Abstract
This paper includes in an unified way several results about existence and uniqueness of solutions of Fokker–Planck equations from (Bogachev et al., J Funct Anal, 256:1269–1298, 2009) [2], (Bogachev et al., J Evol Equ, 10(3):487–509, 2010) [3], (Bogachev et al., Partial Differ Equ, 36:925–939, 2011) [4] and (Bogachev et al., Bull London Math Soc 39:631–640, 2007) [1], using probabilistic methods. Several applications are provided including Burgers and 2D-Navier–Stokes equations perturbed by noise. Some of these applications were also studied by a different analytic approach in (Bogachev et al., J Differ Equ, 259(8):3854–3873, 2015) [5], (Bogachev et al., Ann Sc Norm Super Pisa Cl Sci 14(3):983–1023, 2015) [6], (Da Prato et al., Commun Math Stat, 1(3):281–304, 2013) [11].
Access provided by CONRICYT-eBooks. Download conference paper PDF
Similar content being viewed by others
Keywords
2000 Mathematics Subject Classification AMS
1 Introduction and Setting of the Problem
We are given a separable Hilbert space H with norm \(|\cdot |\) and inner product \(\langle \cdot , \cdot \rangle .\) We denote by \(\mathcal B(H)\) the set of all Borel subsets of H and by \( \mathcal P(H)\) the set of all Borel probability measures on H. If \(\psi :H\rightarrow [0,+\infty ]\) is convex and lower semi-continuous we denote by \(\mathcal P_\psi (H)\) the set of all elements \(\zeta \in \mathcal P(H)\) such that \( \int _H \psi (x)\zeta (dx)<\infty . \) Clearly \(\mathcal P_\psi (H)\) is a convex subset of \(\mathcal P(H)\). We shall assume throughout the paper the following
Hypothesis 1
(i) \(A:D(A)\subset H\rightarrow H\) is self-adjoint and there exists \(\omega >0\) such that \(\langle Ax,x \rangle \le -\omega |x|^2,\;\forall \;x\in D(A).\)
(ii) There exists \(\epsilon _0\in (0,1/2)\) such that \((-A)^{-1+2\epsilon _0}\) is of trace class.
(iii) \(C:H\rightarrow H\) is bounded, symmetric and nonnegative.
(iv) \(F:D(F)\subset [0,T]\times H\rightarrow H\) is Borel, whereas D(F) is a Borel subset of \([0,T]\times H\).
We have taken A and C time independent and A self-adjoint for the sake of simplicity, but more general cases (including: A infinitesimal generator of a \(C_0\) semigroup and time dependent and equations with multiplicative noise) could be considered without important changes in the proofs.
For any \(\beta \in (0,1)\) we shall use the notation \( \Vert x\Vert _{\beta }:=|(-A)^\beta x|,\;\forall \;x\in D((-A)^\beta ). \) We recall that for any \(\beta >0\) the embedding \(D((-A)^\beta )\subset H\) is compact and that there exists \(k_\beta >0\) such that
A probability kernel \( [0,T]\rightarrow \mathcal P(H),\; t\mapsto \mu _t, \) is a mapping \( [0,T]\rightarrow \mathbb R,\; t\mapsto \mu _t(I) \) such that is measurable for any \(I\in \mathcal B(H)\). We identify a probability kernel \((\mu _t)_{t\in [0,T]} \) with the measure \(\underline{\mu }(dx\,dt)=\mu _t(dx)dt\) on \(([0,T]\times H, \mathcal B([0,T])\times H))\) defined as
Let us introduce the space of exponential functions; they will play the role of test functions. For any \(h\in H\) we set \(\varphi _h=e^{i\langle h,x \rangle },\; x\in H.\) Moreover, we denote by \(\mathcal E_A(H)\) the linear span of the real parts of functions \( \varphi _h\) such that \(h\in D(A)\) and by \(\mathcal E_A([0,T]\times H)\) the linear span of all functions of the form \(u(t,x)=g(t)\varphi (x),\;(t,x)\in [0,T]\times H,\) where \(g\in C^1([0,T]), \, g(T)=0\) and \(\varphi \in \mathcal E_A(H)\). Finally, we define the Kolmogorov operator \(K_F\), by setting
for any \(u\in \mathcal E_A([0,T]\times H)\) and \((t,x)\in [0,T]\times H. \)
We consider the following problem. Given \(\zeta \in \mathcal P(H)\), we want to find a probability kernel \(\underline{\mu }=(\mu _t)_{t\in [0,T]}\) such that \(\mu _0=\zeta \) and for all \(u\in \mathcal E_A([0,T]\times H)\), \(K_F u\) is \(\underline{\mu }\)-integrable and it results
where \(H_T=[0,T]\times H.\) \(\underline{\mu }\) will be called a solution of the Fokker–Planck equation (3).
Let us describe the content of the paper. Section 2 is devoted to some preliminaries about the Ornstein–Uhlenbeck semigroup, following [2]. In Sect. 3 we present some existence result for the Fokker–Planck equation (3) under additional assumptions, Hypotheses 2 or 3. The first one is closely related to the paper [3], whereas the slightly different second one allows us to consider more general nonlinearities as Burgers and 2D-Navier–Stokes equations. Finally, Sect. 4 is devoted to uniqueness. Here we start from a basic rank condition, introduced in [1] and distinguish two cases: (i) C is invertible and \(C^{-1}\) is bounded and (ii) C is of trace class. We note that an important consequence of uniqueness is that in this case the Chapman–Kolmogorov equation is well posed, see [3] for a thorough discussion of this fact.
We end this section with some notations used in what follows. By \(C_b(H)\) we mean the space of all real continuous and bounded mappings \(\varphi :H\rightarrow \mathbb R\) endowed with the sup norm \(\Vert \cdot \Vert _{0}\). Moreover, \(C^1_b(H)\) is the subspace of \(C_b(H)\) of all continuously differentiable functions with bounded derivatives. Finally, \(B_b(H)\) is the space of all real Borel mappings \(\varphi :H\rightarrow \mathbb R\) endowed with the sup norm \(\Vert \cdot \Vert _{0}\).
2 Preliminaries on the Ornstein–Uhlenbeck Semigroup
It is well known (see e.g. [8] or [13]) that under Hypothesis 1 the equation
has a unique mild solution given by
where the stochastic convolution \(W_A(t,s)\) is defined as
The following lemma will be used later
Lemma 1
For all \(\beta \in [0,\epsilon _0]\) and \(0\le s\le t\le T\) we have
Proof
We have in fact
\(\square \)
Let us consider the transition evolution operator \(R_{s,t}\), corresponding to Eq. (4), which acts, in particular, on \(C_{b,1}(H)\), the space of all real continuous mappings \(\varphi \) in H such that \(\sup _{x\in H}\,\tfrac{|\varphi (x)|}{1+|x|}<\infty \). For \(0\le s\le t\le T\) we have
Here \(N_{e^{(t-s)A}x,Q_{t-s}}\) represents the Gaussian measure in H with mean \(e^{(t-s)A}x\) and covariance
See e.g. [8]. Let us define a semigroup \(S^{0,1}_\tau ,\,\tau \ge 0,\) in the space
by setting
Note that \(S^{0,1}_\tau =0,\; \forall \;\tau \ge T.\) Let denote by \(\mathcal K^1_0\)Footnote 1 the infinitesimal generator of \(S^{0,1}_\tau \), defined through its resolvent following [7] (see (10) below). The resolvent set of \(\mathcal K^1_0\) coincides with \(\mathbb R\) and its resolvent is given, for all \(f\in C_T([0,T];C_{b,1}(H))\) and all \(\lambda \in \mathbb R\), by
\(\mathcal K^1_0\) can also be defined as follows, following [14]. We say that a function \(u\in C_T([0,T];C_{b,1}(H))\) belongs to the domain of \(\mathcal K^1_0\) if there exists a function \(f\in C_T([0,T];C_{b,1}(H))\) such that
(i)
\(\displaystyle \lim _{\tau \rightarrow 0}\frac{1}{\tau }\;(S^{0,1}_\tau u(t,x)-u(t,x))=f(t,x),\quad \forall \;(t,x)\in [0,T]\times H.\)
(ii) There exists \(C_u>0\) such that
We set \(\mathcal K^1_0 u=f\) and call \(\mathcal K^1_0\) the infinitesimal generator of \(S_\tau \) on \(C_T([0,T];C_{b,1}(H))\). It is not difficult to check that \(\mathcal E_A([0,T]\times H)\subset D(\mathcal K^1_0)\) and that for all \(u\in \mathcal E_A([0,T]\times H)\) it results \(\mathcal K^1_0 \,u=K_0\,u\). So, the abstract operator \(\mathcal K^1_0\) is an extension of the differential Kolmogorov operator \(K_0\). If, in particular, \(u(t,x)=e^{i\langle x,h \rangle }\,g(t),\) \(t\in [0,T],\) \(x\in H,\) where \(h\in D(A),\) \(g\in C^1([0,T])\) and \(g(T)=0\), we have
So, \(K_0u\) has a linear growth in x; this is the reason for introducing the space \(C_{b,1}(H)\) and for requiring that \(h\in D(A)\).
One can show that functions from \(C_T([0,T];C_{b,1}(H))\) can be approximated point-wise by sequences (more precisely multi-sequences) of elements of \(\mathcal E_A([0,T]\times H).\) To describe this approximation, we shall use the following notation
The following three propositions were proven in [2], see also [9].
Proposition 2
For all \(u\in C_T([0,T];C_{b,1}(H))\) there exists \((u_{\mathbf {n}})\subset \mathcal E_A([0,T]\times H)\) such that
(i) \(\displaystyle \lim _{\mathbf {n}\rightarrow \infty }u_{\mathbf {n}}(t,x)=u(t,x),\quad \forall \;(t,x)\in [0,T]\times H.\)
(ii) \(|u_{\mathbf {n}}(t,x)|\le \Vert u\Vert _{C_T([0,T];C_{b,1}(H))}(1+|x|),\quad \forall \;(t,x)\in [0,T]\times H. \)
Proposition 3
For any \(u\in D(\mathcal K^1_0)\) there exists \((u_{\mathbf {n}})\subset \mathcal E_A([0,T]\times H)\) and \(C>0\) such that
(i) \(\displaystyle \lim _{\mathbf {n}\rightarrow \infty }u_{\mathbf {n}}(t,x)=u(t,x),\quad \forall \;(t,x)\in [0,T]\times H.\)
(ii) \(\displaystyle \lim _{\mathbf {n}\rightarrow \infty }K_0\,u_{\mathbf {n}}(t, x)=\mathcal K^1_0\,u(t, x),\quad \forall \;(t,x)\in [0,T]\times H.\)
(iii) \(\displaystyle |u_{\mathbf {n}}(x)|+| K_0u_{\mathbf {n}}(t, x)|\le C(1+|x|),\quad \forall \;(t,x)\in [0,T]\times H\)
We say that \(\mathcal E_A([0,T]\times H)\) is a core for \(\mathcal K^1_0\).
Proposition 4
Let \(f, D_xf\in C_T([0,T];C_{b,1}(H;H))\), \(\lambda \in \mathbb R\) and let \(u=R(\lambda ,\mathcal K^1_0)f\). Then there exists \(u_{\mathbf {n}}\subset \mathcal E_A([0,T]\times H) \) and \(C>0\) such that
(i) \(\displaystyle \lim _{\mathbf {n}\rightarrow \infty }u_{\mathbf {n}}(t,x)=u(t,x),\quad \forall \;(t,x)\in [0,T]\times H.\)
(ii) \(\displaystyle \lim _{\mathbf {n}\rightarrow \infty }K_0\,u_{\mathbf {n}}(t,x)=\mathcal K^1_0 u(t,x),\quad \forall \;(t,x)\in [0,T]\times H.\)
(iii) \(\displaystyle \lim _{\mathbf {n}\rightarrow \infty }D_xu_{\mathbf {n}}(t,x)=D_xu(t,x),\quad \forall \;(t,x)\in [0,T]\times H.\)
(iv) \(\displaystyle |u_{\mathbf {n}}(t,x)|+|K_0u_{\mathbf {n}}(t,x)|+|D_x\varphi _{\mathbf {n}}(t,x)|\le C(1+|x|), \forall \,(t,x)\in [0,T]{\times } H.\)
3 Existence
3.1 Basic Assumptions
We shall first introduce a suitable approximation \(F_\alpha \) of F, \(\alpha \in (0,1]\), such that the problem
has a unique mild solution \(X_\alpha (t,s,x)\), that is
More precisely, we shall assume that
Hypothesis 2
For any \(\alpha \in (0,1]\) there exists a mapping \( F_\alpha :[0,T]\times H\rightarrow H, (t,x)\rightarrow F_\alpha (t,x) \) with the following properties:
(i) \( F_\alpha \) is continuous and bounded and Eq. (11) has a unique mild solution.
(ii) There exists a convex and lower semicontinuous mapping \(V:H\rightarrow [1,+\infty ]\) such that
and
(iii) There exists \(M>0\) such that
where \(P^\alpha \) is the transition evolution operator, \(P^\alpha _{s,t}\,\varphi (x)=\mathbb E[\varphi (X_\alpha (t,s,x))]\), \(0\le s\le t\le T,\) \(\varphi \in B_b(H).\)
Now, fix \(\zeta \in \mathcal P(H);\) then for any \(t\in (0,T]\) and any \(\alpha \in (0,1]\) we define a probability measure \(\mu _t^\alpha \) in H such that (we take here \(s=0\) for simplicity)
and set \( \underline{\mu }^\alpha (dt\,dx)=\mu _t^\alpha (dx)dt.\) Let us deduce some consequences of Hypothesis 2 assuming that \(\int _HV(x)\,\zeta (dx)<\infty \). First from (15) and (16) we find
and so, from (13)
Other useful estimates are provided by the following lemma.
Lemma 5
Assume that Hypotheses 1 and 2 are fulfilled. Let \(x\in H\), \(0<s\le t\le T\) and \(\alpha \in (0,1]\); then \(X_\alpha (t,s,x)\in D((-A)^{\epsilon _0})\), \(\mathbb P\)-a.s.Footnote 2 Moreover there exists \(C>0\) such that
and
Proof
Let us write \(X_\alpha (t,s,x)=X_\alpha (t)\) for short. Then by (12) and Hölder’s inequality we have
Now, taking expectation, recalling (7) and invoking (13) we obtain
Finally, taking into account (15), yields
which implies (19) with a suitable \(C>0\). To prove (20) write
from which, recalling (1) and using that \(|F_\alpha (r,X_\alpha (r))|\le 1+|F_\alpha (r,X_\alpha (r))|^2\), we have
Taking expectation and proceeding as before, we obtain
which easily yields (20). \(\square \)
In the following we set
\(\psi \) is convex and lower semicontinuous.
Corollary 6
Assume that Hypotheses 1 and 2 are fulfilled and let \(\zeta \in \mathcal P_\psi (H)\). Then we have
and
Proof
The proof follows by integrating both sides of (19) and (20) (with \(s=0\)) with respect to \(\zeta \) over H. \(\square \)
We notice now that by Itô’s formula the measure \(\underline{\mu }^\alpha (dt\,dx)=\mu ^\alpha _t(dx)\) solves the approximating equation
where \(K_{F_\alpha }\) is the Kolmogorov operator
We are going to construct a sequence \((\mu ^{\alpha _h}_t)\) of measures weakly convergent to a measure \(\mu _t\) for all \(t\in [0,T]\) and finally, we shall pass to the limit as \(h\rightarrow \infty \) in (24) (with \(\alpha _h\) replacing \(\alpha \)) and prove that \(\underline{\mu }(dx\,dt)=\mu _t(dx)\,dt\) is a solution of the Fokker–Planck equation (3).
3.2 Tightness
Proposition 7
Assume that \(\zeta \in \mathcal P_{\psi }(H)\). Then \((\mu _t^\alpha )_{\alpha \in (0,1]}\) is tight for all \(t\in (0,T]\) and we have
Proof
Thanks to (23), we have for all \(R>0\)
The conclusion follows from the arbitrariness of R because balls in the topology of \(D((-A)^{\epsilon _0})\) are compact in H. \(\square \)
Theorem 8
Assume that Hypotheses 1 and 2 are fulfilled and let \(\zeta \in \mathcal P_{\psi }(H)\).Footnote 3 Then there is a probability kernel \(\underline{\mu }(dt\,dx)=\mu _t(dx)\,dt\) solving (3) and such that \(\mu _t\in \mathcal P_\psi \) for all \(t\in [0,T]\), where \(\psi \) is defined by (21).
Proof
Since for any \(t\in (0,T]\), \((\mu ^\alpha _t)_{\alpha \in (0,1]}\) is tight, there exists a sequence \(\alpha _{h}(t)\rightarrow 0\) and \(\mu _t\in \mathcal P(H)\) such that
By a diagonal extraction argument we can find a sequence \((\alpha _h)\rightarrow 0\) and a family of measures \((\mu _t)_{t\in \mathbb Q}\) such thatFootnote 4
Now we are going to extend the family \((\mu ^{\alpha _h}_t)_{t\in \mathbb Q}\) to the whole interval (0, T] in such a way that the extension is a solution of (3). We shall proceed in three steps.
Step 1. For each \(\varphi \in \mathcal E_A(H)\) there exists \(C(\varphi )>0\) such that for all \(\alpha \in (0,1]\) we have
In fact by Itô’s formula we have
where \(L_t^\alpha ,\,t\in [0,T],\) is the Kolmogorov operator
Now, taking into account (16), write
Let \(C_1(\varphi )>0\) be such that
Therefore from (30), taking into account (15), (13) and (22), we find
So, (28) follows with a suitable constant \(C(\varphi )\).
Step 2. Construction of \(\mu _t\) for all \(t\in (0,T]\).
Let \(t_0\in (0,T]\setminus \mathbb Q\) and let \((t_j)\) be a sequence in \((0,T]\cap \mathbb Q\) convergent to \(t_0\). The sequence \((\mu _{t_j})\) is tight by Proposition 7. Let us choose a limit point of \((\mu _{t_j})\) which we call \(\mu _{t_0}\), that is (for a subsequence which we still call \((t_j)\))
Claim. We have
To prove the claim it is enough to show that
In fact, assume that we have proved (33). Since \((\mu ^{\alpha _h}_{t_0})\) is tight by (26), there exists a subsequence \((\mu ^{\alpha _{h_k}}_{t_0})\) weakly convergent to a probability measure \(\nu \), that is such that
On the other hand, by (33) it follows that
so that
which implies \(\nu =\mu _{t_0}\) because \( \mathcal E_A(H)\) is dense in \(L^1(H,\nu )\).
Now we can prove the Claim. Let \(\varphi \in \mathcal E_A(H)\). Then we have for any \(j\in \mathbb N\)
Taking into account (28), yields
Given \(\delta >0\) there is \(j_\delta \in \mathbb N\) such that \( J_1+J_3<\delta . \) Therefore by (34) we have
Now the conclusion follows from (27) and the arbitrariness of \(\delta \).
Step 3. Conclusion
Let \((\mu _t)_{t\in [0,T]}\) be the family defined in Step 2. We are going to prove that for all \(u\in \mathcal E_A([0,T]\times H)\)
This will imply
so that \((\mu _t)_{t\in [0,T]}\) is a solution of (3). To prove (35) is enough to show that
and
Let us prove (37). Recalling that \(D_xu(t,x)\in D(A)\), that Au is bounded and that \( |x|\,|\langle AD_xu(t,x),x\rangle |\le \Vert ADu\Vert _0\,|x|^2, \) we have for any \(\epsilon >0\), thanks to (19),
Now (37) follows letting first \(h\rightarrow \infty \) and then \(\epsilon \rightarrow 0\).
Next let us prove (38). Set \(g_h(t,x):= \langle Du(t,x),F_{\alpha _h}(t,x)\rangle \) and \( g(t,x):= \langle Du(t,x),F(t,x)\rangle ,\) so that (38) is equivalent to
Notice that by (14) we have
Now write
As far as \(I_h\) is concerned, we have for any \(j\in \mathbb N\),
Therefore, taking into account (40), yields
Now, given \(\epsilon >0\), choose \(j_0\) such that \( 2\alpha _{j_0} \Vert Du\Vert _0\int _{H_T}(1+V(x))\,d\mu _t\,dt<\frac{\epsilon }{2}. \) Then
so that \( \lim _{h\rightarrow \infty }I_h=0, \) for the arbitrariness of \(\epsilon \). Finally, as far as \(J_h\) is concerned, we have
as \(h\rightarrow \infty \). The proof is complete. \(\square \)
Let us present now some examples.
Example 9
(Tr \(C<\infty \)) We assume here that Tr \(C<\infty \), \(F:[0,T]\times H\rightarrow H\) is continuous and there exist \(k>0\) and \(N\in \mathbb N\) such that
Then we set
so that
Let \(X_\alpha \) be the solution to (11). Then by Itô’ formula and (41), for any \(m\in \mathbb N\) there exists \(a_m>0\) such that
Therefore Hypothesis 2 is fulfilled with \(V(x)=C(1+|x|^{4N}),\;x\in H\).
Example 10
(\(C=I\)) We assume here that \(C=I\), \(F:[0,T]\times H\rightarrow H\) is continuous and there exist \(k >0\), \(N_1,\,N\in \mathbb N\) such that for all \((t,x)\in [0,T]\times H\) we have
Then we define \(F_\alpha \) by (42). Let \(X_\alpha \) be the solution to (11). Set \(Y_\alpha (t,s,x):=X_\alpha (t,s,x)-W_A(t,s),\) so that
Now, setting \(Y(t)=Y_\alpha (t,s,x)\), it follows by a simple computation that for any \(m\in \mathbb N\) we have
Recalling [8, Proposition 4.3] it follows that there is \(C(T)>0\) such that \( \frac{d}{dt}\,\mathbb E|Y(t)|^{2m}\le C(T), \) which implies
Therefore Hypothesis 2 is fulfilled with \(V(x)=C(1+|x|^{4N})\).
Example 11
(Reaction–diffusion equations) Here we take \(H=L^2(0,1)\), \(Ax=D^2_\xi \) for all \(x\in D(A)=H^2(0,1)\cap H^1_0(0,1) \) and \( F(t,x)(\xi )=\sum _{i=0}^Na_i(t)(x(\xi ))^i,\; t\in [0,T],\;\xi \in [0,1],\,x\in L^2(0,1), \) where \(N\in \mathbb N\) is odd and greater than 1, \(a_i\in C([0,T])\), \(i=1,\ldots ,N\), and \(a_N(t)<0\) for all \(t\in [0,T].\) Then Hypothesis 1 is fulfilled. Moreover, setting
we have
and
Now Hypothesis 2(i)(ii) are fulfilled with \( V(x)=C\left( 1+|x|^{4N}_{L^{2N}(0,1)}\right) . \) Finally, Hypothesis 2(iii) is fulfilled as well, see [8, Theorem 4.8] and [3, p. 505] where a more general example is also presented. Then Theorem 8 applies.
3.3 Other Assumptions
In this section we set \(G(t,x)=(-A)^{-1/2}F(t,x)\), \(J_\alpha =(1-\alpha A)^{-1}\), \(\alpha \in (0,1].\)
Hypothesis 3
We set \(F_\alpha (t,x)=(-A)^{1/2}J_\alpha G_\alpha (t,x)\) and assume that:
(i) \(G_\alpha \) is continuous and bounded. and the mild equation
has a unique solution.
(ii) There exists a convex and lower semicontinuous mapping \(V:H\rightarrow [1,+\infty ]\) such that
and
(iii) There exists \(M\ge 0\) such that
where \(P^\alpha \) is the transition evolution operator \( P^\alpha _{s,t}\varphi (x)=\mathbb E[\varphi (X_\alpha (t,s,x))],\; 0\le s\le T,\;\varphi \in B_b(H). \)
(Note the difference between Hypothesis 2(iii) and Hypothesis 3(iii); the reason is that from (51) we are not able to estimate \(\mathbb E|X_\alpha (t,s,x)|^2\) independently of \(\alpha \).)
Now fix \(\zeta \in \mathcal P(H)\) and for any \(t\in (0,T]\) and any \(\alpha \in (0,1]\) define the probability measure \(\mu _t^\alpha \) in H setting as in Sect. 3.1
and \(\underline{\mu }^\alpha (dt\, dx)=\mu _t^\alpha (dx)dt.\)
Let us deduce some estimates from Hypothesis 3. First from (54) we have
and
and so, from (52)
Lemma 12
Assume that Hypotheses 1 and 3 are fulfilled and let \(x\in H\), \(0<s\le t\le T\) and \(\alpha \in (0,1]\). Then \(X_\alpha (t,s,x)\in D((-A)^{\epsilon _0})\), \(\mathbb P\)-a.s. and there exists \(C>0\) such that
Proof
Writing \(X_\alpha (t,s,x)=X_\alpha (t)\) we have
Therefore, since \(\Vert J_\alpha \Vert \le 1\) and using that \(|G_\alpha (r,X_\alpha (r))|\le 1+ |G_\alpha (r,X_\alpha (r))|^2\), we have
Taking expectation, yields
Consequently, taking into account (7) and (52) we find
and (58) follows. \(\square \)
In the following we shall set as before
Now, integrating (58) with respect to \(\zeta \) we obtain
Corollary 13
Assume that Hypotheses 1 and 3 are fulfilled. Then for all \(\zeta \in \mathcal P_\psi (H)\), \(t\in (0,T]\), \(x\in H\) and \(\alpha \in (0,1]\) we have
Now by Itô’s formula we see that \(\underline{\mu }^\alpha (dt\,dx)=\mu _t^\alpha (dx)dt\) solves the approximating Fokker–Planck equation (24) where \(K_{F_\alpha }\) is the Kolmogorov operator (25), moreover the tightness of \(\mu _t^\alpha \) follows as in the proof of Proposition 7.
Proposition 14
Assume that \(\zeta \in \mathcal P_{\psi }(H)\). Then \((\mu _t^\alpha )_{\alpha \in (0,1]}\) is tight for all \(t\in (0,T]\) and we have
Theorem 15
Assume that Hypotheses 1 and 3 are fulfilled and let \(\zeta \in \mathcal P_{\psi }(H)\).Then there is a probability kernel \(\underline{\mu }(dt\,dx)=\mu _t(dx)\,dt\) solving (3) and such that \(\mu _t\in \mathcal P_\psi \) for all \(t\in [0,T]\), where \(\psi \) is defined by (60).
Proof
We proceed as in the proof of Theorem 8. The Step 1 is similar: we construct a sequence \((\alpha _h)\rightarrow 0\) and a family of measures \((\mu _t)_{t\in \mathbb Q}\) such that
Then we extend the family \((\mu ^{\alpha _h}_t)_{t\in \mathbb Q}\) to the whole interval (0, T] in such a way that the extension is a solution of (3) proving that
Step 2. For each \(\varphi \in \mathcal E_A(H)\) there exists \(C(\varphi )>0\) such that for all \(\alpha \in (0,1]\) we have
In fact by Itô’s formula we have
where \(L_t^\alpha ,\,t\in [0,T],\) is the Kolmogorov operator
Now we write
and choose \(C_1(\varphi )>0\) such that
Therefore (64) follows arguing as before as well as Step 2. We go now to the Step 3. We have still to show statements (37) and (38); the proof of (37) is exactly the same, whereas for the proof of (38) we have just to write
and notice that \( (-A)^{1/2}D_xu\) is bounded. Then the proof ends as before. \(\square \)
Example 16
(Burgers type equations) We take here \(H=L^2(0,1)\), C of trace class and denote by A the realisation of the Laplace operator in [0, 1] with periodic boundary conditions: \(Ax=D^2_\xi ,\quad D(A)=H^2_\#(0,1),\) where
Moreover we take \(F(t,x)(\xi )=D_\xi (g(t, x(\xi )),\;\;x\in H^1_\#(0,1)\), where \(g:[0,T]\times \mathbb R\rightarrow \mathbb R\) is continuous. Then Hypothesis 1 is fulfilled. Set \(G(t,x)(\xi )=g(t, x(\xi ))\), so that \(F(t,x)(\xi )=(-A)^{1/2}G(t,x),\;\forall \;x\in H^1_\#(0,1). \) Set moreover for \(\alpha \in (0,1]\)
To check Hypothesis 3(i)(ii) we need suitable estimates of the solution \(X_\alpha (t,s,x)\) of Eq. (11). For any \(m\in \mathbb N\) set \( \varphi _m(x)=\frac{1}{2m}\int _0^1|x(\xi )|^{2m}d\xi . \) By Itô’s formula we have
where \((e_h)\) is an orthonormal basis in H and \(Ce_h=c_he_h,\;h\in \mathbb N\). Therefore
In particular, we have \(\displaystyle \frac{d}{dt}\,\mathbb E|X(t)|^2\le -2\omega \mathbb E|X(t)|^2+\text{ Tr }\,C,\) so that
Moreover by (69)
But
and
So,
and taking into account (70) we have
It follows that
Iterating this procedure we find that there exists \(C_m(T)>0\) such that \( \mathbb E[|X(t)|_{L^{2m}(0,1)}^{2m}]\le C_m(T)(1+ |x|^{2m}). \) Therefore Hypothesis 3(i)(ii)(iii) is fulfilled with \( V(x)=C(1+|x|^4_{L^4(0,1)}).\)
Example 17
(Burgers equation perturbed by white noise) We take here \(H\,{=}\,L^2(0,1),\) \(A\,{=}\,Ax\,{=}\,D^2_\xi ,\quad D(A)\,{=}\,H^2_\#(0,1),\) and \(F(t,x)(\xi )=D_\xi (x^2(\xi )),\;\forall \;x\in H^1_\#(0,1)\).Footnote 5 Then Hypothesis 1 is fulfilled. Set \(G(t,x)(\xi )= x^2(\xi ),\;\forall \;x\in H^1_\#(0,1),\) so that \( F(t,x)(\xi )=(-A)^{1/2}G(t,x),\;\forall \;x\in H^1_\#(0,1).\) Set moreover for \(\alpha \in (0,1]\)
Then Hypothesis 3(i)(ii) is fulfilled with \( V(x)=C(1+|x|^4_{L^4(0,1)}).\) Finally, Hypothesis 3(iii) follows from [10, Proposition 2.2].Footnote 6
Example 18
(2D-Navier–Stokes equation perturbed by coloured noise) Let us consider the space \(L_\#^2\) of all real \(2\pi \)-periodic functions in the real variables \(\xi _1\) and \(\xi _2,\) which are measurable and square integrable on \([0,2\pi ]\times [0,2\pi ]\) endowed with the usual scalar product \(\langle \cdot ,\cdot \rangle \) and norm \(|\cdot |\). We denote by \((L_\#^2)^2\) the space consisting of all pairs \( x=\left( \begin{array}{c} x^1\\ x^2 \end{array}\right) \) of elements of \(L_\#^2\) endowed with the inner product \( \displaystyle \langle x,y \rangle =\int _\mathcal O [x^1(\xi )y^1(\xi )+x^2(\xi )y^2(\xi )]d\xi ,\; x,y\in (L_\#^2)^2. \) Moreover, for any \(x\in (L_\#^2)^2\) we set \( |x|= \langle x,x \rangle ^\frac{1}{2}. \) (We shall consider everywhere also the corresponding complexified spaces). Let \( f_{h,k}=\left( \begin{array}{c} e_h\\ e_k \end{array} \right) ,\;h,k\in \mathbb Z^2, \) be the complete orthonormal system of \((L_\#^2)^2,\) where \( e_k(\xi )=\tfrac{1}{2\pi }\;e^{i\langle k,\xi \rangle },\; k=(k_1,k_2)\in \mathbb Z^2,\;\xi =(\xi _1,\xi _2), \) and \(\langle k,\xi \rangle =k_1\xi _1+k_2\xi _2\). Then we shall denote by H the closed subspace of \((L_\#^2)^2\) of all divergence free vectors, that is: \( H=\text{ linear } \text{ span }\;\{g_k:\;k\in \mathbb Z^2\}, \) where
Note in fact that \( \text {div}\;g_k(\xi )=\frac{i}{|k|}\;e_k(-k_1k_2+k_1k_2)=0,\; k\in \mathbb Z^2. \) Any element \(x=\left( \begin{array}{c} x^1\\ x^2 \end{array} \right) \) of H can be developed as a Fourier series \(\displaystyle x=\sum _{k\in \mathbb Z^2}x_k g_k,\) where
Moreover, we shall denote by \(L^p_\#\), \(p\ge 1\), the subspace of \((L_\#^p)^2\) of all divergence free vectors and by \(|\cdot |_{p}\) the norm in \(L^p_\#\). Let us define the Stokes operator \(A:D(A)\rightarrow H\) setting \(Ax=\mathcal P(\varDelta _\xi x-x),\; x\in D(A)=H_\#^{2}, \) A is self-adjoint and \( A g_k=-(1+|k|^2)g_k,\;k\in \mathbb Z^2. \) We also define the non linear operator F setting
For any \(\sigma > 0\) we have
see [8, Chap. 6] for details. If \(x\in H^{1}_\#\), taking into account that div \(x=0\), we have
Set \(G(x)=(-A)^{1/2}F(x), \) and moreover for \(\alpha \in (0,1]\)
and \( G_\alpha (x)=(-A)^{-1/2} F_\alpha (x). \) Then for \(h\in H^{1}_\#\) and \(\alpha \in (0,1]\) we have
Finally, assume that Tr \([CA^{-1]}<\infty \). then it is easy to see that, thanks to [8, Lemma 6.7], Hypothesis 3 is fulfilled with \(G(x)=(1+|x|^4_{L^4_\#})\).
4 Uniqueness
4.1 The Rank Condition
We start with the following crucial result, proved in [1].
Theorem 19
Let \(\psi :H\rightarrow [0,+\infty ]\) be convex and lover semi-continuous and let \(\zeta \in \mathcal P_\psi (H)\). Assume that for any solution \(\underline{\mu }=(\mu _t)_{t\in [0,T]}\) of (3) such that \(\mu _0=\zeta \) the following statement holds
Then Eq. (3) has at most one solution.
The condition (77) is called the rank condition.
Proof
Assume that \(\underline{\mu }_1 =(\mu _{1,t})\) and \(\underline{\mu }_2 =(\mu _{2,t})\) are solution of (3) such that \(\mu _{1,0}=\mu _{2,0}=\zeta \in \mathcal P_\psi (H)\) and set
Then \(\lambda _0=\zeta \) and \( \underline{\mu }_1\ll \underline{\lambda },\quad \underline{\mu }_2\ll \underline{\lambda }. \) Denote by \(\rho _1\) and \(\rho _2\) the respective densities
We claim that
Let us show for instance that \(0\le \rho _1(t,x)\le 2\). In fact, for any \(A\in \mathcal B([0,T]\times H)\) we have \( \underline{\mu }_2(A)=2\underline{\lambda }(A)-\underline{\mu }_1(A), \) so that \( \underline{\mu }_2(A)=\int _A(2-\rho _1)d\underline{\lambda }\), which implies that \(0\le \rho _1(t,x)\le 2\), \(\underline{\lambda }\)-a.e. by the arbitrariness of A.
Now for any \(u\in \mathcal E_A([0,T]\times H)\) we have (recall that \(H_T=[0,T]\times H\)))
Since \(\rho _1-\rho _2\in L^\infty ([0,T]\times H;\underline{\lambda }),\) by the rank condition it follows that \(\rho _1=\rho _2\). \(\square \)
A useful remark for dealing with the rank condition is that for any solution \(\underline{\mu }\) to Eq. (3) such that \(\mu _0\in \mathcal P_\psi (H)\), \(K_F\) is dissipative in \(L^1([0,T]\times H;\underline{\mu })\), as the next proposition shows.
Proposition 20
Let \(F:D(F)\subset [0,T]\times H\rightarrow H\) be Borel and \(\zeta \in \mathcal P(H)\). Assume that \(\underline{\mu }\) is a solution to (3). Then \(K_F\) is dissipative in \(L^p([0,T]\times H;\underline{\mu })\) for all \(p\ge 1\).
Proof
First we show that for any \(u\in \mathcal E_A([0,T]\times H)\) we have
In fact if \(u\in \mathcal E_A([0,T]\times H)\) we have \(u^2\in \mathcal E_A([0,T]\times H)\) as well and
It follows that
Integrating both sides of (80) with respect to \(\underline{\mu }\) over \([0,T]\times H\), yields (79). Now, by (79) we have
which implies the dissipativity of \(K_F\) in \(L^2([0,T]\times H,\underline{\mu })\). The case \(p\ne 2\) follows from standard arguments about diffusions operators, see [12]. \(\square \)
Remark 21
Assume that \(\underline{\mu }\) is a solution to (3). By Proposition 20 it follows that \(K_F\) is closable in \(L^1([0,T]\times H,\underline{\mu })\). We shall denote by \(\overline{K_F}\) its closure. Clearly, if \(\psi :H\rightarrow [0,+\infty ]\) is convex lover semi-continuous and
then the rank condition (77) is fulfilled.
4.2 The Semigroup Associated to a Non Autonomous Problem
We assume here that Hypothesis 2 (resp. Hypothesis 3) is fulfilled. As usual in studying non autonomous systems, it is convenient to consider, besides (11) (resp. (51)), a problem in the unknowns \((H(\tau ),y(\tau ))\) (intended in the mild sense.)Footnote 7
(resp. a similar problem for (51)). To solve problem (81) we set \(y(\tau ,t)=t+\tau \), so that (81) reduces to
We denote by \(H(\tau )=H(\tau ,t,x)\) the mild solution of (81) which, however, is only defined for \(\tau \in [0,T-t]\).
Let us define a semigroup \(S^{F_\alpha }_\tau ,\,\tau \ge 0,\) in the space \(C_T([0,T]\times H)=\{u\in C([0,T]\times H):\;u(T,x)=0,\forall \;x\in H \}\) by setting
Notice that \( S^{F_\alpha }_\tau =0,\; \forall \;\tau \ge T. \) Since the law of \( H(\tau ,t,x)\) coincides, as easily seen, with that of \(X_\alpha (t+\tau ,t,x)\), for any \(x\in H\) and any \(\tau \) such that \(t+\tau \le T\), it results
Let denote by \(\mathcal K_{F_\alpha }\) the infinitesimal generator of \(S^{F_\alpha }_\tau \), defined through its resolvent (see (84) below). Then the resolvent set of \(\mathcal K_{F_\alpha }\) coincides with \(\mathbb R\) and its resolvent is given, for all \(f\in C_T([0,T]\times H)\) and all \(\lambda \in \mathbb R\), by
\(\mathcal K_{F_\alpha }\) can also be defined as follows (following [14]). We say that a function \(u\in C_T([0,T]\times H)\) belongs to the domain of \(\mathcal K_{F_\alpha }\) if there exists a function \(f\in C_T([0,T]\times H)\) such that
(i)
\(\displaystyle \lim _{\tau \rightarrow 0}\frac{1}{\tau }\;(S^{F_\alpha }_\tau u(t,x)-u(t,x))=f(t,x),\quad \forall \;(t,x)\in [0,T]\times H.\)
(ii) There exists \(M_u>0\) such that
We set \(\mathcal K_{F_\alpha }u=f\) and call \(\mathcal K_{F_\alpha }\) the infinitesimal generator of \(S_\tau \) on \(C_T([0,T]\times [0,H])\). As easily seen, the abstract operator \(\mathcal K_{F_\alpha }\) is an extension of the differential Kolmogorov operator \(K_{F_\alpha }\) defined by (2) (with \(F_\alpha \) replacing F).
4.3 The Case When \(C^{-1}\) is Bounded
Theorem 22
Assume, besides Hypotheses 1 and 2 that there exists \(M_1>0\) such that
Let moreover \(\zeta \in \mathcal P_{\psi +V^2}(H)\), where \(\psi \) is defined by (21). Then the Fokker–Planck equation (3) has a unique solution \(\underline{\mu }=(\mu _t)_{t\in [0,T]}\) and \(\mu _t\in \mathcal P_{\psi +V^2}(H)\) for any \(t\ge 0\).
Proof
Let \(f\in \mathcal E_A([0,T]\times H))\) and consider the approximating equation
Thanks to (84) Eq. (86) has a unique solution \(u_\alpha \) given by
and therefore
Step 1. \(u_\alpha \in D(\mathcal K^1_0)\cap C^1_b(H)\) and it results
Fix in fact \(t\in [0,T]\) and \(h>0\) such that \(t+h\le T\). Then, recalling (12) we write
where
and
and
Consequently,
which implies
Now, letting \(h\rightarrow 0\) and taking into account that \( \lim _{h\rightarrow 0}\frac{1}{h}\;g(t+h,t,x)=F_\alpha (t,x), \) yields (89).
Notice that from Step 1 does not follow that \(u_\alpha \in D(K_F)= \mathcal E_A([0,T]\times H\)), but we are going to show in the next step that \(u_\alpha \) belongs to the domain of the closure \(\overline{ K_F}\) of \(K_F.\) (Recall Remark 21).
Step 2. \(u_\alpha \in D(\overline{ K_F})\) and we have
By Proposition 4 there is a multi-sequence \((u_{\alpha ,\mathbf {j}})\) in \(\mathcal E_A([0,T]\times H)\) such that for all \((t,x)\in [0,T]\times H\)
-
(i)
\(\displaystyle \lim _{\mathbf {j}\rightarrow \infty }u_{\alpha ,\mathbf {j}}(t,x)= u_\alpha (t,x),\)
-
(ii)
\(\displaystyle \lim _{\mathbf {j}\rightarrow \infty }K_0 u_{\alpha ,\mathbf {j}}(t,x)= \mathcal K^1_0u_\alpha (t,x), \)
-
(iii)
\(\displaystyle \lim _{\mathbf {j}\rightarrow \infty }D_xu_{\alpha ,\mathbf {j}}(t,x)= D_xu_\alpha (t,x),\)
-
(iv)
\(\displaystyle |u_{\alpha ,\mathbf {j}}(t,x)|+| K_0\, u_{\alpha ,\mathbf {j}}(t,x)|+|D_xu_{\alpha ,\mathbf {j}}(t,x)|\le C(1+|x|),\quad x\in H .\)
Then we have
Since
we have by Hypothesis 2
By the dominated convergence theorem it follows that \(u_\alpha \in D(\overline{K_F})\) and
Therefore \(u_\alpha \in D(\overline{ K_F})\) and (91) is fulfilled .
Step 3. There is \(C(T)>0\) such that
Multiplying both sides of Eq. (91) by \(u_\alpha \) and integrating in \(\underline{\mu }\) over \(H_T\), yields
Taking into account (79), we obtain
Now, recalling (88), we obtain
Now the conclusion follows by standard arguments.
Step 4.
In view of (91) it is enough to show that
We have in fact by Hölder’s inequality, taking into account (85), (88) and (94) that,
as \(\alpha \rightarrow 0\). The proof is complete. \(\square \)
Example 23
We continue here Example 10 using notations and assumptions there. As we have seen, Hypothesis 3 is fulfilled in this case. To apply Theorem 22 it remains to check that (85) is fulfilled. This clearly holds by (47), thus Theorem 22 applies.
Now we assume Hypotheses 1 and 3. Then, to repeat the proof of Theorem 22 we should find an estimate for \(|(-A)^{1/2}D_xu_\alpha |^2\) rather than for \(|D_xu_\alpha |^2\) which seems to be difficult. So, we shall assume (95) and, arguing as before, we can prove
Theorem 24
Assume, besides Hypotheses 1 and 3 that
Let moreover \(\zeta \in \mathcal P_{\psi }(H)\), where \(\psi \) is defined by (21). Then the Fokker–Planck equation (3) has a unique solution \(\underline{\mu }=(\mu _t)_{t\in [0,T]}\) and \(\mu _t\in \mathcal P_{\psi }(H)\) for any \(t\ge 0\).
Example 25
(Burgers equation) We consider here the setting of Example 17. Then statement (96) follows from [10, Lemma 4.1] and so Theorem 24 applies.
4.4 The Case When Tr \(C<\infty \)
Trying to repeat the proof of Theorem 22, whereas Steps 1 and 2 can be repeated without anyproblems, there is a difficulty for the proof of step 3 which requires \(C^{-1}\in L(H)\). The key point is again to prove the statement (95).
Theorem 26
Assume, besides Hypotheses 1 and 3 that
Then there is a unique solution \(\underline{\mu }\) of the Fokker–Planck equation (3).
Example 27
Consider the 2D-Navier–Stokes equation from Example 18 and assume that Tr\( [(-A)C]<\infty \). Then (97) is fulfilled by [8, Eq. 5.4.8]. So, Theorem 26 applies.
Notes
- 1.
The upper index 1 in the definitions of \(S^{0,1}\) and \(\mathcal K^1_0\) recalls the space \(C_{b,1}(H)\).
- 2.
\(\epsilon _0\) was defined in Hypothesis 1(ii).
- 3.
\(\psi \) is defined in (21).
- 4.
\(\mathbb Q\) denotes the set of rational numbers.
- 5.
We take F independent of t for simplicity.
- 6.
The aforementioned paper concerns Dirichlet boundary conditions but all its results generalise easily to periodic ones.
- 7.
We proceed here as in Sect. 2 (with \(F_\alpha =0\)), but working in \(C_b(H)\) rather than in \(C_{1,b}(H)\).
References
Bogachev, V., Da Prato, G., Röckner, M., Stannat, W.: Uniqueness of solutions to weak parabolic equations for measures. Bull. London Math. Soc. 39, 631–640 (2007)
Bogachev, V., Da Prato, G., Röckner, M.: Fokker-Planck equations and maximal dissipativity for Kolmogorov operators with time dependent singular drifts in Hilbert spaces. J. Funct. Anal. 256, 1269–1298 (2009)
Bogachev, V., Da Prato, G., Röckner, M.: Existence and uniqueness of solutions for Fokker-Planck equations on Hilbert spaces. J. Evol. Equ. 10(3), 487–509 (2010)
Bogachev, V., Da Prato, G., Röckner, M.: Uniqueness for solutions of Fokker-Planck equations on infinite dimensional spaces. Commun. Partial Differ. Equ. 36, 925–939 (2011)
Bogachev, V., Da Prato, G., Röckner, M., Shaposhnikov, S.: On the uniqueness of solutions to continuity equations. J. Differ. Equ. 259(8), 3854–3873 (2015)
Bogachev, V., Da Prato, G., Röckner, M., Shaposhnikov, S.: An analytic approach to infinite-dimensional continuity and Fokker-Planck-Kolmogorov equations. Ann. Sc. Norm. Super. Pisa Cl. Sci. 14(3), 983–1023 (2015)
Cerrai, S.: A Hille-Yosida theorem for weakly continuous semigroups. Semigroup Forum 49, 349–367 (1994)
Da Prato, G.: Kolmogorov Equations for Stochastic PDEs. Birkhäuser, Boston (2004)
Da Prato, G., Tubaro, L.: Some results about dissipativity of Kolmogorov operators. Czechoslov. Math. J. 51(126), 685–699 (2001)
Da Prato, G., Debussche, A.: \(m\)-dissipativity of Kolmogorov operators corresponding to Burgers equations with space-time white noise. Potential Anal 26, 31–55 (2007)
Da Prato, G., Flandoli, F., Röckner, M.: Fokker-Planck equations for SPDE with non-trace class noise. Commun. Math. Stat. 1(3), 281–304 (2013)
Eberle, A.: Uniqueness and Non-uniqueness of Semigroups Generated by Singular Diffusion Operators. Lecture Notes in Mathematics, vol. 1718. Springer, Berlin (1999)
Liu, W., Röckner, M.: Stochastic Partial Differential Equations: An Introduction. Universitext. Springer, Berlin (2015)
Priola, E.: On a class of Markov type semigroups in spaces of uniformly continuous and bounded functions. Studia Math. 136, 271–295 (1999)
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2018 Springer International Publishing AG, part of Springer Nature
About this paper
Cite this paper
Da Prato, G. (2018). Fokker–Planck Equations in Hilbert Spaces. In: Eberle, A., Grothaus, M., Hoh, W., Kassmann, M., Stannat, W., Trutnau, G. (eds) Stochastic Partial Differential Equations and Related Fields. SPDERF 2016. Springer Proceedings in Mathematics & Statistics, vol 229. Springer, Cham. https://doi.org/10.1007/978-3-319-74929-7_5
Download citation
DOI: https://doi.org/10.1007/978-3-319-74929-7_5
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-74928-0
Online ISBN: 978-3-319-74929-7
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)