Introduction
We consider the stochastic differential equation (SDE) on R d \mathbb{R}^d R d :
d X t = b ( t , X t ) d t + σ ( t , X t ) d W t , L X 0 = ν , (1) \mathrm{d}X_t=b(t,X_t)\,\mathrm{d}t+\sigma(t,X_t)\,\mathrm{d}W_t,\qquad\mathcal{L}_{X_0}=\nu, \tag*{(1)} d X t = b ( t , X t ) d t + σ ( t , X t ) d W t , L X 0 = ν , (1)
where ν ∈ P ( R d ) \nu\in\mathcal{P}(\mathbb{R}^d) ν ∈ P ( R d ) and L X t \mathcal{L}_{X_t} L X t denotes the law of X t X_t X t for t ≥ 0 t\ge0 t ≥ 0 . Here
b : [ 0 , ∞ ) × R d → R d , σ : [ 0 , ∞ ) × R d → R d × d b:[0,\infty)\times\mathbb{R}^d\to\mathbb{R}^d,\qquad\sigma:[0,\infty)\times\mathbb{R}^d\to\mathbb{R}^{d\times d} b : [ 0 , ∞ ) × R d → R d , σ : [ 0 , ∞ ) × R d → R d × d
are measurable, and W = ( W t ) t ≥ 0 W = (W_t)_{t\geq0} W = ( W t ) t ≥ 0 is a d d d -dimensional standard Brownian motion on a complete filtered probability space ( Ω , F , ( F t ) t ≥ 0 , P ) (\Omega,\mathcal{F},(\mathcal{F}_t)_{t\geq0},\mathbb{P}) ( Ω , F , ( F t ) t ≥ 0 , P ) . We write a : = σ σ ∗ a := \sigma\sigma^* a := σ σ ∗ for the diffusion matrix. The Fokker–Planck equation (FPE) associated with (1.1) is
∂ t μ t + div ( b μ t ) = 1 2 ∑ i , j = 1 d ∂ i j ( a i j μ t ) , μ 0 = ν . (2) \partial_t\mu_t+\operatorname{div}(b\mu_t)=\frac{1}{2}\sum_{i,j=1}^{d}\partial_{ij}(a_{ij}\mu_t),\qquad\mu_0=\nu. \tag*{(2)} ∂ t μ t + div ( b μ t ) = 2 1 i , j = 1 ∑ d ∂ ij ( a ij μ t ) , μ 0 = ν . (2)
It describes how the law of X t X_t X t evolves in time, and arises naturally in probability theory, statistical physics, and fluid dynamics; see, for example, [3 , 29 ] .
When the diffusion term vanishes, that is, σ = 0 \sigma=0 σ = 0 , the equation reduces to the continuity equation
∂ t μ t + div ( b μ t ) = 0 , μ 0 = ν . (3) \partial_t\mu_t+\operatorname{div}(b\mu_t)=0,\qquad\mu_0=\nu. \tag*{(3)} ∂ t μ t + div ( b μ t ) = 0 , μ 0 = ν . (3)
This equation is closely related to the ordinary differential equation
d Y t = b ( t , Y t ) d t , Y 0 = x . (4) \mathrm{d}Y_t=b(t,Y_t)\,\mathrm{d}t,\qquad Y_0=x. \tag*{(4)} d Y t = b ( t , Y t ) d t , Y 0 = x . (4)
Fix T > 0 T>0 T > 0 . In the classical Cauchy–Lipschitz setting, say b ∈ C 0 , 1 ( [ 0 , T ] × R d ; R d ) b\in C^{0,1}([0,T]\times\mathbb{R}^d;\mathbb{R}^d) b ∈ C 0 , 1 ([ 0 , T ] × R d ; R d ) , the method of characteristics represents the solution of (1.3) as
μ t = ν ∘ Y t − 1 , t ∈ [ 0 , T ] . \mu_t=\nu\circ Y_t^{-1},\qquad t\in[0,T]. μ t = ν ∘ Y t − 1 , t ∈ [ 0 , T ] .
This representation is classical; see, for example, [9 ] . In many applications, however, the Lipschitz assumption on b b b must be weakened. Indeed, the continuity equation is often part of many equations coming from physics. For instance, it is part of the inhomogeneous Navier–Stokes equations, where it expresses conservation of mass. In such models the velocity field is determined by the system itself and is in general not Lipschitz in space. This motivates the study of (1.3) with rough vector fields, to which we now turn.
The theory begins with DiPerna and Lions [10 ] , who proved weak well-posedness of (1.3) in L ∞ ( [ 0 , T ] ; L q ( R d ) ) L^\infty([0,T];L^q(\mathbb{R}^d)) L ∞ ([ 0 , T ] ; L q ( R d )) for b ∈ L 1 ( [ 0 , T ] ; W 1 , p ( R d ; R d ) ) b\in L^1([0,T];W^{1,p}(\mathbb{R}^d;\mathbb{R}^d)) b ∈ L 1 ([ 0 , T ] ; W 1 , p ( R d ; R d )) with 1 / p + 1 / q = 1 1/p+1/q=1 1/ p + 1/ q = 1 , under suitable bounds on div b \operatorname{div}b div b and a mild growth condition. Ambrosio [1 ] later extended this theory to vector fields of bounded variation. Both works also established a qualitative stability result for (1.3): if a sequence of vector fields b n b_n b n converges to b b b in a suitable sense, then the corresponding solutions μ n \mu^n μ n converge to the solution μ \mu μ . This stability, however, is qualitative rather than quantitative. The reason is that the proofs rely on the renormalization method, which is not suited to deriving quantitative stability estimates.
Crippa and De Lellis [9 ] obtained the first quantitative stability estimates for the flow of (1.4) with a bounded Sobolev vector field b ∈ L ∞ ( [ 0 , T ] × R d ; R d ) ∩ L 1 ( [ 0 , T ] ; W 1 , p ( R d ; R d ) ) b\in L^\infty([0,T]\times\mathbb{R}^d;\mathbb{R}^d)\cap L^1([0,T];W^{1,p}(\mathbb{R}^d;\mathbb{R}^d)) b ∈ L ∞ ([ 0 , T ] × R d ; R d ) ∩ L 1 ([ 0 , T ] ; W 1 , p ( R d ; R d )) for some p > 1 p>1 p > 1 . Their key idea was to control a logarithmic distance between two flows. More precisely, given two flows Y 1 Y^1 Y 1 and Y 2 Y^2 Y 2 generated by vector fields b 1 b^1 b 1 and b 2 b^2 b 2 , they introduced functionals of the form
g δ ( t ) : = [ ∫ B r ( log ( 1 + ∣ Y 1 ( t , x ) − Y 2 ( t , x ) ∣ δ ) ) p d x ] 1 / p . g_\delta(t):=\left[\int_{B_r}\left(\log\left(1+\frac{\lvert Y^1(t,x)-Y^2(t,x)\rvert}{\delta}\right)\right)^p\mathrm{d}x\right]^{1/p}. g δ ( t ) := [ ∫ B r ( log ( 1 + δ ∣ Y 1 ( t , x ) − Y 2 ( t , x )∣ ) ) p d x ] 1/ p .
where p > 1 p > 1 p > 1 , r , δ > 0 r, \delta> 0 r , δ > 0 , and B r B_r B r is the open ball in R d \mathbb{R}^d R d centered at the origin with radius r r r . With these functionals, they derived stability estimates of the form
∥ Y 1 ( t , ⋅ ) − Y 2 ( t , ⋅ ) ∥ L p ( B r ) ≤ C t ∣ log ( ∥ b 1 − b 2 ∥ L 1 ( [ 0 , t ] ; L p ( B R ) ) ) ∣ − 1 , \left\|Y^1(t,\cdot)-Y^2(t,\cdot)\right\|_{L^p(B_r)} \le C_t\left|\log\left(\left\|b^1-b^2\right\|_{L^1([0,t];L^p(B_R))}\right)\right|^{-1}, Y 1 ( t , ⋅ ) − Y 2 ( t , ⋅ ) L p ( B r ) ≤ C t log ( b 1 − b 2 L 1 ([ 0 , t ] ; L p ( B R )) ) − 1 ,
where R = r + T ∥ b ∥ L ∞ ( [ 0 , T ] × R d ) R=r+T\|b\|_{L^\infty([0,T]\times\mathbb{R}^d)} R = r + T ∥ b ∥ L ∞ ([ 0 , T ] × R d ) , and C t C_t C t is a positive constant depending not only on t t t but also on r r r , ∥ b 1 ∥ L ∞ ( [ 0 , t ] × R d ) \|b^1\|_{L^\infty([0,t]\times\mathbb{R}^d)} ∥ b 1 ∥ L ∞ ([ 0 , t ] × R d ) , ∥ b 2 ∥ L ∞ ( [ 0 , t ] × R d ) \|b^2\|_{L^\infty([0,t]\times\mathbb{R}^d)} ∥ b 2 ∥ L ∞ ([ 0 , t ] × R d ) and ∥ ∇ x b 1 ∥ L 1 ( [ 0 , t ] ; L p ( R d ) ) \|\nabla_x b^1\|_{L^1([0,t];L^p(\mathbb{R}^d))} ∥ ∇ x b 1 ∥ L 1 ([ 0 , t ] ; L p ( R d )) . Motivated by this result, Seis [32 ] used ideas from optimal transport to establish quantitative stability estimates for solutions of the continuity equation. Beyond the Sobolev setting, the regularity of b b b has been relaxed in several directions [27 , 8 , 4 , 15 , 25 ] ; of particular relevance to us, [6 ] treated vector fields whose gradients are singular integrals (Assumption 4.5). This approach was recently extended by [7 ] to vector fields with partial fractional Sobolev regularity, though under a stronger restriction on the fractional order than in Assumption 3.1.
We now turn to the full equation (2) , in which diffusion is present. For smooth coefficients, the weak well-posedness theory is classical, since linearity and regularity give existence and uniqueness of weak solutions through standard a priori estimates [26 ] . The irregular-coefficient case has also been studied extensively [3 , 20 , 23 , 22 , 30 , 24 , 16 ] . The coefficients considered in this paper, however, fall outside the scope of these results. Weak well-posedness for (2) is therefore established here through the following route. A quantitative stability estimate gives uniqueness, and an approximation argument gives existence.
The stability estimate is the main step. We therefore begin by reviewing the existing quantitative stability results for (2) . The first named author and Luo [21 ] proved such estimates for FPEs with Sobolev regularity. A key step in their analysis was Trevisan’s superposition principle (see Theorem 2.5), which let them recast the stability problem for the FPE as one for the associated SDE. They introduced the Kantorovich–Rubinstein distance
D ~ δ ( μ , ν ) : = inf π ∈ C ( μ , ν ) ∫ R d × R d log ( 1 + ∣ x − y ∣ 2 δ 2 ) π ( d x , d y ) , δ > 0 , μ , ν ∈ P ( R d ) , \widetilde{\mathcal{D}}_\delta(\mu,\nu) := \inf_{\pi\in\mathcal{C}(\mu,\nu)} \int_{\mathbb{R}^d\times\mathbb{R}^d} \log\left(1+\frac{|x-y|^2}{\delta^2}\right)\,\pi(\mathrm{d}x,\mathrm{d}y), \qquad\delta>0,\ \mu,\nu\in\mathcal{P}(\mathbb{R}^d), D δ ( μ , ν ) := π ∈ C ( μ , ν ) inf ∫ R d × R d log ( 1 + δ 2 ∣ x − y ∣ 2 ) π ( d x , d y ) , δ > 0 , μ , ν ∈ P ( R d ) ,
where C ( μ , ν ) \mathcal{C}(\mu,\nu) C ( μ , ν ) denotes the set of all couplings of μ \mu μ and ν \nu ν . Combining this distance with a synchronous coupling of the two associated SDEs, they obtained quantitative stability estimates for weak solutions of the FPE. Using the Zvonkin transformation, they further extended their analysis to singular drifts satisfying the Ladyzhenskaya–Prodi–Serrin condition. In this setting, the diffusion coefficient is the identity matrix, that is, σ = I d × d \sigma=I_{d\times d} σ = I d × d , and the drift coefficient satisfies
b ∈ L q ( [ 0 , T ] ; L p ( R d ; R d ) ) , with p > d , q > 2 such that d p + 2 q < 1. b\in L^q([0,T];L^p(\mathbb{R}^d;\mathbb{R}^d)), \qquad\text{with } p>d,\ q>2\ \text{such that }\frac{d}{p}+\frac{2}{q}<1. b ∈ L q ([ 0 , T ] ; L p ( R d ; R d )) , with p > d , q > 2 such that p d + q 2 < 1.
Two other works [23 , 28 ] studied vector fields whose gradients can be written as a singular integral. Such vector fields were first studied at the level of the ODE (4) by [6 ] , and their analysis was later extended to the FPE by [28 ] and by [23 ] . The former requires the diffusion matrix to be the identity, a = I d × d a=I_{d\times d} a = I d × d , while the latter allows a more general σ = σ ( t , x ) \sigma=\sigma(t,x) σ = σ ( t , x ) with Sobolev regularity, at the cost of an additional boundedness assumption on b b b and σ \sigma σ .
The aim of the present paper is to establish the well-posedness and extend the quantitative stability theory for degenerate FPEs in two directions.
First, we treat anisotropic drift vector fields with partial fractional Sobolev regularity of order s s s . For this class, we prove quantitative stability estimates for weak solutions of degenerate FPEs, extending the theory of [7 ] for Lagrangian flows of the ODE (4) to the diffusive setting. Compared with that work, our treatment introduces two key improvements: we drop the growth assumption on the drift b b b , and we cover the full range s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) , rather than the more restrictive range s ∈ ( 1 / 2 , 1 ) s \in(1/2,1) s ∈ ( 1/2 , 1 ) treated there. As a special case, when the diffusion coefficient vanishes (σ = 0 \sigma= 0 σ = 0 ), the equation reduces to the continuity equation (3) , and our result gives a quantitative stability estimate for it, valid for initial data that are absolutely continuous with respect to the Lebesgue measure.
Second, we study drift vector fields whose distributional derivatives are given by singular integrals. The typical example is the velocity field reconstructed from an L 1 L^1 L 1 vorticity through the Biot–Savart law,
b ( x ) = ( k B S ∗ w ) ( x ) , w ∈ L 1 ( R d ) , b(x) = (k_{\mathrm{BS}} * w)(x), \qquad w \in L^1(\mathbb{R}^d), b ( x ) = ( k BS ∗ w ) ( x ) , w ∈ L 1 ( R d ) ,
where the Biot–Savart kernel k B S k_{\mathrm{BS}} k BS is given by
k B S ( z ) = { z ⊥ 2 π ∣ z ∣ 2 , d = 2 , z ≠ 0 , z s d − 1 ∣ z ∣ d , d ≥ 3 , z ≠ 0 , k_{\mathrm{BS}}(z) =
\begin{cases}
\dfrac{z^\perp}{2\pi|z|^2}, & d = 2,\ z \ne0, \\
\dfrac{z}{s_{d-1}|z|^d}, & d \ge3,\ z \ne0,
\end{cases} k BS ( z ) = ⎩ ⎨ ⎧ 2 π ∣ z ∣ 2 z ⊥ , s d − 1 ∣ z ∣ d z , d = 2 , z = 0 , d ≥ 3 , z = 0 ,
with z ⊥ : = ( − z 2 , z 1 ) z^\perp:= (-z_2,z_1) z ⊥ := ( − z 2 , z 1 ) for z = ( z 1 , z 2 ) ∈ R 2 z = (z_1,z_2) \in\mathbb{R}^2 z = ( z 1 , z 2 ) ∈ R 2 , and s d − 1 s_{d-1} s d − 1 the area of the ( d − 1 ) (d-1) ( d − 1 ) -dimensional unit sphere. This kernel is of independent interest and plays a central role in fluid dynamics, as it recovers the velocity field from the vorticity; through this relation, the incompressible Navier–Stokes equation can be written in vorticity form. Drift fields of this type therefore arise naturally in applications. For this class, we again establish quantitative stability estimates for weak solutions of degenerate FPEs. Our result improves on [28 , 23 ] in two respects. First, we allow a time- and space-dependent diffusion matrix σ = σ ( t , x ) \sigma= \sigma(t,x) σ = σ ( t , x ) that may be degenerate, rather than only the identity matrix required in previous works. Second, we remove the boundedness assumption on the coefficients that was imposed in [23 ] . In addition, as before, our estimate reduces to that for the continuity equation (3) when σ = 0 \sigma= 0 σ = 0 .
In both settings, the stability estimate directly implies uniqueness within the corresponding class of weak solutions. Existence follows from the DiPerna–Lions theory, which was first introduced by DiPerna and Lions in their study of the Boltzmann equation [12 , 11 ] . This theory yields a weak solution, but it does not ensure that the solution is a probability measure. We establish this additional property, under the mild assumption of a finite first moment for the initial datum. The point is irrelevant in the classical PDE theory but essential here, since the superposition principle applies precisely to probability-measure-valued solutions. Through it, a stability estimate for (2) becomes one for the associated SDE (1) . At the level of the SDE, the estimate can be obtained by probabilistic tools such as coupling, which do not require the diffusion matrix to be the identity. This is in contrast to the analytic argument of [28 ] , which relies on that assumption, and it allows us to cover a broader, possibly degenerate class of equations. Our results also complement [23 ] , which established uniqueness of weak solutions but not their existence. Together, these results give the weak well-posedness of the degenerate FPEs considered in this paper.
Structure of the paper This paper is organized in the following way. In Section 2, we collect the necessary preliminaries by recalling basic definitions and presenting some useful results that will be used in the following proofs. Section 3 is devoted to the first class of drift coefficients, which have partial fractional Sobolev regularity. The main result (Theorem 3.3) establishes both existence of a weak solution to the degenerate FPE and a quantitative stability estimate for two such solutions. The proof of the existence part relies on a compactness argument via the DiPerna–Lions theory, while the stability estimate is obtained by coupling the associated SDEs and using anisotropic Kantorovich–Rubinstein distances. At the end of the section, in Subsection 3.3, we discuss the rates of convergence that follow from the stability estimate. In Section 4, we treat the second class of drifts, whose distributional gradients are given by singular integrals. The main theorem (Theorem 4.6) provides a quantitative stability estimate for weak solutions under these rough coefficients. Its proof follows a similar SDE-coupling strategy, but requires a refined maximal function estimate adapted to singular kernels. Finally, in Subsection 4.3, we apply the stability result to prove well-posedness of the FPE under an additional structural assumption on the drift, covering the physically relevant case of velocity fields generated from an L 1 L^1 L 1 vorticity via the Biot–Savart law.
Notation Throughout the paper, d ≥ 2 d \ge2 d ≥ 2 is a fixed positive integer, and R d \mathbb{R}^d R d denotes the d d d -dimensional Euclidean space with inner product ⟨ ⋅ , ⋅ ⟩ \langle\cdot,\cdot\rangle ⟨ ⋅ , ⋅ ⟩ and norm ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ , and L d \mathcal{L}^d L d is the Lebesgue measure on R d \mathbb{R}^d R d . For x ∈ R d x \in\mathbb{R}^d x ∈ R d and R > 0 R > 0 R > 0 , B R ( x ) : = { y ∈ R d : ∣ y − x ∣ < R } B_R(x) := \{y \in\mathbb{R}^d : |y-x| < R\} B R ( x ) := { y ∈ R d : ∣ y − x ∣ < R } , and write B R : = B R ( 0 ) B_R := B_R(0) B R := B R ( 0 ) . We use the following conventions and function spaces.
Basic notation
For a set A A A , 1 A \mathbf{1}_A 1 A denotes its indicator function.
For a Polish space M M M , P ( M ) \mathcal{P}(M) P ( M ) is the set of Borel probability measures on M M M , equipped with the weak topology.
For a matrix Q Q Q , denote by Q ∗ Q^* Q ∗ its transpose, by ∥ Q ∥ \|Q\| ∥ Q ∥ its Hilbert–Schmidt norm, by Q i j Q_{ij} Q ij its entries, and (if Q Q Q is square) by tr Q \operatorname{tr} Q tr Q its trace.
f ≲ g f \lesssim g f ≲ g means f ≤ C g f \le Cg f ≤ C g for some constant C > 0 C > 0 C > 0 depending only on fixed parameters.
f ∗ g f * g f ∗ g is the convolution of f f f and g g g . In particular, for n = 1 , 2 , … n = 1,2,\ldots n = 1 , 2 , … , if f f f is R n \mathbb{R}^n R n -valued and g g g is R \mathbb{R} R -valued, then f ∗ g = g ∗ f = ( f 1 ∗ g , … , f n ∗ g ) f * g = g * f = (f_1 * g,\ldots,f_n * g) f ∗ g = g ∗ f = ( f 1 ∗ g , … , f n ∗ g ) .
For a function f f f , f − : = − min { f , 0 } f^- := -\min\{f,0\} f − := − min { f , 0 } .
For a random variable Z Z Z , L Z \mathcal{L}_Z L Z denotes the law of Z Z Z .
∇ \nabla ∇ , div \operatorname{div} div , ∇ 2 \nabla^2 ∇ 2 and ∂ j \partial_j ∂ j denote the spatial gradient, divergence, Hessian, and j j j -th partial derivative, respectively.
Scalar function spaces on R d \mathbb{R}^d R d
W k , p ( R d ) W^{k,p}(\mathbb{R}^d) W k , p ( R d ) (k = 1 , 2 , ⋯ k=1,2,\cdots k = 1 , 2 , ⋯ , 1 ≤ p ≤ ∞ 1\le p\le\infty 1 ≤ p ≤ ∞ ) denotes the Sobolev space of functions whose weak derivatives up to order k k k lie in L p ( R d ) L^p(\mathbb{R}^d) L p ( R d ) , with norm ∥ ⋅ ∥ W k , p ( R d ) \|\cdot\|_{W^{k,p}(\mathbb{R}^d)} ∥ ⋅ ∥ W k , p ( R d ) . W l o c k , p ( R d ) W^{k,p}_{\mathrm{loc}}(\mathbb{R}^d) W loc k , p ( R d ) denotes the local version.
For 0 < s < 1 0<s<1 0 < s < 1 , 1 ≤ p < ∞ 1\le p<\infty 1 ≤ p < ∞ , the fractional Sobolev space W s , p ( R d ) W^{s,p}(\mathbb{R}^d) W s , p ( R d ) is defined by
W s , p ( R d ) : = { f ∈ L p ( R d ) : [ f ] W s , p ( R d ) < ∞ } , W^{s,p}(\mathbb{R}^d):=\left\{f\in L^p(\mathbb{R}^d):[f]_{W^{s,p}(\mathbb{R}^d)}<\infty\right\}, W s , p ( R d ) := { f ∈ L p ( R d ) : [ f ] W s , p ( R d ) < ∞ } ,
[ f ] W s , p ( R d ) p : = ∫ R d ∫ R d ∣ f ( x ) − f ( y ) ∣ p ∣ x − y ∣ d + s p d x d y , [f]_{W^{s,p}(\mathbb{R}^d)}^p:=\int_{\mathbb{R}^d}\int_{\mathbb{R}^d}\frac{|f(x)-f(y)|^p}{|x-y|^{d+sp}}\,dx\,dy, [ f ] W s , p ( R d ) p := ∫ R d ∫ R d ∣ x − y ∣ d + s p ∣ f ( x ) − f ( y ) ∣ p d x d y ,
with norm ∥ f ∥ W s , p ( R d ) : = ∥ f ∥ L p ( R d ) + [ f ] W s , p ( R d ) \|f\|_{W^{s,p}(\mathbb{R}^d)}:=\|f\|_{L^p(\mathbb{R}^d)}+[f]_{W^{s,p}(\mathbb{R}^d)} ∥ f ∥ W s , p ( R d ) := ∥ f ∥ L p ( R d ) + [ f ] W s , p ( R d ) .
C b ( R d ) C_b(\mathbb{R}^d) C b ( R d ) denotes the class of bounded continuous functions. C b k ( R d ) C_b^k(\mathbb{R}^d) C b k ( R d ) (k = 1 , 2 , ⋯ k=1,2,\cdots k = 1 , 2 , ⋯ ) denotes the class of continuous functions bounded with bounded derivatives up to order k k k . C b ∞ ( R d ) C_b^\infty(\mathbb{R}^d) C b ∞ ( R d ) denotes the class of continuous functions bounded with all derivatives (not a Banach space). C ∞ ( R d ) C^\infty(\mathbb{R}^d) C ∞ ( R d ) denotes the class of smooth functions.
Time-space function spaces Let J ⊂ R J\subset\mathbb{R} J ⊂ R be an interval.
C b 0 , 1 ( J × R d ) C_b^{0,1}(J\times\mathbb{R}^d) C b 0 , 1 ( J × R d ) denotes bounded continuous functions together with their first space derivatives. C b 1 , 2 ( J × R d ) C_b^{1,2}(J\times\mathbb{R}^d) C b 1 , 2 ( J × R d ) denotes bounded continuous functions together with their first time derivative and second space derivatives. C c 1 , 2 ( J × R d ) C_c^{1,2}(J\times\mathbb{R}^d) C c 1 , 2 ( J × R d ) denotes the subspace with compact support in the space variable (and vanishing near the time endpoint if J = [ 0 , T ) J=[0,T) J = [ 0 , T ) ).
(Bochner spaces) For a Banach space ( E , ∥ ⋅ ∥ E ) (E,\|\cdot\|_E) ( E , ∥ ⋅ ∥ E ) , the space L q ( J ; E ) L^q(J;E) L q ( J ; E ) (1 ≤ q ≤ ∞ 1\le q\le\infty 1 ≤ q ≤ ∞ ) consists of strongly measurable f : J → E f:J\to E f : J → E with
∥ f ∥ L q ( J ; E ) : = { ( ∫ J ∥ f ( t ) ∥ E q d t ) 1 / q , 1 ≤ q < ∞ , ess sup t ∈ J ∥ f ( t ) ∥ E , q = ∞ . \|f\|_{L^q(J;E)}:=\begin{cases}\left(\int_J\|f(t)\|_E^q\,dt\right)^{1/q},&1\le q<\infty,\\ \operatorname*{ess\,sup}_{t\in J}\|f(t)\|_E,&q=\infty.\end{cases} ∥ f ∥ L q ( J ; E ) := { ( ∫ J ∥ f ( t ) ∥ E q d t ) 1/ q , ess sup t ∈ J ∥ f ( t ) ∥ E , 1 ≤ q < ∞ , q = ∞.
We use the shorthand (mainly in proofs)
L t q L x p : = L q ( J ; L p ( R d ) ) , L t q W x k , p : = L q ( J ; W k , p ( R d ) ) , L t q W x s , p : = L q ( J ; W s , p ( R d ) ) , L_t^qL_x^p:=L^q(J;L^p(\mathbb{R}^d)),\qquad L_t^qW_x^{k,p}:=L^q(J;W^{k,p}(\mathbb{R}^d)),\qquad L_t^qW_x^{s,p}:=L^q(J;W^{s,p}(\mathbb{R}^d)), L t q L x p := L q ( J ; L p ( R d )) , L t q W x k , p := L q ( J ; W k , p ( R d )) , L t q W x s , p := L q ( J ; W s , p ( R d )) ,
and the corresponding norms
∥ ⋅ ∥ L t q L x p , ∥ ⋅ ∥ L t q W x k , p , ∥ ⋅ ∥ L t q W x s , p . \|\cdot\|_{L_t^qL_x^p},\qquad\|\cdot\|_{L_t^qW_x^{k,p}},\qquad\|\cdot\|_{L_t^qW_x^{s,p}}. ∥ ⋅ ∥ L t q L x p , ∥ ⋅ ∥ L t q W x k , p , ∥ ⋅ ∥ L t q W x s , p .
Analogous notation applies when E E E is replaced by any other space.
(Convention for C b ∞ C_b^\infty C b ∞ in Bochner spaces) The notation L q ( J ; C b ∞ ( R d ) ) L^q(J;C_b^\infty(\mathbb{R}^d)) L q ( J ; C b ∞ ( R d )) is an abuse, meaning that for a.e. t ∈ J t \in J t ∈ J , f ( t , ⋅ ) ∈ C b ∞ ( R d ) f(t,\cdot) \in C_b^\infty(\mathbb{R}^d) f ( t , ⋅ ) ∈ C b ∞ ( R d ) , and the L q L^q L q -norm of its derivatives up to the required finite order (which is always fixed by the context) is integrable on J J J .
Weak Lebesgue / Lorentz–Marcinkiewicz spaces For 1 ≤ r < ∞ 1 \le r < \infty 1 ≤ r < ∞ and a measure space ( X , μ ) (\mathcal{X},\mu) ( X , μ ) , L r , ∞ ( X , μ ) L^{r,\infty}(\mathcal{X},\mu) L r , ∞ ( X , μ ) denotes the space of all measurable functions f : X → R f:\mathcal{X}\to\mathbb{R} f : X → R such that
∥ f ∥ L r , ∞ ( X , μ ) : = sup λ > 0 λ μ ( { ∣ f ∣ > λ } ) 1 / r < ∞ . \|f\|_{L^{r,\infty}(\mathcal{X},\mu)} := \sup_{\lambda>0} \lambda\mu(\{|f|>\lambda\})^{1/r} < \infty. ∥ f ∥ L r , ∞ ( X , μ ) := λ > 0 sup λ μ ({ ∣ f ∣ > λ } ) 1/ r < ∞.
For r = ∞ r=\infty r = ∞ , write L ∞ , ∞ : = L ∞ L^{\infty,\infty}:=L^\infty L ∞ , ∞ := L ∞ . On R d \mathbb{R}^d R d , we write L r , ∞ ( R d ) L^{r,\infty}(\mathbb{R}^d) L r , ∞ ( R d ) , and time-dependent versions are given by the Bochner construction above, e.g., L q ( J ; L r , ∞ ( R d ) ) L^q(J;L^{r,\infty}(\mathbb{R}^d)) L q ( J ; L r , ∞ ( R d )) with
∥ f ∥ L q ( J ; L r , ∞ ( X , μ ) ) : = { ( ∫ J ∥ f ( t , ⋅ ) ∥ L r , ∞ ( X , μ ) q d t ) 1 / q , 1 ≤ q < ∞ , ess sup t ∈ J ∥ f ( t , ⋅ ) ∥ L r , ∞ ( X , μ ) , q = ∞ . \|f\|_{L^q(J;L^{r,\infty}(\mathcal{X},\mu))} :=
\begin{cases}
\left(\int_J \|f(t,\cdot)\|_{L^{r,\infty}(\mathcal{X},\mu)}^q\,\mathrm{d}t\right)^{1/q}, & 1\le q<\infty,\\
\operatorname*{ess\,sup}_{t\in J}\|f(t,\cdot)\|_{L^{r,\infty}(\mathcal{X},\mu)}, & q=\infty.
\end{cases} ∥ f ∥ L q ( J ; L r , ∞ ( X , μ )) := ⎩ ⎨ ⎧ ( ∫ J ∥ f ( t , ⋅ ) ∥ L r , ∞ ( X , μ ) q d t ) 1/ q , ess sup t ∈ J ∥ f ( t , ⋅ ) ∥ L r , ∞ ( X , μ ) , 1 ≤ q < ∞ , q = ∞.
Vector- and matrix-valued analogues All the spaces above have natural componentwise extensions to functions taking values in R m \mathbb{R}^m R m or R m × m \mathbb{R}^{m\times m} R m × m (m = 1 , 2 , … m=1,2,\ldots m = 1 , 2 , … ). Norms are taken as the Euclidean norm (for vectors) or the Hilbert–Schmidt norm (for matrices). We often omit the target space when it is clear from the context (e.g., b b b is vector-valued, σ \sigma σ is matrix-valued).
Preliminaries
We begin by recalling the definitions of weak solution for the FPE and for the SDE, respectively. For f ∈ C c 1 , 2 ( ( 0 , T ) × R d ) f\in C_c^{1,2}((0,T)\times\mathbb{R}^d) f ∈ C c 1 , 2 (( 0 , T ) × R d ) , define the operator L L L by
( L f ) ( t , x ) = 1 2 tr ( ( σ σ ∗ ) ( t , x ) [ ∇ 2 f ( t , ⋅ ) ] ( x ) ) + ⟨ b ( t , x ) , [ ∇ f ( t , ⋅ ) ] ( x ) ⟩ , t ∈ ( 0 , T ) , x ∈ R d . (Lf)(t,x)=\frac{1}{2}\operatorname{tr}\left((\sigma\sigma^*)(t,x)[\nabla^2f(t,\cdot)](x)\right)+\langle b(t,x),[\nabla f(t,\cdot)](x)\rangle,\qquad t\in(0,T),\ x\in\mathbb{R}^d. ( L f ) ( t , x ) = 2 1 tr ( ( σ σ ∗ ) ( t , x ) [ ∇ 2 f ( t , ⋅ )] ( x ) ) + ⟨ b ( t , x ) , [ ∇ f ( t , ⋅ )] ( x )⟩ , t ∈ ( 0 , T ) , x ∈ R d .
Definition 2.1 (Weak solution of the FPE). Let T > 0 T>0 T > 0 and ν ∈ P ( R d ) \nu\in\mathcal{P}(\mathbb{R}^d) ν ∈ P ( R d ) . A Borel curve ( μ t ) t ∈ [ 0 , T ] ⊂ P ( R d ) (\mu_t)_{t\in[0,T]}\subset\mathcal{P}(\mathbb{R}^d) ( μ t ) t ∈ [ 0 , T ] ⊂ P ( R d ) is called a weak solution to (1.2) if the following conditions hold:
(i) (Integrability condition)
∫ 0 T ∫ R d ( ∥ a ( t , x ) ∥ + ∣ b ( t , x ) ∣ ) μ t ( d x ) d t < ∞ ; (5) \int_0^T\int_{\mathbb{R}^d}(\|a(t,x)\|+|b(t,x)|)\,\mu_t(\mathrm{d}x)\,\mathrm{d}t<\infty; \tag*{(5)} ∫ 0 T ∫ R d ( ∥ a ( t , x ) ∥ + ∣ b ( t , x ) ∣ ) μ t ( d x ) d t < ∞ ; (5) (ii) (Weak formulation) For every φ ∈ C c 1 , 2 ( [ 0 , T ) × R d ) \varphi\in C_c^{1,2}([0,T)\times\mathbb{R}^d) φ ∈ C c 1 , 2 ([ 0 , T ) × R d ) ,
∫ R d φ ( 0 , x ) ν ( d x ) + ∫ 0 T ∫ R d ( ∂ t φ ( t , x ) + L φ ( t , x ) ) μ t ( d x ) d t = 0. (6) \int_{\mathbb{R}^d}\varphi(0,x)\,\nu(\mathrm{d}x)+\int_0^T\int_{\mathbb{R}^d}(\partial_t\varphi(t,x)+L\varphi(t,x))\,\mu_t(\mathrm{d}x)\,\mathrm{d}t=0. \tag*{(6)} ∫ R d φ ( 0 , x ) ν ( d x ) + ∫ 0 T ∫ R d ( ∂ t φ ( t , x ) + L φ ( t , x )) μ t ( d x ) d t = 0. (6)
If ( μ t ) t ∈ [ 0 , T ] (\mu_t)_{t\in[0,T]} ( μ t ) t ∈ [ 0 , T ] is absolutely continuous with respect to the Lebesgue measure, we write
μ t ( d x ) = ρ ( t , x ) d x . \mu_t(\mathrm{d}x)=\rho(t,x)\,\mathrm{d}x. μ t ( d x ) = ρ ( t , x ) d x .
In this case we also call ρ \rho ρ a weak solution to the FPE (2) .
Definition 2.3 (Weak solution of the SDE). Let T > 0 T>0 T > 0 and ν ∈ P ( R d ) \nu\in\mathcal{P}(\mathbb{R}^d) ν ∈ P ( R d ) . A pair ( X , W ) (X,W) ( X , W ) is called a weak solution to the SDE (1) with initial distribution ν \nu ν if there exists a complete filtered probability space ( Ω , F , ( F t ) t ∈ [ 0 , T ] , P ) (\Omega,\mathcal{F},(\mathcal{F}_t)_{t\in[0,T]},\mathbb{P}) ( Ω , F , ( F t ) t ∈ [ 0 , T ] , P ) such that W = ( W t ) t ∈ [ 0 , T ] W=(W_t)_{t\in[0,T]} W = ( W t ) t ∈ [ 0 , T ] is a d d d -dimensional standard ( F t ) t ∈ [ 0 , T ] (\mathcal{F}_t)_{t\in[0,T]} ( F t ) t ∈ [ 0 , T ] -Brownian motion, X = ( X t ) t ∈ [ 0 , T ] X=(X_t)_{t\in[0,T]} X = ( X t ) t ∈ [ 0 , T ] is an ( F t ) t ∈ [ 0 , T ] (\mathcal{F}_t)_{t\in[0,T]} ( F t ) t ∈ [ 0 , T ] -adapted continuous process, L X 0 = ν \mathcal{L}_{X_0}=\nu L X 0 = ν , the integrability condition
E ∫ 0 T ( ∣ b ( s , X s ) ∣ + ∥ σ ( s , X s ) ∥ 2 ) d s < ∞ \mathbb{E}\int_0^T \left(|b(s,X_s)|+\|\sigma(s,X_s)\|^2\right)\,\mathrm{d}s<\infty E ∫ 0 T ( ∣ b ( s , X s ) ∣ + ∥ σ ( s , X s ) ∥ 2 ) d s < ∞ holds, and, for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , P \mathbb{P} P -a.s.,
X t = X 0 + ∫ 0 t b ( s , X s ) d s + ∫ 0 t σ ( s , X s ) d W s . X_t=X_0+\int_0^t b(s,X_s)\,\mathrm{d}s+\int_0^t \sigma(s,X_s)\,\mathrm{d}W_s. X t = X 0 + ∫ 0 t b ( s , X s ) d s + ∫ 0 t σ ( s , X s ) d W s . For brevity, we shall often write simply ( X , W ) (X,W) ( X , W ) . When the underlying probability space needs to be specified explicitly, we use the full notation ( Ω , F , ( F t ) t ∈ [ 0 , T ] , P ; X , W ) (\Omega,\mathcal{F},(\mathcal{F}_t)_{t\in[0,T]},\mathbb{P};X,W) ( Ω , F , ( F t ) t ∈ [ 0 , T ] , P ; X , W ) .
The link between these two notions of solution is provided by the martingale solutions of the SDE, which we now define. Let C ( [ 0 , T ] ; R d ) C([0,T];\mathbb{R}^d) C ([ 0 , T ] ; R d ) be the space of continuous paths from [ 0 , T ] [0,T] [ 0 , T ] to R d \mathbb{R}^d R d . For 0 ≤ t ≤ T 0\le t\le T 0 ≤ t ≤ T , let
B t ( C ( [ 0 , T ] ; R d ) ) : = σ ( e s : 0 ≤ s ≤ t ) \mathcal{B}_t(C([0,T];\mathbb{R}^d)):=\sigma(e_s:0\le s\le t) B t ( C ([ 0 , T ] ; R d )) := σ ( e s : 0 ≤ s ≤ t )
be the canonical σ \sigma σ -algebra, where e s : C ( [ 0 , T ] ; R d ) → R d e_s:C([0,T];\mathbb{R}^d)\to\mathbb{R}^d e s : C ([ 0 , T ] ; R d ) → R d is the evaluation map at time s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] .
Definition 2.4 (Martingale solution of the SDE). Given ν ∈ P ( R d ) \nu\in\mathcal{P}(\mathbb{R}^d) ν ∈ P ( R d ) , a probability measure Q ∈ P ( C ( [ 0 , T ] ; R d ) ) Q\in\mathcal{P}(C([0,T];\mathbb{R}^d)) Q ∈ P ( C ([ 0 , T ] ; R d )) is called a martingale solution to the SDE (1) with initial distribution ν \nu ν if
Q ∘ e 0 − 1 = ν , Q\circ e_0^{-1}=\nu, Q ∘ e 0 − 1 = ν ,
∫ 0 T ∫ R d ( ∣ b ( t , x ) ∣ + ∥ a ( t , x ) ∥ ) ( Q ∘ e t − 1 ) ( d x ) d t < ∞ , \int_0^T \int_{\mathbb{R}^d} (|b(t,x)|+\|a(t,x)\|)(Q\circ e_t^{-1})(dx)\,dt<\infty, ∫ 0 T ∫ R d ( ∣ b ( t , x ) ∣ + ∥ a ( t , x ) ∥ ) ( Q ∘ e t − 1 ) ( d x ) d t < ∞ ,
and, for every φ ∈ C b 1 , 2 ( [ 0 , T ] × R d ) \varphi\in C_b^{1,2}([0,T]\times\mathbb{R}^d) φ ∈ C b 1 , 2 ([ 0 , T ] × R d ) , the process ( M t φ ) t ∈ [ 0 , T ] (M_t^\varphi)_{t\in[0,T]} ( M t φ ) t ∈ [ 0 , T ] defined by
M t φ ( ω ) : = φ ( t , ω t ) − φ ( 0 , ω 0 ) − ∫ 0 t ( ∂ s φ + L φ ) ( s , ω s ) d s , ω ∈ C ( [ 0 , T ] ; R d ) , M_t^\varphi(\omega):=\varphi(t,\omega_t)-\varphi(0,\omega_0)-\int_0^t(\partial_s\varphi+L\varphi)(s,\omega_s)\,ds,\qquad\omega\in C([0,T];\mathbb{R}^d), M t φ ( ω ) := φ ( t , ω t ) − φ ( 0 , ω 0 ) − ∫ 0 t ( ∂ s φ + L φ ) ( s , ω s ) d s , ω ∈ C ([ 0 , T ] ; R d ) ,
is a B t ( C ( [ 0 , T ] ; R d ) ) \mathcal{B}_t(C([0,T];\mathbb{R}^d)) B t ( C ([ 0 , T ] ; R d )) -martingale under Q Q Q .
It is well known that martingale solutions and weak solutions of the SDE are equivalent in the following sense. If Q ∈ P ( C ( [ 0 , T ] ; R d ) ) Q\in\mathcal{P}(C([0,T];\mathbb{R}^d)) Q ∈ P ( C ([ 0 , T ] ; R d )) is a martingale solution to (1.1), then there exists a weak solution ( Ω , F , ( F t ) t ∈ [ 0 , T ] , P ; X , W ) (\Omega,\mathcal{F},(\mathcal{F}_t)_{t\in[0,T]},\mathbb{P};X,W) ( Ω , F , ( F t ) t ∈ [ 0 , T ] , P ; X , W ) to (1.1) with P ∘ X − 1 = Q \mathbb{P}\circ X^{-1}=Q P ∘ X − 1 = Q . Conversely, if ( X , W ) (X,W) ( X , W ) is a weak solution to (1.1), then its path law P ∘ X − 1 \mathbb{P}\circ X^{-1} P ∘ X − 1 is a martingale solution. For a comprehensive treatment, we refer to the monographs [31 , 18 ] .
On the one hand, for any φ ∈ C c 1 , 2 ( [ 0 , T ) × R d ) \varphi\in C_c^{1,2}([0,T)\times\mathbb{R}^d) φ ∈ C c 1 , 2 ([ 0 , T ) × R d ) , applying Itô’s formula to φ ( t , X t ) \varphi(t,X_t) φ ( t , X t ) gives (2.2). Hence the time marginals of any weak solution to the SDE form a weak solution to the corresponding FPE (1.2). On the other hand, the converse direction follows from the following superposition principle.
Theorem 2.5 (Superposition principle, [33 Theorem 2.5] ). Let ( μ t ) t ∈ [ 0 , T ] ⊂ P ( R d ) (\mu_t)_{t\in[0,T]}\subset\mathcal{P}(\mathbb{R}^d) ( μ t ) t ∈ [ 0 , T ] ⊂ P ( R d ) be a weakly continuous weak solution to the FPE (1.2). Then there exists a martingale solution Q ∈ P ( C ( [ 0 , T ] ; R d ) ) Q\in\mathcal{P}(C([0,T];\mathbb{R}^d)) Q ∈ P ( C ([ 0 , T ] ; R d )) to the SDE (1.1) such that
Q ∘ e t − 1 = μ t , t ∈ [ 0 , T ] . Q\circ e_t^{-1}=\mu_t,\qquad t\in[0,T]. Q ∘ e t − 1 = μ t , t ∈ [ 0 , T ] . Equivalently, there exists a weak solution ( X , W ) (X,W) ( X , W ) to the SDE satisfying
L X t = μ t , t ∈ [ 0 , T ] . \mathcal{L}_{X_t}=\mu_t,\qquad t\in[0,T]. L X t = μ t , t ∈ [ 0 , T ] .
Our stability estimates are formulated in terms of expectations of the weak solutions to the associated SDEs. These weak solutions may, however, be defined on different probability spaces, and hence cannot be compared directly. To address this issue, we use the following lemma (cf. [21 Proposition 3.6] ). It constructs two processes on a common filtered probability space, driven by the same Brownian motion, each having the same law as one of the original weak solutions. This coupling serves as the starting point of our stability analysis.
Lemma 2.6. Let ( Ω , F , ( F t ) t ≥ 0 , P ; X , W ) (\Omega,\mathcal{F},(\mathcal{F}_t)_{t\geq0},\mathbb{P};X,W) ( Ω , F , ( F t ) t ≥ 0 , P ; X , W ) and ( Ω ^ , F ^ , ( F ^ t ) t ≥ 0 , P ^ ; X ^ , W ^ ) (\widehat{\Omega},\widehat{\mathcal{F}},(\widehat{\mathcal{F}}_t)_{t\geq0},\widehat{\mathbb{P}};\widehat{X},\widehat{W}) ( Ω , F , ( F t ) t ≥ 0 , P ; X , W ) be weak solutions to the SDE (1.1) corresponding to the coefficients ( b , σ ) (b,\sigma) ( b , σ ) and ( b ^ , σ ^ ) (\widehat{b},\widehat{\sigma}) ( b , σ ) , with initial distributions ν \nu ν and ν ^ \widehat{\nu} ν , respectively. Then, for any π 0 ∈ C ( ν , ν ^ ) \pi_0\in\mathcal{C}(\nu,\widehat{\nu}) π 0 ∈ C ( ν , ν ) , there exist a filtered probability space ( Ω ~ , F ~ , ( F ~ t ) t ≥ 0 , P ~ ) (\widetilde{\Omega},\widetilde{\mathcal{F}},(\widetilde{\mathcal{F}}_t)_{t\geq0},\widetilde{\mathbb{P}}) ( Ω , F , ( F t ) t ≥ 0 , P ) , a standard d d d -dimensional Brownian motion W ~ = ( W ~ t ) t ≥ 0 \widetilde{W}=(\widetilde{W}_t)_{t\geq0} W = ( W t ) t ≥ 0 adapted to ( F ~ t ) t ≥ 0 (\widetilde{\mathcal{F}}_t)_{t\geq0} ( F t ) t ≥ 0 , and two ( F ~ t ) t ≥ 0 (\widetilde{\mathcal{F}}_t)_{t\geq0} ( F t ) t ≥ 0 -adapted continuous processes Y Y Y and Y ^ \widehat{Y} Y such that L ( Y 0 , Y ^ 0 ) = π 0 \mathcal{L}_{(Y_0,\widehat{Y}_0)}=\pi_0 L ( Y 0 , Y 0 ) = π 0 and the following properties hold:
(i) The pairs ( X , W ) (X,W) ( X , W ) and ( Y , W ~ ) (Y,\widetilde{W}) ( Y , W ) have the same law, and so do the pairs ( X ^ , W ^ ) (\widehat{X},\widehat{W}) ( X , W ) and ( Y ^ , W ~ ) (\widehat{Y},\widetilde{W}) ( Y , W ) .
(ii) The systems ( Ω ~ , F ~ , ( F ~ t ) t ≥ 0 , P ~ ; Y , W ~ ) (\widetilde{\Omega},\widetilde{\mathcal{F}},(\widetilde{\mathcal{F}}_t)_{t\geq0},\widetilde{\mathbb{P}};Y,\widetilde{W}) ( Ω , F , ( F t ) t ≥ 0 , P ; Y , W ) and ( Ω ~ , F ~ , ( F ~ t ) t ≥ 0 , P ~ ; Y ^ , W ^ ~ ) (\widetilde{\Omega},\widetilde{\mathcal{F}},(\widetilde{\mathcal{F}}_t)_{t\geq0},\widetilde{\mathbb{P}};\widehat{Y},\widetilde{\widehat{W}}) ( Ω , F , ( F t ) t ≥ 0 , P ; Y , W ) are weak solutions to the SDE (1.1) with coefficients ( b , σ ) (b,\sigma) ( b , σ ) and ( b ^ , σ ^ ) (\hat{b},\hat{\sigma}) ( b ^ , σ ^ ) , and with initial distributions ν \nu ν and ν ^ \hat{\nu} ν ^ , respectively.
In view of this lemma, we may assume without loss of generality that the weak solutions under consideration are defined on a common filtered probability space and are driven by the same Brownian motion.
Finally, let ( X , μ ) (\mathcal{X},\mu) ( X , μ ) be a measure space. For r ∈ [ 1 , ∞ ] r\in[1,\infty] r ∈ [ 1 , ∞ ] , we denote by L r ( X , μ ) L^r(\mathcal{X},\mu) L r ( X , μ ) the usual Lebesgue space equipped with the norm ∥ ⋅ ∥ L r ( X , μ ) \|\cdot\|_{L^r(\mathcal{X},\mu)} ∥ ⋅ ∥ L r ( X , μ ) , and by L r , ∞ ( X , μ ) L^{r,\infty}(\mathcal{X},\mu) L r , ∞ ( X , μ ) the weak Lebesgue / Lorentz–Marcinkiewicz space endowed with the quasi-norm ∥ ⋅ ∥ L r , ∞ ( X , μ ) \|\cdot\|_{L^{r,\infty}(\mathcal{X},\mu)} ∥ ⋅ ∥ L r , ∞ ( X , μ ) ; precise definitions are given in Notation. Recall the continuous embedding L r ( X , μ ) ⊂ L r , ∞ ( X , μ ) L^r(\mathcal{X},\mu)\subset L^{r,\infty}(\mathcal{X},\mu) L r ( X , μ ) ⊂ L r , ∞ ( X , μ ) , which satisfies
∥ f ∥ L r , ∞ ( X , μ ) ≤ ∥ f ∥ L r ( X , μ ) , f ∈ L r ( X , μ ) . \|f\|_{L^{r,\infty}(\mathcal{X},\mu)}\leq\|f\|_{L^r(\mathcal{X},\mu)},\qquad f\in L^r(\mathcal{X},\mu). ∥ f ∥ L r , ∞ ( X , μ ) ≤ ∥ f ∥ L r ( X , μ ) , f ∈ L r ( X , μ ) .
We shall also make use of the following interpolation inequalities; for their proofs, we refer to [8 , 6 ] .
Lemma 2.7 (Interpolation). Let E ⊂ X E\subset\mathcal{X} E ⊂ X be a measurable set with μ ( E ) < ∞ \mu(E)<\infty μ ( E ) < ∞ , and let f : X → R f:\mathcal{X}\to\mathbb{R} f : X → R be a measurable function. Then the following estimates hold.
(1) For 1 ≤ r < ℓ ≤ ∞ 1\leq r<\ell\leq\infty 1 ≤ r < ℓ ≤ ∞ ,
∥ f ∥ L r ( E , μ ) ≤ ℓ ℓ − r μ ( E ) 1 − r ℓ ∥ f ∥ L ℓ , ∞ ( E , μ ) . (7) \|f\|_{L^r(E,\mu)}\leq\frac{\ell}{\ell-r}\mu(E)^{1-\frac{r}{\ell}}\|f\|_{L^{\ell,\infty}(E,\mu)}. \tag*{(7)} ∥ f ∥ L r ( E , μ ) ≤ ℓ − r ℓ μ ( E ) 1 − ℓ r ∥ f ∥ L ℓ , ∞ ( E , μ ) . (7) (2) For 1 < ℓ ≤ ∞ 1<\ell\leq\infty 1 < ℓ ≤ ∞ ,
∥ f ∥ L 1 ( E , μ ) ≤ ℓ ℓ − 1 ∥ f ∥ L 1 , ∞ ( E , μ ) [ 1 + log ( μ ( E ) 1 − 1 ℓ ∥ f ∥ L ℓ , ∞ ( E , μ ) ∥ f ∥ L 1 , ∞ ( E , μ ) ) ] . (8) \|f\|_{L^1(E,\mu)}\leq\frac{\ell}{\ell-1}\|f\|_{L^{1,\infty}(E,\mu)}
\left[1+\log\left(\frac{\mu(E)^{1-\frac{1}{\ell}}\|f\|_{L^{\ell,\infty}(E,\mu)}}{\|f\|_{L^{1,\infty}(E,\mu)}}\right)\right]. \tag*{(8)} ∥ f ∥ L 1 ( E , μ ) ≤ ℓ − 1 ℓ ∥ f ∥ L 1 , ∞ ( E , μ ) [ 1 + log ( ∥ f ∥ L 1 , ∞ ( E , μ ) μ ( E ) 1 − ℓ 1 ∥ f ∥ L ℓ , ∞ ( E , μ ) ) ] . (8) Here, for ℓ = ∞ \ell=\infty ℓ = ∞ , the factors ℓ ℓ − r \frac{\ell}{\ell-r} ℓ − r ℓ in (2.3) and ℓ ℓ − 1 \frac{\ell}{\ell-1} ℓ − 1 ℓ in (2.4) are interpreted as their limiting values, both equal to 1, and if ∥ f ∥ L 1 , ∞ ( E , μ ) = 0 \|f\|_{L^{1,\infty}(E,\mu)}=0 ∥ f ∥ L 1 , ∞ ( E , μ ) = 0 , the right-hand side of (2.4) is understood as 0 by continuity.
Further details on weak Lebesgue spaces can be found in the monograph [17 ] .
Partially regular drift coefficients
Main results
In this subsection, we present our main results for FPEs whose drift coefficients have partial fractional Sobolev regularity. To this end, we split the state space as
R d = R d 1 × R d 2 , d = d 1 + d 2 , \mathbb{R}^d=\mathbb{R}^{d_1}\times\mathbb{R}^{d_2},\qquad d=d_1+d_2, R d = R d 1 × R d 2 , d = d 1 + d 2 ,
where d 1 , d 2 d_1,d_2 d 1 , d 2 are positive integers. Consider drift coefficients b : [ 0 , T ] × R d 1 × R d 2 → R d 1 × R d 2 b : [0,T]\times\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}\to\mathbb{R}^{d_1}\times\mathbb{R}^{d_2} b : [ 0 , T ] × R d 1 × R d 2 → R d 1 × R d 2 of the form
b ( t , x 1 , x 2 ) = ( b 1 ( t , x 1 ) , b 2 ( t , x 1 , x 2 ) ) . b(t,x_1,x_2)=\left(b_1(t,x_1),b_2(t,x_1,x_2)\right). b ( t , x 1 , x 2 ) = ( b 1 ( t , x 1 ) , b 2 ( t , x 1 , x 2 ) ) .
and assume that the diffusion coefficient σ : [ 0 , T ] × R d 1 × R d 2 → R d × d \sigma: [0,T]\times\mathbb{R}^{d_1}\times\mathbb{R}^{d_2}\to\mathbb{R}^{d\times d} σ : [ 0 , T ] × R d 1 × R d 2 → R d × d has the block-diagonal form
σ ( t , x 1 , x 2 ) = ( σ 1 ( t , x 1 ) 0 0 σ 2 ( t , x 1 , x 2 ) ) . \sigma(t,x_1,x_2)=
\begin{pmatrix}
\sigma_1(t,x_1) & 0\\
0 & \sigma_2(t,x_1,x_2)
\end{pmatrix}. σ ( t , x 1 , x 2 ) = ( σ 1 ( t , x 1 ) 0 0 σ 2 ( t , x 1 , x 2 ) ) .
where t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] and ( x 1 , x 2 ) ∈ R d 1 × R d 2 (x_1,x_2)\in\mathbb{R}^{d_1}\times\mathbb{R}^{d_2} ( x 1 , x 2 ) ∈ R d 1 × R d 2 . Note that b 1 b_1 b 1 and σ 1 \sigma_1 σ 1 do not depend on the second variable x 2 x_2 x 2 .
Under this structural assumption, we impose the following conditions on the coefficients. For convenience, we define the modified drift
b σ : = b − 1 2 div ( σ σ ∗ ) . b^\sigma:=b-\frac{1}{2}\operatorname{div}(\sigma\sigma^*). b σ := b − 2 1 div ( σ σ ∗ ) .
Assumption 3.1. Let p , r ∈ ( 1 , ∞ ] p,r\in(1,\infty] p , r ∈ ( 1 , ∞ ] , s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) , and ν ∈ P ( R d ) \nu\in\mathcal{P}(\mathbb{R}^d) ν ∈ P ( R d ) .
( H σ ) (\mathcal{H}^\sigma) ( H σ ) σ ∈ L 2 ( [ 0 , T ] ; L ∞ ( R d ; R d × d ) ) \sigma\in L^2([0,T];L^\infty(\mathbb{R}^d;\mathbb{R}^{d\times d})) σ ∈ L 2 ([ 0 , T ] ; L ∞ ( R d ; R d × d )) with
( ∑ i , j , k = 1 d 1 ∣ ∂ k ( σ 1 ) i j ∣ 2 ) 1 / 2 ∈ L 2 ( [ 0 , T ] ; L 2 r ( R d 1 ) ) , \left(\sum_{i,j,k=1}^{d_1}\left|\partial_k(\sigma_1)_{ij}\right|^2\right)^{1/2}\in L^2([0,T];L^{2r}(\mathbb{R}^{d_1})), i , j , k = 1 ∑ d 1 ∣ ∂ k ( σ 1 ) ij ∣ 2 1/2 ∈ L 2 ([ 0 , T ] ; L 2 r ( R d 1 )) , ( ∑ k = 1 d ∑ i , j = 1 d 2 ∣ ∂ k ( σ 2 ) i j ∣ 2 ) 1 / 2 ∈ L 2 ( [ 0 , T ] ; L 2 r ( R d ) ) , \left(\sum_{k=1}^{d}\sum_{i,j=1}^{d_2}\left|\partial_k(\sigma_2)_{ij}\right|^2\right)^{1/2}\in L^2([0,T];L^{2r}(\mathbb{R}^d)), ( k = 1 ∑ d i , j = 1 ∑ d 2 ∣ ∂ k ( σ 2 ) ij ∣ 2 ) 1/2 ∈ L 2 ([ 0 , T ] ; L 2 r ( R d )) , where the derivative is understood in the weak sense.
( H b ) (\mathcal{H}^b) ( H b ) b ∈ L l o c p ( [ 0 , T ] × R d ; R d ) b\in L^p_{\mathrm{loc}}([0,T]\times\mathbb{R}^d;\mathbb{R}^d) b ∈ L loc p ([ 0 , T ] × R d ; R d ) with
b 1 ∈ L 1 ( [ 0 , T ] ; W 1 , p ( R d 1 ; R d 1 ) ) ∩ L 1 ( [ 0 , T ] ; L ∞ ( R d 1 ; R d 1 ) ) , b_1\in L^1([0,T];W^{1,p}(\mathbb{R}^{d_1};\mathbb{R}^{d_1}))\cap L^1([0,T];L^\infty(\mathbb{R}^{d_1};\mathbb{R}^{d_1})), b 1 ∈ L 1 ([ 0 , T ] ; W 1 , p ( R d 1 ; R d 1 )) ∩ L 1 ([ 0 , T ] ; L ∞ ( R d 1 ; R d 1 )) , b 2 ∈ L 1 ( [ 0 , T ] × R d 2 ; W s , 1 ( R d 1 ; R d 2 ) ) ∩ L 1 ( [ 0 , T ] × R d 1 ; W 1 , p ( R d 2 ; R d 2 ) ) . b_2\in L^1([0,T]\times\mathbb{R}^{d_2};W^{s,1}(\mathbb{R}^{d_1};\mathbb{R}^{d_2}))\cap L^1([0,T]\times\mathbb{R}^{d_1};W^{1,p}(\mathbb{R}^{d_2};\mathbb{R}^{d_2})). b 2 ∈ L 1 ([ 0 , T ] × R d 2 ; W s , 1 ( R d 1 ; R d 2 )) ∩ L 1 ([ 0 , T ] × R d 1 ; W 1 , p ( R d 2 ; R d 2 )) . ( H b σ ) (\mathcal{H}^{b^\sigma}) ( H b σ ) ( div b σ ) − ∈ L 1 ( [ 0 , T ] ; L ∞ ( R d ) ) (\operatorname{div}b^\sigma)^-\in L^1([0,T];L^\infty(\mathbb{R}^d)) ( div b σ ) − ∈ L 1 ([ 0 , T ] ; L ∞ ( R d )) .
( H ν ) (\mathcal{H}^\nu) ( H ν ) The initial distribution ν \nu ν admits a probability density ρ 0 ∈ L 1 ( R d ) ∩ L ∞ ( R d ) \rho_0\in L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d) ρ 0 ∈ L 1 ( R d ) ∩ L ∞ ( R d ) with
∫ R d ∣ x ∣ ρ 0 ( x ) d x < ∞ . \int_{\mathbb{R}^d}|x|\rho_0(x)\,dx<\infty. ∫ R d ∣ x ∣ ρ 0 ( x ) d x < ∞.
Remark 3.2. (i) The condition ( div b ) − ∈ L 1 ( [ 0 , T ] ; L ∞ ( R d ) ) (\operatorname{div}b)^-\in L^1([0,T];L^\infty(\mathbb{R}^d)) ( div b ) − ∈ L 1 ([ 0 , T ] ; L ∞ ( R d )) is standard in the study of transport equations; see, for instance, [1 , 10 ] . It has a natural interpretation in terms of the flow generated by b b b : the divergence quantifies local compression and expansion. The most familiar case is div b = 0 \operatorname{div}b=0 div b = 0 , which corresponds to an incompressible flow; the condition on ( div b ) − (\operatorname{div}b)^{-} ( div b ) − is weaker, allowing compression but keeping it controlled. In the FPE setting, as shown by Le Bris and Lions [24 ] , the same idea leads naturally to condition ( H b σ ) (H^{b^\sigma}) ( H b σ ) . This assumption provides the fundamental a priori estimate for densities, which is essential for constructing weak solutions when the drift has low regularity.
(ii) The conditions ∫ R d ∣ x ∣ ρ 0 ( x ) d x < ∞ \int_{\mathbb{R}^{d}} |x|\rho_{0}(x)\,\mathrm{d}x<\infty ∫ R d ∣ x ∣ ρ 0 ( x ) d x < ∞ and (2.1) together imply that the solution to the FPE has a finite first moment, namely
ess sup t ∈ [ 0 , T ] ∫ R d ∣ x ∣ μ t ( d x ) < ∞ . \underset{t\in[0,T]}{\operatorname{ess\,sup}}\int_{\mathbb{R}^{d}} |x|\mu_t(\mathrm{d}x)<\infty. t ∈ [ 0 , T ] ess sup ∫ R d ∣ x ∣ μ t ( d x ) < ∞.
Indeed, by the superposition principle, let ( X , W ) (X,W) ( X , W ) be a weak solution to the corresponding SDE with L X t = μ t \mathcal{L}_{X_t}=\mu_t L X t = μ t for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . Condition (2.1) guarantees that ∫ 0 t σ ( s , X s ) d W s \int_0^t \sigma(s,X_s)\,\mathrm{d}W_s ∫ 0 t σ ( s , X s ) d W s is a square-integrable martingale; see, for instance, [18 ] . By Jensen’s inequality and Itô’s isometry, we obtain, for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
∫ R d ∣ x ∣ μ t ( d x ) = E ∣ X t ∣ ≤ E ∣ X 0 ∣ + E ∣ ∫ 0 t σ ( s , X s ) d W s ∣ + E ∣ ∫ 0 t b ( s , X s ) d s ∣ ≤ ∫ R d ∣ x ∣ ρ 0 ( x ) d x + ( ∫ 0 T ∫ R d ∥ σ ( s , x ) ∥ 2 μ s ( d x ) d s ) 1 / 2 + ∫ 0 T ∫ R d ∣ b ( s , x ) ∣ μ s ( d x ) d s < ∞ . \begin{aligned}
\int_{\mathbb{R}^{d}} |x|\mu_t(\mathrm{d}x)
&=\mathbb{E}|X_t|\\
&\leq\mathbb{E}|X_0|+\mathbb{E}\left|\int_0^t \sigma(s,X_s)\,\mathrm{d}W_s\right|+\mathbb{E}\left|\int_0^t b(s,X_s)\,\mathrm{d}s\right|\\
&\leq\int_{\mathbb{R}^{d}} |x|\rho_0(x)\,\mathrm{d}x+
\left(\int_0^T\int_{\mathbb{R}^{d}}\|\sigma(s,x)\|^2\mu_s(\mathrm{d}x)\,\mathrm{d}s\right)^{1/2}
+\int_0^T\int_{\mathbb{R}^{d}}|b(s,x)|\mu_s(\mathrm{d}x)\,\mathrm{d}s\\
&<\infty.
\end{aligned} ∫ R d ∣ x ∣ μ t ( d x ) = E ∣ X t ∣ ≤ E ∣ X 0 ∣ + E ∫ 0 t σ ( s , X s ) d W s + E ∫ 0 t b ( s , X s ) d s ≤ ∫ R d ∣ x ∣ ρ 0 ( x ) d x + ( ∫ 0 T ∫ R d ∥ σ ( s , x ) ∥ 2 μ s ( d x ) d s ) 1/2 + ∫ 0 T ∫ R d ∣ b ( s , x ) ∣ μ s ( d x ) d s < ∞.
Taking the essential supremum in time gives the desired estimate.
(iii) Under assumption ( H σ ) (H^\sigma) ( H σ ) , one readily verifies that for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
σ 1 ( t , ⋅ ) ∈ W l o c 1 , 1 ( R d 1 ; R d 1 × d 1 ) , σ 2 ( t , ⋅ ) ∈ W l o c 1 , 1 ( R d ; R d 2 × d 2 ) . \sigma_1(t,\cdot)\in W_{\mathrm{loc}}^{1,1}(\mathbb{R}^{d_1};\mathbb{R}^{d_1\times d_1}),\qquad
\sigma_2(t,\cdot)\in W_{\mathrm{loc}}^{1,1}(\mathbb{R}^{d};\mathbb{R}^{d_2\times d_2}). σ 1 ( t , ⋅ ) ∈ W loc 1 , 1 ( R d 1 ; R d 1 × d 1 ) , σ 2 ( t , ⋅ ) ∈ W loc 1 , 1 ( R d ; R d 2 × d 2 ) .
This regularity makes Lemma 3.10 applicable.
Corresponding to the structure of the coefficients b , σ b,\sigma b , σ , for δ 1 , δ 2 > 0 \delta_1,\delta_2>0 δ 1 , δ 2 > 0 , we introduce the d × d d\times d d × d diagonal matrix
B = diag ( δ 1 , … , δ 1 ⏟ d 1 t i m e s , δ 2 , … , δ 2 ⏟ d 2 t i m e s ) . (9) B=\operatorname{diag}(\underbrace{\delta_1,\ldots,\delta_1}_{d_1\ \mathrm{times}},\underbrace{\delta_2,\ldots,\delta_2}_{d_2\ \mathrm{times}}). \tag*{(9)} B = diag ( d 1 times δ 1 , … , δ 1 , d 2 times δ 2 , … , δ 2 ) . (9)
which defines an anisotropic Kantorovich–Rubinstein distance
D δ 1 , δ 2 ( μ , ν ) : = inf π ∈ C ( μ , ν ) ∫ R d × R d log ( 1 + ∣ B − 1 ( x − y ) ∣ ) π ( d x , d y ) , μ , ν ∈ P ( R d ) , D_{\delta_1,\delta_2}(\mu,\nu):=\inf_{\pi\in\mathcal{C}(\mu,\nu)}
\int_{\mathbb{R}^{d}\times\mathbb{R}^{d}}\log\left(1+\left|B^{-1}(x-y)\right|\right)\,\pi(\mathrm{d}x,\mathrm{d}y),\qquad\mu,\nu\in\mathcal{P}(\mathbb{R}^{d}), D δ 1 , δ 2 ( μ , ν ) := π ∈ C ( μ , ν ) inf ∫ R d × R d log ( 1 + B − 1 ( x − y ) ) π ( d x , d y ) , μ , ν ∈ P ( R d ) ,
where B − 1 B^{-1} B − 1 denotes the inverse matrix of B B B . When the measures admit densities ρ 1 , ρ 2 \rho_1,\rho_2 ρ 1 , ρ 2 , we also write D δ 1 , δ 2 ( ρ 1 , ρ 2 ) \mathcal{D}_{\delta_1,\delta_2}(\rho_1,\rho_2) D δ 1 , δ 2 ( ρ 1 , ρ 2 ) for the distance between the corresponding distributions.
We now present the main theorem of this section. It has two components: first, an existence result for weak solutions that are probability densities, a property crucial for invoking the superposition principle and thereby linking the FPE to the associated SDE; second, a quantitative stability estimate for any two such solutions. Owing to the local integrability of the coefficients b 1 b_1 b 1 and σ 1 \sigma_1 σ 1 , the latter is formulated in terms of expectations of the corresponding SDE solutions, with trajectories stopped upon exiting a ball in the state space.
Theorem 3.3. The following statements hold.
(Existence) Under Assumption 3.1, the FPE (2) admits a weak solution that is absolutely continuous with respect to the Lebesgue measure, with a probability density ρ ∈ L ∞ ( [ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d ) ) \rho\in L^\infty([0,T];L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d)) ρ ∈ L ∞ ([ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d )) .
(Quantitative stability) Let ( b , σ , ρ 0 ) (b,\sigma,\rho_0) ( b , σ , ρ 0 ) and ( b ^ , σ ^ , ρ ^ 0 ) (\hat b,\hat\sigma,\hat\rho_0) ( b ^ , σ ^ , ρ ^ 0 ) be two triples satisfying Assumption 3.1. Denote by ρ , ρ ^ ∈ L ∞ ( [ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d ) ) \rho,\hat\rho\in L^\infty([0,T];L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d)) ρ , ρ ^ ∈ L ∞ ([ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d )) the corresponding weak solutions to the FPE (2) , and by ( X , W ) (X,W) ( X , W ) and ( X ^ , W ^ ) (\hat X,\hat W) ( X ^ , W ^ ) the associated weak solutions to the SDE (1) . For R > 0 R>0 R > 0 , define the stopping time
τ R : = inf { t ∈ [ 0 , T ] : ∣ X t ∣ ∨ ∣ X ^ t ∣ ≥ R } . \tau_R:=\inf\{t\in[0,T]: |X_t|\vee|\hat X_t|\ge R\}. τ R := inf { t ∈ [ 0 , T ] : ∣ X t ∣ ∨ ∣ X ^ t ∣ ≥ R } . Then, for every δ 2 ∈ ( 0 , 1 ) \delta_2\in(0,1) δ 2 ∈ ( 0 , 1 ) , one can choose δ 1 ∈ ( 0 , δ 2 ) \delta_1\in(0,\delta_2) δ 1 ∈ ( 0 , δ 2 ) such that, for all R > 0 R>0 R > 0 , the following stability estimate holds:
sup t ∈ [ 0 , T ] E { log ( 1 + ∣ B − 1 ( X t ∧ τ R − X ^ t ∧ τ R ) ∣ ) } ≲ D δ 1 , δ 2 ( ρ 0 , ρ ^ 0 ) + 1 δ 1 ∥ b 1 − b ^ 1 ∥ L 1 ( [ 0 , T ] ; L p ( B R ) ) + 1 δ 1 ∥ b 2 − b ^ 2 ∥ L 1 ( [ 0 , T ] ; L 1 ( R d ) ) + 1 δ 1 2 ∥ σ − σ ^ ∥ L 2 ( [ 0 , T ] ; L 2 r ( B R ) ) + 1. \sup_{t\in[0,T]}\mathbb{E}\left\{\log\left(1+\left|B^{-1}(X_{t\wedge\tau_R}-\hat X_{t\wedge\tau_R})\right|\right)\right\}
\lesssim\mathcal{D}_{\delta_1,\delta_2}(\rho_0,\hat\rho_0)+\frac{1}{\delta_1}\|b_1-\hat b_1\|_{L^1([0,T];L^p(B_R))}
+\frac{1}{\delta_1}\|b_2-\hat b_2\|_{L^1([0,T];L^1(\mathbb{R}^d))}
+\frac{1}{\delta_1^2}\|\sigma-\hat\sigma\|_{L^2([0,T];L^{2r}(B_R))}+1. t ∈ [ 0 , T ] sup E { log ( 1 + B − 1 ( X t ∧ τ R − X ^ t ∧ τ R ) ) } ≲ D δ 1 , δ 2 ( ρ 0 , ρ ^ 0 ) + δ 1 1 ∥ b 1 − b ^ 1 ∥ L 1 ([ 0 , T ] ; L p ( B R )) + δ 1 1 ∥ b 2 − b ^ 2 ∥ L 1 ([ 0 , T ] ; L 1 ( R d )) + δ 1 2 1 ∥ σ − σ ^ ∥ L 2 ([ 0 , T ] ; L 2 r ( B R )) + 1. Here the implicit constant in ≲ \lesssim ≲ is independent of δ 1 , δ 2 \delta_1,\delta_2 δ 1 , δ 2 , and depends only on d , T , R d,T,R d , T , R , the L ∞ ( [ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d ) ) L^\infty([0,T];L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d)) L ∞ ([ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d )) -norms of ρ , ρ ^ \rho,\hat\rho ρ , ρ ^ , and
∥ b ∥ L p ( [ 0 , T ] × B R + 1 ) , ∥ b 2 ∥ L 1 ( [ 0 , T ] × R d 2 ; W s , 1 ( R d 1 ) ) , ∥ ∇ x 1 b 1 ∥ L 1 ( [ 0 , T ] ; L p ( R d 1 ) ) , \|b\|_{L^p([0,T]\times B_{R+1})},\quad
\|b_2\|_{L^1([0,T]\times\mathbb{R}^{d_2};W^{s,1}(\mathbb{R}^{d_1}))},\quad
\|\nabla_{x_1}b_1\|_{L^1([0,T];L^p(\mathbb{R}^{d_1}))}, ∥ b ∥ L p ([ 0 , T ] × B R + 1 ) , ∥ b 2 ∥ L 1 ([ 0 , T ] × R d 2 ; W s , 1 ( R d 1 )) , ∥ ∇ x 1 b 1 ∥ L 1 ([ 0 , T ] ; L p ( R d 1 )) , ∥ ∇ x 2 b 2 ∥ L 1 ( [ 0 , T ] × R d 1 ; L p ( R d 2 ) ) , ∥ ∇ x 1 σ 1 ∥ L 2 ( [ 0 , T ] ; L 2 r ( R d 1 ) ) , ∥ ∇ x σ 2 ∥ L 2 ( [ 0 , T ] ; L 2 r ( R d ) ) . \|\nabla_{x_2}b_2\|_{L^1([0,T]\times\mathbb{R}^{d_1};L^p(\mathbb{R}^{d_2}))},\quad
\|\nabla_{x_1}\sigma_1\|_{L^2([0,T];L^{2r}(\mathbb{R}^{d_1}))},\quad
\|\nabla_x\sigma_2\|_{L^2([0,T];L^{2r}(\mathbb{R}^d))}. ∥ ∇ x 2 b 2 ∥ L 1 ([ 0 , T ] × R d 1 ; L p ( R d 2 )) , ∥ ∇ x 1 σ 1 ∥ L 2 ([ 0 , T ] ; L 2 r ( R d 1 )) , ∥ ∇ x σ 2 ∥ L 2 ([ 0 , T ] ; L 2 r ( R d )) .
Remark 3.4. (1) The above stability estimate implies uniqueness of weak solutions to the FPE (2) in the class L ∞ ( [ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d ) ) L^\infty([0,T];L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d)) L ∞ ([ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d )) . Indeed, if b = b ^ b=\hat b b = b ^ , σ = σ ^ \sigma=\hat\sigma σ = σ ^ , and ρ 0 = ρ ^ 0 \rho_0=\hat\rho_0 ρ 0 = ρ ^ 0 , then the right-hand side of the estimate is bounded independently of δ 1 \delta_1 δ 1 and δ 2 \delta_2 δ 2 . Letting δ 1 , δ 2 ↓ 0 \delta_1,\delta_2\downarrow0 δ 1 , δ 2 ↓ 0 , the left-hand side would diverge unless X t ∧ τ R = X ^ t ∧ τ R X_{t\wedge\tau_R}=\hat X_{t\wedge\tau_R} X t ∧ τ R = X ^ t ∧ τ R a.s. for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . Letting R → ∞ R\to\infty R → ∞ gives X t = X ^ t X_t=\hat X_t X t = X ^ t a.s. for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , and hence ρ = ρ ^ \rho=\hat\rho ρ = ρ ^ in L ∞ ( [ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d ) ) L^\infty([0,T];L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d)) L ∞ ([ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d )) .
(2) The assumption b 1 ∈ L 1 ( [ 0 , T ] ; L ∞ ( R d 1 ; R d 1 ) ) b_1\in L^1([0,T];L^\infty(\mathbb{R}^{d_1};\mathbb{R}^{d_1})) b 1 ∈ L 1 ([ 0 , T ] ; L ∞ ( R d 1 ; R d 1 )) in ( H b ) (\mathcal{H}^b) ( H b ) is used only to guarantee the integrability condition (5) . It may be replaced by any other assumption that ensures the same condition.
Proof of Theorem 3.3
The proof of Theorem 3.3 is divided into two parts. We first prove the existence of a weak solution to the FPE, and then establish the stability estimate for weak solutions. The existence proof relies on the classical DiPerna–Lions theory, which is particularly well suited to the case of degenerate noise. In our setting, an additional difficulty arises, namely ensuring that the solution of the FPE remains a probability measure. When adapting the compactness argument, we must therefore show that no mass is lost in the limit. We first recall the compactness criterions on which the proof rests.
Theorem 3.5 (Banach–Alaoglu–Bourbaki, [5 Theorem 3.16] ). Let X X X be a Banach space. The closed unit ball in X ∗ X^* X ∗ is compact in the weak- topology.*
Definition 3.6 (Equi-integrability). Let U ⊂ R d U \subset\mathbb{R}^{d} U ⊂ R d be a bounded Borel set. A family { f i } i ∈ I ⊂ L 1 ( U ) \{f_i\}_{i\in I} \subset L^1(U) { f i } i ∈ I ⊂ L 1 ( U ) is equi-integrable if, for every η > 0 \eta> 0 η > 0 , there exists δ > 0 \delta> 0 δ > 0 such that, for every Borel set E ⊂ U E \subset U E ⊂ U with L d ( E ) ≤ δ \mathcal{L}^{d}(E) \le\delta L d ( E ) ≤ δ , one has
sup i ∈ I ∫ E ∣ f i ∣ d x ≤ η . \sup_{i\in I}\int_E |f_i|\,\mathrm{d}x \le\eta. i ∈ I sup ∫ E ∣ f i ∣ d x ≤ η .
Theorem 3.7 (Dunford–Pettis, [13 Theorem 1.43] ). Let U ⊂ R d U \subset\mathbb{R}^{d} U ⊂ R d be a bounded Borel set, and let the family { f i } i ∈ I ⊂ L 1 ( U ) \{f_i\}_{i\in I} \subset L^1(U) { f i } i ∈ I ⊂ L 1 ( U ) be bounded. Then the following are equivalent:
{ f i } i ∈ I \{f_i\}_{i\in I} { f i } i ∈ I is equi-integrable;
{ f i } i ∈ I \{f_i\}_{i\in I} { f i } i ∈ I is weakly relatively compact in L 1 ( U ) L^1(U) L 1 ( U ) .
Let ϕ ∈ C c ∞ ( R d ) \phi\in C^\infty_c(\mathbb{R}^{d}) ϕ ∈ C c ∞ ( R d ) be a nonnegative standard mollifier. For 0 < ϵ < 1 0 < \epsilon< 1 0 < ϵ < 1 , set
ϕ ϵ ( x ) : = ϵ − d ϕ ( x ϵ ) , x ∈ R d . (10) \phi^\epsilon(x) := \epsilon^{-d}\phi\left(\frac{x}{\epsilon}\right), \qquad x \in\mathbb{R}^{d}. \tag*{(10)} ϕ ϵ ( x ) := ϵ − d ϕ ( ϵ x ) , x ∈ R d . (10)
We define the mollified coefficients and initial datum by
b ϵ : = b ∗ ϕ ϵ , a ϵ : = a ∗ ϕ ϵ , ρ 0 ϵ : = ρ 0 ∗ ϕ ϵ . b^\epsilon:= b * \phi^\epsilon, \qquad a^\epsilon:= a * \phi^\epsilon, \qquad\rho^\epsilon_0 := \rho_0 * \phi^\epsilon. b ϵ := b ∗ ϕ ϵ , a ϵ := a ∗ ϕ ϵ , ρ 0 ϵ := ρ 0 ∗ ϕ ϵ .
We also set
( b σ ) ϵ : = b σ ∗ ϕ ϵ = b ϵ − 1 2 div ( a ϵ ) . (b^\sigma)^\epsilon:= b^\sigma* \phi^\epsilon= b^\epsilon- \frac{1}{2}\operatorname{div}(a^\epsilon). ( b σ ) ϵ := b σ ∗ ϕ ϵ = b ϵ − 2 1 div ( a ϵ ) .
By the integrability assumptions on b b b , the boundedness of σ \sigma σ , and the standard properties of mollifiers, we have for any ϵ ∈ ( 0 , 1 ) \epsilon\in(0,1) ϵ ∈ ( 0 , 1 ) ,
b ϵ ∈ L 1 ( [ 0 , T ] ; C b ∞ ( R d ; R d ) ) , a ϵ ∈ L 1 ( [ 0 , T ] ; C ∞ ( R d ; R d × d ) ) , ρ 0 ϵ ∈ C b ∞ ( R d ) ∩ L 1 ( R d ) . b^\epsilon\in L^1([0,T];C^\infty_b(\mathbb{R}^{d};\mathbb{R}^{d})), \qquad a^\epsilon\in L^1([0,T];C^\infty(\mathbb{R}^{d};\mathbb{R}^{d\times d})), \qquad\rho^\epsilon_0 \in C^\infty_b(\mathbb{R}^{d}) \cap L^1(\mathbb{R}^{d}). b ϵ ∈ L 1 ([ 0 , T ] ; C b ∞ ( R d ; R d )) , a ϵ ∈ L 1 ([ 0 , T ] ; C ∞ ( R d ; R d × d )) , ρ 0 ϵ ∈ C b ∞ ( R d ) ∩ L 1 ( R d ) .
We consider the regularized FPE
{ ∂ t ρ ϵ + div ( b ϵ ρ ϵ ) = 1 2 ∑ i , j = 1 d ∂ i j ( a i j ϵ ρ ϵ ) , in ( 0 , T ) × R d , ρ ϵ ( 0 , x ) = ρ 0 ϵ ( x ) , in R d . \begin{cases}
\partial_t\rho^\epsilon+\operatorname{div}(b^\epsilon\rho^\epsilon)=\dfrac{1}{2}\sum_{i,j=1}^{d}\partial_{ij}(a^\epsilon_{ij}\rho^\epsilon), & \text{in }(0,T)\times\mathbb{R}^{d},\\
\rho^\epsilon(0,x)=\rho^\epsilon_0(x), & \text{in }\mathbb{R}^{d}.
\end{cases} ⎩ ⎨ ⎧ ∂ t ρ ϵ + div ( b ϵ ρ ϵ ) = 2 1 ∑ i , j = 1 d ∂ ij ( a ij ϵ ρ ϵ ) , ρ ϵ ( 0 , x ) = ρ 0 ϵ ( x ) , in ( 0 , T ) × R d , in R d .
By the classical parabolic theory of linear PDEs (see, e.g., [26 ] ), this equation admits a unique solution
ρ ε ∈ C ( [ 0 , T ] ; C 0 2 ( R d ) ) , \rho^\varepsilon\in C([0,T];C_0^2(\mathbb{R}^d)), ρ ε ∈ C ([ 0 , T ] ; C 0 2 ( R d )) ,
where C 0 2 ( R d ) C_0^2(\mathbb{R}^d) C 0 2 ( R d ) is the space of twice continuously differentiable functions vanishing at infinity together with their derivatives up to order 2. Moreover, using the definition of ( b ε ) (b^\varepsilon) ( b ε ) , it can equivalently be written in divergence form as
{ ∂ t ρ ε + div ( ( b ε ) ρ ε ) = 1 2 ∑ i , j = 1 d ∂ i ( a i j ε ∂ j ρ ε ) , in ( 0 , T ) × R d , ρ ε ( 0 , x ) = ρ 0 ε ( x ) , in R d . (11) \begin{cases}
\partial_t\rho^\varepsilon+\operatorname{div}\left((b^\varepsilon)\rho^\varepsilon\right)=\frac{1}{2}\sum_{i,j=1}^{d}\partial_i\left(a_{ij}^\varepsilon\partial_j\rho^\varepsilon\right), & \text{in }(0,T)\times\mathbb{R}^d,\\
\rho^\varepsilon(0,x)=\rho_0^\varepsilon(x), & \text{in }\mathbb{R}^d. \tag*{(11)}
\end{cases} { ∂ t ρ ε + div ( ( b ε ) ρ ε ) = 2 1 ∑ i , j = 1 d ∂ i ( a ij ε ∂ j ρ ε ) , ρ ε ( 0 , x ) = ρ 0 ε ( x ) , in ( 0 , T ) × R d , in R d . (11)
Having made these preparations, we are now in a position to prove the existence part of Theorem 3.3.
Proof of Theorem 3.3(1). We divide the proof into six steps.
Step I: Non-negativity and uniform a priori estimates for the regularised solutions. Let β : R → R \beta:\mathbb{R}\to\mathbb{R} β : R → R be a convex smooth function with β ( 0 ) = 0 \beta(0)=0 β ( 0 ) = 0 , and ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) . At the regularised level, multiplying both sides of (3.3) by β ′ ( ρ ε ) \beta'(\rho^\varepsilon) β ′ ( ρ ε ) , the usual chain rule (see, e.g., [33 , 24 ] ) gives
d d t ∫ R d β ( ρ ε ) d x + ∫ R d ( ρ ε β ′ ( ρ ε ) − β ( ρ ε ) ) div [ ( b ε ) ] d x + ∫ R d β ′ ′ ( ρ ε ) 2 ( ∇ ρ ε ) ∗ a ε ∇ ρ ε d x = 0. (12) \frac{\mathrm{d}}{\mathrm{d}t}\int_{\mathbb{R}^d}\beta(\rho^\varepsilon)\,\mathrm{d}x+\int_{\mathbb{R}^d}\left(\rho^\varepsilon\beta'(\rho^\varepsilon)-\beta(\rho^\varepsilon)\right)\operatorname{div}\left[(b^\varepsilon)\right]\,\mathrm{d}x+\int_{\mathbb{R}^d}\frac{\beta''(\rho^\varepsilon)}{2}(\nabla\rho^\varepsilon)^*a^\varepsilon\nabla\rho^\varepsilon\,\mathrm{d}x=0. \tag*{(12)} d t d ∫ R d β ( ρ ε ) d x + ∫ R d ( ρ ε β ′ ( ρ ε ) − β ( ρ ε ) ) div [ ( b ε ) ] d x + ∫ R d 2 β ′′ ( ρ ε ) ( ∇ ρ ε ) ∗ a ε ∇ ρ ε d x = 0. (12)
Under Assumption 3.1, suitable choices of the convex function β \beta β in the above identity (3.4) give the following estimates. Since the corresponding computations are standard for the regularised equation and can be found in [33 , 24 ] , we omit the detailed calculation here and only state the resulting estimates.
(i) Taking smooth approximations of
β ( u ) = u − , u ∈ R , \beta(u)=u^{-},\qquad u\in\mathbb{R}, β ( u ) = u − , u ∈ R ,
we obtain the non-negativity of the regularised solutions
ρ ε ( t , x ) ≥ 0 , ( t , x ) ∈ [ 0 , T ] × R d , \rho^\varepsilon(t,x)\geq0,\qquad(t,x)\in[0,T]\times\mathbb{R}^d, ρ ε ( t , x ) ≥ 0 , ( t , x ) ∈ [ 0 , T ] × R d ,
since ρ 0 ε ≥ 0 \rho_0^\varepsilon\geq0 ρ 0 ε ≥ 0 .
(ii) Taking smooth approximations of
β ( u ) = ( u − ∥ ρ 0 ∥ L x ∞ ) + , u ∈ R , \beta(u)=\left(u-\|\rho_0\|_{L_x^\infty}\right)^{+},\qquad u\in\mathbb{R}, β ( u ) = ( u − ∥ ρ 0 ∥ L x ∞ ) + , u ∈ R ,
we obtain the uniform L ∞ L^\infty L ∞ estimate
sup ϵ ∈ ( 0 , 1 ) ∥ ρ ϵ ∥ L ∞ ( [ 0 , T ] × R d ) ≤ ∥ ρ 0 ∥ L x ∞ exp ( A ( T ) ) , (13) \sup_{\epsilon\in(0,1)} \|\rho^\epsilon\|_{L^\infty([0,T]\times\mathbb{R}^d)} \le\|\rho_0\|_{L_x^\infty}\exp(A(T)), \tag*{(13)} ϵ ∈ ( 0 , 1 ) sup ∥ ρ ϵ ∥ L ∞ ([ 0 , T ] × R d ) ≤ ∥ ρ 0 ∥ L x ∞ exp ( A ( T )) , (13)
where for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
A ( t ) : = ∫ 0 t ∥ ( div b σ ) − ( s , ⋅ ) ∥ L x ∞ d s . A(t):=\int_0^t \|(\operatorname{div}b^\sigma)^-(s,\cdot)\|_{L_x^\infty}\,\mathrm{d}s. A ( t ) := ∫ 0 t ∥ ( div b σ ) − ( s , ⋅ ) ∥ L x ∞ d s .
(iii) Taking smooth approximations of
β ( u ) = ∣ u ∣ r , u ∈ R , 1 ≤ r < ∞ , \beta(u)=|u|^r,\qquad u\in\mathbb{R},\ 1\le r<\infty, β ( u ) = ∣ u ∣ r , u ∈ R , 1 ≤ r < ∞ ,
and applying Grönwall’s inequality, we derive
sup ϵ ∈ ( 0 , 1 ) ∥ ρ ϵ ∥ L t ∞ L x r ≤ ∥ ρ 0 ∥ L x r exp [ ( 1 − 1 r ) A ( T ) ] , 1 ≤ r < ∞ . (14) \sup_{\epsilon\in(0,1)}\|\rho^\epsilon\|_{L_t^\infty L_x^r}\le\|\rho_0\|_{L_x^r}\exp\left[\left(1-\frac{1}{r}\right)A(T)\right],\qquad1\le r<\infty. \tag*{(14)} ϵ ∈ ( 0 , 1 ) sup ∥ ρ ϵ ∥ L t ∞ L x r ≤ ∥ ρ 0 ∥ L x r exp [ ( 1 − r 1 ) A ( T ) ] , 1 ≤ r < ∞. (14)
Step II: Mass conservation for the regularised solutions. For R > 0 R>0 R > 0 , let χ R ∈ C c ∞ ( R d ) \chi_R\in C_c^\infty(\mathbb{R}^d) χ R ∈ C c ∞ ( R d ) be a cut-off function satisfying
0 ≤ χ R ≤ 1 , χ R = 1 on B R , χ R = 0 outside B 2 R , 0\le\chi_R\le1,\qquad\chi_R=1\ \text{on }B_R,\qquad\chi_R=0\ \text{outside }B_{2R}, 0 ≤ χ R ≤ 1 , χ R = 1 on B R , χ R = 0 outside B 2 R ,
and
∣ ∇ χ R ∣ ≲ 1 R , ∥ ∇ 2 χ R ∥ ≲ 1 R 2 . |\nabla\chi_R|\lesssim\frac{1}{R},\qquad\|\nabla^2\chi_R\|\lesssim\frac{1}{R^2}. ∣∇ χ R ∣ ≲ R 1 , ∥ ∇ 2 χ R ∥ ≲ R 2 1 .
Multiplying (11) by χ R \chi_R χ R and integrating by parts, we obtain for a.e. t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) ,
d d t ∫ R d ρ ϵ χ R d x = ∫ R d ⟨ ( b σ ) ϵ , ∇ χ R ⟩ ρ ϵ d x + 1 2 ∫ R d ρ ϵ div ( ( a ϵ ) ∗ ∇ χ R ) d x . \frac{\mathrm{d}}{\mathrm{d}t}\int_{\mathbb{R}^d}\rho^\epsilon\chi_R\,\mathrm{d}x
=
\int_{\mathbb{R}^d}\langle(b^\sigma)^\epsilon,\nabla\chi_R\rangle\rho^\epsilon\,\mathrm{d}x
+\frac{1}{2}\int_{\mathbb{R}^d}\rho^\epsilon\operatorname{div}\left((a^\epsilon)^*\nabla\chi_R\right)\,\mathrm{d}x. d t d ∫ R d ρ ϵ χ R d x = ∫ R d ⟨( b σ ) ϵ , ∇ χ R ⟩ ρ ϵ d x + 2 1 ∫ R d ρ ϵ div ( ( a ϵ ) ∗ ∇ χ R ) d x .
which gives, for every t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) ,
∫ R d ρ ϵ ( t , ⋅ ) χ R d x − ∫ R d ρ ϵ ( 0 , ⋅ ) χ R d x = ∫ 0 t ∫ R d ⟨ ( b σ ) ϵ , ∇ χ R ⟩ ρ ϵ d x d s + 1 2 ∫ 0 t ∫ R d ρ ϵ div ( ( a ϵ ) ∗ ∇ χ R ) d x d s . (15) \int_{\mathbb{R}^d}\rho^\epsilon(t,\cdot)\chi_R\,\mathrm{d}x-\int_{\mathbb{R}^d}\rho^\epsilon(0,\cdot)\chi_R\,\mathrm{d}x
=
\int_0^t\int_{\mathbb{R}^d}\langle(b^\sigma)^\epsilon,\nabla\chi_R\rangle\rho^\epsilon\,\mathrm{d}x\,\mathrm{d}s
+\frac{1}{2}\int_0^t\int_{\mathbb{R}^d}\rho^\epsilon\operatorname{div}\left((a^\epsilon)^*\nabla\chi_R\right)\,\mathrm{d}x\,\mathrm{d}s. \tag*{(15)} ∫ R d ρ ϵ ( t , ⋅ ) χ R d x − ∫ R d ρ ϵ ( 0 , ⋅ ) χ R d x = ∫ 0 t ∫ R d ⟨( b σ ) ϵ , ∇ χ R ⟩ ρ ϵ d x d s + 2 1 ∫ 0 t ∫ R d ρ ϵ div ( ( a ϵ ) ∗ ∇ χ R ) d x d s . (15)
Since ( b σ ) ϵ ∈ L 1 ( [ 0 , T ] ; C ∞ ( R d ; R d ) ) (b^\sigma)^\epsilon\in L^1([0,T];C^\infty(\mathbb{R}^d;\mathbb{R}^d)) ( b σ ) ϵ ∈ L 1 ([ 0 , T ] ; C ∞ ( R d ; R d )) , a ϵ ∈ L 1 ( [ 0 , T ] ; C b ∞ ( R d ; R d × R d ) ) a^\epsilon\in L^1([0,T];C_b^\infty(\mathbb{R}^d;\mathbb{R}^{d\times\mathbb{R}^d})) a ϵ ∈ L 1 ([ 0 , T ] ; C b ∞ ( R d ; R d × R d )) , ρ ϵ ∈ L ∞ ( [ 0 , T ] ; L 1 ( R d ) ) \rho^\epsilon\in L^\infty([0,T];L^1(\mathbb{R}^d)) ρ ϵ ∈ L ∞ ([ 0 , T ] ; L 1 ( R d )) with ρ ϵ ≥ 0 \rho^\epsilon\ge0 ρ ϵ ≥ 0 , and ∇ χ R , ∇ 2 χ R \nabla\chi_R,\nabla^2\chi_R ∇ χ R , ∇ 2 χ R are supported in B 2 R ∖ B R B_{2R}\setminus B_R B 2 R ∖ B R , the right-hand side of (15) tends to 0 0 0 as R → ∞ R\to\infty R → ∞ by the dominated convergence theorem. Thus,
∫ R d ρ ϵ ( t , ⋅ ) d x = ∫ R d ρ ϵ ( 0 , ⋅ ) d x = 1 , \int_{\mathbb{R}^d}\rho^\epsilon(t,\cdot)\,\mathrm{d}x=\int_{\mathbb{R}^d}\rho^\epsilon(0,\cdot)\,\mathrm{d}x=1, ∫ R d ρ ϵ ( t , ⋅ ) d x = ∫ R d ρ ϵ ( 0 , ⋅ ) d x = 1 ,
so that ρ ϵ ( t , ⋅ ) \rho^\epsilon(t,\cdot) ρ ϵ ( t , ⋅ ) is a probability density for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] .
Step III: Compactness and extraction of a subsequence. From (13) and (14) , ( ρ ϵ ) ϵ ∈ ( 0 , 1 ) (\rho^\epsilon)_{\epsilon\in(0,1)} ( ρ ϵ ) ϵ ∈ ( 0 , 1 ) is uniformly bounded in L ∞ ( [ 0 , T ] ; L r ( R d ) ) L^\infty([0,T];L^r(\mathbb{R}^d)) L ∞ ([ 0 , T ] ; L r ( R d )) for every r ∈ ( 1 , ∞ ] r\in(1,\infty] r ∈ ( 1 , ∞ ] . By the Banach–Alaoglu–Bourbaki theorem (Theorem 3.5 ), up to extracting a subsequence, not relabeled, for every r ∈ ( 1 , ∞ ] r\in(1,\infty] r ∈ ( 1 , ∞ ] , as ϵ → 0 + \epsilon\to0^+ ϵ → 0 + , we have the weak-∗ * ∗ convergence
ρ ϵ ⇀ ∗ ρ in L ∞ ( [ 0 , T ] ; L r ( R d ) ) . (16) \rho^\epsilon\overset{*}{\rightharpoonup} \rho\quad\text{in } L^\infty([0,T];L^r(\mathbb{R}^d)). \tag*{(16)} ρ ϵ ⇀ ∗ ρ in L ∞ ([ 0 , T ] ; L r ( R d )) . (16)
The case r = 1 r=1 r = 1 requires a separate argument, one cannot argue directly by duality, since the dual space of L ∞ L^\infty L ∞ is strictly larger than L 1 L^1 L 1 . Instead, we show that ( ρ ϵ ) ϵ ∈ ( 0 , 1 ) (\rho^\epsilon)_{\epsilon\in(0,1)} ( ρ ϵ ) ϵ ∈ ( 0 , 1 ) is weakly relatively compact in L l o c 1 ( [ 0 , T ] × R d ) L^1_{\mathrm{loc}}([0,T]\times\mathbb{R}^d) L loc 1 ([ 0 , T ] × R d ) . To this end, let ( ρ 0 n ) n ≥ 1 ⊂ C c ∞ ( R d ) (\rho_0^n)_{n\geq1}\subset C_c^\infty(\mathbb{R}^d) ( ρ 0 n ) n ≥ 1 ⊂ C c ∞ ( R d ) be a nonnegative sequence converging to ρ 0 \rho_0 ρ 0 in L 1 ( R d ) L^1(\mathbb{R}^d) L 1 ( R d ) as n → ∞ n\to\infty n → ∞ , and for each n ≥ 1 n\geq1 n ≥ 1 , denote by ρ n , ϵ \rho^{n,\epsilon} ρ n , ϵ the corresponding approximating solutions constructed as above with initial datum ρ 0 n , ϵ : = ρ 0 n ∗ ϕ ϵ \rho_0^{n,\epsilon}:=\rho_0^n*\phi^\epsilon ρ 0 n , ϵ := ρ 0 n ∗ ϕ ϵ . For fixed n ≥ 1 n\geq1 n ≥ 1 and every 1 ≤ r ≤ ∞ 1\leq r\leq\infty 1 ≤ r ≤ ∞ , (3.5) and (3.6) give a uniform bound
∥ ρ n , ϵ ∥ L t ∞ L x r ≤ C ( n , r , T ) , \|\rho^{n,\epsilon}\|_{L^\infty_tL^r_x}\leq C(n,r,T), ∥ ρ n , ϵ ∥ L t ∞ L x r ≤ C ( n , r , T ) ,
with C ( n , r , T ) C(n,r,T) C ( n , r , T ) being a positive constant, depending on n , r , T n,r,T n , r , T and independent of ϵ \epsilon ϵ . Hence, choosing any fixed r > 1 r>1 r > 1 , for every bounded cylinder Q ⊂ [ 0 , T ] × R d Q\subset[0,T]\times\mathbb{R}^d Q ⊂ [ 0 , T ] × R d and every Borel set E ⊂ Q E\subset Q E ⊂ Q ,
sup 0 < ϵ < 1 ∫ E ρ n , ϵ d x d t ≤ C ( n , r , T ) L d + 1 ( E ) 1 / r ′ , \sup_{0<\epsilon<1}\int_E\rho^{n,\epsilon}\,\mathrm{d}x\,\mathrm{d}t\leq C(n,r,T)\mathcal{L}^{d+1}(E)^{1/r'}, 0 < ϵ < 1 sup ∫ E ρ n , ϵ d x d t ≤ C ( n , r , T ) L d + 1 ( E ) 1/ r ′ ,
where r ′ = r r − 1 r'=\frac{r}{r-1} r ′ = r − 1 r . Thus, for each fixed n n n , the family { ρ n , ϵ } 0 < ϵ < 1 \{\rho^{n,\epsilon}\}_{0<\epsilon<1} { ρ n , ϵ } 0 < ϵ < 1 is equi-integrable on every bounded cylinder Q ⊂ [ 0 , T ] × R d Q\subset[0,T]\times\mathbb{R}^d Q ⊂ [ 0 , T ] × R d .
On the other hand, by linearity, ρ ϵ − ρ n , ϵ \rho^\epsilon-\rho^{n,\epsilon} ρ ϵ − ρ n , ϵ solves the same regularized equation with initial datum ρ 0 ϵ − ρ 0 n , ϵ \rho_0^\epsilon-\rho_0^{n,\epsilon} ρ 0 ϵ − ρ 0 n , ϵ . Hence, the corresponding L 1 L^1 L 1 estimate (3.6) with r = 1 r=1 r = 1 gives
∥ ρ ϵ − ρ n , ϵ ∥ L t ∞ L x 1 ≤ C 0 ∥ ρ 0 ϵ − ρ 0 n , ϵ ∥ L x 1 ≤ C 0 ∥ ρ 0 − ρ 0 n ∥ L x 1 . (17) \|\rho^\epsilon-\rho^{n,\epsilon}\|_{L^\infty_tL^1_x}\leq C_0\|\rho_0^\epsilon-\rho_0^{n,\epsilon}\|_{L^1_x}\leq C_0\|\rho_0-\rho_0^n\|_{L^1_x}. \tag*{(17)} ∥ ρ ϵ − ρ n , ϵ ∥ L t ∞ L x 1 ≤ C 0 ∥ ρ 0 ϵ − ρ 0 n , ϵ ∥ L x 1 ≤ C 0 ∥ ρ 0 − ρ 0 n ∥ L x 1 . (17)
where C 0 : = exp ( A ( T ) ) C_0:=\exp(A(T)) C 0 := exp ( A ( T )) .
We now show that { ρ ϵ } 0 < ϵ < 1 \{\rho^\epsilon\}_{0<\epsilon<1} { ρ ϵ } 0 < ϵ < 1 is equi-integrable on a fixed bounded cylinder Q ⊂ [ 0 , T ] × R d Q\subset[0,T]\times\mathbb{R}^d Q ⊂ [ 0 , T ] × R d . Let η > 0 \eta>0 η > 0 . Choose n n n so large that
T C 0 ∥ ρ 0 − ρ 0 n ∥ L x 1 ≤ η 2 . TC_0\|\rho_0-\rho_0^n\|_{L^1_x}\leq\frac{\eta}{2}. T C 0 ∥ ρ 0 − ρ 0 n ∥ L x 1 ≤ 2 η .
For this fixed n n n , choose δ > 0 \delta>0 δ > 0 such that, whenever E ⊂ Q E\subset Q E ⊂ Q and L d + 1 ( E ) ≤ δ \mathcal{L}^{d+1}(E)\leq\delta L d + 1 ( E ) ≤ δ ,
sup 0 < ϵ < 1 ∫ E ρ n , ϵ d x d t ≤ η 2 . \sup_{0<\epsilon<1}\int_E\rho^{n,\epsilon}\,\mathrm{d}x\,\mathrm{d}t\leq\frac{\eta}{2}. 0 < ϵ < 1 sup ∫ E ρ n , ϵ d x d t ≤ 2 η .
Then, combining this with (17) ,
sup 0 < ϵ < 1 ∫ E ρ ϵ d x d t ≤ T C 0 ∥ ρ 0 − ρ 0 n ∥ L x 1 + sup 0 < ϵ < 1 ∫ E ρ n , ϵ d x d t ≤ η . \sup_{0<\epsilon<1}\int_E\rho^\epsilon\,\mathrm{d}x\,\mathrm{d}t\leq TC_0\|\rho_0-\rho_0^n\|_{L^1_x}+\sup_{0<\epsilon<1}\int_E\rho^{n,\epsilon}\,\mathrm{d}x\,\mathrm{d}t\leq\eta. 0 < ϵ < 1 sup ∫ E ρ ϵ d x d t ≤ T C 0 ∥ ρ 0 − ρ 0 n ∥ L x 1 + 0 < ϵ < 1 sup ∫ E ρ n , ϵ d x d t ≤ η .
Therefore, { ρ ϵ } 0 < ϵ < 1 \{\rho^\epsilon\}_{0<\epsilon<1} { ρ ϵ } 0 < ϵ < 1 is equi-integrable on every bounded cylinder. By the Dunford–Pettis theorem (Theorem 3.7) and a diagonal extraction over an increasing sequence of bounded cylinders, we may assume (passing to a further subsequence, still denoted by ρ ϵ \rho^\epsilon ρ ϵ ) that the following weak convergence holds:
ρ ϵ ⇀ ρ in L l o c 1 ( [ 0 , T ] × R d ) . (18) \rho^\epsilon\rightharpoonup\rho\quad\text{in } L^1_{\mathrm{loc}}([0,T]\times\mathbb{R}^d). \tag*{(18)} ρ ϵ ⇀ ρ in L loc 1 ([ 0 , T ] × R d ) . (18)
The limit in (18) coincides with the one in (16) by uniqueness of weak limits. In particular,
ρ ≥ 0 for a.e. ( t , x ) ∈ [ 0 , T ] × R d . \rho\ge0 \quad\text{for a.e. } (t,x)\in[0,T]\times\mathbb{R}^d. ρ ≥ 0 for a.e. ( t , x ) ∈ [ 0 , T ] × R d .
Step IV: Uniform first moment estimate. We aim to establish the following estimate
sup 0 < ϵ < 1 ess sup t ∈ [ 0 , T ] ∫ R d ∣ x ∣ ρ ϵ ( t , x ) d x = : C < ∞ . (19) \sup_{0<\epsilon<1}\operatorname*{ess\,sup}_{t\in[0,T]}\int_{\mathbb{R}^d}|x|\rho^\epsilon(t,x)\,dx =: C < \infty. \tag*{(19)} 0 < ϵ < 1 sup t ∈ [ 0 , T ] ess sup ∫ R d ∣ x ∣ ρ ϵ ( t , x ) d x =: C < ∞. (19)
Let σ ϵ : = ( a ϵ ) 1 / 2 \sigma^\epsilon:= (a^\epsilon)^{1/2} σ ϵ := ( a ϵ ) 1/2 be the symmetric non-negative square root of a ϵ a^\epsilon a ϵ . Since a ϵ a^\epsilon a ϵ belongs to L 1 ( [ 0 , T ] ; C b 2 ( R d ; R d × d ) ) L^1([0,T];C_b^2(\mathbb{R}^d;\mathbb{R}^{d\times d})) L 1 ([ 0 , T ] ; C b 2 ( R d ; R d × d )) and a ϵ a^\epsilon a ϵ is symmetric and non-negative definite, [31 Lemma 3.2.3] implies that σ ϵ \sigma^\epsilon σ ϵ is bounded and Lipschitz in x ∈ R d x\in\mathbb{R}^d x ∈ R d , with Lipschitz constant integrable in t t t . Hence the SDE
d X t ϵ = b ϵ ( t , X t ϵ ) d t + σ ϵ ( t , X t ϵ ) d W t , L X 0 ϵ ( d x ) = ρ 0 ϵ ( x ) d x , dX_t^\epsilon=b^\epsilon(t,X_t^\epsilon)\,dt+\sigma^\epsilon(t,X_t^\epsilon)\,dW_t,\qquad\mathcal{L}_{X_0^\epsilon}(dx)=\rho_0^\epsilon(x)\,dx, d X t ϵ = b ϵ ( t , X t ϵ ) d t + σ ϵ ( t , X t ϵ ) d W t , L X 0 ϵ ( d x ) = ρ 0 ϵ ( x ) d x ,
is well-posed. By Itô’s formula, the law of X t ϵ X_t^\epsilon X t ϵ is a weakly continuous solution to the regularized FPE. Combined with the uniqueness of weakly continuous solutions to the FPE, see Appendix A.4 in [33 ] , this gives
L X t ϵ ( d x ) = ρ ϵ ( t , x ) d x for a.e. t ∈ [ 0 , T ] . \mathcal{L}_{X_t^\epsilon}(dx)=\rho^\epsilon(t,x)\,dx\quad\text{for a.e. }t\in[0,T]. L X t ϵ ( d x ) = ρ ϵ ( t , x ) d x for a.e. t ∈ [ 0 , T ] .
Applying Itô’s formula and taking expectations, we have
ess sup t ∈ [ 0 , T ] ∫ R d ∣ x ∣ ρ ϵ ( t , x ) d x = sup t ∈ [ 0 , T ] E ∣ X t ϵ ∣ ≤ ∫ R d ∣ x ∣ ρ 0 ϵ ( x ) d x + ( ∫ 0 T ∫ R d ∥ σ ϵ ( s , x ) ∥ 2 ρ ϵ ( s , x ) d x d s ) 1 / 2 + ∫ 0 T ∫ R d ∣ b ϵ ( s , x ) ∣ ρ ϵ ( s , x ) d x d s . \begin{aligned}
\operatorname*{ess\,sup}_{t\in[0,T]}\int_{\mathbb{R}^d}|x|\rho^\epsilon(t,x)\,dx
=\sup_{t\in[0,T]}\mathbb{E}|X_t^\epsilon| \\
&\le\int_{\mathbb{R}^d}|x|\rho_0^\epsilon(x)\,dx+
\left(\int_0^T\int_{\mathbb{R}^d}\|\sigma^\epsilon(s,x)\|^2\rho^\epsilon(s,x)\,dx\,ds\right)^{1/2}
+\int_0^T\int_{\mathbb{R}^d}|b^\epsilon(s,x)|\rho^\epsilon(s,x)\,dx\,ds.
\end{aligned} t ∈ [ 0 , T ] ess sup ∫ R d ∣ x ∣ ρ ϵ ( t , x ) d x = t ∈ [ 0 , T ] sup E ∣ X t ϵ ∣ ≤ ∫ R d ∣ x ∣ ρ 0 ϵ ( x ) d x + ( ∫ 0 T ∫ R d ∥ σ ϵ ( s , x ) ∥ 2 ρ ϵ ( s , x ) d x d s ) 1/2 + ∫ 0 T ∫ R d ∣ b ϵ ( s , x ) ∣ ρ ϵ ( s , x ) d x d s .
Since ρ 0 ϵ = ρ 0 ∗ ϕ ϵ \rho_0^\epsilon=\rho_0*\phi^\epsilon ρ 0 ϵ = ρ 0 ∗ ϕ ϵ and 0 < ϵ < 1 0<\epsilon<1 0 < ϵ < 1 ,
∫ R d ∣ x ∣ ρ 0 ϵ ( x ) d x ≤ ∫ R d ∣ x ∣ ρ 0 ( x ) d x + 1. \int_{\mathbb{R}^d}|x|\rho_0^\epsilon(x)\,dx\le\int_{\mathbb{R}^d}|x|\rho_0(x)\,dx+1. ∫ R d ∣ x ∣ ρ 0 ϵ ( x ) d x ≤ ∫ R d ∣ x ∣ ρ 0 ( x ) d x + 1.
Moreover, from ( H σ ) (\mathcal{H}^\sigma) ( H σ ) ,
∫ 0 T ∫ R d ∥ σ ϵ ( s , x ) ∥ 2 ρ ϵ ( s , x ) d x d s ≤ ∥ σ ∥ L t 2 L x ∞ 2 . \int_0^T\int_{\mathbb{R}^d}\|\sigma^\epsilon(s,x)\|^2\rho^\epsilon(s,x)\,dx\,ds\le\|\sigma\|_{L_t^2L_x^\infty}^2. ∫ 0 T ∫ R d ∥ σ ϵ ( s , x ) ∥ 2 ρ ϵ ( s , x ) d x d s ≤ ∥ σ ∥ L t 2 L x ∞ 2 .
For the drift term, using the structure b ϵ = ( b 1 ∗ 1 ϕ ϵ , b 2 ∗ ϕ ϵ ) b^\epsilon=(b_1\ast_1\phi^\epsilon,b_2\ast\phi^\epsilon) b ϵ = ( b 1 ∗ 1 ϕ ϵ , b 2 ∗ ϕ ϵ ) , mass conservation, and (13) , we obtain
∫ 0 T ∫ R d ∣ b ϵ ( s , x ) ∣ ρ ϵ ( s , x ) d x d s ≤ ∥ b 1 ∥ L t 1 L x 1 ∞ + ∥ ρ ϵ ∥ L ∞ ( [ 0 , T ] × R d ) ∥ b 2 ∥ L t 1 L x 1 ≤ ∥ b 1 ∥ L t 1 L x 1 ∞ + C 0 ∥ ρ 0 ∥ L x ∞ ∥ b 2 ∥ L t 1 L x 1 . \int_0^T\int_{\mathbb{R}^d}|b^\epsilon(s,x)|\rho^\epsilon(s,x)\,\mathrm{d}x\,\mathrm{d}s\leq\|b_1\|_{L_t^1L_{x_1}^\infty}+\|\rho^\epsilon\|_{L^\infty([0,T]\times\mathbb{R}^d)}\|b_2\|_{L_t^1L_x^1}
\leq\|b_1\|_{L_t^1L_{x_1}^\infty}+C_0\|\rho_0\|_{L_x^\infty}\|b_2\|_{L_t^1L_x^1}. ∫ 0 T ∫ R d ∣ b ϵ ( s , x ) ∣ ρ ϵ ( s , x ) d x d s ≤ ∥ b 1 ∥ L t 1 L x 1 ∞ + ∥ ρ ϵ ∥ L ∞ ([ 0 , T ] × R d ) ∥ b 2 ∥ L t 1 L x 1 ≤ ∥ b 1 ∥ L t 1 L x 1 ∞ + C 0 ∥ ρ 0 ∥ L x ∞ ∥ b 2 ∥ L t 1 L x 1 .
Combining the above estimates proves (19) . Here the symbol ∗ 1 \ast_1 ∗ 1 denotes convolution with respect to the x 1 x_1 x 1 -variable only. We assume that the mollifier splits as ϕ ( x 1 , x 2 ) = ϕ 1 ( x 1 ) ϕ 2 ( x 2 ) \phi(x_1,x_2)=\phi_1(x_1)\phi_2(x_2) ϕ ( x 1 , x 2 ) = ϕ 1 ( x 1 ) ϕ 2 ( x 2 ) for ( x 1 , x 2 ) ∈ R d 1 × R d 2 (x_1,x_2)\in\mathbb{R}^{d_1}\times\mathbb{R}^{d_2} ( x 1 , x 2 ) ∈ R d 1 × R d 2 , and set ϕ 1 ϵ ( x 1 ) = ϵ − d 1 ϕ 1 ( x 1 / ϵ ) \phi_1^\epsilon(x_1)=\epsilon^{-d_1}\phi_1(x_1/\epsilon) ϕ 1 ϵ ( x 1 ) = ϵ − d 1 ϕ 1 ( x 1 / ϵ ) , where ϕ 1 , ϕ 2 \phi_1,\phi_2 ϕ 1 , ϕ 2 are nonnegative standard mollifiers on R d 1 \mathbb{R}^{d_1} R d 1 and R d 2 \mathbb{R}^{d_2} R d 2 , respectively. Then b 1 ∗ 1 ϕ ϵ : = b 1 ∗ ϕ 1 ϵ b_1\ast_1\phi^\epsilon:=b_1\ast\phi_1^\epsilon b 1 ∗ 1 ϕ ϵ := b 1 ∗ ϕ 1 ϵ .
Step V: Recovery of the total mass of the limit density. For every nonnegative ψ ∈ L ∞ ( [ 0 , T ] ) \psi\in L^\infty([0,T]) ψ ∈ L ∞ ([ 0 , T ]) and the cut-off function χ R \chi_R χ R , the local weak convergence in (18) gives
∫ 0 T ∫ R d ρ ( t , x ) ψ ( t ) χ R ( x ) d x d t = lim ϵ → 0 ∫ 0 T ∫ R d ρ ϵ ( t , x ) ψ ( t ) χ R ( x ) d x d t ≤ ∫ 0 T ψ ( t ) d t . \int_0^T\int_{\mathbb{R}^d}\rho(t,x)\psi(t)\chi_R(x)\,\mathrm{d}x\,\mathrm{d}t=\lim_{\epsilon\to0}\int_0^T\int_{\mathbb{R}^d}\rho^\epsilon(t,x)\psi(t)\chi_R(x)\,\mathrm{d}x\,\mathrm{d}t\leq\int_0^T\psi(t)\,\mathrm{d}t. ∫ 0 T ∫ R d ρ ( t , x ) ψ ( t ) χ R ( x ) d x d t = ϵ → 0 lim ∫ 0 T ∫ R d ρ ϵ ( t , x ) ψ ( t ) χ R ( x ) d x d t ≤ ∫ 0 T ψ ( t ) d t .
Letting R → ∞ R\to\infty R → ∞ yields
∫ R d ρ ( t , x ) d x ≤ 1 , for a.e. t ∈ [ 0 , T ] . \int_{\mathbb{R}^d}\rho(t,x)\,\mathrm{d}x\leq1,\qquad\text{for a.e. }t\in[0,T]. ∫ R d ρ ( t , x ) d x ≤ 1 , for a.e. t ∈ [ 0 , T ] .
Thus, for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , ρ ( t , ⋅ ) \rho(t,\cdot) ρ ( t , ⋅ ) is a sub-probability density.
Conversely, by (19) , for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
∫ R d ρ ϵ ( t , x ) ( 1 − χ R ( x ) ) d x ≤ ∫ ∣ x ∣ ≥ R ρ ϵ ( t , x ) d x ≤ 1 R ∫ R d ∣ x ∣ ρ ϵ ( t , x ) d x ≤ C R . \int_{\mathbb{R}^d}\rho^\epsilon(t,x)(1-\chi_R(x))\,\mathrm{d}x\leq\int_{|x|\geq R}\rho^\epsilon(t,x)\,\mathrm{d}x\leq\frac{1}{R}\int_{\mathbb{R}^d}|x|\rho^\epsilon(t,x)\,\mathrm{d}x\leq\frac{C}{R}. ∫ R d ρ ϵ ( t , x ) ( 1 − χ R ( x )) d x ≤ ∫ ∣ x ∣ ≥ R ρ ϵ ( t , x ) d x ≤ R 1 ∫ R d ∣ x ∣ ρ ϵ ( t , x ) d x ≤ R C .
Since ρ ϵ ( t , ⋅ ) \rho^\epsilon(t,\cdot) ρ ϵ ( t , ⋅ ) is a probability density,
∫ R d ρ ϵ ( t , x ) χ R ( x ) d x = 1 − ∫ R d ρ ϵ ( t , x ) ( 1 − χ R ( x ) ) d x ≥ 1 − C R . \int_{\mathbb{R}^d}\rho^\epsilon(t,x)\chi_R(x)\,\mathrm{d}x=1-\int_{\mathbb{R}^d}\rho^\epsilon(t,x)(1-\chi_R(x))\,\mathrm{d}x\geq1-\frac{C}{R}. ∫ R d ρ ϵ ( t , x ) χ R ( x ) d x = 1 − ∫ R d ρ ϵ ( t , x ) ( 1 − χ R ( x )) d x ≥ 1 − R C .
Multiplying by ψ ( t ) \psi(t) ψ ( t ) , integrating over ( 0 , T ) (0,T) ( 0 , T ) , and passing to the limit ϵ → 0 \epsilon\to0 ϵ → 0 , we obtain
∫ 0 T ∫ R d ρ ( t , x ) ψ ( t ) χ R ( x ) d x d t ≥ ( 1 − C R ) ∫ 0 T ψ ( t ) d t . \int_0^T\int_{\mathbb{R}^d}\rho(t,x)\psi(t)\chi_R(x)\,\mathrm{d}x\,\mathrm{d}t\geq\left(1-\frac{C}{R}\right)\int_0^T\psi(t)\,\mathrm{d}t. ∫ 0 T ∫ R d ρ ( t , x ) ψ ( t ) χ R ( x ) d x d t ≥ ( 1 − R C ) ∫ 0 T ψ ( t ) d t .
Letting R → ∞ R\to\infty R → ∞ gives
∫ 0 T ∫ R d ρ ( t , x ) ψ ( t ) d x d t ≥ ∫ 0 T ψ ( t ) d t . \int_0^T\int_{\mathbb{R}^d}\rho(t,x)\psi(t)\,\mathrm{d}x\,\mathrm{d}t\geq\int_0^T\psi(t)\,\mathrm{d}t. ∫ 0 T ∫ R d ρ ( t , x ) ψ ( t ) d x d t ≥ ∫ 0 T ψ ( t ) d t .
Together with the sub-probability estimate, this proves
∫ R d ρ ( t , x ) d x = 1 , for a.e. t ∈ [ 0 , T ] . \int_{\mathbb{R}^d}\rho(t,x)\,\mathrm{d}x=1,\qquad\text{for a.e. }t\in[0,T]. ∫ R d ρ ( t , x ) d x = 1 , for a.e. t ∈ [ 0 , T ] .
Consequently,
ρ ∈ L ∞ ( [ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d ) ) , \rho\in L^\infty([0,T];L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d)), ρ ∈ L ∞ ([ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d )) ,
and ρ ( t , ⋅ ) \rho(t,\cdot) ρ ( t , ⋅ ) is a probability density for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] .
Step VI: Verification that the limit ρ \rho ρ satisfies (6) . Let φ ∈ C c 1 , 2 ( [ 0 , T ) × R d ) \varphi\in C_c^{1,2}([0,T)\times\mathbb{R}^d) φ ∈ C c 1 , 2 ([ 0 , T ) × R d ) . The regularized solutions satisfy
∫ 0 T ∫ R d ρ ε [ ∂ t φ + b ε ⋅ ∇ φ + 1 2 tr ( a ε ∇ 2 φ ) ] d x d t + ∫ R d ρ 0 ε ( x ) φ ( 0 , x ) d x = 0. \int_0^T\int_{\mathbb{R}^d}\rho^\varepsilon\left[\partial_t\varphi+b^\varepsilon\cdot\nabla\varphi+\frac{1}{2}\operatorname{tr}(a^\varepsilon\nabla^2\varphi)\right]\,\mathrm{d}x\,\mathrm{d}t+\int_{\mathbb{R}^d}\rho_0^\varepsilon(x)\varphi(0,x)\,\mathrm{d}x=0. ∫ 0 T ∫ R d ρ ε [ ∂ t φ + b ε ⋅ ∇ φ + 2 1 tr ( a ε ∇ 2 φ ) ] d x d t + ∫ R d ρ 0 ε ( x ) φ ( 0 , x ) d x = 0.
Using the local convergence of the mollified coefficients, the uniform L ∞ L^\infty L ∞ bound on ρ ε \rho^\varepsilon ρ ε , the weak-∗ * ∗ convergence of ρ ε \rho^\varepsilon ρ ε in L ∞ ( [ 0 , T ] × R d ) L^\infty([0,T]\times\mathbb{R}^d) L ∞ ([ 0 , T ] × R d ) , and the weak convergence in L l o c 1 ( [ 0 , T ] × R d ) L^1_{\mathrm{loc}}([0,T]\times\mathbb{R}^d) L loc 1 ([ 0 , T ] × R d ) , we may pass to the limit in each term. Since ρ 0 ε → ρ 0 \rho_0^\varepsilon\to\rho_0 ρ 0 ε → ρ 0 in L 1 ( R d ) L^1(\mathbb{R}^d) L 1 ( R d ) as ε → 0 + \varepsilon\to0^+ ε → 0 + , we obtain
∫ 0 T ∫ R d ρ [ ∂ t φ + b ⋅ ∇ φ + 1 2 tr ( a ∇ 2 φ ) ] d x d t + ∫ R d ρ 0 ( x ) φ ( 0 , x ) d x = 0. \int_0^T\int_{\mathbb{R}^d}\rho\left[\partial_t\varphi+b\cdot\nabla\varphi+\frac{1}{2}\operatorname{tr}(a\nabla^2\varphi)\right]\,\mathrm{d}x\,\mathrm{d}t+\int_{\mathbb{R}^d}\rho_0(x)\varphi(0,x)\,\mathrm{d}x=0. ∫ 0 T ∫ R d ρ [ ∂ t φ + b ⋅ ∇ φ + 2 1 tr ( a ∇ 2 φ ) ] d x d t + ∫ R d ρ 0 ( x ) φ ( 0 , x ) d x = 0.
By Remark 2.2, the measure-valued map ρ ( t , x ) d x \rho(t,x)\,\mathrm{d}x ρ ( t , x ) d x admits a unique weakly continuous representative with initial value ρ 0 ( x ) d x \rho_0(x)\,\mathrm{d}x ρ 0 ( x ) d x . This completes the construction of a weak solution to the FPE.
Remark 3.8. By Theorem 3.3 (1) and the superposition principle Theorem 2.5 , the FPE solution induces a weak solution to the SDE (1) , which also yields a construction of weak solutions to the associated degenerate SDE.
We now turn to the second part of Theorem 3.3 , the quantitative stability estimate for weak solutions. The proof relies on several auxiliary lemmas, which we state first.
Since our analysis concerns stability under rough coefficients, we use a standard mollification procedure. In particular, we shall use a lemma due to Crippa and Ligabue [7 ] . Under fractional Sobolev regularity, it provides quantitative estimates both for the convergence of a mollified function to the original one and for the blow-up rate of its derivatives.
Lemma 3.9. Let s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) and p ∈ [ 1 , ∞ ) p\in[1,\infty) p ∈ [ 1 , ∞ ) . Then, there exists a constant C > 0 C>0 C > 0 such that for any ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) and f ∈ W s , p ( R d ) f\in W^{s,p}(\mathbb{R}^d) f ∈ W s , p ( R d ) ,
∥ f − f ∗ ϕ ε ∥ L p ( R d ) ≤ C ε s ∥ f ∥ W s , p ( R d ) , ∥ ∇ ( f ∗ ϕ ε ) ∥ L p ( R d ) ≤ C ε s − 1 ∥ f ∥ W s , p ( R d ) , \|f-f*\phi^\varepsilon\|_{L^p(\mathbb{R}^d)}\le C\varepsilon^s\|f\|_{W^{s,p}(\mathbb{R}^d)},\qquad\|\nabla(f*\phi^\varepsilon)\|_{L^p(\mathbb{R}^d)}\le C\varepsilon^{s-1}\|f\|_{W^{s,p}(\mathbb{R}^d)}, ∥ f − f ∗ ϕ ε ∥ L p ( R d ) ≤ C ε s ∥ f ∥ W s , p ( R d ) , ∥∇ ( f ∗ ϕ ε ) ∥ L p ( R d ) ≤ C ε s − 1 ∥ f ∥ W s , p ( R d ) , where ϕ ε \phi^\varepsilon ϕ ε is given by (10) .
To control the error terms below, we shall use the Hardy–Littlewood maximal function. For f ∈ L l o c 1 ( R d ) f\in L^1_{\mathrm{loc}}(\mathbb{R}^d) f ∈ L loc 1 ( R d ) , set
M f ( x ) : = sup r > 0 1 L d ( B r ( x ) ) ∫ B r ( x ) ∣ f ( z ) ∣ d z , x ∈ R d . \mathcal{M}f(x):=\sup_{r>0}\frac{1}{\mathcal{L}^d(B_r(x))}\int_{B_r(x)}|f(z)|\,\mathrm{d}z,\qquad x\in\mathbb{R}^d. M f ( x ) := r > 0 sup L d ( B r ( x )) 1 ∫ B r ( x ) ∣ f ( z ) ∣ d z , x ∈ R d .
It is well known that M \mathcal{M} M is of weak type ( 1 , 1 ) (1,1) ( 1 , 1 ) , i.e., there exists a constant c d > 0 c_d>0 c d > 0 depending only on d d d such that
∥ M f ∥ L 1 , ∞ ( R d ) ≤ c d ∥ f ∥ L 1 ( R d ) , f ∈ L 1 ( R d ) . (20) \|\mathcal{M}f\|_{L^{1,\infty}(\mathbb{R}^d)}\le c_d\|f\|_{L^1(\mathbb{R}^d)},\qquad f\in L^1(\mathbb{R}^d). \tag*{(20)} ∥ M f ∥ L 1 , ∞ ( R d ) ≤ c d ∥ f ∥ L 1 ( R d ) , f ∈ L 1 ( R d ) . (20)
Moreover, for 1 < r ≤ ∞ 1<r\le\infty 1 < r ≤ ∞ , M \mathcal{M} M satisfies the strong ( r , r ) (r,r) ( r , r ) estimate, i.e., there exists a constant c d , r > 0 c_{d,r}>0 c d , r > 0 depending only on d , r d,r d , r such that
∥ M f ∥ L r ( R d ) ≤ c d , r ∥ f ∥ L r ( R d ) , f ∈ L r ( R d ) . \|\mathcal{M}f\|_{L^r(\mathbb{R}^d)}\le c_{d,r}\|f\|_{L^r(\mathbb{R}^d)},\qquad f\in L^r(\mathbb{R}^d). ∥ M f ∥ L r ( R d ) ≤ c d , r ∥ f ∥ L r ( R d ) , f ∈ L r ( R d ) .
We next recall a Morrey-type inequality, which bounds the increments of a Sobolev function in terms of the maximal function of its gradient; see [14 Appendix] or the more precise version [19 Lemma 3.1] for a proof.
Lemma 3.10. There exists a constant C d > 0 C_d>0 C d > 0 such that, for any f ∈ W l o c 1 , 1 ( R d ) f\in W_{\mathrm{loc}}^{1,1}(\mathbb{R}^d) f ∈ W loc 1 , 1 ( R d ) ,
∣ f ( x ) − f ( y ) ∣ ≤ C d ∣ x − y ∣ ( M ( ∇ f ) ( x ) + M ( ∇ f ) ( y ) ) , |f(x)-f(y)|\le C_d|x-y|\left(\mathcal{M}(\nabla f)(x)+\mathcal{M}(\nabla f)(y)\right), ∣ f ( x ) − f ( y ) ∣ ≤ C d ∣ x − y ∣ ( M ( ∇ f ) ( x ) + M ( ∇ f ) ( y ) ) , for a.e. x , y ∈ R d x,y\in\mathbb{R}^d x , y ∈ R d .
The following lemmas give a variant of this Morrey-type estimate, together with the corresponding L r L^r L r -estimate for the modified maximal function. These results were originally proved in [7 ] ; for completeness, we include a brief proof.
Lemma 3.11 (A variant maximal function estimate). There exists a constant C d > 0 C_d>0 C d > 0 such that, for every f ∈ W l o c 1 , 1 ( R d ) f\in W_{\mathrm{loc}}^{1,1}(\mathbb{R}^d) f ∈ W loc 1 , 1 ( R d ) and for a.e. x , y ∈ R d x,y\in\mathbb{R}^d x , y ∈ R d ,
∣ f ( x ) − f ( y ) ∣ ≤ C d ∣ B − 1 ( x − y ) ∣ ( M ~ ( ∇ f ) ( x ) + M ~ ( ∇ f ) ( y ) ) , |f(x)-f(y)|\le C_d|B^{-1}(x-y)|\left(\widetilde{\mathcal{M}}(\nabla f)(x)+\widetilde{\mathcal{M}}(\nabla f)(y)\right), ∣ f ( x ) − f ( y ) ∣ ≤ C d ∣ B − 1 ( x − y ) ∣ ( M ( ∇ f ) ( x ) + M ( ∇ f ) ( y ) ) , where the matrix B B B is given by (3.1), and
M ~ ( ∇ f ) ( x ) : = M ( ∑ i = 1 d ∣ ∂ i f ( B ⋅ ) ∣ ∣ B i i ∣ ) ( B − 1 x ) . \widetilde{\mathcal{M}}(\nabla f)(x):=\mathcal{M}\left(\sum_{i=1}^{d}|\partial_i f(B\cdot)||B_{ii}|\right)(B^{-1}x). M ( ∇ f ) ( x ) := M ( i = 1 ∑ d ∣ ∂ i f ( B ⋅ ) ∣∣ B ii ∣ ) ( B − 1 x ) . Proof. Set f ~ ( x ) : = f ( B x ) \widetilde{f}(x):=f(Bx) f ( x ) := f ( B x ) . By Lemma 3.10, for a.e. ξ , η ∈ R d \xi,\eta\in\mathbb{R}^d ξ , η ∈ R d ,
∣ f ~ ( ξ ) − f ~ ( η ) ∣ ≤ C d ∣ ξ − η ∣ ( M ( ∇ f ~ ) ( ξ ) + M ( ∇ f ~ ) ( η ) ) . |\widetilde{f}(\xi)-\widetilde{f}(\eta)|\le C_d|\xi-\eta|\left(\mathcal{M}(\nabla\widetilde{f})(\xi)+\mathcal{M}(\nabla\widetilde{f})(\eta)\right). ∣ f ( ξ ) − f ( η ) ∣ ≤ C d ∣ ξ − η ∣ ( M ( ∇ f ) ( ξ ) + M ( ∇ f ) ( η ) ) . Since
M ( ∇ f ~ ) ( ξ ) ≤ M ( ∑ i = 1 d ∣ ∂ i f ( B ⋅ ) ∣ ∣ B i i ∣ ) ( ξ ) , \mathcal{M}(\nabla\widetilde{f})(\xi)\le\mathcal{M}\left(\sum_{i=1}^{d}|\partial_i f(B\cdot)||B_{ii}|\right)(\xi), M ( ∇ f ) ( ξ ) ≤ M ( i = 1 ∑ d ∣ ∂ i f ( B ⋅ ) ∣∣ B ii ∣ ) ( ξ ) , we obtain, for a.e. ξ , η ∈ R d \xi,\eta\in\mathbb{R}^d ξ , η ∈ R d ,
∣ f ( B ξ ) − f ( B η ) ∣ ≤ C d ∣ ξ − η ∣ [ M ( ∑ i = 1 d ∣ ∂ i f ( B ⋅ ) ∣ ∣ B i i ∣ ) ( ξ ) + M ( ∑ i = 1 d ∣ ∂ i f ( B ⋅ ) ∣ ∣ B i i ∣ ) ( η ) ] . |f(B\xi)-f(B\eta)|\le C_d|\xi-\eta|\left[\mathcal{M}\left(\sum_{i=1}^{d}|\partial_i f(B\cdot)||B_{ii}|\right)(\xi)+\mathcal{M}\left(\sum_{i=1}^{d}|\partial_i f(B\cdot)||B_{ii}|\right)(\eta)\right]. ∣ f ( B ξ ) − f ( B η ) ∣ ≤ C d ∣ ξ − η ∣ [ M ( i = 1 ∑ d ∣ ∂ i f ( B ⋅ ) ∣∣ B ii ∣ ) ( ξ ) + M ( i = 1 ∑ d ∣ ∂ i f ( B ⋅ ) ∣∣ B ii ∣ ) ( η ) ] . Letting ξ = B − 1 x \xi=B^{-1}x ξ = B − 1 x and η = B − 1 y \eta=B^{-1}y η = B − 1 y completes the proof.
Lemma 3.12 (Operator bounds). Let M ~ ( ∇ f ) \widetilde{\mathcal{M}}(\nabla f) M ( ∇ f ) be defined as in Lemma 3.11. Then the following estimates hold.
If p = 1 p=1 p = 1 and ∣ ∇ f ∣ ∈ L 1 ( R d ) |\nabla f| \in L^{1}(\mathbb{R}^{d}) ∣∇ f ∣ ∈ L 1 ( R d ) , then
∥ M ~ ( ∇ f ) ∥ L 1 , ∞ ( R d ) ≲ δ 1 ∑ i = 1 d 1 ∥ ∂ i f ∥ L 1 ( R d ) + δ 2 ∑ i = d 1 + 1 d ∥ ∂ i f ∥ L 1 ( R d ) . \|\widetilde{\mathcal{M}}(\nabla f)\|_{L^{1,\infty}(\mathbb{R}^{d})} \lesssim\delta_{1}\sum_{i=1}^{d_{1}}\|\partial_{i}f\|_{L^{1}(\mathbb{R}^{d})}+\delta_{2}\sum_{i=d_{1}+1}^{d}\|\partial_{i}f\|_{L^{1}(\mathbb{R}^{d})}. ∥ M ( ∇ f ) ∥ L 1 , ∞ ( R d ) ≲ δ 1 i = 1 ∑ d 1 ∥ ∂ i f ∥ L 1 ( R d ) + δ 2 i = d 1 + 1 ∑ d ∥ ∂ i f ∥ L 1 ( R d ) .
If 1 < p ≤ ∞ 1<p\leq\infty 1 < p ≤ ∞ and ∣ ∇ f ∣ ∈ L p ( R d ) |\nabla f| \in L^{p}(\mathbb{R}^{d}) ∣∇ f ∣ ∈ L p ( R d ) , then
∥ M ~ ( ∇ f ) ∥ L p ( R d ) ≲ δ 1 ∑ i = 1 d 1 ∥ ∂ i f ∥ L p ( R d ) + δ 2 ∑ i = d 1 + 1 d ∥ ∂ i f ∥ L p ( R d ) . \|\widetilde{\mathcal{M}}(\nabla f)\|_{L^{p}(\mathbb{R}^{d})} \lesssim\delta_{1}\sum_{i=1}^{d_{1}}\|\partial_{i}f\|_{L^{p}(\mathbb{R}^{d})}+\delta_{2}\sum_{i=d_{1}+1}^{d}\|\partial_{i}f\|_{L^{p}(\mathbb{R}^{d})}. ∥ M ( ∇ f ) ∥ L p ( R d ) ≲ δ 1 i = 1 ∑ d 1 ∥ ∂ i f ∥ L p ( R d ) + δ 2 i = d 1 + 1 ∑ d ∥ ∂ i f ∥ L p ( R d ) .
Proof. As in Lemma 3.11, consider the rescaled function f ~ ( x ) = f ( B x ) \widetilde{f}(x)=f(Bx) f ( x ) = f ( B x ) . Adapting the argument behind (20) , we obtain
∥ M ~ ( ∇ f ) ∥ L 1 , ∞ ( R d ) = ∥ M ( ∑ i = 1 d ∣ ∂ i f ~ ∣ ( B − 1 ⋅ ) ) ∥ L 1 , ∞ ( R d ) ≲ ∥ ∑ i = 1 d ∣ ∂ i f ~ ∣ ( B − 1 ⋅ ) ∥ L 1 ( R d ) ≲ ∑ i = 1 d ∥ ( ∂ i f ( B ⋅ ) B i i ) ( B − 1 ⋅ ) ∥ L 1 ( R d ) = ∑ i = 1 d B i i ∥ ∂ i f ∥ L 1 ( R d ) . \begin{aligned}
\|\widetilde{\mathcal{M}}(\nabla f)\|_{L^{1,\infty}(\mathbb{R}^{d})}
&=\left\|\mathcal{M}\left(\sum_{i=1}^{d}|\partial_{i}\widetilde{f}|(B^{-1}\cdot)\right)\right\|_{L^{1,\infty}(\mathbb{R}^{d})}\\
&\lesssim\left\|\sum_{i=1}^{d}|\partial_{i}\widetilde{f}|(B^{-1}\cdot)\right\|_{L^{1}(\mathbb{R}^{d})}\\
&\lesssim\sum_{i=1}^{d}\|(\partial_{i}f(B\cdot)B_{ii})(B^{-1}\cdot)\|_{L^{1}(\mathbb{R}^{d})}\\
&=\sum_{i=1}^{d}B_{ii}\|\partial_{i}f\|_{L^{1}(\mathbb{R}^{d})}.
\end{aligned} ∥ M ( ∇ f ) ∥ L 1 , ∞ ( R d ) = M ( i = 1 ∑ d ∣ ∂ i f ∣ ( B − 1 ⋅ ) ) L 1 , ∞ ( R d ) ≲ i = 1 ∑ d ∣ ∂ i f ∣ ( B − 1 ⋅ ) L 1 ( R d ) ≲ i = 1 ∑ d ∥ ( ∂ i f ( B ⋅ ) B ii ) ( B − 1 ⋅ ) ∥ L 1 ( R d ) = i = 1 ∑ d B ii ∥ ∂ i f ∥ L 1 ( R d ) .
The proof of conclusion 2 is similar and is omitted.
With these preparations, we now prove the second part of Theorem 3.3 .
Proof of Theorem 3.3(2). To circumvent the non-differentiability of
x ↦ log ( 1 + ∣ B − 1 x ∣ 2 ) x\mapsto\log\left(1+\sqrt{|B^{-1}x|^{2}}\right) x ↦ log ( 1 + ∣ B − 1 x ∣ 2 )
at x = 0 x=0 x = 0 , we use a standard regularization argument. For α > 0 \alpha>0 α > 0 , introduce the smooth approximation
x ↦ log ( 1 + ∣ B − 1 x ∣ 2 + α ) , x\mapsto\log\left(1+\sqrt{|B^{-1}x|^{2}+\alpha}\right), x ↦ log ( 1 + ∣ B − 1 x ∣ 2 + α ) ,
which is smooth on R d \mathbb{R}^{d} R d and converges pointwise to log ( 1 + ∣ B − 1 x ∣ 2 ) \log\left(1+\sqrt{|B^{-1}x|^{2}}\right) log ( 1 + ∣ B − 1 x ∣ 2 ) as α → 0 \alpha\to0 α → 0 .
By Theorem 3.3 (1), the weak solutions ρ \rho ρ and ρ ^ \hat{\rho} ρ ^ to the FPE (1.2) with corresponding coefficients exist. By the superposition principle (Theorem 2.5 ), each of them gives rise to a weak solution of the SDE (1.1), although they may be defined on different probability spaces with different Brownian motions. Applying Lemma 2.6 to these two weak solutions, we may assume, without loss of generality, that they are realized on a common filtered probability space and driven by the same Brownian motion ( W t ) t ≥ 0 (W_t)_{t\geq0} ( W t ) t ≥ 0 . Moreover, by the same lemma, the initial coupling can be chosen arbitrarily, and we take it to be optimal with respect to D δ 1 , δ 2 \mathcal{D}_{\delta_1,\delta_2} D δ 1 , δ 2 . With these preparations, we consider the coupled SDEs:
d X t = b ( t , X t ) d t + σ ( t , X t ) d W t , L X 0 = ν , \mathrm{d}X_t=b(t,X_t)\,\mathrm{d}t+\sigma(t,X_t)\,\mathrm{d}W_t,\qquad\mathcal{L}_{X_0}=\nu, d X t = b ( t , X t ) d t + σ ( t , X t ) d W t , L X 0 = ν ,
and
d X ^ t = b ^ ( t , X ^ t ) d t + σ ^ ( t , X ^ t ) d W t , L X ^ 0 = ν ^ , \mathrm{d}\widehat{X}_t=\widehat{b}(t,\widehat{X}_t)\,\mathrm{d}t+\widehat{\sigma}(t,\widehat{X}_t)\,\mathrm{d}W_t,\qquad\mathcal{L}_{\widehat{X}_0}=\widehat{\nu}, d X t = b ( t , X t ) d t + σ ( t , X t ) d W t , L X 0 = ν ,
with ν = ρ 0 d x \nu=\rho_0\mathrm{d}x ν = ρ 0 d x and ν ^ = ρ ^ 0 d x \widehat{\nu}=\widehat{\rho}_0\mathrm{d}x ν = ρ 0 d x . Moreover, the superposition principle gives the time marginals
L X t ( d x ) = ρ ( t , x ) d x , L X ^ t ( d x ) = ρ ^ ( t , x ) d x , t ∈ ( 0 , T ) . \mathcal{L}_{X_t}(\mathrm{d}x)=\rho(t,x)\mathrm{d}x,\qquad\mathcal{L}_{\widehat{X}_t}(\mathrm{d}x)=\widehat{\rho}(t,x)\mathrm{d}x,\qquad t\in(0,T). L X t ( d x ) = ρ ( t , x ) d x , L X t ( d x ) = ρ ( t , x ) d x , t ∈ ( 0 , T ) .
We also choose the initial values ( X 0 , X ^ 0 ) (X_0,\widehat{X}_0) ( X 0 , X 0 ) to be optimally coupled, namely
E log ( 1 + ∣ B − 1 ( X 0 − X ^ 0 ) ∣ 2 ) = D δ 1 , δ 2 ( ν , ν ^ ) . (21) \mathbb{E}\log\left(1+\sqrt{\lvert B^{-1}(X_0-\widehat{X}_0)\rvert^2}\right)=\mathcal{D}_{\delta_1,\delta_2}(\nu,\widehat{\nu}). \tag*{(21)} E log ( 1 + ∣ B − 1 ( X 0 − X 0 ) ∣ 2 ) = D δ 1 , δ 2 ( ν , ν ) . (21)
Recall that
τ R : = inf { t ∈ [ 0 , T ] : ∣ X t ∣ ∨ ∣ X ^ t ∣ ≥ R } , R > 0. \tau_R:=\inf\left\{t\in[0,T]:\lvert X_t\rvert\vee\lvert\widehat{X}_t\rvert\geq R\right\},\qquad R>0. τ R := inf { t ∈ [ 0 , T ] : ∣ X t ∣ ∨ ∣ X t ∣ ≥ R } , R > 0.
By Itô’s formula applied to the above SDEs, we obtain
E log ( 1 + ∣ B − 1 ( X t ∧ τ R − X ^ t ∧ τ R ) ∣ 2 + α ) ≤ E log ( 1 + ∣ B − 1 ( X 0 − X ^ 0 ) ∣ 2 + α ) + E ( ∫ 0 t ∧ τ R ∣ B − 1 ( b ( s , X s ) − b ^ ( s , X ^ s ) ) ∣ 1 + ∣ B − 1 ( X s − X ^ s ) ∣ d s ) + 1 2 E ( ∫ 0 t ∧ τ R ∥ B − 1 ( σ ( s , X s ) − σ ^ ( s , X ^ s ) ) ∥ 2 ∣ B − 1 ( X s − X ^ s ) ∣ 2 + α + ∣ B − 1 ( X s − X ^ s ) ∣ 2 + α d s ) . (22) \begin{aligned}
\mathbb{E}\log\left(1+\sqrt{\lvert B^{-1}(X_{t\wedge\tau_R}-\widehat{X}_{t\wedge\tau_R})\rvert^2+\alpha}\right)
&\leq\mathbb{E}\log\left(1+\sqrt{\lvert B^{-1}(X_0-\widehat{X}_0)\rvert^2+\alpha}\right)\\
&\quad+\mathbb{E}\left(\int_0^{t\wedge\tau_R}\frac{\lvert B^{-1}(b(s,X_s)-\widehat{b}(s,\widehat{X}_s))\rvert}{1+\lvert B^{-1}(X_s-\widehat{X}_s)\rvert}\,\mathrm{d}s\right)\\
&\quad+\frac{1}{2}\mathbb{E}\left(\int_0^{t\wedge\tau_R}\frac{\lVert B^{-1}(\sigma(s,X_s)-\widehat{\sigma}(s,\widehat{X}_s))\rVert^2}{\sqrt{\lvert B^{-1}(X_s-\widehat{X}_s)\rvert^2+\alpha}+\lvert B^{-1}(X_s-\widehat{X}_s)\rvert^2+\alpha}\,\mathrm{d}s\right). \tag*{(22)}
\end{aligned} E log ( 1 + ∣ B − 1 ( X t ∧ τ R − X t ∧ τ R ) ∣ 2 + α ) ≤ E log ( 1 + ∣ B − 1 ( X 0 − X 0 ) ∣ 2 + α ) + E ( ∫ 0 t ∧ τ R 1 + ∣ B − 1 ( X s − X s )∣ ∣ B − 1 ( b ( s , X s ) − b ( s , X s ))∣ d s ) + 2 1 E ∫ 0 t ∧ τ R ∣ B − 1 ( X s − X s ) ∣ 2 + α + ∣ B − 1 ( X s − X s ) ∣ 2 + α ∥ B − 1 ( σ ( s , X s ) − σ ( s , X s )) ∥ 2 d s . (22)
We set
I 1 : = E ( ∫ 0 t ∧ τ R ∣ B − 1 ( b ( s , X s ) − b ^ ( s , X ^ s ) ) ∣ 1 + ∣ B − 1 ( X s − X ^ s ) ∣ d s ) , I_1:=\mathbb{E}\left(\int_0^{t\wedge\tau_R}\frac{\lvert B^{-1}(b(s,X_s)-\widehat{b}(s,\widehat{X}_s))\rvert}{1+\lvert B^{-1}(X_s-\widehat{X}_s)\rvert}\,\mathrm{d}s\right), I 1 := E ( ∫ 0 t ∧ τ R 1 + ∣ B − 1 ( X s − X s )∣ ∣ B − 1 ( b ( s , X s ) − b ( s , X s ))∣ d s ) ,
and
I 2 : = E ( ∫ 0 t ∧ τ R ∥ B − 1 ( σ ( s , X s ) − σ ^ ( s , X ^ s ) ) ∥ 2 ∣ B − 1 ( X s − X ^ s ) ∣ 2 + α + ∣ B − 1 ( X s − X ^ s ) ∣ 2 + α d s ) . I_2:=\mathbb{E}\left(\int_0^{t\wedge\tau_R}\frac{\lVert B^{-1}(\sigma(s,X_s)-\widehat{\sigma}(s,\widehat{X}_s))\rVert^2}{\sqrt{\lvert B^{-1}(X_s-\widehat{X}_s)\rvert^2+\alpha}+\lvert B^{-1}(X_s-\widehat{X}_s)\rvert^2+\alpha}\,\mathrm{d}s\right). I 2 := E ∫ 0 t ∧ τ R ∣ B − 1 ( X s − X s ) ∣ 2 + α + ∣ B − 1 ( X s − X s ) ∣ 2 + α ∥ B − 1 ( σ ( s , X s ) − σ ( s , X s )) ∥ 2 d s .
It suffices to estimate I 1 I_1 I 1 and I 2 I_2 I 2 .
Estimate for I 1 I_1 I 1 . We split the estimate of I 1 I_1 I 1 into four steps.
Step 1: Mollification of the drift coefficient b b b . Denote by b 2 ϵ b_{2}^{\epsilon} b 2 ϵ the regularization of b 2 b_{2} b 2 in the x 1 x_{1} x 1 -variable,
b 2 ϵ ( t , x 1 , x 2 ) : = b 2 ( t , ⋅ , x 2 ) ∗ 1 ϕ ϵ ( x 1 ) , t ∈ [ 0 , T ] , x 1 ∈ R d 1 , x 2 ∈ R d 2 , b_{2}^{\epsilon}(t,x_{1},x_{2}) := b_{2}(t,\cdot,x_{2}) *_1 \phi^{\epsilon}(x_{1}), \qquad t \in[0,T],\ x_{1} \in\mathbb{R}^{d_{1}},\ x_{2} \in\mathbb{R}^{d_{2}}, b 2 ϵ ( t , x 1 , x 2 ) := b 2 ( t , ⋅ , x 2 ) ∗ 1 ϕ ϵ ( x 1 ) , t ∈ [ 0 , T ] , x 1 ∈ R d 1 , x 2 ∈ R d 2 ,
where ∗ 1 *_1 ∗ 1 denotes convolution in the x 1 x_{1} x 1 -variable only, with ϕ ϵ \phi^{\epsilon} ϕ ϵ understood as a standard mollifier on R d 1 \mathbb{R}^{d_{1}} R d 1 , and set
b ϵ : = ( b 1 , b 2 ϵ ) . b^{\epsilon} := (b_{1},b_{2}^{\epsilon}). b ϵ := ( b 1 , b 2 ϵ ) .
Step 2: Intermediate estimate via mollification. Let Ω T , R : = [ 0 , T ] × B R × B R \Omega_{T,R} := [0,T] \times B_{R} \times B_{R} Ω T , R := [ 0 , T ] × B R × B R . Introducing the mollified drift as an intermediate term, we obtain
I 1 ≤ ∫ Ω T , R ∣ B − 1 ( b ( s , x ) − b ^ ( s , y ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ π s ( d x , d y ) d s ≤ ∫ Ω T , R { ∣ B − 1 ( b ( s , x ) − b ϵ ( s , x ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ + ∣ B − 1 ( b ϵ ( s , x ) − b ^ ϵ ( s , y ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ } π s ( d x , d y ) d s \begin{aligned}
I_{1} \le\int_{\Omega_{T,R}} \frac{\left|B^{-1}(b(s,x)-\widehat{b}(s,y))\right|}{1+\left|B^{-1}(x-y)\right|}\,\pi_{s}(dx,dy)\,ds \\
&\le\int_{\Omega_{T,R}} \left\{\frac{\left|B^{-1}(b(s,x)-b^{\epsilon}(s,x))\right|}{1+\left|B^{-1}(x-y)\right|}+\frac{\left|B^{-1}(b^{\epsilon}(s,x)-\widehat{b}^{\epsilon}(s,y))\right|}{1+\left|B^{-1}(x-y)\right|}\right\}\pi_{s}(dx,dy)\,ds
\end{aligned} I 1 ≤ ∫ Ω T , R 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ( s , x ) − b ( s , y )) π s ( d x , d y ) d s ≤ ∫ Ω T , R ⎩ ⎨ ⎧ 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ( s , x ) − b ϵ ( s , x )) + 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y )) ⎭ ⎬ ⎫ π s ( d x , d y ) d s
+ ∫ Ω T , R ∣ B − 1 ( b ^ ϵ ( s , y ) − b ^ ( s , y ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ π s ( d x , d y ) d s \quad+\int_{\Omega_{T,R}}\frac{\left|B^{-1}(\widehat{b}^{\epsilon}(s,y)-\widehat{b}(s,y))\right|}{1+\left|B^{-1}(x-y)\right|}\,\pi_{s}(dx,dy)\,ds + ∫ Ω T , R 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ϵ ( s , y ) − b ( s , y )) π s ( d x , d y ) d s
≲ 1 δ 2 ∥ b 2 − b 2 ϵ ∥ L t 1 L x 1 + 1 δ 2 ∥ b ^ 2 − b ^ 2 ϵ ∥ L t 1 L x 1 + ∫ Ω T , R ∣ B − 1 ( b ϵ ( s , x ) − b ^ ϵ ( s , y ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ π s ( d x , d y ) d s , \lesssim\frac{1}{\delta_{2}}\left\|b_{2}-b_{2}^{\epsilon}\right\|_{L_{t}^{1}L_{x}^{1}}+\frac{1}{\delta_{2}}\left\|\widehat{b}_{2}-\widehat{b}_{2}^{\epsilon}\right\|_{L_{t}^{1}L_{x}^{1}}+\int_{\Omega_{T,R}}\frac{\left|B^{-1}(b^{\epsilon}(s,x)-\widehat{b}^{\epsilon}(s,y))\right|}{1+\left|B^{-1}(x-y)\right|}\,\pi_{s}(dx,dy)\,ds, ≲ δ 2 1 ∥ b 2 − b 2 ϵ ∥ L t 1 L x 1 + δ 2 1 b 2 − b 2 ϵ L t 1 L x 1 + ∫ Ω T , R 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y )) π s ( d x , d y ) d s ,
where π s = L ( X s , X ^ s ) \pi_{s}=\mathcal{L}(X_{s},\widehat{X}_{s}) π s = L ( X s , X s ) stands for the joint distribution of ( X s , X ^ s ) (X_{s},\widehat{X}_{s}) ( X s , X s ) for any s ∈ [ 0 , T ] s \in[0,T] s ∈ [ 0 , T ] . The last inequality uses the boundedness of the densities ρ \rho ρ and ρ ^ \widehat{\rho} ρ .
Set
I 1 , 1 : = ∫ Ω T , R ∣ B − 1 ( b ϵ ( s , x ) − b ^ ϵ ( s , y ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ π s ( d x , d y ) d s . I_{1,1}:=\int_{\Omega_{T,R}}\frac{\left|B^{-1}(b^{\epsilon}(s,x)-\widehat{b}^{\epsilon}(s,y))\right|}{1+\left|B^{-1}(x-y)\right|}\,\pi_{s}(dx,dy)\,ds. I 1 , 1 := ∫ Ω T , R 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y )) π s ( d x , d y ) d s .
Then, for any 0 < δ 1 < δ 2 0<\delta_{1}<\delta_{2} 0 < δ 1 < δ 2 ,
I 1 , 1 ≤ ∫ Ω T , R { ∣ B − 1 ( b ϵ ( s , y ) − b ^ ϵ ( s , y ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ + ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ } π s ( d x , d y ) d s ≤ 1 δ 1 ∫ Ω T , R ∣ b 1 ( s , y 1 ) − b ^ 1 ( s , y 1 ) ∣ ρ ^ s ( y 1 , y 2 ) d y 1 d y 2 d s + 1 δ 2 ∥ b 2 − b ^ 2 ∥ L t 1 L x 1 \begin{aligned}
I_{1,1}\le\int_{\Omega_{T,R}}\left\{\frac{\left|B^{-1}(b^{\epsilon}(s,y)-\widehat{b}^{\epsilon}(s,y))\right|}{1+\left|B^{-1}(x-y)\right|}+\frac{\left|B^{-1}(b^{\epsilon}(s,x)-b^{\epsilon}(s,y))\right|}{1+\left|B^{-1}(x-y)\right|}\right\}\pi_{s}(dx,dy)\,ds \\
&\le\frac{1}{\delta_{1}}\int_{\Omega_{T,R}}|b_{1}(s,y_{1})-\widehat{b}_{1}(s,y_{1})|\widehat{\rho}_{s}(y_{1},y_{2})\,dy_{1}\,dy_{2}\,ds+\frac{1}{\delta_{2}}\left\|b_{2}-\widehat{b}_{2}\right\|_{L_{t}^{1}L_{x}^{1}}
\end{aligned} I 1 , 1 ≤ ∫ Ω T , R ⎩ ⎨ ⎧ 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ϵ ( s , y ) − b ϵ ( s , y )) + 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y )) ⎭ ⎬ ⎫ π s ( d x , d y ) d s ≤ δ 1 1 ∫ Ω T , R ∣ b 1 ( s , y 1 ) − b 1 ( s , y 1 ) ∣ ρ s ( y 1 , y 2 ) d y 1 d y 2 d s + δ 2 1 b 2 − b 2 L t 1 L x 1
+ ∫ Ω T , R ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ π s ( d x , d y ) d s \quad+\int_{\Omega_{T,R}}\frac{\left|B^{-1}(b^{\epsilon}(s,x)-b^{\epsilon}(s,y))\right|}{1+\left|B^{-1}(x-y)\right|}\,\pi_{s}(dx,dy)\,ds + ∫ Ω T , R 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y )) π s ( d x , d y ) d s
≲ ∫ Ω T , R ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ π s ( d x , d y ) d s + 1 δ 1 ∥ b 1 − b ^ 1 ∥ L 1 ( [ 0 , T ] ; L p ( B R ) ) \lesssim\int_{\Omega_{T,R}}\frac{\left|B^{-1}(b^{\epsilon}(s,x)-b^{\epsilon}(s,y))\right|}{1+\left|B^{-1}(x-y)\right|}\,\pi_{s}(dx,dy)\,ds+\frac{1}{\delta_{1}}\left\|b_{1}-\widehat{b}_{1}\right\|_{L^{1}([0,T];L^{p}(B_{R}))} ≲ ∫ Ω T , R 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y )) π s ( d x , d y ) d s + δ 1 1 b 1 − b 1 L 1 ([ 0 , T ] ; L p ( B R ))
+ 1 δ 1 ∥ b 2 − b ^ 2 ∥ L t 1 L x 1 , \quad+\frac{1}{\delta_{1}}\left\|b_{2}-\widehat{b}_{2}\right\|_{L_{t}^{1}L_{x}^{1}}, + δ 1 1 b 2 − b 2 L t 1 L x 1 ,
where in the last inequality we used 0 < 1 / δ 2 < 1 / δ 1 0<1/\delta_{2}<1/\delta_{1} 0 < 1/ δ 2 < 1/ δ 1 and the boundedness of the densities ρ \rho ρ and ρ ^ \widehat{\rho} ρ . It remains to estimate
I 1 , 1 , 1 : = ∫ Ω T , R ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y ) ) ∣ 1 + ∣ B − 1 ( x − y ) ∣ π s ( d x , d y ) d s . I_{1,1,1}:=\int_{\Omega_{T,R}}\frac{\left|B^{-1}(b^{\epsilon}(s,x)-b^{\epsilon}(s,y))\right|}{1+\left|B^{-1}(x-y)\right|}\,\pi_{s}(dx,dy)\,ds. I 1 , 1 , 1 := ∫ Ω T , R 1 + ∣ B − 1 ( x − y ) ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y )) π s ( d x , d y ) d s .
Applying Lemma 3.11 to the components b 1 b_1 b 1 and b 2 ϵ b_2^\epsilon b 2 ϵ , together with the subadditivity of M ~ \widetilde{\mathcal{M}} M , we obtain, for a.e. x , y ∈ R d x,y \in\mathbb{R}^d x , y ∈ R d ,
∣ b 1 ( t , x ) − b 1 ( t , y ) ∣ ∣ B − 1 ( x − y ) ∣ ≲ M ~ ( ∇ x 1 b 1 ) ( t , x ) + M ~ ( ∇ x 1 b 1 ) ( t , y ) , \frac{\lvert b_1(t,x)-b_1(t,y)\rvert}{\lvert B^{-1}(x-y)\rvert}
\lesssim\widetilde{\mathcal{M}}(\nabla_{x_1}b_1)(t,x)+\widetilde{\mathcal{M}}(\nabla_{x_1}b_1)(t,y), ∣ B − 1 ( x − y )∣ ∣ b 1 ( t , x ) − b 1 ( t , y )∣ ≲ M ( ∇ x 1 b 1 ) ( t , x ) + M ( ∇ x 1 b 1 ) ( t , y ) ,
and
∣ b 2 ϵ ( t , x ) − b 2 ϵ ( t , y ) ∣ ∣ B − 1 ( x − y ) ∣ ≲ M ~ ( ∇ b 2 ϵ ) ( t , x ) + M ~ ( ∇ b 2 ϵ ) ( t , y ) ≲ M ~ ( ∇ x 1 b 2 ϵ ) ( t , x ) + M ~ ( ∇ x 1 b 2 ϵ ) ( t , y ) + M ~ ( ∇ x 2 b 2 ϵ ) ( t , x ) + M ~ ( ∇ x 2 b 2 ϵ ) ( t , y ) . \begin{aligned}
\frac{\lvert b_2^\epsilon(t,x)-b_2^\epsilon(t,y)\rvert}{\lvert B^{-1}(x-y)\rvert}
&\lesssim\widetilde{\mathcal{M}}(\nabla b_2^\epsilon)(t,x)+\widetilde{\mathcal{M}}(\nabla b_2^\epsilon)(t,y)\\
&\lesssim\widetilde{\mathcal{M}}(\nabla_{x_1}b_2^\epsilon)(t,x)+\widetilde{\mathcal{M}}(\nabla_{x_1}b_2^\epsilon)(t,y)\\
&\quad+\widetilde{\mathcal{M}}(\nabla_{x_2}b_2^\epsilon)(t,x)+\widetilde{\mathcal{M}}(\nabla_{x_2}b_2^\epsilon)(t,y).
\end{aligned} ∣ B − 1 ( x − y )∣ ∣ b 2 ϵ ( t , x ) − b 2 ϵ ( t , y )∣ ≲ M ( ∇ b 2 ϵ ) ( t , x ) + M ( ∇ b 2 ϵ ) ( t , y ) ≲ M ( ∇ x 1 b 2 ϵ ) ( t , x ) + M ( ∇ x 1 b 2 ϵ ) ( t , y ) + M ( ∇ x 2 b 2 ϵ ) ( t , x ) + M ( ∇ x 2 b 2 ϵ ) ( t , y ) .
Since the marginals of π t \pi_t π t are absolutely continuous with respect to the Lebesgue measure, these pointwise estimates hold for a.e. t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] and π t \pi_t π t -a.e. ( x , y ) ∈ R d × R d (x,y) \in\mathbb{R}^d \times\mathbb{R}^d ( x , y ) ∈ R d × R d . Therefore
I 1 , 1 , 1 ≤ ∫ Ω T , R ψ 1 π s ( d x , d y ) d s + ∫ Ω T , R ψ 2 π s ( d x , d y ) d s + ∫ Ω T , R ψ 3 π s ( d x , d y ) d s , I_{1,1,1}\leq\int_{\Omega_{T,R}}\psi_1\,\pi_s(dx,dy)\,ds+\int_{\Omega_{T,R}}\psi_2\,\pi_s(dx,dy)\,ds+\int_{\Omega_{T,R}}\psi_3\,\pi_s(dx,dy)\,ds, I 1 , 1 , 1 ≤ ∫ Ω T , R ψ 1 π s ( d x , d y ) d s + ∫ Ω T , R ψ 2 π s ( d x , d y ) d s + ∫ Ω T , R ψ 3 π s ( d x , d y ) d s ,
where
ψ 1 ( s , x , y ) : = min { ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y ) ) ∣ , 1 δ 1 ( M ~ ( ∇ x 1 b 1 ) ( s , x ) + M ~ ( ∇ x 1 b 1 ) ( s , y ) ) } , \psi_1(s,x,y):=\min\left\{\lvert B^{-1}(b^\epsilon(s,x)-b^\epsilon(s,y))\rvert,\frac{1}{\delta_1}\left(\widetilde{\mathcal{M}}(\nabla_{x_1}b_1)(s,x)+\widetilde{\mathcal{M}}(\nabla_{x_1}b_1)(s,y)\right)\right\}, ψ 1 ( s , x , y ) := min { ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y ))∣ , δ 1 1 ( M ( ∇ x 1 b 1 ) ( s , x ) + M ( ∇ x 1 b 1 ) ( s , y ) ) } ,
ψ 2 ( s , x , y ) : = min { ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y ) ) ∣ , 1 δ 2 ( M ~ ( ∇ x 1 b 2 ϵ ) ( s , x ) + M ~ ( ∇ x 1 b 2 ϵ ) ( s , y ) ) } , \psi_2(s,x,y):=\min\left\{\lvert B^{-1}(b^\epsilon(s,x)-b^\epsilon(s,y))\rvert,\frac{1}{\delta_2}\left(\widetilde{\mathcal{M}}(\nabla_{x_1}b_2^\epsilon)(s,x)+\widetilde{\mathcal{M}}(\nabla_{x_1}b_2^\epsilon)(s,y)\right)\right\}, ψ 2 ( s , x , y ) := min { ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y ))∣ , δ 2 1 ( M ( ∇ x 1 b 2 ϵ ) ( s , x ) + M ( ∇ x 1 b 2 ϵ ) ( s , y ) ) } ,
and
ψ 3 ( s , x , y ) : = min { ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y ) ) ∣ , 1 δ 2 ( M ~ ( ∇ x 2 b 2 ϵ ) ( s , x ) + M ~ ( ∇ x 2 b 2 ϵ ) ( s , y ) ) } . \psi_3(s,x,y):=\min\left\{\lvert B^{-1}(b^\epsilon(s,x)-b^\epsilon(s,y))\rvert,\frac{1}{\delta_2}\left(\widetilde{\mathcal{M}}(\nabla_{x_2}b_2^\epsilon)(s,x)+\widetilde{\mathcal{M}}(\nabla_{x_2}b_2^\epsilon)(s,y)\right)\right\}. ψ 3 ( s , x , y ) := min { ∣ B − 1 ( b ϵ ( s , x ) − b ϵ ( s , y ))∣ , δ 2 1 ( M ( ∇ x 2 b 2 ϵ ) ( s , x ) + M ( ∇ x 2 b 2 ϵ ) ( s , y ) ) } .
Step 3: An estimate for the maximal function. Define a finite measure μ \mu μ on Ω T : = [ 0 , T ] × R d × R d \Omega_T:=[0,T]\times\mathbb{R}^d\times\mathbb{R}^d Ω T := [ 0 , T ] × R d × R d by
μ ( d t , d x , d y ) : = 1 [ 0 , T ] ( t ) 1 B R × B R ( x , y ) π t ( d x , d y ) d t . \mu(dt,dx,dy):=\mathbb{1}_{[0,T]}(t)\mathbb{1}_{B_R\times B_R}(x,y)\pi_t(dx,dy)dt. μ ( d t , d x , d y ) := 1 [ 0 , T ] ( t ) 1 B R × B R ( x , y ) π t ( d x , d y ) d t .
Using the second term in the minimum and the operator bounds in Lemma 3.12, since p ∈ ( 1 , ∞ ] p\in(1,\infty] p ∈ ( 1 , ∞ ] , we deduce
∥ ψ 1 ∥ L 1 ( Ω T , μ ) ≲ 1 δ 1 ∥ M ~ ( ∇ x 1 b 1 ) ∥ L 1 ( Ω T , μ ) ≲ 1 δ 1 ∫ 0 T ∫ B R d 1 M ~ ( ∇ x 1 b 1 ) ( t , x 1 ) d x 1 d t ≲ 1 δ 1 ∫ 0 T ( ∫ B R d 1 M ~ ( ∇ x 1 b 1 ) ( t , x 1 ) p d x 1 ) 1 p d t ≲ 1 δ 1 ∥ M ~ ( ∇ x 1 b 1 ) ∥ L t 1 L x 1 p ≲ ∥ ∇ x 1 b 1 ∥ L t 1 L x 1 p . \begin{aligned}
\|\psi_1\|_{L^1(\Omega_{T,\mu})}
&\lesssim\frac{1}{\delta_1}\|\widetilde{\mathcal{M}}(\nabla_{x_1}b_1)\|_{L^1(\Omega_{T,\mu})}\\
&\lesssim\frac{1}{\delta_1}\int_0^T\int_{B_R^{d_1}}\widetilde{\mathcal{M}}(\nabla_{x_1}b_1)(t,x_1)\,dx_1\,dt\\
&\lesssim\frac{1}{\delta_1}\int_0^T\left(\int_{B_R^{d_1}}\widetilde{\mathcal{M}}(\nabla_{x_1}b_1)(t,x_1)^p\,dx_1\right)^{\frac{1}{p}}dt\\
&\lesssim\frac{1}{\delta_1}\|\widetilde{\mathcal{M}}(\nabla_{x_1}b_1)\|_{L_t^1L_{x_1}^p}\\
&\lesssim\|\nabla_{x_1}b_1\|_{L_t^1L_{x_1}^p}.
\end{aligned} ∥ ψ 1 ∥ L 1 ( Ω T , μ ) ≲ δ 1 1 ∥ M ( ∇ x 1 b 1 ) ∥ L 1 ( Ω T , μ ) ≲ δ 1 1 ∫ 0 T ∫ B R d 1 M ( ∇ x 1 b 1 ) ( t , x 1 ) d x 1 d t ≲ δ 1 1 ∫ 0 T ( ∫ B R d 1 M ( ∇ x 1 b 1 ) ( t , x 1 ) p d x 1 ) p 1 d t ≲ δ 1 1 ∥ M ( ∇ x 1 b 1 ) ∥ L t 1 L x 1 p ≲ ∥ ∇ x 1 b 1 ∥ L t 1 L x 1 p .
Proceeding similarly,
∥ ψ 3 ∥ L 1 ( Ω T , μ ) ≲ ∥ ∇ x 2 b 2 ∥ L 1 ( [ 0 , T ] × R d 1 ; L x 2 p ) . \|\psi_3\|_{L^1(\Omega_T,\mu)} \lesssim\|\nabla_{x_2} b_2\|_{L^1([0,T]\times\mathbb{R}^{d_1};L^p_{x_2})}. ∥ ψ 3 ∥ L 1 ( Ω T , μ ) ≲ ∥ ∇ x 2 b 2 ∥ L 1 ([ 0 , T ] × R d 1 ; L x 2 p ) .
The estimate of ψ 2 \psi_2 ψ 2 requires more care, due to the lack of regularity of b 2 b_2 b 2 in the x 1 x_1 x 1 -variable. On the one hand, Lemma 3.9, the weak (1, 1) bound for M ~ \widetilde{\mathcal{M}} M , and the uniform boundedness of the densities give
∥ ψ 2 ∥ L 1 , ∞ ( Ω T , μ ) ≤ 2 δ 2 ( ∥ ρ ∥ L t ∞ L x ∞ + ∥ ρ ^ ∥ L t ∞ L x ∞ ) ∫ 0 T ∥ M ~ ( ∇ x 1 b 2 ε ) ( t , ⋅ ) ∥ L 1 , ∞ ( R d ) d t ≲ δ 1 δ 2 ∥ ∇ x 1 b 2 ε ∥ L t 1 L x 1 ≲ δ 1 δ 2 ε s − 1 ∥ b 2 ∥ L 1 ( [ 0 , T ] × R d 2 ; W x 1 s , 1 ) ≲ δ 1 δ 2 ε s − 1 . \begin{aligned}
\|\psi_2\|_{L^{1,\infty}(\Omega_T,\mu)}
&\leq\frac{2}{\delta_2}\left(\|\rho\|_{L^\infty_tL^\infty_x}+\|\widehat{\rho}\|_{L^\infty_tL^\infty_x}\right)\int_0^T\|\widetilde{\mathcal{M}}(\nabla_{x_1}b_2^\varepsilon)(t,\cdot)\|_{L^{1,\infty}(\mathbb{R}^d)}\,dt\\
&\lesssim\frac{\delta_1}{\delta_2}\|\nabla_{x_1}b_2^\varepsilon\|_{L^1_tL^1_x}\\
&\lesssim\frac{\delta_1}{\delta_2}\varepsilon^{s-1}\|b_2\|_{L^1([0,T]\times\mathbb{R}^{d_2};W^{s,1}_{x_1})}\\
&\lesssim\frac{\delta_1}{\delta_2}\varepsilon^{s-1}.
\end{aligned} ∥ ψ 2 ∥ L 1 , ∞ ( Ω T , μ ) ≤ δ 2 2 ( ∥ ρ ∥ L t ∞ L x ∞ + ∥ ρ ∥ L t ∞ L x ∞ ) ∫ 0 T ∥ M ( ∇ x 1 b 2 ε ) ( t , ⋅ ) ∥ L 1 , ∞ ( R d ) d t ≲ δ 2 δ 1 ∥ ∇ x 1 b 2 ε ∥ L t 1 L x 1 ≲ δ 2 δ 1 ε s − 1 ∥ b 2 ∥ L 1 ([ 0 , T ] × R d 2 ; W x 1 s , 1 ) ≲ δ 2 δ 1 ε s − 1 .
The right-hand side diverges as ε → 0 + \varepsilon\to0^+ ε → 0 + whenever 0 < s < 1 0<s<1 0 < s < 1 , which highlights the importance of a careful choice of δ 1 \delta_1 δ 1 and δ 2 \delta_2 δ 2 . On the other hand, by the local L p L^p L p -integrability of b b b on [ 0 , T ] × R d [0,T]\times\mathbb{R}^d [ 0 , T ] × R d ,
∥ ψ 2 ∥ L p ( Ω T , μ ) ≲ 1 δ 1 ∥ b ε ∥ L p ( [ 0 , T ] × B R ) ≲ 1 δ 1 ∥ b ∥ L p ( [ 0 , T ] × B R + 1 ) ≲ 1 δ 1 . \begin{aligned}
\|\psi_2\|_{L^p(\Omega_T,\mu)}
&\lesssim\frac{1}{\delta_1}\|b^\varepsilon\|_{L^p([0,T]\times B_R)}\\
&\lesssim\frac{1}{\delta_1}\|b\|_{L^p([0,T]\times B_{R+1})}\\
&\lesssim\frac{1}{\delta_1}.
\end{aligned} ∥ ψ 2 ∥ L p ( Ω T , μ ) ≲ δ 1 1 ∥ b ε ∥ L p ([ 0 , T ] × B R ) ≲ δ 1 1 ∥ b ∥ L p ([ 0 , T ] × B R + 1 ) ≲ δ 1 1 .
Applying the interpolation inequality (2.4) and the bound ∥ ψ 2 ∥ L p , ∞ ( Ω T , μ ) ≤ ∥ ψ 2 ∥ L p ( Ω T , μ ) \|\psi_2\|_{L^{p,\infty}(\Omega_T,\mu)}\leq\|\psi_2\|_{L^p(\Omega_T,\mu)} ∥ ψ 2 ∥ L p , ∞ ( Ω T , μ ) ≤ ∥ ψ 2 ∥ L p ( Ω T , μ ) , we obtain
∥ ψ 2 ∥ L 1 ( Ω T , μ ) ≤ p p − 1 ∥ ψ 2 ∥ L 1 , ∞ ( Ω T , μ ) [ 1 + log ( μ ( Ω T ) 1 − 1 / p ∥ ψ 2 ∥ L p ( Ω T , μ ) ∥ ψ 2 ∥ L 1 , ∞ ( Ω T , μ ) ) ] . \|\psi_2\|_{L^1(\Omega_T,\mu)}
\leq\frac{p}{p-1}\|\psi_2\|_{L^{1,\infty}(\Omega_T,\mu)}
\left[1+\log\left(\frac{\mu(\Omega_T)^{1-1/p}\|\psi_2\|_{L^p(\Omega_T,\mu)}}{\|\psi_2\|_{L^{1,\infty}(\Omega_T,\mu)}}\right)\right]. ∥ ψ 2 ∥ L 1 ( Ω T , μ ) ≤ p − 1 p ∥ ψ 2 ∥ L 1 , ∞ ( Ω T , μ ) [ 1 + log ( ∥ ψ 2 ∥ L 1 , ∞ ( Ω T , μ ) μ ( Ω T ) 1 − 1/ p ∥ ψ 2 ∥ L p ( Ω T , μ ) ) ] .
Since
∥ ψ 2 ∥ L 1 , ∞ ( Ω T , μ ) ≤ μ ( Ω T ) 1 − 1 p ∥ ψ 2 ∥ L p , ∞ ( Ω T , μ ) ≤ μ ( Ω T ) 1 − 1 p ∥ ψ 2 ∥ L p ( Ω T , μ ) , \|\psi_2\|_{L^{1,\infty}(\Omega_T,\mu)}
\leq\mu(\Omega_T)^{1-\frac{1}{p}}\|\psi_2\|_{L^{p,\infty}(\Omega_T,\mu)}
\leq\mu(\Omega_T)^{1-\frac{1}{p}}\|\psi_2\|_{L^p(\Omega_T,\mu)}, ∥ ψ 2 ∥ L 1 , ∞ ( Ω T , μ ) ≤ μ ( Ω T ) 1 − p 1 ∥ ψ 2 ∥ L p , ∞ ( Ω T , μ ) ≤ μ ( Ω T ) 1 − p 1 ∥ ψ 2 ∥ L p ( Ω T , μ ) ,
the right-hand side is nondecreasing as a function of ∥ ψ 2 ∥ L 1 , ∞ ( Ω T , μ ) \|\psi_2\|_{L^{1,\infty}(\Omega_T,\mu)} ∥ ψ 2 ∥ L 1 , ∞ ( Ω T , μ ) . Thus,
∥ ψ 2 ∥ L 1 ( Ω T , μ ) ≲ δ 1 δ 2 ε s − 1 [ 1 + log ( 1 δ 1 δ 1 δ 2 ε s − 1 ) ] . \|\psi_2\|_{L^1(\Omega_T,\mu)}
\lesssim\frac{\delta_1}{\delta_2}\varepsilon^{s-1}
\left[1+\log\left(\frac{1}{\delta_1\frac{\delta_1}{\delta_2}\varepsilon^{s-1}}\right)\right]. ∥ ψ 2 ∥ L 1 ( Ω T , μ ) ≲ δ 2 δ 1 ε s − 1 [ 1 + log ( δ 1 δ 2 δ 1 ε s − 1 1 ) ] .
Combining all the above estimates, we derive
I 1 ≲ 1 δ 2 ∥ b 2 − b ^ 2 ∥ L t 1 L x 1 + 1 δ 2 ∥ b ^ 2 − b ^ 2 ε ∥ L t 1 L x 1 + 1 δ 1 ∥ b 1 − b ^ 1 ∥ L 1 ( [ 0 , T ] ; L p ( B R ) ) . + 1 δ 1 ∥ b 2 − b ^ 2 ∥ L t 1 L x 1 + δ 1 δ 2 ε s − 1 [ 1 + log ( 1 δ 1 δ 1 δ 2 ε s − 1 ) ] + 1 \begin{aligned}
I_1\lesssim\frac{1}{\delta_2}\|b_2-\widehat{b}_2\|_{L^1_tL^1_x}
+\frac{1}{\delta_2}\|\widehat{b}_2-\widehat{b}_2^\varepsilon\|_{L^1_tL^1_x}
+\frac{1}{\delta_1}\|b_1-\widehat{b}_1\|_{L^1([0,T];L^p(B_R))}. \\
&+\frac{1}{\delta_1}\lVert b_2-\widehat{b}_2\rVert_{L_t^1L_x^1}+\frac{\delta_1}{\delta_2}\varepsilon^{s-1}\left[1+\log\left(\frac{1}{\delta_1\frac{\delta_1}{\delta_2}\varepsilon^{s-1}}\right)\right]+1
\end{aligned} I 1 ≲ δ 2 1 ∥ b 2 − b 2 ∥ L t 1 L x 1 + δ 2 1 ∥ b 2 − b 2 ε ∥ L t 1 L x 1 + δ 1 1 ∥ b 1 − b 1 ∥ L 1 ([ 0 , T ] ; L p ( B R )) . + δ 1 1 ∥ b 2 − b 2 ∥ L t 1 L x 1 + δ 2 δ 1 ε s − 1 [ 1 + log ( δ 1 δ 2 δ 1 ε s − 1 1 ) ] + 1
≲ 1 δ 1 ∥ b 1 − b ^ 1 ∥ L 1 ( [ 0 , T ] ; L p ( B R ) ) + 1 δ 1 ∥ b 2 − b ^ 2 ∥ L t 1 L x 1 + ε s δ 2 + δ 1 δ 2 ε s − 1 [ 1 + log ( δ 2 δ 1 2 ε s − 1 ) ] + 1. \lesssim\frac{1}{\delta_1}\lVert b_1-\widehat{b}_1\rVert_{L^1([0,T];L^p(B_R))}+\frac{1}{\delta_1}\lVert b_2-\widehat{b}_2\rVert_{L_t^1L_x^1}+\frac{\varepsilon^s}{\delta_2}
+\frac{\delta_1}{\delta_2}\varepsilon^{s-1}\left[1+\log\left(\frac{\delta_2}{\delta_1^2\varepsilon^{s-1}}\right)\right]+1. ≲ δ 1 1 ∥ b 1 − b 1 ∥ L 1 ([ 0 , T ] ; L p ( B R )) + δ 1 1 ∥ b 2 − b 2 ∥ L t 1 L x 1 + δ 2 ε s + δ 2 δ 1 ε s − 1 [ 1 + log ( δ 1 2 ε s − 1 δ 2 ) ] + 1.
Step 4: Choice of the parameters. Set
β : = δ 1 δ 2 ≪ 1 , \beta:=\frac{\delta_1}{\delta_2}\ll1, β := δ 2 δ 1 ≪ 1 ,
and let ε = β \varepsilon=\beta ε = β . Then
ε s δ 2 + δ 1 δ 2 ε s − 1 [ 1 + log ( δ 2 δ 1 2 ε s − 1 ) ] = β s δ 2 + β s [ 1 + log ( 1 δ 2 β s + 1 ) ] . \frac{\varepsilon^s}{\delta_2}+\frac{\delta_1}{\delta_2}\varepsilon^{s-1}\left[1+\log\left(\frac{\delta_2}{\delta_1^2\varepsilon^{s-1}}\right)\right]=\frac{\beta^s}{\delta_2}+\beta^s\left[1+\log\left(\frac{1}{\delta_2\beta^{s+1}}\right)\right]. δ 2 ε s + δ 2 δ 1 ε s − 1 [ 1 + log ( δ 1 2 ε s − 1 δ 2 ) ] = δ 2 β s + β s [ 1 + log ( δ 2 β s + 1 1 ) ] .
For δ 2 ∈ ( 0 , 1 ) \delta_2\in(0,1) δ 2 ∈ ( 0 , 1 ) ,
β s δ 2 + β s [ 1 + log ( 1 δ 2 β s + 1 ) ] = β s [ 1 δ 2 + 1 − log δ 2 − ( s + 1 ) log β ] ≤ β s [ 1 δ 2 + 1 − log δ 2 ] + 2 β s ( − log β ) . \begin{aligned}
\frac{\beta^s}{\delta_2}+\beta^s\left[1+\log\left(\frac{1}{\delta_2\beta^{s+1}}\right)\right]=\beta^s\left[\frac{1}{\delta_2}+1-\log\delta_2-(s+1)\log\beta\right] \\
&\le\beta^s\left[\frac{1}{\delta_2}+1-\log\delta_2\right]+2\beta^s(-\log\beta).
\end{aligned} δ 2 β s + β s [ 1 + log ( δ 2 β s + 1 1 ) ] = β s [ δ 2 1 + 1 − log δ 2 − ( s + 1 ) log β ] ≤ β s [ δ 2 1 + 1 − log δ 2 ] + 2 β s ( − log β ) .
Since
lim β → 0 2 β s ( − log β ) = 0 , \lim_{\beta\to0}2\beta^s(-\log\beta)=0, β → 0 lim 2 β s ( − log β ) = 0 ,
there exists a constant β 0 ∈ ( 0 , 1 ) \beta_0\in(0,1) β 0 ∈ ( 0 , 1 ) such that, for β ∈ ( 0 , β 0 ] \beta\in(0,\beta_0] β ∈ ( 0 , β 0 ] ,
2 β s ( − log β ) < 1. 2\beta^s(-\log\beta)<1. 2 β s ( − log β ) < 1.
Consequently, there exists
δ 1 ≤ min { δ 2 β 0 , δ 2 ( 1 [ 1 δ 2 + 1 − log ( δ 2 ) ] ) 1 / s } < 1 \delta_1\le\min\left\{\delta_2\beta_0,\delta_2\left(\frac{1}{\left[\frac{1}{\delta_2}+1-\log(\delta_2)\right]}\right)^{1/s}\right\}<1 δ 1 ≤ min ⎩ ⎨ ⎧ δ 2 β 0 , δ 2 [ δ 2 1 + 1 − log ( δ 2 ) ] 1 1/ s ⎭ ⎬ ⎫ < 1
such that
(3.15)
I 1 ≲ 1 δ 1 ∥ b 1 − b ^ 1 ∥ L 1 ( [ 0 , T ] ; L p ( B R ) ) + 1 δ 1 ∥ b 2 − b ^ 2 ∥ L t 1 L x 1 + 1. I_1\lesssim\frac{1}{\delta_1}\lVert b_1-\widehat{b}_1\rVert_{L^1([0,T];L^p(B_R))}+\frac{1}{\delta_1}\lVert b_2-\widehat{b}_2\rVert_{L_t^1L_x^1}+1. I 1 ≲ δ 1 1 ∥ b 1 − b 1 ∥ L 1 ([ 0 , T ] ; L p ( B R )) + δ 1 1 ∥ b 2 − b 2 ∥ L t 1 L x 1 + 1.
Estimate for I 2 I_2 I 2 . Since the diffusion coefficient is the Sobolev-regular part, we estimate I 2 I_2 I 2 directly:
I 2 ≲ ∫ Ω T , R ∥ B − 1 ( σ ( s , x ) − σ ^ ( s , y ) ) ∥ 2 ∣ B − 1 ( x − y ) ∣ 2 + α + ∣ B − 1 ( x − y ) ∣ 2 + α π s ( d x , d y ) d s I_2\lesssim\int_{\Omega_{T,R}}\frac{\lVert B^{-1}(\sigma(s,x)-\widehat{\sigma}(s,y))\rVert^2}{\sqrt{\lvert B^{-1}(x-y)\rvert^2+\alpha}+\lvert B^{-1}(x-y)\rvert^2+\alpha}\,\pi_s(dx,dy)\,ds I 2 ≲ ∫ Ω T , R ∣ B − 1 ( x − y ) ∣ 2 + α + ∣ B − 1 ( x − y ) ∣ 2 + α ∥ B − 1 ( σ ( s , x ) − σ ( s , y )) ∥ 2 π s ( d x , d y ) d s
≲ ∫ Ω T , R ∥ B − 1 ( σ ( s , x ) − σ ( s , y ) ) ∥ 2 ∣ B − 1 ( x − y ) ∣ 2 π s ( d x , d y ) d s + ∫ Ω T , R ∥ B − 1 ( σ ( s , y ) − σ ^ ( s , y ) ) ∥ 2 α π s ( d x , d y ) d s ≲ 1 δ 1 2 α ∫ 0 T ∫ B R ∥ σ ( s , y ) − σ ^ ( s , y ) ∥ 2 ρ ^ ( y ) d y d t + ∫ Ω T , R ∥ B − 1 ( σ ( s , x ) − σ ( s , y ) ) ∥ 2 ∣ B − 1 ( x − y ) ∣ 2 π s ( d x , d y ) d s ≲ 1 δ 1 2 α ∥ σ − σ ^ ∥ L 2 ( [ 0 , T ] ; L 2 r ( B R ) ) 2 + ∫ Ω T , R ∥ B − 1 ( σ ( s , x ) − σ ( s , y ) ) ∥ 2 ∣ B − 1 ( x − y ) ∣ 2 π s ( d x , d y ) d s . \begin{aligned}
&\lesssim\int_{\Omega_{T,R}} \frac{\lVert B^{-1}(\sigma(s,x)-\sigma(s,y))\rVert^2}{\lvert B^{-1}(x-y)\rvert^2}\,\pi_s(dx,dy)\,ds \\
&\quad+ \int_{\Omega_{T,R}} \frac{\lVert B^{-1}(\sigma(s,y)-\widehat{\sigma}(s,y))\rVert^2}{\sqrt{\alpha}}\,\pi_s(dx,dy)\,ds \\
&\lesssim\frac{1}{\delta_1^2\sqrt{\alpha}}\int_0^T\int_{B_R}\lVert\sigma(s,y)-\widehat{\sigma}(s,y)\rVert^2\widehat{\rho}(y)\,dy\,dt \\
&\quad+ \int_{\Omega_{T,R}}\frac{\lVert B^{-1}(\sigma(s,x)-\sigma(s,y))\rVert^2}{\lvert B^{-1}(x-y)\rvert^2}\,\pi_s(dx,dy)\,ds \\
&\lesssim\frac{1}{\delta_1^2\sqrt{\alpha}}\lVert\sigma-\widehat{\sigma}\rVert^2_{L^2([0,T];L^{2r}(B_R))} \\
&\quad+ \int_{\Omega_{T,R}}\frac{\lVert B^{-1}(\sigma(s,x)-\sigma(s,y))\rVert^2}{\lvert B^{-1}(x-y)\rvert^2}\,\pi_s(dx,dy)\,ds.
\end{aligned} ≲ ∫ Ω T , R ∣ B − 1 ( x − y ) ∣ 2 ∥ B − 1 ( σ ( s , x ) − σ ( s , y )) ∥ 2 π s ( d x , d y ) d s + ∫ Ω T , R α ∥ B − 1 ( σ ( s , y ) − σ ( s , y )) ∥ 2 π s ( d x , d y ) d s ≲ δ 1 2 α 1 ∫ 0 T ∫ B R ∥ σ ( s , y ) − σ ( s , y ) ∥ 2 ρ ( y ) d y d t + ∫ Ω T , R ∣ B − 1 ( x − y ) ∣ 2 ∥ B − 1 ( σ ( s , x ) − σ ( s , y )) ∥ 2 π s ( d x , d y ) d s ≲ δ 1 2 α 1 ∥ σ − σ ∥ L 2 ([ 0 , T ] ; L 2 r ( B R )) 2 + ∫ Ω T , R ∣ B − 1 ( x − y ) ∣ 2 ∥ B − 1 ( σ ( s , x ) − σ ( s , y )) ∥ 2 π s ( d x , d y ) d s .
Here, for the penultimate inequality, we use that ρ ∈ L t ∞ L x 1 \rho\in L_t^\infty L_x^1 ρ ∈ L t ∞ L x 1 and δ 1 ∈ ( 0 , δ 2 ) \delta_1\in(0,\delta_2) δ 1 ∈ ( 0 , δ 2 ) , and for the last inequality, we applied Hölder’s inequality since σ , σ ^ ∈ L t 2 L x ∞ \sigma,\widehat{\sigma}\in L_t^2L_x^\infty σ , σ ∈ L t 2 L x ∞ and ρ ^ ∈ L t ∞ L x r ′ \widehat{\rho}\in L_t^\infty L_x^{r'} ρ ∈ L t ∞ L x r ′ with r ′ r' r ′ being the conjugate exponent of r r r .
Set
I 2 , 1 : = ∫ Ω T , R ∥ B − 1 ( σ ( s , x ) − σ ( s , y ) ) ∥ 2 ∣ B − 1 ( x − y ) ∣ 2 π s ( d x , d y ) d s . I_{2,1}:=\int_{\Omega_{T,R}}\frac{\lVert B^{-1}(\sigma(s,x)-\sigma(s,y))\rVert^2}{\lvert B^{-1}(x-y)\rvert^2}\,\pi_s(dx,dy)\,ds. I 2 , 1 := ∫ Ω T , R ∣ B − 1 ( x − y ) ∣ 2 ∥ B − 1 ( σ ( s , x ) − σ ( s , y )) ∥ 2 π s ( d x , d y ) d s .
Applying Lemma 3.11 (see also Remark 3.2(iii)), we have for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] and for a.e. x , y ∈ R d x,y\in\mathbb{R}^d x , y ∈ R d ,
∥ B − 1 ( σ ( t , x ) − σ ( t , y ) ) ∥ 2 ∣ B − 1 ( x − y ) ∣ 2 ≲ ∥ B − 1 ( σ 1 ( t , x ) − σ 1 ( t , y ) ) ∥ 2 ∣ B − 1 ( x − y ) ∣ 2 + ∥ B − 1 ( σ 2 ( t , x ) − σ 2 ( t , y ) ) ∥ 2 ∣ B − 1 ( x − y ) ∣ 2 \frac{\lVert B^{-1}(\sigma(t,x)-\sigma(t,y))\rVert^2}{\lvert B^{-1}(x-y)\rvert^2}
\lesssim
\frac{\lVert B^{-1}(\sigma_1(t,x)-\sigma_1(t,y))\rVert^2}{\lvert B^{-1}(x-y)\rvert^2}
+
\frac{\lVert B^{-1}(\sigma_2(t,x)-\sigma_2(t,y))\rVert^2}{\lvert B^{-1}(x-y)\rvert^2} ∣ B − 1 ( x − y ) ∣ 2 ∥ B − 1 ( σ ( t , x ) − σ ( t , y )) ∥ 2 ≲ ∣ B − 1 ( x − y ) ∣ 2 ∥ B − 1 ( σ 1 ( t , x ) − σ 1 ( t , y )) ∥ 2 + ∣ B − 1 ( x − y ) ∣ 2 ∥ B − 1 ( σ 2 ( t , x ) − σ 2 ( t , y )) ∥ 2
≲ 1 δ 1 2 ( M ~ ( ∇ x 1 σ 1 ) ) 2 ( t , x ) + 1 δ 1 2 ( M ~ ( ∇ x 1 σ 1 ) ) 2 ( t , y ) + 1 δ 2 2 ( M ~ ( ∇ σ 2 ) ) 2 ( t , x ) + 1 δ 2 2 ( M ~ ( ∇ σ 2 ) ) 2 ( t , y ) . \begin{aligned}
&\lesssim\frac{1}{\delta_1^2}\left(\widetilde{\mathcal{M}}(\nabla_{x_1}\sigma_1)\right)^2(t,x)
+\frac{1}{\delta_1^2}\left(\widetilde{\mathcal{M}}(\nabla_{x_1}\sigma_1)\right)^2(t,y)\\
&\quad+\frac{1}{\delta_2^2}\left(\widetilde{\mathcal{M}}(\nabla\sigma_2)\right)^2(t,x)
+\frac{1}{\delta_2^2}\left(\widetilde{\mathcal{M}}(\nabla\sigma_2)\right)^2(t,y).
\end{aligned} ≲ δ 1 2 1 ( M ( ∇ x 1 σ 1 ) ) 2 ( t , x ) + δ 1 2 1 ( M ( ∇ x 1 σ 1 ) ) 2 ( t , y ) + δ 2 2 1 ( M ( ∇ σ 2 ) ) 2 ( t , x ) + δ 2 2 1 ( M ( ∇ σ 2 ) ) 2 ( t , y ) .
Combining the fact that ρ , ρ ^ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) \rho,\widehat{\rho}\in L_t^\infty(L_x^1\cap L_x^\infty) ρ , ρ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) , assumption ( H σ ) (H^\sigma) ( H σ ) , Lemma 3.12, by the similar argument as above, we obtain
I 2 , 1 ≲ 1. I_{2,1}\lesssim1. I 2 , 1 ≲ 1.
Taking α = 1 \alpha=1 α = 1 ,
I 2 ≲ 1 δ 1 2 ∥ σ − σ ^ ∥ L 2 ( [ 0 , T ] ; L 2 r ( B R ) ) 2 + 1. (23) I_2\lesssim\frac{1}{\delta_1^2}\lVert\sigma-\widehat{\sigma}\rVert^2_{L^2([0,T];L^{2r}(B_R))}+1. \tag*{(23)} I 2 ≲ δ 1 2 1 ∥ σ − σ ∥ L 2 ([ 0 , T ] ; L 2 r ( B R )) 2 + 1. (23)
With α = 1 \alpha=1 α = 1 , by (3.13), we have
E log ( 1 + ∣ B − 1 ( X 0 − X ^ 0 ) ∣ 2 + α ) ≤ E log ( 1 + ∣ B − 1 ( X 0 − X ^ 0 ) ∣ ) + 1 = D δ 1 , δ 2 ( ν , ν ^ ) + 1. (24) \begin{aligned}
\mathbb{E}\log\left(1+\sqrt{\lvert B^{-1}(X_0-\widehat{X}_0)\rvert^2+\alpha}\right)
\leq\mathbb{E}\log\left(1+\lvert B^{-1}(X_0-\widehat{X}_0)\rvert\right)+1 \\
&= \mathcal{D}_{\delta_1,\delta_2}(\nu,\hat{\nu})+1. \tag*{(24)}
\end{aligned} E log ( 1 + ∣ B − 1 ( X 0 − X 0 ) ∣ 2 + α ) ≤ E log ( 1 + ∣ B − 1 ( X 0 − X 0 )∣ ) + 1 = D δ 1 , δ 2 ( ν , ν ^ ) + 1. (24)
Finally, combining the estimates (22) , (3.15), (23) and (24) and taking the supremum on [ 0 , T ] [0,T] [ 0 , T ] , we complete the proof.
Rates of convergence
The goal of this subsection is to show that the distance D δ 1 , δ 2 \mathcal{D}_{\delta_1,\delta_2} D δ 1 , δ 2 plays a crucial role in deriving explicit weak convergence rates for solutions to the FPE, which further illustrates its advantage over the classical L 1 L^1 L 1 -Wasserstein distance. To make this point precise, we specialize Assumption 3.1 to the case b 1 = 0 b_1=0 b 1 = 0 and σ = 0 \sigma=0 σ = 0 , with b 2 ∈ L p ( [ 0 , T ] × R d ; R d 2 ) b_2\in L^p([0,T]\times\mathbb{R}^d;\mathbb{R}^{d_2}) b 2 ∈ L p ([ 0 , T ] × R d ; R d 2 ) , so that the integrability now holds over the whole space rather than only locally.
Perturbations of b 2 b_2 b 2 . We consider a sequence b 2 n → b 2 b_2^n\to b_2 b 2 n → b 2 in L t 1 L x 1 L_t^1L_x^1 L t 1 L x 1 as n → ∞ n\to\infty n → ∞ , where each b 2 n b_2^n b 2 n satisfies the conditions in Assumption 3.1, while keeping
ρ 0 n = ρ 0 , b 1 n = b 1 = 0 , σ n = σ = 0. \rho_0^n=\rho_0,\qquad b_1^n=b_1=0,\qquad\sigma^n=\sigma=0. ρ 0 n = ρ 0 , b 1 n = b 1 = 0 , σ n = σ = 0.
For each n ≥ 1 n\ge1 n ≥ 1 , let ρ n , ρ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) \rho^n,\rho\in L_t^\infty(L_x^1\cap L_x^\infty) ρ n , ρ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) denote the corresponding weak solutions to the FPE with drift components b 2 n b_2^n b 2 n and b 2 b_2 b 2 , respectively. Set
ϵ n : = ∥ b 2 n − b 2 ∥ L t 1 L x 1 , \epsilon_n:=\lVert b_2^n-b_2\rVert_{L_t^1L_x^1}, ϵ n := ∥ b 2 n − b 2 ∥ L t 1 L x 1 ,
so that ϵ n → 0 \epsilon_n\to0 ϵ n → 0 as n → ∞ n\to\infty n → ∞ . By Theorem 3.3[2 ] , under the present specialization, the right-hand side of the stability estimate is independent of R R R . Hence, letting R → ∞ R\to\infty R → ∞ , we obtain
sup t ∈ [ 0 , T ] D δ 1 , δ 2 ( ρ n ( t , ⋅ ) , ρ ( t , ⋅ ) ) ≲ ϵ n δ 1 + 1 \sup_{t\in[0,T]}\mathcal{D}_{\delta_1,\delta_2}\bigl(\rho^n(t,\cdot),\rho(t,\cdot)\bigr)\lesssim\frac{\epsilon_n}{\delta_1}+1 t ∈ [ 0 , T ] sup D δ 1 , δ 2 ( ρ n ( t , ⋅ ) , ρ ( t , ⋅ ) ) ≲ δ 1 ϵ n + 1
for every admissible pair ( δ 1 , δ 2 ) (\delta_1,\delta_2) ( δ 1 , δ 2 ) .
To quantify the weak convergence, we choose vanishing admissible parameters ( δ 1 n , δ 2 n ) (\delta_1^n,\delta_2^n) ( δ 1 n , δ 2 n ) such that the corresponding Kantorovich–Rubinstein distance remains uniformly bounded. By the parameter choice in the proof of Theorem 3.3[2 ] , for all sufficiently large n n n , we may take
δ 1 n : = ϵ n , δ 2 n : = ( δ 1 n ) s s + 1 , s ∈ ( 0 , 1 ) , \delta_1^n:=\epsilon_n,\qquad\delta_2^n:=(\delta_1^n)^{\frac{s}{s+1}},\qquad s\in(0,1), δ 1 n := ϵ n , δ 2 n := ( δ 1 n ) s + 1 s , s ∈ ( 0 , 1 ) ,
and therefore,
sup t ∈ [ 0 , T ] D δ 1 n , δ 2 n ( ρ n ( t , ⋅ ) , ρ ( t , ⋅ ) ) ≲ 1 , \sup_{t\in[0,T]}\mathcal{D}_{\delta_1^n,\delta_2^n}\bigl(\rho^n(t,\cdot),\rho(t,\cdot)\bigr)\lesssim1, t ∈ [ 0 , T ] sup D δ 1 n , δ 2 n ( ρ n ( t , ⋅ ) , ρ ( t , ⋅ ) ) ≲ 1 ,
with a constant independent of n n n .
We now show quantitatively that the shrinking of these scales implies weak convergence. Fix 0 < θ < 1 0<\theta<1 0 < θ < 1 and set
η i n : = ( δ i n ) θ , i = 1 , 2. \eta_i^n:=(\delta_i^n)^\theta,\qquad i=1,2. η i n := ( δ i n ) θ , i = 1 , 2.
Then for each i = 1 , 2 i=1,2 i = 1 , 2 , η i n → 0 \eta_i^n\to0 η i n → 0 and η i n / δ i n → ∞ \eta_i^n/\delta_i^n\to\infty η i n / δ i n → ∞ as n → ∞ n\to\infty n → ∞ . For each t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , let π n , t \pi_{n,t} π n , t be an optimal coupling associated with D δ 1 n , δ 2 n ( ρ n ( t , ⋅ ) , ρ ( t , ⋅ ) ) \mathcal{D}_{\delta_1^n,\delta_2^n}\bigl(\rho^n(t,\cdot),\rho(t,\cdot)\bigr) D δ 1 n , δ 2 n ( ρ n ( t , ⋅ ) , ρ ( t , ⋅ ) ) , and let B n B_n B n be the matrix in (3.1) corresponding to ( δ 1 n , δ 2 n ) (\delta_1^n,\delta_2^n) ( δ 1 n , δ 2 n ) . Writing x = ( x 1 , x 2 ) ∈ R d 1 × R d 2 x=(x_1,x_2)\in\mathbb{R}^{d_1}\times\mathbb{R}^{d_2} x = ( x 1 , x 2 ) ∈ R d 1 × R d 2 and y = ( y 1 , y 2 ) ∈ R d 1 × R d 2 y=(y_1,y_2)\in\mathbb{R}^{d_1}\times\mathbb{R}^{d_2} y = ( y 1 , y 2 ) ∈ R d 1 × R d 2 , define
A i , n : = { ( x , y ) : ∣ x i − y i ∣ > η i n } , i = 1 , 2. A_{i,n}:=\{(x,y): |x_i-y_i|>\eta_i^n\},\qquad i=1,2. A i , n := {( x , y ) : ∣ x i − y i ∣ > η i n } , i = 1 , 2.
Then
π n , t ( A i , n ) log ( 1 + η i n δ i n ) ≤ D δ 1 n , δ 2 n ( ρ n ( t , ⋅ ) , ρ ( t , ⋅ ) ) ≲ 1 , i = 1 , 2. \pi_{n,t}(A_{i,n})\log\left(1+\frac{\eta_i^n}{\delta_i^n}\right)\leq\mathcal{D}_{\delta_1^n,\delta_2^n}\bigl(\rho^n(t,\cdot),\rho(t,\cdot)\bigr)\lesssim1,\qquad i=1,2. π n , t ( A i , n ) log ( 1 + δ i n η i n ) ≤ D δ 1 n , δ 2 n ( ρ n ( t , ⋅ ) , ρ ( t , ⋅ ) ) ≲ 1 , i = 1 , 2.
and hence
sup t ∈ [ 0 , T ] π n , t ( A 1 , n ∪ A 2 , n ) ≲ 1 log ( 1 + ( δ 1 n ) θ − 1 ) + 1 log ( 1 + ( δ 2 n ) θ − 1 ) . \sup_{t\in[0,T]}\pi_{n,t}(A_{1,n}\cup A_{2,n})\lesssim\frac{1}{\log\left(1+(\delta_1^n)^{\theta-1}\right)}+\frac{1}{\log\left(1+(\delta_2^n)^{\theta-1}\right)}. t ∈ [ 0 , T ] sup π n , t ( A 1 , n ∪ A 2 , n ) ≲ log ( 1 + ( δ 1 n ) θ − 1 ) 1 + log ( 1 + ( δ 2 n ) θ − 1 ) 1 .
Consequently, for every bounded Lipschitz function f f f on R d \mathbb{R}^d R d , using the coupling π n , t \pi_{n,t} π n , t and splitting the integral over ( A 1 , n ∪ A 2 , n ) c (A_{1,n}\cup A_{2,n})^c ( A 1 , n ∪ A 2 , n ) c and A 1 , n ∪ A 2 , n A_{1,n}\cup A_{2,n} A 1 , n ∪ A 2 , n , we obtain
sup t ∈ [ 0 , T ] ∣ ∫ R d f ( x ) ρ n ( t , x ) d x − ∫ R d f ( x ) ρ ( t , x ) d x ∣ ≤ sup t ∈ [ 0 , T ] ∫ R d × R d ∣ f ( x ) − f ( y ) ∣ π n , t ( d x , d y ) ≤ Lip ( f ) ( η 1 n + η 2 n ) + 2 ∥ f ∥ L ∞ ( R d ) sup t ∈ [ 0 , T ] π n , t ( A 1 , n ∪ A 2 , n ) \begin{aligned}
\sup_{t\in[0,T]}\left|\int_{\mathbb{R}^d}f(x)\rho^n(t,x)\,\mathrm{d}x-\int_{\mathbb{R}^d}f(x)\rho(t,x)\,\mathrm{d}x\right| \\
&\leq\sup_{t\in[0,T]}\int_{\mathbb{R}^d\times\mathbb{R}^d}|f(x)-f(y)|\pi_{n,t}(\mathrm{d}x,\mathrm{d}y) \\
&\leq\operatorname{Lip}(f)(\eta_1^n+\eta_2^n)+2\|f\|_{L^\infty(\mathbb{R}^d)}\sup_{t\in[0,T]}\pi_{n,t}(A_{1,n}\cup A_{2,n})
\end{aligned} t ∈ [ 0 , T ] sup ∫ R d f ( x ) ρ n ( t , x ) d x − ∫ R d f ( x ) ρ ( t , x ) d x ≤ t ∈ [ 0 , T ] sup ∫ R d × R d ∣ f ( x ) − f ( y ) ∣ π n , t ( d x , d y ) ≤ Lip ( f ) ( η 1 n + η 2 n ) + 2∥ f ∥ L ∞ ( R d ) t ∈ [ 0 , T ] sup π n , t ( A 1 , n ∪ A 2 , n )
≲ Lip ( f ) ( e n θ + e n s θ s + 1 ) + 2 ∥ f ∥ L ∞ ( R d ) [ 1 log ( 1 + e n − ( 1 − θ ) ) + 1 log ( 1 + e n − s ( 1 − θ ) s + 1 ) ] \lesssim\operatorname{Lip}(f)\left(e_n^\theta+e_n^{\frac{s\theta}{s+1}}\right)+2\|f\|_{L^\infty(\mathbb{R}^d)}\left[\frac{1}{\log\left(1+e_n^{-(1-\theta)}\right)}+\frac{1}{\log\left(1+e_n^{-\frac{s(1-\theta)}{s+1}}\right)}\right] ≲ Lip ( f ) ( e n θ + e n s + 1 s θ ) + 2∥ f ∥ L ∞ ( R d ) log ( 1 + e n − ( 1 − θ ) ) 1 + log ( 1 + e n − s + 1 s ( 1 − θ ) ) 1
⟶ 0 , as n → ∞ , \longrightarrow0,\qquad\text{as }n\to\infty, ⟶ 0 , as n → ∞ ,
where Lip ( f ) \operatorname{Lip}(f) Lip ( f ) stands for the Lipschitz constant of f f f . Thus, ρ n \rho^n ρ n converges weakly to ρ \rho ρ , uniformly in t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , and the preceding estimate gives an explicit quantitative rate of convergence against bounded Lipschitz test functions.
Perturbations of the initial data. We now assume ρ 0 n → ρ 0 \rho_0^n\to\rho_0 ρ 0 n → ρ 0 in L x 1 ∩ L ∞ L_x^1\cap L^\infty L x 1 ∩ L ∞ as n → ∞ n\to\infty n → ∞ , while b n = b b^n=b b n = b , σ n = σ \sigma^n=\sigma σ n = σ for n ≥ 1. C o m b i n i n g t h e l i n e a r i t y o f t h e F P E w i t h t h e e n e r g y e s t i m a t e s , w e o b t a i n , f o r e v e r y $ 1 ≤ p ≤ ∞ n\geq1. Combining the linearity of the FPE with the energy estimates, we obtain, for every \$1\leq p\leq\infty n ≥ 1. C o mbinin g t h e l in e a r i t y o f t h e F P E w i t h t h ee n er g y es t ima t es , w eo b t ain , f or e v er y $1 ≤ p ≤ ∞ ,
∥ ρ n − ρ ∥ L t ∞ L x p ≲ ∥ ρ 0 n − ρ 0 ∥ L x p . \|\rho^n-\rho\|_{L_t^\infty L_x^p}\lesssim\|\rho_0^n-\rho_0\|_{L_x^p}. ∥ ρ n − ρ ∥ L t ∞ L x p ≲ ∥ ρ 0 n − ρ 0 ∥ L x p .
Consequently, ρ n \rho^n ρ n converges strongly to ρ \rho ρ in L t ∞ L x p L_t^\infty L_x^p L t ∞ L x p for every 1 ≤ p ≤ ∞ 1\leq p\leq\infty 1 ≤ p ≤ ∞ , with convergence rate ∥ ρ 0 n − ρ 0 ∥ L x p \|\rho_0^n-\rho_0\|_{L_x^p} ∥ ρ 0 n − ρ 0 ∥ L x p . This rate is clearly optimal.
Gradient of drift fields as singular integrals
Main result
We introduce several classes of singular kernels of convolution type, and recall the basic properties of the associated singular integral operators; see [17 ] for more details. For a function K K K , we write K ^ \widehat{K} K for its Fourier transform whenever it is well defined.
Definition 4.1 (Singular kernel). Let K ∈ S ′ ( R d ) K \in\mathcal{S}'(\mathbb{R}^{d}) K ∈ S ′ ( R d ) which coincides with a locally integrable function on R d ∖ { 0 } \mathbb{R}^{d} \setminus\{0\} R d ∖ { 0 } . We say that K K K is a singular kernel on R d \mathbb{R}^{d} R d if there exist constants A 1 , A 2 ≥ 0 A_{1}, A_{2} \ge0 A 1 , A 2 ≥ 0 such that
∣ K ^ ( ξ ) ∣ ≤ A 1 , ξ ∈ R d , |\widehat{K}(\xi)| \le A_{1}, \qquad\xi\in\mathbb{R}^{d}, ∣ K ( ξ ) ∣ ≤ A 1 , ξ ∈ R d , ∫ ∣ x ∣ ≥ 2 ∣ y ∣ ∣ K ( x − y ) − K ( x ) ∣ d x ≤ A 2 , y ∈ R d ∖ { 0 } . \int_{|x|\ge2|y|} |K(x-y)-K(x)| \, \mathrm{d}x \le A_{2}, \qquad y \in\mathbb{R}^{d} \setminus\{0\}. ∫ ∣ x ∣ ≥ 2∣ y ∣ ∣ K ( x − y ) − K ( x ) ∣ d x ≤ A 2 , y ∈ R d ∖ { 0 } .
The Calderón–Zygmund theory yields the following results.
Theorem 4.2 (Calderón–Zygmund). Let K K K be a singular kernel as in Definition 4.1 . For u ∈ L 2 ( R d ) u \in L^{2}(\mathbb{R}^{d}) u ∈ L 2 ( R d ) , define the operator associated with K K K as
S u = K ∗ u , Su = K * u, S u = K ∗ u , in the sense of multiplication in the Fourier variable. Then, for every 1 < p < ∞ 1 < p < \infty 1 < p < ∞ , there exists a constant C d , p > 0 C_{d,p} > 0 C d , p > 0 such that
∥ S u ∥ L p ( R d ) ≤ C d , p ( A 1 + A 2 ) ∥ u ∥ L p ( R d ) , u ∈ L p ( R d ) ∩ L 2 ( R d ) . (25) \|Su\|_{L^{p}(\mathbb{R}^{d})} \le C_{d,p}(A_{1}+A_{2})\|u\|_{L^{p}(\mathbb{R}^{d})}, \qquad u \in L^{p}(\mathbb{R}^{d}) \cap L^{2}(\mathbb{R}^{d}). \tag*{(25)} ∥ S u ∥ L p ( R d ) ≤ C d , p ( A 1 + A 2 ) ∥ u ∥ L p ( R d ) , u ∈ L p ( R d ) ∩ L 2 ( R d ) . (25) For p = 1 p=1 p = 1 , there exists a constant C d > 0 C_{d}>0 C d > 0 such that one has the weak-type estimate
∥ S u ∥ L 1 , ∞ ( R d ) ≤ C d ( A 1 + A 2 ) ∥ u ∥ L 1 ( R d ) , u ∈ L 1 ( R d ) ∩ L 2 ( R d ) . (26) \|Su\|_{L^{1,\infty}(\mathbb{R}^{d})} \le C_{d}(A_{1}+A_{2})\|u\|_{L^{1}(\mathbb{R}^{d})}, \qquad u \in L^{1}(\mathbb{R}^{d}) \cap L^{2}(\mathbb{R}^{d}). \tag*{(26)} ∥ S u ∥ L 1 , ∞ ( R d ) ≤ C d ( A 1 + A 2 ) ∥ u ∥ L 1 ( R d ) , u ∈ L 1 ( R d ) ∩ L 2 ( R d ) . (26) Consequently, S S S extends uniquely to a bounded operator on L p ( R d ) L^{p}(\mathbb{R}^{d}) L p ( R d ) for 1 < p < ∞ 1 < p < \infty 1 < p < ∞ , and to a bounded operator from L 1 ( R d ) L^{1}(\mathbb{R}^{d}) L 1 ( R d ) to L 1 , ∞ ( R d ) L^{1,\infty}(\mathbb{R}^{d}) L 1 , ∞ ( R d ) , still denoted by S S S , for which the above estimates hold for all u u u in the respective spaces.
The operator obtained in this way is called the singular integral operator associated with K K K . For p = 1 p=1 p = 1 , the extension S : L 1 ( R d ) → L 1 , ∞ ( R d ) S : L^{1}(\mathbb{R}^{d}) \to L^{1,\infty}(\mathbb{R}^{d}) S : L 1 ( R d ) → L 1 , ∞ ( R d ) may be misleading, since a function in L 1 , ∞ ( R d ) L^{1,\infty}(\mathbb{R}^{d}) L 1 , ∞ ( R d ) is in general not locally integrable and therefore need not define a distribution. To circumvent this, we introduce a different extension into the tempered distributions S d i s : L 1 ( R d ) → S ′ ( R d ) S^{\mathrm{dis}} : L^{1}(\mathbb{R}^{d}) \to\mathcal{S}'(\mathbb{R}^{d}) S dis : L 1 ( R d ) → S ′ ( R d ) , defined via the duality
⟨ S d i s u , φ ⟩ : = ⟨ u , S ~ φ ⟩ , φ ∈ S ( R d ) , (27) \langle S^{\mathrm{dis}}u,\varphi\rangle:= \langle u,\widetilde{S}\varphi\rangle, \qquad\varphi\in\mathcal{S}(\mathbb{R}^{d}), \tag*{(27)} ⟨ S dis u , φ ⟩ := ⟨ u , S φ ⟩ , φ ∈ S ( R d ) , (27) where S ~ \widetilde{S} S is the singular integral operator associated with K ~ ( x ) : = K ( − x ) \widetilde{K}(x):=K(-x) K ( x ) := K ( − x ) . It is important to emphasize that this distributional extension is not the same as the extensions S : L 1 ( R d ) → L 1 , ∞ ( R d ) S : L^{1}(\mathbb{R}^{d}) \to L^{1,\infty}(\mathbb{R}^{d}) S : L 1 ( R d ) → L 1 , ∞ ( R d ) . The two should not be identified.
For our purposes, we use a more regular class of kernels, namely, singular kernels of fundamental type.
Definition 4.3 (Singular kernel of fundamental type). We say that K : R d ∖ { 0 } → R K : \mathbb{R}^{d} \setminus\{0\} \to\mathbb{R} K : R d ∖ { 0 } → R is a singular kernel of fundamental type if it coincides with a C 1 C^{1} C 1 function on R d ∖ { 0 } \mathbb{R}^{d} \setminus\{0\} R d ∖ { 0 } and there exist constants C 1 , C 2 ≥ 0 C_{1}, C_{2} \ge0 C 1 , C 2 ≥ 0 such that
(i) (Size and gradient condition)
∣ K ( x ) ∣ + ∣ x ∣ ∣ ∇ K ( x ) ∣ ≤ C 1 ∣ x ∣ d , x ≠ 0 ; |K(x)|+|x||\nabla K(x)|\le\frac{C_1}{|x|^d},\qquad x\ne0; ∣ K ( x ) ∣ + ∣ x ∣∣∇ K ( x ) ∣ ≤ ∣ x ∣ d C 1 , x = 0 ;
(ii) (Cancellation condition)
sup 0 < R 1 < R 2 < ∞ ∣ ∫ B R 2 ∖ B R 1 K ( x ) d x ∣ ≤ C 2 . \sup_{0<R_1<R_2<\infty}\left|\int_{B_{R_2}\setminus B_{R_1}}K(x)\,\mathrm{d}x\right|\le C_2. 0 < R 1 < R 2 < ∞ sup ∫ B R 2 ∖ B R 1 K ( x ) d x ≤ C 2 .
As a consequence, K K K admits an extension to a tempered distribution on R d \mathbb{R}^d R d , which is a singular kernel and is unique up to adding a constant multiple of the Dirac mass at the origin.
We also recall the following standard notion from the distribution theory.
Definition 4.4. For f ∈ D ′ ( R d ) f\in\mathcal{D}'(\mathbb{R}^d) f ∈ D ′ ( R d ) , its distributional derivative ∂ j f \partial_jf ∂ j f is defined by
⟨ ∂ j f , ψ ⟩ : = − ⟨ f , ∂ j ψ ⟩ , ψ ∈ D ( R d ) . \langle\partial_jf,\psi\rangle:=-\langle f,\partial_j\psi\rangle,\qquad\psi\in\mathcal{D}(\mathbb{R}^d). ⟨ ∂ j f , ψ ⟩ := − ⟨ f , ∂ j ψ ⟩ , ψ ∈ D ( R d ) .
For the statements of our main results, we require the following assumption.
Assumption 4.5. Let K = ( K i j ) 1 ≤ i , j ≤ d K=(K_{ij})_{1\le i,j\le d} K = ( K ij ) 1 ≤ i , j ≤ d be a matrix-valued kernel such that each K i j K_{ij} K ij is a singular kernel of fundamental type, and let w = ( w i ) 1 ≤ i ≤ d w=(w_i)_{1\le i\le d} w = ( w i ) 1 ≤ i ≤ d such that each w i ∈ L 1 ( [ 0 , T ] ; L 1 ( R d ) ) w_i\in L^1([0,T];L^1(\mathbb{R}^d)) w i ∈ L 1 ([ 0 , T ] ; L 1 ( R d )) .
( H ~ ∇ b ) ∇ b = K ∗ w = ( K i j ∗ w i ) 1 ≤ i , j ≤ d (\widetilde{\mathcal{H}}^{\nabla b})\ \nabla b=K*w=(K_{ij}*w_i)_{1\le i,j\le d} ( H ∇ b ) ∇ b = K ∗ w = ( K ij ∗ w i ) 1 ≤ i , j ≤ d , and the convolutions are understood in the distributional sense of (4.3).
( H ~ b ) b ∈ L p , ∞ ( [ 0 , T ] × R d ; R d ) (\widetilde{\mathcal{H}}^b)\ b\in L^{p,\infty}([0,T]\times\mathbb{R}^d;\mathbb{R}^d) ( H b ) b ∈ L p , ∞ ([ 0 , T ] × R d ; R d ) for some p > 1 p>1 p > 1 .
( H ~ σ ) σ ∈ L 2 ( [ 0 , T ] ; L ∞ ( R d ; R d × d ) ) (\widetilde{\mathcal{H}}^\sigma)\ \sigma\in L^2([0,T];L^\infty(\mathbb{R}^d;\mathbb{R}^{d\times d})) ( H σ ) σ ∈ L 2 ([ 0 , T ] ; L ∞ ( R d ; R d × d )) with ( ∑ i , j , k = 1 d ∣ ∂ k σ i j ∣ 2 ) 1 / 2 ∈ L 2 ( [ 0 , T ] ; L 2 r ( R d ) ) \left(\sum_{i,j,k=1}^d|\partial_k\sigma_{ij}|^2\right)^{1/2}\in L^2([0,T];L^{2r}(\mathbb{R}^d)) ( ∑ i , j , k = 1 d ∣ ∂ k σ ij ∣ 2 ) 1/2 ∈ L 2 ([ 0 , T ] ; L 2 r ( R d )) for some r > 1 r>1 r > 1 .
We note that assumption ( H ~ b ) (\widetilde{\mathcal{H}}^b) ( H b ) implies b ∈ L l o c 1 ( [ 0 , T ] × R d ; R d ) b\in L^1_{\mathrm{loc}}([0,T]\times\mathbb{R}^d;\mathbb{R}^d) b ∈ L loc 1 ([ 0 , T ] × R d ; R d ) . This follows immediately by applying the interpolation inequality (2.3) with r = 1 r=1 r = 1 and ℓ = p > 1 \ell=p>1 ℓ = p > 1 on any compact subset of [ 0 , T ] × R d [0,T]\times\mathbb{R}^d [ 0 , T ] × R d . Moreover, under assumption ( H ~ σ ) (\widetilde{\mathcal{H}}^\sigma) ( H σ ) , we have σ ( t , ⋅ ) ∈ W l o c 1 , 1 ( R d ; R d × d ) \sigma(t,\cdot)\in W^{1,1}_{\mathrm{loc}}(\mathbb{R}^d;\mathbb{R}^{d\times d}) σ ( t , ⋅ ) ∈ W loc 1 , 1 ( R d ; R d × d ) for almost every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] .
As before, we consider an increasing logarithmical cost, which generates the Kantorovich–Rubinstein distance
D δ ( μ , ν ) : = inf π ∈ C ( μ , ν ) ∫ R d × R d log ( 1 + ∣ x − y ∣ δ ) π ( d x , d y ) , δ > 0 , μ , ν ∈ P ( R d ) . D_\delta(\mu,\nu):=\inf_{\pi\in\mathcal{C}(\mu,\nu)}\int_{\mathbb{R}^d\times\mathbb{R}^d}\log\left(1+\frac{|x-y|}{\delta}\right)\pi(\mathrm{d}x,\mathrm{d}y),\qquad\delta>0,\quad\mu,\nu\in\mathcal{P}(\mathbb{R}^d). D δ ( μ , ν ) := π ∈ C ( μ , ν ) inf ∫ R d × R d log ( 1 + δ ∣ x − y ∣ ) π ( d x , d y ) , δ > 0 , μ , ν ∈ P ( R d ) .
We now state the stability estimate.
Theorem 4.6. Let ( b , σ , ρ 0 ) (b,\sigma,\rho_0) ( b , σ , ρ 0 ) and ( b ^ , σ ^ , ρ ^ 0 ) (\hat b,\hat\sigma,\hat\rho_0) ( b ^ , σ ^ , ρ ^ 0 ) be two triples satisfying Assumption 4.5 and ( H ν ) (\mathcal{H}^{\nu}) ( H ν ) . Let ρ , ρ ^ ∈ L ∞ ( [ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d ) ) \rho,\hat\rho\in L^\infty([0,T];L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d)) ρ , ρ ^ ∈ L ∞ ([ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d )) be the corresponding weak solutions to the FPE (1.2). Then, for every ϵ > 0 \epsilon>0 ϵ > 0 , there exists a constant C ϵ > 0 C_\epsilon>0 C ϵ > 0 such that, for every δ > 0 \delta>0 δ > 0 ,
sup t ∈ [ 0 , T ] D δ ( ρ ( t , ⋅ ) , ρ ^ ( t , ⋅ ) ) ≲ D δ ( ρ 0 , ρ ^ 0 ) + 1 δ ∫ 0 T ∫ R d ∣ b ( t , x ) − b ^ ( t , x ) ∣ ρ ( t , x ) d x d t + 1 δ 2 ∫ 0 T ∫ R d ∥ σ ( t , x ) − σ ^ ( t , x ) ∥ 2 ρ ( t , x ) d x d t + ϵ [ 1 + log ( 1 ϵ δ ) ] + C ϵ . \sup_{t\in[0,T]}D_\delta(\rho(t,\cdot),\hat\rho(t,\cdot))\lesssim D_\delta(\rho_0,\hat\rho_0)+\frac{1}{\delta}\int_0^T\int_{\mathbb{R}^d}|b(t,x)-\hat b(t,x)|\rho(t,x)\,\mathrm{d}x\mathrm{d}t
+\frac{1}{\delta^2}\int_0^T\int_{\mathbb{R}^d}\|\sigma(t,x)-\hat\sigma(t,x)\|^2\rho(t,x)\,\mathrm{d}x\mathrm{d}t
+\epsilon\left[1+\log\left(\frac{1}{\epsilon\delta}\right)\right]+C_\epsilon. t ∈ [ 0 , T ] sup D δ ( ρ ( t , ⋅ ) , ρ ^ ( t , ⋅ )) ≲ D δ ( ρ 0 , ρ ^ 0 ) + δ 1 ∫ 0 T ∫ R d ∣ b ( t , x ) − b ^ ( t , x ) ∣ ρ ( t , x ) d x d t + δ 2 1 ∫ 0 T ∫ R d ∥ σ ( t , x ) − σ ^ ( t , x ) ∥ 2 ρ ( t , x ) d x d t + ϵ [ 1 + log ( ϵδ 1 ) ] + C ϵ . Here the implicit constant in ≲ \lesssim ≲ is independent of ϵ \epsilon ϵ and δ \delta δ .
The stability estimate in Theorem 4.6 can be restated as follows.
Corollary 4.8. Under the assumptions of Theorem 4.6 , for every ϵ > 0 \epsilon>0 ϵ > 0 , there exists a constant C ϵ > 0 C_\epsilon>0 C ϵ > 0 such that, for every δ > 0 \delta>0 δ > 0 ,
sup t ∈ [ 0 , T ] D δ ( ρ ( t , ⋅ ) , ρ ^ ( t , ⋅ ) ) ≲ D δ ( ρ 0 , ρ ^ 0 ) + 1 δ 2 ∥ σ − σ ^ ∥ L 2 ( [ 0 , T ] ; L ∞ ( R d ) ) 2 ∥ ρ ∥ L ∞ ( [ 0 , T ] ; L 1 ( R d ) ) + 1 δ ∥ b − b ^ ∥ L p , ∞ ( [ 0 , T ] × R d ) ∥ ρ ∥ L 1 ( [ 0 , T ] × R d ) 1 / p ′ ∥ ρ ∥ L ∞ ( [ 0 , T ] × R d ) 1 / p + ϵ [ 1 + log ( 1 ϵ δ ) ] + C ϵ . \begin{aligned}
\sup_{t\in[0,T]}D_\delta(\rho(t,\cdot),\hat\rho(t,\cdot))\lesssim{}& D_\delta(\rho_0,\hat\rho_0)+\frac{1}{\delta^2}\|\sigma-\hat\sigma\|_{L^2([0,T];L^\infty(\mathbb{R}^d))}^2\|\rho\|_{L^\infty([0,T];L^1(\mathbb{R}^d))}\\
&+\frac{1}{\delta}\|b-\hat b\|_{L^{p,\infty}([0,T]\times\mathbb{R}^d)}\|\rho\|_{L^1([0,T]\times\mathbb{R}^d)}^{1/p'}\|\rho\|_{L^\infty([0,T]\times\mathbb{R}^d)}^{1/p}\\
&+\epsilon\left[1+\log\left(\frac{1}{\epsilon\delta}\right)\right]+C_\epsilon.
\end{aligned} t ∈ [ 0 , T ] sup D δ ( ρ ( t , ⋅ ) , ρ ^ ( t , ⋅ )) ≲ D δ ( ρ 0 , ρ ^ 0 ) + δ 2 1 ∥ σ − σ ^ ∥ L 2 ([ 0 , T ] ; L ∞ ( R d )) 2 ∥ ρ ∥ L ∞ ([ 0 , T ] ; L 1 ( R d )) + δ 1 ∥ b − b ^ ∥ L p , ∞ ([ 0 , T ] × R d ) ∥ ρ ∥ L 1 ([ 0 , T ] × R d ) 1/ p ′ ∥ ρ ∥ L ∞ ([ 0 , T ] × R d ) 1/ p + ϵ [ 1 + log ( ϵδ 1 ) ] + C ϵ . where p ′ p' p ′ is the conjugate exponent of p p p , and the implicit constant in ≲ \lesssim ≲ is independent of ϵ \epsilon ϵ and δ \delta δ .
Proof. Let E = [ 0 , T ] × R d E=[0,T]\times\mathbb{R}^d E = [ 0 , T ] × R d . Since ρ ∈ L ∞ ( [ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d ) ) \rho\in L^\infty([0,T];L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d)) ρ ∈ L ∞ ([ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d )) , choosing a product-measurable representative and applying Tonelli’s theorem immediately yield ρ ∈ L 1 ( E ) ∩ L ∞ ( E ) \rho\in L^1(E)\cap L^\infty(E) ρ ∈ L 1 ( E ) ∩ L ∞ ( E ) . By standard real interpolation, it holds that
( L 1 ( E ) , L ∞ ( E ) ) 1 / q , 1 = L q ′ , 1 ( E ) , q ∈ ( 1 , ∞ ) , q ′ : = q q − 1 , \left(L^1(E),L^\infty(E)\right)_{1/q,1}=L^{q',1}(E),\qquad q\in(1,\infty),\quad q':=\frac{q}{q-1}, ( L 1 ( E ) , L ∞ ( E ) ) 1/ q , 1 = L q ′ , 1 ( E ) , q ∈ ( 1 , ∞ ) , q ′ := q − 1 q ,
with an equivalent norm. Hence, ρ ∈ L q , 1 ( E ) \rho\in L^{q,1}(E) ρ ∈ L q , 1 ( E ) for every q ∈ ( 1 , ∞ ) q\in(1,\infty) q ∈ ( 1 , ∞ ) , and there exists a constant C q > 0 C_q>0 C q > 0 such that
∥ ρ ∥ L q , 1 ( E ) ≤ C q ∥ ρ ∥ L 1 ( E ) 1 / q ∥ ρ ∥ L ∞ ( E ) 1 / q ′ . \|\rho\|_{L^{q,1}(E)}\leq C_q\|\rho\|_{L^1(E)}^{1/q}\|\rho\|_{L^\infty(E)}^{1/q'}. ∥ ρ ∥ L q , 1 ( E ) ≤ C q ∥ ρ ∥ L 1 ( E ) 1/ q ∥ ρ ∥ L ∞ ( E ) 1/ q ′ .
Combining this inequality with [17 Exercise 1.4.1(b)] , we obtain for any p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) with p ′ = p p − 1 p' = \frac{p}{p-1} p ′ = p − 1 p ,
∫ 0 T ∫ R d ∣ b ( t , x ) − b ^ ( t , x ) ∣ ρ ( t , x ) d x d t ≤ ∥ b − b ^ ∥ L ∞ , ∞ ( E ) ∥ ρ ∥ L p ′ , 1 ( E ) ≲ ∥ b − b ^ ∥ L p , ∞ ( E ) ∥ ρ ∥ L 1 ( E ) 1 / p ′ ∥ ρ ∥ L ∞ ( E ) 1 / p . \int_0^T \int_{\mathbb{R}^d} |b(t,x)-\hat{b}(t,x)|\rho(t,x)\,\mathrm{d}x\mathrm{d}t \le\|b-\hat{b}\|_{L^{\infty,\infty}(E)}\|\rho\|_{L^{p',1}(E)}
\lesssim\|b-\hat{b}\|_{L^{p,\infty}(E)}\|\rho\|_{L^1(E)}^{1/p'}\|\rho\|_{L^\infty(E)}^{1/p}. ∫ 0 T ∫ R d ∣ b ( t , x ) − b ^ ( t , x ) ∣ ρ ( t , x ) d x d t ≤ ∥ b − b ^ ∥ L ∞ , ∞ ( E ) ∥ ρ ∥ L p ′ , 1 ( E ) ≲ ∥ b − b ^ ∥ L p , ∞ ( E ) ∥ ρ ∥ L 1 ( E ) 1/ p ′ ∥ ρ ∥ L ∞ ( E ) 1/ p .
By the assumption σ , σ ^ ∈ L 2 ( [ 0 , T ] ; L ∞ ( R d ; R d × d ) ) \sigma,\hat{\sigma}\in L^2([0,T];L^\infty(\mathbb{R}^d;\mathbb{R}^{d\times d})) σ , σ ^ ∈ L 2 ([ 0 , T ] ; L ∞ ( R d ; R d × d )) , a direct estimate gives
∫ 0 T ∫ R d ∥ σ ( t , x ) − σ ^ ( t , x ) ∥ 2 ρ ( t , x ) d x d t ≤ ∥ σ − σ ^ ∥ L 2 ( [ 0 , T ] ; L ∞ ( R d ) ) 2 ∥ ρ ∥ L ∞ ( [ 0 , T ] ; L 1 ( R d ) ) . \int_0^T\int_{\mathbb{R}^d}\|\sigma(t,x)-\hat{\sigma}(t,x)\|^2\rho(t,x)\,\mathrm{d}x\mathrm{d}t\le\|\sigma-\hat{\sigma}\|_{L^2([0,T];L^\infty(\mathbb{R}^d))}^2\|\rho\|_{L^\infty([0,T];L^1(\mathbb{R}^d))}. ∫ 0 T ∫ R d ∥ σ ( t , x ) − σ ^ ( t , x ) ∥ 2 ρ ( t , x ) d x d t ≤ ∥ σ − σ ^ ∥ L 2 ([ 0 , T ] ; L ∞ ( R d )) 2 ∥ ρ ∥ L ∞ ([ 0 , T ] ; L 1 ( R d )) .
As an immediate consequence of Theorem 4.6 , we have the following uniqueness.
Corollary 4.9. Under the assumptions of Theorem 4.6 , weak solutions to the FPE (1.2) in the class L ∞ ( [ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d ) ) L^\infty([0,T];L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d)) L ∞ ([ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d )) are unique.
Proof. Let b = b ^ b=\hat{b} b = b ^ , σ = σ ^ \sigma=\hat{\sigma} σ = σ ^ and ρ 0 = ρ ^ 0 \rho_0=\hat{\rho}_0 ρ 0 = ρ ^ 0 . By Theorem 4.6 ,
D δ ( ρ ( t , ⋅ ) , ρ ^ ( t , ⋅ ) ) ≲ ε [ 1 + log ( 1 ε δ ) ] + C ε , t ∈ [ 0 , T ] , ε > 0. D_\delta(\rho(t,\cdot),\hat{\rho}(t,\cdot))\lesssim\varepsilon\left[1+\log\left(\frac{1}{\varepsilon\delta}\right)\right]+C_\varepsilon,\qquad t\in[0,T],\ \varepsilon>0. D δ ( ρ ( t , ⋅ ) , ρ ^ ( t , ⋅ )) ≲ ε [ 1 + log ( ε δ 1 ) ] + C ε , t ∈ [ 0 , T ] , ε > 0.
If 0 < δ < e − 1 0<\delta<e^{-1} 0 < δ < e − 1 , that is, ∣ log δ ∣ > 1 |\log\delta|>1 ∣ log δ ∣ > 1 , then
1 ∣ log δ ∣ [ 1 + log ( 1 ε δ ) ] ≤ 1 + ∣ log δ ∣ + ∣ log ε ∣ ∣ log δ ∣ ≤ 2 + ∣ log ε ∣ . \frac{1}{|\log\delta|}\left[1+\log\left(\frac{1}{\varepsilon\delta}\right)\right]\le\frac{1+|\log\delta|+|\log\varepsilon|}{|\log\delta|}\le2+|\log\varepsilon|. ∣ log δ ∣ 1 [ 1 + log ( ε δ 1 ) ] ≤ ∣ log δ ∣ 1 + ∣ log δ ∣ + ∣ log ε ∣ ≤ 2 + ∣ log ε ∣.
Hence, for an arbitrarily small constant a > 0 a>0 a > 0 , there exists 0 < ε 0 = ε 0 ( a ) < 1 0<\varepsilon_0=\varepsilon_0(a)<1 0 < ε 0 = ε 0 ( a ) < 1 such that
ε 0 ∣ log δ ∣ [ 1 + log ( 1 ε 0 δ ) ] ≤ a 2 . \frac{\varepsilon_0}{|\log\delta|}\left[1+\log\left(\frac{1}{\varepsilon_0\delta}\right)\right]\le\frac{a}{2}. ∣ log δ ∣ ε 0 [ 1 + log ( ε 0 δ 1 ) ] ≤ 2 a .
For this fixed ε 0 > 0 \varepsilon_0>0 ε 0 > 0 , we may choose δ ∈ ( 0 , e − 1 ) \delta\in(0,e^{-1}) δ ∈ ( 0 , e − 1 ) small enough that
C ε 0 ∣ log δ ∣ ≤ a 2 . \frac{C_{\varepsilon_0}}{|\log\delta|}\le\frac{a}{2}. ∣ log δ ∣ C ε 0 ≤ 2 a .
Thus, for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
lim δ → 0 D δ ( ρ ( t , ⋅ ) , ρ ^ ( t , ⋅ ) ) ∣ log δ ∣ = 0. (28) \lim_{\delta\to0}\frac{D_\delta(\rho(t,\cdot),\hat{\rho}(t,\cdot))}{|\log\delta|}=0. \tag*{(28)} δ → 0 lim ∣ log δ ∣ D δ ( ρ ( t , ⋅ ) , ρ ^ ( t , ⋅ )) = 0. (28)
Remark 2.1(2) in [21 ] shows that (28) implies ρ = ρ ^ \rho=\hat{\rho} ρ = ρ ^ in L t ∞ ( L x 1 ∩ L x ∞ ) L^\infty_t(L^1_x\cap L^\infty_x) L t ∞ ( L x 1 ∩ L x ∞ ) .
Proof of Theorem 4.6
Compared with the previous section, the primary challenge in proving Theorem 4.6 is that the Morrey-type estimate from Lemma 3.10 no longer applies to the relevant difference quotients for the drift field b b b introduced in Assumption 4.5. We therefore follow [6 ] and introduce a smooth variant of the usual maximal function. Given φ ∈ C c ∞ ( R d ) \varphi\in C^\infty_c(\mathbb{R}^d) φ ∈ C c ∞ ( R d ) , for ϵ > 0 \epsilon> 0 ϵ > 0 , set
φ ϵ ( x ) : = 1 ϵ d φ ( x ϵ ) , x ∈ R d . \varphi^\epsilon(x) := \frac{1}{\epsilon^d}\varphi\left(\frac{x}{\epsilon}\right), \qquad x \in\mathbb{R}^d. φ ϵ ( x ) := ϵ d 1 φ ( ϵ x ) , x ∈ R d .
For each component, define
M φ ( K i j ∗ w i ) ( t , x ) : = sup ϵ > 0 ∣ ⟨ ( K i j ∗ w i ) ( t , ⋅ ) , φ ϵ ( x − ⋅ ) ⟩ ∣ , ( t , x ) ∈ [ 0 , T ] × R d , \mathcal{M}_{\varphi}(K_{ij} * w_i)(t,x) := \sup_{\epsilon>0} \left|\left\langle(K_{ij} * w_i)(t,\cdot),\varphi^\epsilon(x-\cdot)\right\rangle\right|, \qquad(t,x) \in[0,T] \times\mathbb{R}^d, M φ ( K ij ∗ w i ) ( t , x ) := ϵ > 0 sup ∣ ⟨ ( K ij ∗ w i ) ( t , ⋅ ) , φ ϵ ( x − ⋅ ) ⟩ ∣ , ( t , x ) ∈ [ 0 , T ] × R d ,
in the sense of (4.3), and set
M φ ( K ∗ w ) ( t , x ) : = ( ∑ i , j = 1 d ∣ M φ ( K i j ∗ w i ) ( t , x ) ∣ 2 ) 1 / 2 , ( t , x ) ∈ [ 0 , T ] × R d . \mathcal{M}_{\varphi}(K * w)(t,x) := \left(\sum_{i,j=1}^{d}\left|\mathcal{M}_{\varphi}(K_{ij} * w_i)(t,x)\right|^2\right)^{1/2}, \qquad(t,x) \in[0,T] \times\mathbb{R}^d. M φ ( K ∗ w ) ( t , x ) := ( i , j = 1 ∑ d ∣ M φ ( K ij ∗ w i ) ( t , x ) ∣ 2 ) 1/2 , ( t , x ) ∈ [ 0 , T ] × R d .
The first part of the next result is from [6 Proposition 4.2] , and the second part is an easy consequence of [6 Theorem 3.3] and the Marcinkiewicz interpolation theorem (cf. [17 Theorem 1.3.2] ) since the operator M φ \mathcal{M}_{\varphi} M φ is sublinear.
Lemma 4.10. Let b ∈ L l o c 1 ( [ 0 , T ] × R d ; R d ) b \in L^1_{\mathrm{loc}}([0,T]\times\mathbb{R}^d;\mathbb{R}^d) b ∈ L loc 1 ([ 0 , T ] × R d ; R d ) and assume that ( H ~ ∇ b ) (\widetilde{\mathcal{H}}^{\nabla b}) ( H ∇ b ) holds. Then there exist a constant C d > 0 C_d>0 C d > 0 and a function φ ∈ C c ∞ ( R d ) \varphi\in C^\infty_c(\mathbb{R}^d) φ ∈ C c ∞ ( R d ) such that, for a.e. t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] , there is an L d \mathcal{L}^d L d -negligible set N t ⊂ R d N_t \subset\mathbb{R}^d N t ⊂ R d with
∣ b ( t , x ) − b ( t , y ) ∣ ≤ C d ∣ x − y ∣ ( M φ ( ∇ b ) ( t , x ) + M φ ( ∇ b ) ( t , y ) ) (29) |b(t,x)-b(t,y)| \le C_d|x-y|\left(\mathcal{M}_{\varphi}(\nabla b)(t,x)+\mathcal{M}_{\varphi}(\nabla b)(t,y)\right) \tag*{(29)} ∣ b ( t , x ) − b ( t , y ) ∣ ≤ C d ∣ x − y ∣ ( M φ ( ∇ b ) ( t , x ) + M φ ( ∇ b ) ( t , y ) ) (29) for all x , y ∉ N t x,y \notin N_t x , y ∈ / N t . Moreover, for every 1 < p ≤ 2 1<p\le2 1 < p ≤ 2 , there exists a constant C d , p > 0 C_{d,p}>0 C d , p > 0 such that, for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
∥ M φ ( ∇ b ) ( t , ⋅ ) ∥ L p ( R d ) ≤ C d , p ∥ w ( t , ⋅ ) ∥ L p ( R d ) , (30) \|\mathcal{M}_{\varphi}(\nabla b)(t,\cdot)\|_{L^p(\mathbb{R}^d)}\le C_{d,p}\|w(t,\cdot)\|_{L^p(\mathbb{R}^d)}, \tag*{(30)} ∥ M φ ( ∇ b ) ( t , ⋅ ) ∥ L p ( R d ) ≤ C d , p ∥ w ( t , ⋅ ) ∥ L p ( R d ) , (30) and for p = 1 p=1 p = 1 , there exists a constant C d ′ > 0 C'_d>0 C d ′ > 0 such that, for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
∥ M φ ( ∇ b ) ( t , ⋅ ) ∥ L 1 , ∞ ( R d ) ≤ C d ′ ∥ w ( t , ⋅ ) ∥ L 1 ( R d ) . (31) \|\mathcal{M}_{\varphi}(\nabla b)(t,\cdot)\|_{L^{1,\infty}(\mathbb{R}^d)}\le C'_d\|w(t,\cdot)\|_{L^1(\mathbb{R}^d)}. \tag*{(31)} ∥ M φ ( ∇ b ) ( t , ⋅ ) ∥ L 1 , ∞ ( R d ) ≤ C d ′ ∥ w ( t , ⋅ ) ∥ L 1 ( R d ) . (31)
For the application below, we use the following variant of (4.5).
Lemma 4.11. Under the assumption of Lemma 4.10, for any ϵ > 0 \epsilon>0 ϵ > 0 , there exist measurable functions G ϵ 1 , G ϵ 2 : [ 0 , T ] × R d → R + G^1_\epsilon,G^2_\epsilon:[0,T]\times\mathbb{R}^d\to\mathbb{R}_+ G ϵ 1 , G ϵ 2 : [ 0 , T ] × R d → R + and a constant C ϵ > 0 C_\epsilon>0 C ϵ > 0 such that
∥ G ϵ 1 ∥ L 1 ( [ 0 , T ] ; L 1 , ∞ ( R d ) ) ≤ ϵ , ∥ G ϵ 2 ∥ L 1 ( [ 0 , T ] ; L 2 ( R d ) ) ≤ C ϵ \|G^1_\epsilon\|_{L^1([0,T];L^{1,\infty}(\mathbb{R}^d))}\le\epsilon,\qquad\|G^2_\epsilon\|_{L^1([0,T];L^2(\mathbb{R}^d))}\le C_\epsilon ∥ G ϵ 1 ∥ L 1 ([ 0 , T ] ; L 1 , ∞ ( R d )) ≤ ϵ , ∥ G ϵ 2 ∥ L 1 ([ 0 , T ] ; L 2 ( R d )) ≤ C ϵ and moreover, for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , there exists an L d \mathcal{L}^d L d -negligible set N t ⊂ R d N_t\subset\mathbb{R}^d N t ⊂ R d such that for all x , y ∉ N t x,y\notin N_t x , y ∈ / N t ,
∣ b ( t , x ) − b ( t , y ) ∣ ≤ C d ∣ x − y ∣ ( G ϵ 1 ( t , x ) + G ϵ 1 ( t , y ) + G ϵ 2 ( t , x ) + G ϵ 2 ( t , y ) ) , (32) |b(t,x)-b(t,y)|\le C_d|x-y|\left(G^1_\epsilon(t,x)+G^1_\epsilon(t,y)+G^2_\epsilon(t,x)+G^2_\epsilon(t,y)\right), \tag*{(32)} ∣ b ( t , x ) − b ( t , y ) ∣ ≤ C d ∣ x − y ∣ ( G ϵ 1 ( t , x ) + G ϵ 1 ( t , y ) + G ϵ 2 ( t , x ) + G ϵ 2 ( t , y ) ) , (32) where C d > 0 C_d>0 C d > 0 is a constant depending only on d d d .
Proof. Fix ϵ > 0 \epsilon> 0 ϵ > 0 . Since w ∈ L 1 ( [ 0 , T ] ; L 1 ( R d ; R d ) ) w \in L^{1}([0,T];L^{1}(\mathbb{R}^{d};\mathbb{R}^{d})) w ∈ L 1 ([ 0 , T ] ; L 1 ( R d ; R d )) , by a standard truncation, write w = w 1 η + w 2 η w = w_{1}^{\eta} + w_{2}^{\eta} w = w 1 η + w 2 η with
∥ w 1 η ∥ L t 1 L x 1 ≤ η , ∥ w 2 η ∥ L t 1 L x 2 ≤ C η , \lVert w_{1}^{\eta}\rVert_{L_{t}^{1}L_{x}^{1}} \leq\eta,\qquad\lVert w_{2}^{\eta}\rVert_{L_{t}^{1}L_{x}^{2}} \leq C\eta, ∥ w 1 η ∥ L t 1 L x 1 ≤ η , ∥ w 2 η ∥ L t 1 L x 2 ≤ C η ,
where η > 0 \eta> 0 η > 0 depends on ϵ \epsilon ϵ and is chosen below. For instance, we may choose w 1 η = w 1 { ∣ w ∣ > R } w_{1}^{\eta}=w\mathbb{1}_{\{|w|>R\}} w 1 η = w 1 { ∣ w ∣ > R } for large enough R > 0 R>0 R > 0 to ensure that ∥ w 1 η ∥ L t 1 L x 1 ≤ η \lVert w_{1}^{\eta}\rVert_{L_{t}^{1}L_{x}^{1}}\leq\eta ∥ w 1 η ∥ L t 1 L x 1 ≤ η , and hence, w 2 η = w 1 { ∣ w ∣ ≤ R } w_{2}^{\eta}=w\mathbb{1}_{\{|w|\leq R\}} w 2 η = w 1 { ∣ w ∣ ≤ R } clearly satisfies the desired estimate. Define
G ϵ 1 : = M φ ( K ∗ w 1 η ) , G ϵ 2 : = M φ ( K ∗ w 2 η ) , (33) G_{\epsilon}^{1}:=\mathcal{M}_{\varphi}(K*w_{1}^{\eta}),\qquad G_{\epsilon}^{2}:=\mathcal{M}_{\varphi}(K*w_{2}^{\eta}), \tag*{(33)} G ϵ 1 := M φ ( K ∗ w 1 η ) , G ϵ 2 := M φ ( K ∗ w 2 η ) , (33)
where φ ∈ C c ∞ ( R d ) \varphi\in C_{c}^{\infty}(\mathbb{R}^{d}) φ ∈ C c ∞ ( R d ) . Then (4.8) follows from (4.5) and the sublinearity of M φ \mathcal{M}_{\varphi} M φ . Finally, by the weak ( 1 , 1 ) (1,1) ( 1 , 1 ) estimate (4.7) and the L 2 L^{2} L 2 boundedness (4.6), we may choose η > 0 \eta>0 η > 0 sufficiently small so that the stated bounds on G ϵ 1 G_{\epsilon}^{1} G ϵ 1 and G ϵ 2 G_{\epsilon}^{2} G ϵ 2 hold. □
Proof of Theorem 4.6. As in the proof of Theorem 3.3(2), we reduce to the SDEs via the superposition principle in Theorem 2.5, where the required integrability is guaranteed by our assumptions. Lemma 2.6 then allows us to assume, without loss of generality, that the two SDE solutions are coupled on a common probability space with the same Brownian motion, and with an optimal initial coupling. So, we consider the coupled SDEs driven by the same Brownian motion ( W t ) t ≥ 0 (W_t)_{t\geq0} ( W t ) t ≥ 0 :
d X t = b ( t , X t ) d t + σ ( t , X t ) d W t , L X 0 = ν , dX_t=b(t,X_t)\,dt+\sigma(t,X_t)\,dW_t,\qquad\mathcal{L}_{X_0}=\nu, d X t = b ( t , X t ) d t + σ ( t , X t ) d W t , L X 0 = ν ,
and
d X ^ t = b ^ ( t , X ^ t ) d t + σ ^ ( t , X ^ t ) d W t , L X ^ 0 = ν ^ , d\widehat{X}_t=\widehat{b}(t,\widehat{X}_t)\,dt+\widehat{\sigma}(t,\widehat{X}_t)\,dW_t,\qquad\mathcal{L}_{\widehat{X}_0}=\widehat{\nu}, d X t = b ( t , X t ) d t + σ ( t , X t ) d W t , L X 0 = ν ,
where ν = ρ 0 L d \nu=\rho_0\mathcal{L}^{d} ν = ρ 0 L d and ν ^ = ρ ^ 0 L d \widehat{\nu}=\widehat{\rho}_0\mathcal{L}^{d} ν = ρ 0 L d . The initial coupling is chosen optimally with respect to the Kantorovich–Rubinstein distance, i.e.,
E log ( 1 + ∣ X 0 − X ^ 0 ∣ 2 δ ) = D δ ( ν , ν ^ ) . (34) \mathbb{E}\log\left(1+\frac{\sqrt{|X_0-\widehat{X}_0|^2}}{\delta}\right)=D_{\delta}(\nu,\widehat{\nu}). \tag*{(34)} E log 1 + δ ∣ X 0 − X 0 ∣ 2 = D δ ( ν , ν ) . (34)
By the superposition principle again, the time marginals satisfy
L X t ( d x ) = ρ ( t , x ) d x , L X ^ t ( d x ) = ρ ^ ( t , x ) d x , t ∈ [ 0 , T ] . \mathcal{L}_{X_t}(dx)=\rho(t,x)\,dx,\qquad\mathcal{L}_{\widehat{X}_t}(dx)=\widehat{\rho}(t,x)\,dx,\qquad t\in[0,T]. L X t ( d x ) = ρ ( t , x ) d x , L X t ( d x ) = ρ ( t , x ) d x , t ∈ [ 0 , T ] .
Let α = δ 2 > 0 \alpha=\delta^2>0 α = δ 2 > 0 . By Itô’s formula,
E log ( 1 + ∣ X t − X ^ t ∣ 2 + α δ ) ≤ E log ( 1 + ∣ X 0 − X ^ 0 ∣ 2 + α δ ) + ∫ 0 T ∫ R d × R d ∣ b ( t , x ) − b ^ ( t , y ) ∣ δ + ∣ x − y ∣ π t ( d x , d y ) d t + ∫ 0 T ∫ R d × R d ∥ σ ( t , x ) − σ ^ ( t , y ) ∥ 2 δ ∣ x − y ∣ 2 + α + ∣ x − y ∣ 2 + α π t ( d x , d y ) d t , t ∈ [ 0 , T ] . (35) \begin{aligned}
\mathbb{E}\log\left(1+\frac{\sqrt{|X_t-\widehat{X}_t|^2+\alpha}}{\delta}\right)
&\leq\mathbb{E}\log\left(1+\frac{\sqrt{|X_0-\widehat{X}_0|^2+\alpha}}{\delta}\right) \\
&\quad+\int_{0}^{T}\int_{\mathbb{R}^{d}\times\mathbb{R}^{d}}
\frac{|b(t,x)-\widehat{b}(t,y)|}{\delta+|x-y|}\,\pi_t(dx,dy)\,dt \\
&\quad+\int_{0}^{T}\int_{\mathbb{R}^{d}\times\mathbb{R}^{d}}
\frac{\lVert\sigma(t,x)-\widehat{\sigma}(t,y)\rVert^{2}}
{\delta\sqrt{|x-y|^{2}+\alpha}+|x-y|^{2}+\alpha}\,\pi_t(dx,dy)\,dt,\qquad t\in[0,T]. \tag*{(35)}
\end{aligned} E log 1 + δ ∣ X t − X t ∣ 2 + α ≤ E log 1 + δ ∣ X 0 − X 0 ∣ 2 + α + ∫ 0 T ∫ R d × R d δ + ∣ x − y ∣ ∣ b ( t , x ) − b ( t , y ) ∣ π t ( d x , d y ) d t + ∫ 0 T ∫ R d × R d δ ∣ x − y ∣ 2 + α + ∣ x − y ∣ 2 + α ∥ σ ( t , x ) − σ ( t , y ) ∥ 2 π t ( d x , d y ) d t , t ∈ [ 0 , T ] . (35)
where π t = L ( X t , X ^ t ) \pi_t=\mathcal{L}(X_t,\widehat{X}_t) π t = L ( X t , X t ) for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . Set
J 1 : = ∫ 0 T ∫ R d × R d ∣ b ( t , x ) − b ^ ( t , y ) ∣ δ + ∣ x − y ∣ π t ( d x , d y ) d t , J_1:=\int_0^T\int_{\mathbb{R}^d\times\mathbb{R}^d}\frac{|b(t,x)-\widehat{b}(t,y)|}{\delta+|x-y|}\,\pi_t(dx,dy)dt, J 1 := ∫ 0 T ∫ R d × R d δ + ∣ x − y ∣ ∣ b ( t , x ) − b ( t , y ) ∣ π t ( d x , d y ) d t ,
and
J 2 : = ∫ 0 T ∫ R d × R d ∥ σ ( t , x ) − σ ^ ( t , y ) ∥ 2 δ ∣ x − y ∣ 2 + α + ∣ x − y ∣ 2 + α π t ( d x , d y ) d t . J_2:=\int_0^T\int_{\mathbb{R}^d\times\mathbb{R}^d}\frac{\|\sigma(t,x)-\widehat{\sigma}(t,y)\|^2}{\delta\sqrt{|x-y|^2+\alpha}+|x-y|^2+\alpha}\,\pi_t(dx,dy)dt. J 2 := ∫ 0 T ∫ R d × R d δ ∣ x − y ∣ 2 + α + ∣ x − y ∣ 2 + α ∥ σ ( t , x ) − σ ( t , y ) ∥ 2 π t ( d x , d y ) d t .
We first estimate J 2 J_2 J 2 . By the Morrey-type estimate in Lemma 3.10 , the triangle inequality and Hölder’s inequality,
J 2 ≤ ∫ 0 T ∫ R d × R d { ∥ σ ( t , x ) − σ ^ ( t , x ) ∥ 2 δ α + ∥ σ ^ ( t , x ) − σ ^ ( t , y ) ∥ 2 ∣ x − y ∣ 2 } π t ( d x , d y ) d t ≲ 1 δ 2 ∫ 0 T ∫ R d ∥ σ ( t , x ) − σ ^ ( t , x ) ∥ 2 ρ ( t , x ) d x d t + ∫ 0 T ( ∫ R d ∣ M ( ∇ σ ^ ) ( t , x ) ∣ 2 r d x ) 1 / r ( ∥ ρ ( t , ⋅ ) ∥ L x r ′ + ∥ ρ ^ ( t , ⋅ ) ∥ L x r ′ ) d t ≲ 1 δ 2 ∫ 0 T ∫ R d ∥ σ ( t , x ) − σ ^ ( t , x ) ∥ 2 ρ ( t , x ) d x d t + 1. (36) \begin{aligned}
J_2 &\leq\int_0^T\int_{\mathbb{R}^d\times\mathbb{R}^d}\left\{\frac{\|\sigma(t,x)-\widehat{\sigma}(t,x)\|^2}{\delta\sqrt{\alpha}}+\frac{\|\widehat{\sigma}(t,x)-\widehat{\sigma}(t,y)\|^2}{|x-y|^2}\right\}\pi_t(dx,dy)dt\\
&\lesssim\frac{1}{\delta^2}\int_0^T\int_{\mathbb{R}^d}\|\sigma(t,x)-\widehat{\sigma}(t,x)\|^2\rho(t,x)\,dxdt\\
&\quad+\int_0^T\left(\int_{\mathbb{R}^d}|\mathcal{M}(\nabla\widehat{\sigma})(t,x)|^{2r}\,dx\right)^{1/r}\left(\|\rho(t,\cdot)\|_{L_x^{r'}}+\|\widehat{\rho}(t,\cdot)\|_{L_x^{r'}}\right)dt\\
&\lesssim\frac{1}{\delta^2}\int_0^T\int_{\mathbb{R}^d}\|\sigma(t,x)-\widehat{\sigma}(t,x)\|^2\rho(t,x)\,dxdt+1. \tag*{(36)}
\end{aligned} J 2 ≤ ∫ 0 T ∫ R d × R d { δ α ∥ σ ( t , x ) − σ ( t , x ) ∥ 2 + ∣ x − y ∣ 2 ∥ σ ( t , x ) − σ ( t , y ) ∥ 2 } π t ( d x , d y ) d t ≲ δ 2 1 ∫ 0 T ∫ R d ∥ σ ( t , x ) − σ ( t , x ) ∥ 2 ρ ( t , x ) d x d t + ∫ 0 T ( ∫ R d ∣ M ( ∇ σ ) ( t , x ) ∣ 2 r d x ) 1/ r ( ∥ ρ ( t , ⋅ ) ∥ L x r ′ + ∥ ρ ( t , ⋅ ) ∥ L x r ′ ) d t ≲ δ 2 1 ∫ 0 T ∫ R d ∥ σ ( t , x ) − σ ( t , x ) ∥ 2 ρ ( t , x ) d x d t + 1. (36)
where r ′ r' r ′ is the conjugate number of r r r , and the last inequality is due to the assumption that ρ , ρ ^ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) \rho,\widehat{\rho}\in L_t^\infty(L_x^1\cap L_x^\infty) ρ , ρ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) and ( ∑ i , j , k = 1 d ∣ ∂ k σ ^ i j ∣ 2 ) 1 / 2 ∈ L t 2 L x 2 r \left(\sum_{i,j,k=1}^d|\partial_k\widehat{\sigma}_{ij}|^2\right)^{1/2}\in L_t^2L_x^{2r} ( ∑ i , j , k = 1 d ∣ ∂ k σ ij ∣ 2 ) 1/2 ∈ L t 2 L x 2 r , and the strong ( 2 r , 2 r ) (2r,2r) ( 2 r , 2 r ) boundedness of M \mathcal{M} M .
We next estimate J 1 J_1 J 1 . By the triangle inequality,
J 1 ≤ 1 δ ∫ 0 T ∫ R d ∣ b ( t , x ) − b ^ ( t , x ) ∣ ρ ( t , x ) d x d t + ∫ 0 T ∫ R d × R d ∣ b ^ ( t , x ) − b ^ ( t , y ) ∣ δ + ∣ x − y ∣ π t ( d x , d y ) d t = : 1 δ ∫ 0 T ∫ R d ∣ b ( t , x ) − b ^ ( t , x ) ∣ ρ ( t , x ) d x d t + J 1 , 1 . (37) \begin{aligned}
J_1&\leq\frac{1}{\delta}\int_0^T\int_{\mathbb{R}^d}|b(t,x)-\widehat{b}(t,x)|\rho(t,x)\,dxdt\\
&\quad+\int_0^T\int_{\mathbb{R}^d\times\mathbb{R}^d}\frac{|\widehat{b}(t,x)-\widehat{b}(t,y)|}{\delta+|x-y|}\,\pi_t(dx,dy)dt\\
&=: \frac{1}{\delta}\int_0^T\int_{\mathbb{R}^d}|b(t,x)-\widehat{b}(t,x)|\rho(t,x)\,dxdt+J_{1,1}. \tag*{(37)}
\end{aligned} J 1 ≤ δ 1 ∫ 0 T ∫ R d ∣ b ( t , x ) − b ( t , x ) ∣ ρ ( t , x ) d x d t + ∫ 0 T ∫ R d × R d δ + ∣ x − y ∣ ∣ b ( t , x ) − b ( t , y ) ∣ π t ( d x , d y ) d t =: δ 1 ∫ 0 T ∫ R d ∣ b ( t , x ) − b ( t , x ) ∣ ρ ( t , x ) d x d t + J 1 , 1 . (37)
By (32) with b b b replaced by b ^ \widehat{b} b , for any ϵ > 0 \epsilon>0 ϵ > 0 ,
∣ b ^ ( t , x ) − b ^ ( t , y ) ∣ δ + ∣ x − y ∣ ≲ min { ∣ b ^ ( t , x ) − b ^ ( t , y ) ∣ δ , G ϵ 1 ( t , x ) + G ϵ 1 ( t , y ) } + G ϵ 2 ( t , x ) + G ϵ 2 ( t , y ) \frac{|\widehat{b}(t,x)-\widehat{b}(t,y)|}{\delta+|x-y|}
\lesssim\min\left\{\frac{|\widehat{b}(t,x)-\widehat{b}(t,y)|}{\delta},G_\epsilon^1(t,x)+G_\epsilon^1(t,y)\right\}+G_\epsilon^2(t,x)+G_\epsilon^2(t,y) δ + ∣ x − y ∣ ∣ b ( t , x ) − b ( t , y ) ∣ ≲ min { δ ∣ b ( t , x ) − b ( t , y ) ∣ , G ϵ 1 ( t , x ) + G ϵ 1 ( t , y ) } + G ϵ 2 ( t , x ) + G ϵ 2 ( t , y )
≲ min { ∣ b ^ ( t , x ) ∣ δ , G ϵ 1 ( t , x ) } + min { ∣ b ^ ( t , y ) ∣ δ , G ϵ 1 ( t , y ) } + min { ∣ b ^ ( t , y ) ∣ δ , G ϵ 1 ( t , x ) } + min { ∣ b ^ ( t , x ) ∣ δ , G ϵ 1 ( t , y ) } + G ϵ 2 ( t , x ) + G ϵ 2 ( t , y ) , \begin{aligned}
\lesssim\min\left\{\frac{|\widehat{b}(t,x)|}{\delta},G_\epsilon^1(t,x)\right\}+\min\left\{\frac{|\widehat{b}(t,y)|}{\delta},G_\epsilon^1(t,y)\right\} \\
&+ \min\left\{\frac{|\hat{b}(t,y)|}{\delta},G_{\epsilon}^{1}(t,x)\right\}+\min\left\{\frac{|\hat{b}(t,x)|}{\delta},G_{\epsilon}^{1}(t,y)\right\}
\\
+G_{\epsilon}^{2}(t,x)+G_{\epsilon}^{2}(t,y),
\end{aligned} ≲ min { δ ∣ b ( t , x ) ∣ , G ϵ 1 ( t , x ) } + min { δ ∣ b ( t , y ) ∣ , G ϵ 1 ( t , y ) } + G ϵ 2 ( t , x ) + G ϵ 2 ( t , y ) , + min { δ ∣ b ^ ( t , y ) ∣ , G ϵ 1 ( t , x ) } + min { δ ∣ b ^ ( t , x ) ∣ , G ϵ 1 ( t , y ) }
where the functions G ϵ 1 G_{\epsilon}^{1} G ϵ 1 and G ϵ 2 G_{\epsilon}^{2} G ϵ 2 are defined similarly as (4.9). It is therefore enough to estimate the representative terms.
Set
η 1 ( t , x ) : = min { ∣ b ^ ( t , x ) ∣ δ , G ϵ 1 ( t , x ) } , η 2 ( t , x , y ) : = min { ∣ b ^ ( t , y ) ∣ δ , G ϵ 1 ( t , x ) } , \eta_{1}(t,x):=\min\left\{\frac{|\hat{b}(t,x)|}{\delta},G_{\epsilon}^{1}(t,x)\right\},\qquad\eta_{2}(t,x,y):=\min\left\{\frac{|\hat{b}(t,y)|}{\delta},G_{\epsilon}^{1}(t,x)\right\}, η 1 ( t , x ) := min { δ ∣ b ^ ( t , x ) ∣ , G ϵ 1 ( t , x ) } , η 2 ( t , x , y ) := min { δ ∣ b ^ ( t , y ) ∣ , G ϵ 1 ( t , x ) } ,
and
η 3 ( t , x ) : = G ϵ 2 ( t , x ) . \eta_{3}(t,x):=G_{\epsilon}^{2}(t,x). η 3 ( t , x ) := G ϵ 2 ( t , x ) .
Define a finite measure μ \mu μ on Ω T : = [ 0 , T ] × R d × R d \Omega_{T}:=[0,T]\times\mathbb{R}^{d}\times\mathbb{R}^{d} Ω T := [ 0 , T ] × R d × R d by
μ ( d t , d x , d y ) = 1 [ 0 , T ] ( t ) π t ( d x , d y ) d t . \mu(\mathrm{d}t,\mathrm{d}x,\mathrm{d}y)=\mathbb{1}_{[0,T]}(t)\pi_{t}(\mathrm{d}x,\mathrm{d}y)\mathrm{d}t. μ ( d t , d x , d y ) = 1 [ 0 , T ] ( t ) π t ( d x , d y ) d t .
Since ρ , ρ ^ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) \rho,\hat{\rho}\in L_{t}^{\infty}(L_{x}^{1}\cap L_{x}^{\infty}) ρ , ρ ^ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) , by Lemma 4.11,
∥ η 3 ∥ L 1 ( Ω T , μ ) ≲ ∥ G ϵ 2 ∥ L t 1 L x 2 ≤ C ϵ . (38) \|\eta_{3}\|_{L^{1}(\Omega_{T},\mu)}\lesssim\|G_{\epsilon}^{2}\|_{L_{t}^{1}L_{x}^{2}}\leq C\epsilon. \tag*{(38)} ∥ η 3 ∥ L 1 ( Ω T , μ ) ≲ ∥ G ϵ 2 ∥ L t 1 L x 2 ≤ C ϵ . (38)
We now estimate η 1 \eta_{1} η 1 and η 2 \eta_{2} η 2 . Since η 1 ≤ G ϵ 1 \eta_{1}\leq G_{\epsilon}^{1} η 1 ≤ G ϵ 1 and ρ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) \rho\in L_{t}^{\infty}(L_{x}^{1}\cap L_{x}^{\infty}) ρ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) , by Lemma 4.11,
∥ η 1 ∥ L 1 , ∞ ( Ω T , μ ) = sup λ > 0 λ μ ( { ( t , x , y ) ∈ Ω T : ∣ η 1 ( t , x ) ∣ > λ } ) ≤ sup λ > 0 λ ∫ 0 T ∫ { x : ∣ G ϵ 1 ( t , x ) ∣ > λ } ρ ( t , x ) d x d t ≤ ∥ ρ ∥ L ∞ ( [ 0 , T ] × R d ) sup λ > 0 λ ∫ 0 T L d ( { x : ∣ G ϵ 1 ( t , x ) ∣ > λ } ) d t \begin{aligned}
\|\eta_{1}\|_{L^{1,\infty}(\Omega_{T},\mu)}
=\sup_{\lambda>0}\lambda\mu\left(\{(t,x,y)\in\Omega_{T}:|\eta_{1}(t,x)|>\lambda\}\right) \\
&\leq\sup_{\lambda>0}\lambda\int_{0}^{T}\int_{\{x:|G_{\epsilon}^{1}(t,x)|>\lambda\}}\rho(t,x)\,\mathrm{d}x\mathrm{d}t \\
&\leq\|\rho\|_{L^{\infty}([0,T]\times\mathbb{R}^{d})}\sup_{\lambda>0}\lambda\int_{0}^{T}\mathcal{L}^{d}\left(\{x:|G_{\epsilon}^{1}(t,x)|>\lambda\}\right)\mathrm{d}t
\end{aligned} ∥ η 1 ∥ L 1 , ∞ ( Ω T , μ ) = λ > 0 sup λ μ ( {( t , x , y ) ∈ Ω T : ∣ η 1 ( t , x ) ∣ > λ } ) ≤ λ > 0 sup λ ∫ 0 T ∫ { x : ∣ G ϵ 1 ( t , x ) ∣ > λ } ρ ( t , x ) d x d t ≤ ∥ ρ ∥ L ∞ ([ 0 , T ] × R d ) λ > 0 sup λ ∫ 0 T L d ( { x : ∣ G ϵ 1 ( t , x ) ∣ > λ } ) d t
≲ ∥ G ϵ 1 ∥ L 1 ( [ 0 , T ] ; L 1 , ∞ ( R d ) ) ≲ ϵ . \lesssim\|G_{\epsilon}^{1}\|_{L^{1}([0,T];L^{1,\infty}(\mathbb{R}^{d}))}\lesssim\epsilon. ≲ ∥ G ϵ 1 ∥ L 1 ([ 0 , T ] ; L 1 , ∞ ( R d )) ≲ ϵ .
Similarly,
∥ η 2 ∥ L 1 , ∞ ( Ω T , μ ) ≲ ϵ . \|\eta_{2}\|_{L^{1,\infty}(\Omega_{T},\mu)}\lesssim\epsilon. ∥ η 2 ∥ L 1 , ∞ ( Ω T , μ ) ≲ ϵ .
On the other hand, for every λ > 0 \lambda>0 λ > 0 ,
μ ( { ( t , x , y ) ∈ Ω T : ∣ η 1 ( t , x ) ∣ > λ } ) ≤ μ ( { ( t , x , y ) ∈ Ω T : ∣ b ^ ( t , x ) ∣ δ > λ } ) = ∫ { ( t , x , y ) ∈ Ω T : ∣ b ^ ( t , x ) ∣ δ > λ } π t ( d x , d y ) d t ≤ ∫ { ( t , x ) ∈ [ 0 , T ] × R d : ∣ b ^ ( t , x ) ∣ δ > λ } ρ ( t , x ) d x d t ≤ ∥ ρ ∥ L ∞ ( [ 0 , T ] × R d ) ( L 1 ⊗ L d ) ( { ( t , x ) ∈ [ 0 , T ] × R d : ∣ b ^ ( t , x ) ∣ δ > λ } ) . \begin{aligned}
\mu\left(\{(t,x,y)\in\Omega_{T}:|\eta_{1}(t,x)|>\lambda\}\right) \\
&\leq\mu\left(\left\{(t,x,y)\in\Omega_{T}:\frac{|\hat{b}(t,x)|}{\delta}>\lambda\right\}\right) \\
&=\int_{\left\{(t,x,y)\in\Omega_{T}:\frac{|\hat{b}(t,x)|}{\delta}>\lambda\right\}}\pi_{t}(\mathrm{d}x,\mathrm{d}y)\mathrm{d}t \\
&\leq\int_{\left\{(t,x)\in[0,T]\times\mathbb{R}^{d}:\frac{|\hat{b}(t,x)|}{\delta}>\lambda\right\}}\rho(t,x)\,\mathrm{d}x\mathrm{d}t \\
&\leq\|\rho\|_{L^{\infty}([0,T]\times\mathbb{R}^{d})}(\mathcal{L}^{1}\otimes\mathcal{L}^{d})\left(\left\{(t,x)\in[0,T]\times\mathbb{R}^{d}:\frac{|\hat{b}(t,x)|}{\delta}>\lambda\right\}\right).
\end{aligned} μ ( {( t , x , y ) ∈ Ω T : ∣ η 1 ( t , x ) ∣ > λ } ) ≤ μ ( { ( t , x , y ) ∈ Ω T : δ ∣ b ^ ( t , x ) ∣ > λ } ) = ∫ { ( t , x , y ) ∈ Ω T : δ ∣ b ^ ( t , x ) ∣ > λ } π t ( d x , d y ) d t ≤ ∫ { ( t , x ) ∈ [ 0 , T ] × R d : δ ∣ b ^ ( t , x ) ∣ > λ } ρ ( t , x ) d x d t ≤ ∥ ρ ∥ L ∞ ([ 0 , T ] × R d ) ( L 1 ⊗ L d ) ( { ( t , x ) ∈ [ 0 , T ] × R d : δ ∣ b ^ ( t , x ) ∣ > λ } ) .
Hence
∥ η 1 ∥ L p , ∞ ( Ω T , μ ) ≲ 1 δ ∥ b ^ ∥ L p , ∞ ( [ 0 , T ] × R d ; L 1 ⊗ L d ) ≲ 1 δ . \|\eta_1\|_{L^{p,\infty}(\Omega_T,\mu)} \lesssim\frac{1}{\delta}\|\hat b\|_{L^{p,\infty}([0,T]\times\mathbb{R}^d;L^1\otimes\mathcal{L}^d)} \lesssim\frac{1}{\delta}. ∥ η 1 ∥ L p , ∞ ( Ω T , μ ) ≲ δ 1 ∥ b ^ ∥ L p , ∞ ([ 0 , T ] × R d ; L 1 ⊗ L d ) ≲ δ 1 .
and similarly,
∥ η 2 ∥ L p , ∞ ( Ω T , μ ) ≲ 1 δ . \|\eta_2\|_{L^{p,\infty}(\Omega_T,\mu)} \lesssim\frac{1}{\delta}. ∥ η 2 ∥ L p , ∞ ( Ω T , μ ) ≲ δ 1 .
Combining the estimates on ∥ η i ∥ L 1 , ∞ ( Ω T , μ ) \|\eta_i\|_{L^{1,\infty}(\Omega_T,\mu)} ∥ η i ∥ L 1 , ∞ ( Ω T , μ ) and ∥ η i ∥ L p , ∞ ( Ω T , μ ) \|\eta_i\|_{L^{p,\infty}(\Omega_T,\mu)} ∥ η i ∥ L p , ∞ ( Ω T , μ ) , i = 1 , 2 i=1,2 i = 1 , 2 , by the interpolation inequality (2.4), we obtain
∥ η 1 ∥ L 1 ( Ω T , μ ) + ∥ η 2 ∥ L 1 ( Ω T , μ ) ≲ ϵ [ 1 + log ( 1 ϵ δ ) ] . (39) \|\eta_1\|_{L^1(\Omega_T,\mu)}+\|\eta_2\|_{L^1(\Omega_T,\mu)} \lesssim\epsilon\left[1+\log\left(\frac{1}{\epsilon\delta}\right)\right]. \tag*{(39)} ∥ η 1 ∥ L 1 ( Ω T , μ ) + ∥ η 2 ∥ L 1 ( Ω T , μ ) ≲ ϵ [ 1 + log ( ϵδ 1 ) ] . (39)
Combining the preceding estimates (4.13), (4.14) and (4.15) together, we derive
J 1 ≲ 1 δ ∫ 0 T ∫ R d ∣ b ( t , x ) − b ^ ( t , x ) ∣ ρ ( t , x ) d x d t + ϵ [ 1 + log ( 1 ϵ δ ) ] + C ϵ . (40) J_1\lesssim\frac{1}{\delta}\int_0^T\int_{\mathbb{R}^d}|b(t,x)-\hat b(t,x)|\rho(t,x)\,\mathrm{d}x\mathrm{d}t+\epsilon\left[1+\log\left(\frac{1}{\epsilon\delta}\right)\right]+C_\epsilon. \tag*{(40)} J 1 ≲ δ 1 ∫ 0 T ∫ R d ∣ b ( t , x ) − b ^ ( t , x ) ∣ ρ ( t , x ) d x d t + ϵ [ 1 + log ( ϵδ 1 ) ] + C ϵ . (40)
By (4.10) and the elementary inequality log ( 1 + a 2 + δ 2 / δ ) ≤ log 2 + log ( 1 + a / δ ) \log(1+\sqrt{a^2+\delta^2}/\delta)\le\log2+\log(1+a/\delta) log ( 1 + a 2 + δ 2 / δ ) ≤ log 2 + log ( 1 + a / δ ) for any a ≥ 0 a\ge0 a ≥ 0 , we have
E log ( 1 + ∣ X 0 − X ^ 0 ∣ 2 + δ 2 δ ) ≤ D δ ( ν , ν ^ ) + log 2. (41) \mathbb{E}\log\left(1+\frac{\sqrt{|X_0-\hat X_0|^2+\delta^2}}{\delta}\right)\le D_\delta(\nu,\hat\nu)+\log2. \tag*{(41)} E log 1 + δ ∣ X 0 − X ^ 0 ∣ 2 + δ 2 ≤ D δ ( ν , ν ^ ) + log 2. (41)
Therefore, the desired estimate follows by combining (4.17), (4.11) with the estimates (4.16) on J 1 J_1 J 1 and (4.12) on J 2 J_2 J 2 , since [ 0 , ∞ ) ∋ α ↦ E log [ 1 + ( ∣ X t − X ^ t ∣ 2 + α ) 1 / 2 / δ ] [0,\infty)\ni\alpha\mapsto\mathbb{E}\log[1+(|X_t-\hat X_t|^2+\alpha)^{1/2}/\delta] [ 0 , ∞ ) ∋ α ↦ E log [ 1 + ( ∣ X t − X ^ t ∣ 2 + α ) 1/2 / δ ] is increasing.
An application to well-posedness
In this subsection, we prove well-posedness of weak solutions to the FPE. Since the condition on ∇ b \nabla b ∇ b is insufficient for existence, we impose an additional structural assumption on b b b itself.
Assumption 4.12. For each j = 1 , … , d j=1,\ldots,d j = 1 , … , d , let Φ j : R d → R \Phi_j:\mathbb{R}^d\to\mathbb{R} Φ j : R d → R be a homogeneous function of degree 1 − d 1-d 1 − d , i.e.,
Φ j ( λ x ) = λ 1 − d Φ j ( x ) , λ > 0 , x ∈ R d , \Phi_j(\lambda x)=\lambda^{1-d}\Phi_j(x),\qquad\lambda>0,\ x\in\mathbb{R}^d, Φ j ( λ x ) = λ 1 − d Φ j ( x ) , λ > 0 , x ∈ R d , which satisfies the following conditions:
( Φ 1 ) (\Phi_1) ( Φ 1 ) Φ j ∣ R d ∖ { 0 } ∈ C 2 ( R d ∖ { 0 } ) \Phi_j|_{\mathbb{R}^d\setminus\{0\}}\in C^2(\mathbb{R}^d\setminus\{0\}) Φ j ∣ R d ∖ { 0 } ∈ C 2 ( R d ∖ { 0 }) ;
( Φ 2 ) (\Phi_2) ( Φ 2 ) there exists a constant C > 0 C>0 C > 0 such that for every integer 0 ≤ m ≤ 2 0\le m\le2 0 ≤ m ≤ 2 ,
∣ ∇ m Φ j ( x ) ∣ ≤ C ∣ x ∣ d − 1 + m , x ≠ 0 , |\nabla^m\Phi_j(x)|\le\frac{C}{|x|^{d-1+m}},\qquad x\ne0, ∣ ∇ m Φ j ( x ) ∣ ≤ ∣ x ∣ d − 1 + m C , x = 0 , where ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ is understood as the Euclidean norm if m = 0 , 1 m=0,1 m = 0 , 1 , and the Hilbert–Schmidt norm if m = 2 m=2 m = 2 ;
The drift vector field b = ( b 1 , … , b d ) b=(b_1,\ldots,b_d) b = ( b 1 , … , b d ) is defined componentwise by
b j ( t , x ) : = { Φ j ∗ w j ( t , ⋅ ) } ( x ) , t ∈ [ 0 , T ] , x ∈ R d , b_j(t,x):=\{\Phi_j*w_j(t,\cdot)\}(x),\qquad t\in[0,T],\ x\in\mathbb{R}^d, b j ( t , x ) := { Φ j ∗ w j ( t , ⋅ )} ( x ) , t ∈ [ 0 , T ] , x ∈ R d ,
where w = ( w 1 , … , w d ) ∈ L ∞ ( [ 0 , T ] ; L 1 ( R d ; R d ) ) w=(w_1,\ldots,w_d)\in L^\infty([0,T];L^1(\mathbb{R}^d;\mathbb{R}^d)) w = ( w 1 , … , w d ) ∈ L ∞ ([ 0 , T ] ; L 1 ( R d ; R d )) .
The main result in this subsection is contained in the following theorem.
Theorem 4.14. Assume that Assumption 4.12, ( H σ ) (\mathcal{H}^{\sigma}) ( H σ ) , ( H b σ ) (\mathcal{H}^{b^\sigma}) ( H b σ ) , and ( H ν ) (\mathcal{H}^{\nu}) ( H ν ) hold. Then the FPE (2) admits a unique weak solution in the class L ∞ ( [ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d ) ) L^\infty([0,T];L^1(\mathbb{R}^d)\cap L^\infty(\mathbb{R}^d)) L ∞ ([ 0 , T ] ; L 1 ( R d ) ∩ L ∞ ( R d )) .
Before proving the theorem, we present the following basic regularity.
Lemma 4.15. Let h ∈ L ∞ ( [ 0 , T ] ; L 1 ( R d ) ) h\in L^\infty([0,T];L^1(\mathbb{R}^d)) h ∈ L ∞ ([ 0 , T ] ; L 1 ( R d )) and let ψ : R d → R \psi:\mathbb{R}^d\to\mathbb{R} ψ : R d → R satisfy
∣ ψ ( x ) ∣ ≤ C ∣ x ∣ d − 1 , x ≠ 0 , |\psi(x)|\leq\frac{C}{|x|^{d-1}},\qquad x\neq0, ∣ ψ ( x ) ∣ ≤ ∣ x ∣ d − 1 C , x = 0 , for some constant C > 0 C>0 C > 0 . For ( t , x ) ∈ [ 0 , T ] × R d (t,x)\in[0,T]\times\mathbb{R}^d ( t , x ) ∈ [ 0 , T ] × R d , set V ( t , x ) : = [ h ( t , ⋅ ) ∗ ψ ] ( x ) V(t,x):=[h(t,\cdot)*\psi](x) V ( t , x ) := [ h ( t , ⋅ ) ∗ ψ ] ( x ) . Then
V ∈ L p , ∞ ( [ 0 , T ] × R d ) , p : = d d − 1 . (42) V\in L^{p,\infty}([0,T]\times\mathbb{R}^d),\qquad p:=\frac{d}{d-1}. \tag*{(42)} V ∈ L p , ∞ ([ 0 , T ] × R d ) , p := d − 1 d . (42)
Proof. By the assumption on ψ \psi ψ , we have the pointwise estimate
∣ V ( t , x ) ∣ ≤ ∫ R d C ∣ h ( t , y ) ∣ ∣ x − y ∣ d − 1 d y , t ∈ [ 0 , T ] , |V(t,x)| \le\int_{\mathbb{R}^{d}} \frac{C|h(t,y)|}{|x-y|^{d-1}}\,\mathrm{d}y,\qquad t\in[0,T], ∣ V ( t , x ) ∣ ≤ ∫ R d ∣ x − y ∣ d − 1 C ∣ h ( t , y ) ∣ d y , t ∈ [ 0 , T ] ,
where C > 0 C>0 C > 0 is a constant. For any R > 0 R>0 R > 0 and a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
∣ V ( t , x ) ∣ ≤ C ∫ B R ( x ) ∣ h ( t , y ) ∣ ∣ x − y ∣ d − 1 d y + C ∫ R d ∖ B R ( x ) ∣ h ( t , y ) ∣ ∣ x − y ∣ d − 1 d y ≤ C ( 1 B R 1 ∣ ⋅ ∣ d − 1 ∗ ∣ h ( t , ⋅ ) ∣ ( x ) ) + C N R d − 1 , \begin{aligned}
|V(t,x)| \le C\int_{B_R(x)}\frac{|h(t,y)|}{|x-y|^{d-1}}\,\mathrm{d}y+C\int_{\mathbb{R}^{d}\setminus B_R(x)}\frac{|h(t,y)|}{|x-y|^{d-1}}\,\mathrm{d}y \\
&\le C\left(\mathbb{1}_{B_R}\frac{1}{|\cdot|^{d-1}}*|h(t,\cdot)|(x)\right)+\frac{CN}{R^{d-1}},
\end{aligned} ∣ V ( t , x ) ∣ ≤ C ∫ B R ( x ) ∣ x − y ∣ d − 1 ∣ h ( t , y ) ∣ d y + C ∫ R d ∖ B R ( x ) ∣ x − y ∣ d − 1 ∣ h ( t , y ) ∣ d y ≤ C ( 1 B R ∣ ⋅ ∣ d − 1 1 ∗ ∣ h ( t , ⋅ ) ∣ ( x ) ) + R d − 1 C N ,
for some constant N > 0 N>0 N > 0 , where the last inequality is due to that ∥ h ( t , ⋅ ) ∥ L 1 ≤ N \|h(t,\cdot)\|_{L^1}\le N ∥ h ( t , ⋅ ) ∥ L 1 ≤ N for a.e. t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] since h ∈ L ∞ ( [ 0 , T ] ; L 1 ( R d ) ) h\in L^\infty([0,T];L^1(\mathbb{R}^d)) h ∈ L ∞ ([ 0 , T ] ; L 1 ( R d )) .
Let λ > 0 \lambda>0 λ > 0 be arbitrary and choose R = ( 2 C N / λ ) 1 / ( d − 1 ) R=(2CN/\lambda)^{1/(d-1)} R = ( 2 C N / λ ) 1/ ( d − 1 ) , so that C N / R d − 1 = λ / 2 CN/R^{d-1}=\lambda/2 C N / R d − 1 = λ /2 . Letting μ = L 1 ∣ [ 0 , T ] ⊗ L d \mu=\mathcal{L}^1|_{[0,T]}\otimes\mathcal{L}^d μ = L 1 ∣ [ 0 , T ] ⊗ L d , by Chebyshev’s inequality, we derive
μ ( { ∣ V ∣ > λ } ) ≤ μ ( { ( 1 B R 1 ∣ ⋅ ∣ d − 1 ∗ ∣ h ∣ ) > λ 2 C } ) ≤ 2 C λ ∫ 0 T ∫ R d ( 1 B R 1 ∣ ⋅ ∣ d − 1 ∗ ∣ h ( t , ⋅ ) ∣ ) ( x ) d x d t . \begin{aligned}
\mu(\{|V|>\lambda\})\le\mu\left(\left\{\left(\mathbb{1}_{B_R}\frac{1}{|\cdot|^{d-1}}*|h| \right)>\frac{\lambda}{2C}\right\}\right) \\
&\le\frac{2C}{\lambda}\int_0^T\int_{\mathbb{R}^d}\left(\mathbb{1}_{B_R}\frac{1}{|\cdot|^{d-1}}*|h(t,\cdot)|\right)(x)\,\mathrm{d}x\mathrm{d}t.
\end{aligned} μ ({ ∣ V ∣ > λ }) ≤ μ ( { ( 1 B R ∣ ⋅ ∣ d − 1 1 ∗ ∣ h ∣ ) > 2 C λ } ) ≤ λ 2 C ∫ 0 T ∫ R d ( 1 B R ∣ ⋅ ∣ d − 1 1 ∗ ∣ h ( t , ⋅ ) ∣ ) ( x ) d x d t .
By Young’s convolution inequality and the computation
∥ 1 B R 1 ∣ ⋅ ∣ d − 1 ∥ L 1 ( R d ) = c d R , \left\|\mathbb{1}_{B_R}\frac{1}{|\cdot|^{d-1}}\right\|_{L^1(\mathbb{R}^d)}=c_dR, 1 B R ∣ ⋅ ∣ d − 1 1 L 1 ( R d ) = c d R ,
we obtain
μ ( { ∣ V ∣ > λ } ) ≤ 2 C λ ⋅ ( c d R ) ⋅ ( T N ) ≤ C ′ λ − d d − 1 , \mu(\{|V|>\lambda\})\le\frac{2C}{\lambda}\cdot(c_dR)\cdot(TN)\le C'\lambda^{-\frac{d}{d-1}}, μ ({ ∣ V ∣ > λ }) ≤ λ 2 C ⋅ ( c d R ) ⋅ ( T N ) ≤ C ′ λ − d − 1 d ,
for some constant C ′ > 0 C'>0 C ′ > 0 , where c d c_d c d is the volume of the unit ball in R d \mathbb{R}^d R d . Therefore, with
p = d d − 1 , p=\frac{d}{d-1}, p = d − 1 d ,
sup λ > 0 λ p μ ( { ∣ V ∣ > λ } ) < ∞ , \sup_{\lambda>0}\lambda^p\mu(\{|V|>\lambda\})<\infty, λ > 0 sup λ p μ ({ ∣ V ∣ > λ }) < ∞ ,
which proves (42) .
The next lemma computes the distributional derivative of b b b .
Lemma 4.16 (Distributional derivative of b b b ). Let Assumption 4.12 hold. Then, for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , the distributional derivative of b ( t , ⋅ ) b(t,\cdot) b ( t , ⋅ ) is given by
∂ j b i ( t , x ) = [ ∂ j Φ i ∗ w i ( t , ⋅ ) ] ( x ) − c i j w i ( t , x ) in D ′ ( R d ) , (43) \partial_j b_i(t,x)=[\partial_j\Phi_i*w_i(t,\cdot)](x)-c_{ij}w_i(t,x)\qquad\text{in }\mathcal{D}'(\mathbb{R}^d), \tag*{(43)} ∂ j b i ( t , x ) = [ ∂ j Φ i ∗ w i ( t , ⋅ )] ( x ) − c ij w i ( t , x ) in D ′ ( R d ) , (43) where the convolution is understood in the distributional sense of (27) and c = ( c i j ) 1 ≤ i , j ≤ d c=(c_{ij})_{1\le i,j\le d} c = ( c ij ) 1 ≤ i , j ≤ d is the constant matrix with
c i j = ∫ ∣ x ∣ = 1 Φ i ( x ) x j d ω ( x ) , 1 ≤ i , j ≤ d . c_{ij}=\int_{|x|=1}\Phi_i(x)x_j\,\mathrm{d}\omega(x),\qquad1\le i,j\le d. c ij = ∫ ∣ x ∣ = 1 Φ i ( x ) x j d ω ( x ) , 1 ≤ i , j ≤ d .
Proof. Fix i , j ∈ { 1 , … , d } i,j \in\{1,\ldots,d\} i , j ∈ { 1 , … , d } and take an arbitrary φ ∈ C ∞ ( R d ) \varphi\in C^\infty(\mathbb{R}^d) φ ∈ C ∞ ( R d ) . By assumptions ( Φ 1 ) (\Phi_1) ( Φ 1 ) and ( Φ 2 ) (\Phi_2) ( Φ 2 ) with m = 0 m=0 m = 0 , we have Φ i ∈ L l o c 1 ( R d ) \Phi_i \in L^1_{\mathrm{loc}}(\mathbb{R}^d) Φ i ∈ L loc 1 ( R d ) . For every ϵ > 0 \epsilon>0 ϵ > 0 , applying Green’s formula on { ∣ x ∣ ≥ ϵ } \{|x|\ge\epsilon\} { ∣ x ∣ ≥ ϵ } yields
∫ ∣ x ∣ ≥ ϵ Φ i ( x ) ∂ j φ ( x ) d x = − ∫ ∣ x ∣ ≥ ϵ φ ( x ) ∂ j Φ i ( x ) d x + ∫ ∣ x ∣ = ϵ Φ i ( x ) φ ( x ) x j ∣ x ∣ d ω ( x ) . (44) \int_{|x|\ge\epsilon}\Phi_i(x)\partial_j\varphi(x)\,\mathrm{d}x
=
-\int_{|x|\ge\epsilon}\varphi(x)\partial_j\Phi_i(x)\,\mathrm{d}x
+
\int_{|x|=\epsilon}\Phi_i(x)\varphi(x)\frac{x_j}{|x|}\,\mathrm{d}\omega(x). \tag*{(44)} ∫ ∣ x ∣ ≥ ϵ Φ i ( x ) ∂ j φ ( x ) d x = − ∫ ∣ x ∣ ≥ ϵ φ ( x ) ∂ j Φ i ( x ) d x + ∫ ∣ x ∣ = ϵ Φ i ( x ) φ ( x ) ∣ x ∣ x j d ω ( x ) . (44)
Since Φ i ∂ j φ ∈ L 1 ( R d ) \Phi_i\partial_j\varphi\in L^1(\mathbb{R}^d) Φ i ∂ j φ ∈ L 1 ( R d ) , it is clear that letting ϵ → 0 + \epsilon\to0^+ ϵ → 0 + , the left-hand side of (44) converges to ∫ R d Φ i ( x ) ∂ j φ ( x ) d x \int_{\mathbb{R}^d}\Phi_i(x)\partial_j\varphi(x)\,\mathrm{d}x ∫ R d Φ i ( x ) ∂ j φ ( x ) d x . For the boundary term, changing variables x = ϵ z x=\epsilon z x = ϵz with ∣ z ∣ = 1 |z|=1 ∣ z ∣ = 1 and using the homogeneity Φ i ( ϵ z ) = ϵ 1 − d Φ i ( z ) \Phi_i(\epsilon z)=\epsilon^{1-d}\Phi_i(z) Φ i ( ϵz ) = ϵ 1 − d Φ i ( z ) , we obtain
∫ ∣ x ∣ = ϵ Φ i ( x ) φ ( x ) x j ∣ x ∣ d ω ( x ) = ∫ ∣ z ∣ = 1 Φ i ( z ) φ ( ϵ z ) z j d ω ( z ) ⟶ φ ( 0 ) ∫ ∣ z ∣ = 1 Φ i ( z ) z j d ω ( z ) , as ϵ → 0 + . \int_{|x|=\epsilon}\Phi_i(x)\varphi(x)\frac{x_j}{|x|}\,\mathrm{d}\omega(x)
=
\int_{|z|=1}\Phi_i(z)\varphi(\epsilon z)z_j\,\mathrm{d}\omega(z)
\longrightarrow
\varphi(0)\int_{|z|=1}\Phi_i(z)z_j\,\mathrm{d}\omega(z),
\qquad\text{as }\epsilon\to0^+. ∫ ∣ x ∣ = ϵ Φ i ( x ) φ ( x ) ∣ x ∣ x j d ω ( x ) = ∫ ∣ z ∣ = 1 Φ i ( z ) φ ( ϵz ) z j d ω ( z ) ⟶ φ ( 0 ) ∫ ∣ z ∣ = 1 Φ i ( z ) z j d ω ( z ) , as ϵ → 0 + .
where convergence follows from the dominated convergence theorem, since φ ( ϵ z ) → φ ( 0 ) \varphi(\epsilon z)\to\varphi(0) φ ( ϵz ) → φ ( 0 ) uniformly and Φ i ( z ) z j \Phi_i(z)z_j Φ i ( z ) z j is integrable on { ∣ z ∣ = 1 } \{|z|=1\} { ∣ z ∣ = 1 } . Consequently, the limit of the first integral on the right-hand side of (44) exists, and it is precisely the Cauchy principal value
p.v. ∫ R d φ ( x ) ∂ j Φ i ( x ) d x : = lim ϵ → 0 + ∫ ∣ x ∣ ≥ ϵ φ ( x ) ∂ j Φ i ( x ) d x . \operatorname{p.v.}\int_{\mathbb{R}^d}\varphi(x)\partial_j\Phi_i(x)\,\mathrm{d}x
:=
\lim_{\epsilon\to0^+}\int_{|x|\ge\epsilon}\varphi(x)\partial_j\Phi_i(x)\,\mathrm{d}x. p.v. ∫ R d φ ( x ) ∂ j Φ i ( x ) d x := ϵ → 0 + lim ∫ ∣ x ∣ ≥ ϵ φ ( x ) ∂ j Φ i ( x ) d x .
Thus,
∫ R d Φ i ( x ) ∂ j φ ( x ) d x = − p.v. ∫ R d φ ( x ) ∂ j Φ i ( x ) d x + c i j φ ( 0 ) , \int_{\mathbb{R}^d}\Phi_i(x)\partial_j\varphi(x)\,\mathrm{d}x
=
-\operatorname{p.v.}\int_{\mathbb{R}^d}\varphi(x)\partial_j\Phi_i(x)\,\mathrm{d}x
+
c_{ij}\varphi(0), ∫ R d Φ i ( x ) ∂ j φ ( x ) d x = − p.v. ∫ R d φ ( x ) ∂ j Φ i ( x ) d x + c ij φ ( 0 ) ,
which in the sense of distributions reads
∂ j Φ i = p.v. ( ∂ j Φ i ) − c i j δ 0 . (45) \partial_j\Phi_i=\operatorname{p.v.}(\partial_j\Phi_i)-c_{ij}\delta_0. \tag*{(45)} ∂ j Φ i = p.v. ( ∂ j Φ i ) − c ij δ 0 . (45)
where δ 0 \delta_0 δ 0 is the Dirac mass at the origin, and
c i j : = ∫ ∣ x ∣ = 1 Φ i ( x ) x j d ω ( x ) . c_{ij}:=\int_{|x|=1}\Phi_i(x)x_j\,\mathrm{d}\omega(x). c ij := ∫ ∣ x ∣ = 1 Φ i ( x ) x j d ω ( x ) .
Fix t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . Recall that
⟨ ∂ j Φ i ∗ w i ( t , ⋅ ) , φ ⟩ : = ⟨ w i ( t , ⋅ ) , ∂ j Φ i ~ ∗ φ ⟩ , \langle\partial_j\Phi_i * w_i(t,\cdot),\varphi\rangle
:=
\langle w_i(t,\cdot),\widetilde{\partial_j\Phi_i} * \varphi\rangle, ⟨ ∂ j Φ i ∗ w i ( t , ⋅ ) , φ ⟩ := ⟨ w i ( t , ⋅ ) , ∂ j Φ i ∗ φ ⟩ ,
where
∂ j Φ i ~ ( x ) : = ∂ j Φ i ( − x ) , x ∈ R d . \widetilde{\partial_j\Phi_i}(x):=\partial_j\Phi_i(-x),
\qquad x\in\mathbb{R}^d. ∂ j Φ i ( x ) := ∂ j Φ i ( − x ) , x ∈ R d .
Using (45) , we get
⟨ b i ( t , ⋅ ) , ∂ j φ ⟩ = ∫ R d ∫ R d Φ i ( x − y ) w i ( t , y ) ∂ j φ ( x ) d y d x = ∫ R d ( ∫ R d Φ i ( x − y ) ∂ j φ ( x ) d x ) w i ( t , y ) d y = ∫ R d ( − p . v . ∫ R d φ ( x ) ∂ j Φ i ( x − y ) d x + c i j ⟨ δ 0 ( ⋅ − y ) , φ ⟩ ) w i ( t , y ) d y = − ⟨ ∂ j Φ i ∗ w i ( t , ⋅ ) , φ ⟩ + ⟨ c i j w i ( t , ⋅ ) , φ ⟩ . \begin{aligned}
\langle b_i(t,\cdot),\partial_j\varphi\rangle
=
\int_{\mathbb{R}^d}\int_{\mathbb{R}^d}
\Phi_i(x-y)w_i(t,y)\partial_j\varphi(x)\,\mathrm{d}y\mathrm{d}x \\
&= \int_{\mathbb{R}^d} \left(\int_{\mathbb{R}^d} \Phi_i(x-y)\partial_j\varphi(x)\,dx\right)w_i(t,y)\,dy \\
&= \int_{\mathbb{R}^d} \left(-\mathop{\mathrm{p.v.}}\int_{\mathbb{R}^d}\varphi(x)\partial_j\Phi_i(x-y)\,dx+c_{ij}\langle\delta_0(\cdot-y),\varphi\rangle\right)w_i(t,y)\,dy \\
&=-\langle\partial_j\Phi_i\ast w_i(t,\cdot),\varphi\rangle+\langle c_{ij}w_i(t,\cdot),\varphi\rangle.
\end{aligned} ⟨ b i ( t , ⋅ ) , ∂ j φ ⟩ = ∫ R d ∫ R d Φ i ( x − y ) w i ( t , y ) ∂ j φ ( x ) d y d x = ∫ R d ( ∫ R d Φ i ( x − y ) ∂ j φ ( x ) d x ) w i ( t , y ) d y = ∫ R d ( − p.v. ∫ R d φ ( x ) ∂ j Φ i ( x − y ) d x + c ij ⟨ δ 0 ( ⋅ − y ) , φ ⟩ ) w i ( t , y ) d y = − ⟨ ∂ j Φ i ∗ w i ( t , ⋅ ) , φ ⟩ + ⟨ c ij w i ( t , ⋅ ) , φ ⟩ .
Therefore
∂ j b i ( t , x ) = { ∂ j Φ i ∗ w i ( t , ⋅ ) } ( x ) − c i j w i ( t , x ) in D ′ ( R d ) , \partial_jb_i(t,x)=\{\partial_j\Phi_i\ast w_i(t,\cdot)\}(x)-c_{ij}w_i(t,x)\quad\text{in }\mathcal{D}'(\mathbb{R}^d), ∂ j b i ( t , x ) = { ∂ j Φ i ∗ w i ( t , ⋅ )} ( x ) − c ij w i ( t , x ) in D ′ ( R d ) ,
which proves the lemma.
The zeroth-order term on the right-hand side of (4.19), which originates from the Dirac mass, is more regular than the singular integral, and it is fully controlled by the estimates for the singular part. Hence, for notational convenience and without loss of generality, we write
∇ b ( t , x ) = { ∇ Φ ∗ w ( t , ⋅ ) } ( x ) , in D ′ ( R d ; R d × d ) , \nabla b(t,x)=\{\nabla\Phi\ast w(t,\cdot)\}(x),\quad\text{in }\mathcal{D}'(\mathbb{R}^d;\mathbb{R}^{d\times d}), ∇ b ( t , x ) = { ∇Φ ∗ w ( t , ⋅ )} ( x ) , in D ′ ( R d ; R d × d ) ,
for any t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , where the convolution is understood in the sense of ( H ~ ∇ b ) (\widetilde{\mathcal{H}}^{\nabla}b) ( H ∇ b ) .
Proof of Theorem 4.14. By Corollary 4.9, weak solutions in the class L t ∞ ( L x 1 ∩ L x ∞ ) L^\infty_t(L^1_x\cap L^\infty_x) L t ∞ ( L x 1 ∩ L x ∞ ) are unique. It therefore remains only to prove existence.
Let ( φ ϵ ) ϵ ∈ ( 0 , 1 ) (\varphi^\epsilon)_{\epsilon\in(0,1)} ( φ ϵ ) ϵ ∈ ( 0 , 1 ) be a standard mollifier defined by (3.2). Set
b ϵ : = b ∗ φ ϵ , a ϵ : = a ∗ φ ϵ ∈ L 1 ( [ 0 , T ] ; C b ∞ ( R d ; R d × d ) ) , ρ 0 ϵ : = ρ 0 ∗ φ ϵ ∈ C b ∞ ( R d ) . b^\epsilon:=b\ast\varphi^\epsilon,\qquad a^\epsilon:=a\ast\varphi^\epsilon\in L^1([0,T];C_b^\infty(\mathbb{R}^d;\mathbb{R}^{d\times d})),\qquad\rho_0^\epsilon:=\rho_0\ast\varphi^\epsilon\in C_b^\infty(\mathbb{R}^d). b ϵ := b ∗ φ ϵ , a ϵ := a ∗ φ ϵ ∈ L 1 ([ 0 , T ] ; C b ∞ ( R d ; R d × d )) , ρ 0 ϵ := ρ 0 ∗ φ ϵ ∈ C b ∞ ( R d ) .
By Lemma 4.15 and the interpolation inequality (2.3),
b ∈ L l o c q ( [ 0 , T ] × R d ; R d ) , 1 ≤ q ≤ d d − 1 . b\in L^q_{\mathrm{loc}}([0,T]\times\mathbb{R}^d;\mathbb{R}^d),\qquad1\leq q\leq\frac{d}{d-1}. b ∈ L loc q ([ 0 , T ] × R d ; R d ) , 1 ≤ q ≤ d − 1 d .
Hence b ϵ ( t , ⋅ ) ∈ C ∞ ( R d ; R d ) b^\epsilon(t,\cdot)\in C^\infty(\mathbb{R}^d;\mathbb{R}^d) b ϵ ( t , ⋅ ) ∈ C ∞ ( R d ; R d ) for every ϵ ∈ ( 0 , 1 ) \epsilon\in(0,1) ϵ ∈ ( 0 , 1 ) and t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) . Moreover, Assumption 4.12 gives
b ϵ ∈ L 1 ( [ 0 , T ] ; C b ∞ ( R d ; R d ) ) . b^\epsilon\in L^1([0,T];C_b^\infty(\mathbb{R}^d;\mathbb{R}^d)). b ϵ ∈ L 1 ([ 0 , T ] ; C b ∞ ( R d ; R d )) .
Repeating the proof of Theorem 3.3(1), it remains only to verify that
sup 0 < ϵ < 1 ∫ 0 T ∫ R d ∣ b ϵ ( s , x ) ∣ ρ ϵ ( s , x ) d x d s < ∞ . \sup_{0<\epsilon<1}\int_0^T\int_{\mathbb{R}^d}|b^\epsilon(s,x)|\rho^\epsilon(s,x)\,dx\,ds<\infty. 0 < ϵ < 1 sup ∫ 0 T ∫ R d ∣ b ϵ ( s , x ) ∣ ρ ϵ ( s , x ) d x d s < ∞.
Set
μ ϵ ( d t , d x ) : = 1 [ 0 , T ] ( t ) ρ ϵ ( t , x ) d x d t , μ : = L 1 ∣ [ 0 , T ] ⊗ L d . \mu^\epsilon(dt,dx):=\mathbf{1}_{[0,T]}(t)\rho^\epsilon(t,x)\,dx\,dt,\qquad\mu:=\mathcal{L}^1|_{[0,T]}\otimes\mathcal{L}^d. μ ϵ ( d t , d x ) := 1 [ 0 , T ] ( t ) ρ ϵ ( t , x ) d x d t , μ := L 1 ∣ [ 0 , T ] ⊗ L d .
By (2.3) and the uniform boundedness of ρ ϵ \rho^\epsilon ρ ϵ , we obtain, for p = d d − 1 p=\frac{d}{d-1} p = d − 1 d ,
∥ b ϵ ∥ L 1 ( [ 0 , T ] × R d ; μ ϵ ) ≤ p p − 1 T 1 − 1 / p ∥ b ϵ ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ϵ ) ≲ ∥ b ϵ ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) . \|b^\epsilon\|_{L^1([0,T]\times\mathbb{R}^d;\mu^\epsilon)}
\leq\frac{p}{p-1}T^{1-1/p}\|b^\epsilon\|_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu^\epsilon)}
\lesssim\|b^\epsilon\|_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}. ∥ b ϵ ∥ L 1 ([ 0 , T ] × R d ; μ ϵ ) ≤ p − 1 p T 1 − 1/ p ∥ b ϵ ∥ L p , ∞ ([ 0 , T ] × R d ; μ ϵ ) ≲ ∥ b ϵ ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) .
It remains to control the weak L p L^p L p quasi-norm of b ϵ b^\epsilon b ϵ . We claim that
∥ b ϵ ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) ≲ ∥ b ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) , ϵ ∈ ( 0 , 1 ) . \lVert b^\epsilon\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)} \lesssim\lVert b\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}, \qquad\epsilon\in(0,1). ∥ b ϵ ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) ≲ ∥ b ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) , ϵ ∈ ( 0 , 1 ) .
Indeed, for λ > 0 \lambda>0 λ > 0 , write
b = b 1 + b 2 , b 1 : = b 1 { ∣ b ∣ > λ / 2 } , b 2 : = b 1 { ∣ b ∣ ≤ λ / 2 } . b=b_1+b_2,\qquad b_1:=b\mathbf{1}_{\{|b|>\lambda/2\}},\qquad b_2:=b\mathbf{1}_{\{|b|\le\lambda/2\}}. b = b 1 + b 2 , b 1 := b 1 { ∣ b ∣ > λ /2 } , b 2 := b 1 { ∣ b ∣ ≤ λ /2 } .
Let b 1 ϵ = b 1 ∗ ϕ ϵ b_1^\epsilon=b_1\ast\phi^\epsilon b 1 ϵ = b 1 ∗ ϕ ϵ and b 2 ϵ = b 2 ∗ ϕ ϵ b_2^\epsilon=b_2\ast\phi^\epsilon b 2 ϵ = b 2 ∗ ϕ ϵ . Then ∣ b 2 ϵ ∣ ≤ λ / 2 \lvert b_2^\epsilon\rvert\le\lambda/2 ∣ b 2 ϵ ∣ ≤ λ /2 , and hence
{ ∣ b ϵ ∣ > λ } ⊂ { ∣ b 1 ϵ ∣ > λ / 2 } . \{|b^\epsilon|>\lambda\}\subset\{|b_1^\epsilon|>\lambda/2\}. { ∣ b ϵ ∣ > λ } ⊂ { ∣ b 1 ϵ ∣ > λ /2 } .
Thus,
μ ( { ∣ b ϵ ∣ > λ } ) ≤ 2 λ ∥ b 1 ϵ ∥ L t , x 1 ≤ 2 λ ∥ b 1 ∥ L t , x 1 . \mu(\{|b^\epsilon|>\lambda\})\le\frac{2}{\lambda}\lVert b_1^\epsilon\rVert_{L^1_{t,x}}\le\frac{2}{\lambda}\lVert b_1\rVert_{L^1_{t,x}}. μ ({ ∣ b ϵ ∣ > λ }) ≤ λ 2 ∥ b 1 ϵ ∥ L t , x 1 ≤ λ 2 ∥ b 1 ∥ L t , x 1 .
Since b ∈ L p , ∞ ( [ 0 , T ] × R d ; R d ) b\in L^{p,\infty}([0,T]\times\mathbb{R}^d;\mathbb{R}^d) b ∈ L p , ∞ ([ 0 , T ] × R d ; R d ) with p = d d − 1 > 1 p=\frac{d}{d-1}>1 p = d − 1 d > 1 by Lemma 4.15, applying the layer-cake formula, we obtain
∥ b 1 ∥ L t , x 1 = ∫ 0 ∞ μ ( { ∣ b 1 ∣ > s } ) d s = ∫ 0 λ / 2 μ ( { ∣ b 1 ∣ > s } ) d s + ∫ λ / 2 ∞ μ ( { ∣ b 1 ∣ > s } ) d s = λ 2 μ ( { ∣ b ∣ > λ 2 } ) + ∫ λ / 2 ∞ μ ( { ∣ b ∣ > s } ) d s ≤ λ 2 ∥ b ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) p ( λ 2 ) − p + ∥ b ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) p ∫ λ / 2 ∞ s − p d s = ∥ b ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) p ( λ 2 ) 1 − p + 1 p − 1 ∥ b ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) p ( λ 2 ) 1 − p = p 2 p − 1 p − 1 ∥ b ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) p λ 1 − p . \begin{aligned}
\lVert b_1\rVert_{L^1_{t,x}}
&=\int_0^\infty\mu(\{|b_1|>s\})\,\mathrm{d}s\\
&=\int_0^{\lambda/2}\mu(\{|b_1|>s\})\,\mathrm{d}s+\int_{\lambda/2}^\infty\mu(\{|b_1|>s\})\,\mathrm{d}s\\
&=\frac{\lambda}{2}\mu\left(\left\{|b|>\frac{\lambda}{2}\right\}\right)+\int_{\lambda/2}^\infty\mu(\{|b|>s\})\,\mathrm{d}s\\
&\le\frac{\lambda}{2}\lVert b\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}^p\left(\frac{\lambda}{2}\right)^{-p}
+\lVert b\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}^p\int_{\lambda/2}^\infty s^{-p}\,\mathrm{d}s\\
&=\lVert b\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}^p\left(\frac{\lambda}{2}\right)^{1-p}
+\frac{1}{p-1}\lVert b\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}^p\left(\frac{\lambda}{2}\right)^{1-p}\\
&=\frac{p2^{p-1}}{p-1}\lVert b\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}^p\lambda^{1-p}.
\end{aligned} ∥ b 1 ∥ L t , x 1 = ∫ 0 ∞ μ ({ ∣ b 1 ∣ > s }) d s = ∫ 0 λ /2 μ ({ ∣ b 1 ∣ > s }) d s + ∫ λ /2 ∞ μ ({ ∣ b 1 ∣ > s }) d s = 2 λ μ ( { ∣ b ∣ > 2 λ } ) + ∫ λ /2 ∞ μ ({ ∣ b ∣ > s }) d s ≤ 2 λ ∥ b ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) p ( 2 λ ) − p + ∥ b ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) p ∫ λ /2 ∞ s − p d s = ∥ b ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) p ( 2 λ ) 1 − p + p − 1 1 ∥ b ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) p ( 2 λ ) 1 − p = p − 1 p 2 p − 1 ∥ b ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) p λ 1 − p .
It follows that
μ ( { ∣ b ϵ ∣ > λ } ) ≤ p 2 p p − 1 ∥ b ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) p λ − p . \mu(\{|b^\epsilon|>\lambda\})\le\frac{p2^p}{p-1}\lVert b\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}^p\lambda^{-p}. μ ({ ∣ b ϵ ∣ > λ }) ≤ p − 1 p 2 p ∥ b ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) p λ − p .
Equivalently,
λ [ μ ( { ∣ b ϵ ∣ > λ } ) ] 1 / p ≤ 2 ( p p − 1 ) 1 / p ∥ b ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) . \lambda\left[\mu(\{|b^\epsilon|>\lambda\})\right]^{1/p}\le2\left(\frac{p}{p-1}\right)^{1/p}\lVert b\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}. λ [ μ ({ ∣ b ϵ ∣ > λ }) ] 1/ p ≤ 2 ( p − 1 p ) 1/ p ∥ b ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) .
Taking the supremum over λ > 0 \lambda>0 λ > 0 gives
∥ b ϵ ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) ≤ 2 ( p p − 1 ) 1 / p ∥ b ∥ L p , ∞ ( [ 0 , T ] × R d ; μ ) . \lVert b^\epsilon\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}\le2\left(\frac{p}{p-1}\right)^{1/p}\lVert b\rVert_{L^{p,\infty}([0,T]\times\mathbb{R}^d;\mu)}. ∥ b ϵ ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) ≤ 2 ( p − 1 p ) 1/ p ∥ b ∥ L p , ∞ ([ 0 , T ] × R d ; μ ) .
Arguing as in the proof of Theorem 3.3 (1), we obtain
ρ ∈ L t ∞ ( L x 1 ∩ L x ∞ ) \rho\in L^\infty_t(L^1_x \cap L^\infty_x) ρ ∈ L t ∞ ( L x 1 ∩ L x ∞ )
and ρ ( t , ⋅ ) \rho(t,\cdot) ρ ( t , ⋅ ) is a probability density for a.e. t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] . Moreover, ρ \rho ρ satisfies (6) . Since the drift b b b belongs only to a weak Lebesgue space, it remains to verify the integrability condition (5) . As σ \sigma σ is uniformly bounded, it suffices to show that
∥ b ρ ∥ L 1 ( [ 0 , T ] × R d ; R d ) < ∞ . \|b\rho\|_{L^1([0,T]\times\mathbb{R}^d;\mathbb{R}^d)} < \infty. ∥ b ρ ∥ L 1 ([ 0 , T ] × R d ; R d ) < ∞.
This follows from (42) and the interpolation inequality (7) . The proof is complete.
Acknowledgment
The authors gratefully acknowledge the financial support of the National Key R&D Program of China (No. 2022YFA1006000).
Data Availability Statement
Data sharing is not applicable to this article as no datasets were generated or analysed during the current study.
Declarations
Conflict of Interests. The authors declare that they have no conflict of interest.
Declaration of generative AI and AI-assisted technologies in the manuscript preparation process. During the preparation of this work, the authors used ChatGPT (OpenAI) solely for language editing and polishing. It was not used for research design, mathematical analysis, proofs, interpretation of results, or conclusions. The authors reviewed and edited all relevant content and take full responsibility for the published article.