betaFeedback

WorksMathematics

Feynman–Kac Representation for Parabolic Equations with Memory: A New Class of BSDEs

Abstract

This paper is concerned with the connection between parabolic equations with memory (MPEs, for short) and backward stochastic differential equations (BSDEs, for short). To provide a probabilistic representation of MPEs, we introduce a new class of BSDEs, called BSDEs with memory (MBSDEs, for short). The main feature of MBSDEs is that, at the current time ss, the generator depends not only on the current value Yt,x(s)≡Ys,Xt,x(s)(s)Y^{t,x}(s)\equiv Y^{s,X^{t,x}(s)}(s) but also on the future path r↦Yr,Xt,x(s)(r)r\mapsto Y^{r,X^{t,x}(s)}(r), r∈[s,T]r\in[s,T]. This structure reveals a fundamental difference from classical BSDEs: in the classical setting, the time and space variables are coupled and inseparable, whereas in MBSDEs they are completely independent. We prove the well-posedness of these MBSDEs and show that their adapted solutions provide a probabilistic representation for the solutions of MPEs. Furthermore, under a monotonicity condition, we establish a comparison principle for one-dimensional MBSDEs, which enables us to develop a viscosity solution theory for MPEs. This generalizes the celebrated nonlinear Feynman–Kac formula of Peng [25] and Pardoux–Peng [22] to the case with memory, in both the classical and viscosity solution senses.

AMS subject classifications. 45K05, 35K10, 60H10, 60H20.

Introduction

The nonlinear Feynman–Kac formula, established in the seminal works of Peng [25] and Pardoux–Peng [22], asserts that the solution to the second-order semilinear parabolic equation

{∂tu(t,x)+12tr⁡[∂xx2u(t,x)σ(t)σ(t)T]+g(t,x,u(t,x),∂xu(t,x)σ(t))=0,(t,x)∈[0,T]×Rn,u(T,x)=h(x),x∈Rn,\begin{cases} \partial_t u(t,x) + \frac{1}{2}\operatorname{tr}\left[\partial_{xx}^{2}u(t,x)\sigma(t)\sigma(t)^{\mathsf T}\right] + g\left(t,x,u(t,x),\partial_xu(t,x)\sigma(t)\right)=0, & (t,x)\in[0,T]\times\mathbb{R}^{n},\\ u(T,x)=h(x), & x\in\mathbb{R}^{n}, \end{cases}

admits the following probabilistic representation:

u(t,x)=Yt,x(t),(t,x)∈[0,T]×Rn,u(t,x)=Y^{t,x}(t), \qquad(t,x)\in[0,T]\times\mathbb{R}^{n},

where Yt,x(⋅)Y^{t,x}(\cdot), together with some adapted process Zt,x(⋅)Z^{t,x}(\cdot), is the unique adapted solution to the following Markovian backward stochastic differential equation (BSDE, for short):

dXt,x(s)=σ(s)dW(s),s∈[t,T],dYt,x(s)=−g(s,Xt,x(s),Yt,x(s),Zt,x(s))ds+Zt,x(s)dW(s),s∈[t,T],Xt,x(t)=x,Yt,x(T)=h(Xt,x(T)).(1)\begin{aligned} dX^{t,x}(s) &= \sigma(s)dW(s), \qquad s \in[t,T],\\ dY^{t,x}(s) &= -g\left(s,X^{t,x}(s),Y^{t,x}(s),Z^{t,x}(s)\right)ds+Z^{t,x}(s)dW(s), \qquad s \in[t,T],\\ X^{t,x}(t) &= x,\qquad Y^{t,x}(T)=h\left(X^{t,x}(T)\right). \tag*{(1)} \end{aligned}

This connection extends the Feynman–Kac representation from linear parabolic PDEs to semilinear PDEs, thereby enabling a probabilistic treatment of such PDEs. Moreover, it paves the way for the use of Monte Carlo methods in their numerical solution; see e.g. Ma–Yong [20] and Delarue–Menozzi [12]. Since then, a substantial body of work has been devoted to studying the connections between forward–backward stochastic differential equations (FBSDEs, for short) and parabolic PDEs; see e.g. Ma–Protter–Yong [19], Pardoux–Tang [24], and Delarue [11] for the quasi-linear case, and Cheridito–Soner–Touzi–Victoir [8], Peng [26, 27], and Touzi [28] for the fully nonlinear case. Further developments on the connection between FBSDEs and stochastic PDE, path-dependent PDE, or mean-field PDEs can be found in Pardoux–Peng [23], Hu–Ma–Yong [18], Ekren–Christian–Touzi–Zhang [13], and Buckdahn–Li–Peng–Rainer [5], among others. Recently, this connection has played a crucial role in the study of the master equation in mean-field games and of deep-learning methods for high-dimensional PDEs; see Cardaliaguet–Delarue–Lasry–Lions [6] and Han–Jentzen–E [17], respectively.

The parabolic equation with memory (MPE), a variant of the classical parabolic PDE, serves to model memory effects in dynamical systems. Typical examples include the equations arising in the electromagnetic theory of dielectrics and in heat conduction for materials with memory, which were investigated by Davis [9, 10] in the 1970s. For a detailed account of the background and motivations of MPEs, we refer to the book by Amendola–Fabrizio–Golden [3] and the references therein. The presence of the memory term makes the MPE fundamentally different from its classical counterpart. A representative example is that the flow generated by the heat equation with memory exhibits a hybrid parabolic–hyperbolic character; see Gurtin–Pipkin [16] and Yong–Zhang [33, 34]. Due to this distinctive feature, the MPE has drawn considerable interest in recent decades. For instance, Alabau-Boussouira–Cannarsa–Sforza [2] obtained a decay estimate; Fabrizio–Giorgi–Pata [14] proposed a new approach based on several novel notions; Andrade and Viana [4] established well-posedness and regularity within an abstract framework; Wang–Zhang–Zuazua [29] carried out a detailed analysis of the flow generated by MPEs, to name but a few. The associated control problems have also been extensively studied; see e.g. Aassila–Cavallcanti–Soriani [1], Fu–Yong–Zhang [15], Chaves–Silva–Zhang–Zuazua [7], and Wang–Zhang–Zuazua [30].

Then a natural question arises concerning the probabilistic interpretation of MPEs, namely the corresponding Feynman–Kac formula. This question, however, remains unexplored, and the present paper aims to address this important issue. Several new and interesting phenomena emerge in the process. To be specific, we study the probabilistic representation of the following mm-dimensional MPE:

∂tu(t,x)+12tr⁡[∂xx2u(t,x)σ(t)σ(t)T]+g(t,x,u(t,x),∂xu(t,x)σ(t),M[u](t,x))=0,(t,x)∈[0,T]×Rn,u(T,x)=h(x),x∈Rn,(2)\begin{aligned} \partial_tu(t,x)+\frac{1}{2}\operatorname{tr}\left[\partial_{xx}^{2}u(t,x)\sigma(t)\sigma(t)^{\mathsf{T}}\right] &+g\left(t,x,u(t,x),\partial_xu(t,x)\sigma(t),\mathcal{M}[u](t,x)\right)=0, \qquad(t,x)\in[0,T]\times\mathbb{R}^{n},\\ u(T,x)&=h(x), \qquad x\in\mathbb{R}^{n}, \tag*{(2)} \end{aligned}

where the memory operator M[⋅]\mathcal{M}[\cdot] is defined by

M[u](t,x)=∫tTM(s−t)u(s,x) ds,(t,x)∈[0,T]×Rn,\mathcal{M}[u](t,x)=\int_t^T M(s-t)u(s,x)\,ds,\qquad(t,x)\in[0,T]\times\mathbb{R}^{n},

with M(⋅):[0,T]→Rk×mM(\cdot):[0,T]\to\mathbb{R}^{k\times m} being a Lipschitz continuous function. Observe that, for any fixed point (t,x)(t,x), the quantity M[u](t,x)\mathcal{M}[u](t,x) depends on the values u(s,x)u(s,x) for s∈[t,T]s\in[t,T]. Hence, M[⋅]\mathcal{M}[\cdot] is a non-local operator in the time variable.

To address this non-local term, we introduce a new class of BSDEs:

{dXt,x(s)=σ(s)dW(s),s∈[t,T],dYt,x(s)=−g(s,Xt,x(s),Yt,x(s),Zt,x(s),Mt,x[Y](s))ds+Zt,x(s)dW(s),s∈[t,T],Xt,x(t)=x,Yt,x(T)=h(Xt,x(T)),(3)\left\{ \begin{aligned} dX^{t,x}(s) &= \sigma(s)dW(s), && s \in[t,T],\\ dY^{t,x}(s) &= -g\left(s,X^{t,x}(s),Y^{t,x}(s),Z^{t,x}(s),\mathcal{M}^{t,x}[Y](s)\right)ds+Z^{t,x}(s)dW(s), && s \in[t,T],\\ X^{t,x}(t) &= x,\qquad Y^{t,x}(T)=h\left(X^{t,x}(T)\right), \end{aligned} \right. \tag*{(3)}

where (t,x)∈[0,T]×Rn(t,x) \in[0,T] \times\mathbb{R}^{n} is the initial pair, and for any fixed s∈[t,T]s \in[t,T],

Mt,x[Y](s)=∫sTM(r−s)Yr,Xt,x(s)(r)dr,s∈[t,T].(4)\mathcal{M}^{t,x}[Y](s)=\int_s^T M(r-s)Y^{r,X^{t,x}(s)}(r)dr,\qquad s \in[t,T]. \tag*{(4)}

The key novelty of (3) is the inclusion of the non-local term Mt,x[Y](⋅)\mathcal{M}^{t,x}[Y](\cdot) in the generator gg which is designed to capture the memory term M[u](⋅,⋅)\mathcal{M}[u](\cdot,\cdot) of the MPE (2). The superscript of Yt,x(⋅)Y^{t,x}(\cdot) signifies that the initial pair of the associated diffusion process is (t,x)(t,x). In the integral term (4), the variable rr plays a dual role: it is both the current time and the initial time of Yr,Xt,x(s)(r)Y^{r,X^{t,x}(s)}(r). One may therefore expect that Yr,Xt,x(s)(r)Y^{r,X^{t,x}(s)}(r) is Fs\mathcal{F}_s-measurable for any r∈[s,T]r \in[s,T], which ensures that the memory term (4) remains progressively measurable. To distinguish this equation from classical BSDEs, we refer to it as a BSDE with memory (MBSDE, for short). The identification of this appropriate form constitutes the first contribution of our paper.

Recalling (1) and invoking the flow property of BSDEs, the solution to (1) satisfies

Yt,x(s)=Ys,Xt,x(s)(s)=Ys,x~(s)∣x~=Xt,x(s).Y^{t,x}(s)=Y^{s,X^{t,x}(s)}(s)=Y^{s,\widetilde{x}}(s)\big|_{\widetilde{x}=X^{t,x}(s)}.

Consequently, in BSDE (1), the space variable is always evaluated at the current value of the diffusion process Xt,x(⋅)X^{t,x}(\cdot) at time ss, so the time and space variables are coupled and inseparable. The situation is entirely different for MBSDE (3): there we need to determine Yr,Xt,x(s)(r)=Yr,x~(r)∣x~=Xt,x(s)Y^{r,X^{t,x}(s)}(r)=Y^{r,\widetilde{x}}(r)\big|_{\widetilde{x}=X^{t,x}(s)} for s≠rs\ne r. Hence, in MBSDEs the time and space variables are completely independent, marking an essential difference from the classical BSDE framework. This motivates the more natural notation (Y⋅,⋅(⋅),Z⋅,⋅(⋅))(Y^{\cdot,\cdot}(\cdot),Z^{\cdot,\cdot}(\cdot)), rather than (Yt,x(⋅),Zt,x(⋅))(Y^{t,x}(\cdot),Z^{t,x}(\cdot)), for the solution to (3).

We now turn to the well-posedness of the MBSDE (3). The main obstacle is the non-local term Mt,x[Y](⋅)\mathcal{M}^{t,x}[Y](\cdot). We use an approximation argument to deal with it and thereby construct a solution to (3). Specifically, we partition [0,T][0,T] into NN subintervals [ti,ti+1][t_i,t_{i+1}] with ti=iNt_i=\frac{i}{N}, and on each subinterval we build a suitable standard BSDE to approximate the MBSDE. A key tool in this construction is the Feynman–Kac formula for backward stochastic Volterra integral equations (BSVIEs, for short) established in Wang–Yong–Zhang [32]. Gluing these local approximations together over [0,T][0,T] yields an approximate solution to (3) that depends on NN. Through a series of suitable estimates, we prove that this sequence of approximate solutions is Cauchy in an appropriate norm, and its limit is exactly a solution to (3). This approach not only establishes the existence of a solution to (3) but also reveals its connections with and differences from classical BSDEs. Uniqueness then follows from a standard argument.

Given the unique solution (Y⋅,⋅(⋅),Z⋅,⋅(⋅))(Y^{\cdot,\cdot}(\cdot),Z^{\cdot,\cdot}(\cdot)) to (3), the following function is well-defined:

u‾(t,x)=Yt,x(t),(t,x)∈[0,T]×Rn.\overline{u}(t,x)=Y^{t,x}(t),\qquad(t,x)\in[0,T]\times\mathbb{R}^{n}.

When the coefficients σ(⋅)\sigma(\cdot), g(⋅)g(\cdot), and h(⋅)h(\cdot) are sufficiently smooth, we can show that u‾(⋅,⋅)\overline{u}(\cdot,\cdot) belongs to C1,2([0,T]×Rn;Rm)C^{1,2}([0,T]\times\mathbb{R}^{n};\mathbb{R}^{m}), and hence is a classical solution to (3). This establishes a rigorous connection between the MPE (2) and the MBSDE (3). We emphasize that, unlike analytic methods, our approach does not require the leading coefficient σ(⋅)\sigma(\cdot) to be nondegenerate. To handle the non-local term Mt,x[Y](⋅)\mathcal{M}^{t,x}[Y](\cdot), we need to establish the uniform estimate sup⁡t,s,x∣Yt,x(s)∣\sup_{t,s,x}|Y^{t,x}(s)|.

When the coefficients are less regular, we adopt the viscosity solution framework under the assumption that Y⋅,⋅(⋅)Y^{\cdot,\cdot}(\cdot) is scalar. Since viscosity solutions are local in nature whereas the MPE (2) contains a non-local term, its comparison principle becomes considerably more delicate. Indeed, the comparison principle may fail in general, even for classical solutions. Inspired by Wang–Yong [31] and Wang–Yong–Zhang [32], we impose a monotonicity condition, under which the comparison principle holds naturally for classical solutions. The comparison principle for viscosity solutions is then established by a modification method.

The rest of the paper is organized as follows. § 2 collects some preliminary results and introduces a few elementary notions. § 3 is devoted to identifying the form of MBSDEs. In § 4, we establish the well-posedness and the comparison theorem for MBSDEs. The Feynman–Kac formula is proved in § 5 for classical solutions and in Section 6 for viscosity solutions.

Preliminaries

Let (Ω,F,P)(\Omega,\mathcal{F},\mathbb{P}) be a complete probability space, on which a standard 1-dimensional Brownian motion W(⋅)W(\cdot) is defined. The augmented natural filtration of W(⋅)W(\cdot) is denoted by F={Ft}t≥0\mathbb{F}=\{\mathcal{F}_t\}_{t\geq0}. Let T>0T>0 be a fixed time horizon and t∈[0,T)t\in[0,T). We denote

References

  1. [1]M. Aassila, M. M. Cavalcanti, and J. A. Soriano, Asymptotic stability and energy decay rates for solutions of the wave equation with memory in a star-shaped domain, SIAM J. Control Optim., 38 (2000), 1581–1602.DOI
  2. [2]F. Alabau-Boussouira, P. Cannarsa, and D. Sforza, Decay estimates for second order evolution equations with memory, J. Funct. Anal., 254 (2008), 1342–1372.
  3. [3]G. Amendola, M. Fabrizio, and J. M. Golden, Thermodynamics of Materials With Memory: Theory and Applications, Second Edition, Springer, New York, 2021.
  4. [4]B. Andrade and A. Viana, Abstract Volterra integro-differential equations with applications to parabolic models with memory, Math. Ann., 369 (2017), 1131–1175.
  5. [5]R. Buckdahn, J. Li, S. Peng, and C. Rainer, Mean-field stochastic differential equations and associated PDEs, Ann. Probab., 45 (2017), 824–878.arxiv.org/abs/1407.1215
  6. [6]P. Cardaliaguet, F. Delarue, J. M. Lasry, and P. L. Lions, The master equation and the convergence problem in mean field games, Ann. of Math. Stud., 201, Princeton University Press, 2019.
  7. [7]F. Chaves-Silva, X. Zhang, and E. Zuazua, Controllability of evolution equations with memory, SIAM J. Control Optim. 55 (2017), 2437–2459.
  8. [8]P. Cheridito, H. M. Soner, N. Touzi, and N. Victor, Second-order backward stochastic differential equations and fully nonlinear parabolic PDEs, Comm. Pure Appl. Math., 60 (2007), 1081–1110.DOI
  9. [9]P. L. Davis, On the hyperbolicity of the equations of the linear theory of heat conduction for materials with memory, SIAM J. Appl. Math., 30 (1976), 75–80.DOI
  10. [10]P. L. Davis, On the speed of propagation of solutions of integro-differential equations, SIAM J. Math. Anal., 10 (1979), 570–576.
  11. [11]F. Delarue, On the existence and uniqueness of solutions to FBSDEs in a non-degenerate case, Stochastic Process. Appl., 99 (2002), 209–286.DOI
  12. [12]F. Delarue and S. Menozzi, A forward-backward stochastic algorithm for quasi-linear PDEs, Ann. Appl. Probab., 16 (2006), 140–184.
  13. [13]I. Ekren, C. Keller, N. Touzi, and J. Zhang, On viscosity solutions of path dependent PDEs, Ann. Probab., 42 (2014), 204–236.DOI
  14. [14]M. Fabrizio, C. Giorgi, and V. Pata, A new approach to equations with memory, Arch. Ration. Mech. Anal., 198 (2010), 189–232.
  15. [15]X. Fu, J. Yong, and X. Zhang, Controllability and observability of a heat equation with hyperbolic memory kernel, J. Differential Equations 247 (2009), 2395–2439.DOI
  16. [16]M. E. Gurtin and A. C. Pipkin, A general theory of heat conduction with finite wave speeds, Arch. Ration. Mech. Anal., 31 (1968), 113–126.DOI
  17. [17]J. Han, A. Jentzen, and W. E Solving high-dimensional partial differential equations using deep learning, Proc. Natl. Acad. Sci. USA, 115 (2018), 8505–8510.DOI
  18. [18]Y. Hu, J. Ma, and J. Yong, On semi-linear degenerate backward stochastic partial differential equations, Probab. Theory Related Fields, 123 (2002), 381–411.DOI
  19. [19]J. Ma, P. Protter, and J. Yong, Solving forward-backward stochastic differential equations explicitly — a four step scheme, Probab. Theory Related Fields, 98 (1994), 339–359.DOI
  20. [20]J. Ma and J. Yong, Forward-backward stochastic differential equations and their applications, Lecture Notes in Mathematics, Springer, Berlin, 1999.
  21. [21]J. Ma and J. Zhang, Representation theorems for backward stochastic differential equations, Ann. Appl. Probab., 12 (2002), 1390–1418.DOI
  22. [22]E. Pardoux and S. Peng, Backward stochastic differential equations and quasilinear parabolic partial differential equations, Stochastic partial differential equations and their applications, Springer, Berlin, Heidelberg, (1992), 200–217.
  23. [23]E. Pardoux and S. Peng, Backward doubly stochastic differential equations and systems of quasilinear SPDEs, Probab. Theory Related Fields, 98 (1994), 209–227.DOI
  24. [24]E. Pardoux and S. Tang, Forward-backward stochastic differential equations and quasilinear parabolic PDEs, Probab. Theory Related Fields, 114 (1999), 123–150.DOI
  25. [25]S. Peng, Probabilistic interpretation for systems of quasilinear parabolic partial differential equations, Stochastics Stochastics Rep., 37 (1991), 61–74.DOI
  26. [26]S. Peng, Backward stochastic differential equation, nonlinear expectation and their applications, Proceedings of ICM, 2010, Volume I, 393–432.
  27. [27]S. Peng, Nonlinear expectations and stochastic calculus under uncertainty, Probab. Theory Stoch. Model., Springer, Berlin, 2019.
  28. [28]N. Touzi, Second order backward SDEs, fully nonlinear PDEs, and applications in finance, Proceedings of ICM, 2010, Volume IV, 3132–3150.DOI
  29. [29]G. Wang, Y. Zhang, and E. Zuazua, Flow decomposition for heat equations with memory, J. Math. Pures Appl. 158 (2022), 183–215.DOI
  30. [30]G. Wang, Y. Zhang, and E. Zuazua, Observability for heat equations with time-dependent analytic memory, Arch. Ration. Mech. Anal., 248 (2024), 115.DOI
  31. [31]T. Wang and J. Yong, Comparison theorems for some backward stochastic Volterra integral equations, Stochastic Process. Appl., 125 (2015), 1756–1798.DOI
  32. [32]H. Wang, J. Yong, and J. Zhang, Path dependent Feynman–Kac formula for forward backward stochastic Volterra integral equations, Ann. Inst. Henri Poincaré Probab. Stat., 58 (2022), 603–638.arxiv.org/abs/2004.05825
  33. [33]J. Yong and X. Zhang, Heat equations with memory, Nonlinear Anal., 63 (2005), 99–108.DOI
  34. [34]J. Yong and X. Zhang, Heat equation with memory in anisotropic and non-homogeneous media, Acta Math. Sin. (Engl. Ser.), 27 (2011), 219–254.DOI

Paper details

vlnc:W_bgq6pgrv
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents