Paper: arXiv 2308.14487
Authors: Daniel Bussell, Camilo Andrés García-Trillos
Abstract
We propose a new multistep deep learning-based algorithm for the resolution of moderate to high dimensional nonlinear backward stochastic differential equations (BSDEs) and their corresponding parabolic partial differential equations (PDE). Our algorithm relies on the iterated time discretisation of the BSDE and approximates its solution and gradient using deep neural networks and automatic differentiation at each time step. The approximations are obtained by sequential minimisation of local quadratic loss functions at each time step through stochastic gradient descent. We provide an analysis of approximation error in the case of a network architecture with weight constraints requiring only low regularity conditions on the generator of the BSDE. The algorithm increases accuracy from its single step parent model and has reduced complexity when compared to similar models in the literature.
Complexity vs Empirical Score
- Math Complexity: 9.5/10
- Empirical Rigor: 3.0/10
- Quadrant: Lab Rats — high math complexity, limited empirical testing
Why this score: The paper presents advanced theoretical mathematics involving stochastic calculus, PDEs, and neural network approximation theory, with rigorous error analysis. However, it lacks direct empirical finance application, backtesting, or financial data, focusing instead on numerical algorithm validation for PDEs.
Research Flowchart
flowchart TD A["Research Goal: Solve high-dimensional non-linear PDEs and BSDEs"] --> B["Methodology: Iterated time discretization of BSDE"] B --> C["Data/Inputs: Generator, initial conditions, network architecture with weight constraints"] C --> D["Computational Process: Sequential minimization of local quadratic loss via Stochastic Gradient Descent"] D --> E["Computation: Deep neural networks & automatic differentiation for solution and gradient approximation"] E --> F["Outcome 1: Increased accuracy vs. single-step models"] E --> G["Outcome 2: Reduced computational complexity compared to similar models"] E --> H["Outcome 3: Theoretical error analysis under low regularity conditions"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a new multistep deep learning‑based algorithm for solving moderate to high dimensional nonlinear backward stochastic differential equations (BSDEs) and their corresponding parabolic PDEs.
Source sentence
We propose a new multistep deep learning-based algorithm for the resolution of moderate to high dimensional nonlinear backward stochastic differential equations (BSDEs) and their corresponding parabolic partial differential equations (PDE).
Assumes 4
- X0 is in L3(F0, Rd). model
Source sentence
(i) X0 ∈L3(F0, Rd) (ii) The functions µ and σ are uniformly Lipschitz in t, Lipschitz in x and continuously differ-entiable in x.
- The functions μ and σ are uniformly Lipschitz in t, Lipschitz in x and continuously differentiable in x. model
Source sentence
(i) X0 ∈L3(F0, Rd) (ii) The functions µ and σ are uniformly Lipschitz in t, Lipschitz in x and continuously differ-entiable in x.
- The generator f is Lipschitz continuous in x, y, z and 1/2‑Hölder continuous in time. model
Source sentence
(iii) The generator f is Lipschitz continuous in x, y, z and 1/2-H¨older continuous in time, that is: ∃K > 0 such that for all (t, x, y, z) and (t′, x′, y′, z′) ∈[0, T] × Rd × R × Rd, |f(t, x, y, z) −f(t′, x′, y′, z′)| ≤K(|t −t′|1/2 + |x −x′|2 + |y −y′| + |z −z′|2).
- The function g satisfies a linear growth condition. model
Source sentence
(v) The function g satisfies a linear growth condition.
Data
- Unbounded example (50) synthetic unknown
Source sentence
6.2. Unbounded example. We consider the parameters (50) ( µ = 0, σ = Id √ d f(x, y, z) = k(x) + y √ d(1d.z) + y2 2 DEEP MULTI-STEP MIXED ALGORITHM FOR HIGH-DIMENSIONAL BSDES 15 such that the solution of the PDE is given by u(t, x) = T −t d d X i=1 (sin(xi)1xi<0 + xi1xi>0) + cos d X i=1 ixi ! .
Costs
Not applicable
Finds 3
- In dimension d=1 the DADM scheme achieves an averaged value 1.3735 with relative error 0.29%, outperforming DBDP2 (0.41%) and DBSDE (0.49%). performance
Source sentence
The results are listed in Figure 5 and we show the plots in dimension 1 for the DADM Scheme below Scheme errors Averaged value Standard devia-tion Relative error (%) (DBDP2) 1.3719 0.0040 0.41 (DBSDE) 1.3708 0.0031 0.49 (DADM) 1.3735 0.0041 0.29 Table 4.
- In dimension d=8 the DADM scheme achieves an averaged value 1.0675 with relative error 8.01%, improving over DBDP2 (11.02%) while DBSDE does not converge. performance
Source sentence
16 DANIEL BUSSELL AND CAMILO ANDR´ES GARC´IA TRILLOS Scheme errors Averaged value Standard devia-tion Relative error (DBDP2) 1.0327 0.0081 11.02 (DBSDE) NotConv NotConv NotConv (DADM) 1.0675 0.01241 8.01 Table 5.
- Overall, the DADM scheme outperforms the benchmark DBSDE and its one‑step counterpart DBDP2 in terms of smaller relative error across tested dimensions. theoretical
Source sentence
We can see that,from our example (6.1) ,when the maturity is large our the DADM scheme outperforms the benchmark DBSDE as that method fails to converge due to the high number of times steps in the discretisation, We also note that our scheme outperforms its one step counterpart the DBDP2 scheme in terms of smaller relative error.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 20 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.