betaFeedback

WorksMathematics

Stochastic Maximum Principle for McKean-Vlasov Control with Discrete Path Dependence

Abstract

We study a class of McKean-Vlasov control problems with discrete path dependence. The coefficients and cost functional may depend on finitely many past values of the controlled state, observed at fixed deterministic times, and on their joint law. We establish well-posedness of the controlled state equation and derive necessary and sufficient optimality conditions through a stochastic Pontryagin maximum principle. The adjoint process is characterized by a backward stochastic differential equation with jumps at fixed observation times. Each jump represents the conditional sensitivity of future costs with respect to the corresponding observed state and its distribution. In the linear-quadratic case, we prove global solvability of the resulting forward-backward system by combining a continuation argument with a mean-field Riccati reduction. We finally discuss an application to time-series generation, where the terminal cost is given by a kernel-based discrepancy between the law of the sampled path and a target path distribution.

Introduction

McKean–Vlasov stochastic control has become a fundamental framework for optimization problems involving interactions through probability distributions. In its classical formulation, the dynamics and the cost functional depend on the current state of the controlled process and on its marginal distribution. Such problems arise naturally in mean-field control, economics, engineering, and more recently in machine learning and generative modeling. Over the last decade, a rich theory has been developed for these models, including dynamic programming principles, master equations, and stochastic Pontryagin maximum principles, see among others [5, 12].

Early versions of the stochastic maximum principle for mean-field dynamics go back to [3, 8] for linear dependence on the law of the process and for general non-linear dependence in [10, 1]. In these works, the coefficients of the state equation and of the cost depend on the current state and control and on its marginal law. However, in many situations, the relevant information is not contained in the current marginal distribution alone, but in the joint distribution of observations collected at several dates. Given deterministic observation dates 0<t1<⋯<tN=T0 < t_1 < \cdots< t_N = T, a typical objective functional considered in this paper takes the form

J(α)=E[∫0Tf(t,Xt,L(Xt),αt) dt]+g(L(Xt1,…,XtN)),(1)J(\alpha) = \mathbb{E}\left[\int_0^T f(t,X_t,\mathcal{L}(X_t),\alpha_t)\,dt\right] + g\left(\mathcal{L}(X_{t_1},\ldots,X_{t_N})\right), \tag*{(1)}

where XX is the controlled state process, L(Xt1,…,XtN)\mathcal{L}(X_{t_1},\ldots,X_{t_N}) its joint law at the observation dates, and α\alpha is an admissible control.

Unlike classical McKean–Vlasov control problems, the objective functional (1) depends on the joint law of observations collected along the trajectory. Such dependence naturally arises whenever temporal interactions play a significant role. Indeed, properties such as autocorrelation, lead–lag effects or persistence phenomena are encoded in the joint distribution of observations rather than in their individual marginals. Consequently, controlling only the marginal laws L(Xt)\mathcal{L}(X_t) is generally insufficient for reproducing the temporal structure of the underlying process.

The purpose of this paper is to develop a stochastic maximum principle for McKean–Vlasov control problems with such discrete path dependence. More precisely, we consider controlled diffusions whose coefficients and objective functional may depend on the current state, its current distribution, the vector of previously observed values (Xt1,…,Xtk)(X_{t_1},\ldots,X_{t_k}) for tk≤t<tk+1t_k \leq t < t_{k+1}, and the joint distribution of these observations. The dependence is assumed to be non-anticipative (see Definition 2.1), so that only observations already revealed at time tt may influence the dynamics.

The framework considered here occupies an intermediate position between classical and fully path-dependent McKean–Vlasov control. In classical McKean–Vlasov control, the coefficients depend only on (Xt,L(Xt))(X_t,\mathcal{L}(X_t)). At the other end of the spectrum, recent works have investigated control problems depending on the entire past trajectory X[0,t]X_{[0,t]} and on its law. Such formulations naturally lead to infinite-dimensional state spaces and require sophisticated analytical tools including functional Itô calculus and path-dependent partial differential equations; see [15], [13], and [16]. By contrast, the present paper focuses on dependence through a finite collection of observation dates. This discrete path dependence is rich enough to capture temporal interactions while preserving a finite-dimensional structure.

Recently, path-dependent McKean–Vlasov control problems have been studied in [14] through dynamic programming methods and in [9] through a stochastic maximum principle. The finite-observation structure considered here leads to a different adjoint characterization. A precise comparison with a formulation by a family of NN coupled continuous adjoint processes is given in Remark 3.5.

Our main contributions. We first establish existence and uniqueness of the controlled McKean–Vlasov dynamics under standard Lipschitz and linear-growth assumptions. The main novelty of the present work is the appearance of jumps in the adjoint equation. In the classical McKean–Vlasov setting, the adjoint process is continuous and satisfies a backward stochastic differential equation. In contrast, the adjoint process (Y,Z)(Y,Z) associated with (1) is càdlàg and exhibits jumps at the observation dates t1,…,tN−1t_1,\ldots,t_{N-1}. These jumps reflect the fact that additional path information becomes available whenever a new observation date is reached. More precisely, the jump size at each observation date (tk)1≤k≤N−1(t_k)_{1\leq k\leq N-1} is given by

−ΔYtk=E[∫tkT(∂xkH(t,Θt)+E~[∂x~kδδmH(t,Θ~t)(X)]) dt+∂xkg(X,L(X))+E~[∂x~kδδmg~(X,L(X))(X) | Ftk]].(2)-\Delta Y_{t_k}=\mathbb{E}\left[\int_{t_k}^{T}\left(\partial_{x_k}H(t,\Theta_t)+\widetilde{\mathbb{E}}\left[\partial_{\widetilde{x}_k}\frac{\delta}{\delta m}H(t,\widetilde{\Theta}_t)(X)\right]\right)\,\mathrm{d}t +\partial_{x_k}g(X,\mathcal{L}(X))+\widetilde{\mathbb{E}}\left[\partial_{\widetilde{x}_k}\frac{\delta}{\delta m}\widetilde{g}(X,\mathcal{L}(X))(X)\,\middle|\,\mathcal{F}_{t_k}\right]\right]. \tag*{(2)}

where HH is the Hamiltonian function and where we set Θt=((Xt1,…,Xtl),L(Xt1,…,Xtl),Xt,L(Xt),Yt,Zt,αt)\Theta_t=((X_{t_1},\ldots,X_{t_l}),\mathcal{L}(X_{t_1},\ldots,X_{t_l}),X_t,\mathcal{L}(X_t),Y_t,Z_t,\alpha_t) for tl≤t<tl+1t_l\leq t<t_{l+1} for any 1≤k≤l≤N−11\leq k\leq l\leq N-1, Θ~\widetilde{\Theta} an independent copy and X=(Xt1,…,XtN)\mathbf{X}=(X_{t_1},\ldots,X_{t_N}). We derive the Pontryagin maximum principle providing necessary stochastic maximum conditions for optimality. We characterize the adjoint process as the solution of a backward stochastic differential equation with jump at deterministic times given by (2). Under suitable convexity assumptions on the Hamiltonian and the terminal cost, we further obtain sufficient conditions for optimality. Our second main result concerns solvability of the resulting forward–backward system. In the linear-quadratic case, we prove global solvability by a continuation method, in the spirit of [19, 10]. We also show that the system can be reduced to a mean-field Riccati system, related to the Riccati methods used in linear-quadratic McKean–Vlasov control [23, 4, 22]. This gives a novel tractable description of the optimal feedback control in the linear-quadratic setting. Indeed, the Riccati equations together with the jump relations characterize the affine structure of the adjoint process, while the stability of the conditional expectation of the state process allows the coefficients arising in the jump terms to be identified. These ingredients lead to a recursive backward characterization of the system, in which all the coefficients characterizing the adjoint process can be successively identified.

One motivation for the present framework comes from recent stochastic approaches to generative modeling, see [18, 7, 6]. In applications involving time-series data, the target object is naturally a probability measure on path space, or more precisely the joint law of observations collected at several dates. Matching only marginal distributions generally fails to reproduce the temporal dependence structure of the data. The control framework developed in this paper provides a probabilistic formulation for such problems and yields an optimality system that explicitly incorporates path-distribution dependence. To establish the well-posedness of this optimality system, we consider the Maximum Mean Discrepancy (MMD) distance on the space of probability measures (see [21]), and, under suitable regularity assumptions on the kernel KK, prove local-in-time existence and uniqueness of the resulting forward-backward stochastic differential equation.

Outline of the paper. The remainder of the paper is organized as follows. Section 2 introduces the control problem and establishes well-posedness of the state dynamics. Section 3 derives the variational equations and the adjoint backward stochastic differential equation. Section 4 proves the stochastic maximum principle and presents sufficient conditions for optimality. Section 5 studies solvability of the forward–backward system, first in the linear-quadratic case and then in the application to time-series generation. Section 6 concludes, and additional proofs are gathered in the appendix.

Problem formulation

Notations

Throughout the paper, we fix a finite time horizon T>0T>0 and a complete probability space (Ω,F,P)(\Omega,\mathcal{F},\mathbb{P}) supporting an nn-dimensional Brownian motion W=(Wt)0≤t≤TW=(W_t)_{0\leq t\leq T}. Let F=(Ft)0≤t≤T\mathbb{F}=(\mathcal{F}_t)_{0\leq t\leq T} be the natural filtration generated by WW augmented with the P\mathbb{P}-null sets and with the σ\sigma-algebra generated by a rich enough random variable independent of WW used for randomization of the initial condition of the state process. The expectation under P\mathbb{P} is denoted by E\mathbb{E}. For a given random variable χ\chi over (Ω,F,P)(\Omega,\mathcal{F},\mathbb{P}), we denote by Pχ\mathbb{P}_{\chi} its law under P\mathbb{P}.

We denote by C([0,T];Rd)C([0,T];\mathbb{R}^{d}) the space of continuous functions from [0,T][0,T] into Rd\mathbb{R}^{d}, and

Sp([0,T];Rd):={Y=(Yt)0≤t≤T:Y is F-prog. measurable, caˋdlaˋg and E[sup⁡0≤t≤T∣Yt∣p]<∞},p≥1,S∞([0,T];Rd):={Y=(Yt)0≤t≤T:Y is F-prog. measurable, caˋdlaˋg and ∥sup⁡0≤t≤T∣Yt∣∥L∞<∞},H2([0,T];Rd×n):={Z=(Zt)0≤t≤T:Z is F-prog. measurable and E[∫0T∣Zt∣2 dt]<∞},BMO⁡([0,T];Rd×n):={Z=(Zt)0≤t≤T:Z is F-prog. measurable and sup⁡τ∥E[∫τT∣Zt∣2 dt∣Fτ]∥L∞<∞}.\begin{aligned} \mathcal{S}^{p}([0,T];\mathbb{R}^{d}) &:= \left\{Y=(Y_t)_{0\leq t\leq T}:Y\text{ is }\mathbb{F}\text{-prog. measurable, càdlàg and }\mathbb{E}\left[\sup_{0\leq t\leq T}|Y_t|^p\right]<\infty\right\},\quad p\geq1,\\ \mathcal{S}^{\infty}([0,T];\mathbb{R}^{d}) &:= \left\{Y=(Y_t)_{0\leq t\leq T}:Y\text{ is }\mathbb{F}\text{-prog. measurable, càdlàg and }\left\|\sup_{0\leq t\leq T}|Y_t|\right\|_{L^\infty}<\infty\right\},\\ \mathcal{H}^{2}([0,T];\mathbb{R}^{d\times n}) &:= \left\{Z=(Z_t)_{0\leq t\leq T}:Z\text{ is }\mathbb{F}\text{-prog. measurable and }\mathbb{E}\left[\int_0^T|Z_t|^2\,dt\right]<\infty\right\},\\ \operatorname{BMO}([0,T];\mathbb{R}^{d\times n}) &:= \left\{Z=(Z_t)_{0\leq t\leq T}:Z\text{ is }\mathbb{F}\text{-prog. measurable and }\sup_{\tau}\left\|\mathbb{E}\left[\left.\int_{\tau}^{T}|Z_t|^2\,dt\right|\mathcal{F}_{\tau}\right]\right\|_{L^\infty}<\infty\right\}. \end{aligned}

where the supremum in the definition of BMO is taken over all F\mathbb{F}-stopping times τ\tau valued in [0,T][0,T] and where ∥X∥L∞:=inf⁡{C≥0:∣X∣≤C, P-a.s.}\|X\|_{L^\infty}:=\inf\{C\geq0:|X|\leq C,\ \mathbb{P}\text{-a.s.}\} for any random variable χ\chi over (Ω,F,P)(\Omega,\mathcal{F},\mathbb{P}).

References

  1. [1]B. Acciaio, J. Backhoff-Veraguas, and R. Carmona, Extended mean field control problems: stochastic maximum principle and transport perspective, SIAM journal on Control and Optimization, 57 (2019), pp. 3666–3693.arxiv.org/abs/1802.05754
  2. [2]A. Alouadi, G. Loeper, C. Marsala, O. Mazhar, and H. Pham, SBBTS: A unified Schrödinger-Bass framework for synthetic financial time series, arXiv preprint arXiv:2604.07159, (2026).arxiv.org/abs/2604.07159
  3. [3]D. Andersson and B. Djehiche, A maximum principle for SDEs of mean-field type, Applied Mathematics & Optimization, 63 (2011), pp. 341–356.DOI
  4. [4]M. Basel and H. Pham, A weak martingale approach to linear-quadratic McKean–Vlasov stochastic control problems, Journal of Optimization Theory and Applications, 181 (2019), pp. 347–382.
  5. [5]A. Bensoussan, J. Frehse, and P. Yam, Mean Field Games and Mean Field Type Control Theory, SpringerBriefs in Mathematics, Springer, 2013.DOI
  6. [6]S. E. Boustany, T. Basseras, S. Mekkaoui, A. Alouadi, Y. Hafsi, and H. Pham, Deep-MKV-TS: Path-dependent McKean–Vlasov control for financial time series generation, arXiv preprint arXiv:2608.19394, (2026).
  7. [7]S. E. Boustany, S. Mekkaoui, Y. Hafsi, A. Alouadi, and H. Pham, Learning generative dynamics with soft law constraints: A McKean–Vlasov FBSDE approach, arXiv preprint arXiv:2605.08928, (2026).
  8. [8]R. Buckdahn, B. Djehiche, and J. Li, A general stochastic maximum principle for SDEs of mean-field type, Applied Mathematics & Optimization, 64 (2011), pp. 197–216.DOI
  9. [9]R. Buckdahn, J. Li, J. Li, and C. Xing, Path-dependent controlled mean-field coupled forward-backward SDEs: The associated stochastic maximum principle, SIAM Journal on Control and Optimization, 63 (2025), pp. 2124–2153.DOI
  10. [10]R. Carmona and F. Delarue, Forward-backward stochastic differential equations and controlled McKean Vlasov dynamics, The Annals of Probability, 43 (2015), pp. 2647–2700.
  11. [11]R. Carmona and F. Delarue, Probabilistic Theory of Mean Field Games with Applications I: Mean Field FBSDEs, Control, and Games, vol. 83 of Probability Theory and Stochastic Modelling, Springer, 2018.
  12. [12]R. Carmona and F. Delarue, Probabilistic Theory of Mean Field Games with Applications II: Mean Field Games with Common Noise and Master Equations, vol. 84 of Probability Theory and Stochastic Modelling, Springer, 2018.DOI
  13. [13]R. Cont and D.-A. Fournié, Functional Itô calculus and stochastic integral representation of martingales, The Annals of Probability, 41 (2013), pp. 109–133.DOI
  14. [14]A. Cosso, F. Gozzi, I. Kharroubi, H. Pham, and M. Rosestolato, Optimal control of path-dependent McKean–Vlasov SDES in infinite-dimension, The Annals of Applied Probability, 33 (2023), pp. 2863 – 2918.arxiv.org/abs/2012.14772
  15. [15]B. Dupire, Functional Itô calculus. Portfolio Research Paper, 2009.
  16. [16]I. Ekren, C. Keller, N. Touzi, and J. Zhang, On viscosity solutions of path dependent PDEs, The Annals of Probability, 42 (2014), pp. 204–236.DOI
  17. [17]M. Hamdouche, P. Henry-Labordere, and H. Pham, Nonparametric generative modeling for time series via Schrödinger bridge, Journal of Machine Learning Research (JMLR), 27(58) (2026), pp. 1–23.
  18. [18]J. Ma, Y. Tan, and R. Xu, Schrödinger bridge for generative AI: Soft-constrained formulation and convergence analysis, arXiv preprint arXiv:2510.11829, (2025).arxiv.org/abs/2510.11829
  19. [19]S. Peng and Z. Wu, Fully coupled forward-backward stochastic differential equations and applications to optimal control, SIAM Journal on Control and Optimization, 37 (1999), pp. 825–843.DOI
  20. [20]T. C. Sideris, Ordinary Differential Equations and Dynamical Systems, vol. 2 of Atlantis Studies in Differential Equations, Atlantis Press, 2013.DOI
  21. [21]B. K. Sriperumbudur, A. Gretton, K. Fukumizu, B. Schölkopf, and G. R. Lanckriet, Hilbert space embeddings and metrics on probability measures, The Journal of Machine Learning Research, 11 (2010), pp. 1517–1561.arxiv.org/abs/0907.5309
  22. [22]J. Sun and J. Yong, Stochastic Linear-Quadratic Optimal Control Theory: Open-Loop and Closed-Loop Solutions, SpringerBriefs in Mathematics, Springer, 2020.DOI
  23. [23]J. Yong, Linear-quadratic optimal control problems for mean-field stochastic differential equations, SIAM Journal on Control and Optimization, 51 (2013), pp. 2809–2838.

Paper details

vlnc:W_bysm7bcs
Version
v2 · latest
Published
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