Backward Doubly Stochastic DEs (BDSDEs)

• BDSDEs are backward stochastic systems that incorporate dual noise processes, integrating forward Itô and backward stochastic integrals.
• They generalize classical BSDEs to include jumps, multivalued dynamics, and mean-field extensions, enabling robust representations for various SPDEs.
• They are analyzed under Lipschitz and monotonicity conditions to ensure existence, uniqueness, and convergence of numerical schemes.

A backward doubly stochastic differential equation (BDSDE) is a backward stochastic system on a probability space $(\Omega, \mathcal F, \mathbb P)$ equipped with two mutually independent processes—typically a Brownian motion $W$ (forward-time noise) and an independent Brownian motion or pure-jump Lévy process $B, L$ (backward-time or discontinuous noise)—where the dynamics for an adapted process pair $(Y_t, Z_t)$ for $t\in[0,T]$ incorporates both forward Itô and backward stochastic integrals. BDSDEs generalize classical backward SDEs (BSDEs) to settings involving backward noise channels or jumps, and provide a probabilistic representation for wide classes of semilinear, possibly non-Markovian, stochastic PDEs with continuous and/or discontinuous noise.

1. Canonical Formulation and Filtration Structure

A prototypical BDSDE driven by two independent Brownian motions, in scalar case, is

$$Y_t = \xi + \int_t^T f(s, Y_s, Z_s) ds + \int_t^T g(s, Y_s, Z_s) d\overleftarrow{B}_s - \int_t^T Z_s dW_s, \quad 0 \leq t \leq T,$$

where

• $\xi$ is an $\mathcal F_T$-measurable terminal random variable,
• $f, g$ are measurable generators, typically Lipschitz or with weaker monotonicity/continuity,
• $d\overleftarrow{B}_s$ denotes the backward Itô integral (integration with respect to decreasing time),
• $(Y_t, Z_t)$ is adapted to the two-way filtration $$\mathcal F_t = \sigma \{ W_s: s \leq t \} \vee \sigma \{ B_r-B_t: t \leq r \leq T \} \vee \mathcal N$$, which is neither increasing nor decreasing.

Generalizations include:

• Backward doubly SDEs with jumps, i.e., driven by Brownian motion and Poisson or Lévy processes (Zhu et al., 2010, Ren et al., 2010, Owo et al., 2021).
• Reflected and/or multivalued BDSDEs imposing state constraints (Anis et al., 2014, Elmansouri et al., 30 Sep 2025).
• Mean-field BDSDEs involving functional dependence on the joint law of the solution (Buckdahn et al., 2021).
• BDSDEs with discontinuous, quadratic, anticipated, or time-inhomogeneous coefficients (Zhu et al., 2010, Hu et al., 2022, Xu, 2012).

2. Structural Assumptions and Well-Posedness

Classical well-posedness theory (existence, uniqueness, stability) for BDSDEs, stemming from Pardoux and Peng's foundational work, is established under the following analytic conditions:

• Lipschitz continuity in $(y,z)$ for both the drift $f$ and backward coefficient $g$:

$$|f(t, y_1, z_1) - f(t, y_2, z_2)| \leq L (|y_1 - y_2| + |z_1 - z_2|),$$

$$|g(t, y_1, z_1) - g(t, y_2, z_2)|^2 \leq L (|y_1 - y_2|^2 + \alpha |z_1 - z_2|^2),\quad \alpha < 1.$$

• Monotonicity properties (e.g., one-sided Lipschitz, left-Lipschitz, or monotone continuity in $y$), enabling extensions to discontinuous or only left-continuous coefficients (Zhu et al., 2010, Lin, 2010), as well as sublinear or quadratic growth in $z$ (Hu et al., 2022).
• Filtration assumptions: Existence, uniqueness, and stability theory typically assume the nonstandard, non-monotonic two-way filtration described above.

Key results under such assumptions include existence and uniqueness of adapted solutions $(Y, Z)$ in suitable $L^p$ spaces, a priori estimates, and comparison principles (Aman, 2010, Shi et al., 2010). Weakening these structural hypotheses to allow for discontinuity or only one-sided continuity in $y$, or non-Lipschitz (e.g., sublinear, quadratic, or uniformly continuous) growth in $z$, is feasible via monotone-approximation and sandwiching arguments (Lin, 2010, Hu et al., 2022).

3. Extended Classes: Jumps, Multivaluedness, and Mean-Field

Several substantial generalizations of the classical BDSDE framework are now established:

• BDSDEs with jumps: Incorporation of forward integrals with respect to Poisson random measures, Lévy processes, or Teugels martingales, allowing stochastic coefficients to be discontinuous (with jump-compatibility conditions), and enabling probabilistic representation for stochastic integro-differential equations (SPDIEs) (Zhu et al., 2010, Ren et al., 2010, Owo et al., 2021, Elmansouri et al., 30 Sep 2025).
• Multivalued BDSDEs: Equations in which the drift includes a subdifferential operator of a convex function, yielding a maximal monotone multivalued inclusion. Well-posedness arises via Yosida approximation and contractive estimates (Ren et al., 2010).
• Mean-field BDSDEs: Equations whose coefficients depend on the law of the solution and/or an associated forward process (McKean–Vlasov type), requiring $L^2$-only differentiability and nonlocal PDE representations. Existence and uniqueness hold for arbitrary law-Lipschitz constants in the backward coefficient—a sharp difference from classical theory which restricts the size of the noise-Lipschitz constant (Buckdahn et al., 2021).
• Reflected BDSDEs: Enforcing a Skorokhod minimality condition to keep the solution above a given barrier, with applications to obstacle problems for SPDEs with Neumann or nonlinear boundary conditions (Anis et al., 2014, Elmansouri et al., 30 Sep 2025).

4. Numerical Schemes and Regularity

Discretization of BDSDEs inherits the dual complexity of forward (Itô) and backward (Kunita–Itô or Marcus-type) stochastic integrals. Notable schemes include:

• Backward Euler and time-splitting methods: Explicit and implicit recursive backward schemes for the pair $(Y^n, Z^n)$, with pathwise time discretization in both forward and backward integration (Aman, 2010, Hu et al., 2017, Bao et al., 2021, Matoussi et al., 2014).
• Convergence rates: Under standard regularity, schemes achieve strong $L^2$ rates of $O(\Delta^{1/2})$ or $O(\Delta)$, depending on the interpolation and approximation of conditional expectations (Aman, 2010).
• Regularity estimates: New notions of $L^2$-regularity (modulus of continuity in $Y$ and time-averaged $L^2$-regularity in $Z$) proved for BDSDEs and exploited both analytically (e.g., for the convergence proofs) and numerically (Aman, 2010, Hu et al., 2017). Malliavin calculus techniques yield strong path-regularity in the non-smooth setting (Hu et al., 2017).

5. Probabilistic Representations for SPDEs

BDSDEs supply a comprehensive probabilistic framework for representing, solving, and numerically approximating solutions and stochastic viscosity solutions to a wide array of SPDEs, reflected or not, with continuous or discontinuous paths:

• Semilinear SPDEs: The Markovian BDSDE solution $Y^{t,x}$ at $t$ is $u(t,x)$, the solution to the corresponding semilinear SPDE; $Z^{t,x}$ gives the gradient term through $\sigma^\top \nabla_x u$ (Aman, 2010, Anis et al., 2014, Hu et al., 2022, Matoussi et al., 2014).
• SPDEs with reflection/obstacle/problem: Reflecting BDSDEs map to weak/variational solutions of SPDEs with Dirichlet or nonlinear Neumann conditions, and can characterize the reflection measure via the transition law under stochastic flow (Anis et al., 2014, Elmansouri et al., 30 Sep 2025).
• SPDEs with jumps: BDSDEs driven by Brownian and Poisson/Lévy processes yield probabilistic representations for SPDIEs with nonlocal (integro-differential) terms, even under non-Lipschitz or weak monotonicity assumptions (Zhu et al., 2010, Ren et al., 2010, Owo et al., 2021).

The table below summarizes prototypical BDSDE forms and their primary SPDE connections:

BDSDE Type	Key Features	Associated SPDE
Brownian–Brownian	Both forward and backward noise	Semilinear parabolic
Brownian–Lévy/Poisson	Forward: jumps, Backward: diffusion	SPDIE/nonlocal
Reflected (Barrier)	Skorokhod reflection, obstacle constraint	Obstacle/variational
Multivalued	Subdifferential drift (monotone)	Nonlinear inclusion (MSPDE)
Mean-field	Law-dependence in coefficients	Nonlocal McKean–Vlasov

6. Nonstandard and Advanced Topics

• Quadratic BDSDEs: Theory handles $Z$-quadratic generators with maximal/minimal solution comparison, monotone stability, and full Sobolev SPDE representation (Hu et al., 2022).
• Anticipated BDSDEs: Drivers depending on future values of solutions, with duality to delayed SDEs and extension of contraction mapping principles for path-dependent data (Xu, 2012).
• Kneser-type Theorems: In the continuous-coefficient setting, BDSDEs may possess either unique or continuum-many adapted solutions, depending on the invertibility and range properties of the backward coefficient (Shi et al., 2010).
• Rough Paths Framework: Extension to jointly rough backward SDEs (BDSDEs driven by Brownian and rough-path noise) has been accomplished for discontinuous or finite $q<2$-variation rough paths, via randomized conditional solutions and Picard–BMO iteration (Becherer et al., 26 May 2025).

7. Open Issues, Limitations, and Extensions

A plausible implication is that as BDSDE theory moves toward covering multivalued, mean-field, fully non-Markovian, or infinite-dimensional settings, additional analytical, computational, and measurability challenges become central—for example, obtaining well-defined Itô (or Marcus) backward integrals on path spaces, proving stability under rough/Young perturbations, or achieving regularity/sensitivity results on spaces of probability measures (Becherer et al., 26 May 2025, Buckdahn et al., 2021). Strong links between BDSDEs and the probabilistic representations of SPDEs in divergence or weak forms suggest ongoing advances may further expand the class of SPDEs tractable through stochastic analysis and Monte Carlo, especially under weak regularity, discontinuous coefficients, or in high dimensions (Zhu et al., 2010, Ren et al., 2010, Hu et al., 2022).

BDSDEs now function as a central unifying tool for non-Markovian filtering, stochastic control under dual randomness, nonlinear PDEs with backward input, and reflected/multivalued/obstacle problems, underpinning both mathematical foundations and computational methods at the intersection of stochastic analysis and PDE theory.