Singularly perturbed forward-backward stochastic differential equations: application to the optimal control of bilinear systems

We study linear-quadratic stochastic optimal control problems with bilinear state dependence for which the underlying stochastic differential equation (SDE) consists of slow and fast degrees of freedom. We show that, in the same way in which the underlying dynamics can be well approximated by a reduced order effective dynamics in the time scale limit (using classical homogenziation results), the associated optimal expected cost converges in the time scale limit to an effective optimal cost. This entails that we can well approximate the stochastic optimal control for the whole system by the reduced order stochastic optimal control, which is clearly easier to solve because of lower dimensionality. The approach uses an equivalent formulation of the Hamilton-Jacobi-Bellman (HJB) equation, in terms of forward-backward SDEs (FBSDEs). We exploit the efficient solvability of FBSDEs via a least squares Monte Carlo algorithm and show its applicability by a suitable numerical example.


Introduction.
Stochastic optimal control is one of the important fields in mathematics which has attracted the attention of both pure and applied mathematicians [57,27]. Stochastic control problems also appear in a variety of applications, such as statistics [61,23], financial mathematics [21,55], molecular dynamics [58,34] or materials science [56,4], to mention just a few. For some applictions in science and engineering, such as molecular dynamics [58,62], the high dimensionality of the state space is an important aspect when solving optimal control problems. Another issue when solving optimal control problems by discretising the corresponding dynamic programming equations in space and time are multiscale effects that come into play when the state space dynamics exhibit slow and fast motions.
Here we consider such systems that have slow and fast scales and that are possibly highdimensional. Several techniques have been developed to reduce the spatial dimension of control systems (see e.g. [3,10] and the references therein), but these techniques treat the control as a possibly time-dependent parameter ("open loop control") and do not take into account that the control may be a feedback control that depends on the state variables ("closed loop control"). Clearly, homogenization techniques for stochastic control systems have been extensively studied by applied analysts using a variety of different mathematical tools, including viscosity solutions of the Hamilton-Jacobi-Bellman equation [13,25], backward stochastic differential equations [17,39], or occupation measures [45,46]. The convergence analysis of multiscale stochastic control systems is quite involved and non-constructive, in that the limiting equations of motion are not given in explicit or closed form; see [43,42] for notable exceptions, dealing mainly with the case when the dynamics is linear.
In general, the elimination of variables and solving control problems do not commute, so one of the key questions in control engineering is under which conditions it is possible to eliminate variables before solving an optimal problem. We call this the model reduction problem. In this paper we identify a class of stochastic feedback control problems with bilinear state dependence that have the property that they admit the elimination of variables (i.e. model reduction) before solving the control problem. These systems tuern oput to be relevant in the control of high-dimensional transport PDEs, such as Fokker-Planck equations or the evolution equations of open quantum systems [33,49]. Our approach is based on a Donsker-Varadhan type duality principle between a linear Feynman-Kac PDE and the semi-linear dynamic programming PDE associated with a stochastic control problem [32]. Here we exploit the fact that the dynamic programming PDE can be recast as an uncoupled forward backward stochastic differential equation (see e.g. [54,59]) that can be treated by model reduction techniques, such as averaging or homogenisation.
The relation between semilinear PDEs of Hamilton-Jacobi-Bellman type and forwardbackward stochastic differential equations (FBSDE) is a classical subject that has been first studied by Pardoux and Peng [52] and since then received lot of attention from various sides, e.g. [6,22,24,36,37,44]. The solution theory has its roots in the work of Antonelli [2] and since then has been extended in various directions; see e.g. [5,7,63,47].
From a theoretical point of view, this paper goes beyond our previous works [32,35] in that we prove strong convergence of the value function and the control without relying on compactness or periodicity assumptions for the fast variables, even though we focus on bilinear systems only, which is the weakest form of nonlinearity. (Many nonlinear systems however can be represented as bilinear systems by a so-called Carleman linearisation.) It also goes beyond the classical works [43,42] that treat systems that are either fully linear or linear in the fast variables. We stress that we are mainly aiming at the model reduction problem, but we discuss alongside with the theoretical results some ideas to discretise the corresponding FBSDE [8,19,15,9,38], since one of the main motivations for doing model reduction is to reduce the numerical complexity of solving optimal control problems.
1.1. Set-up and problem statement. We briefly discuss the technical set-up of the control problem considered in this paper. In this paper, we consider the linear-quadratic (LQ) stochastic control problem of the following form: minimize the expected cost over all admissible controls u ∈ U and subject to: Here τ < ∞ is a bounded stopping time (specified below), and the set of admissible controls U is chosen such that (2) has a unique strong solution. The denomination linear-quadratic for (1)-(2) is due to the specific dependence of the system on the control variable u. The state vector x ∈ R n is assumed to be high-dimensional, which is why we seek a low-dimensional approximation of (1)-(2). 2 Specifically, we consider the case that q 0 and q 1 are quadratic in x, a is linear and σ is constant, and the control term is an affine function of x, i.e., In this case the system is called bilinear (including linear systems as a special case), and the aim is to replace (2) by a lower dimensional bilinear system with statesx ∈ R ns , n s ≪ n and an associated reduced cost functional that is solved instead of (1)- (2). Letting v * denote the minimizer ofJ , we require that v * is a good approximation of the minimizer u * of the original problem where "good approximation" is understood in the sense that In the last equation, closeness must be suitably interpreted, e.g. uniformly on all compact subsets of R n × [0, T ) for some T < ∞. One situation in which the above approximation property holds is when u * ≈ v * uniformly in t and the cost is continuous in the control, but it turns out that this requirement will be too strong in general and overly restrictive. We will discuss alternative criteria in the course of this paper.
1.2. Outline. The paper is organised as follows: In Section 2 we introduce the bilinear stochastic control problem studied in this paper and derive the corresponding forwardbackward stochastic differential equation (FBSDE). Section 3 contains the main result, a convergence result for the value function of a singularly perturbed control problem with bilinear state dependence, based on an FBSDE formulation. In Section 4 we present a numerical example to illustrate the theoretical findings and discuss the numerical discretization of the FBSDE. The article concludes in Section 5 with a short summary and a discussion of future work. The proof of the main result and some technical lemmas are recorded in the Appendix.
2. Singularly perturbed bilinear control systems. We now specify the system dynamics (2) and the corresponding cost functional (1). Let (x 1 , x 2 ) ∈ R ns × R n f with n s + n f = n denote a decomposition of the state vector x ∈ R n into relevant (slow) and irrelevant (fast) components. Further let W = (W t ) t≥0 denote R m -valued Brownian motion on a probability space (Ω, F, P ) that is endowed with the filtration (F t ) t≥0 generated by W . For any initial condition x ∈ R n and any A-valued admissible control u ∈ U , with A ⊂ R, we consider the following system of Itô stochastic differential equations that depends parametrically on a parameter ǫ > 0 via the coefficients where for brevity we also drop the dependence of the process on the control u, i.e. X ǫ s = X u,ǫ s . The stiffness matrix A in (3) is assumed to be of the form with n = n s + n f . Control and noise coefficients are given by where N x + B ∈ range(C) for all x ∈ R n ; often we will consider either the case m = 1 with C i = √ ρB i , ρ > 0, or m = n, with C being a multiple of the identity when ǫ = 1. All block matrices A ij , N ij , B i and C j are assumed to be order 1 and independent of ǫ.
The above ǫ-scaling of coefficients is natural for a system with n s slow and n f fast degrees of freedom and arises, for example, as a result of a balancing transformation applied to a large-scale system of equations; see e.g. [31,33]. A special case of (3) is the linear system (7) dX ǫ s = (AX ǫ s + Bu s ) ds + CdW s . Our goal is to control the stochastic dynamics (3)-or (7) as a special variant-so that a given cost criterion is optimised. Specifically, given two symmetric positive semidefinite matrices Q 0 , Q 1 ∈ R ns×ns , we consider the quadratic cost functional that we seek to minimize subject to the dynamics (3). Here the expectation is understood over all realisations of (X ǫ s ) s∈[t,τ ] starting at X ǫ t = x, and as a consequence J is a function of the initial data (t, x). The stopping time is defined as the minimum of some time T < ∞ and the first exit time of a domain D = D s × R n f ⊂ R ns × R n f where D s is an open and bounded set with smooth boundary. Specifically, we set τ = min{τ D , T }, with In other words, τ is the stopping time that is defined by the event that either s = T or X ǫ s leaves the set D = D s × R n f , whichever comes first. Note that the cost function does not explicitly depend on the fast variables x 2 . We define the corresponding value function by Remark. As a consequence of the boundedness of D s ⊂ R ns , we may assume that all coefficients in our control problem are bounded or Lipschitz continuous, which makes some of the proofs in the paper more transparent.
We further note that all of the following considerations trivially carry over to the case N = 0 and a multi-dimensional control variable, i.e., u ∈ R k and B ∈ R n×k .
2.1. From stochastic control to forward-backward stochastic differential equations. We suppose that the matrix pair (A, C) satisfies the Kalman rank condition (10) rank(C|AC|A 2 C| . . . |A n−1 C) = n .
A necessary-and in this case sufficient-condition for optimality of our optimal control problem is that the value function (9) solves a semilinear parabolic partial differential equation of Hamilton-Jacobi-Bellman type (a.k.a. dynamic programming equation) [26] Here L ǫ is the infinitesimal generator of the control-free process, and the nonlinearity f is independent of ǫ and given by Note that f is furthermore independent of y and that the Moore-Penrose pseudoinverse is unambiguously defined since z = C ⊤ ∇V ǫ and (N x + B) ∈ range(C), which by noting that C ⊤ ♯ C ⊤ is the orthogonal projection onto range(C) implies that The specific semilinear form of the equation is a consequence of the control problem being linear-quadratic. As a consequence, the dynamic programming equation (11) admits a representation in form of an uncoupled forward-backward stochastic differential equation (FBSDE). To appreciate this point, consider the control-free process X ǫ s = X ǫ,u=0 s with infinitesimal generator L ǫ and define an adapted process (We abuse notation and denote both the controlled and the uncontrolled process by X ǫ s .) Then, by definition, Y ǫ t = V ǫ (x, t). Moreover, by Itô's formula and the dynamic programming equation (11), the pair (X ǫ s , Y ǫ s ) s∈[t,τ ] can be shown to solve the system of equations Here, the second equation is only meaningful if interpreted as a backward equation, since only in this case Z ǫ s is uniquely defined. To see this, let f = 0 and q 1 (x) = x and note that the ansatz (14) implies that Y ǫ s is adapted to the filtration generated by the forward process X ǫ s . If the second equation was just a time- But such a solution would not be adapted, because Y ǫ s for s < τ would depend on the future value X ǫ τ of the forward process. Remark. Equation (15) is called an uncoupled FBSDE because the forward equation for X ǫ s is independent of Y ǫ s or Z ǫ s . The fact that the FBSDE is uncoupled furnishes a wellknown duality relation between the value function of an LQ optimal control problem and the cumulate generating function of the cost [18,20]; specifically, in the case that N = 0, B = C and the pair (A, B) being completely controllable, it holds that Here the expectation on the right hand side is taken over all realisations of the control-free process X ǫ s = X ǫ,u=0 s , starting at X ǫ t = x. By the Feynman-Kac theorem, the function ψ ǫ = exp(−V ǫ ) solves the linear parabolic boundary value problem which is equivalent to the corresponding dynamic programming equation (11).
3. Model reduction. The idea now is to exploit the fact that (15) is uncoupled, which allows us to derive an FBSDE for the slow variablesX ǫ s = X ǫ 1,s only, by standard singular perturbation methods. The reduced FBSDE as ǫ → 0 will then be of the form where the limiting form of the backward SDE follows from the corresponding properties of the forward SDE. Specifically, assuming that the solution of the associated SDE that is governing the fast dynamics as ǫ → 0, is ergodic with unique Gaussian invariant measure π = N (0, Σ), where Σ = Σ ⊤ > 0 is the unique solution to the Lyapunov equation we obtain that, asymptotically as ǫ → 0, As a consequence, the limiting SDE governing the evolution of the slow process X ǫ 1,s -in other words: the forward part of (18)-has the coefficients as following from standard homogenisation arguments [53]; a formal derivation is given in the appendix. By a similar reasoning we find that the driver of the limiting backward SDE reads specifically, The limiting backward SDE is equipped with a terminal conditionq 1 that equals q 1 , namely, Interpretation as an optimal control problem. It is possible to interpret the reduced FBSDE again as the probabilistic version of a dynamic programming equation. To this end, note that (10) implies that the matrix pair (Ā,C) satisfies the Kalman rank condition [1] rank(C|AC|A 2C | . . . |A ns−1C ) = n s .
As a consequence, the semilinear partial differential equation has a classical solution V ∈ C 1,2 ([0, T )×D)∩C 0,1 (E + s ). LettingȲ s := V (s,X s ), 0 t s τ , with initial dataX t = x 1 andZ s =C ⊤ ∇V (s,X s ), the limiting FBSDE (18) can be readily seen to be equivalent to (27). The latter is the dynamic programming equation of the following LQ optimal control problem: minimize the cost functional where (w s ) s≥0 denotes standard Brownian motion in R ns and we have introduced the new control coefficientsM =CN andD =CB.
3.1. Convergence of the control value. Before we state our main result and discuss its implications for the model reduction of linear and bilinear systems, we recall that basic assumptions that we impose on the system dynamics. Specifically, we say that the dynamics (3) and the corresponding cost functional (8) (15) is bounded; for simplicity we set Q 1 = 0 in (8). Assumption 2 implies that the fast subsystem (19) has a unique Gaussian invariant measure π = N (0, Σ) with full topological support, i.e., we have Σ = Σ ⊤ > 0. According to [11,Prop. 3.1] and [44], existence and uniqueness of (15) is guaranteed by Assumptions 3 and 4 and the controllability of (A, C) and the range condition, which imply that the transition probability densities of the (controlled or uncontrolled) forward process X ǫ s are smooth and strictly positive. As a consequence of the complete controllability of the original system, the reduced system (30) is completely controllable too, which guarantees existence and uniqueness of a classical solution of the limiting dynamic programming equation (27); see, e.g., [50].

The terminal condition in
Uniform convergence of the value function V ǫ → V is now entailed by the strong convergence of the solution to the corresponding FBSDE as is expressed by the following Theorem. The proof of the Theorem is given in Appendix A.2. For the reader's convenience, we present a formal derivation of the limit equation in the next subsection.

Formal derivation of the limiting FBSDE.
Our derivation of the limit FBSDE follows standard homogenisation arguments (see [29,41,53]), taking advantage of the fact that the FBSDE is uncoupled. To this end we consider the following linear evolution equation is the generator associated with the control-free forward process X ǫ s in (15). We follow the standard procedure of [53] and consider the perturbative expansion that we insert into the Kolmogorov equation (31). Equating different powers of ǫ we find a hierarchy of equations, the first three of which read Assumption 2 on page 8 implies that L 0 has a one-dimensional nullspace that is spanned by functions that are constant in x 2 , and thus the first of the three equations implies that φ 0 is independent of x 1 . Hence the second equation-the cell problem-reads The last equation has a solution by the Fredholm alternative, since the right hand side averages to zero under the invariant measure π of the fast dynamics that is generated by the operator L 0 , in other words, the right hand side of the linear equation is orthogonal to the nullspace of L * 0 spanned by the density of π. 1 The form of the equation suggests the general ansatz where the function R plays no role in what follows, so we set it equal to zero. Since L 0 ψ = −(A 12 x 2 ) ⊤ , the function ψ must be of the form ψ = Qx 2 with a matrix Q ∈ R ns×n f . Hence 1 Here L * 0 is the formal L 2 adjoint of the operator L0, defined on a suitable dense subspace of L 2 . 9 Now, solvability of the last of the three equations requires again that the right hand side averages to zero under π, i.e.
which formally yields the limiting equation for φ = φ 0 (x 1 , t). Since π is a Gaussian measure with mean 0 and covariance Σ given by (20), the integral (36) can be explicitly computed: whereL is given by (28) and the initial condition φ(·, 0) = g is a consequence of the fact that the initial condition in (31) is independent of ǫ. By the controllability of the pair (Ā,C), the limiting equation (37) has a unique classical solution and uniform convergence φ ǫ → φ is guaranteed by standard results, e.g., [53,Thm. 20.1].
Since the backward part of (15) is independent of ǫ, the final form of the homogenised FBSDE (18) is found by averaging over x 2 , with the unique solution of the corresponding backward SDE satisfying Z 2,s = 0 as the averaged backward process is independent of x 2 .
4. Numerical studies. In this section we presents numerical results for linear and bilinear control systems and discuss the numerical discretisation of uncoupled FBSDE associated with LQ stochastic control problems. We begin with the latter.

Numerical FBSDE discretisation.
The fact that (15) or (18) are decoupled entails that they can be discretised by an explicit time-stepping algorithm. Here we utilize a variant of the least-squares Monte Carlo algorithm proposed in [9]; see also [30]. The convergence of numerical schemes for FBSDE with quadratic nonlinearities in the driver has been analysed in [60].
The least-squares Monte Carlo scheme is based on the Euler discretisation of (15): where (X n ,Ŷ n ) denotes the numerical discretisation of the joint process (X ǫ s , Y ǫ s ), where we set X ǫ s = X ǫ τ D for s ∈ (τ D , T ] when τ D < T , and (ξ k ) k≥1 is an i.i.d. sequence of normalised Gaussian random variables. Now let F n = σ Ŵ k : 0 k n be the σ-algebra generated by the discrete Brownian motionŴ n := √ ∆t i n ξ i . By definition the joint process (X ǫ s , Y ǫ s ) is adapted to the filtration generated by (W r ) 0 r s , therefore where we have used thatẐ n is independent of ξ n+1 . In order to computeŶ n fromŶ n+1 we use the identification of Z ǫ s with C ⊤ ∇V ǫ (s, X ǫ s ) and replaceẐ n in (39) by (40)Ẑ n = C ⊤ ∇V ǫ (t n ,X n ) , which, the parametric ansatz (42) for the V ǫ makes the overall scheme explicit inX n andŶ n .
Least-squares solution of the backward SDE. In order to evaluate the conditional ex-pectationŶ n = E[·|F n ] we recall that a conditional expectation can be characterised as the solution to the following quadratic minimisation problem: Given M independent realisationsX (i) n , i = 1, . . . , M of the forward processX n , this suggests the approximation scheme (Note that N = N D is random.) For simplicity, we assume in what follows that the terminal value is zero, i.e., we set q 1 = 0. (Recall that the existence and uniqueness result from [44] requires q 1 to be bounded.) To representŶ n as a function Y (X n ) we use the ansatz with coefficients α 1 (·), . . . , α K (·) ∈ R and suitable basis functions ϕ 1 , . . . , ϕ K : R n → R (e.g. Gaussians). Note that the coefficients α k are the unknowns in the least-squares problem (41) and thus are independent of the realisation. Now the least-squares problem that has to be solved in the n-th step of the backward iteration is of the form with coefficients ,...,M ;k=1,...,K and data Assuming that the coefficient matrix A n ∈ R M ×K , K M defined by (44) has maximum rank K, then the solution to the least-squares problem (43) is given by The thus defined scheme is strongly convergent of order 1/2 as ∆t → 0 and M, K → ∞ as has been analysed by [9]. Controlling the approximation quality for finite values ∆t, M, K, however, requires a careful adjustment of the simulation parameters and appropriate basis functions, especially with regard to the condition number of the matrix A n . 11 4.2. Numerical example. Illustrating our theoretical findings of Theorem 3.1, we consider a linear system of form (7) where the matrices A, B and C are given by and This is an instance of a controlled Langevin equation with friction and noise coeffcient γ, σ > 0 which are assumed to fulfill the fluctation-dissipation relation 2γ = σ 2 .
In the example we let γ = 1/2 and σ = 1. The quadratic cost functional (8) is determined by the running cost via Q 0 = I n×n ∈ R n×n and we apply no terminal cost, i.e. Q 1 = 0. The associated effective equations are given by (29)-(30), wherē We apply the previously described FBSDE scheme (38), (42),(43)- (46), which was shown to yield good results in [40], to both the full and the reduced system, and we choose n = 3, i.e the full system is six dimensional. To this end we choose the basis functions where δ = 0.1 is fixed but µ k = µ k (n) changes in each timestep such that the basis follows the forward process. For this, we simulate K additional forward trajectories X (k) , k = 1, . . . , K and set µ k (n) = X (k) n . We choose the parameters for the numerics as follows. The number of basis functions K is given by K = 9 for the reduced system and K ǫ = 40 for the full system. We choose these values because the maximally observed rank of the matrices A n defined in (44) is 9 for the reduced system and we want these matrices to have rank K. For the full system we could have used a greater values for K, but we want to keep the computational effort reasonable. Further, we choose ∆t = 5 · 10 −5 , the final time T = 0.5 and the number of realisations M = 400. We let the whole algorithm run five times and compute the distance between the value functions of the full and reduced systems for which convergence of order 1/2 was found in the proof of Theorem 3.1. Indeed, this is the order of convergence which we also observe in the numerics of our example as can be seen in figure 1 where we depict the mean and standard deviation of E(ǫ).

Discussion.
We shall now discuss the implications of the above simple example when it comes to more complicated dynamical systems. As a general remark the results show that it is possible to to apply model reduction before solving the corresponding optimal control problem where the control variable in the original equation can simply be treated as a parameter. This is in accordance with the general model reduction strategy in control engineering; see e.g. [3,10] and the references therein. Our results not only guarantee convergence of the value function via convergence of Y ǫ , but they also imply strong convergence of the optimal control, by the convergence of the control process Z ǫ in L 2 . (See the appendix for details.) This means that in the case of a system with time scale separation, our result is highly valuable since we can resort to the reduced system for finding the optimal control which can then be applied to the full systems dynamics.
We stress that our results carry over to fully nonlinear stochastic control problems which have a similar LQ structure [32]. Clearly, for realistic (i.e. high-dimensional or nonlinear) systems the identification of a small parameter ǫ remains challenging, and one has to resort to e.g. semi-empirical approaches, such as [28,48].
If the dynamics is linear, as is the case here, small parameters may be identified using system theoretic arguments based on balancing transformations (see, e.g., [31,33]). These approaches require that the dynamics is either linear or bilinear in the state variables, but the aforementioned duality for the quasi-linear dynamic programming equation can be used here as well in order to change the drift of the forward SDE from some nonlinear vector field, say, b to a linear vector field b 0 = Ax. Assuming that the noise coefficient C is square and invertible and ignoring ǫ and the boundary condition for the moment, it is easy to see that 13 the dynamic programming PDE (11) can be recast as is the generator of a forward SDE with nonlinear drift b, and is the driver of the corresponding backward SDE. Even though the change of drift is somewhat arbitrary, it shows that by changing the driver in the backward SDE it is possible to reduce the control problem to one with linear drift that falls within the category that is considered in this paper, at the expense of having a possibly non-quadratic cost functional.
Remark. Changing the drift may be advantageous in connection with the numerical FB-SDE solver. In the martingale basis approach of Bender and Steiner [9], the authors have suggested to use basis functions that are defined as conditional expectations of certain linearly independent candidate functions over the forward process, which makes the basis functions martingales. Computing the martingale basis, however, comes with a large computational overhead, which is why the authors consider only cases in which the conditional expectations can be computed analytically. Changing the drift of the forward SDE may thus be used to simplify the forward dynamics so that its distribution becomes analytically tractable.

Conclusions and outlook.
We have given a proof of concept that model reduction methods for singularly perturbed bilinear control systems can be applied to the dynamics, before solving the corresponding optimal control problem. The key idea that to connect the HJB corresponding to our stochastic optimal control which is a semi-linear PDE to a perturbed forward-backward SDE which is decoupled, so we benefit from this end to derive a reduced FBSDE as the perturbation parameter ǫ goes to zero. The reduced FBSDE can then be interpreted as a reduced stochastic control problem, and we have proved uniform convergence of the corresponding value function. As an auxiliary result, we obtain that the optimal control converges as well in a strong sense, which implies that the optimal control computed from the reduced system can be used to control the original dynamics.
We presented numerical results for linear control system and we discussed the numerical discretisation of uncoupled FBSDE, based on the computation of conditional expectations. For the latter the choice of the basis functions plays an essential role, and how to cleverly choose the ansatz functions, possibly exploiting that the forward SDE has an explicit solution (see e.g. [9]) is an important topic, especially for high dimensional problems. We leave the question regarding the adaptive choice of ansatz functions to future work.
Another class of important problems, that we have not considered in this article, are slowfast systems with vanishing noise. The natural question here is how the limit equation depend on the order in which noise and time scale parameters go to zero. This question has important consequences for the associated deterministic control problems and its regularisation by noise. We leave this topic for future work too. 14 Appendix A. Proofs and technical lemmas. The idea of the proof of Theorem 3.1 closely follows the work [16], with the main differences being (a) that we consider slow-fast systems exhibiting three time scales, in particular the slow equation contains singular O(ǫ −1/2 ) terms, and (b) that the coefficients of the fast dynamics are not periodic, with the fast process being asymptotically Gaussian as ǫ → 0; in particular the n f -dimensional fast process lives on the unbounded domain R n f .
A.1. Poisson equation Lemma. Theorem 3.1 rests on the following Lemma that is similar to a result in [12]. . Further assume that h is centered with respect to the invariant measure π of the fast process. Then for every t ∈ [0, T ] and initial conditions (X ǫ 1,u , X ǫ 2,u ) = (x 1 , x 2 ) ∈ R ns × R n f , 0 u < t, we have Proof. We remind the reader of the definition (33) of the differential operators L 0 , L 1 and L 2 , and consider the Poisson equation where M 1 and M 2 are square integrable martingales with respect to the natural filtration generated by the Brownian motion W s : M 1 (u, v) = v u (∂ t ψ + L 2 ψ)(s, X ǫ 1,s , X ǫ 2,s )ds + 1 √ ǫ v u L 1 ψ(s, X ǫ 1,s , X ǫ 2,s )ds By the properties of the solution to (48) the first three integrals on the right hand side are uniformly bounded in u and v, and thus v u h(s, X ǫ 1,s , X ǫ 2,s )ds = − ǫδψ(u, v) + ǫ v u (∂ t ψ + L 2 ψ)(s, X ǫ 1,s , X ǫ 2,s )ds + √ ǫ v u L 1 ψ(s, X ǫ 1,s , X ǫ 2,s )ds + ǫM 1 (u, v) + √ ǫM 2 (u, v) .
By the Itô isometry and the boundedness of the derivatives ∇ x 1 ψ and ∇ x 2 ψ, the martingale term can be bounded by Hence E v u h(s, X ǫ 1,s , X ǫ 2,s )ds 2 Cǫ , with a generic constant 0 < C < ∞ that is independent of u, v and ǫ.
A.2. Convergence of the value function.
Lemma A.2. Suppose that Condition LQ from page 8 holds. Then where V ǫ is the solution of the original dynamic programming equation (11) and V is the solution of the limiting dynamic programming equation (27).
The constant and C depends on x and t, but is finite on every compact subset of D × [0, T ].