Abstract
We consider a simultaneous small noise limit for a singularly perturbed coupled diffusion described by
where \(B_t, W_t\) are independent Brownian motions on \({\mathbb R}^d\) and \({\mathbb R}^m\) respectively, \(b : \mathbb {R}^d \times \mathbb {R}^m \rightarrow \mathbb {R}^d\), \(U : \mathbb {R}^d \times \mathbb {R}^m \rightarrow \mathbb {R}\) and \(s :(0,\infty ) \rightarrow (0,\infty )\). We impose regularity assumptions on b, U and let \(0< \alpha < 1.\) When \(s(\varepsilon )\) goes to zero slower than a prescribed rate as \(\varepsilon \rightarrow 0\), we characterize all weak limit points of \(X^{\varepsilon }\), as \(\varepsilon \rightarrow 0\), as solutions to a differential equation driven by a measurable vector field. Under an additional assumption on the behaviour of \(U(x, \cdot )\) at its global minima we characterize all limit points as Filippov solutions to the differential equation.
Similar content being viewed by others
Notes
Carathéodory solutions relax the classical requirement that the solution must follow the direction of the vector field at all times: the differential equation need not be satisfied on some set of measure zero on [0, T]. See [28] for a precise definition.
Proposition B.1, presented in Appendix, is a more general nonlinear filtering equation and could be of independent interest.
References
Arapostathis, A., Borkar, V.S., Ghosh, M.K.: Ergodic Control of Diffusion Processes, Volume 143 of Encyclopedia of Mathematics and its Applications. Cambridge University Press, Cambridge (2012)
Athreya, K .B., Hwang, Chii-Ruey: Gibbs measures asymptotics. Sankhya A 72(1), 191–207 (2010)
Allinger, D.F., Mitter, Sanjoy K.: New results on the innovations problem for non-linear filtering. Stochastics 4(4), 339–348 (1981)
Athreya, S.R., Sunder, V.S.: Measure & Probability. Universities Press, CRC Press, Hyderabad (2008)
Biswas, A., Borkar, Vivek S.: Small noise asymptotics for invariant densities for a class of diffusions: a control theoretic view. J. Math. Anal. Appl. 360(2), 476–484 (2009)
Bain, A., Crisan, Dan: Fundamentals of stochastic filtering, volume 60 of Stochastic Modelling and Applied Probability, vol. 60. Springer, New York (2009)
Bovier, A., Eckhoff, M., Gayrard, V., Klein, Markus: Metastability and small eigenvalues in Markov chains. J. Phys. A 33(46), L447–L451 (2000)
Bovier, A., Eckhoff, M., Gayrard, V., Klein, M.: Metastability in stochastic dynamics of disordered mean-field models. Prob. Theory Relat. Fields 119(1), 99–161 (2001)
Bovier, A., Eckhoff, M., Gayrard, V., Klein, Markus: Metastability in reversible diffusion processes. I. Sharp asymptotics for capacities and exit times. J. Eur. Math. Soc. (JEMS) 6(4), 399–424 (2004)
Bakry, D., Gentil, I., Ledoux, Michel: Analysis and Geometry of Markov Diffusion Operators, Volume 348 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 348. Springer, Cham (2014)
Billingsley, Patrick: Convergence of Probability Measures. Wiley, New York (1968)
Buckdahn, R., Ouknine, Y., Quincampoix, M.: On limiting values of stochastic differential equations with small noise intensity tending to zero. Bull. Sci. Math. 133(3), 229–237 (2009)
Borkar, Vivek S: Probability Theory. Universitext. Springer, New York (1995). An advanced course
Chiang, T.-S., Hwang, C.-R., Sheu, Shuenn Jyi: Diffusion for global optimization in \({\bf R}^n\). SIAM J. Control Optim. 25(3), 737–753 (1987)
Delarue, F., Flandoli, Franco: The transition point in the zero noise limit for a 1D Peano example. Discret. Contin. Dyn. Syst. Ser. A 34, 4071–4084 (2014)
Eckhoff, Michael: Precise asymptotics of small eigenvalues of reversible diffusions in the metastable regime. Ann. Prob. 33(1), 244–299 (2005)
Eckmann, J.-P., Ruelle, D.: Ergodic theory of chaos and strange attractors. Rev. Mod. Phys. 57(3, part 1), 617–656 (1985)
Fleming, W.H., Soner, H Mete: Controlled Markov Processes and Viscosity Solutions, Volume 25 of Stochastic Modelling and Applied Probability, second edn. Springer, New York (2006)
Freidlin, M.I., Wentzell, A.D.: Random perturbations of dynamical systems, volume 260 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer, Heidelberg, 3rd edition, (2012). Translated from the 1979 Russian original by Joseph Szücs
Foster, D., Young, Peyton: Stochastic evolutionary game dynamics. Theor. Popul. Biol. 38(2), 219–232 (1990)
Gelfand, S.B., Mitter, Sanjoy K.: Recursive stochastic algorithms for global optimization in \({ R}^d\). SIAM J. Control Optim. 29(5), 999–1018 (1991)
Gailus, S., Spiliopoulos, Konstantinos: Statistical inference for perturbed multiscale dynamical systems. Stoch. Process. Appl. 127(2), 419–448 (2017)
Haussmann, U.G., de Zakai, L’équation: et le problème séparé du contrôle optimal stochastique. In: Séminaire de probabilités, XIX, 1983/84, volume 1123 of Lecture Notes in Math., pp. 37–62. Springer, Berlin, (1985)
Hwang, C.-R., Sheu, Shuenn Jyi: Large-time behavior of perturbed diffusion Markov processes with applications to the second eigenvalue problem for Fokker-Planck operators and simulated annealing. Acta Appl. Math. 19(3), 253–295 (1990)
Hwang, Chii-Ruey: Laplace’s method revisited: weak convergence of probability measures. Ann. Prob. 8(6), 1177–1182 (1980)
Ikeda, N., Watanabe, S.: Stochastic Differential Equations and Diffusion Processes, Volume 24 of North-Holland Mathematical Library, 2nd edn. North-Holland Publishing Co., Kodansha, Ltd., Amsterdam (1989)
Jongen, H.T., Weber, G.-W.: On parametric non-linear programming. Ann. Opr. Res 27, 253–284 (1990)
John, S.S., Biles, Daniel C: A comparison of the Carathéodory and Filippov solution sets. J. Math. Anal. Appl. 198(2), 571–580 (1996)
Kurtz, T.G., Ocone, D.L.: Unique characterization of conditional distributions in nonlinear filtering. Ann. Prob. 16(1), 80–107 (1988)
Kabanov, Y., Pergamenshchikov, S.: Two-Scale Stochastic Systems, Volume 49 of Applications of Mathematics. Asymptotic Analysis and Control, Stochastic Modelling and Applied Probability, vol. 49. Springer, Berlin (2003)
Kunita, H.: Cauchy problem for stochastic partial differential equations arising in nonlinear filtering theory. Systems Control Lett. 1(1), 37–41 (1981/82)
Liptser, R.: Large deviations for two scaled diffusions. Prob. Theory Relat. Fields 106(1), 71–104 (1996)
Metafune, Giorgio, Pallara, Diego, Rhandi, Abdelaziz: Global properties of invariant measures. J. Funct. Anal. 223(2), 396–424 (2005)
Morse, M.R., Spiliopoulos, Konstantinos: Moderate deviations for systems of slow-fast diffusions. Asymptot. Anal. 105,3–4, 97–135 (2017)
Puhalskii, A.: On large deviations of coupled diffusions with time scale separation. Ann. Prob. 44(64), 3111–3186 (2016)
Priola, E., Wang, F.-Y.: Gradient estimates for diffusion semigroups with singular coefficients. J. Funct. Anal. 236, 244–264 (2006)
Renardy, M., Rogers, R.C.: An Introduction to Partial Differential Equations, vol. 13. Springer, New York (2006)
Sastry, S Shankar: The effects of small noise on implicitly defined nonlinear dynamical systems. IEEE Trans. Circ. Syst. 30(9), 651–663 (1983)
Spiliopoulos, Konstantinos: Large deviations and importance sampling for systems of slow-fast motion. Appl. Math. Optim. 67(1), 123–161 (2013)
Spiliopoulos, Konstantinos: Fluctuation analysis and short time asymptotics for multiple scales diffusion processes. Stochastics and Dynamics 14(03), 1350026 (2014)
Ventcel, A.D.: The asymptotic behavior of the eigenvalues of matrices with elements of the order \({\rm exp}\{-V_{ij}/(2\varepsilon ^{2})\}\). Dokl. Akad. Nauk SSSR 202, 263–265 (1972)
Veretennikov, AYu.: Large deviations in averaging principle for stochastic differential equation systems (noncompact case). Stoch. Stoch. Rep. 48(1–2), 83–96 (1994)
Veretennikov, A.Y.: On large deviations in the averaging principle for SDEs with a “full dependence”. Ann. Prob. 27(1), 284–296 (1999)
Veretennikov, A.Y.: On large deviations for SDEs with small diffusion and averaging. Stoch. Process. Appl. 89(1), 69–79 (2000)
Veretennikov, Alexander: On large deviations in the averaging principle for SDE’s with a “full dependence”, revisited [mr1681106]. Discret. Contin. Dyn. Syst. Ser. B 18(2), 523–549 (2013)
Wong, E.: Representation of martingales, quadratic variation and applications. SIAM J. Control 9, 621–633 (1971)
Acknowledgements
Research of S.R.A. was supported in part by ISF-UGC grant, research of V.S.B. was supported in part by a J. C. Bose Fellowship, research of K.S.K. was supported in part by the grant MTR/2017/000416 from SERB and research of R.S. was supported in part by RBCCPS-IISc. S.R.A., V.S.B. and R.S. would like to thank the International Centre for Theoretical Sciences (ICTS) for hospitality during the Large deviation theory in statistical physics: Recent advances and future challenges (Code:ICTS/Prog-ldt/2017/8). The authors thank Sanjoy Mitter for pointing out the reference [29], Laurent Miclo and Patrick Cattiaux for suggestions on the spectral gap estimate in Proposition 2.1(b), and Konstantinos Spiliopoulos for pointing out several references in the literature.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix A: Existence, Uniqueness, and Gradient Estimates
In this section we show that the coupled system (4) and (5) has a unique strong solution. We begin with a technical lemma.
Lemma A.1
Under (U1), (8) and (9) in assumption (U2) there is \(K_4 >0\) and \(R^\prime \ge R\) such that
Also, there exists a nonnegative continuous function \(g: (0,\infty ) \rightarrow (0,\infty )\) such that
Proof
We proceed as follows. Let \(\mathbb {B}_a\) denote the closed ball of radius a centred at the origin. For any y with \(||y|| > R\), writing \(\nabla _y U(x,y) = \nabla _y U(x,0) + \int _0^1 D^2_y U(x,ty) y ~dt\), we have
for any \(K_4 > 0\) and \(\Vert y\Vert \ge R^\prime \ge R + M(1+R)/K_3 + K_4 / K_3\). In the second inequality above, we have used (9) for the line segment joining 0 to y that lies within \(\mathbb {B}_R\) and (8) for the remaining line segment. This establishes (70).
Next, for any \(y,z \in {\mathbb R}^m\), define \(t_0(y,z)\) to be the fractional length of the line segment joining y to z that is within \(\mathbb {B}_R\). Take \(R_1 = R(1+2Mm/K_3)\). With \(r = ||y-z||\), we can write
The inequality in (72) follows because:
-
(a)
from (9), on account of \(||D^2_yU(x,y+t(z-y))|| \le M\) when \(y+t(z-y) \in \mathbb {B}_R\), we easily obtain the simple inequality \(\langle (z-y), D_y^2(x, y+t(z-y))(z-y) \rangle \ge -Mmr^2\) using which the first term is obtained; and
-
(b)
from (8), \(\langle (z-y), D_y^2(x, y+t(z-y))(z-y) \rangle \ge K_3 r^2\) when \(y+t(z-y)\) is outside \(\mathbb {B}_R\).
The inequality in (73) follows from the easily verifiable fact
for every y, z such that one of them is outside \(\mathbb {B}_{R_1}\).
From (73), it is clear that we may take \(g(\cdot )\) to be any continuous function that dominates the function \(Mmr \cdot \mathbf{1}\{ r \le 2R_1\}\), and there is at least one such \(g(\cdot )\) that satisfies \(\int _0^{\infty } g(s) ds < \infty \). \(\square \)
Given Brownian motion \(B_t\) on \({\mathbb R}^d\) and an independent Brownian motion \(W_t\) on a filtered probability space \((\Omega , \mathcal{F}, {\mathbb {P}})\), a strong solution to the coupled system (4) and (5) is a continuous process \((X^{\varepsilon }_t, Y^{\varepsilon }_t)\) that is adapted to the complete filtration generated by B, W, and satisfies (4) and (5). We say that strong uniqueness holds for the coupled system (4) and (5) if whenever \((X^{\varepsilon }_t, Y^{\varepsilon }_t)\) and \((\tilde{X}^{\varepsilon }_t, \tilde{Y}^{\varepsilon }_t)\) are two strong solutions of the coupled system (4) and (5) with the common initial condition \(x_0,y_0\), then \({\mathbb {P}}((X^{\varepsilon }_t, Y^{\varepsilon }_t) = (\tilde{X}^{\varepsilon }_t, \tilde{Y}^{\varepsilon }_t) \text{ for } \text{ all } t \ge 0) = 1.\)
Lemma A.2
Assume (B1), (U1) and (U2). Let \(\varepsilon >0, 0< \alpha < 1\) and \(s(\varepsilon ) >0\) be given. The coupled system given by (4) and (5) has a unique strong solution.
Proof
By assumptions (B1) and (U1), we know that \(b: {\mathbb R}^d \times {\mathbb R}^m { \rightarrow \mathbb {R}^d}\) and \(\nabla _yU: {\mathbb R}^d \times {\mathbb R}^m { \rightarrow \mathbb {R}^m}\) are locally Lipschitz functions. By [26, page 178 Theorem 3.1] there exist a unique strong solution, \(({X}^{\varepsilon }_t, {Y}^{\varepsilon }_t)_{0 \le t < \zeta }\) where
We will now establish nonexplosiveness of the process. Let \(f : {\mathbb R}^d \times {\mathbb R}^m \rightarrow [0, \infty )\) be given by \(f(x,y) = ~\Vert x \Vert ^2 + \Vert y \Vert ^2\). Let
Clearly \(\sigma _n \le \zeta \) almost surely for all \(n \ge 1\). Let \(t >0\) be given. Applying Ito’s formula at time \(\sigma _n \wedge t\), we obtain that
Using the fact that b is bounded from assumption (B1) we have, for \(r >0\),
Using (9) from assumption (U2) and (70) derived above we have, for \(r >0\),
Substituting (76) and (77) in (75) we have
where the penultimate inequality uses \(c_3 {\mathbb E}[\Vert X^{\varepsilon }_r\Vert ] \le c_5(1+r)\) for a suitable \(c_5\), a fact that follows from (4) and the boundedness assumption on b in (U1). As \(\sigma _n \rightarrow \zeta \) almost surely, the above would imply
Thus if \(\zeta < t\) then we have a contradiction, as the left-hand side is infinity and the right-hand side is finite. As \(t >0\) was arbitrary, we have \(\zeta = \infty \) almost surely. This establishes nonexplosiveness of the process and completes the proof of strong uniqueness. \(\square \)
Appendix B: Nonlinear Filtering Equation
Let \(x_0 \in {\mathbb R}^d\), \(y_0\in {\mathbb R}^m\), \(\sigma _1 >0\) and \(\sigma _2 >0\). On the probability space \((\Omega , \mathcal{F}, {\mathbb {P}})\) let \(\{B_t\}_{t \ge 0}\) and \(\{W_t\}_{t \ge 0}\) be Brownian motions on \({\mathbb R}^d\) and \({\mathbb R}^m\) respectively. In this section, we consider the coupled diffusion \((X_t, Y_t)_{t \in [0,T]}\) on \(\mathbb {R}^d\times \mathbb {R}^m\) described by
where \(0 \le t \le T\), \(b_1 : \mathbb {R}^d \times \mathbb {R}^m \rightarrow \mathbb {R}^d\) and \(b_2 : \mathbb {R}^d \times \mathbb {R}^m \rightarrow \mathbb {R}^m.\) We will make the following assumptions:
-
\(b_1 \in C_b({\mathbb R}^d \times {\mathbb R}^m)\) is locally Lipschitz continuous in y-variable and is uniformly (w.r.t. y) Lipschitz continuous in x-variable, i.e. \(\exists K_1 > 0 \) such that \(\forall \ x, x' \in {\mathbb R}^d, y \in {\mathbb R}^m\)
$$\begin{aligned} \parallel b_1(x, y) - b_1(x', y)\parallel ~ \le K_1 \parallel x-x'\parallel . \end{aligned}$$ -
\(b_2 \in C^1(\mathbb {R}^d \times \mathbb {R}^m).\) Further, \(b_2(x, y)\) is uniformly (w.r.t. y) Lipschitz continuous in x-variable, i.e. \(\exists K_2 > 0 \) such that \(\forall \ x, x' \in {\mathbb R}^d\), \( y \in {\mathbb R}^m\),
$$\begin{aligned} \parallel b_2(x, y) - b_2(x', y)\parallel ~ \le K_2 \parallel x-x'\parallel . \end{aligned}$$ -
There exists \(R>0, M>0\) such that for all \(x \in {\mathbb R}^d\)
$$\begin{aligned} \sup _{\Vert y \Vert \le R} \Vert b_2(x,y)\Vert \le M. \end{aligned}$$Further, there exist \(K_4 >0\) and \(R^\prime \ge R\) such that for all \(x \in {\mathbb R}^d\)
$$\begin{aligned} - \langle b_2(x,y) , y \rangle> K_4 \Vert y\Vert ^2, \quad \Vert y \Vert > R^\prime . \end{aligned}$$
Using the above assumptions in the same proof as in Lemma A.2, it is standard to see that the above coupled system has a unique strong solution. With \(f \in C_b^2({\mathbb R}^m)\), for \(y \in {\mathbb R}^m\) let
where x is treated as a parameter. Let \({\mathcal F}_t^X = \sigma (X_s, s \le t)\), \({\mathcal F}^X = \bigvee _{t \ge 0} {\mathcal F}_t^X\). Define \({\mathcal F}_t^{X,Y}\) and \({\mathcal F}^{X,Y}\) analogously. Define \(\pi _t(dy)\) as the conditional law of \(Y_t\) given \({\mathcal F}_t^X\) so that
Proposition B.1
(Nonlinear Filtering Equation) The measure valued process \(\pi \) is the unique solution to the (Fujisaki-Kallianpur-Kunita) nonlinear filtering equation
where \(\tilde{B}\) is a standard Brownian motion.
Remark B.2
\({\tilde{B}}\) is explicitly defined later in the proof. It is called the ‘innovations process’ and, under mild technical conditions, is known to generate the same increasing \(\sigma \)-fields as B [3].
We will closely mimic the arguments in [6], Chapter 3] proved for the case when \(b_1(x,y)\) is a function of the first argument alone. In our setting, \(b_1(x,y)\) is a function of both arguments.
Set
and
Define the probability measure Q by
This consistently defines Q on \({\mathcal F}^{X,Y}\). As \(b_1\) is bounded, by the Cameron-Martin-Girsanov theorem, it follows that \(\bar{B}_\cdot \) is an \(\mathbb {R}^d\)-valued standard Brownian motion under Q. Under Q, the joint process (X, Y) given by (79) - (80) takes the form
Before we begin the proof we need some preliminary lemmas.
Lemma B.3
For \(t > 0\), let Z be a Q-integrable \({\mathcal F}_t^{X,Y}\)-measurable \(\mathbb {R}^d\)-valued random variable. Then
Proof
Set
Then \({\mathcal F}^X = \tilde{\mathcal F}^X_t\vee {\mathcal F}^X_t\), and since \(X_s = \sigma _1 \bar{B}_s\), an \(\{{\mathcal F}_s^X\}\)-Wiener process under Q, \(\tilde{\mathcal F}^X_t\) is independent of \({\mathcal F}_t^X\) under Q. Hence
This completes the proof of the lemma. \(\square \)
Lemma B.4
Let \(\{{\alpha }_t, ~t \ge 0\}\) be an \(\{{\mathcal F}_t^{X,Y}\}\)-progressively measurable \(\mathbb {R}\)-valued process such that
Then
Proof
Using Lemma B.3, it follows that
are \({\mathcal F}^X_t\)-measurable. Hence using the ‘density result’ of Krylov and Rozovskii, see [6], Lemma B.39, p.355], it is enough to show
for all process \(\beta (\cdot )\) of the form
for a deterministic \(r \in L^{\infty }([0, t] ; \mathbb {R}^d)\). Consider
This completes the proof of the lemma.
Lemma B.5
Let \(x \in {\mathbb R}^d\). Let \(\{{\alpha }_t, ~t \ge 0\}\) be \(\{{\mathcal F}_t^{X,Y}\}\)-progressively measurable process such that
where
and \(\langle M^f \rangle _t\) is its quadratic variation. Then
Proof
Via Itô’s formula, we first obtain
Under P, this is driven by a Brownian motion independent of \(B_t\), which leads to
and hence \(Q-\)almost surely. Using this, the proof follows along the lines of the proof of Lemma B.4.
Set
and for \(g \in C^2(\mathbb {R}^d \times \mathbb {R}^m)\) with a little abuse of notation denote
We then have the following.
Lemma B.6
(Kallianpur–Striebel formula) For \(g \in C^2(\mathbb {R}^d \times \mathbb {R}^m)\),
Proof
In view of Lemma B.3, it is enough to show that
Since both left and right sides are \({\mathcal F}_t^X\)-measurable, it is enough to show that
for all \({\mathcal F}_t^X\)-measurable \(\beta \). This is now easily verified since, for such \(\beta \), we have
This completes the proof of the lemma. \(\square \)
We are now ready to prove Proposition B.1. We shall derive first the Zakai equation solved by certain unnormalized conditional laws. Then we shall show existence to the Fujisaki-Kallianpur-Kunita) nonlinear filtering equation (81), followed by uniqueness.
Proof of Proposition B.1
Observe that \(\{\Lambda _t, t \ge 0\}\) is given by the solution of the SDE
for \(t \ge 0\). Hence by a routine application of It\(\hat{\mathrm{o}}\)’s formula it follows that
From this, since \(X_t\) is driven by \(B_t\) and \(Y_t\) is driven by \(W_t\), for \(f \in C_b^2({\mathbb R}^m)\), the cross-variation \(\langle \tilde{\Lambda }, f(Y_{\cdot }) \rangle _t = 0\) P-a.s. and hence Q-a.s. Using It\(\hat{\mathrm{o}}\)’s formula again, we get
and hence, using (85) and (87), we get
Taking conditional expectation \(E^Q[ \ \cdot \ | {\mathcal F}^X]\) in (88) we have using Lemma B.5 we have
For \(g \in C({\mathbb R}^d\times {\mathbb R}^m)\), denoting
in (88) and using Lemma B.6, we arrive at the Zakai equation
For \(\mathbf 1 :=\) the constant function identically equal to 1, we see that \(\rho _t(\mathbf 1 ) = E^Q[ \tilde{\Lambda }_t | {{\mathcal F}^X} ]\), and hence
The nonnegative measure valued process \(\{\rho _t\}_{t \ge 0}\) is called the process of unnormalized conditional laws in view of (91).
Now we are ready to prove the existence theorem for the Fujisaki-Kallianpur-Kunita (FKK) equation, (81). From the Zakai equation (90) we get
In particular, one can deduce that
Using It\(\hat{\mathrm{o}}\)’s formula, we get
Note that the cross-variation
From It\(\hat{\mathrm{o}}\)’s formula for the product of \(\rho _t(f)\) and \(\frac{1}{\rho _t(\mathbf 1 )}\) we get
Substituting (92),(94), and (95) in the above we have
From (91), using some simple algebra in the above we obtain
Let
the so called ‘innovation process’. For \(0 \le s < t \), we have
Thus \(\{{I}_t | t \ge 0\}\) is an \(\{{\mathcal F}_t^X\}\)-martingale with mean 0 and quadratic variation \(\sigma _1^2t\). Thus by Levy’s characterization, I is a scaled Brownian motion. Define
So \(\tilde{B}_t\) is a \(\{{\mathcal F}_t^X\}\)-adapted standard Brownian motion under P. Therefore we have shown that,
with \(\tilde{B}_s\) being a standard Brownian motion. Thus we have shown existence of a solution to the FKK equation. Uniqueness of the FKK equation in the sense of martingale problem follows from Theorem 3.3 of Kurtz and Ocone [29]. Note that while Kurtz and Ocone [29] cite nonlinear filtering as an example of this theorem, they consider the classical formulation (see [29, Theorem 4.1]) which is more restrictive than ours. However the aforementioned theorem ([29, Theorem 3.3]) is general enough to cover our problem. \(\square \)
From (93) we have
Since \(\pi _t(f) = \frac{\rho _t(f)}{\rho _t(\mathbf 1 )},\)
Thus solutions \(\pi , \rho \) of FKK, resp. Zakai equations are in one-one correspondence and uniqueness of one implies that of the other.
Remark B.7
It is interesting to note that some of the earlier uniqueness arguments for the classical framework such as one using multiple Wiener integral expansion due to [31] or via the Clark-Davis ‘pathwise’ filter as in [23], do not work for our case. (The latter would work only if \(b_1(x, \cdot ) = \nabla F(x, \cdot )\) for a suitable F.)
Appendix C: Laplace’s Principle
We now characterize weak limit points of the sequence of invariant measures for the fast process \(\nu ^{\varepsilon _n, x^{\varepsilon _n}}\) when \(\varepsilon _n \rightarrow 0\) and for deterministic \(x^{\varepsilon _n} \rightarrow x\). This is used in the proof of Proposition 2.4(c,d).
Lemma C.1
Let \(n \ge 1, 0< \varepsilon _n < 1, x^{\varepsilon _n} \in {\mathbb R}^d\), and \(x \in \mathbb {R}^d\). Suppose \(\varepsilon _n \rightarrow 0\) and \(x^{\varepsilon _n} \rightarrow x \) as \(n\rightarrow \infty \).
-
(a)
Then the sequence of measures \(\nu ^{\varepsilon _n, x^{\varepsilon _n}}\) is tight and any limit point is supported on \(\arg \min \{U(x, \cdot )\}\).
-
(b)
Assume (U4) and let \(x \in D_L^\circ \) for some \(L \ge 1\). Then \(\nu ^{\varepsilon _n, x^{\varepsilon _n}}\) converges weakly to \(\nu ^{0,x}\), where \(\nu ^{0,x}\) is given by
$$\begin{aligned} \sum _{i=1}^{L} \delta _{y_i(x)}\frac{\left( \text{ Det }\left[ D_y^2 U(x,y_i(x))\right] \right) ^{-\frac{1}{2}}}{\sum _{j=1}^{L} \left( \text{ Det }\left[ D_y^2 U(x,y_j(x))\right] \right) ^{-\frac{1}{2}}}. \end{aligned}$$
Proof
(a) Using (8) and (9) it is easy to see (see Lemma A.1 in Appendix A) that there is \(K_4 >0\) and an \(R^\prime \ge R\) such that
Let \(h: {\mathbb R}^m \rightarrow {\mathbb R}\) be given by
Using (99), there is \(R^\prime >0\) such that
for all \(\parallel y \parallel > R^\prime \) and \(n \ge 1.\) We can assume without loss of generality that \(\max \{\Vert x^{\varepsilon _n} \Vert , \Vert x\Vert \} \le R^\prime \). So by regularity assumption on U from (U1) we have
for some \(K \equiv K(m,s,R^\prime ,U)>0\). Using (100), (101) along with Proposition 2.4 in [33] and its proof, we have for all \(n \ge 1\)
Define \(g: {\mathbb R}^m \rightarrow {\mathbb R}\) by \(g(y) = 2K_4 \parallel y \parallel ^2 - m s(\varepsilon )^{2}\). Using (100) and (102) we have
As \(g(y) \rightarrow \infty \) when \(\parallel y \parallel \rightarrow \infty \) we can conclude that the sequence of measures \(\{\nu ^{\varepsilon _n, x_n}\}_{n \ge 1}\) is tight. We will now show that any limit point \(\nu \) is supported on \(\arg \min U(x, \cdot )\).
Let \(z \in {\mathbb R}^m, z \not \in \arg \min \{U(x, \cdot ) \}.\) As \(U(x^{\varepsilon _n},\cdot )\) converges to \(U(x,\cdot )\) uniformly on compact sets, there exists \(\delta >0 \) and \(r >0\) such that
and
Therefore, for \(n \ge 1\),
Therefore,
Hence any limit point \(\nu \) is supported on the \(\arg \min \{ U(x,\cdot )\}.\)
(b) Let \(x \in D_L^\circ \) for some \(L \ge 1\). Under (U4) the global minima \(y_i(x), 1 \le i \le L\) are nondegenerate, i.e., the matrix \(D^2_yU(x, y_i(x))\) is positive definite for \(1 \le i \le L\). Since \(D_L^\circ \) is open, using \(U \in C^2({\mathbb R}^m \times {\mathbb R}^d)\) in (U1), with a suitable relabelling of the minima if necessary, we have \(y_i(x^{\varepsilon _n}) \rightarrow y_i(x) \ \forall 1 \le i \le L\) as \(n \rightarrow \infty \). Let \(B_i\) be the ball centered at \(y_i(x)\) with radius 1 for each i. Let n be sufficiently large so that \(y_i(x^{\varepsilon _n})\) are in a ball centered at \(y_i(x)\) with radius \(\frac{1}{2}\). Let \(B^n_i\) be ball centered at \(y_i(x^{\varepsilon _n})\) with radius \(\frac{1}{4}\) for each i. Note that \(\nabla _yU(x^{\varepsilon _n}, y_i(x^{\varepsilon _n})) = 0\) and \(U(x^{\varepsilon _n}, y_i(x^{\varepsilon _n})) = \min U(x^{\varepsilon _n}, \cdot ):= u_{\min }\) (say) as it does not depend on i. Using Taylor’s expansion up to second order, we have that for each \( y \in B_i\) there is a \(\tilde{y}_i(x^{\varepsilon _n}) \in B_i\) such that
The above and standard fact about Gaussian random variables implies:
and
where \(A_{i,n} = \left\{ z \in {\mathbb R}^m : \left\| \left( D^2_yU(x^{\varepsilon _n}, y_i(x^{\varepsilon _n})) \right) ^{-1/2} z \right\| \le \frac{1}{2\sqrt{2}} \right\} \) and \(Z^m \) is a standard \(m-\)dimensional Gaussian random variable. Note that \(A_{i,n}\) is a bounded set in \({\mathbb R}^m\) and by (U1), \(U \in C^2({\mathbb R}^d \times {\mathbb R}^m)\). So as \(n \rightarrow \infty \) we have
for all i. Therefore using a standard sandwich argument we can conclude that, for balls \(B_i\) and \(B_j\), \(1 \le i, j \le L\),
From (a) we know that the sequence of measures \(\{\nu ^{\varepsilon _n,x^{\varepsilon _n}}\}_{n \ge 1}\) are tight and all limit points are measures supported on the \(\arg \min U(x, \cdot )\). Consequently by (106) we have that any limit point \(\nu ^{0,x}\) is given by
Since all subsequential limit points are the same we have the result. \(\square \)
Rights and permissions
About this article
Cite this article
Athreya, S.R., Borkar, V.S., Kumar, K.S. et al. Simultaneous Small Noise Limit for Singularly Perturbed Slow-Fast Coupled Diffusions. Appl Math Optim 83, 2327–2374 (2021). https://doi.org/10.1007/s00245-019-09630-w
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00245-019-09630-w
Keywords
- Averaging principle
- Slow–fast motion
- Carathéodory solution
- Filippov solution
- Small noise limit
- Nonlinear filter
- Spectral gap
- Reversible diffusion