Solving high-dimensional partial differential equations using deep learning
The Deep BSDE method solved 100-dimensional PDEs with 0.17%–0.46% relative error.
Key Findings
Methodology
The paper reformulates semilinear parabolic PDEs as forward–backward stochastic differential equations. The state process is discretized with Euler–Maruyama, while fully connected neural networks at each time slice approximate Z=σᵀ∇u. The initial value and gradient are trainable parameters, and all parameters are optimized through the terminal mismatch E|g(X_T)−Û_T|². Implementations use ReLU, batch normalization, Adam, and TensorFlow.
Key Results
- For the 100-dimensional nonlinear Black–Scholes equation with default risk, the method achieved 0.46% relative error in 1,607 seconds. With T=1, δ=2/3, and R=0.02, the reference price was 57.300, whereas omitting default risk produced 60.781.
- For the 100-dimensional HJB equation, the relative error was 0.17% at λ=1 in 330 seconds. For the 100-dimensional Allen–Cahn equation at t=0.3, the error was 0.30% in 647 seconds.
- A depth ablation reduced error from 2.29% with 29 trainable layers to 0.53% with 145 layers; intermediate values were 0.90%, 0.60%, and 0.56%, showing that richer gradient representations improve accuracy.
Significance
The work converts high-dimensional nonlinear PDE solving from a grid-based numerical problem into learning over stochastic trajectories, avoiding explicit tensor grids whose size grows exponentially with dimension. It offers a practical route to model 100 assets, agents, or resources jointly, rather than imposing representative-agent or low-dimensional interaction assumptions. The sub-percent errors on three distinct PDE families suggest that deep networks can function as scalable numerical solvers, not merely generic regressors, with direct relevance to finance, control, operations research, and physics.
Technical Contribution
The central contribution is the Deep BSDE Solver: time-indexed neural networks approximate the BSDE control variable Z, while Euler propagation couples them into a single computational graph. Unlike direct regression of u or grid discretization, the method learns Z=σᵀ∇u and trains all time slices through one terminal-condition objective. The architecture contains (H+1)(N−1) trainable layers when each subnetwork has H hidden layers, turning gradient estimation into explicit representation learning and avoiding repeated Hessian construction.
Novelty
The novelty is not the isolated use of BSDEs or neural networks, but their structured integration. A separate network represents the gradient-related quantity at every time step, and stochastic dynamics propagate these local predictions into a terminal value. Compared with Feynman–Kac Monte Carlo for linear equations, branching diffusion for restricted nonlinearities, and grid methods for low dimensions, the approach targets a substantially broader class of nonlinear parabolic PDEs.
Limitations
- Only several benchmark equations at dimension 100 are studied. The method relies on stochastic optimization, time discretization, and network capacity, and the paper does not provide a universal dimension-independent error guarantee.
- Training cost grows with the number of time slices, batch size, and network width. The method is mainly demonstrated for pointwise or regional values and may become unstable with nonsmooth terminal data or severe nonlinearities.
- The paper explicitly does not solve quantum many-body problems involving the Pauli exclusion principle; its reference solutions are also numerical approximations.
Future Work
Promising directions include adaptive time stepping, temporal parameter sharing, variance reduction, GPU parallelization, and rigorous error decompositions. The framework could be extended to hundreds or thousands of dimensions, coupled systems, constrained controls, and nonsmooth terminal conditions. The authors particularly identify multi-agent economics, full-portfolio finance, and large-scale resource allocation as important application areas, while the quantum many-body setting remains an open challenge.
AI Executive Summary
High-dimensional PDEs describe baskets of financial assets, multi-agent control systems, and many-particle physics. Yet conventional finite-difference, spectral, and tensor-grid methods suffer from the curse of dimensionality: their computational burden grows exponentially with dimension. Linear parabolic equations can often be treated with Feynman–Kac and Monte Carlo, but nonlinear equations remain difficult. Han, Jentzen, and E address this gap with the Deep BSDE method.
The method rewrites a semilinear parabolic PDE as a forward–backward stochastic differential equation. Euler–Maruyama simulates the forward state process, while a separate fully connected neural network at each time slice approximates Z=σᵀ∇u. The initial value and gradient are trainable variables. These components are assembled into a deep computational graph, and Adam minimizes the squared mismatch between the propagated terminal output and g(X_T). ReLU activation and batch normalization are used in the reported implementation.
The experiments are deliberately high-dimensional: all three benchmark families use dimension 100. The nonlinear Black–Scholes problem reaches 0.46% relative error in 1,607 seconds; the HJB problem reaches 0.17% in 330 seconds; and the Allen–Cahn problem reaches 0.30% in 647 seconds. In the pricing example, including default risk gives a reference value of 57.300, compared with 60.781 when default is ignored. The results establish a compelling computational proof of concept for joint multi-asset pricing, dynamic control, and reaction–diffusion simulation, while leaving general theory, extreme scaling, and complex physical constraints unresolved.
Deep Analysis
Background
High-dimensional PDEs arise in quantum many-body systems, basket derivatives, dynamic programming, and resource allocation. Classical polynomial, piecewise, wavelet, spectral, and grid methods commonly encounter exponential scaling. Feynman–Kac plus Monte Carlo handles linear parabolic equations; Darbon–Osher’s Hopf-formula methods address selected inviscid Hamilton–Jacobi problems; multilevel Picard iteration and branching diffusion extend some nonlinear cases. Their applicability, stability, or efficiency remains limited for general nonlinear parabolic PDEs.
Core Problem
The target is u_t+1/2Tr(σσᵀHess u)+∇u·μ+f(t,x,u,σᵀ∇u)=0 with terminal condition u(T,x)=g(x), evaluated at u(0,ξ). The difficulty is simultaneous high dimensionality, nonlinear dependence on the gradient, exponentially large grids, and unstable estimation of derivatives. A useful solver must learn values and dynamic sensitivities without enumerating the full state space.
Innovation
The paper contributes four linked ideas: 1) a Markovian BSDE representation of the PDE; 2) direct approximation of Z=σᵀ∇u rather than a global value function; 3) stacking time-indexed subnetworks with Euler state propagation into one end-to-end graph; and 4) a terminal-condition loss that trains every time slice jointly. This combines stochastic simulation, control-like gradient learning, and deep compositional approximation.
Methodology
- �� Sampling: simulate X_{t_{n+1}}=X_{t_n}+μΔt+σΔW with d-dimensional Brownian increments.
- �� Representation: use a multilayer feedforward network at t_n to approximate (σᵀ∇u)(t_n,X_{t_n}).
- �� Propagation: update Y_{n+1}=Y_n−f_nΔt+Z_nᵀΔW_n.
- �� Optimization: train θ={θu0,θ∇u0,θ1,…,θN−1} using E|g(X_T)−Ŷ_T|².
- �� Implementation: each subnetwork has four layers, two d+10 hidden layers, ReLU and batch normalization; batch size is 64 and Adam is used. Experiments use N=20 or 40 and report five independent runs.
Experiments
All principal tests use dimension 100. The Black–Scholes setup uses T=1, δ=2/3, R=0.02, μ̄=0.02, σ̄=0.2, terminal payoff min_i x_i, N=40, and learning rate 0.008. HJB uses g(x)=ln((1+||x||²)/2), N=20, λ=1, and learning rate 0.01. Allen–Cahn uses g(x)=1/(2+0.4||x||²), N=20, and learning rate 0.0005. References come from multilevel Picard, Monte Carlo, or branching diffusion.
Results
The solver achieves sub-percent errors across three qualitatively different problems: 0.46% for nonlinear Black–Scholes, 0.17% for HJB, and 0.30% for Allen–Cahn. Runtime is 1,607, 330, and 647 seconds, respectively. HJB costs decrease as control strength λ increases, matching optimal-control intuition. In the depth study, error falls from 2.29% at 29 trainable layers to 0.53% at 145 layers, with reported standard deviations around 0.0014–0.0026.
Applications
Financial institutions can price claims on many underlying assets while incorporating nonlinear default risk. Control and operations-research systems can optimize many devices, agents, or resources jointly. Physics researchers can simulate high-dimensional reaction–diffusion dynamics. Preconditions include a tractable diffusion model, computable terminal condition, and evaluable drift, diffusion, and nonlinear generator.
Limitations & Outlook
The method depends on temporal discretization, trajectory coverage, and nonconvex stochastic optimization. Gradient noise, nonsmooth terminal conditions, degenerate diffusion, and strong nonlinearities may cause instability. The demonstrations use dimension 100 and run on a 2.9 GHz Intel Core i5 MacBook Pro with 16 GB memory, so industrial-scale claims are premature. The authors also acknowledge that Pauli-constrained quantum many-body problems remain outside the method’s reach.
Plain Language Accessible to non-experts
Imagine managing a gigantic factory with hundreds of moving parts. A traditional method would place a sensor at every possible position of every machine. With many machines, the number of sensor combinations becomes impossibly large. The Deep BSDE approach instead sends many inspectors along randomly chosen production routes. At each checkpoint, a small learning system looks at the current situation and recommends how the factory should adjust next. When an inspector reaches the end, the system compares the final outcome with the required target and changes all earlier recommendations to improve future routes.
The important trick is that the system does not memorize the whole factory. It learns useful decisions along many representative journeys. In the paper, this idea reaches 0.46% error for a 100-asset pricing problem, 0.17% for a 100-dimensional control problem, and 0.30% for a 100-dimensional physics problem. It is not magic: training still requires many trials, and unusual rules may confuse the learner. But it replaces an impossible complete map with a practical collection of guided journeys.
ELI14 Explained like you're 14
Imagine a video game with 100 control knobs. Every second, the game moves randomly, and you must choose actions so your final score is as good as possible. A normal strategy would try to list every possible game situation. With 100 knobs, that list becomes absurdly huge. The paper’s method trains an AI adventure team instead. At every checkpoint, a small neural network sees the current situation and suggests the next adjustment. The team plays until the end, compares its final score with the target, and updates all the advice.
This is called the Deep BSDE method. BSDE is basically a mathematical way to describe how information travels backward from the final score while the simulated game moves forward. Adam is the coach that improves the networks after each batch of trials. The results are impressive: the error is 0.46% in a 100-dimensional finance problem, 0.17% in a control problem, and 0.30% in a physics problem.
It is like learning how to navigate a giant maze without drawing every corridor. Deeper networks helped too: increasing trainable layers from 29 to 145 reduced error from 2.29% to 0.53%. Still, the method is not unbeatable. Training can be noisy, and special physics rules—such as quantum particles refusing to occupy forbidden states—remain difficult!
Glossary
Deep BSDE method
A neural numerical method for high-dimensional nonlinear parabolic PDEs. It combines BSDE simulation, time-indexed subnetworks, and terminal-loss optimization.
The paper’s central solver.
Backward stochastic differential equation (BSDE)
A stochastic integral equation constrained by a terminal condition and interpreted backward in time. Its Markovian solution encodes PDE values and gradients.
The bridge between PDEs and learning.
Feynman–Kac formula
A representation of linear parabolic PDE solutions as expectations over stochastic processes. BSDEs extend this principle to important nonlinear settings.
Discussed as prior methodology.
Hamilton–Jacobi–Bellman equation
A PDE for the value function of an optimal-control problem. Its nonlinear gradient term reflects optimization over controls.
Used in the 100-dimensional LQG experiment.
Curse of dimensionality
The phenomenon in which computational or representational complexity grows exponentially with dimension. Tensor grids are a canonical example.
The main obstacle addressed by the paper.
Euler–Maruyama discretization
A finite-step approximation for stochastic differential equations using drift, diffusion, and Brownian increments. It turns continuous paths into a trainable computation graph.
Used for forward and backward propagation.
Open Questions Unanswered questions from this research
- 1 A general theory for approximation error, stability, and dimension dependence in nonlinear Deep BSDE solvers remains incomplete; the paper’s strongest evidence is empirical.
- 2 It is unclear whether training remains reliable in hundreds or thousands of dimensions, with nonsmooth terminal conditions, degenerate diffusion, or severe nonlinearities. Adaptive sampling and variance reduction are needed.
- 3 Embedding complex physical constraints, especially the Pauli exclusion principle in quantum many-body systems, remains an explicit open problem.
Applications
Immediate Applications
Multi-asset default-risk pricing
Banks can model many underlying assets jointly in a nonlinear Black–Scholes equation with default intensity Q(u). The paper’s reference price is 57.300 with default risk versus 60.781 when it is ignored, demonstrating material pricing impact.
High-dimensional dynamic control
Operations and energy researchers can use an HJB formulation to coordinate many devices or resources. The method is most suitable when the diffusion process and terminal cost can be simulated and evaluated reliably.
Long-term Vision
Joint multi-agent and multiparticle modeling
The framework could replace representative-agent approximations with models containing many interacting agents, instruments, or particles. Major obstacles are scalable training, coupled dynamics, uncertainty quantification, and hard physical constraints.
Large-scale scientific computing
With adaptive architectures and accelerator hardware, Deep BSDE solvers may become practical for high-dimensional reaction–diffusion and stochastic control problems. Rigorous reliability estimates will be essential for safety-critical deployment.
Abstract
Developing algorithms for solving high-dimensional partial differential equations (PDEs) has been an exceedingly difficult task for a long time, due to the notoriously difficult problem known as the "curse of dimensionality". This paper introduces a deep learning-based approach that can handle general high-dimensional parabolic PDEs. To this end, the PDEs are reformulated using backward stochastic differential equations and the gradient of the unknown solution is approximated by neural networks, very much in the spirit of deep reinforcement learning with the gradient acting as the policy function. Numerical results on examples including the nonlinear Black-Scholes equation, the Hamilton-Jacobi-Bellman equation, and the Allen-Cahn equation suggest that the proposed algorithm is quite effective in high dimensions, in terms of both accuracy and cost. This opens up new possibilities in economics, finance, operational research, and physics, by considering all participating agents, assets, resources, or particles together at the same time, instead of making ad hoc assumptions on their inter-relationships.