C-PINN: A neural network framework based on the Cordès condition for solving linear and fully nonlinear equations in non-divergence form and its applications

1. Problem Statement

The paper addresses the numerical challenges associated with solving Partial Differential Equations (PDEs) in non-divergence form, specifically:

• Linear Non-Divergence Equations: Equations of the form $Lu = A(x):D^2u + b(x)\cdot\nabla u - c(x)u = f(x)$.
• Fully Nonlinear Equations: Including Hamilton-Jacobi-Bellman (HJB) equations (arising in stochastic control) and Monge-Ampère (MA) equations (central to optimal transport and geometry).

Key Challenges:

• Optimization Instability: Standard Physics-Informed Neural Networks (PINNs) struggle with these equations because the loss landscape is highly non-convex. Direct minimization of high-order derivatives (Hessians) leads to gradient pathologies, poor convergence, and low accuracy.
• Curse of Dimensionality: Traditional mesh-based methods (Finite Element/Difference) become computationally intractable in high dimensions, while standard PINNs often fail to converge in these settings for non-divergence operators.
• Lack of Variational Structure: Unlike divergence-form equations, non-divergence equations do not naturally possess a variational structure suitable for standard energy minimization approaches in deep learning.

2. Methodology: The C-PINN Framework

The authors propose C-PINN, a novel framework that integrates the Cordès condition from elliptic PDE theory into the neural network training process to precondition the optimization problem.

A. Theoretical Foundation: The Cordès Condition

The Cordès condition is a structural constraint on the coefficient matrix $A(x)$ that ensures the operator behaves similarly to the Laplacian, guaranteeing well-posedness and regularity.

• Condition: It restricts the anisotropy of the eigenvalues of $A(x)$, ensuring the ratio of the squared trace to the squared Frobenius norm is bounded:
  $$\frac{(\text{tr}(A))^2}{\|A\|_F^2} \geq d - 1 + \epsilon$$
  where $d$ is the spatial dimension and $\epsilon \in (0, 1]$.
• Preconditioning: By defining a spatially varying scaling function $\alpha(x) = \frac{\text{tr}(A)}{\|A\|_F^2}$, the original equation is transformed. Theoretically, this transforms the operator into a strict contraction mapping in the Sobolev space $H^2(\Omega)$.

B. Loss Function Reformulation

Instead of minimizing the raw PDE residual $|Lu - f|^2$, C-PINN minimizes a Cordès-weighted residual:
$$L_{\text{Cordès}}(\theta) = \frac{1}{N} \sum \left| \frac{\text{tr}(A)}{\text{tr}(A^2) + \delta} (L u_\theta - f) \right|^2$$

• Effect: This weighting smooths the loss landscape, eliminates extreme local curvatures, and prevents gradient explosions, effectively turning a non-convex problem into a well-conditioned one.

C. Extension to Nonlinear Equations (Dual-Loop Framework)

To handle fully nonlinear equations (HJB and MA), the authors introduce a Dual-Loop Newton Framework:

1. Outer Loop (Linearization):
   • HJB: Uses Semi-smooth Newton iteration. The non-differentiable $\sup/\inf$ operator is handled by identifying the active control branch $\alpha^*$, freezing the coefficients, and linearizing the equation.
   • MA: Uses Newton-Picard linearization. The determinant operator is linearized using the Fréchet derivative (cofactor matrix), converting the nonlinear equation into a sequence of linear non-divergence equations.
2. Inner Loop (C-PINN Solver):
   • The resulting linear surrogate equation is solved using the C-PINN framework with the Cordès preconditioner applied to the frozen coefficient matrix.

3. Key Contributions

1. Novel Loss Formulation: Introduced a mathematically rigorous preconditioning mechanism based on the Cordès condition that stabilizes the training of PINNs for non-divergence form PDEs.
2. Unified Framework for Nonlinear PDEs: Successfully extended the method to fully nonlinear HJB and MA equations via a dual-loop Newton linearization strategy, unifying the treatment of these distinct equation classes.
3. High-Dimensional Scalability: Demonstrated the ability to solve equations in dimensions up to 20D, effectively circumventing the curse of dimensionality that plagues mesh-based methods.
4. Optimal Transport Applications: Applied the framework to solve Monge-Ampère equations for optimal transport problems, including complex geometries (cortical surfaces, 3D models).

4. Numerical Results

The paper validates C-PINN through extensive experiments comparing it against standard PINNs and other benchmarks:

• Linear Non-Divergence Equations:
  • Accuracy: C-PINN achieved significantly lower $L_2$ and $L_\infty$ errors compared to standard PINNs across smooth, weakly singular, and discontinuous coefficient cases.
  • Stability: Optimization dynamics showed a smoother loss landscape and stable gradient norms, whereas standard PINNs exhibited chaotic oscillations.
• High-Dimensional Problems:
  • Successfully solved 3D, 5D, and 20D elliptic equations.
  • In the 20D case, C-PINN maintained stable convergence, while standard PINN performance degraded significantly.
• Nonlinear Equations (HJB & MA):
  • C-PINN solved HJB equations with high accuracy, outperforming PINN in both error metrics and convergence speed.
  • For the Monge-Ampère equation, the method accurately captured the convex solution and the associated transport maps.
• Optimal Transport:
  • Applied to square-to-square transport and complex 3D surface mappings (human cortex, lion head).
  • The method successfully corrected area distortions in conformal parameterizations by redistributing mass according to the target density.

5. Significance and Impact

• Bridging Theory and Practice: The work successfully bridges classical functional analysis (Cordès condition, contraction mapping) with modern deep learning, providing a theoretical guarantee for the stability of neural network solvers for a class of PDEs previously considered difficult.
• Robustness: The method is robust against discontinuous coefficients, weak singularities, and complex geometries.
• Generality: By reducing fully nonlinear problems to a sequence of preconditioned linear problems, the framework offers a generalizable approach for a wide range of challenging PDEs in stochastic control, geometry, and physics.
• Practical Utility: The ability to solve high-dimensional optimal transport problems opens new avenues for applications in generative modeling, domain adaptation, and resource allocation where traditional methods fail.

In conclusion, C-PINN represents a significant advancement in the field of scientific machine learning, offering a mathematically grounded solution to the instability issues inherent in solving non-divergence and fully nonlinear PDEs with neural networks.