betaFeedback

Reading v1View latest version

WorksMathematics

Rough backward stochastic differential equations

Abstract

We develop an intrinsic well-posedness theory for nonlinear backward stochastic differential equations (BSDEs) driven simultaneously by Brownian motion and a (level-22) rough path of finite pp-variation. Unlike earlier approaches based on smooth approximation or transformation methods, we formulate the equation directly by viewing its solution as a rough semimartingale (in the sense of \cite{friz2023rough}). This framework is particularly suited to BSDEs, whose Brownian martingale component is only implicitly defined and lacks the a priori time regularity required by stochastic-sewing-based controlled rough path methods \cite{fhl21,allan2024rough}. We establish comparison, existence, uniqueness, and stability under the natural regularity condition H∈CbγH\in C_b^γ, γ>pγ>p. The main analytical difficulty, namely, the loss of integrability arising from nonlinear composition, is overcome through conditional pp-variation norms with BMO-type properties. Finally, by randomizing the rough driver as a Brownian rough path, we establish a direct correspondence between rough BSDEs and backward doubly stochastic differential equations (BDSDEs).

2020 Mathematics Subject Classification. 60L20, 60H10.

Introduction

Let X\mathbf{X} be a pp-rough path with 2<p<32 < p < 3. On a filtered probability space carrying a Brownian motion WW, we consider the following rough backward stochastic differential equation (rough BSDE)

Yt=ξ+∫tTf(r,Yr,Zr) dr+∫tTH(Yr) dXr−∫tTZr dWr,t∈[0,T].(1)Y_t = \xi+ \int_t^T f(r,Y_r,Z_r)\,\mathrm{d}r + \int_t^T H(Y_r)\,\mathrm{d}\mathbf{X}_r - \int_t^T Z_r\,\mathrm{d}W_r,\qquad t \in[0,T]. \tag*{(1)}

This equation was originally introduced by Diehl and Friz [7]. Their construction starts with smooth paths Xn\mathbf{X}^n whose canonical rough path lifts Xn\mathbf{X}^n converge to X\mathbf{X}. The corresponding classical BSDEs are solved via a Doss–Sussmann transformation, and the solution to (1) is then defined as the limit of the resulting pairs (Yn,Zn)(Y^n,Z^n). Formally, the rough integral is interpreted as

∫tTH(Yr) dXr:=lim⁡n→∞∫tTH(Yrn) dXrn,(2)\int_t^T H(Y_r)\,\mathrm{d}\mathbf{X}_r := \lim_{n\to\infty}\int_t^T H(Y_r^n)\,\mathrm{d}\mathbf{X}_r^n, \tag*{(2)}

where the limit is independent of the approximating sequence and gives a robust solution map in the rough path topology. It does not, however, provide an intrinsic definition of the integral in (1). In particular, H(Y)H(Y) is not treated as an integrand within a rough stochastic integration theory, and the dynamics of YY are characterized only through approximation.

A direct theory for (1) is available in the Young regime. Diehl and Zhang [8] consider drivers of finite qq-variation with q<2q < 2 and establish well-posedness via a fixed-point argument. In that setting, the integral involving YY is a pathwise Young integral and the equation is satisfied directly. Song et al. [27, 28] further develop the nonlinear Young regime, providing a probabilistic interpretation of stochastic partial differential equations (SPDEs) driven by space-time noise in the Young regime. A general jump case is studied by Becherer and Sun [2]. For genuinely rough drivers, Liang and Tang [22] extend the approximation and flow approach of [7] to multidimensional BSDEs, under either a smallness condition or a diagonal structural assumption on the nonlinear rough coefficient. They also develop an intrinsic theory for multidimensional equations but with a linear rough drift. Their construction uses stochastic controlled rough paths, a pp-variation rough stochastic integral, and the stochastic sewing lemma. Stochastic controlled rough paths and stochastic sewing also play a central role in recent developments for rough (forward) stochastic differential equations; see, for example, [11, 1]. A reflected rough BSDE theory is introduced by Li et al. [21], who also study the associated rough PDEs with obstacles, as well as optimal stopping and American option pricing problems.

The aim of the present paper is to solve (1) intrinsically with a nonlinear rough coefficient H(Y)H(Y) appearing in the rough integral, without resorting to Doss–Sussmann transformation or defining the solution through smooth approximations. We show that the pair (Y,−H(Y))(Y,-H(Y)) belongs to an appropriate space of rough semimartingales (RSMs), originally introduced in [16]. Within this space, the integral against X\mathbf{X} in (1) is defined intrinsically as a rough stochastic integral via compensated Riemann sums:

∫tTH(Yr) dXr:=lim⁡∣π∣→0∑[u,v]∈π(H(Yu) δXu,v−DH(Yu)H(Yu) Xu,v),(3)\int_t^T H(Y_r)\,\mathrm{d}\mathbf{X}_r := \lim_{|\pi|\to0}\sum_{[u,v]\in\pi}\bigl(H(Y_u)\,\delta\mathbf{X}_{u,v}-DH(Y_u)H(Y_u)\,\mathbb{X}_{u,v}\bigr), \tag*{(3)}

where the limit is taken in probability. Thus, the solution pair (Y,Z)(Y,Z) satisfies (1) as a hybrid rough stochastic integral equation in the precise sense of Definition 4.1.

A notable feature of our work is that we are able to obtain all required estimates in the RSM setting. In particular, neither the construction of (3) nor the well-posedness proof uses the stochastic sewing lemma. This differs from the stochastic controlled rough path approach of Liang and Tang [22] and from the rough SDE theory developed in [11, 1]. The present work therefore provides an original alternative to stochastic sewing for natural classes of equations that combine rough signals with (implicit) martingale components.

The distinction is relevant for BSDEs because the martingale integrand ZZ is part of the solution and is known only implicitly as a general predictable process. Stochastic controlled rough path methods require suitable time regularity of the martingale component in an LkL^k-scale. Such regularity is not available from the definition of the classical BSDE and must normally be established as part of the analysis. A rough semimartingale, by contrast, is specified through a Doob–Meyer-type decomposition and does not impose time regularity on its martingale part. Moreover, its definition is independent of a prescribed LkL^k-integrability exponent. This allows the notion of solution to be formulated before any integrability estimates are proved.

The intrinsic formulation also allows the comparison principle and stability estimates to be established directly from the equation. Strict comparison, for example, is not generally preserved under passage to the limit: approximation readily gives non-strict ordering, but strictness may be lost. Similarly, direct difference estimates are difficult when the rough integral is known only through the approximation (2). Working in the RSM space, we prove a strict comparison theorem, an a priori estimate, existence and uniqueness, and local stability with respect to the terminal condition, the drift coefficient, and the driving rough path.

Our method also reduces the regularity required for the rough coefficient, and we only need

H∈Cbγ,for some γ>p.(4)H \in C^\gamma_b, \qquad\text{for some } \gamma> p. \tag*{(4)}

This is the standard regularity threshold for the corresponding RDE. In contrast, the flow transformation approach in [7] requires the stronger condition H∈Cbγ+2H \in C^{\gamma+2}_b. The additional two derivatives are needed to control the transformed stochastic equation and are not intrinsic to the rough dynamics. Since no transformation is used in the present approach, the strict comparison theorem, the a priori estimate, and the well-posedness result all hold under the natural condition (4).

We now describe the main estimates. Suppose first that YY is essentially bounded. To control the RSM norm of (Y,−H(Y))(Y,-H(Y)), one needs estimates for the remainder in the rough expansion of H(Y)H(Y). In contrast to the deterministic RDE case, this remainder contains a martingale component, which can be controlled via a recent result on pp-variation estimates for martingale transforms [16]. Its conditional version Lemma A.2 gives the required control of the mixed conditional moments of the pp-variation norm; see (32) in Proposition 3.5. This is the main input in the composition estimates for rough semimartingales.

An essential bound for YY is obtained through comparison. The difference of two solutions satisfies a linear rough BSDE. We derive an explicit exponential representation of this linear equation and combine it with a John–Nirenberg inequality for VMOp-var^{p\text{-var}} processes. This yields the comparison theorem and, by comparison with suitable deterministic RDEs, the required uniform bound. The same representation also yields strict comparison under the corresponding strict ordering assumptions.

The remaining difficulty is the loss of integrability caused by nonlinear composition. As observed in the stochastic controlled rough path setting [11], estimates for H(Y)−H(Y~)H(Y)-H(\widetilde{Y}) naturally require a higher integrability exponent than the one initially available for Y−Y~Y-\widetilde{Y}. We address this by proving an energy estimate for suitable increasing processes. It implies that the conditional norms ∥ ⁣∥⋅∥ ⁣∥p-var ∥ ∞;k,∞\|\!\| \cdot\|\!\|_{p\text{-var}\,\|\,\infty;k,\infty}, defined by conditional moments of the pp-variation norm, are equivalent for different values of kk. These norms have the same self-improving feature as classical kk-BMO norms, and we therefore refer to them as BMO-type norms. Their equivalence compensates for the integrability loss in the nonlinear composition estimate. This leads to a local contraction and, by iteration, to existence and uniqueness of the solution on the full time interval. The same estimates yield continuous dependence of the solution on the data.

Finally, we relate rough BSDEs to backward doubly stochastic differential equations (BDSDEs), famously introduced by Pardoux and Peng [24], as a counterpart in the BSDE theory for SPDEs. To this end, we first introduce a backward causality property for processes with rough parameters to describe their measurability with respect to the backward filtration. Then we show that the solution to rough BSDEs has such a measurable version with respect to the σ\sigma-algebra generated by the forward Brownian motion and the backward rough path flow, via the measurable selection initiated by Dellacherie and Meyer [4] and recently developed by Denkert et al. [6]. Randomizing XX by a Brownian rough path then produces a BDSDE, while the intrinsic rough integral identifies the corresponding backward stochastic integral. This, for the first time, provides a direct connection between the rough BSDE and BDSDE formulations without passing through smooth approximations of the additional Brownian signal. (Applications to semilinear rough and stochastic PDEs are also possible but will be pursued elsewhere.)

In conclusion, our contributions are threefold. Firstly, we introduce the intrinsic notion of solution to rough BSDEs in the framework of rough semimartingales, as well as the corresponding rough stochastic analysis (without reliance on - and complementary to - the work [11]). Secondly, we establish the well-posedness and overcome the problem of degenerating integrability by establishing energy estimates for BMO-type processes. Thirdly, we connect rough BSDEs and BDSDEs via a novel direct randomization argument.

Orginzation of Paper. In Section 2.1, we introduce notations and assumptions. Section 2.2 recalls the RSM framework and the construction of rough stochastic integrals. In Section 3.1, we prove energy estimates for increasing processes and the equivalence of the BMO-type pp-variation norms. Composition and integration estimates for RSMs are established in Section 3.2, and the RSM seminorm used in the stability analysis is introduced in Section 3.3. Section 4 contains the well-posedness theory. We first prove the comparison theorem in Section 4.1, derive the a prioria\ priori estimate in Section 4.2, and then establish existence, uniqueness, and stability under the condition (1.4) in Section 4.3. Section 5 studies the connection with BDSDEs, including backward causality in Section 5.1 and Brownian randomization in Section 5.2. Appendix A collects Lépinglé-type inequalities, conditional pp-variation estimates for martingale transforms, and estimates for RDEs with unbounded drift. The symbolic index is provided in Appendix B.

Mathematical Roadmap. First and importantly, Lemma 3.2 implies that the qq-variation of a continuous process is a BMO-process. Consequently, the conditional norms ∥⋅∥q-var∣∞;κ,∞\lVert\cdot\rVert_{q\text{-var}\mid\infty;\kappa,\infty} are comparable for all k>1k>1. Together with the conditional martingale-transform estimate of Lemma A.2, this underlies two parallel lines of the argument. The first controls a single solution: Proposition 3.5 bounds the remainder of H(Y)H(Y), Corollary 3.8 converts this into a bound for ∫H(Y) dX\int H(Y)\,dX, and Propositions 4.7 and 4.8 then show that, once ∥Y∥∞\lVert Y\rVert_\infty is controlled, both ∥δMY∥BMO\lVert\delta M^Y\rVert_{\mathrm{BMO}} and the RSM seminorm are bounded. The missing bound for ∥Y∥∞\lVert Y\rVert_\infty comes from a different direction: the comparison theorem (Theorem 4.3), proved via the exponential representation of the linearized equation and the John–Nirenberg inequality, compares YY with two deterministic RDEs and yields the a prioria\ priori estimate of Theorem 4.9. The second line of argument controls differences: Proposition 3.10, in which the equivalence of norms compensates for the integrability loss in the nonlinear composition, feeds into Lemma 3.17 and then into the contraction property of Proposition 4.16. Theorem 4.17 combines the contraction on small intervals with the a prioria\ priori estimate to obtain global well-posedness, and the same difference estimates give the stability result, Theorem 4.19. Section 5 relies on the preceding theory only for well-posedness, stability, and the RDE bound underlying the a prioria\ priori estimate.

Acknowledgment PKF and HZ acknowledge support from DFG CRC/TRR 388 “Rough Analysis, Stochastic Dynamics and Related Fields”, Projects A07, B04 and B05. Part of this work was carried out during a visit of the first author to Shandong University. JS is partially supported by National Natural Science Foundation of China (Grant Numbers 12521001, 12471142); JS and HZ are supported by the Fundamental Research Funds for the Central Universities. HZ is partially supported by NSF of Shandong (ZR2023MA026).

Preliminaries

Notations and Assumptions

We collect notations that will be used in this article. Let (E,∥⋅∥E)(E,\|\cdot\|_E) be a Banach space, and let VV and KK be Euclidean spaces. We denote by ∣⋅∣|\cdot| the Euclidean norm for both vectors and matrices. Some of the definitions introduced below depend on the underlying time interval or terminal time. For simplicity of notation, we omit this dependence when the interval is [0,T][0,T] (with terminal time TT) and no confusion arises.

  • Probability space: Fix a finite time horizon T>0T>0. For an integer d≥1d\ge1, let {Wt}t∈[0,T]\{W_t\}_{t\in[0,T]} be a dd-dimensional standard Brownian motion on the probability space (Ω,F,P)(\Omega,\mathcal{F},\mathbb{P}), and let (Ft)t∈[0,T](\mathcal{F}_t)_{t\in[0,T]} be the augmented filtration generated by WW. Denote the conditional expectation by Et[⋅]:=E[⋅∣Ft]\mathbb{E}_t[\cdot]:=\mathbb{E}[\cdot\mid\mathcal{F}_t].

  • Partition: For 0≤s≤t≤T0\le s\le t\le T, set Δ[s,t]:={(u,v):s≤u≤v≤t}\Delta_{[s,t]}:=\{(u,v):s\le u\le v\le t\} and Δ:=Δ[0,T]\Delta:=\Delta_{[0,T]}. Let P[s,t]\mathcal{P}_{[s,t]} denote the set of all partitions π={[ti,ti+1]; s=t0<t1<⋯<tn=t}\pi=\{[t_i,t_{i+1}];\ s=t_0<t_1<\cdots<t_n=t\} of [s,t][s,t], and define the mesh size of π\pi by ∣π∣:=max⁡0≤i<n(ti+1−ti)|\pi|:=\max_{0\le i<n}(t_{i+1}-t_i).

  • Control function: For a function w:Δ→R+w:\Delta\to\mathbb{R}^{+}, we call ww a control function if ww is continuous and satisfies that w(s,u)+w(u,t)≤w(s,t)w(s,u)+w(u,t)\le w(s,t) for s≤u≤ts\le u\le t. In particular, w(s,s)=0w(s,s)=0 for all ss.

  • Linear map: We denote by L(V,K)\mathcal{L}(V,K) the space of linear maps from VV to KK, endowed with the operator norm ∥F∥L(V,K)=sup⁡x∈V,∣x∣≤1∣F(x)∣\|F\|_{\mathcal{L}(V,K)}=\sup_{x\in V,|x|\le1}|F(x)|. We write V⊗KV\otimes K for the (algebraic) tensor product of VV and KK, equipped with the projective norm. In finite dimensions, the spaces L(V,K)\mathcal{L}(V,K) and K⊗VK\otimes V are canonically isomorphic, and the above norms are equivalent.

  • Space of functions: For any closed set U⊂VU\subset V and a measurable function f:U→Ef:U\to E, define the uniform norm by

∥f∥∞;U:=ess sup⁡r∈U∥f(r)∥E.\|f\|_{\infty;U}:=\operatorname*{ess\,sup}_{r\in U}\|f(r)\|_E.

If UU is clear from the context, we simply write ∥f∥∞\|f\|_\infty for ∥f∥∞;U\|f\|_{\infty;U}. Denote by C(U,E)C(U,E) the space of continuous functions f:U→Ef:U\to E. Denote by Cb(U,E)C_b(U,E) the space of functions f∈C(U,E)f\in C(U,E) with finite uniform norm.

For (s,t)∈Δ(s,t)\in\Delta and k≥1k\ge1, denote by Lk([s,t],E)L^k([s,t],E) the space of measurable functions f:[s,t]→Ef:[s,t]\to E such that

∥f∥Lk([s,t]):={∫st∣fr∣k dr}1/k<∞.\|f\|_{L^k([s,t])}:=\left\{\int_s^t |f_r|^k\,\mathrm{d}r\right\}^{1/k}<\infty.

For λ≥1\lambda\ge1, denote by Cλ-var(Δ[s,t],E)C^{\lambda\text{-var}}(\Delta_{[s,t]},E) the space of continuous functions f∈C(Δ[s,t],E)f\in C(\Delta_{[s,t]},E) such that

∥f∥λ-var;[s,t]:=sup⁡π∈P[s,t]{∑[u,v]∈π∥fu,v∥Eλ}1/λ<∞.\|f\|_{\lambda\text{-var};[s,t]}:=\sup_{\pi\in\mathcal{P}_{[s,t]}}\left\{\sum_{[u,v]\in\pi}\|f_{u,v}\|_E^\lambda\right\}^{1/\lambda}<\infty.

Also denote the λ\lambda-variation norm of ff as a two-parameter function defined by

References

  1. [1]Andrew L Allan and Jost Pieper. Rough stochastic analysis with jumps. arXiv preprint arXiv:2408.06978, 2024.arxiv.org/abs/2408.06978
  2. [2]Dirk Becherer and Yuchen Sun. Rough backward SDEs with discontinuous Young drivers. arXiv preprint arXiv:2505.20437, 2025.arxiv.org/abs/2505.20437
  3. [3]Jannis R Dause, Peter K. Friz, Arnulf Jentzen, and Jian Song. Controlled fields, rough stochastic calculus, and Itô-Wentzell-Alekseev-Grönber identities. arXiv preprint arXiv:2603.05388, 2026.arxiv.org/abs/2603.05388
  4. [4]Claude Dellacherie and Paul-André Meyer. Probabilities and potential, volume 29 of North-Holland Mathematics Studies. North-Holland Publishing Co., Amsterdam-New York, 1978.
  5. [5]Claude Dellacherie and Paul-André Meyer. Probabilities and Potential, volume 29 of North-Holland Mathematics Studies. North-Holland Publishing Company, Amsterdam, New York, Oxford, 1978. Translation of Probabilités et potentiel.
  6. [6]Robert Denkert, Idris Kharroubi, and Huyên Pham. A randomisation method for mean-field control problems with common noise. Electron. J. Probab., 30:Paper No. 189, 42, 2025.
  7. [7]Joscha Diehl and Peter K. Friz. Backward stochastic differential equations with rough drivers. Ann. Probab., 40(4):1715–1758, 2012.
  8. [8]Joscha Diehl and Jianfeng Zhang. Backward stochastic differential equations with Young drift. Probab. Uncertain. Quant. Risk, 2:Paper No. 5, 17, 2017.
  9. [9]Nicole El Karoui, Shige Peng, and Marie-Claire Quenez. Backward stochastic differential equations in finance. Math. Finance, 7(1):1–71, 1997.
  10. [10]Peter K. Friz and Martin Hairer. A course on rough paths. Universitext. Springer, Cham, second edition, [2020] ©2020. With an introduction to regularity structures.
  11. [11]Peter K. Friz, Antoine Hocquet, and Khoa Lê. Rough stochastic differential equations. arXiv preprint arXiv:2106.10340, 2021.arxiv.org/abs/2106.10340
  12. [12]Peter K. Friz, Khoa Lê, and Huilin Zhang. Controlled rough SDEs, pathwise stochastic control and dynamic programming principles. Annals of Probability (to appear), 2026. See https://imstat.org/journals-and-publications/annals-of-probability/annals-of-probability-future-papers/ and arXiv preprint: arXiv:2412.05698v2.arxiv.org/abs/2412.05698
  13. [13]Peter K. Friz, Khoa Lê, and Huilin Zhang. Randomisation of rough stochastic differential equations. In Dan Crisan, Ilya Chevyrev, Thomas Cass, James Foster, Christian Litterer, and Christopher Salvi, editors, Stochastic Analysis and Applications 2025, pages 51–71, Cham, 2026. Springer Nature Switzerland.
  14. [14]Peter K. Friz and Nicolas B. Victoir. Multidimensional stochastic processes as rough paths, volume 120 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2010. Theory and applications.
  15. [15]Peter K. Friz and Huilin Zhang. Differential equations driven by rough paths with jumps. J. Differential Equations, 264(10):6226–6301, 2018.
  16. [16]Peter K. Friz and Pavel Zorin-Kranich. Rough semimartingales and p-variation estimates for martingale transforms. Ann. Probab., 51(2):397–441, 2023.DOI
  17. [17]Ioannis Karatzas and Steven E. Shreve. Brownian motion and stochastic calculus, volume 113 of Graduate Texts in Mathematics. Springer-Verlag, New York, second edition, 1991.
  18. [18]Norihiko Kazamaki. Continuous exponential martingales and BMO, volume 1579 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1994.
  19. [19]Hiroshi Kunita. Stochastic flows and stochastic differential equations, volume 24. Cambridge university press, 1990.
  20. [20]Khoa Lê. Quantitative John–Nirenberg inequality for stochastic processes of bounded mean oscillation. arXiv preprint arXiv:2210.15736, 2022.arxiv.org/abs/2210.15736
  21. [21]Hanwu Li, Huilin Zhang, and Kuan Zhang. Reflected backward stochastic differential equations with rough drivers. Stochastic Process. Appl., 195:Paper No. 104874, 36, 2026.
  22. [22]Jiahao Liang and Shanjian Tang. Multidimensional backward stochastic differential equations with rough drifts. Trans. Amer. Math. Soc., 378(1):201–257, 2025.
  23. [23]David Nualart and Panqiu Xia. On nonlinear rough paths. ALEA Lat. Am. J. Probab. Math. Stat., 17(1):545–587, 2020.
  24. [24]Étienne Pardoux and Shige Peng. Backward doubly stochastic differential equations and systems of quasilinear SPDEs. Probab. Theory Related Fields, 98(2):209–227, 1994.
  25. [25]Sebastian Riedel and Michael Scheutzow. Rough differential equations with unbounded drift term. J. Differential Equations, 262(1):283–312, 2017.arxiv.org/abs/1605.05604
  26. [26]Jian Song, Xiaoming Song, and Qi Zhang. Nonlinear Feynman-Kac formulas for stochastic partial differential equations with space-time noise. SIAM J. Math. Anal., 51(2):955–990, 2019.DOI
  27. [27]Jian Song, Huilin Zhang, and Kuan Zhang. Backward stochastic differential equations with nonlinear Young drivers I. arXiv preprint arXiv:2504.18632, 2025.arxiv.org/abs/2504.18632
  28. [28]Jian Song, Huilin Zhang, and Kuan Zhang. Backward stochastic differential equations with nonlinear Young drivers II. arXiv preprint arXiv:2509.05183, 2025.arxiv.org/abs/2509.05183
  29. [29]Josef Teichmann. Another approach to some rough and stochastic partial differential equations. Stoch. Dyn., 11(2-3):535–550, 2011.
  30. [30]Jianfeng Zhang. Backward stochastic differential equations, volume 86 of Probability Theory and Stochastic Modelling. Springer, New York, 2017. From linear to fully nonlinear theory.

Paper details

vlnc:W_q3jg7hj7
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