Introduction and main results
We consider the stochastic heat equation
∂ t u t ( x ) = 1 2 ∂ x x u t ( x ) + b ( t , x , u t ( x ) ) + σ ( t , x , u t ( x ) ) W ˙ ( t , x ) , t > 0 , x ∈ R , (1) \partial_t u_t(x) = \frac{1}{2}\partial_{xx}u_t(x) + b(t,x,u_t(x)) + \sigma(t,x,u_t(x))\dot{W}(t,x), \qquad t > 0,\ x \in\mathbb{R}, \tag*{(1)} ∂ t u t ( x ) = 2 1 ∂ xx u t ( x ) + b ( t , x , u t ( x )) + σ ( t , x , u t ( x )) W ˙ ( t , x ) , t > 0 , x ∈ R , (1)
where W ˙ \dot{W} W ˙ is space-time Gaussian white noise and b , σ : R + × R 2 → R b,\sigma: \mathbb{R}_+ \times\mathbb{R}^2 \to\mathbb{R} b , σ : R + × R 2 → R are Borel functions. The initial function u 0 u_0 u 0 is deterministic and belongs to
C t e m : = { f ∈ C ( R ) : ∣ f ∣ μ < ∞ for every μ > 0 } , ∣ f ∣ μ : = sup x ∈ R e − μ ∣ x ∣ ∣ f ( x ) ∣ . C_{\mathrm{tem}} := \{f \in C(\mathbb{R}) : |f|_\mu< \infty\text{ for every } \mu> 0\}, \qquad|f|_\mu:= \sup_{x \in\mathbb{R}} e^{-\mu|x|}|f(x)|. C tem := { f ∈ C ( R ) : ∣ f ∣ μ < ∞ for every μ > 0 } , ∣ f ∣ μ := x ∈ R sup e − μ ∣ x ∣ ∣ f ( x ) ∣.
We equip C t e m C_{\mathrm{tem}} C tem with the topology induced by these seminorms.
For Lipschitz coefficients of linear growth, existence and pathwise uniqueness follow by Picard iteration; see Walsh [11, Chapter 3] on a bounded interval and Shiga [10, Theorem 2.2] on R \mathbb{R} R .
Branching particle models give a reason to weaken the Lipschitz condition on σ \sigma σ . Konno and Shiga [5, Theorem 1.4] derived the equation for the nonnegative density u t u_t u t of a one-dimensional super-Brownian motion X t ( d x ) = u t ( x ) d x X_t(dx) = u_t(x)\,dx X t ( d x ) = u t ( x ) d x . With the usual normalization, u u u satisfies (1.1) with b = 0 b = 0 b = 0 and σ ( z ) = z \sigma(z) = \sqrt{z} σ ( z ) = z for z ≥ 0 z \ge0 z ≥ 0 . Exponential duality gives uniqueness in law among nonnegative solutions, but pathwise uniqueness remains open (see [7 Example 1] ). See also Xiong [12 ] for a related SPDE formulation and Chen [1 ] for super-Brownian motion with immigration.
If particles branch at a rate r ( z ) r(z) r ( z ) determined by the local density z ≥ 0 z \ge0 z ≥ 0 , the noise coefficient becomes z r ( z ) \sqrt{z}\,r(z) z r ( z ) (see [8 Section 1] ). For instance, the choice r ( z ) = z r(z) = \sqrt{z} r ( z ) = z gives σ ( z ) = z 3 / 4 \sigma(z) = z^{3/4} σ ( z ) = z 3/4 . Particles can also interact by annihilating on contact. In the model of Mueller, Mytnik and Perkins [7 Example 4] , each collision between particles of opposite sign removes equal positive and negative mass. If u + u^{+} u + and u − u^{-} u − are the two densities in a scaling limit, their difference u = u + − u − u=u^{+}-u^{-} u = u + − u − can take either sign and satisfies (1.1) with b = 0 b=0 b = 0 and σ ( z ) = ∣ z ∣ \sigma(z)=\sqrt{|z|} σ ( z ) = ∣ z ∣ for z ∈ R z\in\mathbb{R} z ∈ R . These models lead us to study Hölder continuous diffusion coefficients. It is natural to ask the following question.
How far can the regularity of σ \sigma σ be weakened while retaining pathwise uniqueness among real-valued solutions?
For Hölder diffusion coefficients, this problem remained open only at the critical exponent γ = 3 / 4 \gamma=3/4 γ = 3/4 . Mytnik and Perkins [8 Theorem 1.2] proved pathwise uniqueness for (1.1) when σ \sigma σ is locally Hölder continuous of order γ > 3 / 4 \gamma>3/4 γ > 3/4 , under their growth hypotheses. For every 0 < γ < 3 / 4 0<\gamma<3/4 0 < γ < 3/4 , Mueller, Mytnik and Perkins [7 Theorem 1.1] proved nonuniqueness in law and pathwise nonuniqueness by constructing a nontrivial real-valued solution with b = 0 b=0 b = 0 , σ ( z ) = ∣ z ∣ γ \sigma(z)=|z|^{\gamma} σ ( z ) = ∣ z ∣ γ and zero initial data.
Our main aim is to settle the critical case γ = 3 / 4 \gamma=3/4 γ = 3/4 . We give a concise proof for all γ ∈ [ 3 / 4 , 1 ] \gamma\in[3/4,1] γ ∈ [ 3/4 , 1 ] using the short-time linear comparison method introduced in Subsection 1.1.
Condition (H). For some γ ∈ [ 3 / 4 , 1 ] \gamma\in[3/4,1] γ ∈ [ 3/4 , 1 ] and every T < ∞ T<\infty T < ∞ , there exist constants C , L , K , a , m ≥ 0 C,L,K,a,m\geq0 C , L , K , a , m ≥ 0 such that the following hold for 0 ≤ t ≤ T 0\leq t\leq T 0 ≤ t ≤ T and x , z , w ∈ R x,z,w\in\mathbb{R} x , z , w ∈ R .
(i) The coefficients have linear growth:
∣ b ( t , x , z ) ∣ + ∣ σ ( t , x , z ) ∣ ≤ C ( 1 + ∣ z ∣ ) . |b(t,x,z)|+|\sigma(t,x,z)|\leq C(1+|z|). ∣ b ( t , x , z ) ∣ + ∣ σ ( t , x , z ) ∣ ≤ C ( 1 + ∣ z ∣ ) .
(ii) The drift is Lipschitz:
∣ b ( t , x , z ) − b ( t , x , w ) ∣ ≤ L ∣ z − w ∣ . |b(t,x,z)-b(t,x,w)|\leq L|z-w|. ∣ b ( t , x , z ) − b ( t , x , w ) ∣ ≤ L ∣ z − w ∣.
(iii) For ∣ z − w ∣ ≤ 1 |z-w|\leq1 ∣ z − w ∣ ≤ 1 , the diffusion coefficient satisfies
∣ σ ( t , x , z ) − σ ( t , x , w ) ∣ ≤ K e a ∣ x ∣ ( 1 + ∣ z ∣ + ∣ w ∣ ) m ∣ z − w ∣ γ . |\sigma(t,x,z)-\sigma(t,x,w)|\leq Ke^{a|x|}(1+|z|+|w|)^{m}|z-w|^{\gamma}. ∣ σ ( t , x , z ) − σ ( t , x , w ) ∣ ≤ K e a ∣ x ∣ ( 1 + ∣ z ∣ + ∣ w ∣ ) m ∣ z − w ∣ γ .
All filtered probability spaces ( Ω , F , ( F t ) t ≥ 0 , P ) (\Omega,\mathcal{F},(\mathcal{F}_{t})_{t\geq0},\mathbb{P}) ( Ω , F , ( F t ) t ≥ 0 , P ) considered below satisfy the usual conditions. For a space-time Gaussian white noise W W W with intensity d t d x dt\,dx d t d x on such a space, we require that the noise up to time t t t is F t \mathcal{F}_{t} F t -measurable and its future increments are independent of F t \mathcal{F}_{t} F t . The natural filtration of W W W is understood to be completed and right-continuous.
On a measure space ( E , ν ) (E,\nu) ( E , ν ) we identify measurable functions f : E → R f:E\to\mathbb{R} f : E → R that agree ν \nu ν -almost everywhere. For 1 ≤ p < ∞ 1\leq p<\infty 1 ≤ p < ∞ , set
L p ( E , ν ) : = { f : ∥ f ∥ L p ( E , ν ) < ∞ } , ∥ f ∥ L p ( E , ν ) : = ( ∫ E ∣ f ∣ p d ν ) 1 / p . L^{p}(E,\nu):=\{f:\|f\|_{L^{p}(E,\nu)}<\infty\},\qquad\|f\|_{L^{p}(E,\nu)}:=\left(\int_{E}|f|^{p}\,d\nu\right)^{1/p}. L p ( E , ν ) := { f : ∥ f ∥ L p ( E , ν ) < ∞ } , ∥ f ∥ L p ( E , ν ) := ( ∫ E ∣ f ∣ p d ν ) 1/ p .
For p = ∞ p=\infty p = ∞ the norm is the ν \nu ν -essential supremum of ∣ f ∣ |f| ∣ f ∣ . For λ > 0 \lambda>0 λ > 0 , set ρ ( x ) : = e − λ 1 + x 2 \rho(x):=e^{-\lambda\sqrt{1+x^{2}}} ρ ( x ) := e − λ 1 + x 2 . On R \mathbb{R} R we write ∥ f ∥ p : = ∥ f ∥ L p ( R , d x ) \|f\|_{p}:=\|f\|_{L^{p}(\mathbb{R},dx)} ∥ f ∥ p := ∥ f ∥ L p ( R , d x ) and ∥ f ∥ p , ρ : = ∥ f ∥ L p ( R , ρ ( x ) d x ) \|f\|_{p,\rho}:=\|f\|_{L^{p}(\mathbb{R},\rho(x)\,dx)} ∥ f ∥ p , ρ := ∥ f ∥ L p ( R , ρ ( x ) d x ) . For f g ∈ L 1 ( R , d x ) fg\in L^{1}(\mathbb{R},dx) f g ∈ L 1 ( R , d x ) we write
⟨ f , g ⟩ : = ∫ R f ( x ) g ( x ) d x . \langle f,g\rangle:=\int_{\mathbb{R}}f(x)g(x)\,dx. ⟨ f , g ⟩ := ∫ R f ( x ) g ( x ) d x .
Definition 1.1 (Weak and strong solutions). Let u 0 ∈ C t e m u_{0}\in C_{\mathrm{tem}} u 0 ∈ C tem be deterministic, and let u = ( u t ( x ) ) t ≥ 0 , x ∈ R u=(u_{t}(x))_{t\geq0,x\in\mathbb{R}} u = ( u t ( x ) ) t ≥ 0 , x ∈ R be a real-valued predictable random field on some filtered probability space.
(i) We call u u u a weak solution of (1.1) with initial function u 0 u_{0} u 0 if ( t ↦ u t ) ∈ C ( R + ; C t e m ) (t\mapsto u_{t})\in C(\mathbb{R}_{+};C_{\mathrm{tem}}) ( t ↦ u t ) ∈ C ( R + ; C tem ) almost surely and there exists a space-time Gaussian white noise W W W on the same probability space such that, for each ϕ ∈ C c ∞ ( R ) \phi\in C_{c}^{\infty}(\mathbb{R}) ϕ ∈ C c ∞ ( R ) , all integrals below are well defined and the following identity holds almost surely for all t ≥ 0 t\geq0 t ≥ 0 :
⟨ u t , ϕ ⟩ = ⟨ u 0 , ϕ ⟩ + ∫ 0 t [ 1 2 ⟨ u s , ϕ ′ ′ ⟩ + ⟨ b ( s , ⋅ , u s ) , ϕ ⟩ ] d s + ∫ 0 t ∫ R ϕ ( x ) σ ( s , x , u s ( x ) ) W ( d s , d x ) . \langle u_{t},\phi\rangle=\langle u_{0},\phi\rangle+\int_{0}^{t}\left[\frac{1}{2}\langle u_{s},\phi^{\prime\prime}\rangle+\langle b(s,\cdot,u_{s}),\phi\rangle\right]\,ds+\int_{0}^{t}\int_{\mathbb{R}}\phi(x)\sigma(s,x,u_{s}(x))W(ds,dx). ⟨ u t , ϕ ⟩ = ⟨ u 0 , ϕ ⟩ + ∫ 0 t [ 2 1 ⟨ u s , ϕ ′′ ⟩ + ⟨ b ( s , ⋅ , u s ) , ϕ ⟩ ] d s + ∫ 0 t ∫ R ϕ ( x ) σ ( s , x , u s ( x )) W ( d s , d x ) .
(ii) A weak solution u u u driven by W W W is called a strong solution if u u u is adapted to the natural filtration of W W W .
Unless otherwise specified, a solution of (1.1) means a weak solution. We say that pathwise uniqueness holds if any two solutions on the same filtered probability space, driven by the same white noise and with the same initial function, are indistinguishable.
Theorem 1.2. Suppose that (H) holds. Then, for every deterministic u 0 ∈ C t e m u_0 \in C_{\mathrm{tem}} u 0 ∈ C tem , equation (1.1) has a pathwise unique strong solution with initial function u 0 u_0 u 0 .
The joint law of ( u , W ) (u,W) ( u , W ) is therefore uniquely determined by u 0 u_0 u 0 . In particular, consider
∂ t u = 1 2 ∂ x x u + ∣ u ∣ 3 / 4 W ˙ \partial_t u = \frac{1}{2}\partial_{xx}u + |u|^{3/4}\dot{W} ∂ t u = 2 1 ∂ xx u + ∣ u ∣ 3/4 W ˙
with zero initial function. Since ∣ ∣ z ∣ 3 / 4 − ∣ w ∣ 3 / 4 ∣ ≤ ∣ z − w ∣ 3 / 4 \left||z|^{3/4}-|w|^{3/4}\right| \leq|z-w|^{3/4} ∣ z ∣ 3/4 − ∣ w ∣ 3/4 ≤ ∣ z − w ∣ 3/4 and u ≡ 0 u \equiv0 u ≡ 0 is a solution, Theorem 1.2 gives u ≡ 0 u \equiv0 u ≡ 0 almost surely for every real-valued solution.
It is worth noting that we impose no nondegeneracy condition on σ \sigma σ . This is important for branching models, where the noise vanishes at zero population density. Under uniform nondegeneracy, Han [2 ] proved uniqueness in law for stochastic heat equations on bounded intervals with locally Hölder drift and γ \gamma γ -Hölder diffusion, γ > 3 / 4 \gamma> 3/4 γ > 3/4 . His recent work [3 Theorem 1.1] extends uniqueness in law to every positive global Hölder order for bounded, uniformly nondegenerate diffusion in the drift-free scalar equation on the one-dimensional torus.
The short-time linear comparison method
In this subsection, we briefly describe our method for proving pathwise uniqueness. Let u , v u,v u , v be two solutions driven by the same white noise W W W . At each time r r r and position x x x , define the solution difference D r ( x ) D_r(x) D r ( x ) , the noise coefficient difference A r ( x ) A_r(x) A r ( x ) , and the drift difference G r ( x ) G_r(x) G r ( x ) by
D r ( x ) : = u r ( x ) − v r ( x ) , D_r(x) := u_r(x)-v_r(x), D r ( x ) := u r ( x ) − v r ( x ) ,
A r ( x ) : = σ ( r , x , u r ( x ) ) − σ ( r , x , v r ( x ) ) , A_r(x) := \sigma(r,x,u_r(x))-\sigma(r,x,v_r(x)), A r ( x ) := σ ( r , x , u r ( x )) − σ ( r , x , v r ( x )) ,
G r ( x ) : = b ( r , x , u r ( x ) ) − b ( r , x , v r ( x ) ) . G_r(x) := b(r,x,u_r(x))-b(r,x,v_r(x)). G r ( x ) := b ( r , x , u r ( x )) − b ( r , x , v r ( x )) .
Fix λ > 0 \lambda> 0 λ > 0 . For equal initial functions, our goal is to show that E ∥ D t ∥ 1 , ρ = 0 \mathbb{E}\|D_t\|_{1,\rho}=0 E ∥ D t ∥ 1 , ρ = 0 for every t t t .
Mytnik and Perkins [8 ] use the Yamada–Watanabe method [13 ] to prove pathwise uniqueness for γ > 3 / 4 \gamma> 3/4 γ > 3/4 . To control the local-time term arising from Itô’s formula, they obtain bounds of order a n γ − 3 / 4 a_n^{\gamma-3/4} a n γ − 3/4 , where a n ↓ 0 a_n \downarrow0 a n ↓ 0 (see [8, (3.18)]). At γ = 3 / 4 \gamma= 3/4 γ = 3/4 , this factor provides no decay.
To treat the endpoint γ = 3 / 4 \gamma= 3/4 γ = 3/4 , we introduce the short-time linear comparison method . Fix t > 0 t > 0 t > 0 , put h = t / N h=t/N h = t / N and t j = j h t_j=jh t j = j h . On each interval [ t j , t j + 1 ] [t_j,t_{j+1}] [ t j , t j + 1 ] , we introduce the following auxiliary linear equation for the linear comparison process Z j Z^j Z j .
∂ r Z j = 1 2 ∂ x x Z j + G + β ( h ) Z j W ˙ , Z t j j = D t j . \partial_r Z^j = \frac{1}{2}\partial_{xx}Z^j + G + \beta^{(h)}Z^j\dot{W}, \qquad Z^j_{t_j}=D_{t_j}. ∂ r Z j = 2 1 ∂ xx Z j + G + β ( h ) Z j W ˙ , Z t j j = D t j .
Here β ( h ) \beta^{(h)} β ( h ) is the quotient A / D A/D A / D (zero where D = 0 D=0 D = 0 ) truncated to [ − 1 2 h − 1 / 4 , 1 2 h − 1 / 4 ] \left[-\frac{1}{2}h^{-1/4},\frac{1}{2}h^{-1/4}\right] [ − 2 1 h − 1/4 , 2 1 h − 1/4 ] . We compare Z j Z^j Z j with the difference D = u − v D=u-v D = u − v , which satisfies
∂ r D = 1 2 ∂ x x D + G + A W ˙ . \partial_r D = \frac{1}{2}\partial_{xx}D + G + A\dot{W}. ∂ r D = 2 1 ∂ xx D + G + A W ˙ .
Subtracting the equation for Z j Z^j Z j cancels the drift G G G and gives
∂ r ( D − Z j ) = 1 2 ∂ x x ( D − Z j ) + [ β ( h ) ( D − Z j ) + ( A − β ( h ) D ) ] W ˙ , ( D − Z j ) t j = 0. \partial_r(D-Z^j)=\frac{1}{2}\partial_{xx}(D-Z^j)+\left[\beta^{(h)}(D-Z^j)+(A-\beta^{(h)}D)\right]\dot{W}, \qquad(D-Z^j)_{t_j}=0. ∂ r ( D − Z j ) = 2 1 ∂ xx ( D − Z j ) + [ β ( h ) ( D − Z j ) + ( A − β ( h ) D ) ] W ˙ , ( D − Z j ) t j = 0.
We have
E ∥ D t j + 1 ∥ 1 , ρ ≤ E ∥ Z t j + 1 j ∥ 1 , ρ + E ∥ D t j + 1 − Z t j + 1 j ∥ 1 , ρ . \mathbb{E}\|D_{t_{j+1}}\|_{1,\rho}\leq\mathbb{E}\|Z^j_{t_{j+1}}\|_{1,\rho}+\mathbb{E}\|D_{t_{j+1}}-Z^j_{t_{j+1}}\|_{1,\rho}. E ∥ D t j + 1 ∥ 1 , ρ ≤ E ∥ Z t j + 1 j ∥ 1 , ρ + E ∥ D t j + 1 − Z t j + 1 j ∥ 1 , ρ .
Step 1. We first estimate E ∥ Z t j + 1 j ∥ 1 , ρ \mathbb{E}\lVert Z^{j}_{t_{j+1}}\rVert_{1,\rho} E ∥ Z t j + 1 j ∥ 1 , ρ in terms of the difference at t j t_j t j and the drift on [ t j , t j + 1 ] [t_j,t_{j+1}] [ t j , t j + 1 ] . As h ↓ 0 h \downarrow0 h ↓ 0 , the stochastic integral is taken over a shorter time interval, while ∣ β ( h ) ∣ |\beta^{(h)}| ∣ β ( h ) ∣ may grow. Therefore we need a precise estimate that controls both effects. Proposition 2.1(ii) gives
E ∥ Z t j + 1 j ∥ 1 , ρ ≤ e c h E ∥ D t j ∥ 1 , ρ + L ∫ t j t j + 1 e c ( t j + 1 − s ) E ∥ D s ∥ 1 , ρ d s . (2) \mathbb{E}\lVert Z^{j}_{t_{j+1}}\rVert_{1,\rho} \le e^{ch}\mathbb{E}\lVert D_{t_j}\rVert_{1,\rho}+L\int_{t_j}^{t_{j+1}}e^{c(t_{j+1}-s)}\mathbb{E}\lVert D_s\rVert_{1,\rho}\,ds. \tag*{(2)} E ∥ Z t j + 1 j ∥ 1 , ρ ≤ e c h E ∥ D t j ∥ 1 , ρ + L ∫ t j t j + 1 e c ( t j + 1 − s ) E ∥ D s ∥ 1 , ρ d s . (2)
Here c = λ 2 / 2 c=\lambda^2/2 c = λ 2 /2 depends only on the weight ρ \rho ρ , and L L L is the Lipschitz constant of the drift. Both are independent of h h h and β ( h ) \beta^{(h)} β ( h ) .
Step 2. We next show that the sum of the errors ε j \varepsilon_j ε j tends to zero as h ↓ 0 h\downarrow0 h ↓ 0 . Under the bound ∣ A ∣ ≤ K ∣ D ∣ γ |A|\le K|D|^\gamma ∣ A ∣ ≤ K ∣ D ∣ γ with γ ≥ 3 / 4 \gamma\ge3/4 γ ≥ 3/4 , summing the errors over the t / h t/h t / h intervals as in (3.20) gives
∑ j = 0 N − 1 ε j = ( t h ) 3 / 4 o ( h γ ) = o ( h γ − 3 / 4 ) ⟶ 0 as h ↓ 0. (3) \sum_{j=0}^{N-1}\varepsilon_j=\left(\frac{t}{h}\right)^{3/4}o(h^\gamma)=o(h^{\gamma-3/4})\longrightarrow0\qquad\text{as }h\downarrow0. \tag*{(3)} j = 0 ∑ N − 1 ε j = ( h t ) 3/4 o ( h γ ) = o ( h γ − 3/4 ) ⟶ 0 as h ↓ 0. (3)
Substituting (1.2) into the bound for E ∥ D t j + 1 ∥ 1 , ρ \mathbb{E}\lVert D_{t_{j+1}}\rVert_{1,\rho} E ∥ D t j + 1 ∥ 1 , ρ , multiplying by e − c t j + 1 e^{-ct_{j+1}} e − c t j + 1 , and summing over j j j , we obtain, with Y ( s ) = e − c s E ∥ D s ∥ 1 , ρ Y(s)=e^{-cs}\mathbb{E}\lVert D_s\rVert_{1,\rho} Y ( s ) = e − cs E ∥ D s ∥ 1 , ρ ,
Y ( t ) ≤ Y ( 0 ) + L ∫ 0 t Y ( s ) d s + ∑ j = 0 N − 1 e − c t j + 1 ε j . Y(t)\le Y(0)+L\int_0^tY(s)\,ds+\sum_{j=0}^{N-1}e^{-ct_{j+1}}\varepsilon_j. Y ( t ) ≤ Y ( 0 ) + L ∫ 0 t Y ( s ) d s + j = 0 ∑ N − 1 e − c t j + 1 ε j .
The sum tends to zero by (1.3). Since Y ( 0 ) = 0 Y(0)=0 Y ( 0 ) = 0 , Gronwall’s inequality gives Y ( t ) = 0 Y(t)=0 Y ( t ) = 0 for every t t t .
The linear comparison estimate
Let W W W be a space-time Gaussian white noise on the given filtered probability space with filtration ( F r ) r ≥ 0 (\mathcal{F}_r)_{r\geq0} ( F r ) r ≥ 0 . Let ( P r ) r ≥ 0 (P_r)_{r\geq0} ( P r ) r ≥ 0 denote the heat semigroup on R \mathbb{R} R for 1 2 ∂ x x \frac{1}{2}\partial_{xx} 2 1 ∂ xx . Thus P 0 f = f P_0f=f P 0 f = f and, for r > 0 r>0 r > 0 ,
( P r f ) ( x ) = ∫ R p r ( x − y ) f ( y ) d y , p r ( z ) : = ( 2 π r ) − 1 / 2 e − z 2 / ( 2 r ) . (P_rf)(x)=\int_{\mathbb{R}}p_r(x-y)f(y)\,dy,\qquad p_r(z):=(2\pi r)^{-1/2}e^{-z^2/(2r)}. ( P r f ) ( x ) = ∫ R p r ( x − y ) f ( y ) d y , p r ( z ) := ( 2 π r ) − 1/2 e − z 2 / ( 2 r ) .
Fix λ > 0 \lambda>0 λ > 0 and put c : = λ 2 / 2 c:=\lambda^2/2 c := λ 2 /2 . Throughout this section, equalities and inequalities between random functions are understood in L 2 ( R , ρ ( x ) d x ) L^2(\mathbb{R},\rho(x)\,dx) L 2 ( R , ρ ( x ) d x ) almost surely at each fixed time.
Fix 0 ≤ s < t 0\leq s<t 0 ≤ s < t . Let ξ \xi ξ be an F s \mathcal{F}_s F s -measurable L 2 ( R , ρ ( x ) d x ) L^2(\mathbb{R},\rho(x)\,dx) L 2 ( R , ρ ( x ) d x ) -valued random variable, and let G , β G,\beta G , β be predictable real-valued random fields on ( s , t ] × R (s,t]\times\mathbb{R} ( s , t ] × R . Consider the linear equation
Z r ( x ) = ( P r − s ξ ) ( x ) + ∫ s r ( P r − q G q ) ( x ) d q + ∫ s r ∫ R p r − q ( x − y ) β q ( y ) Z q ( y ) W ( d q , d y ) , s ≤ r ≤ t . (4) \begin{aligned}
Z_r(x)={}&(P_{r-s}\xi)(x)+\int_s^r(P_{r-q}G_q)(x)\,dq\\
&+\int_s^r\int_{\mathbb{R}}p_{r-q}(x-y)\beta_q(y)Z_q(y)\,W(dq,dy),\qquad s\leq r\leq t. \tag*{(4)}
\end{aligned} Z r ( x ) = ( P r − s ξ ) ( x ) + ∫ s r ( P r − q G q ) ( x ) d q + ∫ s r ∫ R p r − q ( x − y ) β q ( y ) Z q ( y ) W ( d q , d y ) , s ≤ r ≤ t . (4)
Proposition 2.1 (Linear comparison estimate). Assume that E ∥ ξ ∥ 2 , ρ 2 < ∞ \mathbb{E}\lVert\xi\rVert_{2,\rho}^2<\infty E ∥ ξ ∥ 2 , ρ 2 < ∞ , ∫ s t E ∥ G q ∥ 2 , ρ 2 d q < ∞ \int_s^t\mathbb{E}\lVert G_q\rVert_{2,\rho}^2\,dq<\infty ∫ s t E ∥ G q ∥ 2 , ρ 2 d q < ∞ and ∣ β ∣ ≤ K β |\beta|\leq K_\beta ∣ β ∣ ≤ K β for some deterministic constant K β < ∞ K_\beta<\infty K β < ∞ . Then (2.1) has a unique predictable solution Z Z Z in the class sup s ≤ r ≤ t E ∥ Z r ∥ 2 , ρ 2 < ∞ \sup_{s\leq r\leq t}\mathbb{E}\lVert Z_r\rVert_{2,\rho}^2<\infty sup s ≤ r ≤ t E ∥ Z r ∥ 2 , ρ 2 < ∞ .
(i) If ξ \xi ξ and G G G are nonnegative, then Z r ≥ 0 Z_r\geq0 Z r ≥ 0 for each r ∈ [ s , t ] r\in[s,t] r ∈ [ s , t ] .
(ii) For signed data and each r ∈ [ s , t ] r\in[s,t] r ∈ [ s , t ] , we have
E ∥ Z r ∥ 1 , ρ ≤ e c ( r − s ) E ∥ ξ ∥ 1 , ρ + ∫ s r e c ( r − q ) E ∥ G q ∥ 1 , ρ d q . (5) \mathbb{E}\lVert Z_r\rVert_{1,\rho}\leq e^{c(r-s)}\mathbb{E}\lVert\xi\rVert_{1,\rho}+\int_s^r e^{c(r-q)}\mathbb{E}\lVert G_q\rVert_{1,\rho}\,dq. \tag*{(5)} E ∥ Z r ∥ 1 , ρ ≤ e c ( r − s ) E ∥ ξ ∥ 1 , ρ + ∫ s r e c ( r − q ) E ∥ G q ∥ 1 , ρ d q . (5)
To prove (2.2), we first establish nonnegativity for the equation with spatially smoothed noise and pass to the white-noise limit. For signed data, linearity then bounds ∣ Z ∣ |Z| ∣ Z ∣ by the solution with initial value ∣ ξ ∣ |\xi| ∣ ξ ∣ and drift ∣ G ∣ |G| ∣ G ∣ . Taking expectations in its mild equation removes the stochastic integral and gives (2.2).
Lemma 2.2. For q ≥ 0 q \ge0 q ≥ 0 and f ∈ L 2 ( R , ρ ( x ) d x ) f \in L^2(\mathbb{R},\rho(x)\,dx) f ∈ L 2 ( R , ρ ( x ) d x ) , we have
P q ρ ≤ e c q ρ , ∥ P q f ∥ 2 , ρ 2 ≤ e c q ∥ f ∥ 2 , ρ 2 . P_q\rho\le e^{cq}\rho,\qquad\|P_qf\|_{2,\rho}^2 \le e^{cq}\|f\|_{2,\rho}^2. P q ρ ≤ e c q ρ , ∥ P q f ∥ 2 , ρ 2 ≤ e c q ∥ f ∥ 2 , ρ 2 . Proof. Differentiating p q ( z ) p_q(z) p q ( z ) for q > 0 q > 0 q > 0 , we obtain
∂ q p q ( z ) = ( z 2 2 q 2 − 1 2 q ) p q ( z ) = 1 2 ∂ z z p q ( z ) . (6) \partial_q p_q(z)=\left(\frac{z^2}{2q^2}-\frac{1}{2q}\right)p_q(z)=\frac{1}{2}\partial_{zz}p_q(z). \tag*{(6)} ∂ q p q ( z ) = ( 2 q 2 z 2 − 2 q 1 ) p q ( z ) = 2 1 ∂ z z p q ( z ) . (6) Using integration by parts and (6) , we obtain
∂ q ( P q ρ ) ( x ) = ∫ R ∂ q p q ( x − y ) ρ ( y ) d y = 1 2 ∫ R ∂ y y p q ( x − y ) ρ ( y ) d y = 1 2 ( P q ρ ′ ′ ) ( x ) . (7) \partial_q(P_q\rho)(x)=\int_{\mathbb{R}}\partial_qp_q(x-y)\rho(y)\,dy
=\frac{1}{2}\int_{\mathbb{R}}\partial_{yy}p_q(x-y)\rho(y)\,dy=\frac{1}{2}(P_q\rho^{\prime\prime})(x). \tag*{(7)} ∂ q ( P q ρ ) ( x ) = ∫ R ∂ q p q ( x − y ) ρ ( y ) d y = 2 1 ∫ R ∂ y y p q ( x − y ) ρ ( y ) d y = 2 1 ( P q ρ ′′ ) ( x ) . (7) Differentiating ρ \rho ρ twice, we obtain
ρ ′ ′ ( x ) 2 ρ ( x ) = λ 2 x 2 2 ( 1 + x 2 ) − λ 2 ( 1 + x 2 ) 3 / 2 ≤ c . (8) \frac{\rho^{\prime\prime}(x)}{2\rho(x)}=\frac{\lambda^2x^2}{2(1+x^2)}-\frac{\lambda}{2(1+x^2)^{3/2}}\le c. \tag*{(8)} 2 ρ ( x ) ρ ′′ ( x ) = 2 ( 1 + x 2 ) λ 2 x 2 − 2 ( 1 + x 2 ) 3/2 λ ≤ c . (8) By (7) , (8) and positivity of P q P_q P q , we obtain
∂ q ( e − c q P q ρ ) = e − c q P q ( 1 2 ρ ′ ′ − c ρ ) ≤ 0. \partial_q(e^{-cq}P_q\rho)=e^{-cq}P_q\left(\frac{1}{2}\rho^{\prime\prime}-c\rho\right)\le0. ∂ q ( e − c q P q ρ ) = e − c q P q ( 2 1 ρ ′′ − c ρ ) ≤ 0. By monotonicity and lim q ↓ 0 P q ρ = ρ \lim_{q\downarrow0}P_q\rho=\rho lim q ↓ 0 P q ρ = ρ , we obtain the first bound. Finally, applying Jensen’s inequality and the first bound, we obtain
∥ P q f ∥ 2 , ρ 2 ≤ ∫ R ρ ( x ) P q ( ∣ f ∣ 2 ) ( x ) d x = ⟨ P q ρ , ∣ f ∣ 2 ⟩ ≤ e c q ∥ f ∥ 2 , ρ 2 , \|P_qf\|_{2,\rho}^2\le\int_{\mathbb{R}}\rho(x)P_q(|f|^2)(x)\,dx=\langle P_q\rho,|f|^2\rangle\le e^{cq}\|f\|_{2,\rho}^2, ∥ P q f ∥ 2 , ρ 2 ≤ ∫ R ρ ( x ) P q ( ∣ f ∣ 2 ) ( x ) d x = ⟨ P q ρ , ∣ f ∣ 2 ⟩ ≤ e c q ∥ f ∥ 2 , ρ 2 , which is the second bound.
Lemma 2.3. Fix 0 ≤ s < r 0\le s<r 0 ≤ s < r and let B = ( B q ( y ) ) s < q < r , y ∈ R B=(B_q(y))_{s<q<r,\ y\in\mathbb{R}} B = ( B q ( y ) ) s < q < r , y ∈ R be a real-valued predictable random field satisfying
∫ s r E ∥ B q ∥ 2 , ρ 2 2 π ( r − q ) d q < ∞ . (9) \int_s^r\frac{\mathbb{E}\|B_q\|_{2,\rho}^2}{2\sqrt{\pi(r-q)}}\,dq<\infty. \tag*{(9)} ∫ s r 2 π ( r − q ) E ∥ B q ∥ 2 , ρ 2 d q < ∞. (9) Set H q , x ( y ) : = p r − q ( x − y ) B q ( y ) H_{q,x}(y):=p_{r-q}(x-y)B_q(y) H q , x ( y ) := p r − q ( x − y ) B q ( y ) . Then the integral below is well defined in L 2 ( R , ρ ( x ) d x ) L^2(\mathbb{R},\rho(x)\,dx) L 2 ( R , ρ ( x ) d x ) , and we have
E ∥ ∫ s r ∫ R H q , ⋅ ( y ) W ( d q , d y ) ∥ 2 , ρ 2 = ∫ R ρ ( x ) E ∫ s r ∥ H q , x ∥ 2 2 d q d x \mathbb{E}\left\|\int_s^r\int_{\mathbb{R}}H_{q,\cdot}(y)\,W(dq,dy)\right\|_{2,\rho}^2
=\int_{\mathbb{R}}\rho(x)\mathbb{E}\int_s^r\|H_{q,x}\|_2^2\,dq\,dx E ∫ s r ∫ R H q , ⋅ ( y ) W ( d q , d y ) 2 , ρ 2 = ∫ R ρ ( x ) E ∫ s r ∥ H q , x ∥ 2 2 d q d x ≤ ∫ s r e c ( r − q ) / 2 2 π ( r − q ) E ∥ B q ∥ 2 , ρ 2 d q . (10) \le\int_s^r\frac{e^{c(r-q)/2}}{2\sqrt{\pi(r-q)}}\mathbb{E}\|B_q\|_{2,\rho}^2\,dq. \tag*{(10)} ≤ ∫ s r 2 π ( r − q ) e c ( r − q ) /2 E ∥ B q ∥ 2 , ρ 2 d q . (10) Proof. Using p ℓ 2 = p ℓ / 2 / ( 2 π ℓ ) p_\ell^2=p_{\ell/2}/(2\sqrt{\pi\ell}) p ℓ 2 = p ℓ /2 / ( 2 π ℓ ) and Lemma 2.2, we obtain
∫ R ρ ( x ) p ℓ ( x − y ) 2 d x = ( P ℓ / 2 ρ ) ( y ) 2 π ℓ ≤ e c ℓ / 2 ρ ( y ) 2 π ℓ , ℓ > 0. (11) \int_{\mathbb{R}}\rho(x)p_\ell(x-y)^2\,dx
=\frac{(P_{\ell/2}\rho)(y)}{2\sqrt{\pi\ell}}
\le\frac{e^{c\ell/2}\rho(y)}{2\sqrt{\pi\ell}},\qquad\ell>0. \tag*{(11)} ∫ R ρ ( x ) p ℓ ( x − y ) 2 d x = 2 π ℓ ( P ℓ /2 ρ ) ( y ) ≤ 2 π ℓ e c ℓ /2 ρ ( y ) , ℓ > 0. (11)
Taking ℓ = r − q \ell= r-q ℓ = r − q in (11) gives
∫ s r ∫ R E ∣ B q ( y ) ∣ 2 ( ∫ R ρ ( x ) p r − q ( x − y ) 2 d x ) d y d q ≤ ∫ s r e c ( r − q ) / 2 2 π ( r − q ) E ∥ B q ∥ 2 , ρ 2 d q < ∞ , \int_s^r \int_{\mathbb{R}} \mathbb{E}|B_q(y)|^2\left(\int_{\mathbb{R}}\rho(x)p_{r-q}(x-y)^2\,\mathrm{d}x\right)\mathrm{d}y\,\mathrm{d}q
\leq\int_s^r \frac{e^{c(r-q)/2}}{2\sqrt{\pi(r-q)}}\mathbb{E}\|B_q\|_{2,\rho}^2\,\mathrm{d}q < \infty, ∫ s r ∫ R E ∣ B q ( y ) ∣ 2 ( ∫ R ρ ( x ) p r − q ( x − y ) 2 d x ) d y d q ≤ ∫ s r 2 π ( r − q ) e c ( r − q ) /2 E ∥ B q ∥ 2 , ρ 2 d q < ∞ ,
where the last integral is finite by (9) and e c ( r − q ) / 2 ≤ e c ( r − s ) / 2 e^{c(r-q)/2}\leq e^{c(r-s)/2} e c ( r − q ) /2 ≤ e c ( r − s ) /2 . Since ρ > 0 \rho>0 ρ > 0 , we have E ∫ s r ∥ H q , x ∥ 2 d q < ∞ \mathbb{E}\int_s^r\|H_{q,x}\|^2\,\mathrm{d}q<\infty E ∫ s r ∥ H q , x ∥ 2 d q < ∞ for almost every x x x . The white-noise isometry now gives the equality in (10) .
Lemma 2.4. Fix ε > 0 \varepsilon>0 ε > 0 . Then the random field
( r , z ) ⟼ ∫ 0 r ∫ R p ε ( z − y ) W ( d q , d y ) , ( r , z ) ∈ R + × R , (r,z)\longmapsto\int_0^r\int_{\mathbb{R}}p_\varepsilon(z-y)\,W(\mathrm{d}q,\mathrm{d}y),\qquad(r,z)\in\mathbb{R}_+\times\mathbb{R}, ( r , z ) ⟼ ∫ 0 r ∫ R p ε ( z − y ) W ( d q , d y ) , ( r , z ) ∈ R + × R , admits a jointly continuous version, which we denote by ( W r ε ( z ) ) r ≥ 0 , z ∈ R (W_r^\varepsilon(z))_{r\geq0,z\in\mathbb{R}} ( W r ε ( z ) ) r ≥ 0 , z ∈ R . For each fixed z z z , ( W r ε ( z ) ) r ≥ 0 (W_r^\varepsilon(z))_{r\geq0} ( W r ε ( z ) ) r ≥ 0 is an ( F r ) r ≥ 0 (\mathcal{F}_r)_{r\geq0} ( F r ) r ≥ 0 -Brownian motion with variance rate p 2 ε ( 0 ) p_{2\varepsilon}(0) p 2 ε ( 0 ) .
Proof. Every finite linear combination of the values of the field is a white-noise integral with a deterministic integrand and hence is centered Gaussian. By the white-noise isometry and the heat semigroup property, we obtain
E [ W r ε ( x ) W q ε ( z ) ] = ( r ∧ q ) ∫ R p ε ( x − y ) p ε ( z − y ) d y = ( r ∧ q ) p 2 ε ( x − z ) . \mathbb{E}[W_r^\varepsilon(x)W_q^\varepsilon(z)]=(r\wedge q)\int_{\mathbb{R}}p_\varepsilon(x-y)p_\varepsilon(z-y)\,\mathrm{d}y=(r\wedge q)p_{2\varepsilon}(x-z). E [ W r ε ( x ) W q ε ( z )] = ( r ∧ q ) ∫ R p ε ( x − y ) p ε ( z − y ) d y = ( r ∧ q ) p 2 ε ( x − z ) . For r , q ≤ T r,q\leq T r , q ≤ T , this gives
E ∣ W r ε ( x ) − W q ε ( z ) ∣ 2 = ∣ r − q ∣ p 2 ε ( 0 ) + 2 ( r ∧ q ) ( p 2 ε ( 0 ) − p 2 ε ( x − z ) ) ≤ C ε , T ( ∣ r − q ∣ + ∣ x − z ∣ 2 ) . \mathbb{E}|W_r^\varepsilon(x)-W_q^\varepsilon(z)|^2=|r-q|p_{2\varepsilon}(0)+2(r\wedge q)\bigl(p_{2\varepsilon}(0)-p_{2\varepsilon}(x-z)\bigr)
\leq C_{\varepsilon,T}\bigl(|r-q|+|x-z|^2\bigr). E ∣ W r ε ( x ) − W q ε ( z ) ∣ 2 = ∣ r − q ∣ p 2 ε ( 0 ) + 2 ( r ∧ q ) ( p 2 ε ( 0 ) − p 2 ε ( x − z ) ) ≤ C ε , T ( ∣ r − q ∣ + ∣ x − z ∣ 2 ) . Using Gaussian sixth moments and Kolmogorov’s continuity theorem in ( r , x ) (r,x) ( r , x ) , we obtain the stated jointly continuous version. For fixed z z z and q < r q<r q < r , the increment W r ε ( z ) − W q ε ( z ) W_r^\varepsilon(z)-W_q^\varepsilon(z) W r ε ( z ) − W q ε ( z ) uses only the white noise on ( q , r ] × R (q,r]\times\mathbb{R} ( q , r ] × R . It is independent of F q \mathcal{F}_q F q and has law N ( 0 , ( r − q ) p 2 ε ( 0 ) ) N(0,(r-q)p_{2\varepsilon}(0)) N ( 0 , ( r − q ) p 2 ε ( 0 )) . Since W 0 ε ( z ) = 0 W_0^\varepsilon(z)=0 W 0 ε ( z ) = 0 , this proves the Brownian motion assertion.
We use martingale measures in the sense of Walsh [11 Chapter 2, pp. 287–291] . For each bounded Borel set A ⊂ R A\subset\mathbb{R} A ⊂ R , M ( A ) = ( M r ( A ) ) r ≥ 0 M(A)=(M_r(A))_{r\geq0} M ( A ) = ( M r ( A ) ) r ≥ 0 is a square-integrable martingale starting at zero. At each time r r r , the map A ↦ M r ( A ) A\mapsto M_r(A) A ↦ M r ( A ) is countably additive in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) . We call M M M continuous if each M ( A ) M(A) M ( A ) has continuous paths. For a rectangle D = A × B × ( a , b ] D=A\times B\times(a,b] D = A × B × ( a , b ] , write Q ( D ) : = ⟨ M ( A ) , M ( B ) ⟩ b − ⟨ M ( A ) , M ( B ) ⟩ a Q(D):=\langle M(A),M(B)\rangle_b-\langle M(A),M(B)\rangle_a Q ( D ) := ⟨ M ( A ) , M ( B ) ⟩ b − ⟨ M ( A ) , M ( B ) ⟩ a . The martingale measure is worthy if ∣ Q ( D ) ∣ ≤ K ( D ) |Q(D)|\leq K(D) ∣ Q ( D ) ∣ ≤ K ( D ) on rectangles for a random measure K K K that is symmetric and positive definite in the space variables and has finite expected mass on bounded sets, with r ↦ K ( A × B × ( 0 , r ] ) r\mapsto K(A\times B\times(0,r]) r ↦ K ( A × B × ( 0 , r ]) predictable for each A , B A,B A , B .
For H q ( z ) = η 1 ( a , b ] ( q ) 1 A ( z ) H_q(z)=\eta1_{(a,b]}(q)1_A(z) H q ( z ) = η 1 ( a , b ] ( q ) 1 A ( z ) with η \eta η bounded and F a \mathcal{F}_a F a -measurable, the Walsh integral is η ( M b ( A ) − M a ( A ) ) \eta(M_b(A)-M_a(A)) η ( M b ( A ) − M a ( A )) . For a predictable field H H H on ( 0 , T ] × R (0,T]\times\mathbb{R} ( 0 , T ] × R , set
∥ H ∥ M 2 : = E ∫ R 2 × ( 0 , T ] ∣ H q ( x ) H q ( y ) ∣ K ( d x , d y , d q ) . \|H\|_M^2:=\mathbb{E}\int_{\mathbb{R}^2\times(0,T]}|H_q(x)H_q(y)|K(\mathrm{d}x,\mathrm{d}y,\mathrm{d}q). ∥ H ∥ M 2 := E ∫ R 2 × ( 0 , T ] ∣ H q ( x ) H q ( y ) ∣ K ( d x , d y , d q ) .
When ∥ H ∥ M < ∞ \|H\|_M<\infty ∥ H ∥ M < ∞ , we approximate H H H in this norm by finite sums of elementary fields and take the limit of their integrals in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) (see [11 Chapter 2, pp. 292–295] ).
For ε ≥ 0 \varepsilon\geq0 ε ≥ 0 , r ≥ 0 r\geq0 r ≥ 0 and bounded Borel sets A ⊂ R A\subset\mathbb{R} A ⊂ R , define
M r ε ( A ) : = ∫ 0 r ∫ R ( P ε 1 A ) ( y ) W ( d q , d y ) . M_r^\varepsilon(A):=\int_0^r\int_{\mathbb{R}}(P_\varepsilon1_A)(y)\,W(\mathrm{d}q,\mathrm{d}y). M r ε ( A ) := ∫ 0 r ∫ R ( P ε 1 A ) ( y ) W ( d q , d y ) .
For ε > 0 \varepsilon> 0 ε > 0 this gives the spatial convolution of W W W with p ε p_{\varepsilon} p ε . Since P 0 P_0 P 0 is the identity, M 0 M^0 M 0 is the original white noise W W W . In particular, for 0 ≤ a < b 0 \le a < b 0 ≤ a < b we have
M b 0 ( A ) − M a 0 ( A ) = W ( ( a , b ] × A ) . M_b^0(A) - M_a^0(A) = W((a,b] \times A). M b 0 ( A ) − M a 0 ( A ) = W (( a , b ] × A ) .
Lemma 2.5. For every ε ≥ 0 \varepsilon\ge0 ε ≥ 0 , M ε M^{\varepsilon} M ε is a continuous worthy martingale measure. If 0 ≤ s < r < ∞ 0 \le s < r < \infty 0 ≤ s < r < ∞ and H H H is predictable with E ∫ s r ∥ H q ∥ 2 2 d q < ∞ \mathbb{E}\int_s^r \lVert H_q\rVert_2^2\,dq < \infty E ∫ s r ∥ H q ∥ 2 2 d q < ∞ , then both integrals below are square integrable and
∫ s r ∫ R H q ( z ) M ε ( d q , d z ) = ∫ s r ∫ R ( P ε H q ) ( y ) W ( d q , d y ) a.s. (12) \int_s^r \int_{\mathbb{R}} H_q(z) M^{\varepsilon}(dq,dz) = \int_s^r \int_{\mathbb{R}} (P_{\varepsilon}H_q)(y) W(dq,dy) \quad\text{a.s.} \tag*{(12)} ∫ s r ∫ R H q ( z ) M ε ( d q , d z ) = ∫ s r ∫ R ( P ε H q ) ( y ) W ( d q , d y ) a.s. (12)
Proof. For ε = 0 \varepsilon= 0 ε = 0 , the white-noise martingales M 0 ( A ) M^0(A) M 0 ( A ) have continuous paths and covariations r ∣ A ∩ B ∣ r\lvert A \cap B\rvert r ∣ A ∩ B ∣ . Their covariance measure is supported on the spatial diagonal and serves as a dominating measure. Identity (12) follows from P 0 H q = H q P_0H_q = H_q P 0 H q = H q . Now fix ε > 0 \varepsilon> 0 ε > 0 . For a bounded Borel set A A A , the stochastic Fubini theorem [11 Theorem 2.6] gives
∫ A W r ε ( z ) d z = ∫ A ( ∫ 0 r ∫ R p ε ( z − y ) W ( d q , d y ) ) d z = ∫ 0 r ∫ R ( ∫ A p ε ( z − y ) d z ) W ( d q , d y ) = ∫ 0 r ∫ R ( P ε 1 A ) ( y ) W ( d q , d y ) = M r ε ( A ) . (13) \begin{aligned}
\int_A W_r^{\varepsilon}(z)\,dz
= \int_A \left(\int_0^r \int_{\mathbb{R}} p_{\varepsilon}(z-y) W(dq,dy)\right)dz \\
&= \int_0^r \int_{\mathbb{R}} \left(\int_A p_{\varepsilon}(z-y)\,dz\right) W(dq,dy) \\
&= \int_0^r \int_{\mathbb{R}} (P_{\varepsilon}\mathbf{1}_A)(y) W(dq,dy) = M_r^{\varepsilon}(A). \tag*{(13)}
\end{aligned} ∫ A W r ε ( z ) d z = ∫ A ( ∫ 0 r ∫ R p ε ( z − y ) W ( d q , d y ) ) d z = ∫ 0 r ∫ R ( ∫ A p ε ( z − y ) d z ) W ( d q , d y ) = ∫ 0 r ∫ R ( P ε 1 A ) ( y ) W ( d q , d y ) = M r ε ( A ) . (13)
The integrability condition in [11 Theorem 2.6] is
∫ A ∫ 0 r ∫ R p ε ( z − y ) 2 d y d q d z = r ∣ A ∣ p 2 ε ( 0 ) < ∞ . \int_A \int_0^r \int_{\mathbb{R}} p_{\varepsilon}(z-y)^2\,dy\,dq\,dz = r\lvert A\rvert p_{2\varepsilon}(0) < \infty. ∫ A ∫ 0 r ∫ R p ε ( z − y ) 2 d y d q d z = r ∣ A ∣ p 2 ε ( 0 ) < ∞.
By its white-noise representation, each M ε ( A ) M^{\varepsilon}(A) M ε ( A ) is a continuous square-integrable martingale. For bounded Borel sets A A A , B B B ,
⟨ M ε ( A ) , M ε ( B ) ⟩ r = r ⟨ P ε 1 A , P ε 1 B ⟩ = r ∫ A ∫ B p 2 ε ( z − y ) d y d z . \left\langle M^{\varepsilon}(A), M^{\varepsilon}(B)\right\rangle_r
= r\langle P_{\varepsilon}\mathbf{1}_A, P_{\varepsilon}\mathbf{1}_B\rangle
= r\int_A\int_B p_{2\varepsilon}(z-y)\,dy\,dz. ⟨ M ε ( A ) , M ε ( B ) ⟩ r = r ⟨ P ε 1 A , P ε 1 B ⟩ = r ∫ A ∫ B p 2 ε ( z − y ) d y d z .
The bound E ∣ M r ε ( A ) ∣ 2 = r ∥ P ε 1 A ∥ 2 2 ≤ r ∣ A ∣ \mathbb{E}\lvert M_r^{\varepsilon}(A)\rvert^2 = r\lVert P_{\varepsilon}\mathbf{1}_A\rVert_2^2 \le r\lvert A\rvert E ∣ M r ε ( A ) ∣ 2 = r ∥ P ε 1 A ∥ 2 2 ≤ r ∣ A ∣ gives countable additivity in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) on bounded sets. The covariance measure p 2 ε ( z − y ) d z d y d q p_{2\varepsilon}(z-y)\,dz\,dy\,dq p 2 ε ( z − y ) d z d y d q also serves as the dominating measure, so M ε M^{\varepsilon} M ε is worthy.
For H q ( z ) = η 1 ( a , b ] ( q ) 1 A ( z ) H_q(z) = \eta\mathbf{1}_{(a,b]}(q)\mathbf{1}_A(z) H q ( z ) = η 1 ( a , b ] ( q ) 1 A ( z ) , with s ≤ a < b ≤ r s \le a < b \le r s ≤ a < b ≤ r and η \eta η bounded and F a \mathcal{F}_a F a -measurable, both sides of (12) equal
η ( M b ε ( A ) − M a ε ( A ) ) = η ∫ a b ∫ R ( P ε 1 A ) ( y ) W ( d q , d y ) . \eta\left(M_b^{\varepsilon}(A) - M_a^{\varepsilon}(A)\right)
= \eta\int_a^b\int_{\mathbb{R}} (P_{\varepsilon}\mathbf{1}_A)(y) W(dq,dy). η ( M b ε ( A ) − M a ε ( A ) ) = η ∫ a b ∫ R ( P ε 1 A ) ( y ) W ( d q , d y ) .
For general H H H , extended by zero outside ( s , r ] (s,r] ( s , r ] , the covariance kernel gives
∥ H ∥ M ε 2 = E ∫ s r ∥ P ε ∣ H q ∣ ∥ 2 2 d q ≤ E ∫ s r ∥ H q ∥ 2 2 d q . \lVert H\rVert_{M^{\varepsilon}}^2
= \mathbb{E}\int_s^r \lVert P_{\varepsilon}\lvert H_q\rvert\ \rVert_2^2\,dq
\le\mathbb{E}\int_s^r \lVert H_q\rVert_2^2\,dq. ∥ H ∥ M ε 2 = E ∫ s r ∥ P ε ∣ H q ∣ ∥ 2 2 d q ≤ E ∫ s r ∥ H q ∥ 2 2 d q .
The white-noise isometry and L 2 L^2 L 2 -contractivity of P ε P_{\varepsilon} P ε give the same upper bound for the second moment of the right-hand side of (12) . Approximating H H H in predictable L 2 L^2 L 2 by finite sums of the elementary fields above proves the identity.
We now fix s s s , t t t , ξ \xi ξ , G G G and β \beta β as in the hypotheses of Proposition 2.1. For ε ≥ 0 \varepsilon\ge0 ε ≥ 0 , we consider the mild equation
Z r ε ( x ) = ( P r − s ξ ) ( x ) + ∫ s r ( P r − q G q ) ( x ) d q + ∫ s r ∫ R p r − q ( x − y ) β q ( y ) Z q ε ( y ) M ε ( d q , d y ) , s ≤ r ≤ t . (14) \begin{aligned}
Z_r^{\varepsilon}(x) = (P_{r-s}\xi)(x) + \int_s^r (P_{r-q}G_q)(x)\,dq \\
&+ \int_s^r \int_{\mathbb{R}} p_{r-q}(x-y)\beta_q(y)Z_q^{\varepsilon}(y) M^{\varepsilon}(dq,dy),
\qquad s \le r \le t. \tag*{(14)}
\end{aligned} Z r ε ( x ) = ( P r − s ξ ) ( x ) + ∫ s r ( P r − q G q ) ( x ) d q + ∫ s r ∫ R p r − q ( x − y ) β q ( y ) Z q ε ( y ) M ε ( d q , d y ) , s ≤ r ≤ t . (14)
The stochastic integral in (14) is the Walsh integral of Lemma 2.5.
Lemma 2.6. (i) For each ε ≥ 0 \varepsilon\ge0 ε ≥ 0 , (2.11) has a unique predictable solution Z ε Z^{\varepsilon} Z ε in the class sup s ≤ r ≤ t E ∥ Z r ε ∥ 2 , ρ 2 < ∞ \sup_{s\le r\le t}\mathbb{E}\|Z^{\varepsilon}_{r}\|_{2,\rho}^{2}<\infty sup s ≤ r ≤ t E ∥ Z r ε ∥ 2 , ρ 2 < ∞ . Moreover,
sup ε ≥ 0 sup s ≤ r ≤ t E ∥ Z r ε ∥ 2 , ρ 2 < ∞ . \sup_{\varepsilon\ge0}\sup_{s\le r\le t}\mathbb{E}\|Z^{\varepsilon}_{r}\|_{2,\rho}^{2}<\infty. ε ≥ 0 sup s ≤ r ≤ t sup E ∥ Z r ε ∥ 2 , ρ 2 < ∞. (ii) For each fixed r ∈ [ s , t ] r\in[s,t] r ∈ [ s , t ] , we have
lim ε ↓ 0 E ∥ Z r ε − Z r 0 ∥ 2 , ρ 2 = 0. (15) \lim_{\varepsilon\downarrow0}\mathbb{E}\|Z^{\varepsilon}_{r}-Z^{0}_{r}\|_{2,\rho}^{2}=0. \tag*{(15)} ε ↓ 0 lim E ∥ Z r ε − Z r 0 ∥ 2 , ρ 2 = 0. (15)
Proof. (i) We construct the solution by a contraction argument. For υ > 0 \upsilon>0 υ > 0 and predictable real-valued random fields Y Y Y on [ s , t ] × R [s,t]\times\mathbb{R} [ s , t ] × R , set
∣ ∣ ∣ Y ∣ ∣ ∣ υ : = sup s ≤ r ≤ t e − υ ( r − s ) / 2 ( E ∥ Y r ∥ 2 , ρ 2 ) 1 / 2 , X s , t : = { Y : ∣ ∣ ∣ Y ∣ ∣ ∣ υ < ∞ } . \left|\!\left|\!\left|Y\right|\!\right|\!\right|_{\upsilon}:=\sup_{s\le r\le t}e^{-\upsilon(r-s)/2}\left(\mathbb{E}\|Y_{r}\|_{2,\rho}^{2}\right)^{1/2},\qquad\mathcal{X}_{s,t}:=\{Y:\left|\!\left|\!\left|Y\right|\!\right|\!\right|_{\upsilon}<\infty\}. ∣ ∣ ∣ Y ∣ ∣ ∣ υ := s ≤ r ≤ t sup e − υ ( r − s ) /2 ( E ∥ Y r ∥ 2 , ρ 2 ) 1/2 , X s , t := { Y : ∣ ∣ ∣ Y ∣ ∣ ∣ υ < ∞ } .
Since [ s , t ] [s,t] [ s , t ] is bounded, ∣ ∣ ∣ Y ∣ ∣ ∣ υ < ∞ \left|\!\left|\!\left|Y\right|\!\right|\!\right|_{\upsilon}<\infty ∣ ∣ ∣ Y ∣ ∣ ∣ υ < ∞ is equivalent to sup s ≤ r ≤ t E ∥ Y r ∥ 2 , ρ 2 < ∞ \sup_{s\le r\le t}\mathbb{E}\|Y_{r}\|_{2,\rho}^{2}<\infty sup s ≤ r ≤ t E ∥ Y r ∥ 2 , ρ 2 < ∞ , so X s , t \mathcal{X}_{s,t} X s , t does not depend on υ \upsilon υ . Write (2.11) as Z ε = I + T ε Z ε Z^{\varepsilon}=I+\mathcal{T}_{\varepsilon}Z^{\varepsilon} Z ε = I + T ε Z ε , where
I r : = P r − s ξ + ∫ s r P r − q G q d q , and I_{r}:=P_{r-s}\xi+\int_s^r P_{r-q}G_q\,dq,\qquad\text{and} I r := P r − s ξ + ∫ s r P r − q G q d q , and
( T ε Y ) r ( x ) : = ∫ s r ∫ R p r − q ( x − y ) β q ( y ) Y q ( y ) M ε ( d q , d y ) . (\mathcal{T}_{\varepsilon}Y)_{r}(x):=\int_s^r\int_{\mathbb{R}}p_{r-q}(x-y)\beta_q(y)Y_q(y)\,M^{\varepsilon}(dq,dy). ( T ε Y ) r ( x ) := ∫ s r ∫ R p r − q ( x − y ) β q ( y ) Y q ( y ) M ε ( d q , d y ) .
By Jensen’s inequality and Lemma 2.2, we have
sup s ≤ r ≤ t E ∥ I r ∥ 2 , ρ 2 ≤ 2 sup s ≤ r ≤ t ( E ∥ P r − s ξ ∥ 2 , ρ 2 + ( r − s ) ∫ s r E ∥ P r − q G q ∥ 2 , ρ 2 d q ) \sup_{s\le r\le t}\mathbb{E}\|I_r\|_{2,\rho}^{2}\le2\sup_{s\le r\le t}\left(\mathbb{E}\|P_{r-s}\xi\|_{2,\rho}^{2}+(r-s)\int_s^r\mathbb{E}\|P_{r-q}G_q\|_{2,\rho}^{2}\,dq\right) s ≤ r ≤ t sup E ∥ I r ∥ 2 , ρ 2 ≤ 2 s ≤ r ≤ t sup ( E ∥ P r − s ξ ∥ 2 , ρ 2 + ( r − s ) ∫ s r E ∥ P r − q G q ∥ 2 , ρ 2 d q )
≤ 2 sup s ≤ r ≤ t ( e c ( r − s ) E ∥ ξ ∥ 2 , ρ 2 + ( r − s ) ∫ s r e c ( r − q ) E ∥ G q ∥ 2 , ρ 2 d q ) \le2\sup_{s\le r\le t}\left(e^{c(r-s)}\mathbb{E}\|\xi\|_{2,\rho}^{2}+(r-s)\int_s^r e^{c(r-q)}\mathbb{E}\|G_q\|_{2,\rho}^{2}\,dq\right) ≤ 2 s ≤ r ≤ t sup ( e c ( r − s ) E ∥ ξ ∥ 2 , ρ 2 + ( r − s ) ∫ s r e c ( r − q ) E ∥ G q ∥ 2 , ρ 2 d q )
≤ 2 e c ( t − s ) ( E ∥ ξ ∥ 2 , ρ 2 + ( t − s ) ∫ s t E ∥ G q ∥ 2 , ρ 2 d q ) . \le2e^{c(t-s)}\left(\mathbb{E}\|\xi\|_{2,\rho}^{2}+(t-s)\int_s^t\mathbb{E}\|G_q\|_{2,\rho}^{2}\,dq\right). ≤ 2 e c ( t − s ) ( E ∥ ξ ∥ 2 , ρ 2 + ( t − s ) ∫ s t E ∥ G q ∥ 2 , ρ 2 d q ) .
Hence I ∈ X s , t I\in\mathcal{X}_{s,t} I ∈ X s , t . For Y ∈ X s , t Y\in\mathcal{X}_{s,t} Y ∈ X s , t and s < r ≤ t s<r\le t s < r ≤ t , set H q , x ( y ) : = p r − q ( x − y ) β q ( y ) Y q ( y ) H_{q,x}(y):=p_{r-q}(x-y)\beta_q(y)Y_q(y) H q , x ( y ) := p r − q ( x − y ) β q ( y ) Y q ( y ) . Applying (2.7) with B = β Y B=\beta Y B = β Y and using
∫ s r d q 2 π ( r − q ) = r − s π , \int_s^r\frac{dq}{2\sqrt{\pi(r-q)}}=\sqrt{\frac{r-s}{\pi}}, ∫ s r 2 π ( r − q ) d q = π r − s ,
we obtain
∫ R ρ ( x ) E ∫ s r ∥ H q , x ∥ 2 2 d q d x ≤ K β 2 e c ( r − s ) / 2 r − s π sup s ≤ q < r E ∥ Y q ∥ 2 , ρ 2 < ∞ . (16) \int_{\mathbb{R}}\rho(x)\mathbb{E}\int_s^r\|H_{q,x}\|_{2}^{2}\,dq\,dx\le K_{\beta}^{2}e^{c(r-s)/2}\sqrt{\frac{r-s}{\pi}}\sup_{s\le q<r}\mathbb{E}\|Y_q\|_{2,\rho}^{2}<\infty. \tag*{(16)} ∫ R ρ ( x ) E ∫ s r ∥ H q , x ∥ 2 2 d q d x ≤ K β 2 e c ( r − s ) /2 π r − s s ≤ q < r sup E ∥ Y q ∥ 2 , ρ 2 < ∞. (16)
Since ρ > 0 \rho>0 ρ > 0 , for almost every x x x we may apply (2.9) to H ⋅ , x H_{\cdot,x} H ⋅ , x . By the white-noise isometry, L 2 L^{2} L 2 -contractivity of P ε P_{\varepsilon} P ε and (2.7), we obtain
E ∥ ( T ε Y ) r ∥ 2 , ρ 2 = ∫ R ρ ( x ) E ∫ s r ∥ P ε H q , x ∥ 2 2 d q d x ≤ ∫ R ρ ( x ) E ∫ s r ∥ H q , x ∥ 2 2 d q d x ≤ K β 2 ∫ s r e c ( r − q ) / 2 2 π ( r − q ) E ∥ Y q ∥ 2 , ρ 2 d q . (17) \begin{aligned}
\mathbb{E}\|(\mathcal{T}_{\varepsilon}Y)_r\|_{2,\rho}^{2}=\int_{\mathbb{R}}\rho(x)\mathbb{E}\int_s^r\|P_{\varepsilon}H_{q,x}\|_{2}^{2}\,dq\,dx \\
&\le\int_{\mathbb{R}}\rho(x)\mathbb{E}\int_s^r\|H_{q,x}\|_{2}^{2}\,dq\,dx \\
&\le K_{\beta}^{2}\int_s^r\frac{e^{c(r-q)/2}}{2\sqrt{\pi(r-q)}}\mathbb{E}\|Y_q\|_{2,\rho}^{2}\,dq. \tag*{(17)}
\end{aligned} E ∥ ( T ε Y ) r ∥ 2 , ρ 2 = ∫ R ρ ( x ) E ∫ s r ∥ P ε H q , x ∥ 2 2 d q d x ≤ ∫ R ρ ( x ) E ∫ s r ∥ H q , x ∥ 2 2 d q d x ≤ K β 2 ∫ s r 2 π ( r − q ) e c ( r − q ) /2 E ∥ Y q ∥ 2 , ρ 2 d q . (17)
Choose υ > c / 2 + 4 K β 4 \upsilon>c/2+4K_{\beta}^{4} υ > c /2 + 4 K β 4 . Multiplying (2.14) by e − υ ( r − s ) e^{-\upsilon(r-s)} e − υ ( r − s ) and using E ∥ Y q ∥ 2 , ρ 2 ≤ e υ ( q − s ) ∣ ∣ ∣ Y ∣ ∣ ∣ υ 2 \mathbb{E}\|Y_q\|_{2,\rho}^{2}\le e^{\upsilon(q-s)}\left|\!\left|\!\left|Y\right|\!\right|\!\right|_{\upsilon}^{2} E ∥ Y q ∥ 2 , ρ 2 ≤ e υ ( q − s ) ∣ ∣ ∣ Y ∣ ∣ ∣ υ 2 , we obtain
∥ T ε Y ∥ υ 2 ≤ K β 2 sup s ≤ r ≤ t e − υ ( r − s ) ∫ s r e c ( r − q ) / 2 2 π ( r − q ) E ∥ Y q ∥ 2 , ρ 2 d q ≤ K β 2 ∥ Y ∥ υ 2 sup s ≤ r ≤ t ∫ 0 r − s e − ( υ − c / 2 ) ℓ 2 π ℓ d ℓ ≤ K β 2 ∥ Y ∥ υ 2 ∫ 0 ∞ e − ( υ − c / 2 ) ℓ 2 π ℓ d ℓ = K β 2 2 υ − c / 2 ∥ Y ∥ υ 2 ≤ 1 4 ∥ Y ∥ υ 2 . \begin{aligned}
\lVert T_\varepsilon Y\rVert_\upsilon^2 \le K_\beta^2 \sup_{s\le r\le t} e^{-\upsilon(r-s)} \int_s^r \frac{e^{c(r-q)/2}}{2\sqrt{\pi(r-q)}}\mathbb{E}\lVert Y_q\rVert_{2,\rho}^2\,dq \\
&\le K_\beta^2\lVert Y\rVert_\upsilon^2 \sup_{s\le r\le t}\int_0^{r-s}\frac{e^{-(\upsilon-c/2)\ell}}{2\sqrt{\pi\ell}}\,d\ell\\
&\le K_\beta^2\lVert Y\rVert_\upsilon^2\int_0^\infty\frac{e^{-(\upsilon-c/2)\ell}}{2\sqrt{\pi\ell}}\,d\ell\\
&= \frac{K_\beta^2}{2\sqrt{\upsilon-c/2}}\lVert Y\rVert_\upsilon^2\le\frac14\lVert Y\rVert_\upsilon^2.
\end{aligned} ∥ T ε Y ∥ υ 2 ≤ K β 2 s ≤ r ≤ t sup e − υ ( r − s ) ∫ s r 2 π ( r − q ) e c ( r − q ) /2 E ∥ Y q ∥ 2 , ρ 2 d q ≤ K β 2 ∥ Y ∥ υ 2 s ≤ r ≤ t sup ∫ 0 r − s 2 π ℓ e − ( υ − c /2 ) ℓ d ℓ ≤ K β 2 ∥ Y ∥ υ 2 ∫ 0 ∞ 2 π ℓ e − ( υ − c /2 ) ℓ d ℓ = 2 υ − c /2 K β 2 ∥ Y ∥ υ 2 ≤ 4 1 ∥ Y ∥ υ 2 .
For Y 1 , Y 2 ∈ X s , t Y_1,Y_2\in\mathcal{X}_{s,t} Y 1 , Y 2 ∈ X s , t , linearity gives
∥ ( I + T ε Y 1 ) − ( I + T ε Y 2 ) ∥ υ = ∥ T ε ( Y 1 − Y 2 ) ∥ υ ≤ 1 2 ∥ Y 1 − Y 2 ∥ υ . \lVert(I+T_\varepsilon Y_1)-(I+T_\varepsilon Y_2)\rVert_\upsilon=\lVert T_\varepsilon(Y_1-Y_2)\rVert_\upsilon\le\frac12\lVert Y_1-Y_2\rVert_\upsilon. ∥( I + T ε Y 1 ) − ( I + T ε Y 2 ) ∥ υ = ∥ T ε ( Y 1 − Y 2 ) ∥ υ ≤ 2 1 ∥ Y 1 − Y 2 ∥ υ .
Thus Y ↦ I + T ε Y Y\mapsto I+T_\varepsilon Y Y ↦ I + T ε Y is a contraction of the Banach space X s , t \mathcal{X}_{s,t} X s , t into itself. Banach’s fixed point theorem gives the unique fixed point Z ε = I + T ε Z ε Z^\varepsilon=I+T_\varepsilon Z^\varepsilon Z ε = I + T ε Z ε . In particular, ∥ Z ε ∥ υ ≤ 2 ∥ I ∥ υ \lVert Z^\varepsilon\rVert_\upsilon\le2\lVert I\rVert_\upsilon ∥ Z ε ∥ υ ≤ 2 ∥ I ∥ υ for every ε ≥ 0 \varepsilon\ge0 ε ≥ 0 , which gives the uniform second-moment bound.
(ii) Write Z : = Z 0 Z:=Z^0 Z := Z 0 . Subtracting (14) at ε = 0 \varepsilon=0 ε = 0 from the same equation at ε > 0 \varepsilon>0 ε > 0 , we obtain
Z ε − Z = T ε ( Z ε − Z ) + ( T ε − T 0 ) Z . Z^\varepsilon-Z=T_\varepsilon(Z^\varepsilon-Z)+(T_\varepsilon-T_0)Z. Z ε − Z = T ε ( Z ε − Z ) + ( T ε − T 0 ) Z .
Fix s < r ≤ t s<r\le t s < r ≤ t and take Y = Z Y=Z Y = Z in the definition of H q , x H_{q,x} H q , x in (i). By (12) and the white-noise isometry, we obtain
E ∥ [ ( T ε − T 0 ) Z ] r ∥ 2 , ρ 2 = ∫ s r ∫ R ρ ( x ) E ∥ P ε H q , x − H q , x ∥ 2 2 d x d q ⟶ 0 ( ε ↓ 0 ) . \mathbb{E}\lVert[(T_\varepsilon-T_0)Z]_r\rVert_{2,\rho}^2=\int_s^r\int_{\mathbb{R}}\rho(x)\mathbb{E}\lVert P_\varepsilon H_{q,x}-H_{q,x}\rVert_2^2\,dx\,dq\longrightarrow0\qquad(\varepsilon\downarrow0). E ∥[( T ε − T 0 ) Z ] r ∥ 2 , ρ 2 = ∫ s r ∫ R ρ ( x ) E ∥ P ε H q , x − H q , x ∥ 2 2 d x d q ⟶ 0 ( ε ↓ 0 ) .
Indeed, by strong continuity on L 2 ( R , d y ) L^2(\mathbb{R},dy) L 2 ( R , d y ) , the squared difference tends to zero for almost every ( ω , q , x ) (\omega,q,x) ( ω , q , x ) . By contractivity it is bounded by 4 ∥ H q , x ∥ 2 2 4\lVert H_{q,x}\rVert_2^2 4 ∥ H q , x ∥ 2 2 , which is integrable by (16) . It remains to control T ε ( Z ε − Z ) T_\varepsilon(Z^\varepsilon-Z) T ε ( Z ε − Z ) . For any sequence ε n ↓ 0 \varepsilon_n\downarrow0 ε n ↓ 0 , set d ( r ) : = lim sup n → ∞ E ∥ Z r ε n − Z r ∥ 2 , ρ 2 d(r):=\limsup_{n\to\infty}\mathbb{E}\lVert Z_r^{\varepsilon_n}-Z_r\rVert_{2,\rho}^2 d ( r ) := lim sup n → ∞ E ∥ Z r ε n − Z r ∥ 2 , ρ 2 . Since the other term tends to zero in mean square, the L 2 L^2 L 2 triangle inequality and (17) give
d ( r ) ≤ K β 2 lim sup n → ∞ ∫ s r e c ( r − q ) / 2 2 π ( r − q ) E ∥ Z q ε n − Z q ∥ 2 , ρ 2 d q ≤ K β 2 ∫ s r e c ( r − q ) / 2 2 π ( r − q ) d ( q ) d q . \begin{aligned}
d(r)\le K_\beta^2\limsup_{n\to\infty}\int_s^r\frac{e^{c(r-q)/2}}{2\sqrt{\pi(r-q)}}\mathbb{E}\lVert Z_q^{\varepsilon_n}-Z_q\rVert_{2,\rho}^2\,dq \\
&\le K_\beta^2\int_s^r\frac{e^{c(r-q)/2}}{2\sqrt{\pi(r-q)}}d(q)\,dq.
\end{aligned} d ( r ) ≤ K β 2 n → ∞ lim sup ∫ s r 2 π ( r − q ) e c ( r − q ) /2 E ∥ Z q ε n − Z q ∥ 2 , ρ 2 d q ≤ K β 2 ∫ s r 2 π ( r − q ) e c ( r − q ) /2 d ( q ) d q .
The last inequality uses the integrable bound C / r − q C/\sqrt{r-q} C / r − q from the uniform second-moment estimate in (i). Using the same exponential weight as in (i), we obtain
sup s ≤ r ≤ t e − υ ( r − s ) d ( r ) ≤ 1 4 sup s ≤ r ≤ t e − υ ( r − s ) d ( r ) . \sup_{s\le r\le t}e^{-\upsilon(r-s)}d(r)\le\frac14\sup_{s\le r\le t}e^{-\upsilon(r-s)}d(r). s ≤ r ≤ t sup e − υ ( r − s ) d ( r ) ≤ 4 1 s ≤ r ≤ t sup e − υ ( r − s ) d ( r ) .
Since d d d is nonnegative and bounded, we have d ( r ) = 0 d(r)=0 d ( r ) = 0 for every r r r . The sequence ε n ↓ 0 \varepsilon_n\downarrow0 ε n ↓ 0 was arbitrary, which proves (15) .
Lemma 2.7. Under the assumptions of Proposition 2.1, fix ε > 0 \varepsilon>0 ε > 0 and let Z ε Z^\varepsilon Z ε solve (14) . If ξ \xi ξ and G G G are nonnegative, then Z r ε ≥ 0 Z_r^\varepsilon\ge0 Z r ε ≥ 0 for each r ∈ [ s , t ] r\in[s,t] r ∈ [ s , t ] .
Proof. To detect the negative part, set z − : = max { − z , 0 } z^-:=\max\{-z,0\} z − := max { − z , 0 } and ψ ( z ) : = ( z − ) 3 / ( 1 + z − ) 2 \psi(z):=(z^-)^3/(1+z^-)^2 ψ ( z ) := ( z − ) 3 / ( 1 + z − ) 2 . Then ψ ∈ C 2 ( R ) \psi\in C^2(\mathbb{R}) ψ ∈ C 2 ( R ) , ψ ( z ) = 0 \psi(z)=0 ψ ( z ) = 0 exactly when z ≥ 0 z\ge0 z ≥ 0 , and − 1 ≤ ψ ′ ≤ 0 -1\le\psi'\le0 − 1 ≤ ψ ′ ≤ 0 . Since ψ ′ ′ ( z ) = 6 z − / ( 1 + z − ) 4 \psi''(z)=6z^-/(1+z^-)^4 ψ ′′ ( z ) = 6 z − / ( 1 + z − ) 4 , we have
0 ≤ z 2 ψ ′ ′ ( z ) = 6 ψ ( z ) ( 1 + z − ) 2 ≤ 6 ψ ( z ) . (18) 0\le z^2\psi''(z)=\frac{6\psi(z)}{(1+z^-)^2}\le6\psi(z). \tag*{(18)} 0 ≤ z 2 ψ ′′ ( z ) = ( 1 + z − ) 2 6 ψ ( z ) ≤ 6 ψ ( z ) . (18)
Let K q ( x , y ) : = β q ( x ) Z q ε ( x ) p ε ( x − y ) K_q(x,y) := \beta_q(x)Z_q^\varepsilon(x)p_\varepsilon(x-y) K q ( x , y ) := β q ( x ) Z q ε ( x ) p ε ( x − y ) . By (12) , the stochastic integral in (14) has integrand P r − q x K q P^x_{r-q}K_q P r − q x K q against W W W , where P h x P^x_h P h x acts on the first space variable. Since ∥ p ε ∥ 2 2 = 1 / ( 2 π ε ) \|p_\varepsilon\|_2^2 = 1/(2\sqrt{\pi\varepsilon}) ∥ p ε ∥ 2 2 = 1/ ( 2 π ε ) , Lemma 2.6 (i) gives
E ∫ s t ∫ R ρ ( x ) ∥ K q ( x , ⋅ ) ∥ 2 2 d x d q ≤ K β 2 2 π ε ∫ s t E ∥ Z q ε ∥ 2 , ρ 2 d q < ∞ . \mathbb{E}\int_s^t\int_{\mathbb{R}}\rho(x)\|K_q(x,\cdot)\|_2^2\,dx\,dq \le\frac{K_\beta^2}{2\sqrt{\pi\varepsilon}}\int_s^t\mathbb{E}\|Z_q^\varepsilon\|_{2,\rho}^2\,dq < \infty. E ∫ s t ∫ R ρ ( x ) ∥ K q ( x , ⋅ ) ∥ 2 2 d x d q ≤ 2 π ε K β 2 ∫ s t E ∥ Z q ε ∥ 2 , ρ 2 d q < ∞.
For δ > 0 \delta>0 δ > 0 , set Z ~ q : = P δ Z q ε \widetilde{Z}_q:=P_\delta Z_q^\varepsilon Z q := P δ Z q ε and K q δ : = P δ x K q K_q^\delta:=P_\delta^xK_q K q δ := P δ x K q . Applying P δ P_\delta P δ to (14) and using the semigroup property, we obtain
Z ~ r ( x ) = ( P δ + r − s ξ ) ( x ) + ∫ s r ( P δ + r − q G q ) ( x ) d q + ∫ s r ∫ R ( P δ + r − q x K q ) ( x , y ) W ( d q , d y ) . \widetilde{Z}_r(x)=(P_{\delta+r-s}\xi)(x)+\int_s^r(P_{\delta+r-q}G_q)(x)\,dq+\int_s^r\int_{\mathbb{R}}(P_{\delta+r-q}^xK_q)(x,y)\,W(dq,dy). Z r ( x ) = ( P δ + r − s ξ ) ( x ) + ∫ s r ( P δ + r − q G q ) ( x ) d q + ∫ s r ∫ R ( P δ + r − q x K q ) ( x , y ) W ( d q , d y ) .
By the heat equation, we have
P δ + r − q = P δ + 1 2 ∫ q r ∂ x x P δ + ℓ − q d ℓ . P_{\delta+r-q}=P_\delta+\frac{1}{2}\int_q^r\partial_{xx}P_{\delta+\ell-q}\,d\ell. P δ + r − q = P δ + 2 1 ∫ q r ∂ xx P δ + ℓ − q d ℓ .
Fix x ∈ R x\in\mathbb{R} x ∈ R . Substituting this identity and applying the stochastic Fubini theorem [11 Theorem 2.6] , we obtain a version of ( Z ~ r ( x ) ) s ≤ r ≤ t (\widetilde{Z}_r(x))_{s\le r\le t} ( Z r ( x ) ) s ≤ r ≤ t satisfying
Z ~ r ( x ) = ( P δ ξ ) ( x ) + ∫ s r ( 1 2 ∂ x x Z ~ q ( x ) + ( P δ G q ) ( x ) ) d q + N r x . (19) \widetilde{Z}_r(x)=(P_\delta\xi)(x)+\int_s^r\left(\frac{1}{2}\partial_{xx}\widetilde{Z}_q(x)+(P_\delta G_q)(x)\right)\,dq+N_r^x. \tag*{(19)} Z r ( x ) = ( P δ ξ ) ( x ) + ∫ s r ( 2 1 ∂ xx Z q ( x ) + ( P δ G q ) ( x ) ) d q + N r x . (19)
N r x : = ∫ s r ∫ R K q δ ( x , y ) W ( d q , d y ) . N_r^x:=\int_s^r\int_{\mathbb{R}}K_q^\delta(x,y)\,W(dq,dy). N r x := ∫ s r ∫ R K q δ ( x , y ) W ( d q , d y ) .
For the spatial interchange, f ↦ ( P δ f ) ( x ) f\mapsto(P_\delta f)(x) f ↦ ( P δ f ) ( x ) is bounded on L 2 ( R , ρ ( z ) d z ) L^2(\mathbb{R},\rho(z)\,dz) L 2 ( R , ρ ( z ) d z ) and Lemma 2.2 applies to P r − q x K q P^x_{r-q}K_q P r − q x K q . For the time interchange, we use the square integrability of K K K above and
sup δ ≤ h ≤ δ + t − s ∫ R ∣ ∂ x x p h ( x − z ) ∣ 2 / ρ ( z ) d z < ∞ . \sup_{\delta\le h\le\delta+t-s}\int_{\mathbb{R}}|\partial_{xx}p_h(x-z)|^2/\rho(z)\,dz<\infty. δ ≤ h ≤ δ + t − s sup ∫ R ∣ ∂ xx p h ( x − z ) ∣ 2 / ρ ( z ) d z < ∞.
The same pointwise bounds give an absolutely integrable drift and E ∫ s t ∥ K q δ ( x , ⋅ ) ∥ 2 2 d q < ∞ \mathbb{E}\int_s^t\|K_q^\delta(x,\cdot)\|_2^2\,dq<\infty E ∫ s t ∥ K q δ ( x , ⋅ ) ∥ 2 2 d q < ∞ . Thus N x N^x N x is a real-valued continuous square-integrable martingale with [ N x ] r = ∫ s r ∥ K q δ ( x , ⋅ ) ∥ 2 2 d q [N^x]_r=\int_s^r\|K_q^\delta(x,\cdot)\|_2^2\,dq [ N x ] r = ∫ s r ∥ K q δ ( x , ⋅ ) ∥ 2 2 d q (see [11 Theorem 2.5, (2.10)] ). Applying the one-dimensional Itô formula to the continuous semimartingale in (19) , we obtain
ψ ( Z ~ r ( x ) ) − ψ ( Z ~ s ( x ) ) = ∫ s r ψ ′ ( Z ~ q ( x ) ) ( 1 2 ∂ x x Z ~ q ( x ) + ( P δ G q ) ( x ) ) d q + ∫ s r ψ ′ ( Z ~ q ( x ) ) d N q x + 1 2 ∫ s r ψ ′ ′ ( Z ~ q ( x ) ) ∥ K q δ ( x , ⋅ ) ∥ 2 2 d q . \begin{aligned}
\psi(\widetilde{Z}_r(x))-\psi(\widetilde{Z}_s(x))
={}&\int_s^r\psi'(\widetilde{Z}_q(x))\left(\frac{1}{2}\partial_{xx}\widetilde{Z}_q(x)+(P_\delta G_q)(x)\right)\,dq\\
&+\int_s^r\psi'(\widetilde{Z}_q(x))\,dN_q^x
+\frac{1}{2}\int_s^r\psi''(\widetilde{Z}_q(x))\|K_q^\delta(x,\cdot)\|_2^2\,dq.
\end{aligned} ψ ( Z r ( x )) − ψ ( Z s ( x )) = ∫ s r ψ ′ ( Z q ( x )) ( 2 1 ∂ xx Z q ( x ) + ( P δ G q ) ( x ) ) d q + ∫ s r ψ ′ ( Z q ( x )) d N q x + 2 1 ∫ s r ψ ′′ ( Z q ( x )) ∥ K q δ ( x , ⋅ ) ∥ 2 2 d q .
We now take expectations and integrate against ρ ( x ) d x \rho(x)\,dx ρ ( x ) d x . The integral against N x N^x N x has mean zero since ∣ ψ ′ ∣ ≤ 1 |\psi'|\le1 ∣ ψ ′ ∣ ≤ 1 . We have ψ ( P δ ξ ) = 0 \psi(P_\delta\xi)=0 ψ ( P δ ξ ) = 0 and ψ ′ ( Z ~ q ) P δ G q ≤ 0 \psi'(\widetilde{Z}_q)P_\delta G_q\le0 ψ ′ ( Z q ) P δ G q ≤ 0 . Using integration by parts and (8) , we obtain
1 2 ⟨ ρ , ψ ′ ( Z ~ q ) ∂ x x Z ~ q ⟩ = 1 2 ⟨ ρ ′ ′ , ψ ( Z ~ q ) ⟩ − 1 2 ⟨ ρ , ψ ′ ′ ( Z ~ q ) ∣ ∂ x Z ~ q ∣ 2 ⟩ ≤ c ⟨ ρ , ψ ( Z ~ q ) ⟩ . \frac{1}{2}\langle\rho,\psi'(\widetilde{Z}_q)\partial_{xx}\widetilde{Z}_q\rangle
=\frac{1}{2}\langle\rho'',\psi(\widetilde{Z}_q)\rangle-\frac{1}{2}\langle\rho,\psi''(\widetilde{Z}_q)|\partial_x\widetilde{Z}_q|^2\rangle
\le c\langle\rho,\psi(\widetilde{Z}_q)\rangle. 2 1 ⟨ ρ , ψ ′ ( Z q ) ∂ xx Z q ⟩ = 2 1 ⟨ ρ ′′ , ψ ( Z q )⟩ − 2 1 ⟨ ρ , ψ ′′ ( Z q ) ∣ ∂ x Z q ∣ 2 ⟩ ≤ c ⟨ ρ , ψ ( Z q )⟩ .
Here we may remove compact cutoffs since ∣ ρ ′ ∣ + ∣ ρ ′ ′ ∣ ≤ C λ ρ |\rho'|+|\rho''|\le C_\lambda\rho ∣ ρ ′ ∣ + ∣ ρ ′′ ∣ ≤ C λ ρ and ∣ ∂ x j p δ ∣ ≤ C δ , j p 2 δ |\partial_x^j p_\delta|\le C_{\delta,j}p_{2\delta} ∣ ∂ x j p δ ∣ ≤ C δ , j p 2 δ for j = 1 , 2 j=1,2 j = 1 , 2 . It follows that
E ⟨ ρ , ψ ( Z ~ r ) ⟩ ≤ c ∫ s r E ⟨ ρ , ψ ( Z ~ q ) ⟩ d q + 1 2 E ∫ s r ∫ R ρ ( x ) ψ ′ ′ ( Z ~ q ( x ) ) ∥ K q δ ( x , ⋅ ) ∥ 2 2 d x d q . \mathbb{E}\langle\rho,\psi(\widetilde{Z}_r)\rangle
\le c\int_s^r\mathbb{E}\langle\rho,\psi(\widetilde{Z}_q)\rangle\,dq
+\frac{1}{2}\mathbb{E}\int_s^r\int_{\mathbb{R}}\rho(x)\psi''(\widetilde{Z}_q(x))\|K_q^\delta(x,\cdot)\|_2^2\,dx\,dq. E ⟨ ρ , ψ ( Z r )⟩ ≤ c ∫ s r E ⟨ ρ , ψ ( Z q )⟩ d q + 2 1 E ∫ s r ∫ R ρ ( x ) ψ ′′ ( Z q ( x )) ∥ K q δ ( x , ⋅ ) ∥ 2 2 d x d q .
By density and Lemma 2.2 , P δ P_\delta P δ converges strongly to the identity in L 2 ( R , ρ ( x ) d x ) L^2(\mathbb{R},\rho(x)\,dx) L 2 ( R , ρ ( x ) d x ) , also for L 2 ( R , d y ) L^2(\mathbb{R},dy) L 2 ( R , d y ) -valued functions. Thus Z ~ q → Z q ε \widetilde{Z}_q\to Z_q^\varepsilon Z q → Z q ε in weighted mean square at each q q q and after integration in q q q . Likewise K δ → K K^\delta\to K K δ → K in weighted mean square after integration in q q q , so ∥ K q δ ( x , ⋅ ) ∥ 2 2 → ∥ K q ( x , ⋅ ) ∥ 2 2 \|K_q^\delta(x,\cdot)\|_2^2\to\|K_q(x,\cdot)\|_2^2 ∥ K q δ ( x , ⋅ ) ∥ 2 2 → ∥ K q ( x , ⋅ ) ∥ 2 2 in L 1 ( d P d q ρ ( x ) d x ) L^1(d\mathbb{P}\,dq\,\rho(x)\,dx) L 1 ( d P d q ρ ( x ) d x ) . Since ψ ′ ′ \psi'' ψ ′′ is bounded and continuous, ψ ′ ′ ( Z ~ ) → ψ ′ ′ ( Z ε ) \psi''(\widetilde{Z})\to\psi''(Z^\varepsilon) ψ ′′ ( Z ) → ψ ′′ ( Z ε ) in measure, which justifies passing to the limit in the Itô correction. Set m ( r ) : = E ⟨ ρ , ψ ( Z r ε ) ⟩ m(r) := \mathbb{E}\langle\rho,\psi(Z_r^\varepsilon)\rangle m ( r ) := E ⟨ ρ , ψ ( Z r ε )⟩ . Letting δ ↓ 0 \delta\downarrow0 δ ↓ 0 and using ∥ K q ( x , ⋅ ) ∥ 2 2 = ∣ β q ( x ) Z q ε ( x ) ∣ 2 / ( 2 π ε ) \|K_q(x,\cdot)\|_2^2 = |\beta_q(x)Z_q^\varepsilon(x)|^2/(2\sqrt{\pi\varepsilon}) ∥ K q ( x , ⋅ ) ∥ 2 2 = ∣ β q ( x ) Z q ε ( x ) ∣ 2 / ( 2 π ε ) and (18) , we obtain
m ( r ) ≤ ( c + 3 K β 2 2 π ε ) ∫ s r m ( q ) d q . m(r) \le\left(c+\frac{3K_\beta^2}{2\sqrt{\pi\varepsilon}}\right)\int_s^r m(q)\,dq. m ( r ) ≤ ( c + 2 π ε 3 K β 2 ) ∫ s r m ( q ) d q .
Gronwall’s inequality gives m ( r ) = 0 m(r)=0 m ( r ) = 0 . Since ρ > 0 \rho>0 ρ > 0 and ψ ( z ) > 0 \psi(z)>0 ψ ( z ) > 0 for z < 0 z<0 z < 0 , we obtain Z r ε ≥ 0 Z_r^\varepsilon\geq0 Z r ε ≥ 0 .
Proof of Proposition 2.1. Since M 0 = W M^0=W M 0 = W , equation (14) at ε = 0 \varepsilon=0 ε = 0 coincides with (4) . Lemma 2.6(i) therefore gives the asserted existence and uniqueness, with Z = Z 0 Z=Z^0 Z = Z 0 .
( i ) (i) ( i ) Fix r ∈ [ s , t ] r\in[s,t] r ∈ [ s , t ] . For ξ , G ≥ 0 \xi,G\geq0 ξ , G ≥ 0 , Lemma 2.7 gives Z r ε ≥ 0 Z_r^\varepsilon\geq0 Z r ε ≥ 0 for every ε > 0 \varepsilon>0 ε > 0 . Recall that Z r − Z_r^- Z r − denotes the negative part of Z r Z_r Z r . By (15) , we have
E ∥ Z r − ∥ 2 , ρ 2 ≤ E ∥ Z r − Z r ε ∥ 2 , ρ 2 ⟶ 0. \mathbb{E}\|Z_r^-\|_{2,\rho}^2\leq\mathbb{E}\|Z_r-Z_r^\varepsilon\|_{2,\rho}^2\longrightarrow0. E ∥ Z r − ∥ 2 , ρ 2 ≤ E ∥ Z r − Z r ε ∥ 2 , ρ 2 ⟶ 0.
Hence Z r ≥ 0 Z_r\geq0 Z r ≥ 0 .
( i i ) (ii) ( ii ) For signed data, apply the existence assertion just proved to (4) with initial value ∣ ξ ∣ |\xi| ∣ ξ ∣ , drift ∣ G ∣ |G| ∣ G ∣ and the same β \beta β and W W W . Denote the resulting solution by Z ‾ \overline{Z} Z . By linearity, Z ‾ ± Z \overline{Z}\pm Z Z ± Z solve the equations with nonnegative data ∣ ξ ∣ ± ξ |\xi|\pm\xi ∣ ξ ∣ ± ξ and ∣ G ∣ ± G |G|\pm G ∣ G ∣ ± G . Applying (i) to both gives ∣ Z r ∣ ≤ Z ‾ r |Z_r|\leq\overline{Z}_r ∣ Z r ∣ ≤ Z r . Pairing the mild equation for Z ‾ r \overline{Z}_r Z r with ρ \rho ρ and taking expectations, we obtain
E ∥ Z r ∥ 1 , ρ ≤ E ⟨ ρ , Z ‾ r ⟩ = E ⟨ P r − s ρ , ∣ ξ ∣ ⟩ + ∫ s r E ⟨ P r − q ρ , ∣ G q ∣ ⟩ d q . \mathbb{E}\|Z_r\|_{1,\rho}\leq\mathbb{E}\langle\rho,\overline{Z}_r\rangle
=\mathbb{E}\langle P_{r-s}\rho,|\xi|\rangle+
\int_s^r\mathbb{E}\langle P_{r-q}\rho,|G_q|\rangle\,dq. E ∥ Z r ∥ 1 , ρ ≤ E ⟨ ρ , Z r ⟩ = E ⟨ P r − s ρ , ∣ ξ ∣ ⟩ + ∫ s r E ⟨ P r − q ρ , ∣ G q ∣ ⟩ d q .
By Lemma 2.3, the stochastic convolution is centered and square integrable in L 2 ( R , ρ ( x ) d x ) L^2(\mathbb{R},\rho(x)\,dx) L 2 ( R , ρ ( x ) d x ) . Hence its pairing with ρ \rho ρ has mean zero, since ∣ ⟨ ρ , f ⟩ ∣ ≤ ∥ ρ ∥ 1 1 / 2 ∥ f ∥ 2 , ρ |\langle\rho,f\rangle|\leq\|\rho\|_1^{1/2}\|f\|_{2,\rho} ∣ ⟨ ρ , f ⟩ ∣ ≤ ∥ ρ ∥ 1 1/2 ∥ f ∥ 2 , ρ . Applying the first bound in Lemma 2.2 gives (5) .
Stability and pathwise uniqueness
Let u u u be a solution in the sense of Definition 1.1(i) with initial function u 0 u_0 u 0 and driving noise W W W . Then, for each fixed r > 0 r>0 r > 0 and x ∈ R x\in\mathbb{R} x ∈ R , the equivalence of the weak and mild formulations [10 Theorem 2.1] gives the almost sure identity
u r ( x ) = ( P r u 0 ) ( x ) + ∫ 0 r ∫ R p r − τ ( x − y ) b ( τ , y , u τ ( y ) ) d y d τ + ∫ 0 r ∫ R p r − τ ( x − y ) σ ( τ , y , u τ ( y ) ) W ( d τ , d y ) . u_r(x)=(P_ru_0)(x)+\int_0^r\int_{\mathbb{R}}p_{r-\tau}(x-y)b(\tau,y,u_\tau(y))\,dy\,d\tau+\int_0^r\int_{\mathbb{R}}p_{r-\tau}(x-y)\sigma(\tau,y,u_\tau(y))\,W(d\tau,dy). u r ( x ) = ( P r u 0 ) ( x ) + ∫ 0 r ∫ R p r − τ ( x − y ) b ( τ , y , u τ ( y )) d y d τ + ∫ 0 r ∫ R p r − τ ( x − y ) σ ( τ , y , u τ ( y )) W ( d τ , d y ) .
For a predictable field B B B for which the integral is defined, write
W s , r ( B ) ( x ) : = ∫ s r ∫ R p r − q ( x − y ) B q ( y ) W ( d q , d y ) , 0 ≤ s ≤ r . W_{s,r}(B)(x):=\int_s^r\int_{\mathbb{R}}p_{r-q}(x-y)B_q(y)\,W(dq,dy),\qquad0\leq s\leq r. W s , r ( B ) ( x ) := ∫ s r ∫ R p r − q ( x − y ) B q ( y ) W ( d q , d y ) , 0 ≤ s ≤ r .
On [ s , t ] [s,t] [ s , t ] , let Z Z Z be the auxiliary process with initial value ξ \xi ξ , drift G G G and noise coefficient β \beta β from Proposition 2.1. We can now write its mild equation (4) as
Z r = P r − s ξ + ∫ s r P r − q G q d q + W s , r ( β Z ) , s ≤ r ≤ t . Z_r=P_{r-s}\xi+\int_s^rP_{r-q}G_q\,dq+W_{s,r}(\beta Z),\qquad s\leq r\leq t. Z r = P r − s ξ + ∫ s r P r − q G q d q + W s , r ( β Z ) , s ≤ r ≤ t .
Proposition 3.1. Suppose that (H) holds. Let u , v u,v u , v be solutions with initial functions u 0 , v 0 ∈ C t e m u_0,v_0\in C_{\mathrm{tem}} u 0 , v 0 ∈ C tem on the same filtered space and driven by the same white noise. Fix T < ∞ T<\infty T < ∞ and take the constants L , a L,a L , a in (H) on [ 0 , T ] [0,T] [ 0 , T ] . Then, for every λ > 8 a \lambda>8a λ > 8 a and the corresponding weight ρ \rho ρ , we have, for 0 ≤ t ≤ T 0\leq t\leq T 0 ≤ t ≤ T ,
E ∥ u t − v t ∥ 1 , ρ ≤ e ( λ 2 / 2 + L ) t ∥ u 0 − v 0 ∥ 1 , ρ . (20) \mathbb{E}\|u_t-v_t\|_{1,\rho}\leq e^{(\lambda^2/2+L)t}\|u_0-v_0\|_{1,\rho}. \tag*{(20)} E ∥ u t − v t ∥ 1 , ρ ≤ e ( λ 2 /2 + L ) t ∥ u 0 − v 0 ∥ 1 , ρ . (20)
Proof of Proposition 3.1. Step 1. We localize the two solutions and establish the integrability bounds needed for the comparison argument.
Fix 0 < t ≤ T 0 < t \le T 0 < t ≤ T and λ > 8 a \lambda> 8a λ > 8 a . We may take K ≥ 2 C K \ge2C K ≥ 2 C and m ≥ 1 m \ge1 m ≥ 1 in (H) . Choose κ > 0 \kappa> 0 κ > 0 sufficiently small that
8 ( a + m κ ) + 2 κ < λ . (21) 8(a + m\kappa) + 2\kappa< \lambda. \tag*{(21)} 8 ( a + mκ ) + 2 κ < λ . (21)
For the two solutions u , v u,v u , v in the proposition, define the coefficient differences by
A r ( x ) = σ ( r , x , u r ( x ) ) − σ ( r , x , v r ( x ) ) , G r ( x ) = b ( r , x , u r ( x ) ) − b ( r , x , v r ( x ) ) . A_r(x) = \sigma(r,x,u_r(x)) - \sigma(r,x,v_r(x)), \qquad G_r(x) = b(r,x,u_r(x)) - b(r,x,v_r(x)). A r ( x ) = σ ( r , x , u r ( x )) − σ ( r , x , v r ( x )) , G r ( x ) = b ( r , x , u r ( x )) − b ( r , x , v r ( x )) .
For each integer n > ∣ u 0 ∣ κ ∨ ∣ v 0 ∣ κ n > |u_0|_\kappa\vee|v_0|_\kappa n > ∣ u 0 ∣ κ ∨ ∣ v 0 ∣ κ , define the stopping time and the stopped coefficients by
τ n = T ∧ inf { r ≥ 0 : ∣ u r ∣ κ ∨ ∣ v r ∣ κ ≥ n } , ( A r n , G r n ) = 1 { r ≤ τ n } ( A r , G r ) . \tau_n = T \wedge\inf\{r \ge0 : |u_r|_\kappa\vee|v_r|_\kappa\ge n\}, \qquad(A_r^n,G_r^n) = \mathbf{1}_{\{r \le\tau_n\}}(A_r,G_r). τ n = T ∧ inf { r ≥ 0 : ∣ u r ∣ κ ∨ ∣ v r ∣ κ ≥ n } , ( A r n , G r n ) = 1 { r ≤ τ n } ( A r , G r ) .
For w = u , v w=u,v w = u , v , define
w r n : = { w r , 0 ≤ r ≤ τ n , P r − τ n w τ n , τ n < r ≤ T . w_r^n :=
\begin{cases}
w_r, & 0 \le r \le\tau_n,\\
P_{r-\tau_n}w_{\tau_n}, & \tau_n < r \le T.
\end{cases} w r n := { w r , P r − τ n w τ n , 0 ≤ r ≤ τ n , τ n < r ≤ T .
Using the weak equation up to τ n \tau_n τ n and the heat equation afterwards, we obtain the mild form
w r n ( x ) = ( P r w 0 ) ( x ) + ∫ 0 r ∧ τ n ∫ R p r − q ( x − y ) b ( q , y , w q n ( y ) ) d y d q + ∫ 0 r ∧ τ n ∫ R p r − q ( x − y ) σ ( q , y , w q n ( y ) ) W ( d q , d y ) , 0 ≤ r ≤ T . w_r^n(x) = (P_rw_0)(x) + \int_0^{r\wedge\tau_n}\int_{\mathbb{R}} p_{r-q}(x-y)b(q,y,w_q^n(y))\,\mathrm{d}y\,\mathrm{d}q
+ \int_0^{r\wedge\tau_n}\int_{\mathbb{R}} p_{r-q}(x-y)\sigma(q,y,w_q^n(y))\,W(\mathrm{d}q,\mathrm{d}y), \qquad0 \le r \le T. w r n ( x ) = ( P r w 0 ) ( x ) + ∫ 0 r ∧ τ n ∫ R p r − q ( x − y ) b ( q , y , w q n ( y )) d y d q + ∫ 0 r ∧ τ n ∫ R p r − q ( x − y ) σ ( q , y , w q n ( y )) W ( d q , d y ) , 0 ≤ r ≤ T .
Set D n : = u n − v n D^n := u^n-v^n D n := u n − v n . Subtracting the two mild equations and using u n = u u^n=u u n = u , v n = v v^n=v v n = v up to τ n \tau_n τ n , we obtain
D r n = P r ( u 0 − v 0 ) + ∫ 0 r P r − τ G τ n d τ + W 0 , r ( A n ) , 0 ≤ r ≤ T . (22) D_r^n = P_r(u_0-v_0) + \int_0^r P_{r-\tau}G_\tau^n\,\mathrm{d}\tau+ \mathcal{W}_{0,r}(A^n), \qquad0 \le r \le T. \tag*{(22)} D r n = P r ( u 0 − v 0 ) + ∫ 0 r P r − τ G τ n d τ + W 0 , r ( A n ) , 0 ≤ r ≤ T . (22)
By construction, almost surely for all ( r , x ) ∈ [ 0 , T ] × R (r,x) \in[0,T]\times\mathbb{R} ( r , x ) ∈ [ 0 , T ] × R , we have
1 { r ≤ τ n } D r n ( x ) = 1 { r ≤ τ n } ( u r ( x ) − v r ( x ) ) . (23) \mathbf{1}_{\{r\le\tau_n\}}D_r^n(x) = \mathbf{1}_{\{r\le\tau_n\}}(u_r(x)-v_r(x)). \tag*{(23)} 1 { r ≤ τ n } D r n ( x ) = 1 { r ≤ τ n } ( u r ( x ) − v r ( x )) . (23)
On { r ≤ τ n } \{r\le\tau_n\} { r ≤ τ n } , we have ∣ u r ( x ) ∣ , ∣ v r ( x ) ∣ ≤ n e κ ∣ x ∣ |u_r(x)|, |v_r(x)| \le n e^{\kappa|x|} ∣ u r ( x ) ∣ , ∣ v r ( x ) ∣ ≤ n e κ ∣ x ∣ , so 1 + ∣ u r ( x ) ∣ + ∣ v r ( x ) ∣ ≤ ( 1 + 2 n ) e κ ∣ x ∣ 1+|u_r(x)|+|v_r(x)| \le(1+2n)e^{\kappa|x|} 1 + ∣ u r ( x ) ∣ + ∣ v r ( x ) ∣ ≤ ( 1 + 2 n ) e κ ∣ x ∣ . Set
J n ( x ) : = K ( 1 + 2 n ) m e ( a + m κ ) ∣ x ∣ . (24) J_n(x) := K(1+2n)^m e^{(a+m\kappa)|x|}. \tag*{(24)} J n ( x ) := K ( 1 + 2 n ) m e ( a + mκ ) ∣ x ∣ . (24)
Using (H) (i), K ≥ 2 C K \ge2C K ≥ 2 C and m ≥ 1 m \ge1 m ≥ 1 , we obtain on this event
∣ A r n ( x ) ∣ + ∣ G r n ( x ) ∣ ≤ 2 C ( 1 + n e κ ∣ x ∣ ) ≤ J n ( x ) . (25) |A_r^n(x)|+|G_r^n(x)| \le2C(1+ne^{\kappa|x|}) \le J_n(x). \tag*{(25)} ∣ A r n ( x ) ∣ + ∣ G r n ( x ) ∣ ≤ 2 C ( 1 + n e κ ∣ x ∣ ) ≤ J n ( x ) . (25)
On { r ≤ τ n , ∣ D r n ( x ) ∣ ≤ 1 } \{r\le\tau_n,\ |D_r^n(x)|\le1\} { r ≤ τ n , ∣ D r n ( x ) ∣ ≤ 1 } , applying (H) (iii) and (3.4), we obtain
∣ A r n ( x ) ∣ ≤ K e a ∣ x ∣ ( 1 + ∣ u r ( x ) ∣ + ∣ v r ( x ) ∣ ) m ∣ D r n ( x ) ∣ γ ≤ J n ( x ) ∣ D r n ( x ) ∣ γ . (26) |A_r^n(x)| \le K e^{a|x|}(1+|u_r(x)|+|v_r(x)|)^m |D_r^n(x)|^\gamma
\le J_n(x)|D_r^n(x)|^\gamma. \tag*{(26)} ∣ A r n ( x ) ∣ ≤ K e a ∣ x ∣ ( 1 + ∣ u r ( x ) ∣ + ∣ v r ( x ) ∣ ) m ∣ D r n ( x ) ∣ γ ≤ J n ( x ) ∣ D r n ( x ) ∣ γ . (26)
Combining these estimates and using A n = G n = 0 A^n=G^n=0 A n = G n = 0 after τ n \tau_n τ n , we obtain almost surely for all ( r , x ) ∈ [ 0 , T ] × R (r,x)\in[0,T]\times\mathbb{R} ( r , x ) ∈ [ 0 , T ] × R
∣ A r n ( x ) ∣ ≤ J n ( x ) ( 1 ∧ ∣ D r n ( x ) ∣ γ ) , ∣ G r n ( x ) ∣ ≤ L ∣ D r n ( x ) ∣ , (27) |A_r^n(x)| \le J_n(x)(1\wedge|D_r^n(x)|^\gamma), \qquad|G_r^n(x)| \le L|D_r^n(x)|, \tag*{(27)} ∣ A r n ( x ) ∣ ≤ J n ( x ) ( 1 ∧ ∣ D r n ( x ) ∣ γ ) , ∣ G r n ( x ) ∣ ≤ L ∣ D r n ( x ) ∣ , (27)
where the second inequality follows from (H) (ii).
Using (24) and 1 + x 2 ≥ ∣ x ∣ \sqrt{1+x^2}\ge|x| 1 + x 2 ≥ ∣ x ∣ , we obtain
ρ ( x ) J n ( x ) 8 e 2 κ ∣ x ∣ = e − λ 1 + x 2 [ K ( 1 + 2 n ) m e ( a + m κ ) ∣ x ∣ ] 8 e 2 κ ∣ x ∣ ≤ K 8 ( 1 + 2 n ) 8 m e − [ λ − 8 ( a + m κ ) − 2 κ ] ∣ x ∣ . \rho(x)J_n(x)^8e^{2\kappa|x|}=e^{-\lambda\sqrt{1+x^2}}\left[K(1+2n)^m e^{(a+m\kappa)|x|}\right]^8e^{2\kappa|x|}
\le K^8(1+2n)^{8m}e^{-[\lambda-8(a+m\kappa)-2\kappa]|x|}. ρ ( x ) J n ( x ) 8 e 2 κ ∣ x ∣ = e − λ 1 + x 2 [ K ( 1 + 2 n ) m e ( a + mκ ) ∣ x ∣ ] 8 e 2 κ ∣ x ∣ ≤ K 8 ( 1 + 2 n ) 8 m e − [ λ − 8 ( a + mκ ) − 2 κ ] ∣ x ∣ .
By (21) , λ − 8 ( a + m κ ) − 2 κ > 0 \lambda-8(a+m\kappa)-2\kappa>0 λ − 8 ( a + mκ ) − 2 κ > 0 . Integrating this bound over R \mathbb{R} R therefore gives
∫ R ρ ( x ) J n ( x ) 8 d x ≤ ∫ R ρ ( x ) J n ( x ) 8 e 2 κ ∣ x ∣ d x ≤ 2 K 8 ( 1 + 2 n ) 8 m λ − 8 ( a + m κ ) − 2 κ < ∞ . (28) \int_{\mathbb{R}}\rho(x)J_n(x)^8\,dx\le\int_{\mathbb{R}}\rho(x)J_n(x)^8e^{2\kappa|x|}\,dx\le\frac{2K^8(1+2n)^{8m}}{\lambda-8(a+m\kappa)-2\kappa}<\infty. \tag*{(28)} ∫ R ρ ( x ) J n ( x ) 8 d x ≤ ∫ R ρ ( x ) J n ( x ) 8 e 2 κ ∣ x ∣ d x ≤ λ − 8 ( a + mκ ) − 2 κ 2 K 8 ( 1 + 2 n ) 8 m < ∞. (28)
As ρ ∈ L 1 ( R , d x ) \rho\in L^1(\mathbb{R},dx) ρ ∈ L 1 ( R , d x ) , we also have ∥ J n ∥ 2 , ρ < ∞ \|J_n\|_{2,\rho}<\infty ∥ J n ∥ 2 , ρ < ∞ . For 0 ≤ r ≤ T 0\le r\le T 0 ≤ r ≤ T , applying Lemma 2.2 to the defining formula for D n D^n D n gives
∥ D r n ∥ 2 , ρ 2 = ∥ P r − ( r ∧ τ n ) ( u r ∧ τ n − v r ∧ τ n ) ∥ 2 , ρ 2 ≤ e c T ∥ u r ∧ τ n − v r ∧ τ n ∥ 2 , ρ 2 ≤ 4 n 2 e c T ∫ R ρ ( x ) e 2 κ ∣ x ∣ d x < ∞ . \|D_r^n\|_{2,\rho}^2=\|P_{r-(r\wedge\tau_n)}(u_{r\wedge\tau_n}-v_{r\wedge\tau_n})\|_{2,\rho}^2
\le e^{cT}\|u_{r\wedge\tau_n}-v_{r\wedge\tau_n}\|_{2,\rho}^2
\le4n^2e^{cT}\int_{\mathbb{R}}\rho(x)e^{2\kappa|x|}\,dx<\infty. ∥ D r n ∥ 2 , ρ 2 = ∥ P r − ( r ∧ τ n ) ( u r ∧ τ n − v r ∧ τ n ) ∥ 2 , ρ 2 ≤ e c T ∥ u r ∧ τ n − v r ∧ τ n ∥ 2 , ρ 2 ≤ 4 n 2 e c T ∫ R ρ ( x ) e 2 κ ∣ x ∣ d x < ∞.
Here 2 κ < λ 2\kappa<\lambda 2 κ < λ by (21) , and the final bound is deterministic and uniform in r r r . Using ∣ A n ∣ + ∣ G n ∣ ≤ J n |A^n|+|G^n|\le J_n ∣ A n ∣ + ∣ G n ∣ ≤ J n , we also obtain
∥ A r n ∥ 2 , ρ 2 + ∥ G r n ∥ 2 , ρ 2 = ∫ R ρ ( x ) ( ∣ A r n ( x ) ∣ 2 + ∣ G r n ( x ) ∣ 2 ) d x ≤ ∥ J n ∥ 2 , ρ 2 < ∞ . \|A_r^n\|_{2,\rho}^2+\|G_r^n\|_{2,\rho}^2=\int_{\mathbb{R}}\rho(x)(|A_r^n(x)|^2+|G_r^n(x)|^2)\,dx\le\|J_n\|_{2,\rho}^2<\infty. ∥ A r n ∥ 2 , ρ 2 + ∥ G r n ∥ 2 , ρ 2 = ∫ R ρ ( x ) ( ∣ A r n ( x ) ∣ 2 + ∣ G r n ( x ) ∣ 2 ) d x ≤ ∥ J n ∥ 2 , ρ 2 < ∞.
Taking expectations and then the supremum over r ≤ T r\le T r ≤ T in these bounds, we conclude that
sup r ≤ T E [ ∥ A r n ∥ 2 , ρ 2 + ∥ G r n ∥ 2 , ρ 2 + ∥ D r n ∥ 2 , ρ 2 ] < ∞ . (29) \sup_{r\le T}\mathbb{E}\left[\|A_r^n\|_{2,\rho}^2+\|G_r^n\|_{2,\rho}^2+\|D_r^n\|_{2,\rho}^2\right]<\infty. \tag*{(29)} r ≤ T sup E [ ∥ A r n ∥ 2 , ρ 2 + ∥ G r n ∥ 2 , ρ 2 + ∥ D r n ∥ 2 , ρ 2 ] < ∞. (29)
Step 2. We construct the auxiliary processes and reduce the desired estimate to controlling the sum of the comparison errors.
Fix n n n . Put h : = t / N h:=t/N h := t / N and t j : = j h t_j:=jh t j := j h . To construct the auxiliary processes with a bounded coefficient, define
β : = { A n / D n , D n ≠ 0 , 0 , D n = 0 , β ( h ) : = { − 1 2 h − 1 / 4 , β < − 1 2 h − 1 / 4 , β , ∣ β ∣ ≤ 1 2 h − 1 / 4 , 1 2 h − 1 / 4 , β > 1 2 h − 1 / 4 . \beta:=\begin{cases}
A^n/D^n, & D^n\ne0,\\
0, & D^n=0,
\end{cases}
\qquad
\beta^{(h)}:=\begin{cases}
-\frac{1}{2}h^{-1/4}, & \beta<-\frac{1}{2}h^{-1/4},\\
\beta, & |\beta|\le\frac{1}{2}h^{-1/4},\\
\frac{1}{2}h^{-1/4}, & \beta>\frac{1}{2}h^{-1/4}.
\end{cases} β := { A n / D n , 0 , D n = 0 , D n = 0 , β ( h ) := ⎩ ⎨ ⎧ − 2 1 h − 1/4 , β , 2 1 h − 1/4 , β < − 2 1 h − 1/4 , ∣ β ∣ ≤ 2 1 h − 1/4 , β > 2 1 h − 1/4 .
For j = 0 , … , N − 1 j=0,\ldots,N-1 j = 0 , … , N − 1 , consider the equation
Z r j = P r − t j D t j n + ∫ t j r P r − q G q n d q + W t j , r ( β ( h ) Z j ) , t j ≤ r ≤ t j + 1 . Z_r^j=P_{r-t_j}D_{t_j}^n+\int_{t_j}^rP_{r-q}G_q^n\,dq+W_{t_j,r}(\beta^{(h)}Z^j),\qquad t_j\le r\le t_{j+1}. Z r j = P r − t j D t j n + ∫ t j r P r − q G q n d q + W t j , r ( β ( h ) Z j ) , t j ≤ r ≤ t j + 1 .
This is (4) with initial value D t j n D_{t_j}^n D t j n , drift G n G^n G n and coefficient β ( h ) \beta^{(h)} β ( h ) . By (29) , we have
E ∥ D t j n ∥ 2 , ρ 2 < ∞ , ∫ t j t j + 1 E ∥ G q n ∥ 2 , ρ 2 d q ≤ h sup q ≤ T E ∥ G q n ∥ 2 , ρ 2 < ∞ . \mathbb{E}\|D_{t_j}^n\|_{2,\rho}^2<\infty,\qquad\int_{t_j}^{t_{j+1}}\mathbb{E}\|G_q^n\|_{2,\rho}^2\,dq\le h\sup_{q\le T}\mathbb{E}\|G_q^n\|_{2,\rho}^2<\infty. E ∥ D t j n ∥ 2 , ρ 2 < ∞ , ∫ t j t j + 1 E ∥ G q n ∥ 2 , ρ 2 d q ≤ h q ≤ T sup E ∥ G q n ∥ 2 , ρ 2 < ∞.
Using these bounds and ∣ β ( h ) ∣ ≤ 1 2 h − 1 / 4 |\beta^{(h)}|\le\frac{1}{2}h^{-1/4} ∣ β ( h ) ∣ ≤ 2 1 h − 1/4 , we apply Proposition 2.1 to obtain a unique predictable solution Z j Z^j Z j satisfying
sup t j ≤ r ≤ t j + 1 E ∥ Z r j ∥ 2 , ρ 2 < ∞ . (30) \sup_{t_j\le r\le t_{j+1}}\mathbb{E}\|Z_r^j\|_{2,\rho}^2<\infty. \tag*{(30)} t j ≤ r ≤ t j + 1 sup E ∥ Z r j ∥ 2 , ρ 2 < ∞. (30)
and
E ∥ Z t j + 1 j ∥ 1 , ρ ≤ e c h E ∥ D t j n ∥ 1 , ρ + ∫ t j t j + 1 e c ( t j + 1 − q ) E ∥ G q n ∥ 1 , ρ d q . (31) \mathbb{E}\|Z_{t_{j+1}}^j\|_{1,\rho}\le e^{ch}\mathbb{E}\|D_{t_j}^n\|_{1,\rho}+\int_{t_j}^{t_{j+1}}e^{c(t_{j+1}-q)}\mathbb{E}\|G_q^n\|_{1,\rho}\,dq. \tag*{(31)} E ∥ Z t j + 1 j ∥ 1 , ρ ≤ e c h E ∥ D t j n ∥ 1 , ρ + ∫ t j t j + 1 e c ( t j + 1 − q ) E ∥ G q n ∥ 1 , ρ d q . (31)
To compare D n D^n D n with Z j Z^j Z j , set R h : = A n − β ( h ) D n R_h:=A^n-\beta^{(h)}D^n R h := A n − β ( h ) D n and X r : = D r n − Z r j X_r:=D_r^n-Z_r^j X r := D r n − Z r j for t j ≤ r ≤ t j + 1 t_j\le r\le t_{j+1} t j ≤ r ≤ t j + 1 . Restarting (22) at t j t_j t j and subtracting the mild equation for Z j Z^j Z j , we cancel the initial and drift terms and obtain
X r = W t j , r ( β ( h ) X + R h ) , t j ≤ r ≤ t j + 1 . (32) X_r=W_{t_j,r}(\beta^{(h)}X+R_h),\qquad t_j\le r\le t_{j+1}. \tag*{(32)} X r = W t j , r ( β ( h ) X + R h ) , t j ≤ r ≤ t j + 1 . (32)
To transfer the L 1 L^1 L 1 estimate to D n D^n D n , set
ε j : = E ∥ X t j + 1 ∥ 1 , ρ . \varepsilon_j := \mathbb{E}\lVert X_{t_{j+1}}\rVert_{1,\rho}. ε j := E ∥ X t j + 1 ∥ 1 , ρ .
Applying (3.12) and the second bound in (3.8), we obtain
E ∥ D t j + 1 n ∥ 1 , ρ ≤ E ∥ Z t j + 1 j ∥ 1 , ρ + ε j ≤ e c h E ∥ D t j n ∥ 1 , ρ + L ∫ t j t j + 1 e c ( t j + 1 − q ) E ∥ D q n ∥ 1 , ρ d q + ε j . \begin{aligned}
\mathbb{E}\lVert D^n_{t_{j+1}}\rVert_{1,\rho}
&\le\mathbb{E}\lVert Z^j_{t_{j+1}}\rVert_{1,\rho}+\varepsilon_j\\
&\le e^{ch}\mathbb{E}\lVert D^n_{t_j}\rVert_{1,\rho}
+L\int_{t_j}^{t_{j+1}}e^{c(t_{j+1}-q)}\mathbb{E}\lVert D^n_q\rVert_{1,\rho}\,dq+\varepsilon_j.
\end{aligned} E ∥ D t j + 1 n ∥ 1 , ρ ≤ E ∥ Z t j + 1 j ∥ 1 , ρ + ε j ≤ e c h E ∥ D t j n ∥ 1 , ρ + L ∫ t j t j + 1 e c ( t j + 1 − q ) E ∥ D q n ∥ 1 , ρ d q + ε j .
Multiplying by e − c t j + 1 e^{-ct_{j+1}} e − c t j + 1 and summing over j j j , we obtain
e − c t E ∥ D t n ∥ 1 , ρ ≤ ∥ u 0 − v 0 ∥ 1 , ρ + L ∫ 0 t e − c q E ∥ D q n ∥ 1 , ρ d q + ∑ j = 0 N − 1 e − c t j + 1 ε j . (33) e^{-ct}\mathbb{E}\lVert D^n_t\rVert_{1,\rho}
\le\lVert u_0-v_0\rVert_{1,\rho}
+L\int_0^t e^{-cq}\mathbb{E}\lVert D^n_q\rVert_{1,\rho}\,dq
+\sum_{j=0}^{N-1}e^{-ct_{j+1}}\varepsilon_j. \tag*{(33)} e − c t E ∥ D t n ∥ 1 , ρ ≤ ∥ u 0 − v 0 ∥ 1 , ρ + L ∫ 0 t e − c q E ∥ D q n ∥ 1 , ρ d q + j = 0 ∑ N − 1 e − c t j + 1 ε j . (33)
For fixed n n n , the process D n D^n D n and the constants c , L c,L c , L do not depend on N N N . Thus, to obtain a Gronwall inequality, it suffices to prove
∑ j = 0 N − 1 ε j ⟶ 0 as N → ∞ . (34) \sum_{j=0}^{N-1}\varepsilon_j\longrightarrow0
\qquad\text{as }N\to\infty. \tag*{(34)} j = 0 ∑ N − 1 ε j ⟶ 0 as N → ∞. (34)
Step 3. We next show that (3.15) holds and complete the proof.
On the truncation event { ∣ β q ( x ) ∣ > 1 2 h − 1 / 4 } \{|\beta_q(x)|>\frac{1}{2}h^{-1/4}\} { ∣ β q ( x ) ∣ > 2 1 h − 1/4 } , we have D q n ( x ) ≠ 0 D^n_q(x)\ne0 D q n ( x ) = 0 by the definition of β q ( x ) \beta_q(x) β q ( x ) . Since 0 ≤ 4 γ − 3 ≤ 1 0\le4\gamma-3\le1 0 ≤ 4 γ − 3 ≤ 1 , (3.6) and (3.8) give
∣ β ∣ 4 γ ∣ A n ∣ = ( ∣ A n ∣ ∣ D n ∣ γ ) 4 ∣ A n ∣ 4 γ − 3 ≤ J n ( x ) 4 ∣ A n ∣ 4 γ − 3 ≤ C n J n ( x ) 4 e κ ∣ x ∣ . (35) |\beta|^{4\gamma}|A^n|
=\left(\frac{|A^n|}{|D^n|\gamma}\right)^4|A^n|^{4\gamma-3}
\le J_n(x)^4|A^n|^{4\gamma-3}
\le C_nJ_n(x)^4e^{\kappa|x|}. \tag*{(35)} ∣ β ∣ 4 γ ∣ A n ∣ = ( ∣ D n ∣ γ ∣ A n ∣ ) 4 ∣ A n ∣ 4 γ − 3 ≤ J n ( x ) 4 ∣ A n ∣ 4 γ − 3 ≤ C n J n ( x ) 4 e κ ∣ x ∣ . (35)
Using the definitions of β ( h ) \beta^{(h)} β ( h ) and R h R_h R h and applying (3.16) in the last line, we obtain
h − γ ∣ R h ∣ = h − γ ∣ D n ∣ ∣ β − β ( h ) ∣ = h − γ ∣ D n ∣ ( ∣ β ∣ − 1 2 h − 1 / 4 ) + ≤ h − γ ∣ A n ∣ 1 { ∣ β ∣ > 1 2 h − 1 / 4 } ≤ 2 4 γ ∣ β ∣ 4 γ ∣ A n ∣ 1 { ∣ β ∣ > 1 2 h − 1 / 4 } ≤ C n J n ( x ) 4 e κ ∣ x ∣ 1 { ∣ β ∣ > 1 2 h − 1 / 4 } . (36) \begin{aligned}
h^{-\gamma}|R_h|
&=h^{-\gamma}|D^n|\,|\beta-\beta^{(h)}|\\
&=h^{-\gamma}|D^n|\left(|\beta|-\tfrac12h^{-1/4}\right)^+\\
&\le h^{-\gamma}|A^n|\mathbf{1}_{\{|\beta|>\frac12h^{-1/4}\}}\\
&\le2^{4\gamma}|\beta|^{4\gamma}|A^n|\mathbf{1}_{\{|\beta|>\frac12h^{-1/4}\}}\\
&\le C_nJ_n(x)^4e^{\kappa|x|}\mathbf{1}_{\{|\beta|>\frac12h^{-1/4}\}}. \tag*{(36)}
\end{aligned} h − γ ∣ R h ∣ = h − γ ∣ D n ∣ ∣ β − β ( h ) ∣ = h − γ ∣ D n ∣ ( ∣ β ∣ − 2 1 h − 1/4 ) + ≤ h − γ ∣ A n ∣ 1 { ∣ β ∣ > 2 1 h − 1/4 } ≤ 2 4 γ ∣ β ∣ 4 γ ∣ A n ∣ 1 { ∣ β ∣ > 2 1 h − 1/4 } ≤ C n J n ( x ) 4 e κ ∣ x ∣ 1 { ∣ β ∣ > 2 1 h − 1/4 } . (36)
To bound ε j \varepsilon_j ε j on I j = [ t j , t j + 1 ] I_j=[t_j,t_{j+1}] I j = [ t j , t j + 1 ] , we first square (3.17) to obtain
h − 2 γ E ∥ R h ( q ) ∥ 2 , ρ 2 ≤ C n E ∫ R ρ ( x ) J n ( x ) 8 e 2 κ ∣ x ∣ 1 { ∣ β q ( x ) ∣ > 1 2 h − 1 / 4 } d x . h^{-2\gamma}\mathbb{E}\lVert R_h(q)\rVert_{2,\rho}^2
\le C_n\mathbb{E}\int_{\mathbb{R}}\rho(x)J_n(x)^{8}e^{2\kappa|x|}
\mathbf{1}_{\{|\beta_q(x)|>\frac12h^{-1/4}\}}\,dx. h − 2 γ E ∥ R h ( q ) ∥ 2 , ρ 2 ≤ C n E ∫ R ρ ( x ) J n ( x ) 8 e 2 κ ∣ x ∣ 1 { ∣ β q ( x ) ∣ > 2 1 h − 1/4 } d x .
The indicator tends to zero pointwise as h ↓ 0 h\downarrow0 h ↓ 0 . By (3.9) and dominated convergence,
sup 0 ≤ q ≤ t E ∥ R h ( q ) ∥ 2 , ρ 2 = O ( h 2 γ ) , ∫ 0 t E ∥ R h ( q ) ∥ 2 , ρ 2 d q = o ( h 2 γ ) . (37) \sup_{0\le q\le t}\mathbb{E}\lVert R_h(q)\rVert_{2,\rho}^2=O(h^{2\gamma}),
\qquad
\int_0^t\mathbb{E}\lVert R_h(q)\rVert_{2,\rho}^2\,dq=o(h^{2\gamma}). \tag*{(37)} 0 ≤ q ≤ t sup E ∥ R h ( q ) ∥ 2 , ρ 2 = O ( h 2 γ ) , ∫ 0 t E ∥ R h ( q ) ∥ 2 , ρ 2 d q = o ( h 2 γ ) . (37)
Take h h h small enough that e c h / 2 ≤ π e^{ch/2}\le\sqrt{\pi} e c h /2 ≤ π . Applying (2.7) to (3.13) gives, for r ∈ I j r\in I_j r ∈ I j ,
E ∥ X r ∥ 2 , ρ 2 ≤ ∫ t j r e c ( r − q ) / 2 2 π ( r − q ) E ∥ β q ( h ) X q + R h ( q ) ∥ 2 , ρ 2 d q ≤ ∫ t j r E ∥ β q ( h ) X q ∥ 2 , ρ 2 + E ∥ R h ( q ) ∥ 2 , ρ 2 r − q d q ≤ ( 1 2 h − 1 / 4 ) 2 ∫ t j r d q r − q ⏟ = 1 2 ( r − t j ) / h ≤ 1 / 2 sup s ∈ I j E ∥ X s ∥ 2 , ρ 2 + ∫ t j r E ∥ R h ( q ) ∥ 2 , ρ 2 r − q d q . \begin{aligned}
\mathbb{E}\lVert X_r\rVert_{2,\rho}^2
&\le\int_{t_j}^{r}\frac{e^{c(r-q)/2}}{2\sqrt{\pi(r-q)}}\,
\mathbb{E}\lVert\beta_q^{(h)}X_q+R_h(q)\rVert_{2,\rho}^2\,dq\\
&\le\int_{t_j}^{r}\frac{\mathbb{E}\lVert\beta_q^{(h)}X_q\rVert_{2,\rho}^2+
\mathbb{E}\lVert R_h(q)\rVert_{2,\rho}^2}{\sqrt{r-q}}\,dq\\
&\le\left(\tfrac12h^{-1/4}\right)^2
\underbrace{\int_{t_j}^{r}\frac{dq}{\sqrt{r-q}}}_{=\frac12\sqrt{(r-t_j)/h}\le1/2}
\sup_{s\in I_j}\mathbb{E}\lVert X_s\rVert_{2,\rho}^2
+\int_{t_j}^{r}\frac{\mathbb{E}\lVert R_h(q)\rVert_{2,\rho}^2}{\sqrt{r-q}}\,dq.
\end{aligned} E ∥ X r ∥ 2 , ρ 2 ≤ ∫ t j r 2 π ( r − q ) e c ( r − q ) /2 E ∥ β q ( h ) X q + R h ( q ) ∥ 2 , ρ 2 d q ≤ ∫ t j r r − q E ∥ β q ( h ) X q ∥ 2 , ρ 2 + E ∥ R h ( q ) ∥ 2 , ρ 2 d q ≤ ( 2 1 h − 1/4 ) 2 = 2 1 ( r − t j ) / h ≤ 1/2 ∫ t j r r − q d q s ∈ I j sup E ∥ X s ∥ 2 , ρ 2 + ∫ t j r r − q E ∥ R h ( q ) ∥ 2 , ρ 2 d q .
where the last inequality uses ∣ β q ( h ) ∣ ≤ 1 2 h − 1 / 4 |\beta_q^{(h)}| \le\frac{1}{2}h^{-1/4} ∣ β q ( h ) ∣ ≤ 2 1 h − 1/4 . The supremum is finite by (29) and (30) . Squaring the last integral and using symmetry gives
( ∫ t j r E ∥ R h ( q ) ∥ 2 , ρ 2 r − q d q ) 2 = 2 ∫ t j r E ∥ R h ( q ) ∥ 2 , ρ 2 r − q ∫ q r E ∥ R h ( v ) ∥ 2 , ρ 2 r − v d v d q ≤ 2 r − q sup 0 ≤ s ≤ t E ∥ R h ( s ) ∥ 2 , ρ 2 ≤ 4 ( sup 0 ≤ s ≤ t E ∥ R h ( s ) ∥ 2 , ρ 2 ) ∫ I j E ∥ R h ( q ) ∥ 2 , ρ 2 d q . \left(\int_{t_j}^{r}\frac{\mathbb{E}\|R_h(q)\|_{2,\rho}^{2}}{\sqrt{r-q}}\,\mathrm{d}q\right)^2
=2\int_{t_j}^{r}\frac{\mathbb{E}\|R_h(q)\|_{2,\rho}^{2}}{\sqrt{r-q}}\int_q^r\frac{\mathbb{E}\|R_h(v)\|_{2,\rho}^{2}}{\sqrt{r-v}}\,\mathrm{d}v\,\mathrm{d}q
\le2\sqrt{r-q}\sup_{0\le s\le t}\mathbb{E}\|R_h(s)\|_{2,\rho}^{2}
\le4\left(\sup_{0\le s\le t}\mathbb{E}\|R_h(s)\|_{2,\rho}^{2}\right)\int_{I_j}\mathbb{E}\|R_h(q)\|_{2,\rho}^{2}\,\mathrm{d}q. ( ∫ t j r r − q E ∥ R h ( q ) ∥ 2 , ρ 2 d q ) 2 = 2 ∫ t j r r − q E ∥ R h ( q ) ∥ 2 , ρ 2 ∫ q r r − v E ∥ R h ( v ) ∥ 2 , ρ 2 d v d q ≤ 2 r − q 0 ≤ s ≤ t sup E ∥ R h ( s ) ∥ 2 , ρ 2 ≤ 4 ( 0 ≤ s ≤ t sup E ∥ R h ( s ) ∥ 2 , ρ 2 ) ∫ I j E ∥ R h ( q ) ∥ 2 , ρ 2 d q .
Taking the supremum over r ∈ I j r\in I_j r ∈ I j in the estimate for E ∥ X r ∥ 2 , ρ 2 \mathbb{E}\|X_r\|_{2,\rho}^{2} E ∥ X r ∥ 2 , ρ 2 and moving the first term to the left gives
1 2 sup r ∈ I j E ∥ X r ∥ 2 , ρ 2 ≤ 2 ( sup 0 ≤ s ≤ t E ∥ R h ( s ) ∥ 2 , ρ 2 ∫ I j E ∥ R h ( q ) ∥ 2 , ρ 2 d q ) 1 / 2 . (38) \frac{1}{2}\sup_{r\in I_j}\mathbb{E}\|X_r\|_{2,\rho}^{2}
\le2\left(\sup_{0\le s\le t}\mathbb{E}\|R_h(s)\|_{2,\rho}^{2}\int_{I_j}\mathbb{E}\|R_h(q)\|_{2,\rho}^{2}\,\mathrm{d}q\right)^{1/2}. \tag*{(38)} 2 1 r ∈ I j sup E ∥ X r ∥ 2 , ρ 2 ≤ 2 ( 0 ≤ s ≤ t sup E ∥ R h ( s ) ∥ 2 , ρ 2 ∫ I j E ∥ R h ( q ) ∥ 2 , ρ 2 d q ) 1/2 . (38)
By Cauchy–Schwarz,
ε j 2 = ( E ∥ X t j + 1 ∥ 1 , ρ ) 2 ≤ ∥ ρ ∥ 1 E ∥ X t j + 1 ∥ 2 , ρ 2 . \varepsilon_j^2=\left(\mathbb{E}\|X_{t_{j+1}}\|_{1,\rho}\right)^2\le\|\rho\|_1\mathbb{E}\|X_{t_{j+1}}\|_{2,\rho}^{2}. ε j 2 = ( E ∥ X t j + 1 ∥ 1 , ρ ) 2 ≤ ∥ ρ ∥ 1 E ∥ X t j + 1 ∥ 2 , ρ 2 .
Applying (38) gives
ε j ≤ 2 ∥ ρ ∥ 1 1 / 2 [ ( sup 0 ≤ s ≤ t E ∥ R h ( s ) ∥ 2 , ρ 2 ) ∫ I j E ∥ R h ( q ) ∥ 2 , ρ 2 d q ] 1 / 4 . \varepsilon_j\le2\|\rho\|_1^{1/2}\left[\left(\sup_{0\le s\le t}\mathbb{E}\|R_h(s)\|_{2,\rho}^{2}\right)\int_{I_j}\mathbb{E}\|R_h(q)\|_{2,\rho}^{2}\,\mathrm{d}q\right]^{1/4}. ε j ≤ 2∥ ρ ∥ 1 1/2 [ ( 0 ≤ s ≤ t sup E ∥ R h ( s ) ∥ 2 , ρ 2 ) ∫ I j E ∥ R h ( q ) ∥ 2 , ρ 2 d q ] 1/4 .
Summing over the N = t / h N=t/h N = t / h intervals and using the concavity of z ↦ z 1 / 4 z\mapsto z^{1/4} z ↦ z 1/4 , we obtain
∑ j = 0 N − 1 ε j ≤ 2 ∥ ρ ∥ 1 1 / 2 ( sup 0 ≤ s ≤ t E ∥ R h ( s ) ∥ 2 , ρ 2 ) 1 / 4 ∑ j = 0 N − 1 ( ∫ I j E ∥ R h ( q ) ∥ 2 , ρ 2 d q ) 1 / 4 \sum_{j=0}^{N-1}\varepsilon_j\le2\|\rho\|_1^{1/2}\left(\sup_{0\le s\le t}\mathbb{E}\|R_h(s)\|_{2,\rho}^{2}\right)^{1/4}\sum_{j=0}^{N-1}\left(\int_{I_j}\mathbb{E}\|R_h(q)\|_{2,\rho}^{2}\,\mathrm{d}q\right)^{1/4} j = 0 ∑ N − 1 ε j ≤ 2∥ ρ ∥ 1 1/2 ( 0 ≤ s ≤ t sup E ∥ R h ( s ) ∥ 2 , ρ 2 ) 1/4 j = 0 ∑ N − 1 ( ∫ I j E ∥ R h ( q ) ∥ 2 , ρ 2 d q ) 1/4
≤ 2 ∥ ρ ∥ 1 1 / 2 ( sup 0 ≤ s ≤ t E ∥ R h ( s ) ∥ 2 , ρ 2 ) 1 / 4 N ( 1 N ∑ j = 0 N − 1 ∫ I j E ∥ R h ( q ) ∥ 2 , ρ 2 d q ) 1 / 4 = 2 ∥ ρ ∥ 1 1 / 2 ( t h ) 3 / 4 [ ( sup 0 ≤ s ≤ t E ∥ R h ( s ) ∥ 2 , ρ 2 ) ⏟ O ( h 2 γ ) ∫ 0 t E ∥ R h ( q ) ∥ 2 , ρ 2 d q ⏟ o ( h 2 γ ) ] 1 / 4 = o ( h γ − 3 / 4 ) (39) \begin{aligned}
\le2\|\rho\|_1^{1/2}\left(\sup_{0\le s\le t}\mathbb{E}\|R_h(s)\|_{2,\rho}^{2}\right)^{1/4}N\left(\frac{1}{N}\sum_{j=0}^{N-1}\int_{I_j}\mathbb{E}\|R_h(q)\|_{2,\rho}^{2}\,\mathrm{d}q\right)^{1/4} \\
&=2\|\rho\|_1^{1/2}\left(\frac{t}{h}\right)^{3/4}\left[\underbrace{\left(\sup_{0\le s\le t}\mathbb{E}\|R_h(s)\|_{2,\rho}^{2}\right)}_{O(h^{2\gamma})}\underbrace{\int_0^t\mathbb{E}\|R_h(q)\|_{2,\rho}^{2}\,\mathrm{d}q}_{o(h^{2\gamma})}\right]^{1/4}=o(h^{\gamma-3/4}) \tag*{(39)}
\end{aligned} ≤ 2∥ ρ ∥ 1 1/2 ( 0 ≤ s ≤ t sup E ∥ R h ( s ) ∥ 2 , ρ 2 ) 1/4 N ( N 1 j = 0 ∑ N − 1 ∫ I j E ∥ R h ( q ) ∥ 2 , ρ 2 d q ) 1/4 = 2∥ ρ ∥ 1 1/2 ( h t ) 3/4 O ( h 2 γ ) ( 0 ≤ s ≤ t sup E ∥ R h ( s ) ∥ 2 , ρ 2 ) o ( h 2 γ ) ∫ 0 t E ∥ R h ( q ) ∥ 2 , ρ 2 d q 1/4 = o ( h γ − 3/4 ) (39)
as h ↓ 0 h\downarrow0 h ↓ 0 , by (37) .
Applying (39) in (33) and letting N → ∞ N\to\infty N → ∞ with n n n and t t t fixed, we obtain
e − c t E ∥ D t n ∥ 1 , ρ ≤ ∥ u 0 − v 0 ∥ 1 , ρ + L ∫ 0 t e − c q E ∥ D q n ∥ 1 , ρ d q , 0 ≤ t ≤ T . e^{-ct}\mathbb{E}\|D_t^n\|_{1,\rho}\le\|u_0-v_0\|_{1,\rho}+L\int_0^t e^{-cq}\mathbb{E}\|D_q^n\|_{1,\rho}\,\mathrm{d}q,\qquad0\le t\le T. e − c t E ∥ D t n ∥ 1 , ρ ≤ ∥ u 0 − v 0 ∥ 1 , ρ + L ∫ 0 t e − c q E ∥ D q n ∥ 1 , ρ d q , 0 ≤ t ≤ T .
By (29) , r ↦ E ∥ D r n ∥ 1 , ρ r\mapsto\mathbb{E}\|D_r^n\|_{1,\rho} r ↦ E ∥ D r n ∥ 1 , ρ is bounded on [ 0 , T ] [0,T] [ 0 , T ] , so Gronwall’s inequality gives
E ∥ D t n ∥ 1 , ρ ≤ e ( c + L ) t ∥ u 0 − v 0 ∥ 1 , ρ . \mathbb{E}\|D_t^n\|_{1,\rho}\le e^{(c+L)t}\|u_0-v_0\|_{1,\rho}. E ∥ D t n ∥ 1 , ρ ≤ e ( c + L ) t ∥ u 0 − v 0 ∥ 1 , ρ .
Almost surely the paths are bounded in ∣ ⋅ ∣ κ |\cdot|_\kappa ∣ ⋅ ∣ κ on [ 0 , T ] [0,T] [ 0 , T ] , so eventually τ n = T \tau_n=T τ n = T and D n = u − v D^n=u-v D n = u − v on [ 0 , T ] [0,T] [ 0 , T ] by (23) . Fatou’s lemma gives
E ∥ u t − v t ∥ 1 , ρ ≤ lim inf n → ∞ E ∥ D t n ∥ 1 , ρ ≤ e ( c + L ) t ∥ u 0 − v 0 ∥ 1 , ρ . \mathbb{E}\|u_t-v_t\|_{1,\rho}\le\liminf_{n\to\infty}\mathbb{E}\|D_t^n\|_{1,\rho}\le e^{(c+L)t}\|u_0-v_0\|_{1,\rho}. E ∥ u t − v t ∥ 1 , ρ ≤ n → ∞ lim inf E ∥ D t n ∥ 1 , ρ ≤ e ( c + L ) t ∥ u 0 − v 0 ∥ 1 , ρ .
If u 0 = v 0 u_0=v_0 u 0 = v 0 , spatial continuity gives equality for every x x x at each rational time almost surely, and time continuity gives indistinguishability.
Existence and uniqueness in law
First, we regularize the diffusion coefficient and construct approximate solutions by Picard iteration. Lemma 4.1 gives the convergence of the iteration and uniform moment bounds. Second, Lemma 4.2 proves weak existence. Finally, we establish Theorem 1.2 by combining weak existence from Lemma 4.2 with pathwise uniqueness from Proposition 3.1.
Let π n z = ( − n ) ∨ ( z ∧ n ) \pi_n z = (-n) \vee(z \wedge n) π n z = ( − n ) ∨ ( z ∧ n ) . Choose a smooth probability density φ \varphi φ supported on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] . Set φ n ( v ) : = n φ ( n v ) \varphi_n(v) := n\varphi(nv) φ n ( v ) := n φ ( n v ) and put
σ n ( r , x , z ) : = ∫ R φ n ( v ) σ ( r , x , π n ( z − v ) ) d v . \sigma_n(r,x,z) := \int_{\mathbb{R}} \varphi_n(v)\sigma(r,x,\pi_n(z-v))\,\mathrm{d}v. σ n ( r , x , z ) := ∫ R φ n ( v ) σ ( r , x , π n ( z − v )) d v .
Fix u 0 ∈ C t e m u_0 \in C_{\mathrm{tem}} u 0 ∈ C tem and T > 0 T > 0 T > 0 . By (H)(i) and (iii), we have
∣ σ n ( r , x , z ) ∣ ≤ C ( 2 + ∣ z ∣ ) , ∣ ∂ z σ n ( r , x , z ) ∣ ≤ C ( 1 + n ) ∥ φ n ∥ 1 (40) |\sigma_n(r,x,z)| \le C(2+|z|), \qquad|\partial_z\sigma_n(r,x,z)| \le C(1+n)\|\varphi_n\|_1 \tag*{(40)} ∣ σ n ( r , x , z ) ∣ ≤ C ( 2 + ∣ z ∣ ) , ∣ ∂ z σ n ( r , x , z ) ∣ ≤ C ( 1 + n ) ∥ φ n ∥ 1 (40)
and
sup r ≤ T , ∣ x ∣ , ∣ z ∣ ≤ R ∣ σ n ( r , x , z ) − σ ( r , x , z ) ∣ ≤ K e a R ( 2 + 2 R ) m n − γ , n > R + 1. \sup_{r\le T,\ |x|,|z|\le R}|\sigma_n(r,x,z)-\sigma(r,x,z)| \le Ke^{aR}(2+2R)m n^{-\gamma}, \qquad n>R+1. r ≤ T , ∣ x ∣ , ∣ z ∣ ≤ R sup ∣ σ n ( r , x , z ) − σ ( r , x , z ) ∣ ≤ K e a R ( 2 + 2 R ) m n − γ , n > R + 1.
Let W ( d s , d u ) W(\mathrm{d}s,\mathrm{d}u) W ( d s , d u ) be a space-time white noise with intensity d s d u \mathrm{d}s\,\mathrm{d}u d s d u .
Lemma 4.1. For each n ≥ 1 n \ge1 n ≥ 1 , equation (1.1) with σ \sigma σ replaced by σ n \sigma_n σ n admits a predictable mild solution u n u^n u n with initial condition u 0 u_0 u 0 , driven by the given white noise W W W . Moreover, for each T > 0 T>0 T > 0 , μ > 0 \mu>0 μ > 0 and p > 8 p>8 p > 8 ,
sup n ≥ 1 , 0 ≤ t ≤ T , x ∈ R e − 2 μ ∣ x ∣ ∥ u t n ( x ) ∥ L p ( Ω ) 2 < ∞ . (41) \sup_{n\ge1,\ 0\le t\le T,\ x\in\mathbb{R}}e^{-2\mu|x|}\|u_t^n(x)\|_{L^p(\Omega)}^2<\infty. \tag*{(41)} n ≥ 1 , 0 ≤ t ≤ T , x ∈ R sup e − 2 μ ∣ x ∣ ∥ u t n ( x ) ∥ L p ( Ω ) 2 < ∞. (41)
Proof. Fix p > 8 p>8 p > 8 and μ > 0 \mu>0 μ > 0 . For each n n n , set V n 0 ( t , x ) : = P t u 0 ( x ) V_n^0(t,x):=P_tu_0(x) V n 0 ( t , x ) := P t u 0 ( x ) , with P 0 u 0 = u 0 P_0u_0=u_0 P 0 u 0 = u 0 . For j ≥ 0 j\ge0 j ≥ 0 , set V n j + 1 ( 0 , x ) = u 0 ( x ) V_n^{j+1}(0,x)=u_0(x) V n j + 1 ( 0 , x ) = u 0 ( x ) . For 0 < t ≤ T 0<t\le T 0 < t ≤ T , define
V n j + 1 ( t , x ) : = P t u 0 ( x ) + ∫ 0 t ∫ R p t − v ( x − y ) b ( v , y , V n j ( v , y ) ) d y d v + ∫ 0 t ∫ R p t − v ( x − y ) σ n ( v , y , V n j ( v , y ) ) W ( d v , d y ) . (42) \begin{aligned}
V_n^{j+1}(t,x) := {}& P_tu_0(x)+\int_0^t\int_{\mathbb{R}}p_{t-v}(x-y)b(v,y,V_n^j(v,y))\,\mathrm{d}y\,\mathrm{d}v \\
&+\int_0^t\int_{\mathbb{R}}p_{t-v}(x-y)\sigma_n(v,y,V_n^j(v,y))W(\mathrm{d}v,\mathrm{d}y). \tag*{(42)}
\end{aligned} V n j + 1 ( t , x ) := P t u 0 ( x ) + ∫ 0 t ∫ R p t − v ( x − y ) b ( v , y , V n j ( v , y )) d y d v + ∫ 0 t ∫ R p t − v ( x − y ) σ n ( v , y , V n j ( v , y )) W ( d v , d y ) . (42)
Let
D n , j ( t ) = sup x e − 2 μ ∣ x ∣ ∥ V n j + 1 ( t , x ) − V n j ( t , x ) ∥ L p ( Ω ) 2 . D_{n,j}(t)=\sup_x e^{-2\mu|x|}\|V_n^{j+1}(t,x)-V_n^j(t,x)\|_{L^p(\Omega)}^2. D n , j ( t ) = x sup e − 2 μ ∣ x ∣ ∥ V n j + 1 ( t , x ) − V n j ( t , x ) ∥ L p ( Ω ) 2 .
Let L n = C ( 1 + n ) ∥ φ n ′ ∥ 1 L_n=C(1+n)\|\varphi_n'\|_1 L n = C ( 1 + n ) ∥ φ n ′ ∥ 1 . For j ≥ 1 j\ge1 j ≥ 1 , define the coefficient differences
Δ b v ( y ) = b ( v , y , V n j ( v , y ) ) − b ( v , y , V n j − 1 ( v , y ) ) , \Delta b_v(y)=b(v,y,V_n^j(v,y))-b(v,y,V_n^{j-1}(v,y)), Δ b v ( y ) = b ( v , y , V n j ( v , y )) − b ( v , y , V n j − 1 ( v , y )) ,
Δ σ v ( y ) = σ n ( v , y , V n j ( v , y ) ) − σ n ( v , y , V n j − 1 ( v , y ) ) . \Delta\sigma_v(y)=\sigma_n(v,y,V_n^j(v,y))-\sigma_n(v,y,V_n^{j-1}(v,y)). Δ σ v ( y ) = σ n ( v , y , V n j ( v , y )) − σ n ( v , y , V n j − 1 ( v , y )) .
It follows from (H)(ii) and (4.1) that
∥ Δ b v ( y ) ∥ L p ( Ω ) ≤ L ∥ V n j ( v , y ) − V n j − 1 ( v , y ) ∥ L p ( Ω ) ≤ L e μ ∣ y ∣ D n , j − 1 ( v ) 1 / 2 , \|\Delta b_v(y)\|_{L^p(\Omega)}\le L\|V_n^j(v,y)-V_n^{j-1}(v,y)\|_{L^p(\Omega)}
\le Le^{\mu|y|}D_{n,j-1}(v)^{1/2}, ∥Δ b v ( y ) ∥ L p ( Ω ) ≤ L ∥ V n j ( v , y ) − V n j − 1 ( v , y ) ∥ L p ( Ω ) ≤ L e μ ∣ y ∣ D n , j − 1 ( v ) 1/2 ,
∥ Δ σ v ( y ) ∥ L p ( Ω ) ≤ L n e μ ∣ y ∣ D n , j − 1 ( v ) 1 / 2 . \|\Delta\sigma_v(y)\|_{L^p(\Omega)}\le L_ne^{\mu|y|}D_{n,j-1}(v)^{1/2}. ∥Δ σ v ( y ) ∥ L p ( Ω ) ≤ L n e μ ∣ y ∣ D n , j − 1 ( v ) 1/2 .
By (4.3),
V n j + 1 ( t , x ) − V n j ( t , x ) = I t ( x ) + J t ( x ) , (43) V_n^{j+1}(t,x)-V_n^j(t,x)=I_t(x)+J_t(x), \tag*{(43)} V n j + 1 ( t , x ) − V n j ( t , x ) = I t ( x ) + J t ( x ) , (43)
where
I t ( x ) = ∫ 0 t ∫ R p t − v ( x − y ) Δ b v ( y ) d y d v , I_t(x)=\int_0^t\int_{\mathbb{R}}p_{t-v}(x-y)\Delta b_v(y)\,\mathrm{d}y\,\mathrm{d}v, I t ( x ) = ∫ 0 t ∫ R p t − v ( x − y ) Δ b v ( y ) d y d v ,
J t ( x ) = ∫ 0 t ∫ R p t − v ( x − y ) Δ σ v ( y ) W ( d v , d y ) . J_t(x)=\int_0^t\int_{\mathbb{R}}p_{t-v}(x-y)\Delta\sigma_v(y)W(\mathrm{d}v,\mathrm{d}y). J t ( x ) = ∫ 0 t ∫ R p t − v ( x − y ) Δ σ v ( y ) W ( d v , d y ) .
Observe that there is a constant C ~ T > 0 \widetilde{C}_{T}>0 C T > 0 such that
∥ e μ ∣ ⋅ ∣ p s ∥ 1 + s 1 / 2 ∥ e μ ∣ ⋅ ∣ p s ∥ 2 2 ≤ C ~ T , 0 < s ≤ T . \left\|e^{\mu|\cdot|}p_s\right\|_1+s^{1/2}\left\|e^{\mu|\cdot|}p_s\right\|_2^2\leq\widetilde{C}_{T},\qquad0<s\leq T. e μ ∣ ⋅ ∣ p s 1 + s 1/2 e μ ∣ ⋅ ∣ p s 2 2 ≤ C T , 0 < s ≤ T .
Then Minkowski’s inequality and Cauchy-Schwarz’s inequality give
e − μ ∣ x ∣ ∥ I t ( x ) ∥ L p ( Ω ) ≤ L ∫ 0 t ∥ e μ ∣ ⋅ ∣ p t − v ∥ 1 D n , j − 1 ( v ) 1 / 2 d v ≤ L C ~ T t 1 / 2 [ ∫ 0 t D n , j − 1 ( v ) d v ] 1 / 2 . (44) e^{-\mu|x|}\|I_t(x)\|_{L^p(\Omega)}\leq L\int_0^t\left\|e^{\mu|\cdot|}p_{t-v}\right\|_1D_{n,j-1}(v)^{1/2}\,dv\leq L\widetilde{C}_Tt^{1/2}\left[\int_0^tD_{n,j-1}(v)\,dv\right]^{1/2}. \tag*{(44)} e − μ ∣ x ∣ ∥ I t ( x ) ∥ L p ( Ω ) ≤ L ∫ 0 t e μ ∣ ⋅ ∣ p t − v 1 D n , j − 1 ( v ) 1/2 d v ≤ L C T t 1/2 [ ∫ 0 t D n , j − 1 ( v ) d v ] 1/2 . (44)
Applying the Burkholder–Davis–Gundy inequality and then Minkowski’s inequality one obtains
e − 2 μ ∣ x ∣ ∥ J t ( x ) ∥ L p ( Ω ) 2 ≤ C p e − 2 μ ∣ x ∣ ∥ ∫ 0 t ∫ R p t − v ( x − y ) 2 ∣ Δ σ v ( y ) ∣ 2 d y d v ∥ L p / 2 ( Ω ) ≤ C p e − 2 μ ∣ x ∣ ∫ 0 t ∫ R p t − v ( x − y ) 2 ∥ Δ σ v ( y ) ∥ L p ( Ω ) 2 d y d v ≤ C p L n 2 ∫ 0 t ∥ e μ ∣ ⋅ ∣ p t − v ∥ 2 2 D n , j − 1 ( v ) d v ≤ C p L n 2 C ~ T ∫ 0 t ( t − v ) − 1 / 2 D n , j − 1 ( v ) d v . (45) \begin{aligned}
e^{-2\mu|x|}\|J_t(x)\|_{L^p(\Omega)}^2\leq C_pe^{-2\mu|x|}\left\|\int_0^t\int_{\mathbb{R}}p_{t-v}(x-y)^2|\Delta\sigma_v(y)|^2\,dy\,dv\right\|_{L^{p/2}(\Omega)} \\
&\leq C_pe^{-2\mu|x|}\int_0^t\int_{\mathbb{R}}p_{t-v}(x-y)^2\|\Delta\sigma_v(y)\|_{L^p(\Omega)}^2\,dy\,dv \\
&\leq C_pL_n^2\int_0^t\left\|e^{\mu|\cdot|}p_{t-v}\right\|_2^2D_{n,j-1}(v)\,dv \\
&\leq C_pL_n^2\widetilde{C}_T\int_0^t(t-v)^{-1/2}D_{n,j-1}(v)\,dv. \tag*{(45)}
\end{aligned} e − 2 μ ∣ x ∣ ∥ J t ( x ) ∥ L p ( Ω ) 2 ≤ C p e − 2 μ ∣ x ∣ ∫ 0 t ∫ R p t − v ( x − y ) 2 ∣Δ σ v ( y ) ∣ 2 d y d v L p /2 ( Ω ) ≤ C p e − 2 μ ∣ x ∣ ∫ 0 t ∫ R p t − v ( x − y ) 2 ∥Δ σ v ( y ) ∥ L p ( Ω ) 2 d y d v ≤ C p L n 2 ∫ 0 t e μ ∣ ⋅ ∣ p t − v 2 2 D n , j − 1 ( v ) d v ≤ C p L n 2 C T ∫ 0 t ( t − v ) − 1/2 D n , j − 1 ( v ) d v . (45)
for some constant C p > 0 C_p>0 C p > 0 . Combining (44) and (45) gives, for some C n C_n C n independent of j j j and t t t ,
D n , j ( t ) ≤ C n ∫ 0 t ( 1 + ( t − v ) − 1 / 2 ) D n , j − 1 ( v ) d v , j ≥ 1 , 0 ≤ t ≤ T . (46) D_{n,j}(t)\leq C_n\int_0^t\left(1+(t-v)^{-1/2}\right)D_{n,j-1}(v)\,dv,\qquad j\geq1,\quad0\leq t\leq T. \tag*{(46)} D n , j ( t ) ≤ C n ∫ 0 t ( 1 + ( t − v ) − 1/2 ) D n , j − 1 ( v ) d v , j ≥ 1 , 0 ≤ t ≤ T . (46)
Let n ≥ 1 n\geq1 n ≥ 1 be fixed. Define
∥ X ∥ β , T : = sup 0 ≤ t ≤ T , x ∈ R e − β t / 2 − μ ∣ x ∣ ∥ X ( t , x ) ∥ L p ( Ω ) . \|X\|_{\beta,T}:=\sup_{0\leq t\leq T,\ x\in\mathbb{R}}e^{-\beta t/2-\mu|x|}\|X(t,x)\|_{L^p(\Omega)}. ∥ X ∥ β , T := 0 ≤ t ≤ T , x ∈ R sup e − β t /2 − μ ∣ x ∣ ∥ X ( t , x ) ∥ L p ( Ω ) .
Multiplying (46) by e − β t e^{-\beta t} e − β t and taking the supremum gives
∥ V n j + 1 − V n j ∥ β , T 2 ≤ q ∥ V n j − V n j − 1 ∥ β , T 2 , q = C n ∫ 0 ∞ e − β s ( 1 + s − 1 / 2 ) d s = C n ( β − 1 + π / β ) . \|V_n^{j+1}-V_n^j\|_{\beta,T}^2\leq q\|V_n^j-V_n^{j-1}\|_{\beta,T}^2,\qquad q=C_n\int_0^\infty e^{-\beta s}(1+s^{-1/2})\,ds=C_n(\beta^{-1}+\sqrt{\pi/\beta}). ∥ V n j + 1 − V n j ∥ β , T 2 ≤ q ∥ V n j − V n j − 1 ∥ β , T 2 , q = C n ∫ 0 ∞ e − β s ( 1 + s − 1/2 ) d s = C n ( β − 1 + π / β ) .
Choose β \beta β so that q < 1 q<1 q < 1 . The linear growth bounds and ∣ P v u 0 ( y ) ∣ ≤ C ~ T ∣ u 0 ∣ μ e μ ∣ y ∣ |P_vu_0(y)|\leq\widetilde{C}_T|u_0|_\mu e^{\mu|y|} ∣ P v u 0 ( y ) ∣ ≤ C T ∣ u 0 ∣ μ e μ ∣ y ∣ give ∥ V n 1 − V n 0 ∥ β , T < ∞ \|V_n^1-V_n^0\|_{\beta,T}<\infty ∥ V n 1 − V n 0 ∥ β , T < ∞ by the same convolution estimates. Thus the successive differences are summable in ∥ ⋅ ∥ β , T \|\cdot\|_{\beta,T} ∥ ⋅ ∥ β , T and absolutely summable almost surely at each ( t , x ) (t,x) ( t , x ) . Their pointwise sum, added to V n 0 V_n^0 V n 0 and defined as zero where the series diverges, gives a predictable field u n u^n u n with ∥ V n j − u n ∥ β , T → 0 \|V_n^j-u^n\|_{\beta,T}\to0 ∥ V n j − u n ∥ β , T → 0 . Applying (44) and (45) to the coefficient differences between V n j V_n^j V n j and u n u^n u n shows that both convolutions in (42) converge in L p ( Ω ) L^p(\Omega) L p ( Ω ) . Hence u n u^n u n satisfies the mild equation with coefficients ( b , σ n ) (b,\sigma_n) ( b , σ n ) . The recursion uses the same noise on each finite horizon, so its limits on overlapping horizons agree and define u n u^n u n on R + \mathbb{R}_+ R + .
Set M n ( t ) = sup x e − 2 μ ∣ x ∣ ∥ u n ( t , x ) ∥ L p ( Ω ) 2 M_n(t)=\sup_x e^{-2\mu|x|}\|u^n(t,x)\|_{L^p(\Omega)}^2 M n ( t ) = sup x e − 2 μ ∣ x ∣ ∥ u n ( t , x ) ∥ L p ( Ω ) 2 . By (H)(i) and (40) , there is a constant C 1 > 0 C_1>0 C 1 > 0 independent of n n n such that
e − μ ∣ y ∣ max { ∥ b ( v , y , u n ( v , y ) ) ∥ L p ( Ω ) , ∥ σ n ( v , y , u n ( v , y ) ) ∥ L p ( Ω ) } ≤ C 1 ( 2 + M n ( v ) 1 / 2 ) . e^{-\mu|y|}\max\left\{\|b(v,y,u^n(v,y))\|_{L^p(\Omega)},\|\sigma_n(v,y,u^n(v,y))\|_{L^p(\Omega)}\right\}\leq C_1(2+M_n(v)^{1/2}). e − μ ∣ y ∣ max { ∥ b ( v , y , u n ( v , y )) ∥ L p ( Ω ) , ∥ σ n ( v , y , u n ( v , y )) ∥ L p ( Ω ) } ≤ C 1 ( 2 + M n ( v ) 1/2 ) .
Then, by the estimates in (44) and (45) , the mild equation gives
M n ( t ) ≤ C 3 + C 3 ∫ 0 t ( 1 + ( t − v ) − 1 / 2 ) M n ( v ) d v (47) M_n(t)\leq C_3+C_3\int_0^t\left(1+(t-v)^{-1/2}\right)M_n(v)\,dv \tag*{(47)} M n ( t ) ≤ C 3 + C 3 ∫ 0 t ( 1 + ( t − v ) − 1/2 ) M n ( v ) d v (47)
for a constant C 3 > 0 C_3>0 C 3 > 0 independent of n n n . Put B n = sup 0 ≤ t ≤ T e − θ t M n ( t ) B_n=\sup_{0\leq t\leq T}e^{-\theta t}M_n(t) B n = sup 0 ≤ t ≤ T e − θ t M n ( t ) , which is finite for each n n n by Picard convergence. Multiplying (47) by e − θ t e^{-\theta t} e − θ t and taking the supremum gives
B n ≤ C 3 + C 3 ( θ − 1 + π θ − 1 / 2 ) B n . B_n\leq C_3+C_3(\theta^{-1}+\sqrt{\pi}\theta^{-1/2})B_n. B n ≤ C 3 + C 3 ( θ − 1 + π θ − 1/2 ) B n .
Choose one θ > 0 \theta>0 θ > 0 so that the coefficient of B n B_n B n is at most 1 / 2 1/2 1/2 . Then (41) follows.
Lemma 4.2 (Weak existence). Assume (H). Then, for every deterministic u 0 ∈ C t e m u_0 \in C_{\mathrm{tem}} u 0 ∈ C tem , there is a filtered probability space carrying a space-time white noise W W W and a weak solution u u u with initial function $u_0 in the sense of Definition 1.1 .
Proof. Let u n u^n u n be the approximate solutions from Lemma 4.1. The proof has two steps. First, we use the Kolmogorov–Chentsov theorem for continuous modifications and its tightness corollary for local tightness [4 Theorem 3.23 and Corollary 16.9] , then use a spatial tail estimate to obtain tightness in C ( R + ; C t e m ) C(\mathbb{R}_+; C_{\mathrm{tem}}) C ( R + ; C tem ) . Second, we take a joint limit of the approximate solutions and the Brownian coordinates of the noise. We pass the martingale and covariation identities to the limit to recover the Walsh integral in the original equation. This is the approximation strategy of [8 proof of Theorem 1.1] .
Step 1. The estimates below hold for every choice of μ > 0 \mu> 0 μ > 0 ; we may therefore choose the weight parameter separately for each seminorm of C t e m C_{\mathrm{tem}} C tem . Fix p > 8 p > 8 p > 8 and T > 0 T > 0 T > 0 , and write Y t n = u t n − P t u 0 Y_t^n = u_t^n - P_tu_0 Y t n = u t n − P t u 0 . The constant C ^ \widehat C C below is uniform in n n n , j j j and the space-time variables. It may depend on p p p , T T T , μ \mu μ , ∣ u 0 ∣ μ |u_0|_\mu ∣ u 0 ∣ μ and the coefficient bounds. By (41) , (H)(i) and (40) ,
∥ b ( v , z , u n ( v , z ) ) ∥ L p ( Ω ) + ∥ σ n ( v , z , u n ( v , z ) ) ∥ L p ( Ω ) ≤ C ^ e μ ∣ z ∣ , v ≤ T . \lVert b(v,z,u^n(v,z))\rVert_{L^p(\Omega)}+\lVert\sigma_n(v,z,u^n(v,z))\rVert_{L^p(\Omega)}\leq\widehat C e^{\mu|z|},\qquad v\leq T. ∥ b ( v , z , u n ( v , z )) ∥ L p ( Ω ) + ∥ σ n ( v , z , u n ( v , z )) ∥ L p ( Ω ) ≤ C e μ ∣ z ∣ , v ≤ T .
For i = 0 , 1 , 2 i=0,1,2 i = 0 , 1 , 2 , q = 1 , 2 q=1,2 q = 1 , 2 , the Gaussian kernel satisfies
∥ e μ ∣ ⋅ ∣ ∂ x i p a ∥ q ≤ C ^ a − i / 2 − ( 1 − 1 / q ) / 2 , 0 < a ≤ T . \lVert e^{\mu|\cdot|}\partial_x^i p_a\rVert_q\leq\widehat C a^{-i/2-(1-1/q)/2},\qquad0<a\leq T. ∥ e μ ∣ ⋅ ∣ ∂ x i p a ∥ q ≤ C a − i /2 − ( 1 − 1/ q ) /2 , 0 < a ≤ T .
For j ∈ Z j\in\mathbb{Z} j ∈ Z , fix x , y ∈ [ j , j + 1 ] x,y\in[j,j+1] x , y ∈ [ j , j + 1 ] , put d = ∣ x − y ∣ d=|x-y| d = ∣ x − y ∣ , and take 0 ≤ s < t ≤ T 0\leq s<t\leq T 0 ≤ s < t ≤ T with ℓ = t − s \ell=t-s ℓ = t − s . The triangle inequality for the exponential weight and the spatial derivative estimate give
∥ e μ ∣ ⋅ ∣ ( p a ( ⋅ + x − y ) − p a ) ∥ q ≤ C ^ a − ( 1 − 1 / q ) / 2 min { 1 , d / a } . \lVert e^{\mu|\cdot|}(p_a(\cdot+x-y)-p_a)\rVert_q\leq\widehat C a^{-(1-1/q)/2}\min\{1,d/\sqrt a\}. ∥ e μ ∣ ⋅ ∣ ( p a ( ⋅ + x − y ) − p a ) ∥ q ≤ C a − ( 1 − 1/ q ) /2 min { 1 , d / a } .
Similarly, ∂ a p a = 1 2 ∂ x x p a \partial_a p_a=\frac{1}{2}\partial_{xx}p_a ∂ a p a = 2 1 ∂ xx p a gives
∥ e μ ∣ ⋅ ∣ ( p a + ℓ − p a ) ∥ q ≤ C ^ a − ( 1 − 1 / q ) / 2 min { 1 , ℓ / a } , a + ℓ ≤ T . \lVert e^{\mu|\cdot|}(p_{a+\ell}-p_a)\rVert_q\leq\widehat C a^{-(1-1/q)/2}\min\{1,\ell/a\},\qquad a+\ell\leq T. ∥ e μ ∣ ⋅ ∣ ( p a + ℓ − p a ) ∥ q ≤ C a − ( 1 − 1/ q ) /2 min { 1 , ℓ / a } , a + ℓ ≤ T .
The same estimates as in (44) and (45) give
∥ Y t n ( x ) − Y s n ( y ) ∥ L p ( Ω ) ≤ C 1 e μ ∣ j ∣ [ d + ℓ ( 2 + log ( T / ℓ ) ) + ( ∫ 0 T a − 1 / 2 min { 1 , d 2 / a } d a ) 1 / 2 + ( ∫ 0 T a − 1 / 2 min { 1 , ℓ 2 / a 2 } d a + 2 ℓ ) 1 / 2 ] ≤ C 2 e μ ∣ j ∣ ( d 1 / 2 + ℓ 1 / 4 ) \begin{aligned}
\lVert Y_t^n(x)-Y_s^n(y)\rVert_{L^p(\Omega)}
&\leq C_1e^{\mu|j|}\left[d+\ell(2+\log(T/\ell))\right.\\
&\quad+\left(\int_0^T a^{-1/2}\min\{1,d^2/a\}\,da\right)^{1/2}\\
&\quad\left.+\left(\int_0^T a^{-1/2}\min\{1,\ell^2/a^2\}\,da+2\sqrt{\ell}\right)^{1/2}\right]\\
&\leq C_2e^{\mu|j|}(d^{1/2}+\ell^{1/4})
\end{aligned} ∥ Y t n ( x ) − Y s n ( y ) ∥ L p ( Ω ) ≤ C 1 e μ ∣ j ∣ [ d + ℓ ( 2 + log ( T / ℓ )) + ( ∫ 0 T a − 1/2 min { 1 , d 2 / a } d a ) 1/2 + ( ∫ 0 T a − 1/2 min { 1 , ℓ 2 / a 2 } d a + 2 ℓ ) 1/2 ≤ C 2 e μ ∣ j ∣ ( d 1/2 + ℓ 1/4 )
for some constants C 1 , C 2 > 0 C_1,C_2>0 C 1 , C 2 > 0 independent of n n n . On Q j = [ 0 , T ] × [ j , j + 1 ] Q_j=[0,T]\times[j,j+1] Q j = [ 0 , T ] × [ j , j + 1 ] , the increment estimate reads
E ∣ Y n ( z ) − Y n ( z ′ ) ∣ p ≤ C 3 e p μ ∣ j ∣ ∣ z − z ′ ∣ p / 4 , z , z ′ ∈ Q j . (48) \mathbb{E}|Y^n(z)-Y^n(z')|^p\leq C_3e^{p\mu|j|}|z-z'|^{p/4},\qquad z,z'\in Q_j. \tag*{(48)} E ∣ Y n ( z ) − Y n ( z ′ ) ∣ p ≤ C 3 e p μ ∣ j ∣ ∣ z − z ′ ∣ p /4 , z , z ′ ∈ Q j . (48)
for some constant C 3 > 0 C_3>0 C 3 > 0 independent of n n n and j j j . For 0 < α < 1 0<\alpha<1 0 < α < 1 , let C α ( Q j ) C^\alpha(Q_j) C α ( Q j ) be the space of functions f : Q j → R f:Q_j\to\mathbb{R} f : Q j → R with finite norm
∥ f ∥ C α ( Q j ) : = sup z ∈ Q j ∣ f ( z ) ∣ + sup z , z ′ ∈ Q j z ≠ z ′ ∣ f ( z ) − f ( z ′ ) ∣ ∣ z − z ′ ∣ α . \lVert f\rVert_{C^\alpha(Q_j)}:=\sup_{z\in Q_j}|f(z)|+\sup_{\substack{z,z'\in Q_j\\z\neq z'}}\frac{|f(z)-f(z')|}{|z-z'|^\alpha}. ∥ f ∥ C α ( Q j ) := z ∈ Q j sup ∣ f ( z ) ∣ + z , z ′ ∈ Q j z = z ′ sup ∣ z − z ′ ∣ α ∣ f ( z ) − f ( z ′ ) ∣ .
where ∣ z − z ′ ∣ |z-z'| ∣ z − z ′ ∣ is the Euclidean distance. By the dyadic estimate in the proof of [4 Theorem 3.23] and Y n ( 0 , j ) = 0 Y^n(0,j)=0 Y n ( 0 , j ) = 0 , ( t , x ) ↦ Y t n ( x ) (t,x)\mapsto Y_t^n(x) ( t , x ) ↦ Y t n ( x ) has a continuous modification for each n n n , and
E ∥ Y n ∥ C α ( Q j ) p ≤ C α e p μ ∣ j ∣ , 0 < α < 1 4 − 2 p . (49) \mathbb{E}\lVert Y^n\rVert_{C^\alpha(Q_j)}^p\leq C_\alpha e^{p\mu|j|},\qquad0<\alpha<\frac14-\frac2p. \tag*{(49)} E ∥ Y n ∥ C α ( Q j ) p ≤ C α e p μ ∣ j ∣ , 0 < α < 4 1 − p 2 . (49)
for some constant C α > 0 C_{\alpha}>0 C α > 0 independent of n n n and j j j . By [4 Corollary 16.9] , the laws of Y n ∣ Q j Y^{n}|_{Q_j} Y n ∣ Q j are tight in C ( Q j ) C(Q_j) C ( Q j ) .
For λ > μ \lambda>\mu λ > μ , let H R n = sup ∣ j ∣ > R , ( t , x ) ∈ Q j e − λ ∣ x ∣ ∣ Y t n ( x ) ∣ H_R^n=\sup_{|j|>R,\,(t,x)\in Q_j}e^{-\lambda|x|}|Y_t^n(x)| H R n = sup ∣ j ∣ > R , ( t , x ) ∈ Q j e − λ ∣ x ∣ ∣ Y t n ( x ) ∣ . Since ∣ x ∣ ≥ ∣ j ∣ − 1 |x|\geq|j|-1 ∣ x ∣ ≥ ∣ j ∣ − 1 on Q j Q_j Q j , the union bound, Markov’s inequality and (49) give
P ( H R n > δ ) ≤ δ − p ∑ ∣ j ∣ > R e − p λ ( ∣ j ∣ − 1 ) E ∥ Y n ∥ C α ( Q j ) p ≤ C α e p λ δ − p ∑ ∣ j ∣ > R e − p ( λ − μ ) ∣ j ∣ ⟶ 0 (50) \mathbb{P}(H_R^n>\delta)\leq\delta^{-p}\sum_{|j|>R}e^{-p\lambda(|j|-1)}\mathbb{E}\|Y^n\|_{C^\alpha(Q_j)}^p\leq C_\alpha e^{p\lambda}\delta^{-p}\sum_{|j|>R}e^{-p(\lambda-\mu)|j|}\longrightarrow0 \tag*{(50)} P ( H R n > δ ) ≤ δ − p ∣ j ∣ > R ∑ e − p λ ( ∣ j ∣ − 1 ) E ∥ Y n ∥ C α ( Q j ) p ≤ C α e p λ δ − p ∣ j ∣ > R ∑ e − p ( λ − μ ) ∣ j ∣ ⟶ 0 (50)
uniformly in n n n as R → ∞ R\to\infty R → ∞ . For each n n n , H R n H_R^n H R n decreases in R R R , so the above probability bound implies H R n ↓ 0 H_R^n\downarrow0 H R n ↓ 0 almost surely. For t , s ∈ [ 0 , T ] t,s\in[0,T] t , s ∈ [ 0 , T ] and integers R ≥ 1 R\geq1 R ≥ 1 ,
∣ Y t n − Y s n ∣ λ ≤ sup ∣ x ∣ ≤ R + 1 ∣ Y t n ( x ) − Y s n ( x ) ∣ + 2 H R n . (51) |Y_t^n-Y_s^n|_\lambda\leq\sup_{|x|\leq R+1}|Y_t^n(x)-Y_s^n(x)|+2H_R^n. \tag*{(51)} ∣ Y t n − Y s n ∣ λ ≤ ∣ x ∣ ≤ R + 1 sup ∣ Y t n ( x ) − Y s n ( x ) ∣ + 2 H R n . (51)
The first term of (51) tends to zero as s → t s\to t s → t by local continuity. Letting R → ∞ R\to\infty R → ∞ in (51) proves continuity in ∣ ⋅ ∣ λ |\cdot|_\lambda ∣ ⋅ ∣ λ . Taking T = m T=m T = m and λ = 1 / k \lambda=1/k λ = 1/ k for m , k ≥ 1 m,k\geq1 m , k ≥ 1 , with μ < 1 / k \mu<1/k μ < 1/ k , gives Y n ∈ C ( R + ; C t e m ) Y^n\in C(\mathbb{R}_+;C_{\mathrm{tem}}) Y n ∈ C ( R + ; C tem ) almost surely, and hence the same for u n = Y n + P ⋅ u 0 u^n=Y^n+P\cdot u_0 u n = Y n + P ⋅ u 0 .
Local tightness on K R = [ 0 , m ] × [ − R , R ] K_R=[0,m]\times[-R,R] K R = [ 0 , m ] × [ − R , R ] , together with (50) and the Arzelà–Ascoli theorem [9 Theorem 7.25] , gives tightness of { Y n } n ≥ 1 \{Y^n\}_{n\geq1} { Y n } n ≥ 1 on [ 0 , m ] [0,m] [ 0 , m ] in the norm ∣ ⋅ ∣ 1 / k |\cdot|_{1/k} ∣ ⋅ ∣ 1/ k . A diagonal argument over m , k ∈ N m,k\in\mathbb{N} m , k ∈ N then gives tightness of ( Y n ) (Y^n) ( Y n ) , and hence of ( u n ) (u^n) ( u n ) , in the Polish space C ( R + ; C t e m ) C(\mathbb{R}_+;C_{\mathrm{tem}}) C ( R + ; C tem ) .
Step 2. Let ( e k ) k ≥ 1 (e_k)_{k\geq1} ( e k ) k ≥ 1 be a real orthonormal basis of L 2 ( R , d x ) L^2(\mathbb{R},dx) L 2 ( R , d x ) and set B t k : = W ( 1 [ 0 , t ] e k ) B_t^k:=W(1_{[0,t]}e_k) B t k := W ( 1 [ 0 , t ] e k ) . The joint laws of ( u n , ( B k ) k ≥ 1 ) (u^n,(B^k)_{k\geq1}) ( u n , ( B k ) k ≥ 1 ) are tight in C ( R + ; C t e m ) × C ( R + ; R ) N C(\mathbb{R}_+;C_{\mathrm{tem}})\times C(\mathbb{R}_+;\mathbb{R})^{\mathbb{N}} C ( R + ; C tem ) × C ( R + ; R ) N , since the law of the Brownian coordinates is fixed. By the Skorokhod representation theorem, after taking a subsequence and changing probability space, we may assume that ( u n , ( B n , k ) k ≥ 1 ) → ( u , ( B k ) k ≥ 1 ) (u^n,(B^{n,k})_{k\geq1})\to(u,(B^k)_{k\geq1}) ( u n , ( B n , k ) k ≥ 1 ) → ( u , ( B k ) k ≥ 1 ) almost surely. For ϕ ∈ C c ∞ ( R ) \phi\in C^\infty_c(\mathbb{R}) ϕ ∈ C c ∞ ( R ) , define
M t ϕ : = ⟨ u t − u 0 , ϕ ⟩ − ∫ 0 t [ 1 2 ⟨ u v , ϕ ′ ′ ⟩ + ⟨ b ( v , ⋅ , u v ) , ϕ ⟩ ] d v , g v ( x ) : = ϕ ( x ) σ ( v , x , u v ( x ) ) . M_t^\phi:=\langle u_t-u_0,\phi\rangle-\int_0^t\left[\frac12\langle u_v,\phi''\rangle+\langle b(v,\cdot,u_v),\phi\rangle\right]\,dv,
\qquad
g_v(x):=\phi(x)\sigma(v,x,u_v(x)). M t ϕ := ⟨ u t − u 0 , ϕ ⟩ − ∫ 0 t [ 2 1 ⟨ u v , ϕ ′′ ⟩ + ⟨ b ( v , ⋅ , u v ) , ϕ ⟩ ] d v , g v ( x ) := ϕ ( x ) σ ( v , x , u v ( x )) .
Define M n , ϕ M^{n,\phi} M n , ϕ and g n g^n g n by the same formulas with ( u , σ ) (u,\sigma) ( u , σ ) replaced by ( u n , σ n ) (u^n,\sigma_n) ( u n , σ n ) . Write W n W^n W n for the white noise represented by ( B n , k ) k ≥ 1 (B^{n,k})_{k\geq1} ( B n , k ) k ≥ 1 in the filtration of the n n n th pair. The martingale identities below hold for each represented pair because they depend only on its joint law. Testing the mild equation for u n u^n u n against ϕ \phi ϕ (see [10 Theorem 2.1] ) gives
M t n , ϕ = ∫ 0 t ∫ R g v n ( x ) W n ( d v , d x ) . M_t^{n,\phi}=\int_0^t\int_{\mathbb{R}}g_v^n(x)\,W^n(dv,dx). M t n , ϕ = ∫ 0 t ∫ R g v n ( x ) W n ( d v , d x ) .
It follows that
[ B n , i , B n , k ] t = 1 { i = k } t , [ M n , ϕ ] t = ∫ 0 t ∥ g v n ∥ 2 2 d v , [ M n , ϕ , B n , k ] t = ∫ 0 t ⟨ g v n , e k ⟩ d v . [B^{n,i},B^{n,k}]_t=1_{\{i=k\}}t,\qquad[M^{n,\phi}]_t=\int_0^t\|g_v^n\|_2^2\,dv,\qquad[M^{n,\phi},B^{n,k}]_t=\int_0^t\langle g_v^n,e_k\rangle\,dv. [ B n , i , B n , k ] t = 1 { i = k } t , [ M n , ϕ ] t = ∫ 0 t ∥ g v n ∥ 2 2 d v , [ M n , ϕ , B n , k ] t = ∫ 0 t ⟨ g v n , e k ⟩ d v .
The locally uniform convergence u n → u u^n\to u u n → u , the local convergence σ n → σ \sigma_n\to\sigma σ n → σ , and (H)(ii)–(iii) imply sup t ≤ T ∣ M t n , ϕ − M t ϕ ∣ → 0 \sup_{t\leq T}|M_t^{n,\phi}-M_t^\phi|\to0 sup t ≤ T ∣ M t n , ϕ − M t ϕ ∣ → 0 and ∫ 0 T ∥ g v n − g v ∥ 2 2 d v → 0 \int_0^T\|g_v^n-g_v\|_2^2\,dv\to0 ∫ 0 T ∥ g v n − g v ∥ 2 2 d v → 0 as n → ∞ n\to\infty n → ∞ almost surely. Test increments of the martingales and their compensated products, such as M t n , ϕ B t n , k − ∫ 0 t ⟨ g v n , e k ⟩ d v M_t^{n,\phi}B_t^{n,k}-\int_0^t\langle g_v^n,e_k\rangle\,dv M t n , ϕ B t n , k − ∫ 0 t ⟨ g v n , e k ⟩ d v , against bounded continuous functions of the joint past. At each fixed time, (4.2) and the Brownian moment bounds give uniform integrability. Passing to the limit shows that M ϕ M^\phi M ϕ and the B k B^k B k are continuous martingales in the joint filtration, with [ B i , B k ] t = 1 { i = k } t [B^i,B^k]_t=1_{\{i=k\}}t [ B i , B k ] t = 1 { i = k } t . Hence [ M ϕ ] t = ∫ 0 t ∥ g v ∥ 2 2 d v [M^\phi]_t=\int_0^t\|g_v\|_2^2\,dv [ M ϕ ] t = ∫ 0 t ∥ g v ∥ 2 2 d v and [ M ϕ , B k ] t = ∫ 0 t ⟨ g v , e k ⟩ d v [M^\phi,B^k]_t=\int_0^t\langle g_v,e_k\rangle\,dv [ M ϕ , B k ] t = ∫ 0 t ⟨ g v , e k ⟩ d v . Lévy’s characterization makes the B k B^k B k independent Brownian motions in this filtration. They define a space-time white noise W W W on the basis ( e k ) k ≥ 1 (e_k)_{k\geq1} ( e k ) k ≥ 1 . By the bracket identities and Parseval’s identity,
E ∣ M t ϕ − ∑ k = 1 l ∫ 0 t ⟨ g v , e k ⟩ d B v k ∣ 2 = E ∫ 0 t ( ∥ g v ∥ 2 2 − ∑ k = 1 l ∣ ⟨ g v , e k ⟩ ∣ 2 ) d v ⟶ 0 \mathbb{E}\left|M_t^\phi-\sum_{k=1}^{l}\int_0^t\langle g_v,e_k\rangle\,dB_v^k\right|^2
=\mathbb{E}\int_0^t\left(\|g_v\|_2^2-\sum_{k=1}^{l}|\langle g_v,e_k\rangle|^2\right)\,dv\longrightarrow0 E M t ϕ − k = 1 ∑ l ∫ 0 t ⟨ g v , e k ⟩ d B v k 2 = E ∫ 0 t ( ∥ g v ∥ 2 2 − k = 1 ∑ l ∣ ⟨ g v , e k ⟩ ∣ 2 ) d v ⟶ 0
as l → ∞ l \to\infty l → ∞ . Hence
M t ϕ = ∫ 0 t ∫ R ϕ ( x ) σ ( v , x , u w ( x ) ) W ( d v , d x ) . M_t^\phi= \int_0^t \int_{\mathbb{R}} \phi(x)\sigma(v,x,u_w(x)) W(dv,dx). M t ϕ = ∫ 0 t ∫ R ϕ ( x ) σ ( v , x , u w ( x )) W ( d v , d x ) .
Thus u u u is a weak solution to (1.1). □ \square □
Proof of Theorem 1.2. Lemma 4.2 supplies a weak solution, while Proposition 3.1 gives pathwise uniqueness for solutions driven by the same white noise in the same filtration. For each t t t , the noise increments after t t t are independent of the solution and the noise up to t t t , so the solution is temporally compatible with the noise. Kurtz’s Yamada–Watanabe theorem for stochastic models [6 Theorem 1.5 and Lemma 2.10] then yields a solution measurable with respect to the prescribed noise (and hence adapted to its natural filtration by [6 Proposition 2.13] ), together with uniqueness of the joint law of the solution and its driving noise. □ \square □