Partial differential equations (PDEs) are the mathematical bedrock of the physical and financial sciences. They describe how quantum particles move, how stock options change in value, and how fluids flow. Yet, for decades, scientists have hit a hard mathematical wall when trying to solve these equations in high dimensions.
This wall is the curse of dimensionality. If you want to solve a PDE across 100 dimensions using a traditional grid-based method (like finite differences or finite elements), and you place a meager 10 grid points along each axis, you suddenly need to compute solutions at $10^{100}$ distinct points—a number that vastly exceeds the number of atoms in the observable universe.
A rapidly growing body of research has fundamentally shifted this paradigm. By marrying deep learning with stochastic calculus, researchers have developed mesh-free methods capable of solving non-linear, high-dimensional PDEs with computational costs that scale polynomially rather than exponentially. Here is the intuition behind why these methods work.
The Probabilistic Escape Hatch
To understand why deep learning can break the curse of dimensionality, we first have to look at a classical loophole in numerical analysis: the connection between diffusion PDEs and stochastic processes.
Consider a linear parabolic PDE like the Black-Scholes equation used in quantitative finance. The classical Feynman–Kac formula establishes that the solution to this PDE at a specific starting point can be represented as the expected value of a random path (a stochastic differential equation, or SDE).
$$\text{PDE Solution at } (x, t) = \mathbb{E}\left[ \text{Payoff of a random trajectory starting at } x \right]$$
This realization changes the game. Instead of building a massive grid over the entire 100-dimensional space, we can simply simulate random trajectories starting from our point of interest using Monte Carlo sampling. Crucially, the error rate of a Monte Carlo simulation scales as $\mathcal{O}(1/\sqrt{M})$, where $M$ is the number of samples. This error rate is completely independent of the number of dimensions $d$.
While this probabilistic approach works beautifully for linear equations, real-world problems—such as risk management or fluid dynamics—are often non-linear. To address this, Christian Beck et al. in On nonlinear Feynman–Kac formulas for viscosity solutions of semilinear parabolic partial differential equations extended this linkage to the non-linear realm. They proved that solutions to certain stochastic fixed-point equations correspond rigorously to "viscosity solutions" of semilinear parabolic PDEs, providing the theoretical justification needed to deploy high-dimensional stochastic solvers like Multilevel Picard (MLP) algorithms. Furthermore, Martin Hutzenthaler et al. demonstrated in Overcoming the curse of dimensionality in the numerical approximation of parabolic partial differential equations with gradient-dependent nonlinearities that even when the non-linearity depends on the gradient of the solution, stochastic approximation algorithms can achieve polynomial complexity scaling.
Neural Networks as Continuous, Mesh-Free Approximators
If Monte Carlo methods can estimate the solution at a single point without a grid, how do we learn the solution over an entire domain? This is where deep neural networks come in. Instead of storing values at discrete grid vertices, we train a neural network $u(x; \theta)$ to act as a continuous, mesh-free function approximator.
Neural networks are uniquely suited for this because they possess a high degree of expressivity and can adaptively discover low-dimensional structures embedded within high-dimensional spaces. A foundational theoretical milestone was achieved by Julius Berner et al. in Analysis of the generalization error: Empirical risk minimization over deep artificial neural networks overcomes the curse of dimensionality in the numerical approximation of Black-Scholes partial differential equations. The authors provided a rigorous mathematical proof showing that deep neural networks trained via empirical risk minimization can approximate solutions to Black-Scholes and Kolmogorov equations without suffering from exponential complexity scaling.
To train these networks, researchers leverage variational principles or residual minimization:
- The Deep Ritz Method (DRM): This approach reformulates the PDE as an energy minimization problem. The network is trained by minimizing a variational functional over the domain. Yuling Jiao et al. provided a rigorous convergence analysis for this approach in Error Analysis of Three-Layer Neural Network Trained with PGD for Deep Ritz Method, proving that optimizing these networks with gradient descent reliably converges to the true physical solution.
- Physics-Informed Neural Networks (PINNs): This framework minimizes the residual of the PDE directly at a set of randomly sampled collocation points. Because the points are sampled randomly rather than on a fixed grid, the training cost scales gracefully with dimension.
Correcting Imperfections with Martingales
One practical challenge of using neural networks as PDE solvers is that optimization is rarely perfect; neural networks can get stuck in local minima or suffer from generalization errors. How can we trust their predictions in high-stakes environments like financial hedging?
An elegant solution is to use the neural network not as the final answer, but as a baseline to construct an unbiased estimator. In Unbiased Deep Solvers for Linear Parametric PDEs, Marc Sabaté Vidales et al. combined deep neural networks with the Martingale Representation Theorem.
The intuition is as follows: in a financial market or a diffusion process, any smooth function of the underlying assets can be decomposed into an expected value plus a martingale (a zero-mean noise term representing the fluctuations of the paths). By training a neural network to approximate both the solution and its gradient, the authors constructed a "control variate"—a variance reduction technique. Even if the neural network's approximation is slightly flawed, the remaining error can be integrated out via Monte Carlo simulations of the SDE, yielding a completely unbiased estimate of both the derivative price and its corresponding hedging strategy.
Exploiting Structural Separability
In certain domains, such as plasma physics and gas dynamics, PDEs govern particles across both spatial coordinates and velocity coordinates simultaneously (e.g., the Boltzmann or Vlasov-Poisson equations). This creates a massive 6D or 7D state space.
To make neural solvers even more efficient in these scenarios, researchers exploit the structural properties of the physical domain, such as coordinate separability.
Standard Neural Network:
(x, y, z, vx, vy, vz) ───► [ Deep MLP ] ───► Solution u
Separable Architecture (SPINN):
(x, y, z) ──────────────► [ Spatial MLP ] ───┐
├──► Tensor Product ──► Solution u
(vx, vy, vz) ───────────► [ Velocity MLP ] ──┘
In Separable Physics-informed Neural Networks for Solving the BGK Model of the Boltzmann Equation, Jaemin Oh et al. introduced a framework that decouples the spatial and velocity dimensions using tensor decompositions. Instead of passing all high-dimensional coordinates into a single massive network, they process spatial and velocity inputs through separate, lower-dimensional networks and combine their outputs via a tensor product. This drastically reduces the number of required collocation points, allowing complex 3D gas dynamics to be simulated efficiently on a single GPU. Similarly, Tunan Kao et al. developed pETNNs: Partial Evolutionary Tensor Neural Networks for Solving Time-Dependent Partial Differential Equations, which pairs tensor neural networks with evolutionary parameter updates to isolate and update only a subset of network parameters over time, slashing computational overhead while preserving high accuracy.
Summary
The transition from traditional grid-based numerical analysis to deep learning-based PDE solving relies on a simple, powerful idea: replacing deterministic discretization with probabilistic sampling and continuous optimization. By leveraging the dimension-independent convergence of Monte Carlo methods, the expressivity of deep neural networks, and structural priors like tensor separability, modern deep solvers have effectively turned the curse of dimensionality into a manageable computational problem.