Smoothing properties of McKean-Vlasov SDEs

In this article, we develop integration by parts formulae on Wiener space for solutions of SDEs with general McKean-Vlasov interaction and uniformly elliptic coefficients. These integration by parts formulae hold both for derivatives with respect to a real variable and derivatives with respect to a measure understood in the sense of Lions. They allows us to prove the existence of a classical solution to a related PDE with irregular terminal condition. We also develop bounds for the derivatives of the density of the solutions of McKean-Vlasov SDEs.


Introduction
The main object of study in this paper is the McKean-Vlasov stochastic differential equation (MVSDE) driven by a Brownian motion B = B 1 , . . . , B d , with coefficients V 0 , . . . , V d : R N × P 2 (R N ) → R N and initial condition θ, a square-integrable random variable independent of B. Here and throughout, we denote by [ξ] the law of a random variable ξ and by P 2 (R N ) the set of probability measures on R N with finite second moment. MVSDEs are equations whose coefficients depend on the law of the solution. They are also referred to as mean-field SDEs and their solutions are often called nonlinear diffusions. These MVSDEs provide a probabilistic representation to the solutions of a class of nonlinear PDEs. A particular example of such nonlinear PDEs was first studied by McKean [29]. These equations describe the limiting behaviour of an individual particle evolving within a large system of particles undergoing diffusive motion and interacting in a 'mean-field' sense, as the population size grows to infinity. A particular characteristic of the limiting behaviour of the system, is that any finite subset of particles become asymptotically independent of each other. This propagation of chaos phenomenon was studied by McKean [30] and Sznitman [34] among many other authors. Existence and uniqueness results, the theory of propagation of chaos and numerical methods have been studied in a variety of settings (see, for example, [6,7,21,31]).
As MVSDEs can be interpreted as limiting equations for large systems, they are widely used as models in statistical physics [7,31] as well as in the study of large-scale social interactions within the theory of mean-field games [26,27,28,19,20,10,11]. Recently, these equations have also appeared in the mathematical finance literature in the specification and calibration of multi-factor stochastic volatility and hybrid models [5,17].
In this paper, we develop several new integration by parts formulae for solutions of MVSDE. In turn, these formulae enable us to use MVSDE to define the solution of a class of partial differential equations that has the form (∂ t − L) U (t, x, [θ]) = 0 for (t, x, [θ]) ∈ (0, T ] × R N × P 2 (R N ) for (x, [θ]) ∈ R N × P 2 (R N ), (1.2) where g : R N × P 2 (R N ) → R and the operator L acts on sufficiently enough functions F : R N × P 2 (R N ) → R N and is defined where σ(z, µ) is the N × d matrix with columns V 1 (z, µ), . . . , V d (z, µ). The last two terms in the description of LF (x, [θ]) involve the derivative with respect to the measure variable as introduced by Lions in his seminal lectures at the Collège de France (see [9] for details), which we describe in Section 2.3. Papers [3,4,22] present further details of the relevance of the class of nonlinear partial differential equations (1.2) For linear parabolic PDEs on [0, T ] × R N it is well known from classical works such as [16,18] that under uniform ellipticity or Hörmander condition, there exist classical solutions even when the initial condition is not differentiable. In this paper, we explore to what extent the same is true for the PDE (1.2) under a uniform ellipticity assumption. That is, we consider the question of whether the PDE (1.2) has classical solutions when the initial condition g is not differentiable. For this we exploit a probabilistic representation for the classical solution 1 of the PDE (1.2) given in terms of a functional of X θ t and of the solution of the following de-coupled equation: We say that this equation is de-coupled as the law appearing in the coefficients is X θ s (the solution of equation (1.1)), rather than the law of X solves the PDE (1.2). A similar result has been proved in [8,12] under different conditions than ours and for an initial condition g that is sufficiently smooth. For the stochastic flow (X x t ) t≥0 solving a classical SDE with initial condition x ∈ R N , the standard strategy to show that the function u(t, x) := E g(X x t ) is a classical solution of a linear PDE is to show, using the flow property of X x t , that for h > 0, u(t + h, x) = E [u(t, X x h )] and then show that u is regular enough to apply Itô's formula to u(t, X x h ). Expanding this process using Itô's formula and sending h → 0 shows that u does indeed solve the related PDE. For MVSDEs, one can develop a similar approach. In this setting, to expand a function depending not only on the process (X x,[θ] t ) t≥0 (where we can use the usual Itô formula) but also on the flow of measures [X θ t ] t≥0 , we require an extension of the classical chain rule and we use here the chain rule proved in [12]. Our main focus is therefore to provide conditions under which U , defined in (1.4), is regular enough to apply the Itô formula and the extended chain rule.
For a general Lipschitz continuous function g : R N × P 2 (R N ) → R, we cannot expect for the mapping (x,

Outline & Main Results
In Section 2, we introduce the notation and the basic results related to MVSDEs. In particular, when describing the smoothness of the coefficients in equations (1.1) and (1.3) in our assumptions, we introduce the notation C k,k b,Lip (R N ×P 2 (R N ); R N ) for functions k-times differentiable with bounded, Lipschitz derivatives, which we introduce precisely in Section 2.3. Similarly, we use the notation K q r (E, M ) to denote processes taking values in a Hilbert space E which are smooth in both Euclidean and measure variables as well as in the Malliavin sense and M denotes how many times the process can be differentiated. This class, which we call the class of Kusuoka-Stroock processes, is introduced in Section 2.4. The class represents a generalization of the class of processes introduced in [25] and analysed in [14].
In Section 3, we prove some results on the differentiability of X . This is proved in the Appendix A.2. We then introduce the uniform ellipticity assumption (UE) in Assumption 3.3, used throughout the rest of the paper. The rest of the section details several corollaries, where we analyse the processes that will play the rôle of Malliavin weights in the integration by parts formulas and identify the class K q r (E, M ) of Kusuoka-Stroock processes to which they belong. With the main technical results complete, in Section 4 we develop integration by parts formulas for derivatives of ( ) under (UE) and the assumption that V 0 , . . . , V d ∈ C k,k b,Lip (R N × P 2 (R N ); R N ). We do this for derivatives with respect to x and with respect to µ. In particular we show that (see Propositions 4.1 and 4.2), for where I 3 α I 2 β (Ψ) and I 3 α I 2 β (Ψ) are defined is defined in Section 4.1 and We also consider integration by parts formulas for derivatives of the function x → Ef (X x,δx t ) (see Theorem 4.4). In Section 5, we return our attention to the PDE (1.2). In Definition 5.3, we introduce the class (IC) of nondifferentiable initial conditions g for which we are able to prove (x, ] is differentiable. We do this by extending the integration by parts formulas of Section 4 to cover this class. Then, for g in this class and assuming uniform ellipticity, and the coefficients V 0 , . . . , V d ∈ C 3,3 b,Lip (R N × P 2 (R N ); R N ) (and possibly bounded depending on the exact form of g) we are able to prove the existence and uniqueness of solutions to the PDE (1.2). In particular, we show (see Theorem 5.8) that function U , defined in (1.4), is a classical solution of the PDE (1.2). Moreover, U is unique among all of the classical solutions satisfying the polynomial growth condition Finally, in Section 6, we apply the integration by parts formulae to the study of the density function of X x,δx t . We study the smoothness of the density function and obtain estimates on its derivatives. The main result (See Theorem 6.1) states that, under suitable conditions, X x,δx t has a density p(t, x, z) such that (x, z) → p(t, x, z) is differentiable a number of times dependent on the regularity of the coefficients. Indeed, when these derivatives exist, there exist a constant C such that where µ = 4|α| + 3|β| + 3N and ν = 1 2 (N + |α| + |β|). Moreover, if V 0 , . . . , V d are bounded then the following Gaussian type estimate holds

Comparison with other works
As mentioned previously, the PDE (1.2) is also studied in [8] and [12]. Let us explain the relationship between the results in those works and the results in this paper.
In [8], the authors prove that derivatives of (x, [θ]) → X x,[θ] t exist up to second order. We also prove this as part of Theorem 3.2, although we extend this to derivatives of any order (assuming sufficient smoothness of the coefficients). In [8], the hypotheses on the continuity and differentiability of the coefficients are the same as ours The authors then consider initial conditions g : R N × P 2 (R N ) → R N for which the derivatives up to second order exist and are bounded, which they use to prove regularity of U . Since g is sufficiently smooth, they do not need to impose any non-degeneracy condition on the coefficients. In our work we remove the constraint on the smoothness of g at the expense of assuming non-degeneracy condition on the coefficients of the MVSDEs. In this sense, their results are complementary to ours.
The paper [12] has a completely different scope. The authors are interested in a nonlinear PDE on [0, T ] × R N × P 2 (R N ), called the master equation in reference to the theory of mean-field games. The PDE we consider is a special case of this, although again they assume that the function g is twice differentiable. Their strategy for proving regularity of U is also different. In their setting, the authors prove that derivatives of the lifted flow exist up to second order (with derivatives in the variable θ being Fréchet derivatives on the Hilbert space L 2 (Ω)) where X x,[θ] t is the forward component in a coupled forward-backward system. They use this result, along with sufficient smoothness of g, to prove that the lifted function U , defined on on [0, T ]×R N ×L 2 (Ω) is sufficiently regular in the Fréchet sense. They then prove a result which allows them to recover regularity of the second order derivatives of U from properties of the second order Fréchet derivatives of U . Using their strategy, the authors of [12] are able to impose hypotheses which only involve conditions on derivatives of the coefficients This is in contrast to our assumptions which impose conditions on More recently, two other works [2,13] give some partial results related to the smoothness of the solutions of McKean-Vlasov SDEs. In [2], the Malliavin differentiability of McKean-Vlasov SDEs is studied using a stochastic perturbation approach of Bismut type. In [13], the strong well-posedness of a McKean-Vlasov SDEs is proven when the diffusion matrix is Lipschitz with respect to both the space and measure arguments and uniformly elliptic and the drift is bounded in space and Hölder continuous in the measure direction. Both works restrict themselves to the particular case when the coefficient dependence on the law of the solution is of scalar type. We obtain some related results in [15], under the same scalar dependence restriction, but under the more general Hörmander condition.
We base our results on the use of Malliavin calculus techniques. The new integration by parts formulae and, more importantly, the identification of the processes appearing in these formulae as Kusuoka-Stroock processes is key to our analysis. The use of Kusuoka-Stroock processes is a very versatile tool. Not only that it enables us to identify the solution of the PDE (1.2), but the also allows us to study the density of X x,δx t and obtain both polynomial and Gaussian local bounds for their derivatives. We are not aware of similar bounds obtained elsewhere in the literature for densities of solutions of MVSDEs.

Notation & Basic Setup
We work on a filtered probability space (Ω, F , F = {F t } t∈[0,T ] , P) which supports an F-adapted d-dimensional Brownian Motion, B = (B 1 , . . . , B d ). We also often denote B 0 (s) = s for s ∈ [0, T ]. We assume that there is a sufficiently rich sub-σ-algebra G ⊂ F independent of B such that all measures µ ∈ P 2 (R N ) correspond to the law of a random variable in L 2 ((Ω, G, P); R N ). Then, we define F to be the filtration generated by B, completed and augmented by G. This is to ensure that in the sequel when we consider processes starting from arbitrary initial conditions θ ∈ L 2 (Ω; R N ) these processes will be F-adapted. We denote the L p norm on (Ω, F , P) by · p and we also introduce the space S p T of continuous F-adapted processes ϕ on [0, T ], satisfying In addition to the probability space (Ω, F , P), we will also make use of other probability spaces (Ω,F ,P) and ( Ω, F , P) when performing the lifting operation associated with the Lions derivative. We assume that these satisfy the same conditions as (Ω, F , P). We denote the L p norm on each of these spaces by · p unless we want to emphasise which space we are working on, in which case we use · L p ( Ω) etc. We use | · | to denote the Euclidean norm. Throughout we denote by α and β multi-indices on {1, . . . , N } including the empty multi-index. We denote by Id N the N × N identity matrix. We also use some terminology from Malliavin calculus: we denote by D the Malliavin derivative and by δ its adjoint, the Skorohod integral. We outline very briefly the basic operators of Malliavin calculus in Appendix A.1.

Basic results on McKean-Vlasov SDEs
We study McKean-Vlasov SDEs with general Lipschitz interaction. The coefficients are functions from R N ×P 2 (R N ) to R N , where P 2 (R N ) denotes the space of probability measures on R N with finite second moment. We equip this space with the 2-Wasserstein metric, W 2 . For a general metric space (M, d), we define the 2-Wasserstein metric on P 2 (M ) by where P µ,ν denotes the set of measures on M × M with marginals µ and ν. When we refer to the Lipschitz property of the coefficients, it is with respect to product distance on R N × P 2 (R N ).
Proposition 2.1 (Existence, Uniqueness and L p estimates). Suppose that θ ∈ L 2 (Ω) and V 0 , . . . , V d are uniformly Lipschitz continuous, then there exists a unique, strong solution to the equation

1)
and there exists a constant C = C(T ), such that Similarly, there exists a unique, strong solution to the equation

3)
and there exists a constant C = C(p, T ), such that for all p ≥ 1,

5)
and X x, Finally, we have the following flow property for any t ∈ [0, T ), s ∈ (t, T ], x ∈ R N and θ ∈ L 2 (Ω), Proof. The proof is standard and we leave it to the reader. We note that the proof of existence and uniqueness of a solution to equation (2.1) was proved in [34] for first-order McKean-Vlasov interaction. The case of a generic Lipschitz McKean-Vlasov interaction is covered in [21].

Differentiation in P 2 (R N )
In Section 5, we study an SDE with a general McKean-Vlasov dependence. We will be interested in differentiability of the stochastic flow associated to this SDE and an associated PDE on [0, T ] × R N × P 2 (R N ). We thus need a notion of derivative for a function on a space of probability measures. The notion of differentiability we use was introduced by P.-L. Lions in his lectures at the Collège de France, recorded in a set of notes by Cardaliaguet [9]. The underlying idea is very well exposed in [11], which we draw on here.
Lions' notion of differentiability is based on the lifting of functions U : P 2 (R N ) → R into functionsŨ defined on the Hilbert space L 2 (Ω; R N ) over some probability space (Ω,F ,P),Ω being a Polish space andP an atomless measure, by settingŨ (X) = U ([X]) forX ∈ L 2 (Ω; R N ). Then, a function U is said to be differentiable at µ 0 ∈ P 2 (R N ) if there exists a random variableX 0 with law µ 0 such that the lifted functionŨ is Fréchet differentiable atX 0 . Whenever this is the case, the Fréchet derivative ofŨ atX 0 can be viewed as an element of L 2 (Ω; R N ) by identifying L 2 (Ω; R N ) and its dual. The derivative in a directionγ ∈ L 2 (Ω; R N ) is given by It then turns out (see Section 6 in [9] for details.) that the distribution of DŨ (X 0 ) ∈ L 2 (Ω; R N ) depends only upon the law µ 0 and not upon the particular random variableX 0 having distribution µ 0 . It is shown in [9] that, as a random variable, DŨ (X 0 ) is of the form g µ0 (X 0 ), where g µ0 : R N → R N is a deterministic measurable function which is uniquely defined µ 0 -almost everywhere on R N , and is square-integrable with respect to the measure µ 0 . We call ∂ µ U (µ 0 ) := g µ0 the derivative of U at µ 0 . We use the notation This holds for any random variableX 0 with distribution µ 0 , irrespective of the probability space on which it is defined.
In the sequel, we will consider functions which are differentiable globally on P 2 (R N ). Moreover, we will consider functions where for each µ ∈ P 2 (R N ), there exists a version of the derivative ∂ µ U (µ) which is assumed to be a priori continuous as a function In this case such a version is unique since, for each θ ∈ L 2 (Ω; (dv)-a.e., so taking a Gaussian random variable G independent of θ, and ǫ > 0, ∂ µ U ([θ + ǫG], v) is defined (dv)-a.e. and taking ǫ → 0 and using the continuity of ∂ µ U , identifies ∂ µ U ([θ], v) uniquely. We show how this definition works in practice in Examples 2.5 and 2.6.
For a function f : P 2 (R N ) → R N , we can straightforwardly apply the above discussion to each component of f = (f 1 , . . . , f N ). To extend to higher derivatives we note that ∂ µ f i takes values in R N , so we denote its components by (∂ µ f i ) j : P 2 (R N ) × R N → R for j = 1, . . . , N and, for a fixed v ∈ R N , we can discuss again the differentiability of If the derivative of this function exists and there is continuous version of then it is unique. It makes sense to use the multi-index notation ∂ (j,k) We always denote these variables, by v 1 , . . . , v n , so When there is no possibility of confusion, we will abbreviate (v 1 , . . . , v n ) to v, so that with | · | the Euclidean norm on R N . It then makes sense to discuss derivatives of the function ∂ α µ f i0 with respect to the variables v 1 , . . . , v n . If, for some j ∈ {1, . . . , N } and all (µ, v 1 , . is l-times continuously differentiable, we denote the derivatives ∂ βj vj ∂ α µ f i0 , for β j a multi-index on {1, . . . , N } with |β j | ≤ l. Similar to the above, we will denote by β the n-tuple of multi-indices (β 1 , . . . , β n ). We also associate a length to β by |β| := |β 1 | + . . . + |β n |, and denote #β := n. Then, we denote by B n the collection of all such β with #β := n, and B := ∪ n≥1 B n . Again, to lighten notation, we will use The coefficients in equations (2.1) and (2.3) are of the type V 0 , . . . , V d : R N × P 2 (R N ) → R N , so depend on a Euclidean variable as well as a measure variable. Considering functions on R N × P 2 (R N ) raises a question about whether the order in which we take derivatives matters. A result from [8] says that derivatives commute when the mixed derivatives are Lipschitz continuous.
both exist and are Lipschitz continuous: i.e. there exists a constant C > 0 such that Then, the functions ∂ x ∂ µ g and ∂ µ ∂ x are identical.
With this in mind, we can introduce the following definition.
(c) We say that h ∈ C n b,Lip (P 2 (R N ); R N ) if h : P 2 (R N ) → R N does not depend on a Euclidean variable but otherwise satisfy the conditions in part (b).

For functions
) with the first argument being considered pointwise by ω and the second depending on the random variable ξ through its law.
2. From the bounds in Definition 2.3(a), we have the following simple consequences for the Fréchet derivative of the liftingṼ of V : for all x, x ′ ∈ R N and θ, θ ′ , γ, γ ′ ∈ L 2 (Ω), 3. Note that we cannot interchange the order of due to Lemma 2.2.

We now introduce some concrete examples of functions
In Section 4, we develop integration by parts formulas modelled on those developed in works of Kusuoka [24] along with Stroock [25] for solutions of classical SDEs. These integration by parts formulas take the form for processes Ψ, Ψ α , Ψ β belonging to a specific class. We work with a class of processes similar to one introduced in [25], which we call the class of Kusuoka-Stroock processes.

For any
Remark 2.9. This definition is different to that in [25] in the following ways: 1. The processes depend on a parameter µ ∈ P 2 (R N ).
2. We keep track of polynomial growth in x of the D m,p -norm through a parameter q > 0 instead of requiring it to be uniformly bounded.
3. We require continuity in L p (Ω) rather than almost surely.
1. The number M denotes how many times the Kusuoka-Stroock process can be differentiated; q measures the polynomial growth of the D m,p -norm of the process in (x, [θ]), and r measures the growth in t.
2. In the definition, we are able to stipulate that the D m,p -norm of all the derivatives will be uniformly bounded w.r.t. v because in the sequel the only dependence on v in any Kusuoka-Stroock processes will come from (v) is bounded w.r.t v and this carries over to the D m,p -norm.
To analyse the density of solutions of the MVSDE (2.1) started from a fixed initial point in R N , it is useful to have notation for Kusuoka-Stroock processes which do not depend on a measure µ ∈ P 2 (R N ). We denote this class by K q r (R, M ). The following lemma says that if we take a Kusuoka-Stroock process on R N × P 2 (R N ) and evaluate its measure argument at a Dirac mass, then this forms a Kusuoka-Stroock process on R N . Its proof is straightforward.

Regularity of Solutions of McKean-Vlasov SDEs
This section contains some basic results about solutions of the equations involved, their integrability and their differentiability with respect to parameters. Existence and uniqueness of solutions to (1.3) is covered in Section 2.2.
. Then the following hold: (a) There exists a modification of X [θ] and note that it solves the following SDE there exists a linear continuous map DX θ t : L 2 (Ω) → L 2 (Ω) such that for all γ ∈ L 2 (Ω), and similarly for X . These processes satisfy the following stochastic differential equations (v) exists and it satisfies the following equation

4)
where Xθ s is copy of X θ s on the probability space (Ω,F ,P) driven by the Brownian motionB and with initial conditionθ. Similarly, s driven by the Brownian motionB and ∂ µ Xθ ,[θ] Finally, the following representation holds for all γ ∈ L 2 (Ω): (v) solves equation (3.4). We first re-write the equation for DX θ t (γ) in terms of ∂ µ V i instead of the Fréchet derivative of the liftingṼ i , as follows Consider the equation satisfied by ∂ µ Xθ ,[θ] s (v), evaluated at v = θ and multiplied by γ with both random variables defined on a probability space ( Ω, F , P). Taking expectation with respect to P, we get (3.8) In the above equation, we are able to take γ inside the Itô integral with no problem since it is defined on a separate probability space to the Brownian motion, B. We are also able to interchange the order of the Itô integral and expectation with respect to P using a stochastic Fubini theorem (see for example [33,Theorem 65]). Again, since ( θ, γ) are defined on a separate probability space, which we can replace in equation (3.8) to get: Now, taking equation (3.1), satisfied by ∂ x X x,[θ] t and evaluating at x = θ, multiplying by γ and adding to One can therefore see that the equation satisfied by ( θ) γ is the same as equation (3.7) satisfied by DX θ t (γ), so by uniqueness they are equal. This representation also makes clear the linearity and continuity of γ → DX θ t (γ).
Following essentially the same procedure shows that E ∂ µ X (v) exists and satisfies equation (3.4). This representation also makes clear the linearity and continuity of γ → DX (c) Let X θ,n denote the Picard approximation of the solution to the McKean-Vlasov SDE (2.1), given by . Since each derivative process satisfies a linear equation (whose exact form is not important for our purposes) the proof is quite mechanical and reserved to the Appendix A.2. Now we introduce some operators acting on Kusuoka-Stroock processes. These are the building blocks of the integration by parts formulae to come. For the rest of this section, we will need the following uniform ellipticity assumption. We make the assumption that there exists ǫ > 0 such that, for all ξ ∈ R N , z ∈ R N and µ ∈ P 2 (R N ), Now, for a multi-index α on {1, . . . , N }, we introduce the following operators acting on elements of K q r (R, n), defined for α = (i), by For α = (α 1 , . . . , α n ) we inductively define and make analogous definitions for each of the other operators. The following result states that these operators are well-defined and describes how each operator transforms a given Kusuoka-Stroock process. The proof is contained in Appendix A.2.

Integration by parts formulae for the de-coupled equation
Having introduced some operators acting on Kusuoka-Stroock processes, we now show how to use these operators to construct Malliavin weights in integration by parts formulas. We first develop integration by parts formulas for ). In the last part of this section, we will show how to combine these results to construct integration by parts formulas for derivatives of the function x → E f (X x,δx t ).
1. First, we note that equation ( are the same except their initial conditions. It therefore follows that for r ≤ t, This allows us to make the following computations for f ∈ C ∞ b (R N ; R), where we have used Malliavin integration by parts E Dφ, u H d = E [φ δ(u)] in the last line. This proves the result for |α| = 1. By Proposition 3.4, I 1 α (Ψ) ∈ K q+2 r (R, (k ∧ n) − 1) when |α| = 1. We can therefore iterate this argument another |α| − 1 times to obtain the result for all α satisfying |α| ≤ [n ∧ k].
4. This follows from parts 2 and 3.

Integration by parts in the measure variable
We now consider derivatives of the function Proof. 1. We use again that for r ≤ t, This allows us to make the following computations for where we have used Malliavin integration by parts E Dφ, u H d = E [φ δ(u)] in the last line. This proves the claim for |β| = 1. For general β, it follows by iterating this integration by parts |β| times.
3. This follows from parts 1 and 2.

Integration by parts for McKean-Vlasov SDE with fixed initial condition
We now consider developing integration by parts formulae for derivatives of the function We introduce the following operator acting on elements of K q r (R, M ), the set of Kusuoka-Stroock processes on R N . For α = (i) x, δ x ) and inductively, for α = (α 1 , . . . , α n ), Moreover, if Φ ∈ K 0 r (R, k) and V 0 , . . . , V d are uniformly bounded, then Proof. This is a direct result of Proposition 3.4 and Lemma 2.11.
In particular, we get the following bound Proof. By the above discussion, Now, we apply the IBPFs developed earlier in Proposition 4.1 part 3 and Theorem 4.2 part 3.

Connection with PDE
We return our attention to the PDE (1.2). The results of the last section suggest that for initial conditions g(z, µ) = g(z), which do not depend on the measure, we can still expect there to be a classical solution, even if g is not differentiable. Indeed, we spell out the conditions under which this is true in Theorem 5.8. But first, let us consider whether the same can be true for initial conditions which do depend on the measure.
The above example shows that for a function g : R N × P 2 (R N ) → R which is Lipschitz continuous, we cannot, in general, expect [θ] → E g X , [X θ t ] take when g is smooth. The following result is Lemma 5.1 from [8].
Lemma 5.2. We assume that the function g : R N × P 2 (R N ) → R N admits continuous derivatives ∂ x g and ∂ µ g satisfying for some q > 0 and 0 ≤ p < 2 Then, ∂ µ U exists and takes the following form: Now we introduce a class of initial conditions g : R N × P 2 (R N ) → R for which we will be able to develop integration by parts formulas.
There exists a sequence of functions (g l ) l≥1 , g l : R N × P 2 (R N ) → R with polynomial growth such that g l → g uniformly on compacts and ∂ x g l exists and also has polynomial growth for each l ≥ 1. 3. For each l ≥ 1 there exists a function G l : , µ, v). Moreover, each G l and its derivatives satisfies the growth condition: there exist q > 0 and 0 ≤ r < 1 such that for all (x, [θ], v) ∈ R N × P 2 (R N ) × R N : In addition, we assume that for all (x, µ, v) the pointwise limit lim l→∞ G l (x, µ, v) exists and the function G defined by G(x, µ, v) := lim l→∞ G l (x, µ, v) is continuous and satisfies the same growth condition.
If ∂ µ g l = ∂ x G l we say g is in the class (IC) x . If ∂ µ g l = ∂ v G l , we say g is in the class (IC) v .
We give some examples of functions g in the class (IC).

Functions with no dependence on the measure:
Suppose that g(x, µ) = ϕ(x) where ϕ ∈ C p (R N ; R). Then, let (ϕ l ) l≥1 be a sequence of mollifications of ϕ and (g l ) l≥1 the corresponding functions defined in the same way. Then, ∂ µ g l (x, µ, v) = 0. So, g belongs to the class (IC) x and G in this case would be G ≡ 0.

Second order interaction:
Suppose g(x, µ) := ϕ(x, y, z)µ(dy)µ(dz) where ϕ : R 3N → R is continuous with |ϕ(x, y, z)| ≤ C(1 + |x| q + |y| r + |z| r ) for some q > 0 and 0 ≤ r < 1. Then, let (ϕ l ) l≥1 be a sequence of mollifications of ϕ and (g l ) l≥1 the corresponding functions defined in the same way. Then, ∂ µ g l (x, µ, v) = [∂ v ϕ l (x, v, y) + ∂ v ϕ l (x, y, v)] µ(dy). So, g belongs to the class (IC) v and G in this case would be 5. Polynomials on the Wasserstein space: ϕ i (x, y)µ(dy), where n ≥ 1 and each ϕ i : R N × R N → R is continuous with |ϕ i (x, y)| ≤ C(1 + |x| q ) for some q > 0. Then, let (ϕ i,l ) l≥1 be a sequence of mollifications of ϕ i and (g l ) l≥1 the corresponding functions defined in the same way. Then, Therefore g belongs to the class (IC) v and G in this case would be Now, we introduce the hypotheses under which we will be able to prove existence and uniqueness of a solution to the PDE (1.2).
exist and are continuous. Moreover, for all compacts K ⊂ P 2 (R N ) Proof. Under both (H1) and (H2), g is in the class (IC), so there is a sequence of functions (g l ) l≥1 approximating . From Proposition 4.1 we know that for i, j ∈ {1, . . . , N } By the growth assumption on g l , Hölder's inequality and the moment estimates already obtained for the processes X is a Kusuoka-Stroock process) and the continuity of g.
To lighten notation, we restrict to the case N = 1 through the rest of this proof. First, we assume (H1) holds, so g is in the class (IC) x . Note that g l satisfies the hypotheses of Lemma 5.2, which gives Now, we recall the following identity connecting D r X So, and, applying Proposition 4.1 part 2, we get Similarly, and applying Proposition 4.1 part 2 again, we get So, in this case, (5.2) can be rewritten as we note that all processes on the right hands side of (5.3) have moments of all orders bounded polynomially in θ 2 except Xθ t in the final term. For the final term, by the growth conditions on G l , Clearly this is bounded in [θ] over compacts in P 2 (R N ). Now, we consider the derivative ∂ v ∂ µ U l . We note that in the definition of , v) exists and we obtain:

(5.4)
We again use that Of course, this identity also holds for 'tilde' processes defined on Ω ,F,P and we denote by D the Malliavin derivative on this space. So, using the above identity and the Malliavin chain rule, we obtain and, applying the integration by parts formula in Proposition 4.1 on the space Ω ,F ,P , we get So, (5.4) becomes We can check each expectation above is finite by using the growth conditions on the functions g l , G l and their derivatives along with Hölder's inequality and the moment estimates on the processes involved, similar to before. In particular, note that we can obtain estimates on (5.3) and (5.5) independently of l. This allows us to use dominated convergence to pass to the limit in these equations. Now, suppose that (H2) holds instead of (H1). Under (H2), g in the class (IC) v . By Lemma 5.2, we have an expression for ∂ µ U l and using the special form of ∂ µ g l for initial conditions in the class (IC) v , we get We again use that Of course, this identity also holds for 'tilde' processes defined on Ω ,F,P and we denote by D the Malliavin derivative on this space. So, using the above identity and the Malliavin chain rule, we obtain r and, applying the integration by parts formula in Proposition 4.1 on the space Ω ,F ,P , we get Similarly, and applying the integration by parts formula in Proposition 4.2 on the space Ω ,F ,P , we get Here we explain the reason for insisting that the coefficients V 0 , . . . , V d are bounded: the Kusuoka-Stroock process Putting the above integration by parts formulas together and using Proposition 4.2 on the space (Ω, F , P) for the first term on the right hand side of (5.6), we see that it can be re-written as and we note the RHS does not depend on derivatives of the functions g and G. Also, so that, Hence, The idea is to expand the first term using the chain rule introduced in [12] and the second term using Itô's formula. Then, dividing by h and sending it to 0, along with continuity of the terms appearing in the expansion, will prove that U indeed solves the PDE (1.2). Lemma 5.5 guarantees that we can apply the chain rule proved in [12]. We apply it to the function U (t, x, ·) to get Itô's formula applied to U (t, ·, [X θ h ]) gives We want the final term to be square integrable, so that it is a true martingale with zero expectation. We have that for some q > 0, so that for all p ≥ 1, and by the linear growth of V i j , we have Hence, the final term is indeed square integrable, and has zero expectation. Putting the expansions back into (5.11), we get By the earlier results on continuity of U and its derivatives and the a priori continuity of the coefficients V 0 , . . . , V d we see that the integrand on the right hand side is a continuous function of h. Dividing by h and sending it to zero, we see that U solves the PDE (1.2). Uniqueness: Fix any t ∈ (0, T ] and any classical solution W with polynomial growth. Set δ > 0, so the smoothness of p(t, x, z) in the forward variable, z. However, they do not cover the smoothness of the function p(t, x, z) in the backward variable, x. The density p(t, x, z) has also been studied by Antonelli & Kohatsu-Higa in [1] under a Hörmander condition on the coefficients. In this case, they establish smoothness of the density in the forward variable, z, but do not establish estimates on the derivatives of this function. The theorem which follows esatblishes the smoothness of p(t, x, z) in the variables (x, z) and we also obtain estimates on its derivatives.
Theorem 6.1. Let α, β be multi-indices on {1, . . . , N } and let k ≥ |α| + |β| + N + 2. Then, for all t ∈ (0, T ] and θ ∈ L 2 (Ω), X x,δx t has a density p(t, x, ·) such that (x, z) → ∂ α x ∂ β z p(t, x, z) exists and is continuous. Moreover, there exists a constant C which depends on T , N and bounds on the coefficients, such that for all t ∈ (0, T ] where µ = 4|α| + 3|β| + 3N and ν = 1 2 (N + |α| + |β|). If V 0 , . . . , V d are bounded then the following estimate holds Proof. Let η = (1, 2, . . . , N ) and introduce the multi-dimensional indicator function and satisfies ∂ η f = g. Now, we first focus on p(t, x, ·), the density of X x,δx where we have used at each step respectively: ∂ η f = g; Corollary 4.5 ; equation (6.3), and Fubini's theorem. It then follows that, for any R > 0 and t ∈ (0, T ], there exists C = C(R, t) > 0 such that Then, it is a result from Taniguchi [35, Lemma 3.1] that X x,δx t has a density function, p(t, x, ·) and that ∂ α x ∂ β z p(t, x, z) exists. Once we know that a smooth density exists, it follows from (6.4) that we can identify ∂ α x ∂ β z p(t, x, z) as Now, the following estimates come from each term's membership of the Kusuoka-Stroock class, as guaranteed by Proposition 3.4 and Corollary 4.5: This proves the estimate (6.1). In addition, if V 0 , . . . , V d are bounded, we can estimate We can therefore apply the exponential martingale inequality to obtain Then, we use (a + b) 2 ≥ a 2 2 − b 2 , which is re-arrangement of Young's inequality, to get This establishes (6.2).

A.1 Elements of Malliavin Calculus
As indicated in the introduction, we will use some tools from Malliavin Calculus to develop integration by parts formulas. Here we introduce the basic terminology. We follow the exposition in [14], with all proofs contained in the book by Nualart [32]. We denote H d := L 2 ([0, T ]; R d ). and use this space to define the Malliavin derivative.
Definition A.1 (Malliavin Derivative). Let f ∈ C ∞ p (R n ; R), for some n ∈ N, h 1 , . . . , h n ∈ H d and F : Ω → R be the functional given by: Any functional of the form (A.1) is called smooth and we denote the class of all such functionals by S. Then the Malliavin derivative of F , denoted by DF ∈ L 2 (Ω; H d ) is given by: We note the isometry L 2 (Ω × [0, T ]; R d ) ≃ L 2 (Ω; H d ). This allows us to identify DF with a process (D r F ) r∈[0,T ] taking values in R d , which we often do. We also denote by D j r F r∈[0,T ] , j = 1, . . . , d, the components of this process.
The set of smooth functionals (random variables) S is dense in L p (Ω), for any p ≥ 1 and D is closable as operator from L p (Ω) to L p (Ω; H d ). We define D 1,p is the closure of the set S within L p (Ω; R d ) with respect to the norm: The higher order Malliavin derivatives are defined in a similar manner. For smooth random variables, we denote the iterated derivative by D (k) F , k ≥ 2, which is a random variable with values in H ⊗k The above expression for D (k) F coincides with that obtained by iteratively applying the Malliavin derivative. In an analogous way, one can close the operator D (k) from L p (Ω) to L p (Ω; H ⊗k d ). So, for any p ≥ 1 and natural k ≥ 1, we define D k,p to be the closure of S with respect to the norm: Moreover, there is nothing which restricts consideration to R d -valued random variables. Indeed, one can consider more general Hilbert space-valued random variables, and the theory would extend in an appropriate way. To this end, denote D k,p (E) to be the appropriate space of E-valued random variables, where E is some separable Hilbert space. For more details, see [32], where also the proof of the following chain rule formula can be found: Proposition A.2 (Chain Rule for the Malliavin Derivative). If ϕ : R m → R is a continuously differentiable function with bounded partial derivatives, and F = (F 1 , . . . , F m ) is a random vector with components belonging to D 1,p for some p ≥ 1. Then ϕ(F ) ∈ D 1,p , with where ∇ϕ is the row vector (∂ 1 ϕ, . . . , ∂ m ϕ) and DF is the matrix (D j F i ) 1≤i≤m,1≤j≤d .  The divergence operator -which is the adjoint of the Malliavin derivative -plays a vital role in the construction of our integration by parts formula. This operator is also called the Skorohod integral. It coincides with a generalisation of the Itô integral to anticipating integrands. A detailed discussion of the divergence operator can be found in Nualart [32]. 2. For every u ∈ Dom δ, then δ(u) ∈ L 2 (Ω) satisfies: E(F δ(u)) = E( DF, u H d ).
Remark A.5. If u = (u 1 , ..., u d ) ∈ Dom δ is F-adapted, then the adjoint δ(u), is nothing more than the Itô integral of u with respect to the d-dimensional Brownian motion B t = (B 1 t , . . . , B d t ). i.e.

A.2 Proofs from Section 3
The first goal of this section is to prove Theorem 3.2. Since each type of derivative (w.r.t. x, µ or v) of X x,[θ] t satisfies a linear equation, we will introduce a general linear equation and, first, derive some a priori L p estimates on the solution. Then, we will show this linear equation is again differentiable under certain assumptions on the coefficients. In the following, we consider an equation with coefficients a 1 , a 2 , a 3 , which depend on (t, x, [θ], v) ∈ [0, T ] × R N × P 2 (R N ) × (R N ) #v with initial condition given by a constant value a 0 3 Below, we denote v r as one element of the tuple v = (v 1 , . . . , v #v ). Then, for i = 1, . . . , d, g i ∈ K q