💰 quantitative finance

Convergence of the Markovian Iteration for Coupled FBSDEs via a Differentiation Approach

This paper proposes a novel differentiation-based approach to rigorously prove the convergence of a Markovian iteration method for solving coupled forward-backward stochastic differential equations with fully coupled drift terms, overcoming previous challenges in controlling Lipschitz constants for $Z$ -coupled systems.

Original authors: Zhipeng Huang, Cornelis W. Oosterlee

Published 2026-07-16

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhipeng Huang, Cornelis W. Oosterlee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe as a giant, chaotic dance floor where particles and prices are constantly bumping into each other, pushed by invisible hands and random gusts of wind. In the world of mathematics and finance, we try to predict where these dancers will end up using equations called Stochastic Differential Equations. These aren't your average high school algebra problems; they involve "forward" steps (predicting where a particle goes next) and "backward" steps (figuring out what value something has right now based on where it will end up in the future). When these two steps are tangled together—where the future depends on the present, and the present depends on the future—we call it a "coupled" system. Solving these is like trying to untangle a knot while wearing blindfolds and running on a treadmill. It's crucial for things like pricing complex financial derivatives or modeling how heat spreads through a material, but the math is notoriously difficult, especially when the "backward" part depends on a hidden variable that represents the speed of change.

This paper tackles a specific, stubborn knot in that dance: a situation where the backward step depends on a hidden variable called "Z," which acts like a scaled derivative or a measure of how sensitive the system is to random jolts. Previous methods could handle cases where the backward step only depended on the current value, but adding this "Z" dependency made the math explode with complexity. The authors, Zhipeng Huang and Cornelis W. Oosterlee, propose a clever new way to untangle this. Instead of trying to guess the "Z" value separately, they use a mathematical trick based on differentiation (finding the slope of a curve). They force the "Z" value to be exactly the slope of the "Y" value, just as physics dictates it should be. By doing this, they prove that their step-by-step guessing game (called Markovian iteration) will eventually settle down to the correct answer, rather than spinning out of control. They also built a computer algorithm to test this, and the results show it works accurately, offering a new, reliable tool for solving these tricky equations.

The Dance of Chance and the Tangled Knot

Let's break down the problem the authors are solving. Imagine you are trying to navigate a boat through a stormy sea. You have a map (the "forward" equation) that tells you how the wind and waves will push your boat forward. But here's the twist: the map itself changes depending on how you steer, and your steering depends on a forecast of where you'll be tomorrow (the "backward" equation). In many real-world scenarios, like pricing a complex financial option, the "forecast" isn't just about your current position; it also depends on how violently the waves are hitting you right now. This "violence" is the Z variable.

For a long time, mathematicians had a good way to solve these problems if the forecast only depended on your position. But when the forecast depended on the waves (Z) too, the existing methods hit a wall. The math became so unstable that the "guess-and-check" process used to solve it would often fail to converge, meaning the answers would bounce around forever without ever finding the truth. The authors of this paper asked: How can we make this guessing game stable when the waves are part of the equation?

The Magic of the Slope

The authors' solution is elegant and relies on a fundamental relationship between position and speed. In the world of these equations, there is a deep connection (known as the Feynman-Kac formula) between the value of the backward process (Y) and the "Z" process. Think of Y as the height of a hill and Z as the steepness of the slope at that exact spot. If you know the shape of the hill perfectly, you automatically know the slope. You don't need to measure the slope separately; it's mathematically locked to the shape of the hill.

Previous methods tried to guess the shape of the hill (Y) and the slope (Z) independently, like two people trying to guess a secret code without talking to each other. This led to confusion and errors. Huang and Oosterlee realized that if they forced the computer to calculate the slope (Z) directly from the shape of the hill (Y) using calculus (differentiation), the two would stay perfectly in sync.

They proposed a new iteration scheme where, at every step of the calculation:

They guess the shape of the hill (Y).
They immediately calculate the slope (Z) by taking the derivative of that guess.
They use this linked pair to update the forward path.
They repeat the process, refining the guess each time.

By doing this, they reduced the problem from managing two chaotic, independent variables to managing just one. This "differentiation approach" acts like a safety rail, keeping the calculations from wandering off the cliff.

Proving the Path is Safe

The paper doesn't just say, "Hey, this looks like it works." The authors went through a rigorous mathematical proof to show that this method is guaranteed to work under certain conditions. They had to prove that the "Lipschitz constants" (a fancy math way of saying "how much the answer can wiggle if the input wiggles") stay under control.

In the old methods, the wiggle-factor could grow uncontrollably with every step of the iteration, causing the whole system to collapse. The authors proved that by using their differentiation trick, they could keep these wiggle-factors bounded. They showed that as long as the coupling between the forward and backward parts isn't too extreme (a condition they call "weak coupling"), the sequence of guesses will converge to a single, stable solution.

They also derived error estimates, which are like a "margin of error" label on their solution. They proved that if you make your time steps smaller (taking more, smaller steps in the simulation), the error shrinks predictably. This gives users confidence that the answer they get is close to the true mathematical reality.

Testing the Theory in the Lab

To make sure their theory wasn't just beautiful math on paper, the authors built a computer algorithm and ran numerical experiments. They compared their new "Differentiation Approach" against the old "Direct Extension" method (which guesses Y and Z separately).

They tested two scenarios:

A simple case: A system where the backward part only depended on Z (the waves), not on Y (the position).
A complex case: A system where the forward part depended on both Y and Z, creating a fully tangled knot.

In both cases, their new method outperformed the old one. The results showed that their algorithm was not only stable but also highly accurate. The errors decreased as they refined the simulation, matching the predictions of their mathematical proofs. They found that by enforcing the relationship between the value and its slope, they could solve problems that were previously too difficult or unstable to handle with standard iterative methods.

Why This Matters

This paper is a significant step forward for anyone trying to solve complex, coupled systems of equations. Whether you are a financial engineer trying to price a derivative that reacts to market volatility, or a physicist modeling how heat and particles interact, the ability to solve these equations reliably is essential. The authors didn't just find a new way to solve a specific problem; they provided a framework that could be used to stabilize other numerical methods, including the popular "Deep BSDE" methods that use neural networks. By proving that these methods can be made to converge even with the tricky Z-coupling, they have opened the door to more accurate and reliable simulations in science and finance.

In short, Huang and Oosterlee took a tangled, unstable knot of equations and found a way to straighten it out by remembering a simple rule: if you know the shape, you know the slope. This simple insight, backed by rigorous proof and tested by computer, makes the impossible solvable.

Technical Summary: Convergence of the Markovian Iteration for Coupled FBSDEs via a Differentiation Approach

Problem Statement
This paper addresses the numerical solution of fully coupled forward-backward stochastic differential equations (FBSDEs) where the drift term of the forward process depends on the backward process $Y$ and its scaled derivative $Z$ . Specifically, the system is given by:
$\begin{cases} dX_t = b(t, X_t, Y_t, Z_t)dt + \sigma(t, X_t, Y_t)dW_t, \\ dY_t = -f(t, X_t, Y_t, Z_t)dt + Z_t dW_t, \\ Y_T = g(X_T). \end{cases}$
While previous work by Bender and Zhang (2008) established convergence for Markovian iteration schemes in $Y$ -coupled FBSDEs (where the forward drift does not depend on $Z$ ), extending these results to systems with $Z$ -coupling in the forward drift has remained challenging. The primary difficulty lies in obtaining uniform control over the Lipschitz constants of the decoupling fields across both time steps and iteration indices within a fixed-point framework. Without such control, the well-posedness and convergence of the discretized system cannot be rigorously guaranteed.

Methodology
The authors propose a novel differentiation-based approach to handle the $Z$ process. Instead of treating the approximation of the backward process $Y$ (denoted $u$ ) and the $Z$ process (denoted $v$ ) as two independent sequences of functions to be optimized, the method enforces a structural relationship derived from the nonlinear Feynman-Kac formula:
$v(t, x) = \partial_x u(t, x) \sigma(t, x, u(t, x)).$
In the proposed Markovian iteration scheme, the approximation $v_i^{\pi, m}$ at each time step is defined strictly as the spatial derivative of the approximation $u_i^{\pi, m}$ multiplied by the diffusion coefficient. This construction links the regularity of $v$ directly to that of $u$ .

The convergence analysis proceeds under the following framework:

Assumptions: The paper assumes standard Lipschitz and linear growth conditions on the coefficients $b, \sigma, f, g$ , along with generalized weak coupling and monotonicity conditions (e.g., small time duration $T$ , weak coupling of $Y$ into the forward SDE, or strong monotonicity in $f$ or $b$ ). Crucially, it assumes the existence of a classical solution to the associated quasi-linear PDE.
Uniform Bounds: By enforcing the differentiation relation, the authors derive uniform bounds for the Lipschitz constants and linear growth coefficients of the decoupling fields $u^{\pi, m}$ and $v^{\pi, m}$ across all iterations $m$ and time steps. This reduces the problem of controlling two intertwined sequences to controlling a single sequence ( $u$ ).
Convergence Proof: Using discrete Gronwall inequalities and contraction mapping arguments, the paper proves that the sequence of approximations converges to a unique solution of the discretized scheme. The error estimate combines the error from the time discretization ( $O(h)$ ) and the error from the iteration steps ( $O(c^m)$ for some contraction factor $c < 1$ ).

Key Contributions

Extension of Markovian Iteration: The paper extends the Markovian iteration framework of Bender and Zhang (2008) to fully coupled FBSDEs where the forward drift depends on $Z$ .
Differentiation-Based Regularity Control: The core theoretical contribution is the introduction of a differentiation-based treatment for the $Z$ process. This approach resolves the difficulty of controlling Lipschitz constants in the presence of $Z$ -coupling by reducing the dependency to a single sequence of decoupling fields.
Rigorous Convergence Analysis: The authors provide a rigorous proof of convergence for the proposed iteration scheme under generalized weak coupling and monotonicity conditions, deriving explicit error estimates for both the iteration count and the time step size.
Algorithmic Implementation: An efficient numerical algorithm is developed that preserves the differentiation relation by construction. This is achieved by parameterizing $Y$ and $Z$ with shared parameters and computing $Z$ via the automatic differentiation of the $Y$ -approximator, rather than solving separate regression problems.

Results

Theoretical: The paper establishes that under the stated assumptions, the discretized FBSDE system is well-posed. The Markovian iteration converges to the unique solution of the discretized system, and the global error is bounded by $C(m c^m + h)$ , where $m$ is the iteration number and $h$ is the time step.
Numerical: Numerical experiments compare the proposed algorithm against a "Direct Extension" of the method in Bender and Zhang (2008), where $Y$ and $Z$ are approximated independently. The experiments confirm the theoretical convergence rates with respect to both time discretization and iteration steps. The proposed method demonstrates superior stability and accuracy, particularly in scenarios where the forward drift depends on $Z$ .

Significance and Claims
The paper claims that its differentiation-based approach provides the necessary theoretical foundation to extend Markovian iteration methods to fully coupled FBSDEs. A significant byproduct of this analysis is that it offers theoretical support for the application of Deep BSDE methods (which rely on neural network approximations of decoupling fields) to FBSDEs with $Z$ -coupling. Prior to this work, convergence guarantees for Deep BSDE methods were largely limited to $Y$ -coupled cases. By proving the well-posedness of the discretized system with $Z$ -coupling, the authors provide a justification for using such deep learning-based solvers in more complex, fully coupled settings. The paper does not claim to solve the general non-Markovian case or remove the need for regularity assumptions on the PDE solution, but rather focuses on establishing convergence within the Markovian framework under specific structural conditions.

The Dance of Chance and the Tangled Knot

The Magic of the Slope

Proving the Path is Safe

Testing the Theory in the Lab

Why This Matters

More like this