Differential equations are the standard language for describing physical systems, but they are notoriously difficult to solve directly. Whether through numerical methods or machine learning, differentiation is a "roughening" operation: it amplifies noise, requires high-order smoothness, and is sensitive to local singularities.
A growing body of research suggests a more robust alternative: transforming differential problems into equivalent integral equations. By using tools like Green’s functions, boundary integral operators, and generalized inverses, researchers are shifting the computational burden from local derivatives to global integrals. This shift provides inherent stability, reduces dimensionality, and allows for the solving of equations that are otherwise mathematically "broken."
The Smoothing Power of Inversion
The core intuition behind this research direction is that integration is a smoothing operation. While estimating a high-order derivative from noisy data is numerically unstable, integrating that same data suppresses noise.
In Green’s matching, Jianbin Tan and colleagues demonstrate this by using Green’s functions to invert differential operators in complex dynamic systems. Instead of trying to calculate acceleration (a second derivative) from noisy position data, they transform the system into an integral form. This "first smooth, then estimate" approach avoids the errors inherent in high-order differentiation, achieving mathematically optimal accuracy.
Similarly, in the study of fractional differential equations, Bangti Jin and co-authors handle the "roughness" of non-integer derivatives by reformulating them into systems where a standard, smooth differential operator dominates, effectively treating the nonlocal fractional terms as lower-order integral components.
Dimensionality Reduction via Boundary Integrals
One of the most powerful applications of this transformation is the reduction of a problem's dimensionality. If a solution satisfies a linear differential equation within a volume, potential theory tells us that the entire solution can be represented as an integral over the boundary of that volume.
Guochang Lin and Jun Wang leverage this in BINet and BI-GreenNet. By representing the solution as a potential integral over the boundary, the neural network only needs to learn the "density" on the boundary manifold rather than the values across the entire 3D volume. This ensures the physics are satisfied "by construction" inside the domain, making the model far more data-efficient and robust than standard physics-informed neural networks (PINNs).
This boundary-centric view is also the key to the Quantum Query Complexity of Elliptic PDE. By reformulating PDEs as integral equations using Green's functions, quantum algorithms can approximate solutions at specific points or submanifolds with a significant speedup over classical methods, as the task becomes one of high-dimensional integration rather than grid-based differentiation.
Managing Non-invertibility and Mixed Derivatives
Not all differential operators are easily invertible. Equations with mixed time-space derivatives or periodic domains often involve operators with "null spaces"—constants that disappear during differentiation and thus cannot be easily recovered.
Shun Sato and Takayasu Matsuo address this by using Moore-Penrose generalized inverses. In their work on evolutionary equations with mixed derivatives, they show that even when a spatial derivative operator is not strictly invertible, a generalized inverse can transform the implicit evolution into a solvable integral form. This ensures that physical invariants (like conservation of mass) are preserved even when the numerical scheme steps through time.
Bypassing Theoretical Obstacles
Integral reformulations also allow researchers to sidestep classical mathematical limitations. For example, the "maximum principle" is a standard tool for proving the symmetry of solutions to second-order PDEs, but it often fails for higher-order or fractional equations.
In Symmetry of solutions to higher and fractional order semilinear equations, Guozhen Lu and Jianxiong Wang convert these complex differential equations into equivalent integral equations using Green's functions. This transformation allows them to apply the "moving plane method" directly to the integral form, establishing symmetry results that were previously unreachable through differential analysis alone.
Embedding Constraints into Neural Operators
The shift toward integral forms is culminating in new architectures for machine learning. Traditional neural networks often struggle to enforce boundary conditions, relying on "penalty terms" in the loss function that are difficult to tune.
The Fredholm Integral Equations Neural Operator (FIE-NO), developed by Haoyang Jiang, solves this by embedding the boundary constraints directly into the mathematical structure of the model. By reformulating the boundary value problem as a Fredholm integral equation, the network learns a mapping from boundary data to interior points that is mathematically guaranteed to respect the constraints, even on irregular geometries.
By treating differential equations not as local constraints to be satisfied, but as global integral operators to be inverted, this research direction provides a unified framework for stability, efficiency, and mathematical rigor across scientific computing and artificial intelligence.