition 4.3 (Well-posedness). For every deterministic u 0 = M 0 f 0 + z 0 ∈ H N u_0=M_0f_0+z_0\in\mathcal{H}_N u 0 = M 0 f 0 + z 0 ∈ H N , equation (35) has a unique global variational solution with continuous H N \mathcal{H}_N H N paths. Its mass is
M t = M 0 exp { σ β 0 ( t ) − 1 2 σ 2 t } , u ( t ) = M t f 0 + z ( t ) . (36) M_t=M_0\exp\left\{\sigma\beta_0(t)-\frac{1}{2}\sigma^2t\right\},\qquad u(t)=M_tf_0+z(t). \tag*{(36)} M t = M 0 exp { σ β 0 ( t ) − 2 1 σ 2 t } , u ( t ) = M t f 0 + z ( t ) . (36) For every finite T T T the solution satisfies
E sup s ≤ T ∥ u ( s ) ∥ H N 2 + E ∫ 0 T ∥ u ( s ) ∥ p p d s < ∞ . \mathbb{E}\sup_{s\leq T}\|u(s)\|_{\mathcal{H}_N}^2+\mathbb{E}\int_0^T\|u(s)\|_p^p\,ds<\infty. E s ≤ T sup ∥ u ( s ) ∥ H N 2 + E ∫ 0 T ∥ u ( s ) ∥ p p d s < ∞. If M 0 ≠ 0 M_0\neq0 M 0 = 0 , then u ( t ) ≠ 0 u(t)\neq0 u ( t ) = 0 for every finite t t t , on one event of probability one.
Proof. Testing the equation by the constant 1 1 1 gives
d M t = σ M t d β 0 ( t ) , dM_t=\sigma M_t\,d\beta_0(t), d M t = σ M t d β 0 ( t ) ,
since every other noise mode and the Neumann drift have zero mass. Its unique solution is (36) . Conversely, start with this mass process and put c ( t ) = M t / ∣ O ∣ c(t)=M_t/|\mathcal{O}| c ( t ) = M t /∣ O ∣ . Construct z z z in the fixed triple
L 0 p ( O ) ⊂ H 0 ⊂ ( L 0 p ( O ) ) ∗ L_0^p(\mathcal{O})\subset H_0\subset(L_0^p(\mathcal{O}))^* L 0 p ( O ) ⊂ H 0 ⊂ ( L 0 p ( O ) ) ∗
from
d z = Δ N Φ ( z + c ( t ) ) d t + M t G 0 d W t ′ + C z d W t ′ , z ( 0 ) = z 0 . (37) dz=\Delta_N\Phi(z+c(t))\,dt+M_tG_0\,dW'_t+Cz\,dW'_t,\qquad z(0)=z_0. \tag*{(37)} d z = Δ N Φ ( z + c ( t )) d t + M t G 0 d W t ′ + C z d W t ′ , z ( 0 ) = z 0 . (37)
Here c ( t ) c(t) c ( t ) is a spatial constant and the two displayed stochastic terms act in disjoint coordinates of W ′ W' W ′ . For any fixed real c c c , define A c A_c A c on L 0 p L_0^p L 0 p by
⟨ A c ( z ) , w ⟩ = ∫ O Φ ( z + c ) w d x . \langle A_c(z),w\rangle=\int_{\mathcal{O}}\Phi(z+c)w\,dx. ⟨ A c ( z ) , w ⟩ = ∫ O Φ ( z + c ) w d x .
For z , w ∈ L 0 p z,w\in L_0^p z , w ∈ L 0 p ,
⟨ A c ( z ) − A c ( w ) , z − w ⟩ = ∫ O [ Φ ( z + c ) − Φ ( w + c ) ] ( z − w ) d x ≥ 0 , \langle A_c(z)-A_c(w),z-w\rangle=\int_{\mathcal{O}}[\Phi(z+c)-\Phi(w+c)](z-w)\,dx\geq0, ⟨ A c ( z ) − A c ( w ) , z − w ⟩ = ∫ O [ Φ ( z + c ) − Φ ( w + c )] ( z − w ) d x ≥ 0 ,
⟨ A c ( z ) , z ⟩ = ∫ O [ Φ ( z + c ) − Φ ( c ) ] z d x ≥ 2 1 − m ∥ z ∥ p p . (38) \langle A_c(z),z\rangle=\int_{\mathcal{O}}[\Phi(z+c)-\Phi(c)]z\,dx\geq2^{1-m}\|z\|_p^p. \tag*{(38)} ⟨ A c ( z ) , z ⟩ = ∫ O [ Φ ( z + c ) − Φ ( c )] z d x ≥ 2 1 − m ∥ z ∥ p p . (38)
The last inequality is the scalar uniform monotonicity of r ↦ ∣ r ∣ p − 2 r r\mapsto|r|^{p-2}r r ↦ ∣ r ∣ p − 2 r . Also
∥ A c ( z ) ∥ ( L 0 p ) ∗ p / ( p − 1 ) ≤ C ( ∥ z ∥ p p + ∣ c ∣ p ) . \|A_c(z)\|_{(L_0^p)^*}^{p/(p-1)}\leq C(\|z\|_p^p+|c|^p). ∥ A c ( z ) ∥ ( L 0 p ) ∗ p / ( p − 1 ) ≤ C ( ∥ z ∥ p p + ∣ c ∣ p ) .
Since p = m + 1 ≥ 2 p=m+1\geq2 p = m + 1 ≥ 2 , the standard uniform monotonicity inequality
( ∣ a ∣ p − 2 a − ∣ b ∣ p − 2 b ) ( a − b ) ≥ 2 2 − p ∣ a − b ∣ p , a , b ∈ R , (|a|^{p-2}a-|b|^{p-2}b)(a-b)\geq2^{2-p}|a-b|^p,\qquad a,b\in\mathbb{R}, ( ∣ a ∣ p − 2 a − ∣ b ∣ p − 2 b ) ( a − b ) ≥ 2 2 − p ∣ a − b ∣ p , a , b ∈ R ,
applied with a = z ( x ) + c a=z(x)+c a = z ( x ) + c and b = c b=c b = c , gives pointwise
[ Φ ( z + c ) − Φ ( c ) ] z ≥ 2 2 − p ∣ z ∣ p = 2 1 − m ∣ z ∣ p . [\Phi(z+c)-\Phi(c)]z\geq2^{2-p}|z|^p=2^{1-m}|z|^p. [ Φ ( z + c ) − Φ ( c )] z ≥ 2 2 − p ∣ z ∣ p = 2 1 − m ∣ z ∣ p .
Since z z z has zero mean, ∫ O Φ ( c ) z d x = 0 \int_{\mathcal{O}}\Phi(c)z\,dx=0 ∫ O Φ ( c ) z d x = 0 , and integration proves the second line of (38) . The estimate is valid for every real c c c and every signed z z z .
The growth estimate follows from the embedding of the zero-mean dual into the full L p ′ L^{p'} L p ′ dual:
∥ A c ( z ) ∥ ( L 0 p ) ∗ ≤ ∥ Φ ( z + c ) ∥ p ′ = ∥ z + c ∥ p p − 1 , \|A_c(z)\|_{(L_0^p)^*}\leq\|\Phi(z+c)\|_{p'}=\|z+c\|_p^{p-1}, ∥ A c ( z ) ∥ ( L 0 p ) ∗ ≤ ∥Φ ( z + c ) ∥ p ′ = ∥ z + c ∥ p p − 1 ,
∥ A c ( z ) ∥ ( L 0 p ) ∗ p ′ ≤ ∥ z + c ∥ p p ≤ 2 p − 1 ( ∥ z ∥ p p + ∣ O ∣ ∣ c ∣ p ) , p ′ = p p − 1 . \|A_c(z)\|_{(L_0^p)^*}^{p'}\leq\|z+c\|_p^p\leq2^{p-1}(\|z\|_p^p+|\mathcal{O}||c|^p),\qquad p'=\frac{p}{p-1}. ∥ A c ( z ) ∥ ( L 0 p ) ∗ p ′ ≤ ∥ z + c ∥ p p ≤ 2 p − 1 ( ∥ z ∥ p p + ∣ O ∣∣ c ∣ p ) , p ′ = p − 1 p .
Thus the claimed bound holds after absorbing the fixed factor ∣ O ∣ |\mathcal{O}| ∣ O ∣ into the constant.
The mass process is adapted and has moments of every order on finite intervals; its paths are bounded on each compact interval. Consequently c ( t ) c(t) c ( t ) is an adapted time-dependent parameter and the growth bound is integrable. The map z ↦ M t G 0 + C z z \mapsto M_tG_0+Cz z ↦ M t G 0 + C z is affine with bounded linear part on the zero-mean space. These are the variational hypotheses of [13 ] , and so an adapted solution of (37) is obtained. This argument uses monotonicity only within the prescribed-mass equation.
Next, we consider energy, reconstruction, and uniqueness . Let
Σ M ′ ( s , z ) = M s G 0 + C z . \Sigma'_M(s,z)=M_sG_0+Cz. Σ M ′ ( s , z ) = M s G 0 + C z .
Since the ( i 0 ) (i0) ( i 0 ) and ( i j ) (ij) ( ij ) channels are orthogonal in the noise space, or simply by ( a + b ) 2 ≤ 2 a 2 + 2 b 2 (a+b)^2\leq2a^2+2b^2 ( a + b ) 2 ≤ 2 a 2 + 2 b 2 ,
∥ Σ M ′ ( s , z ) ∥ L 2 ( U ′ , H 0 ) 2 ≤ 2 ∣ M s ∣ 2 ∥ G 0 ∥ L 2 2 + 2 ∥ C ∥ L ( H 0 , L 2 ) 2 ∥ z ∥ H 0 2 . \|\Sigma'_M(s,z)\|_{\mathcal{L}_2(U',H_0)}^2\leq2|M_s|^2\|G_0\|_{\mathcal{L}_2}^2+2\|C\|_{\mathcal{L}(H_0,L^2)}^2\|z\|_{H_0}^2. ∥ Σ M ′ ( s , z ) ∥ L 2 ( U ′ , H 0 ) 2 ≤ 2∣ M s ∣ 2 ∥ G 0 ∥ L 2 2 + 2∥ C ∥ L ( H 0 , L 2 ) 2 ∥ z ∥ H 0 2 .
Applying Itô’s formula to ∥ z ∥ H 0 2 \|z\|_{H_0}^2 ∥ z ∥ H 0 2 and using (38) gives
∥ z ( t ) ∥ H 0 2 + 2 2 − m ∫ 0 t ∥ z ( s ) ∥ p p d s ≤ ∥ z 0 ∥ H 0 2 + ∫ 0 t ∥ Σ M ′ ( s , z ( s ) ) ∥ L 2 2 d s + 2 ∫ 0 t ( z ( s ) , Σ M ′ ( s , z ( s ) ) d W s ′ ) H 0 . \|z(t)\|_{H_0}^2+2^{2-m}\int_0^t\|z(s)\|_p^p\,\mathrm{d}s\leq\|z_0\|_{H_0}^2+\int_0^t\|\Sigma'_M(s,z(s))\|_{\mathcal{L}_2}^2\,\mathrm{d}s
+2\int_0^t(z(s),\Sigma'_M(s,z(s))\,\mathrm{d}W'_s)_{H_0}. ∥ z ( t ) ∥ H 0 2 + 2 2 − m ∫ 0 t ∥ z ( s ) ∥ p p d s ≤ ∥ z 0 ∥ H 0 2 + ∫ 0 t ∥ Σ M ′ ( s , z ( s )) ∥ L 2 2 d s + 2 ∫ 0 t ( z ( s ) , Σ M ′ ( s , z ( s )) d W s ′ ) H 0 .
Define the stopping time τ R = inf { s : ∥ z ( s ) ∥ H 0 ≥ R } ∧ T \tau_R=\inf\{s:\|z(s)\|_{H_0}\geq R\}\wedge T τ R = inf { s : ∥ z ( s ) ∥ H 0 ≥ R } ∧ T . The BDG and Young inequalities imply
E sup θ ≤ t ∧ τ R ∣ 2 ∫ 0 θ ( z , Σ M ′ ( s , z ) d W s ′ ) H 0 ∣ ≤ 1 2 E sup θ ≤ t ∧ τ R ∥ z ( θ ) ∥ H 0 2 + C E ∫ 0 t ∧ τ R ( ∣ M s ∣ 2 + ∥ z ( s ) ∥ H 0 2 ) d s . \mathbb{E}\sup_{\theta\leq t\wedge\tau_R}\left|2\int_0^\theta(z,\Sigma'_M(s,z)\,\mathrm{d}W'_s)_{H_0}\right|
\leq\frac{1}{2}\mathbb{E}\sup_{\theta\leq t\wedge\tau_R}\|z(\theta)\|_{H_0}^2
+C\mathbb{E}\int_0^{t\wedge\tau_R}(|M_s|^2+\|z(s)\|_{H_0}^2)\,\mathrm{d}s. E θ ≤ t ∧ τ R sup 2 ∫ 0 θ ( z , Σ M ′ ( s , z ) d W s ′ ) H 0 ≤ 2 1 E θ ≤ t ∧ τ R sup ∥ z ( θ ) ∥ H 0 2 + C E ∫ 0 t ∧ τ R ( ∣ M s ∣ 2 + ∥ z ( s ) ∥ H 0 2 ) d s .
Therefore Gronwall’s inequality gives, with a constant independent of R R R ,
E sup s ≤ T ∧ τ R ∥ z ( s ) ∥ H 0 2 + E ∫ 0 T ∧ τ R ∥ z ( s ) ∥ p p d s ≤ C T ( ∥ z 0 ∥ H 0 2 + E ∫ 0 T ∣ M s ∣ 2 d s ) . \mathbb{E}\sup_{s\leq T\wedge\tau_R}\|z(s)\|_{H_0}^2+\mathbb{E}\int_0^{T\wedge\tau_R}\|z(s)\|_p^p\,\mathrm{d}s
\leq C_T\left(\|z_0\|_{H_0}^2+\mathbb{E}\int_0^T|M_s|^2\,\mathrm{d}s\right). E s ≤ T ∧ τ R sup ∥ z ( s ) ∥ H 0 2 + E ∫ 0 T ∧ τ R ∥ z ( s ) ∥ p p d s ≤ C T ( ∥ z 0 ∥ H 0 2 + E ∫ 0 T ∣ M s ∣ 2 d s ) .
Letting R ↑ ∞ R\uparrow\infty R ↑ ∞ proves the same estimate without stopping. Finally,
∥ u ( s ) ∥ p p = ∥ z ( s ) + M s / ∣ O ∣ ∥ p p ≤ 2 p − 1 ( ∥ z ( s ) ∥ p p + ∣ O ∣ 1 − p ∣ M s ∣ p ) , \|u(s)\|_p^p=\|z(s)+M_s/|\mathcal{O}|\|_p^p\leq2^{p-1}\left(\|z(s)\|_p^p+|\mathcal{O}|^{1-p}|M_s|^p\right), ∥ u ( s ) ∥ p p = ∥ z ( s ) + M s /∣ O ∣ ∥ p p ≤ 2 p − 1 ( ∥ z ( s ) ∥ p p + ∣ O ∣ 1 − p ∣ M s ∣ p ) ,
and the geometric Brownian motion M M M has moments of every positive order on finite intervals. This yields the asserted energy bound for u u u . Uniqueness follows first for the scalar mass equation and then for the zero-mean equation with the same prescribed mass, where (38) is used again. The two coordinates reconstruct a solution of (35) with the boundary condition in (32) . If M 0 ≠ 0 M_0\neq0 M 0 = 0 , the explicit exponential in (36) never vanishes. Since u ( t ) = 0 u(t)=0 u ( t ) = 0 would force M t = 0 M_t=0 M t = 0 , non-hitting follows.
Theorem 4.4 (Neumann non-hitting from nonzero initial mass). If M ( u 0 ) ≠ 0 M(u_0)\neq0 M ( u 0 ) = 0 , the solution in Proposition 4.3 satisfies P ( u ( t ) ≠ 0 for every finite t ≥ 0 ) = 1 \mathbb{P}(u(t)\neq0\text{ for every finite }t\geq0)=1 P ( u ( t ) = 0 for every finite t ≥ 0 ) = 1 .
Proof. For Brownian paths in the usual full-probability continuity event, exp { σ β 0 ( t ) − σ 2 t / 2 } \exp\{\sigma\beta_0(t)-\sigma^2t/2\} exp { σ β 0 ( t ) − σ 2 t /2 } is strictly positive for every finite t t t . Formula (36) gives M t ≠ 0 M_t\neq0 M t = 0 at all such times. The zero state has mass zero, so it cannot occur.
Theorem 4.4 covers precisely the class of nonzero-mass initial states. The mass equation starts at zero and stays there if M ( u 0 ) = 0 M(u_0)=0 M ( u 0 ) = 0 . In that class, the presence of K i j K_{ij} K ij modes and their nonlocal vacuum ellipticity property do not by themselves provide non-hitting.
Partial Malliavin calculus with the mass fixed
The full Malliavin derivative with respect to β 0 \beta_0 β 0 changes the mass and creates an indefinite cross term in the zero-mean energy. The support argument does not need this direction. We therefore make the conditional probability space explicit and differentiate only in the independent zero-mean noise.
We introduce the Product Wiener space. Because β 0 \beta_0 β 0 and W ′ W' W ′ are independent, we may define the driving probability space as
( Ω , F , P ) = ( Ω 0 × Ω ′ , F 0 ⊗ F ′ , P 0 ⊗ P ′ ) . (\Omega,\mathcal{F},\mathbb{P})=(\Omega_0\times\Omega',\mathcal{F}_0\otimes\mathcal{F}',\mathbb{P}_0\otimes\mathbb{P}'). ( Ω , F , P ) = ( Ω 0 × Ω ′ , F 0 ⊗ F ′ , P 0 ⊗ P ′ ) .
where β 0 = β 0 ( ω 0 ) \beta_0=\beta_0(\omega_0) β 0 = β 0 ( ω 0 ) and W ′ = W ′ ( ω ′ ) W'=W'(\omega') W ′ = W ′ ( ω ′ ) . Since the initial datum and coefficients are deterministic and Proposition 4.3 gives pathwise uniqueness, passing to this canonical product realization does not change the law of the solution, see [5 ] . Let U ′ \mathcal{U}' U ′ be the noise space generated by the ( i 0 ) (i0) ( i 0 ) and ( i j ) (ij) ( ij ) modes, let H ′ = L 2 ( 0 , T ; U ′ ) \mathcal{H}'=L^2(0,T;\mathcal{U}') H ′ = L 2 ( 0 , T ; U ′ ) , and denote by D ′ D' D ′ the Malliavin derivative on the Wiener space ( Ω ′ , F ′ , P ′ ) (\Omega',\mathcal{F}',\mathbb{P}') ( Ω ′ , F ′ , P ′ ) . For P 0 \mathbb{P}_0 P 0 -almost every ω 0 \omega_0 ω 0 ,
M ω 0 ( s ) = M 0 exp { σ β 0 ( s , ω 0 ) − 1 2 σ 2 s } M^{\omega_0}(s)=M_0\exp\left\{\sigma\beta_0(s,\omega_0)-\frac{1}{2}\sigma^2s\right\} M ω 0 ( s ) = M 0 exp { σ β 0 ( s , ω 0 ) − 2 1 σ 2 s }
is a bounded continuous deterministic function of s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] . With c ω 0 = M ω 0 / ∣ O ∣ c^{\omega_0}=M^{\omega_0}/|\mathcal{O}| c ω 0 = M ω 0 /∣ O ∣ , equation (37) is then a variational SPDE on Ω ′ \Omega' Ω ′ with deterministic time-dependent mass parameter. We write E ′ \mathbb{E}' E ′ for expectation on Ω ′ \Omega' Ω ′ . All conditional Malliavin assertions below mean statements on this second Wiener space for P 0 \mathbb{P}_0 P 0 -almost every fixed ω 0 \omega_0 ω 0 .
The reason for using D ′ D' D ′ can also be seen from the energy. If a derivative Y Y Y has nonzero mass, the positive-mode H 0 H_0 H 0 pairing of the linearized drift with its zero-mean part contains
− ∫ O a Y 2 d x + Y ‾ ∫ O a Y d x , a = m ∣ u ∣ m − 1 , -\int_{\mathcal{O}}aY^2\,dx+\overline{Y}\int_{\mathcal{O}}aY\,dx,\qquad a=m|u|^{m-1}, − ∫ O a Y 2 d x + Y ∫ O aY d x , a = m ∣ u ∣ m − 1 ,
and the second term has no sign. In contrast, D ′ M = 0 D'M=0 D ′ M = 0 and every D ′ z D'z D ′ z has zero mass, so the cross term vanishes.
Proposition 4.5 (Conditional terminal estimate). There exists a set Ω 0 ∗ ∈ F 0 \Omega_0^*\in\mathcal{F}_0 Ω 0 ∗ ∈ F 0 with P 0 ( Ω 0 ∗ ) = 1 \mathbb{P}_0(\Omega_0^*)=1 P 0 ( Ω 0 ∗ ) = 1 such that the following holds. Fix ω 0 ∈ Ω 0 ∗ \omega_0\in\Omega_0^* ω 0 ∈ Ω 0 ∗ , t > 0 t>0 t > 0 , and φ ∈ C c ∞ ( O ) \varphi\in C_c^\infty(\mathcal{O}) φ ∈ C c ∞ ( O ) . Let
F t , φ ω 0 = ⟨ u ω 0 ( t ) , φ ⟩ . F_{t,\varphi}^{\omega_0}=\langle u^{\omega_0}(t),\varphi\rangle. F t , φ ω 0 = ⟨ u ω 0 ( t ) , φ ⟩ . Then F t , φ ω 0 ∈ D 1 , 2 ( Ω ′ ) F_{t,\varphi}^{\omega_0}\in\mathbb{D}^{1,2}(\Omega') F t , φ ω 0 ∈ D 1 , 2 ( Ω ′ ) and
E ′ [ ( 1 δ ∫ t − δ t ∥ D ′ F t , φ ω 0 − Θ t , φ ′ , ω 0 ∥ U ′ 2 d r ) q ] ⟶ 0 ( δ ↓ 0 ) , (39) \mathbb{E}'\left[\left(\frac{1}{\delta}\int_{t-\delta}^{t}\left\|D'F_{t,\varphi}^{\omega_0}-\Theta_{t,\varphi}^{\prime,\omega_0}\right\|_{\mathcal{U}'}^2\,dr\right)^q\right]\longrightarrow0\qquad(\delta\downarrow0), \tag*{(39)} E ′ [ ( δ 1 ∫ t − δ t D ′ F t , φ ω 0 − Θ t , φ ′ , ω 0 U ′ 2 d r ) q ] ⟶ 0 ( δ ↓ 0 ) , (39) where q = ( m + 1 ) / ( 2 m ) q=(m+1)/(2m) q = ( m + 1 ) / ( 2 m ) as in (3.4) and
Θ t , φ ′ , ω 0 = ( ( M t ω 0 b i ⟨ f i , φ ⟩ ) i , ( a i j ℓ j ( z ω 0 ( t ) ) ⟨ f i , φ ⟩ ) i , j ) . (40) \Theta_{t,\varphi}^{\prime,\omega_0}
=\left(\left(M_t^{\omega_0}b_i\langle f_i,\varphi\rangle\right)_i,\left(a_{ij}\ell_j\left(z^{\omega_0}(t)\right)\langle f_i,\varphi\rangle\right)_{i,j}\right). \tag*{(40)} Θ t , φ ′ , ω 0 = ( ( M t ω 0 b i ⟨ f i , φ ⟩ ) i , ( a ij ℓ j ( z ω 0 ( t ) ) ⟨ f i , φ ⟩ ) i , j ) . (40) Equivalently, (39) is the conditional terminal estimate given the mass Brownian path, but its formulation above does not rely on choosing a version of a regular conditional probability.
Proof. Fix ω 0 ∈ Ω 0 ∗ \omega_0\in\Omega_0^* ω 0 ∈ Ω 0 ∗ and suppress the superscript ω 0 \omega_0 ω 0 . Thus M M M and c = M / ∣ O ∣ c=M/|\mathcal{O}| c = M /∣ O ∣ are deterministic, continuous and bounded on [ 0 , T ] [0,T] [ 0 , T ] . The proof is divided into five steps.
Step 1: conditional state estimates and Galerkin convergence. Let P n N P_n^N P n N be the projection onto the first n n n positive Neumann modes and truncate W ′ W' W ′ to its first n n n coordinates in a fixed enumeration. Denote the resulting zero-mean solution by z n z_n z n , put u n = z n + c u_n=z_n+c u n = z n + c , and write
Σ M ′ ( s , z ) = M s G 0 + C z . \Sigma'_M(s,z)=M_sG_0+Cz. Σ M ′ ( s , z ) = M s G 0 + C z .
where G 0 G_0 G 0 and C C C are extended by zero to the complementary coordinates of U ′ U' U ′ . For the fixed path ω 0 \omega_0 ω 0 , set
K M ( ω 0 ) = sup 0 ≤ s ≤ T ∣ M s ω 0 ∣ 2 < ∞ . K_M(\omega_0)=\sup_{0\leq s\leq T}|M_s^{\omega_0}|^2<\infty. K M ( ω 0 ) = 0 ≤ s ≤ T sup ∣ M s ω 0 ∣ 2 < ∞.
The Galerkin equation has the form
d z n + A c ( s ) , n ( z n ) d s = P n N Σ M ′ ( s , z n ) Π n d W s ′ . \mathrm{d}z_n+A_{c(s),n}(z_n)\,\mathrm{d}s=P_n^N\Sigma'_M(s,z_n)\Pi_n\,\mathrm{d}W'_s. d z n + A c ( s ) , n ( z n ) d s = P n N Σ M ′ ( s , z n ) Π n d W s ′ .
By using (38) , contractivity of P n N P_n^N P n N , Π n \Pi_n Π n , and boundedness of G 0 , C G_0,C G 0 , C , we get
2 ⟨ A c ( s ) , n ( z n ) , z n ⟩ ≥ 2 2 − m ∥ z n ∥ p p , 2\langle A_{c(s),n}(z_n),z_n\rangle\geq2^{2-m}\|z_n\|_p^p, 2 ⟨ A c ( s ) , n ( z n ) , z n ⟩ ≥ 2 2 − m ∥ z n ∥ p p ,
∥ P n N Σ M ′ ( s , z n ) Π n ∥ L 2 2 ≤ C 0 K M ( ω 0 ) + C 1 ∥ z n ∥ H 0 2 , \|P_n^N\Sigma'_M(s,z_n)\Pi_n\|_{L_2}^2\leq C_0K_M(\omega_0)+C_1\|z_n\|_{H_0}^2, ∥ P n N Σ M ′ ( s , z n ) Π n ∥ L 2 2 ≤ C 0 K M ( ω 0 ) + C 1 ∥ z n ∥ H 0 2 ,
where C 0 , C 1 C_0,C_1 C 0 , C 1 are deterministic and independent of n n n . Consequently, Itô’s formula gives
∥ z n ( t ) ∥ H 0 2 + 2 2 − m ∫ 0 t ∥ z n ( s ) ∥ p p d s ≤ ∥ P n N z 0 ∥ H 0 2 + C 0 t K M ( ω 0 ) + C 1 ∫ 0 t ∥ z n ( s ) ∥ H 0 2 d s + 2 ∫ 0 t ( z n , P n N Σ M ′ ( s , z n ) Π n d W s ′ ) H 0 . \|z_n(t)\|_{H_0}^2+2^{2-m}\int_0^t\|z_n(s)\|_p^p\,\mathrm{d}s\leq\|P_n^Nz_0\|_{H_0}^2+C_0tK_M(\omega_0)+C_1\int_0^t\|z_n(s)\|_{H_0}^2\,\mathrm{d}s
+2\int_0^t(z_n,P_n^N\Sigma'_M(s,z_n)\Pi_n\,\mathrm{d}W'_s)_{H_0}. ∥ z n ( t ) ∥ H 0 2 + 2 2 − m ∫ 0 t ∥ z n ( s ) ∥ p p d s ≤ ∥ P n N z 0 ∥ H 0 2 + C 0 t K M ( ω 0 ) + C 1 ∫ 0 t ∥ z n ( s ) ∥ H 0 2 d s + 2 ∫ 0 t ( z n , P n N Σ M ′ ( s , z n ) Π n d W s ′ ) H 0 .
Let
τ R n = inf { s : ∥ z n ( s ) ∥ H 0 ≥ R } ∧ T . \tau_R^n=\inf\{s:\|z_n(s)\|_{H_0}\geq R\}\wedge T. τ R n = inf { s : ∥ z n ( s ) ∥ H 0 ≥ R } ∧ T .
BDG and Young’s inequality give
E ′ sup θ ≤ t ∧ τ R n ∣ 2 ∫ 0 θ ( z n , P n N Σ M ′ ( s , z n ) Π n d W s ′ ) H 0 ∣ ≤ 1 2 E ′ sup θ ≤ t ∧ τ R n ∥ z n ( θ ) ∥ H 0 2 + C E ′ ∫ 0 t ∧ τ R n ( K M ( ω 0 ) + ∥ z n ( s ) ∥ H 0 2 ) d s . \mathbb{E}'\sup_{\theta\leq t\wedge\tau_R^n}\left|2\int_0^\theta(z_n,P_n^N\Sigma'_M(s,z_n)\Pi_n\,\mathrm{d}W'_s)_{H_0}\right|
\leq\frac{1}{2}\mathbb{E}'\sup_{\theta\leq t\wedge\tau_R^n}\|z_n(\theta)\|_{H_0}^2+C\mathbb{E}'\int_0^{t\wedge\tau_R^n}(K_M(\omega_0)+\|z_n(s)\|_{H_0}^2)\,\mathrm{d}s. E ′ θ ≤ t ∧ τ R n sup 2 ∫ 0 θ ( z n , P n N Σ M ′ ( s , z n ) Π n d W s ′ ) H 0 ≤ 2 1 E ′ θ ≤ t ∧ τ R n sup ∥ z n ( θ ) ∥ H 0 2 + C E ′ ∫ 0 t ∧ τ R n ( K M ( ω 0 ) + ∥ z n ( s ) ∥ H 0 2 ) d s .
After absorbing the first term and applying Gronwall inequality, we have
E ′ sup s ≤ T ∧ τ R n ∥ z n ( s ) ∥ H 0 2 + E ′ ∫ 0 T ∧ τ R n ∥ z n ( s ) ∥ p p d s ≤ C T ( ∥ z 0 ∥ H 0 2 + T K M ( ω 0 ) ) . \mathbb{E}'\sup_{s\leq T\wedge\tau_R^n}\|z_n(s)\|_{H_0}^2+\mathbb{E}'\int_0^{T\wedge\tau_R^n}\|z_n(s)\|_p^p\,\mathrm{d}s\leq C_T(\|z_0\|_{H_0}^2+TK_M(\omega_0)). E ′ s ≤ T ∧ τ R n sup ∥ z n ( s ) ∥ H 0 2 + E ′ ∫ 0 T ∧ τ R n ∥ z n ( s ) ∥ p p d s ≤ C T ( ∥ z 0 ∥ H 0 2 + T K M ( ω 0 )) .
The right-hand side is independent of n n n and R R R . Fatou’s lemma and monotone convergence as R ↑ ∞ R\uparrow\infty R ↑ ∞ prove
sup n E ′ [ sup s ≤ T ∥ z n ( s ) ∥ H 0 2 + ∫ 0 T ∥ z n ( s ) ∥ p p d s ] ≤ C T ( ω 0 ) < ∞ . (41) \sup_n\mathbb{E}'\left[\sup_{s\leq T}\|z_n(s)\|_{H_0}^2+\int_0^T\|z_n(s)\|_p^p\,\mathrm{d}s\right]\leq C_T(\omega_0)<\infty. \tag*{(41)} n sup E ′ [ s ≤ T sup ∥ z n ( s ) ∥ H 0 2 + ∫ 0 T ∥ z n ( s ) ∥ p p d s ] ≤ C T ( ω 0 ) < ∞. (41)
The quantity required in the terminal-strip estimate is the L p L^p L p norm of u n = z n + c u_n=z_n+c u n = z n + c . The elementary convexity inequality ∣ a + b ∣ p ≤ 2 p − 1 ( ∣ a ∣ p + ∣ b ∣ p ) |a+b|^p\leq2^{p-1}(|a|^p+|b|^p) ∣ a + b ∣ p ≤ 2 p − 1 ( ∣ a ∣ p + ∣ b ∣ p ) gives pointwise
∣ z n ( s , x ) + c ( s ) ∣ p ≤ 2 p − 1 ( ∣ z n ( s , x ) ∣ p + ∣ c ( s ) ∣ p ) . |z_n(s,x)+c(s)|^p\leq2^{p-1}(|z_n(s,x)|^p+|c(s)|^p). ∣ z n ( s , x ) + c ( s ) ∣ p ≤ 2 p − 1 ( ∣ z n ( s , x ) ∣ p + ∣ c ( s ) ∣ p ) .
Integrating over O \mathcal{O} O yields
∥ z n ( s ) + c ( s ) ∥ p p ≤ 2 p − 1 ( ∥ z n ( s ) ∥ p p + ∣ O ∣ ∣ c ( s ) ∣ p ) . (42) \|z_n(s)+c(s)\|_p^p\leq2^{p-1}(\|z_n(s)\|_p^p+|\mathcal{O}||c(s)|^p). \tag*{(42)} ∥ z n ( s ) + c ( s ) ∥ p p ≤ 2 p − 1 ( ∥ z n ( s ) ∥ p p + ∣ O ∣∣ c ( s ) ∣ p ) . (42)
Since the fixed continuous mass path satisfies sup s ≤ T ∣ c ( s ) ∣ < ∞ \sup_{s\leq T}|c(s)|<\infty sup s ≤ T ∣ c ( s ) ∣ < ∞ , integrating this inequality in time and using (41) gives
sup n E ′ ∫ 0 T ∥ u n ( s ) ∥ p p d s ≤ 2 p − 1 [ sup n E ′ ∫ 0 T ∥ z n ( s ) ∥ p p d s + T ∣ O ∣ sup s ≤ T ∣ c ( s ) ∣ p ] < ∞ . \sup_n \mathbb{E}'\int_0^T \|u_n(s)\|_p^p\,\mathrm{d}s \leq2^{p-1}\left[\sup_n \mathbb{E}'\int_0^T \|z_n(s)\|_p^p\,\mathrm{d}s+T|\mathcal{O}|\sup_{s\leq T}|c(s)|^p\right]<\infty. n sup E ′ ∫ 0 T ∥ u n ( s ) ∥ p p d s ≤ 2 p − 1 [ n sup E ′ ∫ 0 T ∥ z n ( s ) ∥ p p d s + T ∣ O ∣ s ≤ T sup ∣ c ( s ) ∣ p ] < ∞.
Moreover, for every fixed r ≥ 2 r\geq2 r ≥ 2 , applying Itô’s formula to ( 1 + ∥ z n ∥ H 0 2 ) r / 2 (1+\|z_n\|_{H_0}^2)^{r/2} ( 1 + ∥ z n ∥ H 0 2 ) r /2 yields
sup n E ′ sup s ≤ T ( 1 + ∥ z n ( s ) ∥ H 0 2 ) r / 2 ≤ C r , T ( ω 0 ) < ∞ , (43) \sup_n\mathbb{E}'\sup_{s\leq T}(1+\|z_n(s)\|_{H_0}^2)^{r/2}\leq C_{r,T}(\omega_0)<\infty, \tag*{(43)} n sup E ′ s ≤ T sup ( 1 + ∥ z n ( s ) ∥ H 0 2 ) r /2 ≤ C r , T ( ω 0 ) < ∞ , (43)
Indeed, the first derivative of Ψ r ( z ) = ( 1 + ∥ z ∥ H 0 2 ) r / 2 \Psi_r(z)=(1+\|z\|_{H_0}^2)^{r/2} Ψ r ( z ) = ( 1 + ∥ z ∥ H 0 2 ) r /2 is a multiple of ( 1 + ∥ z ∥ 2 ) r / 2 − 1 z (1+\|z\|^2)^{r/2-1}z ( 1 + ∥ z ∥ 2 ) r /2 − 1 z , and its Hessian is bounded by C r ( 1 + ∥ z ∥ 2 ) r / 2 − 1 C_r(1+\|z\|^2)^{r/2-1} C r ( 1 + ∥ z ∥ 2 ) r /2 − 1 in operator norm. The drift term is nonpositive because ⟨ A c ( s ) , n ( z n ) , z n ⟩ ≥ 0 \langle A_{c(s),n}(z_n),z_n\rangle\geq0 ⟨ A c ( s ) , n ( z n ) , z n ⟩ ≥ 0 . Furthermore,
∥ Σ M ′ ( s , z n ) ∥ L 2 2 ≤ C ( K M ( ω 0 ) + ∥ z n ∥ H 0 2 ) . \|\Sigma'_M(s,z_n)\|_{\mathcal{L}_2}^2\leq C(K_M(\omega_0)+\|z_n\|_{H_0}^2). ∥ Σ M ′ ( s , z n ) ∥ L 2 2 ≤ C ( K M ( ω 0 ) + ∥ z n ∥ H 0 2 ) .
Consequently both Itô correction terms are bounded by C r , ω 0 ( 1 + Ψ r ( z n ) ) C_{r,\omega_0}(1+\Psi_r(z_n)) C r , ω 0 ( 1 + Ψ r ( z n )) . After using stopping time trick, the BDG inequality controls the martingale by one half of E ′ sup q ≤ s Ψ r ( z n ( q ) ) \mathbb{E}'\sup_{q\leq s}\Psi_r(z_n(q)) E ′ sup q ≤ s Ψ r ( z n ( q )) plus the same integral term. By using the Gronwall inequality and removal of the stopping times, it is easy to get (43) .
The same monotonicity–Minty argument as in Proposition 2.7 , now on the fixed triple L 0 p ⊂ H 0 ⊂ ( L 0 p ) ∗ L_0^p\subset H_0\subset(L_0^p)^* L 0 p ⊂ H 0 ⊂ ( L 0 p ) ∗ , gives after passage to a subsequence
z n ⇀ z in L p ( Ω ′ × ( 0 , T ) × O ) ∩ L 2 ( Ω ′ × ( 0 , T ) ; H 0 ) . (44) z_n\rightharpoonup z\quad\text{in }L^p(\Omega'\times(0,T)\times\mathcal{O})\cap L^2(\Omega'\times(0,T);H_0). \tag*{(44)} z n ⇀ z in L p ( Ω ′ × ( 0 , T ) × O ) ∩ L 2 ( Ω ′ × ( 0 , T ) ; H 0 ) . (44)
Uniqueness of the fixed-mass equation identifies the limit with the conditional solution z z z . Since the limit is unique, every subsequence has the same weak limit.
Step 2: finite-dimensional partial Malliavin derivative. For the fixed mass path, the Galerkin drift is − A c ( s ) , n -A_{c(s),n} − A c ( s ) , n , where
( D A c ( s ) , n ( z ) y , y ) H 0 = ∫ O m ∣ z + c ( s ) ∣ m − 1 y 2 d x ≥ 0. (DA_{c(s),n}(z)y,y)_{H_0}=\int_{\mathcal{O}}m|z+c(s)|^{m-1}y^2\,\mathrm{d}x\geq0. ( D A c ( s ) , n ( z ) y , y ) H 0 = ∫ O m ∣ z + c ( s ) ∣ m − 1 y 2 d x ≥ 0.
To verify this identity, differentiating Φ ( z + θ y + c ( s ) ) \Phi(z+\theta y+c(s)) Φ ( z + θ y + c ( s )) at θ = 0 \theta=0 θ = 0 , we have
d d θ Φ ( z + θ y + c ( s ) ) ∣ θ = 0 = Φ ′ ( z + c ( s ) ) y = m ∣ z + c ( s ) ∣ m − 1 y . \left.\frac{\mathrm{d}}{\mathrm{d}\theta}\Phi(z+\theta y+c(s))\right|_{\theta=0}=\Phi'(z+c(s))y=m|z+c(s)|^{m-1}y. d θ d Φ ( z + θ y + c ( s )) θ = 0 = Φ ′ ( z + c ( s )) y = m ∣ z + c ( s ) ∣ m − 1 y .
Pairing the resulting derivative of A c ( s ) , n A_{c(s),n} A c ( s ) , n with the zero-mean Galerkin vector y y y gives exactly the displayed integral, and the positive Neumann projection disappears in this pairing because y y y lies in its range. Thus A c ( s ) , n ( z ) A_{c(s),n}(z) A c ( s ) , n ( z ) is C 1 C^1 C 1 , locally Lipschitz and one-sided Lipschitz with constant zero. The diffusion P n N Σ M ′ ( s , z ) P_n^N\Sigma'_M(s,z) P n N Σ M ′ ( s , z ) is affine with bounded state derivative P n N C P_n^NC P n N C . The moment bound (43) verifies the r > 2 r>2 r > 2 integrability required in [12 Corollary 3.5] . Hence the finite-dimensional conditional solution is Malliavin differentiable on Ω ′ \Omega' Ω ′ without any state cutoff.
For Y r , k n = D r ′ k z n Y_{r,k}^n=D_r^{\prime k}z_n Y r , k n = D r ′ k z n and s ≥ r s\geq r s ≥ r ,
d Y r , k n = P n N Δ N ( a n Y r , k n ) d s + P n N C Y r , k n d W s ′ , Y r , k n ( r ) = P n N Σ M ′ ( r , z n ( r ) ) e k , a n = m ∣ z n + c ∣ m − 1 . (45) \begin{aligned}
\mathrm{d}Y_{r,k}^n&=P_n^N\Delta_N(a_nY_{r,k}^n)\,\mathrm{d}s+P_n^NCY_{r,k}^n\,\mathrm{d}W'_s,\\
Y_{r,k}^n(r)&=P_n^N\Sigma'_M(r,z_n(r))e_k,\qquad a_n=m|z_n+c|^{m-1}. \tag*{(45)}
\end{aligned} d Y r , k n Y r , k n ( r ) = P n N Δ N ( a n Y r , k n ) d s + P n N C Y r , k n d W s ′ , = P n N Σ M ′ ( r , z n ( r )) e k , a n = m ∣ z n + c ∣ m − 1 . (45)
Here the notation C Y d W ′ CY\,\mathrm{d}W' C Y d W ′ means that C C C acts only in the ( i j ) (ij) ( ij ) coordinates; it is zero in the ( i 0 ) (i0) ( i 0 ) coordinates. Every Y r , k n Y_{r,k}^n Y r , k n is zero mean. Consequently
( P n N Δ N ( a n Y ) , Y ) H 0 = − ∫ O a n Y 2 d x . (P_n^N\Delta_N(a_nY),Y)_{H_0}=-\int_{\mathcal{O}}a_nY^2\,\mathrm{d}x. ( P n N Δ N ( a n Y ) , Y ) H 0 = − ∫ O a n Y 2 d x .
For each fixed k k k , Itô’s formula gives
E ′ ∥ Y r , k n ( s ) ∥ H 0 2 + 2 E ′ ∫ r s ∫ O a n ∣ Y r , k n ∣ 2 d x d q = E ′ ∥ Y r , k n ( r ) ∥ H 0 2 + E ′ ∫ r s ∥ P n N C Y r , k n ( q ) ∥ L 2 2 d q . \mathbb{E}'\|Y_{r,k}^{n}(s)\|_{H_0}^{2}+2\mathbb{E}'\int_{r}^{s}\int_{\mathcal{O}}a_n|Y_{r,k}^{n}|^{2}\,dx\,dq=\mathbb{E}'\|Y_{r,k}^{n}(r)\|_{H_0}^{2}+\mathbb{E}'\int_{r}^{s}\|P_n^{N}CY_{r,k}^{n}(q)\|_{\mathcal{L}_2}^{2}\,dq. E ′ ∥ Y r , k n ( s ) ∥ H 0 2 + 2 E ′ ∫ r s ∫ O a n ∣ Y r , k n ∣ 2 d x d q = E ′ ∥ Y r , k n ( r ) ∥ H 0 2 + E ′ ∫ r s ∥ P n N C Y r , k n ( q ) ∥ L 2 2 d q .
because the stochastic integral has zero expectation after the standard localization. Since
∥ P n N C Y ∥ L 2 2 ≤ ∥ C ∥ 2 ∥ Y ∥ H 0 2 , \|P_n^{N}CY\|_{\mathcal{L}_2}^{2}\leq\|C\|^{2}\|Y\|_{H_0}^{2}, ∥ P n N C Y ∥ L 2 2 ≤ ∥ C ∥ 2 ∥ Y ∥ H 0 2 ,
Gronwall inequality implies
sup r ≤ s ≤ T E ′ ∥ Y r , k n ( s ) ∥ H 0 2 ≤ e ∥ C ∥ 2 T E ′ ∥ Y r , k n ( r ) ∥ H 0 2 . \sup_{r\leq s\leq T}\mathbb{E}'\|Y_{r,k}^{n}(s)\|_{H_0}^{2}\leq e^{\|C\|^{2}T}\mathbb{E}'\|Y_{r,k}^{n}(r)\|_{H_0}^{2}. r ≤ s ≤ T sup E ′ ∥ Y r , k n ( s ) ∥ H 0 2 ≤ e ∥ C ∥ 2 T E ′ ∥ Y r , k n ( r ) ∥ H 0 2 .
Substitution back into the energy identity controls the space–time integral by the same initial energy. Summing over k k k and using
∑ k ∥ Y r , k n ( r ) ∥ H 0 2 = ∑ k ∥ P n N Σ M ′ ( r , z n ( r ) ) e k ∥ H 0 2 ≤ ∥ Σ M ′ ( r , z n ( r ) ) ∥ L 2 2 , \sum_k\|Y_{r,k}^{n}(r)\|_{H_0}^{2}=\sum_k\|P_n^{N}\Sigma'_M(r,z_n(r))e_k\|_{H_0}^{2}\leq\|\Sigma'_M(r,z_n(r))\|_{\mathcal{L}_2}^{2}, k ∑ ∥ Y r , k n ( r ) ∥ H 0 2 = k ∑ ∥ P n N Σ M ′ ( r , z n ( r )) e k ∥ H 0 2 ≤ ∥ Σ M ′ ( r , z n ( r )) ∥ L 2 2 ,
we obtain
sup r ≤ s ≤ T ∑ k E ′ ∥ Y r , k n ( s ) ∥ H 0 2 + ∑ k E ′ ∫ r T ∫ O a n ∣ Y r , k n ∣ 2 d x d s ≤ C T E ′ ∥ Σ M ′ ( r , z n ( r ) ) ∥ L 2 ( U ′ , H 0 ) 2 ≤ C T ( ω 0 ) . (46) \sup_{r\leq s\leq T}\sum_k\mathbb{E}'\|Y_{r,k}^{n}(s)\|_{H_0}^{2}+\sum_k\mathbb{E}'\int_{r}^{T}\int_{\mathcal{O}}a_n|Y_{r,k}^{n}|^{2}\,dx\,ds
\leq C_T\mathbb{E}'\|\Sigma'_M(r,z_n(r))\|_{\mathcal{L}_2(U',H_0)}^{2}\leq C_T(\omega_0). \tag*{(46)} r ≤ s ≤ T sup k ∑ E ′ ∥ Y r , k n ( s ) ∥ H 0 2 + k ∑ E ′ ∫ r T ∫ O a n ∣ Y r , k n ∣ 2 d x d s ≤ C T E ′ ∥ Σ M ′ ( r , z n ( r )) ∥ L 2 ( U ′ , H 0 ) 2 ≤ C T ( ω 0 ) . (46)
The last inequality follows from
∥ Σ M ′ ( r , z n ( r ) ) ∥ L 2 2 ≤ C 0 K M ( ω 0 ) + C 1 ∥ z n ( r ) ∥ H 0 2 \|\Sigma'_M(r,z_n(r))\|_{\mathcal{L}_2}^{2}\leq C_0K_M(\omega_0)+C_1\|z_n(r)\|_{H_0}^{2} ∥ Σ M ′ ( r , z n ( r )) ∥ L 2 2 ≤ C 0 K M ( ω 0 ) + C 1 ∥ z n ( r ) ∥ H 0 2
and (41) . Thus the constant is uniform in n n n and r r r . The constants in (41) –(46) may depend on the fixed mass path, but never on n n n or on the terminal-window length δ \delta δ .
Step 3: uniform terminal remainder estimate. Set
φ ‾ = ∣ O ∣ − 1 ∫ O φ d x , ψ = φ − φ ‾ , ψ n = P n N ψ . \overline{\varphi}=|\mathcal{O}|^{-1}\int_{\mathcal{O}}\varphi\,dx,\qquad\psi=\varphi-\overline{\varphi},\qquad\psi_n=P_n^{N}\psi. φ = ∣ O ∣ − 1 ∫ O φ d x , ψ = φ − φ , ψ n = P n N ψ .
By the Neumann part of Lemma 2.6, sup n ∥ Δ ψ n ∥ ∞ < ∞ \sup_n\|\Delta\psi_n\|_{\infty}<\infty sup n ∥Δ ψ n ∥ ∞ < ∞ and ψ n → ψ \psi_n\to\psi ψ n → ψ in H 1 H^1 H 1 . For F n ( t ) = ⟨ z n ( t ) , ψ n ⟩ + M t φ ‾ F_n(t)=\langle z_n(t),\psi_n\rangle+M_t\overline{\varphi} F n ( t ) = ⟨ z n ( t ) , ψ n ⟩ + M t φ , define
b n ( r ) = Π n Σ M ′ ( r , z n ( r ) ) ∗ ψ n . b_n(r)=\Pi_n\Sigma'_M(r,z_n(r))^*\psi_n. b n ( r ) = Π n Σ M ′ ( r , z n ( r ) ) ∗ ψ n .
The derivative equation gives
D r ′ F n ( t ) − b n ( r ) = R r n + S r n , D'_rF_n(t)-b_n(r)=R_r^n+S_r^n, D r ′ F n ( t ) − b n ( r ) = R r n + S r n ,
with the drift and stochastic remainders defined exactly as in (26) , with a n a_n a n , ψ n \psi_n ψ n and the zero-mean noise replacing their Dirichlet counterparts. Let
A n , δ ′ = 1 δ ∫ t − δ t ∑ k ∫ r t ∫ O a n ∣ Y r , k n ∣ 2 d x d s d r , Z n = ∫ 0 T ∥ u n ( s ) ∥ p d s . A'_{n,\delta}=\frac{1}{\delta}\int_{t-\delta}^{t}\sum_k\int_{r}^{t}\int_{\mathcal{O}}a_n|Y_{r,k}^{n}|^{2}\,dx\,ds\,dr,\qquad Z_n=\int_{0}^{T}\|u_n(s)\|^{p}\,ds. A n , δ ′ = δ 1 ∫ t − δ t k ∑ ∫ r t ∫ O a n ∣ Y r , k n ∣ 2 d x d s d r , Z n = ∫ 0 T ∥ u n ( s ) ∥ p d s .
For each r ∈ [ t − δ , t ] r\in[t-\delta,t] r ∈ [ t − δ , t ] and each k k k , weighted Cauchy–Schwarz gives
∣ R r , k n ∣ 2 ≤ ( ∫ r t ∫ O a n ∣ Y r , k n ∣ 2 d x d s ) ( ∫ r t ∫ O a n ∣ Δ ψ n ∣ 2 d x d s ) . |R_{r,k}^{n}|^{2}\leq\left(\int_{r}^{t}\int_{\mathcal{O}}a_n|Y_{r,k}^{n}|^{2}\,dx\,ds\right)\left(\int_{r}^{t}\int_{\mathcal{O}}a_n|\Delta\psi_n|^{2}\,dx\,ds\right). ∣ R r , k n ∣ 2 ≤ ( ∫ r t ∫ O a n ∣ Y r , k n ∣ 2 d x d s ) ( ∫ r t ∫ O a n ∣Δ ψ n ∣ 2 d x d s ) .
Summing in k k k and enlarging the second time interval to [ t − δ , t ] [t-\delta,t] [ t − δ , t ] yields
1 δ ∫ t − δ t ∥ R r n ∥ U ′ 2 d r ≤ A n , δ ′ C n , δ ′ . \frac{1}{\delta}\int_{t-\delta}^{t}\|R_r^n\|_{U'}^{2}\,dr\leq A'_{n,\delta}C'_{n,\delta}. δ 1 ∫ t − δ t ∥ R r n ∥ U ′ 2 d r ≤ A n , δ ′ C n , δ ′ .
where
C n , δ ′ : = ∫ t − δ t ∫ O a n ∣ Δ ψ n ∣ 2 d x d s . C'_{n,\delta}:=\int_{t-\delta}^{t}\int_{\mathcal{O}}a_n|\Delta\psi_n|^2\,dx\,ds. C n , δ ′ := ∫ t − δ t ∫ O a n ∣Δ ψ n ∣ 2 d x d s .
By Lemma 2.6, sup n ∥ Δ ψ n ∥ ∞ < ∞ \sup_n\|\Delta\psi_n\|_\infty<\infty sup n ∥Δ ψ n ∥ ∞ < ∞ , and hence
C n , δ ′ ≤ C φ ∫ t − δ t ∫ O ∣ u n ∣ m − 1 d x d s . C'_{n,\delta}\le C_\varphi\int_{t-\delta}^{t}\int_{\mathcal{O}}|u_n|^{m-1}\,dx\,ds. C n , δ ′ ≤ C φ ∫ t − δ t ∫ O ∣ u n ∣ m − 1 d x d s .
Put α = ( m − 1 ) / ( m + 1 ) \alpha=(m-1)/(m+1) α = ( m − 1 ) / ( m + 1 ) . Hölder’s inequality on [ t − δ , t ] × O [t-\delta,t]\times\mathcal{O} [ t − δ , t ] × O gives
∫ t − δ t ∫ O ∣ u n ∣ m − 1 ≤ ( δ ∣ O ∣ ) 1 − α ( ∫ t − δ t ∫ O ∣ u n ∣ m + 1 d x d s ) α ≤ C δ 1 − α Z n α . \int_{t-\delta}^{t}\int_{\mathcal{O}}|u_n|^{m-1}\le(\delta|\mathcal{O}|)^{1-\alpha}\left(\int_{t-\delta}^{t}\int_{\mathcal{O}}|u_n|^{m+1}\,dx\,ds\right)^\alpha\le C\delta^{1-\alpha}Z_n^\alpha. ∫ t − δ t ∫ O ∣ u n ∣ m − 1 ≤ ( δ ∣ O ∣ ) 1 − α ( ∫ t − δ t ∫ O ∣ u n ∣ m + 1 d x d s ) α ≤ C δ 1 − α Z n α .
Therefore
C n , δ ′ ≤ C φ δ 1 − α Z n α . C'_{n,\delta}\le C_\varphi\delta^{1-\alpha}Z_n^\alpha. C n , δ ′ ≤ C φ δ 1 − α Z n α .
By (46) , sup n , δ E ′ A n , δ ′ < ∞ \sup_{n,\delta}\mathbb{E}'A'_{n,\delta}<\infty sup n , δ E ′ A n , δ ′ < ∞ , while (42) gives
sup n E ′ Z n < ∞ . \sup_n\mathbb{E}'Z_n<\infty. n sup E ′ Z n < ∞.
With q = ( m + 1 ) / ( 2 m ) q=(m+1)/(2m) q = ( m + 1 ) / ( 2 m ) , the same Hölder calculation as in (27) yields
sup n E ′ ( A n , δ ′ C n , δ ′ ) q ≤ C ( ω 0 ) δ 1 / m . \sup_n\mathbb{E}'(A'_{n,\delta}C'_{n,\delta})^q\le C(\omega_0)\delta^{1/m}. n sup E ′ ( A n , δ ′ C n , δ ′ ) q ≤ C ( ω 0 ) δ 1/ m .
Itô isometry for the stochastic remainder and (46) give
sup n E ′ [ ( 1 δ ∫ t − δ t ∥ S n r ∥ U 2 d r ) q ] ≤ C ( ω 0 ) δ q . \sup_n\mathbb{E}'\left[\left(\frac{1}{\delta}\int_{t-\delta}^{t}\|S_n^r\|_{\mathcal U}^2\,dr\right)^q\right]\le C(\omega_0)\delta^q. n sup E ′ [ ( δ 1 ∫ t − δ t ∥ S n r ∥ U 2 d r ) q ] ≤ C ( ω 0 ) δ q .
Therefore
sup n E ′ [ ( 1 δ ∫ t − δ t ∥ D r ′ F n ( t ) − b n ( r ) ∥ U 2 d r ) q ] ≤ C ( ω 0 ) ( δ 1 / m + δ q ) . (47) \sup_n\mathbb{E}'\left[\left(\frac{1}{\delta}\int_{t-\delta}^{t}\|D'_rF_n(t)-b_n(r)\|_{\mathcal U}^2\,dr\right)^q\right]\le C(\omega_0)(\delta^{1/m}+\delta^q). \tag*{(47)} n sup E ′ [ ( δ 1 ∫ t − δ t ∥ D r ′ F n ( t ) − b n ( r ) ∥ U 2 d r ) q ] ≤ C ( ω 0 ) ( δ 1/ m + δ q ) . (47)
*Step 4: fixed-time weak convergence and the closed graph of D ′ D' D ′ . The weak convergence (44) alone is not enough for Malliavin calculus at the prescribed time t t t , so we record the fixed-time step. For the scalar Galerkin observations, the equation, the uniform flux bound, BDG and (41) –(42) imply
sup n E ′ ∣ F n ( t ) − F n ( s ) ∣ ≤ C ω 0 , φ ( ∣ t − s ∣ 1 / p + ∣ t − s ∣ 1 / 2 ) . \sup_n\mathbb{E}'|F_n(t)-F_n(s)|\le C_{\omega_0,\varphi}(|t-s|^{1/p}+|t-s|^{1/2}). n sup E ′ ∣ F n ( t ) − F n ( s ) ∣ ≤ C ω 0 , φ ( ∣ t − s ∣ 1/ p + ∣ t − s ∣ 1/2 ) .
The one-sided time-averaging argument used in Step 3 of Proposition 2.7 therefore gives
F n ( t ) ⇀ F ( t ) in L 2 ( Ω ′ ) . (48) F_n(t)\rightharpoonup F(t)\quad\text{in }L^2(\Omega'). \tag*{(48)} F n ( t ) ⇀ F ( t ) in L 2 ( Ω ′ ) . (48)
The estimate (46) , paired with ψ n \psi_n ψ n , gives
sup n E ′ ∫ 0 t ∥ D r ′ F n ( t ) ∥ U 2 d r < ∞ . \sup_n\mathbb{E}'\int_0^t\|D'_rF_n(t)\|_{\mathcal U}^2\,dr<\infty. n sup E ′ ∫ 0 t ∥ D r ′ F n ( t ) ∥ U 2 d r < ∞.
Hence, after extraction,
D ′ F n ( t ) ⇀ V in L 2 ( Ω ′ ; H ′ ) . D'F_n(t)\rightharpoonup V\quad\text{in }L^2(\Omega';\mathcal H'). D ′ F n ( t ) ⇀ V in L 2 ( Ω ′ ; H ′ ) .
The Malliavin derivative
D ′ : L 2 ( Ω ′ ) ⊃ D 1 , 2 ( Ω ′ ) → L 2 ( Ω ′ ; H ′ ) D':L^2(\Omega')\supset\mathbb{D}^{1,2}(\Omega')\to L^2(\Omega';\mathcal H') D ′ : L 2 ( Ω ′ ) ⊃ D 1 , 2 ( Ω ′ ) → L 2 ( Ω ′ ; H ′ )
is a closed linear operator [15 ] ; its graph is therefore a closed linear subspace of a Hilbert space and is weakly closed. Together with (48) , this implies
F ( t ) ∈ D 1 , 2 ( Ω ′ ) , V = D ′ F ( t ) . F(t) \in\mathbb{D}^{1,2}(\Omega'), \qquad V = D'F(t). F ( t ) ∈ D 1 , 2 ( Ω ′ ) , V = D ′ F ( t ) .
For fixed δ > 0 \delta> 0 δ > 0 , affine linearity of Σ M ′ \Sigma'_M Σ M ′ and the strong convergence of the spectral/noise projections imply b n ⇀ b b_n \rightharpoonup b b n ⇀ b in L 2 ( Ω ′ × ( t − δ , t ) ; U ′ ) L^2(\Omega' \times(t-\delta,t);U') L 2 ( Ω ′ × ( t − δ , t ) ; U ′ ) , where
b ( r ) = Σ M ′ ( r , z ( r ) ) ∗ ψ . b(r) = \Sigma'_M(r,z(r))^*\psi. b ( r ) = Σ M ′ ( r , z ( r ) ) ∗ ψ .
Let ρ = 2 q = ( m + 1 ) / m > 1 \rho= 2q = (m+1)/m > 1 ρ = 2 q = ( m + 1 ) / m > 1 . The continuous embedding
L 2 ( Ω ′ ; L 2 ( t − δ , t ; U ′ ) ) ↪ L ρ ( Ω ′ ; L 2 ( t − δ , t ; U ′ ) ) L^2(\Omega';L^2(t-\delta,t;U')) \hookrightarrow L^\rho(\Omega';L^2(t-\delta,t;U')) L 2 ( Ω ′ ; L 2 ( t − δ , t ; U ′ )) ↪ L ρ ( Ω ′ ; L 2 ( t − δ , t ; U ′ ))
transports weak convergence to the latter reflexive Banach space. Its norm is weakly lower semicontinuous, so (47) passes to the limit:
E ′ [ ( 1 δ ∫ t − δ t ∥ D r ′ F ( t ) − b ( r ) ∥ U ′ 2 d r ) q ] ≤ C ( ω 0 ) ( δ 1 / m + δ q ) . (49) \mathbb{E}'\left[\left(\frac{1}{\delta}\int_{t-\delta}^{t}\lVert D'_rF(t)-b(r)\rVert_{U'}^2\,dr\right)^q\right] \leq C(\omega_0)(\delta^{1/m}+\delta^q). \tag*{(49)} E ′ [ ( δ 1 ∫ t − δ t ∥ D r ′ F ( t ) − b ( r ) ∥ U ′ 2 d r ) q ] ≤ C ( ω 0 ) ( δ 1/ m + δ q ) . (49)
Step 5: replacement of the center by its terminal value. The fixed-mass solution belongs to L 2 ( Ω ′ ; C ( [ 0 , T ] ; H 0 ) ) L^2(\Omega';C([0,T];H_0)) L 2 ( Ω ′ ; C ([ 0 , T ] ; H 0 )) , M M M is continuous, and Σ M ′ \Sigma'_M Σ M ′ is affine and bounded. Hence
ω t , ω 0 ( δ ) : = 1 δ ∫ t − δ t E ′ ∥ b ( r ) − b ( t ) ∥ U ′ 2 d r ⟶ 0. \omega_{t,\omega_0}(\delta) := \frac{1}{\delta}\int_{t-\delta}^{t}\mathbb{E}'\lVert b(r)-b(t)\rVert_{U'}^2\,dr \longrightarrow0. ω t , ω 0 ( δ ) := δ 1 ∫ t − δ t E ′ ∥ b ( r ) − b ( t ) ∥ U ′ 2 d r ⟶ 0.
Since 0 < q < 1 0 < q < 1 0 < q < 1 , the squared triangle inequality followed by ( a + b ) q ≤ a q + b q (a+b)^q \leq a^q+b^q ( a + b ) q ≤ a q + b q gives from (49)
E ′ [ ( 1 δ ∫ t − δ t ∥ D r ′ F ( t ) − b ( r ) ∥ U ′ 2 d r ) q ] ≤ C ( ω 0 ) ( δ 1 / m + δ q + ω t , ω 0 ( δ ) q ) ⟶ 0. \mathbb{E}'\left[\left(\frac{1}{\delta}\int_{t-\delta}^{t}\lVert D'_rF(t)-b(r)\rVert_{U'}^2\,dr\right)^q\right] \leq C(\omega_0)\left(\delta^{1/m}+\delta^q+\omega_{t,\omega_0}(\delta)^q\right) \longrightarrow0. E ′ [ ( δ 1 ∫ t − δ t ∥ D r ′ F ( t ) − b ( r ) ∥ U ′ 2 d r ) q ] ≤ C ( ω 0 ) ( δ 1/ m + δ q + ω t , ω 0 ( δ ) q ) ⟶ 0.
Finally,
F ( t ) = ⟨ z ( t ) , φ − φ ‾ ⟩ + M t φ ‾ , D ′ M t = 0 , F(t) = \langle z(t),\varphi-\overline{\varphi}\rangle+ M_t\overline{\varphi}, \qquad D'M_t=0, F ( t ) = ⟨ z ( t ) , φ − φ ⟩ + M t φ , D ′ M t = 0 ,
and b ( t ) b(t) b ( t ) is precisely the vector in (40) . This proves (39) .
Theorem 4.6 (Neumann full-domain support for nonzero initial mass). Assume (33) –(34) and let u 0 ∈ H N u_0 \in\mathcal{H}_N u 0 ∈ H N be deterministic with M ( u 0 ) ≠ 0 M(u_0) \ne0 M ( u 0 ) = 0 . For every fixed t > 0 t > 0 t > 0 and every nonzero φ ∈ C c ∞ ( O ) \varphi\in C^\infty_c(O) φ ∈ C c ∞ ( O ) , the scalar observation ⟨ u ( t ) , φ ⟩ \langle u(t),\varphi\rangle ⟨ u ( t ) , φ ⟩ has an absolutely continuous law. Moreover,
P ( supp O u ( t , ⋅ ) = O ‾ ) = 1. (50) \mathbb{P}\bigl(\operatorname{supp}_O u(t,\cdot)=\overline{O}\bigr)=1. \tag*{(50)} P ( supp O u ( t , ⋅ ) = O ) = 1. (50) No separate non-hitting or Malliavin assumption is imposed.
Proof. By Theorem 4.4 , M t ≠ 0 M_t \ne0 M t = 0 for every finite t t t on a P 0 \mathbb{P}_0 P 0 -full set of mass paths. Fix such a path in the full set Ω 0 ∗ \Omega^*_0 Ω 0 ∗ of Proposition 4.5 . Since every b i b_i b i is positive and the span of ( f i ) (f_i) ( f i ) is dense in L 0 2 ( O ) L^2_0(O) L 0 2 ( O ) , a nonzero φ ∈ C c ∞ ( O ) \varphi\in C^\infty_c(O) φ ∈ C c ∞ ( O ) satisfies
∥ Θ t ′ , ω 0 , φ ∥ U ′ 2 ≥ ( M t ω 0 ) 2 ∑ i b i 2 ∣ ⟨ f i , φ ⟩ ∣ 2 > 0. \lVert\Theta_t^{\prime,\omega_0,\varphi}\rVert_{U'}^2 \geq(M_t^{\omega_0})^2\sum_i b_i^2\lvert\langle f_i,\varphi\rangle\rvert^2 > 0. ∥ Θ t ′ , ω 0 , φ ∥ U ′ 2 ≥ ( M t ω 0 ) 2 i ∑ b i 2 ∣⟨ f i , φ ⟩ ∣ 2 > 0.
Indeed, if every pairing vanished then the zero-mean part of φ \varphi φ would vanish, so φ \varphi φ would be constant on the connected domain; compact support then forces φ = 0 \varphi=0 φ = 0 .
Now, we apply Lemma 2.12 on the Wiener space ( Ω ′ , F ′ , P ′ ) (\Omega',\mathcal{F}',\mathbb{P}') ( Ω ′ , F ′ , P ′ ) , with D ′ D' D ′ in place of D D D and with the terminal estimate of Proposition 4.5 . The conditional Malliavin covariance is strictly positive P ′ \mathbb{P}^{\prime} P ′ -almost surely, hence F t , φ ω 0 F_{t,\varphi}^{\omega_0} F t , φ ω 0 has a density with respect to Lebesgue measure. Thus for every Lebesgue-null Borel set A A A ,
P ′ ( F t , φ ω 0 ∈ A ) = 0 for P 0 -a.e. ω 0 . \mathbb{P}^{\prime}\left(F_{t,\varphi}^{\omega_0}\in A\right)=0 \qquad\text{for }\mathbb{P}_0\text{-a.e. }\omega_0. P ′ ( F t , φ ω 0 ∈ A ) = 0 for P 0 -a.e. ω 0 .
Fubini’s theorem on the product realization gives
P ( F t , φ ∈ A ) = ∫ Ω 0 P ′ ( F t , φ ω 0 ∈ A ) P 0 ( d ω 0 ) = 0 , \mathbb{P}\left(F_{t,\varphi}\in A\right)=\int_{\Omega_0}\mathbb{P}^{\prime}\left(F_{t,\varphi}^{\omega_0}\in A\right)\mathbb{P}_0(\mathrm{d}\omega_0)=0, P ( F t , φ ∈ A ) = ∫ Ω 0 P ′ ( F t , φ ω 0 ∈ A ) P 0 ( d ω 0 ) = 0 ,
which proves the unconditional absolute continuity. In particular P ( F t , φ = 0 ) = 0 \mathbb{P}(F_{t,\varphi}=0)=0 P ( F t , φ = 0 ) = 0 for every nonzero localized test. Lemma 2.14 then yields (50) .
Remark 4.7. (i) The nonzero-mass restriction in Theorem 4.6 is essential to this proof. The mass formula alone says nothing about non-hitting for nonzero data of zero mass. In the nonzero-mass class, the K i j K_{ij} K ij modes are not needed for the conclusion: one may set them all to zero and retain the same proof. They were included to make the entire family satisfy nonlocal vacuum ellipticity also at nonzero zero-mass states, as verified in Proposition 4.2 .
(ii) If also σ = 0 \sigma=0 σ = 0 , mass is constant and the K i 0 K_{i0} K i 0 forcing becomes the additive coefficient M 0 G 0 M_0G_0 M 0 G 0 along the solution. For σ ≠ 0 \sigma\ne0 σ = 0 , the preceding conditional proof treats its random amplitude. In both cases the original state-to-noise maps v ↦ M ( v ) b i f i v\mapsto M(v)b_if_i v ↦ M ( v ) b i f i are linear and spatially nonlocal.
Remark 4.8. (i) The distinction between the Dirichlet and Neumann conclusions lies in whether the zero state can be excluded. In the Dirichlet variational space H D = ( H 0 1 ) ∗ H_D=(H_0^1)^* H D = ( H 0 1 ) ∗ , the constant function 1 1 1 is not an admissible test function, so the mass argument used for the Neumann problem does not apply. Under (21) and nonlocal vacuum ellipticity, the Dirichlet argument gives the terminal estimate but does not establish non-hitting. In the Neumann product space H N = R ⊕ H 0 H_N=\mathbb{R}\oplus H_0 H N = R ⊕ H 0 , the mass satisfies a closed scalar SDE. Nonzero initial mass therefore rules out the zero state.
(ii) The Malliavin arguments also use different noise directions. The Dirichlet estimate involves all driving directions and controls the terminal quantity B ( u ( t ) ) ∗ φ B(u(t))^*\varphi B ( u ( t ) ) ∗ φ . In the Neumann case, we condition on the mass Brownian motion and differentiate only along the independent zero-mean noise directions. These directions suffice to obtain strictly positive conditional covariance.
Numerical Example
This section gives a one-dimensional numerical illustration of the propagation mechanisms discussed above. The experiment is designed to distinguish three effects: finite propagation for the deterministic porous medium equation, immediate excitation of an interior vacuum by additive noise, and immediate excitation of the same vacuum by a genuinely nonlocal multiplicative coefficient. The purpose is illustrative rather than a numerical proof of the Malliavin-support theorems.
The bounded-domain theory is formulated for a bounded connected C 2 C^2 C 2 domain O = B 1 ( 0 ) ⊂ R 2 \mathcal{O}=B_1(0)\subset\mathbb{R}^2 O = B 1 ( 0 ) ⊂ R 2 . The purpose is to visualize the same three mechanisms as in the analytical results: persistence of an interior vacuum for the deterministic porous-medium flow, immediate vacuum activation by additive noise, and vacuum activation by a spatially nonlocal multiplicative coefficient.
Let
m = 2 , w 0 ( x ) = ( 1 − ∣ x ∣ 2 0.30 2 ) + m=2,\qquad w_0(x)=\left(1-\frac{|x|^2}{0.30^2}\right)_+ m = 2 , w 0 ( x ) = ( 1 − 0.3 0 2 ∣ x ∣ 2 ) +
Thus supp u 0 ⊂ B 0.30 ( 0 ) \operatorname{supp} u_0 \subset B_{0.30}(0) supp u 0 ⊂ B 0.30 ( 0 ) . We compare
∂ t u D = Δ Φ ( u D ) , d u A = Δ Φ ( u A ) d t + ∑ k = 1 7 σ k A ψ k d β k ( t ) , d u N = Δ Φ ( u N ) d t + ∑ k = 1 7 σ k N Ψ k ( u N ) d β k ( t ) . \begin{aligned}
\partial_t u^{\mathrm{D}} &= \Delta\Phi(u^{\mathrm{D}}), \qquad\mathrm{d}u^{\mathrm{A}} = \Delta\Phi(u^{\mathrm{A}})\,\mathrm{d}t + \sum_{k=1}^{7} \sigma_k^{\mathrm{A}}\psi_k\,\mathrm{d}\beta_k(t),\\
\mathrm{d}u^{\mathrm{N}} &= \Delta\Phi(u^{\mathrm{N}})\,\mathrm{d}t + \sum_{k=1}^{7} \sigma_k^{\mathrm{N}}\Psi_k(u^{\mathrm{N}})\,\mathrm{d}\beta_k(t).
\end{aligned} ∂ t u D d u N = ΔΦ ( u D ) , d u A = ΔΦ ( u A ) d t + k = 1 ∑ 7 σ k A ψ k d β k ( t ) , = ΔΦ ( u N ) d t + k = 1 ∑ 7 σ k N Ψ k ( u N ) d β k ( t ) .
where Φ ( r ) = ∣ r ∣ r \Phi(r)=|r|r Φ ( r ) = ∣ r ∣ r . For the numerical additive modes we take
ψ k ( x ) = ( 1 − ∣ x ∣ 2 ) + 1 2 π ℓ 2 exp ( − ∣ x − q k ∣ 2 2 ℓ 2 ) , ℓ = 0.25 , \psi_k(x)=\left(1-|x|^2\right)_+\frac{1}{2\pi\ell^2}\exp\left(-\frac{|x-q_k|^2}{2\ell^2}\right), \qquad\ell=0.25, ψ k ( x ) = ( 1 − ∣ x ∣ 2 ) + 2 π ℓ 2 1 exp ( − 2 ℓ 2 ∣ x − q k ∣ 2 ) , ℓ = 0.25 ,
with shifts
q k ∈ { 0 , ± 0.4 e 1 , ± 0.4 e 2 , ( 0.3 , 0.3 ) , ( − 0.3 , − 0.3 ) } , q_k \in\{0,\ \pm0.4e_1,\ \pm0.4e_2,\ (0.3,0.3),\ (-0.3,-0.3)\}, q k ∈ { 0 , ± 0.4 e 1 , ± 0.4 e 2 , ( 0.3 , 0.3 ) , ( − 0.3 , − 0.3 )} ,
and weights σ k A = 0.035 k − 1.1 \sigma_k^{\mathrm{A}}=0.035k^{-1.1} σ k A = 0.035 k − 1.1 . For the multiplicative model we use the finite-mode nonlocal coefficient
Ψ k ( v ) ( x ) = ( 1 − ∣ x ∣ 2 ) + ( g q k ∗ v ) ( x ) , σ k N = 0.80 k − 1.1 . \Psi_k(v)(x)=\left(1-|x|^2\right)_+(g_{q_k}*v)(x), \qquad\sigma_k^{\mathrm{N}}=0.80k^{-1.1}. Ψ k ( v ) ( x ) = ( 1 − ∣ x ∣ 2 ) + ( g q k ∗ v ) ( x ) , σ k N = 0.80 k − 1.1 .
The cutoff only enforces compatibility with the boundary in the finite-grid experiment; the map remains spatially nonlocal in the interior because its value at x x x depends on the state over the whole disk.
A Cartesian grid with spacing Δ x = 0.02 \Delta x=0.02 Δ x = 0.02 is masked by the unit disk and zero values are imposed outside the disk. We use Δ t = 10 − 5 \Delta t=10^{-5} Δ t = 1 0 − 5 , T = 2 × 10 − 4 T=2\times10^{-4} T = 2 × 1 0 − 4 , and Euler–Maruyama for the stochastic increments. Define the initially empty remote annulus
A r e m ( 2 ) = { x : 0.75 ≤ ∣ x ∣ ≤ 0.90 } A_{\mathrm{rem}}^{(2)}=\{x:0.75\le|x|\le0.90\} A rem ( 2 ) = { x : 0.75 ≤ ∣ x ∣ ≤ 0.90 }
and
A 2 ( t ) = ∫ A r e m ( 2 ) ∣ u ( t , x ) ∣ d x . A_2(t)=\int_{A_{\mathrm{rem}}^{(2)}} |u(t,x)|\,\mathrm{d}x. A 2 ( t ) = ∫ A rem ( 2 ) ∣ u ( t , x ) ∣ d x .
The numerical values produced by the accompanying Python code are summarized in Table 5.1 . The deterministic computation retains an exact remote vacuum on the grid over the displayed time interval: A 2 ( T ) = 0 A_2(T)=0 A 2 ( T ) = 0 . In contrast, both stochastic models create a nonzero far-field signal after the small time. The deterministic remote annulus remains empty on the displayed time interval, while both stochastic models activate it immediately, see Figure 5.1 . In order to see it more clearly, the exact value of u ( t , x , 0 ) u(t,x,0) u ( t , x , 0 ) is given in Figure 5.2 . Moreover, for three models, the logarithmic magnitude of one realization is shown in Figure 5.3 .
Table 5.1. Two-dimensional unit-disk computation. A grid point is called active when ∣ u ∣ > 10 − 8 |u|>10^{-8} ∣ u ∣ > 1 0 − 8 .
Figure 5.1. Remote-annulus activity in the two-dimensional bounded-domain experiment. The deterministic remote annulus remains empty on the displayed time interval, while both stochastic models activate it immediately.
Figure 5.2. Absolute value along the diameter y = 0 y = 0 y = 0 at the final time.
Figure 5.3. (a)–(c) respectively represent the logarithmic magnitude of one realization of deterministic, additive, and nonlocal multiplicative models.
Lastly, we give some remarks. The experiment reproduces the qualitative distinction proved analytically. For the deterministic PME, the diffusion coefficient degenerates at the vacuum and the interface moves only a finite distance during a short time interval. Additive noise is independent of the current state, so its Dirichlet eigenmodes directly inject a signal into a remote vacuum. The operator g q k ∗ v g_{q_k}*v g q k ∗ v is state dependent but spatially nonlocal.
There are three limitations: First, a spatial grid and a finite noise truncation cannot prove the continuum statement supp O u ( t ) = O \operatorname{supp}_{\mathcal{O}} u(t) = \mathcal{O} supp O u ( t ) = O almost surely. They only visualize its mechanism. Second, a support radius in a stochastic signed solution is threshold dependent. Third, the finite nonlocal truncation does not itself satisfy the infinite-dimensional all-test nonlocal vacuum ellipticity condition. The analytical Malliavin argument, rather than this experiment, is what upgrades terminal nondegeneracy to the fixed-time full-support conclusion.
Conclusion
We have studied the spatial support of solutions to porous medium equations on bounded domains under additive and nonlocal multiplicative noise. For additive noise, we prove that the solution has full distributional support at each fixed positive time. Since the noise need not preserve nonnegativity, the argument uses test functions rather than the positivity set of the solution.
For the Dirichlet problem with nonlocal multiplicative noise, we prove that the solution at each fixed positive time is either identically zero or has full distributional support. A non-hitting condition excludes the zero state and gives full support. In the Neumann model, the total mass satisfies a closed scalar equation; nonzero initial mass excludes the zero state without a separate non-hitting assumption. These results show how the support conclusion depends on the noise and the boundary condition. Whether a similar conclusion holds for porous medium equations with reflection [16 ] remains an interesting question.
During the preparation of this work, the authors used GPT-5.6 Sol to assist with language, the python code and the organization of this paper.
Acknowledgments
This work was supported by the National Natural Science Foundation of China grant 12171247.
Conflict of interest
The authors declare there is no conflict of interest.
Data Availability Statement
The numerical simulation data are available from the corresponding author upon reasonable request.