betaFeedback

WorksMathematics

On the Pontryagin Maximum Principle for Rough Stochastic Differential Equations

Abstract

We establish first- and second-order Pontryagin (a.k.a. stochastic) maximum principles for rough stochastic control problems with general nonlinear Lipγ\mathrm{Lip}^γ-rough vector fields. This complements the dynamic programming approach of Friz-Lê-Zhang [FLZ26b], and extends the maximum principle of Horst-Zhang [HZ26] beyond affine rough vector fields. The paper concludes with applications to pathwise stochastic optimal control.

Introduction

Classical stochastic control concerns the choice of an adapted control that minimizes an expected cost associated with a stochastic differential equation. Its two principal approaches are dynamic programming and Pontryagin’s maximum principle; see [40, 39, 10]. Dynamic programming describes the value function through a Hamilton–Jacobi–Bellman equation, whereas the maximum principle expresses optimality through the state and backward adjoint processes. The theory of backward stochastic differential equations (BSDEs), initiated by Pardoux and Peng [37], provides the natural language for these adjoints; see also [41]. When the diffusion depends on the control and the control space is nonconvex, the second-order maximum principle of Peng [38] shows that a second adjoint is needed to account for the case of spike kindly perturbed diffusion. An extension of this setting arises when the state is also affected by an exogenous signal whose realization is fixed throughout the optimization. The remaining randomness is still described by Brownian motions, but the fixed signal may be too irregular for ordinary differential calculus. Rough stochastic differential equations (RSDEs) provide a framework in which the two sources of noise can be treated simultaneously: one retains Itô integration for the Brownian motion and uses rough integration for the fixed signal. This separates the probabilistic information available to the controller from the temporal irregularity of the exogenous environment. Additionally, separation between fixed rough noise and stochastic dynamics is useful beyond control. It underlies robust filtering with correlated noise and multidimensional observations [13], rough filtering equations and rough Kalman–Bucy theory [9], and conditional pricing in local stochastic volatility models [4]. It also appears in McKean–Vlasov equations with rough common noise [23], linear-quadratic mean-field games with rough common noise [22], and the compactification approach to rough common-noise games [5]. These developments motivate optimality conditions for the underlying controlled RSDEs beyond the affine setting.

For mixed Brownian and rough dynamics, Diehl, Oberhauser and Riedel [17] constructed a joint lift of Brownian motion and a deterministic rough path, with applications to filtering and rough PDEs. In recent years, Lê’s stochastic sewing lemma [Lê20] made it possible to exploit adaptedness and conditional cancellations, leading Friz, Hocquet and Lê [19] to rough stochastic integration and further to an intrinsically formulated RSDE theory with stochastic controlled vector fields. The linear RSDE theory was developed by Bugini, Coghi and Nilssen [8], Horst and Zhang [29]. Related pp-variation estimates and integration theories were established by Friz and Zorin-Kranich [21] and, in a setting allowing jumps, by Allan and Pieper [1]. Within the RSDE framework, Friz, Lê and Zhang [25] established a rough dynamic programming principle and investigated the associated rough HJB equations. Their work on randomization [24] clarifies how a fixed rough signal can be replaced by a random one and how the resulting dynamics relate to conditional laws. Subsequently, Horst and Zhang [29] initiated the maximum-principle approach for systems with an affine rough coefficient. By employing a flow transformation they derived necessary and sufficient conditions which accommodate unbounded coefficients, including linear-quadratic models. In the present paper, we extend the maximum-principle approach by establishing necessary and sufficient conditions for such systems with a nonlinear rough coefficient. The drift and Brownian diffusion may both depend on the control, and the coefficients may be random. We develop a direct variational approach that combines estimates for forward rough SDEs, well-posedness of backward adjoint equations, and rough stochastic integration by parts. Our results cover both linear perturbations of Euclidean-valued controls and spike perturbations in a general separable metric control space. The essential point is to carry out the first- and second-order duality arguments when the rough coefficients of the variational and adjoint equations depend on the stochastic state.

Controlled stochastic systems from a pathwise viewpoint

The pathwise interpretation on stochastic control has a long history. Davis and Burstein [14] studied anticipative control through deterministic control problems parametrized by Wiener paths and suitable Lagrange multipliers. Lions and Souganidis [34, 35] developed a theory of fully nonlinear stochastic PDEs with applications to pathwise stochastic control. Buckdahn and Ma [BM07] established a dynamic programming principle and a stochastic viscosity characterization of the value field, using a Doss–Sussmann transformation to relate the pathwise problem to an auxiliary stochastic control problem. Further connections between pathwise control and possibly non-adapted stochastic PDEs were investigated in [6]. These works explain why conditioning on an exogenous noise leads to questions that are not covered directly by the usual Markovian control theory. Rough path theory, introduced by Lyons [36], gives a deterministic meaning to differential equations driven by irregular signals with extra algebraic (Lie) group structure. Gubinelli’s controlled-path formulation [28] is particularly convenient for the expansions and integration arguments used here; we refer to [18, 20] for systematic description on both formulations of rough paths. A related application to fully nonlinear stochastic PDEs was developed by Caruana, Friz and Oberhauser [11]. For the application of rough paths in optimal control, Diehl, Friz and Gassiat [16] established a maximum principle, a rough HJB characterization, and a stochastic control duality formula for controlled rough differential equations. Allan and Cohen [2] studied pathwise control in connection with robust filtering, including the degeneracy that can arise when the rough coefficient is directly controlled. Relaxed controls and rough HJB equations are considered by Chakraborty, Honnappa and Tindel [12]; see also [3]. Zheng and Hu [42] studied partially observed systems driven by Brownian and fractional Brownian motions, using transformations for systems whose fractional-noise coefficients are linear in the state. Recently, Lew [32] derived a first-order maximum principle for deterministic controls and Gaussian rough drivers, with a pathwise adjoint construction and an application to indirect shooting.

We here derive maximum principles for a pathwise stochastic control problem where the system is driven by an internal Brownian noise and another external Brownian noise. In this setting, controls remain adapted to the internal Brownian motion driving the Itô integral and may or may not depend (see Remark 5.2) on the entire trajectory of the external noise, thereby allowing anticipation of its future evolution. The objective is to minimize the conditional expected cost given this trajectory. By lifting the external Brownian motion to its geometric rough path and freezing its realization, we reduce the problem to the rough stochastic control framework developed in earlier sections. The main results then yield Hamiltonian necessary conditions for conditional optimality, expressed through rough adjoint BSDEs, together with sufficient conditions under the stated convexity and rough-remainder assumptions. These expand the maximum-principle approach to pathwise control of [29] to nonlinear rough vector fields, and complement the dynamic programming approach of [25] through a forward–backward characterization of optimal controls.

Problem formulation and main results

Let WW be a ww-dimensional Brownian motion and let η∈Cgρ,α([0,T];Rℓ)\eta\in\mathcal{C}^{\rho,\alpha}_g([0,T];\mathbb{R}^{\ell}) be a (deterministic) geometric rough path with α∈(13,12)\alpha\in\left(\frac{1}{3},\frac{1}{2}\right). For an admissible control u∈Au \in\mathcal{A}, we consider the Rd\mathbb{R}^{d}-valued state process governed by the rough stochastic differential equation (RSDE)

dXsη,u=b(s,Xsη,u,us) ds+σ(s,Xsη,u,us) dWs+(f,f′)(s,Xsη,u) dηs,s∈[t,T],Xtη,u=x.(1)\begin{aligned} \mathrm{d}X_s^{\eta,u} &= b(s,X_s^{\eta,u},u_s)\,\mathrm{d}s + \sigma(s,X_s^{\eta,u},u_s)\,\mathrm{d}W_s + (f,f')(s,X_s^{\eta,u})\,\mathrm{d}\eta_s,\qquad s \in[t,T],\\ X_t^{\eta,u} &= x. \tag*{(1)} \end{aligned}

Here the Brownian integral is understood in the Itô sense, and (f,f′)(f,f') is a stochastic controlled vector field relative to η\eta, with the prime representing its Gubinelli derivative (see e.g. [20]). This formulation allows us to admit progressively measurable controls with respect to the Brownian filtration, while retaining control dependence in the Brownian diffusion. Throughout, we suppress the dependence of the coefficients on ω\omega when no confusion can arise.

The objective is to minimize the cost functional

Jη(u;t,x)=E[∫tTh(s,Xsη,u,us) ds+g(XTη,u)].(2)\mathcal{J}^{\eta}(u;t,x) = \mathbb{E}\left[\int_t^T h(s,X_s^{\eta,u},u_s)\,\mathrm{d}s+g(X_T^{\eta,u})\right]. \tag*{(2)}

over a class A\mathcal{A} of progressively measurable controls on [t,T][t,T]. The corresponding value function is

Vη(t,x):=inf⁡u∈AJη(u;t,x).(3)V^{\eta}(t,x):=\inf_{u\in\mathcal{A}}\mathcal{J}^{\eta}(u;t,x). \tag*{(3)}

The analytic results we obtained lead to two forms of Pontryagin’s maximum principle. The first gives a derivative formula and a Hamiltonian minimization condition for Euclidean-valued controls and coefficients differentiable in the state and control variables. The second is a rough stochastic counterpart of the second-order principle in [38], allowing a nonconvex control space and a controlled Brownian diffusion, differentiability is required only in the state variable, with stronger spatial regularity to justify the second-order expansion.

For the first maximum principle, we consider linear perturbations uϵ=u+ϵvu^\epsilon=u+\epsilon v. Under Assumptions (A1)–(A3) and 4.5, we prove Gateux differentiability of the cost functional and represent its derivative through a first adjoint equation. More precisely, let

H(ω,s,x,y,z,u):=y⊤b(ω,s,x,u)+Tr⁡(z⊤σ(ω,s,x,u))+h(ω,s,x,u).H(\omega,s,x,y,z,u):=y^\top b(\omega,s,x,u)+\operatorname{Tr}\left(z^\top\sigma(\omega,s,x,u)\right)+h(\omega,s,x,u).

Writing X=Xη,uX=X^{\eta,u}, the adjoint pair ((Y,Y′),Z)((Y,Y'),Z), solves

Ys=∂xg(XT)⊤+∫sT∂xH(r,Xr,Yr,Zr,ur)⊤ dr+∫sT∂xf(r,Xr)⊤Yr dηr−∫sTZr dWr.Y_s=\partial_xg(X_T)^\top+\int_s^T\partial_xH(r,X_r,Y_r,Z_r,u_r)^\top\,\mathrm{d}r+\int_s^T\partial_xf(r,X_r)^\top Y_r\,\mathrm{d}\eta_r-\int_s^T Z_r\,\mathrm{d}W_r.

The rough integral is understood with the associated controlled derivative (computable from Yr′=∂xf(r,Xr)⊤YrY'_r=\partial_xf(r,X_r)^\top Y_r). The following is a loose formulation of our first findings, with full details left to Proposition 4.7 and Theorem 4.8.

Theorem 1.1 (Linear perturbation). The functional Jη\mathcal{J}^{\eta} is Gateux differentiable, and for every v∈Av\in\mathcal{A} its derivative at the point u∈Au\in\mathcal{A} is given by

ddϵJη(u+ϵv;t,x)∣ϵ=0=E[∫tT∂uH(s,Xs,Ys,Zs,us)⋅vs ds].\left.\frac{\mathrm{d}}{\mathrm{d}\epsilon}\mathcal{J}^{\eta}(u+\epsilon v;t,x)\right|_{\epsilon=0} =\mathbb{E}\left[\int_t^T\partial_uH(s,X_s,Y_s,Z_s,u_s)\cdot v_s\,\mathrm{d}s\right].

Thus, an optimal control (ut)t∈[0,T]∈A(u_t)_{t\in[0,T]}\in\mathcal{A} of Jη\mathcal{J}^{\eta} satisfies the usual stationarity condition

∂uH(s,Xs,Ys,Zs,us)=0,P⊗Leb⁡(ds)–a.s.\partial_uH(s,X_s,Y_s,Z_s,u_s)=0,\quad\mathbb{P}\otimes\operatorname{Leb}(\mathrm{d}s)\text{--a.s.}

If additionally HH is convex in the control variable, this yields the standard pointwise minimization of the Hamiltonian, where for every u∈Rκu\in\mathbb{R}^{\kappa} we have

H(s,Xs,Ys,Zs,us)≤H(s,Xs,Ys,Zs,u),P⊗Leb⁡(ds)–a.s.H(s,X_s,Y_s,Z_s,u_s)\leq H(s,X_s,Y_s,Z_s,u),\quad\mathbb{P}\otimes\operatorname{Leb}(\mathrm{d}s)\text{--a.s.}

Notice that although the rough coefficient does not enter HH explicitly, it affects both the state and the adjoint processes.

Our second maximum principle applies to controls with values in a general separable metric space UU. We replace an optimal control uˉ\bar{u} by another admissible control on an interval of length ϵ\epsilon. The resulting spike perturbation of the Brownian integral has size ϵ1/2\epsilon^{1/2}, so its quadratic contribution must be retained when expanding the cost to order ϵ\epsilon. Under Assumptions (B1)–(B4) and 4.15 we construct first and second variations VϵV^\epsilon and ζϵ\zeta^\epsilon and prove

Vϵ=O(ϵ1/2),ζϵ=O(ϵ),Xη,uϵ=Xˉ+Vϵ+ζϵ+o(ϵ),V^\epsilon=O(\epsilon^{1/2}),\qquad\zeta^\epsilon=O(\epsilon),\qquad X^{\eta,u^\epsilon}=\bar{X}+V^\epsilon+\zeta^\epsilon+o(\epsilon),

in the moment and stochastic controlled pp-variation norms specified in Lemma 4.14. Using again the same Hamiltonian, but now denoted with Hˉ\bar H due to the domain difference at the control argument, we consequently denote the corresponding first adjoint by ((Yˉ,Yˉ′),Zˉ)((\bar Y,\bar Y'),\bar Z). Furthermore, the second-order duality argument introduces a matrix-valued second adjoint pair ((Pˉ,Pˉ′),Qˉ)((\bar P,\bar P'),\bar Q), which in addition to terms linear in Pˉ\bar P, its rough driver contains the contraction of the first adjoint with ∂xx2f(⋅,Xˉ)\partial^2_{xx}f(\cdot,\bar X); see (102). This is precisely the additional term caused by the nonlinearity of the rough coefficient. Once again, we summarize our main findings as stated in Theorem 4.16 in the following informal Theorem.

References

  1. [1][AP26] Andrew L Allan and Jost Pieper. “Rough stochastic analysis with jumps”. In: Electronic Journal of Probability 31 (2026), pp. 1–62.arxiv.org/abs/2408.06978
  2. [2][AC20] Andrew L. Allan and Samuel N. Cohen. “Pathwise stochastic control with applications to robust filtering”. In: The Annals of Applied Probability 30.5 (2020), pp. 2274–2310. DOI: 10.1214/19-AAP1558.DOI
  3. [3][Ash+26] Estepan Ashkarian, Prakash Chakraborty, Harsha Honnappa, and Samy Tindel. The Pontryagin maximum principle and Q-functions in rough environments. 2026. arXiv: 2601.05354 [math.OC].arxiv.org/abs/2601.05354
  4. [4]Peter Bank, Christian Bayer, Peter K Friz, and Luca Pelizzari. “Rough PDEs for local stochastic volatility models”. In: Mathematical Finance 35.3 (2025), pp. 661–681.
  5. [5]Erhan Bayraktar, Xihao He, Xiang Yu, and Fengyi Yuan. Mean-field games with rough common noise: the compactification approach. 2026. arXiv: 2602.22602 [math.PR].arxiv.org/abs/2602.22602
  6. [6]Neeraj Bhauryal, Ana Bela Cruzeiro, and Carlos Oliveira. “Pathwise stochastic control and a class of stochastic partial differential equations”. In: Journal of Optimization Theory and Applications 203.2 (2024), pp. 1967–1990.arxiv.org/abs/2301.09214
  7. [8]Fabio Bugini, Michele Coghi, and Torstein Nilssen. “Malliavin calculus for rough stochastic differential equations”. In: arXiv preprint arXiv:2402.12056 (2024).arxiv.org/abs/2402.12056
  8. [9]Fabio Bugini, Peter K Friz, Khoa Lê, and Huilin Zhang. “Rough stochastic filtering”. In: arXiv preprint arXiv:2509.11825 (2025).arxiv.org/abs/2509.11825
  9. [10]René Carmona and François Delarue. Probabilistic theory of mean field games with applications I: Mean Field FBSDEs, Control and Games. Vol. 83. Springer, 2018.
  10. [11]Michael Caruana, Peter K. Friz, and Harald Oberhauser. “A (rough) pathwise approach to a class of non-linear stochastic partial differential equations”. In: Annales de l’Institut Henri Poincaré C, Analyse non linéaire 28.1 (Feb. 2011), pp. 27–46. DOI: 10.1016/j.anihpc.2010.11.002.DOI
  11. [12]Prakash Chakraborty, Harsha Honnappa, and Samy Tindel. Pathwise Relaxed Optimal Control of Rough Differential Equations. 2024. arXiv: 2402.17900 [math.OC].arxiv.org/abs/2402.17900
  12. [13]D. Crisan, J. Diehl, P. K. Friz, and H. Oberhauser. “Robust filtering: correlated noise and multi-dimensional observation”. In: Ann. Appl. Probab. 23.5 (2013), pp. 2139–2160. DOI: 10.1214/12-AAP896.DOI
  13. [14]Mark Davis and Gabriel Burstein. “A deterministic approach to stochastic optimal control with application to anticipative control”. In: Stochastics: An International Journal of Probability and Stochastic Processes 40.3-4 (1992), pp. 203–256.DOI
  14. [15]Joscha Diehl and Peter Friz. “Backward stochastic differential equations with rough drivers”. In: The Annals of Probability 40.4 (July 2012). DOI: 10.1214/11-aop660.DOI
  15. [16]Joscha Diehl, Peter K Friz, and Paul Gassiat. “Stochastic control with rough paths”. In: Applied Mathematics & Optimization 75.2 (2017), pp. 285–315.arxiv.org/abs/1303.7160
  16. [17]Joscha Diehl, Harald Oberhauser, and Sebastian Riedel. “A Lévy area between Brownian motion and rough paths with applications to robust nonlinear filtering and rough partial differential equations”. In: Stochastic processes and their applications 125.1 (2015), pp. 161–181.DOI
  17. [18]P. Friz and N. Victoir. Multidimensional Stochastic Processes as Rough Paths. Vol. 120. Cambridge Studies in Advanced Mathematics. Cambridge: Cambridge University Press, 2010, pp. xiv+670. DOI: 10.1017/CBO9780511845079.DOI
  18. [19]Peter Friz, Antoine Hocquet, and Khoa Lê. “Rough stochastic differential equations”. In: arXiv preprint arXiv:2106.10340, v7 (2026).
  19. [20]Peter K Friz and Martin Hairer. A Course on Rough Paths: With an Introduction to Regularity Structures. Springer Nature, 2020.DOI
  20. [21]Peter K Friz and Pavel Zorin-Kranich. “Rough semimartingales and p-variation estimates for martingale transforms”. In: The Annals of Probability 51.2 (2023), pp. 397–441.DOI
  21. [22]Peter K. Friz, Ioannis Gasteratos, Ulrich Horst, and Stefanos Theodorakopoulos. Mean-field games with rough common noise: the linear-quadratic case. 2026. arXiv: 2602.19210v3 [math.PR].arxiv.org/abs/2602.19210
  22. [23]Peter K. Friz, Antoine Hocquet, and Khoa Lê. “McKean–Vlasov equations with rough common noise”. In: arXiv preprint arXiv:2507.13149 (2025).
  23. [24]Peter K. Friz, Khoa Le, and Huilin Zhang. “Randomisation of Rough Stochastic Differential Equations”. In: Stochastic Analysis and Applications 2025: In Honour of Terry Lyons (SAAT 2025). Springer Proceedings in Mathematics and Statistics. Springer, 2026, pp. 55–71. DOI: 10.1007/978-3-032-03914-9_2.DOI
  24. [25]Peter K. Friz, Khoa Lê, and Huilin Zhang. “Controlled rough SDEs, pathwise stochastic control and dynamic programming principles”. In: The Annals of Probability 54.6 (2026), pp. 2952–2983. DOI: 10.1214/26-AOP1852.DOI
  25. [26]Peter K. Friz, Jian Song, Huilin Zhang, and Kuan Zhang. Rough backward stochastic differential equations. 2026. arXiv: 2609.06606 [math.PR].arxiv.org/abs/2609.06606
  26. [27]Peter K. Friz and Huilin Zhang. “Differential Equations Driven by Rough Paths with Jumps”. In: Journal of Differential Equations 264.10 (2018), pp. 6226–6301. DOI: 10.1016/j.jde.2018.01.031.DOI
  27. [28]Massimiliano Gubinelli. “Controlling rough paths”. In: Journal of Functional Analysis 216.1 (2004), pp. 86–140. DOI: 10.1016/j.jfa.2004.01.002.DOI
  28. [29]Ulrich Horst and Huilin Zhang. “Pontryagin Maximum Principle for Rough Stochastic Systems and Pathwise Stochastic Control”. In: SIAM Journal on Control and Optimization 64.5 (2026), pp. 3609–3644. DOI: 10.1137/25M1749670.DOI
  29. [30]Ioannis Karatzas and Steven Shreve. Brownian motion and stochastic calculus. Springer, 2014.
  30. [32]Thomas Lew. “Rough Stochastic Pontryagin Maximum Principle and an Indirect Shooting Method”. In: SIAM Journal on Control and Optimization 64.2 (2026), pp. 869–905. DOI: 10.1137/25M1742163.DOI
  31. [33]Jiahao Liang and Shanjian Tang. “Multidimensional Backward Stochastic Differential Equations with Rough Drifts”. In: Transactions of the American Mathematical Society 378.1 (2025), pp. 547–586. DOI: 10.1090/tran/9237.DOI
  32. [34]Pierre-Louis Lions and Panagiotis E. Souganidis. “Fully nonlinear stochastic partial differential equations”. In: C. R. Acad. Sci. Paris Sér. I Math. 326.9 (1998), pp. 1085–1092. DOI: 10.1016/S0764-4442(98)80067-0.DOI
  33. [35]Pierre-Louis Lions and Panagiotis E. Souganidis. “Fully nonlinear stochastic partial differential equations: non-smooth equations and applications”. In: Comptes Rendus de l’Académie des Sciences - Series I - Mathematics 327.8 (1998), pp. 735–741. DOI: 10.1016/S0764-4442(98)80161-4.DOI
  34. [36]Terry J. Lyons. “Differential equations driven by rough signals”. In: Revista Matemática Iberoamericana 14.2 (1998), pp. 215–310. DOI: 10.4171/RMI/240.DOI
  35. [37]Etienne Pardoux and Shige Peng. “Adapted solution of a backward stochastic differential equation”. In: Systems & control letters 14.1 (1990), pp. 55–61.DOI
  36. [38]Shige Peng. “A General Stochastic Maximum Principle for Optimal Control Problems”. In: SIAM Journal on Control and Optimization 28.4 (1990), pp. 966–979. DOI: 10.1137/0328054.
  37. [39]Huyên Pham. Continuous-time stochastic control and optimization with financial applications. Vol. 61. Springer Science & Business Media, 2009.DOI
  38. [40]Jiongmin Yong and Xun Yu Zhou. Stochastic controls: Hamiltonian systems and HJB equations. Vol. 43. Springer Science & Business Media, 1999.DOI
  39. [41]Jianfeng Zhang. Backward Stochastic Differential Equations: From Linear to Fully Nonlinear Theory. Springer, 2017.DOI
  40. [42]Yueyang Zheng and Yaozhong Hu. “The Global Maximum Principle for Optimal Control of Partially Observed Stochastic Systems Driven by Fractional Brownian Motion”. In: SIAM Journal on Control and Optimization 62.1 (2024), pp. 509–538. DOI: 10.1137/22M1543203.DOI

Paper details

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

views · 5

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