🤖 machine learning

Deep Operator BSDE: a Numerical Scheme to Approximate Solution Operators

This paper proposes and analyzes a neural network-based numerical scheme, combining Wiener chaos decomposition and the Euler method, to approximate solution operators of Backward Stochastic Differential Equations (BSDEs) with proven convergence and demonstrated accuracy.

Original authors: Pere Diaz-Lozano, Giulia Di Nunno

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pere Diaz-Lozano, Giulia Di Nunno

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 you are a master chef trying to perfect a recipe for a magical soup. In the world of finance, this "soup" is a complex mathematical tool called a Backward Stochastic Differential Equation (BSDE). It helps predict the future value of financial products, like options, by working backward from a known end date to the present.

Usually, if a chef wants to make a soup with a slightly different ingredient (a different "terminal condition"), they have to start from scratch: measure the pot, chop the vegetables, and stir the broth all over again. This is how traditional computer methods work. They solve the equation for one specific scenario, and if you want to know the answer for a different scenario, you have to run the whole expensive calculation again.

The Big Idea: The "Universal Soup Pot"
The authors of this paper, Pere Diaz Lozano and Giulia Di Nunno, asked a bold question: What if we could build a machine that learns the recipe itself, rather than just cooking one specific soup?

They propose a new method called Deep Operator BSDE. Instead of solving for one specific outcome, their method learns the Solution Operator. Think of this as a "Universal Soup Pot" that can instantly tell you how the soup will taste for any combination of ingredients you throw in, without needing to restart the cooking process.

How They Built the Machine
To make this work, they used two main ingredients:

Wiener Chaos Decomposition: Imagine the "ingredients" (the random future events) as a giant, messy pile of Lego bricks. This math trick breaks that messy pile down into neat, organized stacks of specific shapes (called Hermite polynomials). It turns a chaotic, infinite problem into a structured one that a computer can handle.
Neural Networks: These are the "smart cooks." Once the ingredients are organized, they use artificial intelligence (neural networks) to learn the relationship between the ingredients and the final taste.

The "Euler" Step-by-Step
The method works like a game of "telephone" played backward in time. They break the time period (say, one year) into small steps. At each step, the computer guesses the next move based on the current state and the "noise" of the market (modeled by something called Brownian motion). They proved mathematically that if you make the time steps small enough, this backward game will eventually land on the correct answer.

What They Ruled Out
It is important to note what this paper is not doing.

It is not just another way to solve a single equation faster. The authors explicitly argue against the idea that you need to re-run the algorithm from scratch for every new financial product. Their method is designed to handle a whole family of different products at once.
They do not claim this works for every possible mathematical scenario without limits. They admit that if the "ingredients" (the terminal conditions) are too wild or complex, the computer might struggle.

How Sure Are They?
The authors are very confident in the math. They have proved that their method converges (meaning it gets closer and closer to the right answer) under very mild assumptions. They also derived a specific "rate of convergence," which is a mathematical guarantee of how fast the error shrinks as they make the time steps smaller.

However, when it comes to the actual performance on real-world data, they rely on simulations. They didn't just guess; they built the algorithm and tested it on a computer.

The Results: A Taste Test
They put their "Universal Soup Pot" to the test with 22 different types of financial options (like Call, Put, and Asian options) in a standard market setting, and later with 28 different types in a more complex two-asset market.

The Score: In the first test, the error for the price prediction ( $Y$ ) was tiny, around 3.85 ± 1.05 (scaled by $10^{-3}$ ) for Call options. For the "hedging" part ( $Z$ , which is like the sensitivity of the soup to temperature changes), the error was around 1.27 ± 0.70 (scaled by $10^{-2}$ ).
The Comparison: They compared their "Universal Pot" against a traditional method (Monte Carlo simulation) that solves each recipe individually. Their method was able to learn the entire landscape of options at once. While the errors weren't zero, they were competitive, especially considering the massive advantage of not having to re-calculate everything from scratch.
The Catch: The authors warn that as the number of ingredients (dimensions) grows, the computer gets overwhelmed. This is the "curse of dimensionality." If you try to learn a soup with too many different types of spices at once, the neural network needs millions of parameters and a lot of computing power to get it right.

The Bottom Line
The paper demonstrates that it is possible to train a single AI model to understand the entire "recipe book" of financial derivatives, rather than just one dish. They proved the math works, and their simulations show it performs well across many different types of options. It's a step toward a future where financial risk calculators don't just give you a single number, but understand the whole family of possibilities instantly.

Technical Summary: Deep Operator BSDE: A Numerical Scheme to Approximate Solution Operators

Problem Statement
The paper addresses the numerical approximation of the solution operator for Backward Stochastic Differential Equations (BSDEs). While extensive literature exists on approximating the solution $(Y, Z)$ of a BSDE for a fixed terminal condition $\xi$ , existing methods typically require re-execution of the algorithm whenever $\xi$ changes. The authors aim to approximate the operator itself:
$\xi \mapsto (Y^\xi_t, Z^\xi_t)_{t \in [0,T]}$
where $(Y^\xi, Z^\xi)$ solves the BSDE with terminal condition $\xi$ . This is motivated by applications in dynamic risk measures and $g$ -conditional expectations, where one needs to evaluate the solution for a collection of terminal conditions rather than a single instance. The challenge lies in handling the infinite-dimensional nature of the space of terminal conditions $L^2(\mathcal{F}_T)$ and the computational difficulty of calculating conditional expectations in high dimensions.

Methodology
The proposed framework, termed Deep Operator BSDE, combines three main ingredients:

Wiener Chaos Decomposition: The terminal condition $\xi \in L^2(\mathcal{F}_T)$ is decomposed using an orthonormal basis of $L^2([0,T]; \mathbb{R}^d)$ and Hermite polynomials. This allows $\xi$ to be represented as a series of coefficients $d_a$ associated with basis functions $\Phi_a$ .
Infinite-Dimensional FBSDE Representation: The authors demonstrate that the solution to the BSDE with a general terminal condition $\xi$ can be viewed as the solution to a Forward-Backward SDE (FBSDE) system where the forward process $X_t$ is infinite-dimensional. Specifically, $X_t$ satisfies a linear SDE driven by the chaos basis, and the terminal condition of the backward equation becomes a linear functional of the final value of $X_T$ .
Markovian Representation and Operator Euler Scheme:
- Markovian Property: It is proven that the solution operators $Y_t(\xi)$ and $Z_t(\xi)$ admit a Markovian representation of the form $u(t, X_t, \xi)$ and $v(t, X_t, \xi)$ , where $u$ and $v$ are jointly measurable maps.
- Operator Euler Scheme: A recursive numerical scheme is derived, analogous to the classical backward Euler scheme for BSDEs. However, instead of computing random variables at each time step, the scheme computes operators (functions) $Y^\pi_i$ and $Z^\pi_i$ that map the current state of the infinite-dimensional process and the terminal condition to the solution.
- Finite-Dimensional Approximation: To make the problem computationally tractable, the infinite-dimensional chaos expansion is truncated to a finite order $p$ and a finite basis size $M$ . This reduces the problem to approximating finite-dimensional maps.
- Neural Network Implementation: Due to the high dimensionality of the truncated chaos space (the "curse of dimensionality"), the authors employ neural networks to approximate the measurable maps $u$ and $v$ . The training objective is a least-squares minimization of the BSDE residual over a compact set of terminal conditions defined by the truncated chaos coefficients.

Key Contributions

First Operator Approximation Scheme: The paper presents what it claims to be the first numerical scheme specifically designed to approximate the solution operator of a BSDE, rather than the solution for a fixed terminal condition.
Theoretical Convergence: The authors prove the convergence of the Operator Euler scheme to the true solution operator under very mild assumptions (uniform Lipschitz continuity of the generator). They provide a rate of convergence in terms of the time-step size $|\pi|$ and the regularity of the solution components ( $R_Y, R_Z$ ).
Joint Measurability: A significant theoretical contribution is the proof that the solution operators can be represented as jointly measurable functions of time, the infinite-dimensional state, and the terminal condition $\xi$ . This justifies the use of a single learning framework for varying $\xi$ .
Error Bounds: The paper establishes error bounds for the finite-dimensional truncation, showing that the approximation error depends on the tail of the Wiener chaos expansion of $\xi$ .

Results
The method was implemented and tested on two numerical examples:

Linear Generator (1D): Pricing and hedging of 22 different families of options (including European, barrier, Asian, and lookback options) in a Black-Scholes setting. The Deep Operator BSDE was trained on a high-dimensional set of coefficients and evaluated on specific families. Results showed mean relative errors for $Y_0$ (price) generally below $0.5\%$ and for $Z_0$ (hedge) generally below $2\%$ , comparable to baseline Monte Carlo methods.
Nonlinear Generator (2D): Pricing in a 2-dimensional Black-Scholes model with different borrowing and lending rates (nonlinear generator). The method was tested on 28 families of options. The results demonstrated the ability to handle structurally different terminal conditions (e.g., basket, spread, max/min options) within a single unified framework.

Significance and Claims
The paper claims that the primary significance of this work lies in its operator-based perspective. Unlike existing parametric methods that learn solutions for specific parameter vectors or fixed terminal conditions, this method learns the mapping from the entire space of terminal conditions to the solution.

Unified Framework: The authors state that their method allows for the approximation of solutions across "several families of terminal conditions" (e.g., both path-dependent and state-dependent payoffs) within a single training process.
Generality: By working directly with the BSDE formulation and Wiener chaos, the method handles a broader class of terminal conditions than Markovian FBSDE approaches, which rely on the terminal condition being a function of a forward diffusion's final value.
Limitations: The authors modestly acknowledge the "curse of dimensionality" as a key limitation. The dimension of the input space for the neural networks grows rapidly with the chaos order $p$ and basis size $M$ , making the optimization challenging and computationally expensive for very high-dimensional settings. They do not claim the method solves the curse of dimensionality but rather provides a framework to manage it via neural networks for moderately high dimensions.

The paper concludes that while the method may not always perfectly match baselines for specific instances due to the generalization required, it successfully demonstrates the feasibility of learning BSDE solution operators, offering a distinct advantage for problems requiring evaluation across diverse terminal conditions.

More like this